Merge remote-tracking branch 'remotes/origin/master' into brl_reflect

appstub.mod/appstub.linux.c

@@ -1,4 +1,4 @@
-
+#define _GNU_SOURCE
 #include <brl.mod/blitz.mod/blitz.h>
 
 #include <signal.h>

appstub.mod/debugger.stdio.glue.c

@@ -30,7 +30,7 @@ unsigned int bmx_debugger_DebugDeclKind(struct BBDebugDecl * decl) {
 }
 
 struct BBDebugDecl * bmx_debugger_DebugDeclNext( struct BBDebugDecl * decl ) {
-	return ((char *)decl) + sizeof(struct BBDebugDecl);
+	return decl + 1;
 }
 
 void * bmx_debugger_DebugDecl_VarAddress( struct BBDebugDecl * decl ) {

blitz.mod/blitz.bmx

@@ -8,12 +8,14 @@ bbdoc: BASIC/BlitzMax runtime
 End Rem
 Module BRL.Blitz
 
-ModuleInfo "Version: 1.25"
+ModuleInfo "Version: 1.26"
 ModuleInfo "Author: Mark Sibly"
 ModuleInfo "License: zlib/libpng"
 ModuleInfo "Copyright: Blitz Research Ltd"
 ModuleInfo "Modserver: BRL"
 '
+ModuleInfo "History: 1.26"
+ModuleInfo "History: Added extended string to number conversion methods"
 ModuleInfo "History: 1.25"
 ModuleInfo "History: Added suport for risc-v"
 ModuleInfo "History: 1.24"
@@ -98,6 +100,9 @@ ModuleInfo "CC_OPTS: -DJAVA_FINALIZATION -DNO_EXECUTE_PERMISSION"
 ModuleInfo "CC_OPTS: -DBMX_DEBUG"
 ?
 
+' c++11 required for fast float
+ModuleInfo "CPP_OPTS: -std=c++11"
+
 ' uncomment to enable allocation counting
 'ModuleInfo "CC_OPTS: -DBBCC_ALLOCCOUNT"
 
@@ -118,6 +123,7 @@ Import "blitz_gc.c"
 Import "blitz_unicode.c"
 Import "blitz_enum.c"
 Import "blitz_coverage.c"
+Import "blitz_string_ex.cpp"
 
 ?coverage
 Import "hashmap/hashmap.c"
@@ -323,6 +329,47 @@ Function DebugLog( message:String )
 	OnDebugLog message
 End Function
 
+Rem
+bbdoc: Enables parsing of scientific notation in extended string to number conversion methods.
+End Rem
+Const CHARSFORMAT_SCIENTIFIC:ULong = 1 Shl 0
+Rem
+bbdoc: Enables parsing of fixed point notation in extended string to number conversion methods.
+End Rem
+Const CHARSFORMAT_FIXED:ULong = 1 Shl 2
+Rem
+bbdoc: Enables parsing of hexadecimal notation in extended string to number conversion methods.
+End Rem
+Const CHARSFORMAT_HEX:ULong = 1 Shl 3
+Rem
+bbdoc: Do not allow Infinity or NaN in extended string to number conversion methods.
+End Rem
+Const CHARSFORMAT_NOINFNAN:ULong = 1 Shl 4
+Rem
+bbdoc: Enforces JSON number format in extended string to number conversion methods.
+End Rem
+Const CHARSFORMAT_JSON:ULong = 1 Shl 5 | CHARSFORMAT_FIXED | CHARSFORMAT_SCIENTIFIC | CHARSFORMAT_NOINFNAN
+Rem
+bbdoc: Parses JSON number format or Infinity or NaN in extended string to number conversion methods.
+End Rem
+Const CHARSFORMAT_JSONORINFNAN:ULong = 1 Shl 5 | CHARSFORMAT_FIXED | CHARSFORMAT_SCIENTIFIC
+Rem
+bbdoc: Allows parsing of Fortran-style numbers in extended string to number conversion methods.
+End Rem
+Const CHARSFORMAT_FORTRAN:ULong = 1 Shl 6 | CHARSFORMAT_FIXED | CHARSFORMAT_SCIENTIFIC
+Rem
+bbdoc: The default format for extended string to number conversion methods.
+End Rem
+Const CHARSFORMAT_GENERAL:ULong = CHARSFORMAT_FIXED | CHARSFORMAT_SCIENTIFIC
+Rem
+bbdoc: Allows leading plus sign in extended string to number conversion methods.
+End Rem
+Const CHARSFORMAT_ALLOWLEADINGPLUS:ULong = 1 Shl 7
+Rem
+bbdoc: Enables skipping leading whitespace in extended string to number conversion methods.
+End Rem
+Const CHARSFORMAT_SKIPWHITESPACE:ULong = 1 Shl 8
+
 Extern
 
 Rem

blitz.mod/blitz_app.c

@@ -9,8 +9,6 @@ BBString*	bbAppTitle=BBNULLSTRING;
 BBString*	bbLaunchDir=BBNULLSTRING;
 BBArray*	bbAppArgs=BBNULLARRAY;
 
-void **bbGCStackTop;
-
 char * bbArgv0 = NULL;
 
 void bbEnd(){
@@ -467,15 +465,10 @@ void bbStartup( int argc,char *argv[],void *dummy1,void *dummy2 ){
 
 #elif __linux__
 
-	char *ebp;
 	char buf[PATH_MAX];
 	char lnk[PATH_MAX];
 	pid_t pid;
-	
-	// asm( "movl %%ebp,%0;":"=r"(ebp) );//::"%ebp" );
-	
-	bbGCStackTop=ebp+28;
-	
+		
 	bbThreadPreStartup();
 	bbGCStartup();
 	bbThreadStartup();
@@ -510,12 +503,6 @@ void bbStartup( int argc,char *argv[],void *dummy1,void *dummy2 ){
 	
 	CFURLRef url;
 	char buf[PATH_MAX],*e;
-	
-//#if BB_ARGP
-//	bbGCStackTop=bbArgp(0);
-//#else
-	bbGCStackTop=&argc;
-//#endif
 
 	bbGCStartup();
 	bbThreadPreStartup();

blitz.mod/blitz_app.h

@@ -20,8 +20,6 @@ extern BBString*	bbAppTitle;
 extern BBString*	bbLaunchDir;
 extern BBArray*	bbAppArgs;
 
-extern void**		bbGCStackTop;
-
 extern char * bbArgv0;
 
 void		bbEnd();

blitz.mod/blitz_array.c

@@ -92,7 +92,7 @@ static int arrayCellSize(const char * type, unsigned short data_size, int * flag
 	return size;
 }
 
-static BBArray *allocateArray( const char *type,int dims,int *lens, unsigned short data_size ){
+BBArray *bbAllocateArray( const char *type,int dims,int *lens, unsigned short data_size ){
 	int k,*len;
 	unsigned int size=4;
 	int length=1;
@@ -182,7 +182,7 @@ BBArray *bbArrayNew( const char *type,int dims,... ){
 	}
 	va_end(lengths);
 
-	BBArray *arr=allocateArray( type,dims, lens, 0 );
+	BBArray *arr=bbAllocateArray( type,dims, lens, 0 );
 	
 	initializeArray( arr, 0, 0 );
 	
@@ -203,7 +203,7 @@ BBArray *bbArrayNewStruct( const char *type, unsigned short data_size, BBArraySt
 	}
 	va_end(lengths);
 
-	BBArray *arr=allocateArray( type,dims, lens, data_size );
+	BBArray *arr=bbAllocateArray( type,dims, lens, data_size );
 	
 	initializeArray( arr, init, 0 );
 	
@@ -223,7 +223,7 @@ BBArray *bbArrayNewEnum( const char *type, BBEnum * bbEnum, int dims, ... ){
 	}
 	va_end(lengths);
 	
-	BBArray *arr=allocateArray( bbEnum->type,dims, lens, 0 );
+	BBArray *arr=bbAllocateArray( bbEnum->type,dims, lens, 0 );
 	arr->type=bbEnum->atype;
 	
 	initializeArray( arr, 0, bbEnum );
@@ -233,7 +233,7 @@ BBArray *bbArrayNewEnum( const char *type, BBEnum * bbEnum, int dims, ... ){
 
 BBArray *bbArrayNewEx( const char *type,int dims,int *lens ){
 
-	BBArray *arr=allocateArray( type,dims,lens,0 );
+	BBArray *arr=bbAllocateArray( type,dims,lens,0 );
 	
 	initializeArray( arr, 0, 0 );
 	
@@ -242,7 +242,7 @@ BBArray *bbArrayNewEx( const char *type,int dims,int *lens ){
 
 BBArray *bbArrayNew1D( const char *type,int length ){
 
-	BBArray *arr=allocateArray( type,1,&length, 0 );
+	BBArray *arr=bbAllocateArray( type,1,&length, 0 );
 	
 	initializeArray( arr, 0, 0 );
 	
@@ -250,12 +250,12 @@ BBArray *bbArrayNew1D( const char *type,int length ){
 }
 
 BBArray *bbArrayNew1DNoInit( const char *type,int length ){
-	return allocateArray( type,1,&length, 0 );
+	return bbAllocateArray( type,1,&length, 0 );
 }
 
 BBArray *bbArrayNew1DStruct( const char *type,int length, unsigned short data_size, BBArrayStructInit init ){
 
-	BBArray *arr=allocateArray( type,1,&length, data_size );
+	BBArray *arr=bbAllocateArray( type,1,&length, data_size );
 	
 	initializeArray( arr, init, 0 );
 	
@@ -264,7 +264,7 @@ BBArray *bbArrayNew1DStruct( const char *type,int length, unsigned short data_si
 
 BBArray *bbArrayNew1DEnum( const char *type,int length, BBEnum * bbEnum ){
 
-	BBArray *arr=allocateArray( bbEnum->type,1,&length, 0 );
+	BBArray *arr=bbAllocateArray( bbEnum->type,1,&length, 0 );
 	arr->type=bbEnum->atype;
 	
 	initializeArray( arr, 0, bbEnum );
@@ -285,7 +285,7 @@ BBArray *bbArraySliceStruct( const char *type,BBArray *inarr,int beg,int end, un
 
 	if( length<=0 ) return &bbEmptyArray;
 	
-	arr=allocateArray( type,1,&length,data_size );
+	arr=bbAllocateArray( type,1,&length,data_size );
 
 	el_size=arr->size/length;
 	
@@ -386,7 +386,7 @@ BBArray *bbArrayConcat( const char *type,BBArray *x,BBArray *y ){
 		brl_blitz_RuntimeError(bbStringFromCString("Incompatible array element types for concatenation"));
 	}
 
-	arr=allocateArray( type,1,&length, data_size );
+	arr=bbAllocateArray( type,1,&length, data_size );
 	
 	data=(char*)BBARRAYDATA( arr,1 );
 	
@@ -406,7 +406,7 @@ BBArray *bbArrayFromDataSize( const char *type,int length,void *data, unsigned s
 
 	if( length<=0 ) return &bbEmptyArray;
 	
-	arr=allocateArray( type,1,&length,data_size );
+	arr=bbAllocateArray( type,1,&length,data_size );
 
 	memcpy( BBARRAYDATA( arr,1 ),data,arr->size );
 
@@ -419,7 +419,7 @@ BBArray *bbArrayFromDataStruct( const char *type,int length,void *data, unsigned
 
 	if( length<=0 ) return &bbEmptyArray;
 	
-	arr=allocateArray( type,1,&length, data_size );
+	arr=bbAllocateArray( type,1,&length, data_size );
 
 	memcpy( BBARRAYDATA( arr,1 ),data,arr->size );
 

blitz.mod/blitz_array.h

@@ -15,6 +15,61 @@ extern "C"{
 #define BBARRAYDATA(p,n) ((void*)((char*)(p)+((BBArray*)(p))->data_start))
 #define BBARRAYDATAINDEX(p,n,i) bbArrayIndex(p,n,i)
 
+#define BBARRAYNEW1DSTRUCT_FUNC(FUNC_SUFFIX, STRUCT_TYPE, CONSTRUCTOR_FUNC, TYPE_STRING) \
+BBArray *bbArrayNew1DStruct_##FUNC_SUFFIX(int length) { \
+    BBArray *arr = bbAllocateArray(TYPE_STRING, 1, &length, sizeof(struct STRUCT_TYPE)); \
+    if (!arr->size) return arr; \
+    struct STRUCT_TYPE *p = (struct STRUCT_TYPE *)(BBARRAYDATA(arr, arr->dims)); \
+    memset(p, 0, arr->size); \
+    struct STRUCT_TYPE *s = p; \
+    for (int k = arr->scales[0]; k > 0; --k) { \
+        CONSTRUCTOR_FUNC(s); \
+        s++; \
+    } \
+    return arr; \
+}
+
+#define BBARRAYSLICESTRUCT_FUNC(FUNC_SUFFIX, STRUCT_TYPE, CONSTRUCTOR_FUNC, TYPE_STRING) \
+BBArray *bbArraySliceStruct_##FUNC_SUFFIX(BBArray *inarr, int beg, int end) { \
+    int k; \
+    int length = end - beg; \
+    if (length <= 0) return &bbEmptyArray; \
+    BBArray *arr = bbAllocateArray(TYPE_STRING, 1, &length, sizeof(struct STRUCT_TYPE)); \
+    int el_size = sizeof(struct STRUCT_TYPE); \
+    struct STRUCT_TYPE *p = (struct STRUCT_TYPE *)BBARRAYDATA(arr, 1); \
+    int n = -beg; \
+    if (n > 0) { \
+        if (beg + n > end) n = end - beg; \
+        memset(p, 0, n * el_size); \
+        struct STRUCT_TYPE *s = p; \
+        for (k = 0; k < n; ++k) { \
+            CONSTRUCTOR_FUNC(s); \
+            s++; \
+        } \
+        p += n; \
+        beg += n; \
+        if (beg == end) return arr; \
+    } \
+    n = inarr->scales[0] - beg; \
+    if (n > 0) { \
+        if (beg + n > end) n = end - beg; \
+        memcpy(p, (struct STRUCT_TYPE *)BBARRAYDATA(inarr, inarr->dims) + beg, n * el_size); \
+        p += n; \
+        beg += n; \
+        if (beg == end) return arr; \
+    } \
+    n = end - beg; \
+    if (n > 0) { \
+        memset(p, 0, n * el_size); \
+        struct STRUCT_TYPE *s = p; \
+        for (k = 0; k < n; ++k) { \
+            CONSTRUCTOR_FUNC(s); \
+            s++; \
+        } \
+    } \
+    return arr; \
+}
+
 struct BBArray{
 	//extends BBObject
 	BBClass*        clas;
@@ -87,6 +142,8 @@ void bbArrayCopy(BBArray * srcArr, int srcPos, BBArray * dstArr, int dstPos, int
 
 int bbObjectIsEmptyArray(BBObject * o);
 
+BBArray *bbAllocateArray( const char *type,int dims,int *lens, unsigned short data_size );
+
 #ifdef __cplusplus
 }
 #endif

blitz.mod/blitz_classes.i

@@ -15,7 +15,7 @@ String^Object{
 	@length:Int
 
 	-ToString:String()="bbStringToString"
-	-Compare:Int(otherObject:Object)="bbStringCompare"
+	-Compare:Int(otherString:String)="bbStringCompare"
 	
 	-Find:Int( subString:String,startIndex=0 )="bbStringFind"
 	-FindLast:Int( subString:String,startIndex=0 )="bbStringFindLast"
@@ -75,6 +75,16 @@ String^Object{
 	+FromLongInt:String( longIntValue:LongInt )="bbStringFromLongInt"
 	-ToULongInt:ULongInt()="bbStringToULongInt"
 	+FromULongInt:String( ulongIntValue:ULongInt )="bbStringFromULongInt"
+
+	-ToDoubleEx:Int( val:Double Var,startPos:Int=0,endPos:Int=-1,format:Int=5,sep:String="." )="bbStringToDoubleEx"
+	-ToFloatEx:Int( val:Float Var,startPos:Int=0,endPos:Int=-1,format:Int=5,sep:String="." )="bbStringToFloatEx"
+	-ToIntEx:Int( val:Int Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToIntEx"
+	-ToUIntEx:Int( val:UInt Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToUIntEx"
+	-ToLongEx:Int( val:Long Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToLongEx"
+	-ToULongEx:Int( val:ULong Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToULongEx"
+	-ToSizeTEx:Int( val:Size_T Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToSizeTEx"
+	-ToLongIntEx:Int( val:LongInt Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToLongIntEx"
+	-ToULongIntEx:Int( val:ULongInt Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToULongIntEx"
 }AF="bbStringClass"
 
 ___Array^Object{

blitz.mod/blitz_classes.win32.i

@@ -15,7 +15,7 @@ String^Object{
 	@length:Int
 
 	-ToString:String()="bbStringToString"
-	-Compare:Int(otherObject:Object)="bbStringCompare"
+	-Compare:Int(otherString:String)="bbStringCompare"
 	
 	-Find:Int( subString:String,startIndex=0 )="bbStringFind"
 	-FindLast:Int( subString:String,startIndex=0 )="bbStringFindLast"
@@ -80,6 +80,16 @@ String^Object{
 	+FromLongInt:String( longIntValue:LongInt )="bbStringFromLongInt"
 	-ToULongInt:ULongInt()="bbStringToULongInt"
 	+FromULongInt:String( ulongIntValue:ULongInt )="bbStringFromULongInt"
+
+	-ToDoubleEx:Int( val:Double Var,startPos:Int=0,endPos:Int=-1,format:Int=5,sep:String="." )="bbStringToDoubleEx"
+	-ToFloatEx:Int( val:Float Var,startPos:Int=0,endPos:Int=-1,format:Int=5,sep:String="." )="bbStringToFloatEx"
+	-ToIntEx:Int( val:Int Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToIntEx"
+	-ToUIntEx:Int( val:UInt Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToUIntEx"
+	-ToLongEx:Int( val:Long Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToLongEx"
+	-ToULongEx:Int( val:ULong Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToULongEx"
+	-ToSizeTEx:Int( val:Size_T Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToSizeTEx"
+	-ToLongIntEx:Int( val:LongInt Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToLongIntEx"
+	-ToULongIntEx:Int( val:ULongInt Var,startPos:Int=0,endPos:Int=-1,format:Int=5,base:Int=10 )="bbStringToULongIntEx"
 }AF="bbStringClass"
 
 ___Array^Object{

blitz.mod/blitz_gc.c

@@ -134,7 +134,7 @@ int bbGCValidate( void *q ){
 			}
 		}
 		// maybe an array?
-		if (clas == &bbArrayClass) {
+		if (clas == (BBClass *)&bbArrayClass) {
 			return 1;
 		}
 	}

blitz.mod/blitz_memory.c

@@ -4,27 +4,29 @@
 #define SIZEALIGN 16
 #define ALIGNMASK (SIZEALIGN-1)
 
-/* use malloc/free() in Debug mode, otherwise use the GC heap */
-void *bbMemAlloc( size_t size ){
-	void *p;
-#ifdef BMX_DEBUG
-	p=malloc( size );
-#else
-	p=GC_MALLOC_ATOMIC_UNCOLLECTABLE( size );
-	#ifdef BBCC_ALLOCCOUNT
-	++bbGCAllocCount;
-	#endif
-#endif
-	return p;
-	
+void *bbMemAlloc(size_t size) {
+    size_t totalSize = size + SIZEALIGN - 1 + sizeof(void*);
+    void *p = malloc(totalSize);
+    if (!p) {
+        GC_gcollect();
+        p = malloc(totalSize);
+        if (!p) return NULL;
+    }
+    
+    uintptr_t rawAddr = (uintptr_t)p + sizeof(void*);
+    uintptr_t alignedAddr = (rawAddr + SIZEALIGN - 1) & ~(uintptr_t)ALIGNMASK;
+    
+    // Store the original pointer just before the aligned memory.
+    ((void**)alignedAddr)[-1] = p;
+    return (void*)alignedAddr;
 }
 
-void bbMemFree( void *p ){
-#ifdef BMX_DEBUG
-	if ( p ) free(p);
-#else
-	if( p ) GC_free( p );
-#endif
+void bbMemFree(void *p) {
+    if (p) {
+        // Get the original pointer stored before the aligned block and free it.
+        void *original = ((void**)p)[-1];
+        free(original);
+    }
 }
 
 void *bbMemExtend( void *mem,size_t size,size_t new_size ){

blitz.mod/blitz_object.c

@@ -109,7 +109,7 @@ void bbObjectReserved(){
 }
 
 BBObject *bbObjectStringcast( BBObject *o ){
-	if (o->clas == &bbStringClass) {
+	if (o->clas == (BBClass *)&bbStringClass) {
 		return o;
 	} else {
 		return (BBObject *)&bbEmptyString;
@@ -117,11 +117,11 @@ BBObject *bbObjectStringcast( BBObject *o ){
 }
 
 int bbObjectIsString( BBObject *o ){
-	return o->clas == &bbStringClass;
+	return o->clas == (BBClass *)&bbStringClass;
 }
 
 BBObject *bbObjectArraycast( BBObject *o ){
-	if (o->clas == &bbArrayClass) {
+	if (o->clas == (BBClass *)&bbArrayClass) {
 		return o;
 	} else {
 		return (BBObject *)&bbEmptyArray;
@@ -129,13 +129,13 @@ BBObject *bbObjectArraycast( BBObject *o ){
 }
 
 int bbObjectIsArray( BBObject *o ){
-	return o->clas == &bbArrayClass;
+	return o->clas == (BBClass *)&bbArrayClass;
 }
 
 BBObject *bbObjectDowncast( BBObject *o,BBClass *t ){
 	BBClass *p=o->clas;
 	while( p && p!=t ) p=p->super;
-	return p ? o : (t==&bbStringClass) ? (BBObject *)&bbEmptyString : (t==&bbArrayClass) ? (BBObject *)&bbEmptyArray : &bbNullObject;
+	return p ? o : (t==(BBClass *)&bbStringClass) ? (BBObject *)&bbEmptyString : (t==(BBClass *)&bbArrayClass) ? (BBObject *)&bbEmptyArray : &bbNullObject;
 }
 
 void bbObjectRegisterType( BBClass *clas ){

blitz.mod/blitz_string.c

@@ -8,7 +8,7 @@
 #define XXH_IMPLEMENTATION
 #define XXH_STATIC_LINKING_ONLY
 
-#include "hash/xxh3.h"
+#include "hash/xxhash.h"
 
 static void bbStringFree( BBObject *o );
 
@@ -316,49 +316,105 @@ BBString *bbStringFromUTF8String( const unsigned char *p ){
 	return p ? bbStringFromUTF8Bytes( p,strlen((char*)p) ) : &bbEmptyString;
 }
 
-BBString *bbStringFromUTF8Bytes( const unsigned char *p,int n ){
-	int c;
-	unsigned short *d,*q;
-	BBString *str;
-
-	if( !p || n <= 0 ) return &bbEmptyString;
-	
-	d=(unsigned short*)malloc( n*2 );
-	q=d;
-	
-	while( n-- && (c=*p++ & 0xff)){
-		if( c<0x80 ){
-			*q++=c;
-		}else{
-			if (!n--) break;
-			int d=*p++ & 0x3f;
-			if( c<0xe0 ){
-				*q++=((c&31)<<6) | d;
-			}else{
-				if (!n--) break;
-				int e=*p++ & 0x3f;
-				if( c<0xf0 ){
-					*q++=((c&15)<<12) | (d<<6) | e;
-				}else{
-					if (!n--) break;
-					int f=*p++ & 0x3f;
-					int v=((c&7)<<18) | (d<<12) | (e<<6) | f;
-					if( v & 0xffff0000 ) {
-						v -= 0x10000;
-						d = ((v >> 10) & 0x7ff) + 0xd800;
-						e = (v & 0x3ff) + 0xdc00;
-						*q++=d;
-						*q++=e;
-					}else{
-						*q++=v;
-					}
-				}
-			}
-		}
-	}
-	str=bbStringFromShorts( d,q-d );
-	free( d );
-	return str;
+#define REPLACEMENT_CHAR 0xFFFD
+
+BBString *bbStringFromUTF8Bytes(const unsigned char *p, int n) {
+    if (!p || n <= 0) return &bbEmptyString;
+
+    // Allocate worst-case: one output code unit per input byte.
+    unsigned short *buffer = (unsigned short*)malloc(n * sizeof(unsigned short));
+    if (!buffer) return &bbEmptyString; // Allocation failed
+
+    unsigned short *dest = buffer;
+    const unsigned char *end = p + n;
+
+    while (p < end) {
+        unsigned int codepoint;
+        unsigned char byte = *p++;
+
+        if (byte < 0x80) {
+            // 1-byte (ASCII)
+            *dest++ = byte;
+        } else if (byte < 0xC0) {
+            // Unexpected continuation byte; insert replacement.
+            *dest++ = REPLACEMENT_CHAR;
+        } else if (byte < 0xE0) {
+            // 2-byte sequence: 110xxxxx 10xxxxxx
+            if (p >= end) {
+                *dest++ = REPLACEMENT_CHAR;
+                break;
+            }
+            unsigned char byte2 = *p++;
+            if ((byte2 & 0xC0) != 0x80) {
+                *dest++ = REPLACEMENT_CHAR;
+                continue;
+            }
+            codepoint = ((byte & 0x1F) << 6) | (byte2 & 0x3F);
+            if (codepoint < 0x80) { // Overlong encoding
+                *dest++ = REPLACEMENT_CHAR;
+            } else {
+                *dest++ = (unsigned short)codepoint;
+            }
+        } else if (byte < 0xF0) {
+            // 3-byte sequence: 1110xxxx 10xxxxxx 10xxxxxx
+            if (p + 1 >= end) {
+                *dest++ = REPLACEMENT_CHAR;
+                break;
+            }
+            unsigned char byte2 = *p++;
+            unsigned char byte3 = *p++;
+            if ((byte2 & 0xC0) != 0x80 || (byte3 & 0xC0) != 0x80) {
+                *dest++ = REPLACEMENT_CHAR;
+                continue;
+            }
+            codepoint = ((byte & 0x0F) << 12) |
+                        ((byte2 & 0x3F) << 6) |
+                        (byte3 & 0x3F);
+            // Reject overlong sequences and surrogate halves.
+            if (codepoint < 0x800 || (codepoint >= 0xD800 && codepoint <= 0xDFFF)) {
+                *dest++ = REPLACEMENT_CHAR;
+            } else {
+                *dest++ = (unsigned short)codepoint;
+            }
+        } else if (byte < 0xF8) {
+            // 4-byte sequence: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+            if (p + 2 >= end) {
+                *dest++ = REPLACEMENT_CHAR;
+                break;
+            }
+            unsigned char byte2 = *p++;
+            unsigned char byte3 = *p++;
+            unsigned char byte4 = *p++;
+            if ((byte2 & 0xC0) != 0x80 ||
+                (byte3 & 0xC0) != 0x80 ||
+                (byte4 & 0xC0) != 0x80) {
+                *dest++ = REPLACEMENT_CHAR;
+                continue;
+            }
+            codepoint = ((byte & 0x07) << 18) |
+                        ((byte2 & 0x3F) << 12) |
+                        ((byte3 & 0x3F) << 6) |
+                        (byte4 & 0x3F);
+            // Ensure codepoint is within valid range.
+            if (codepoint < 0x10000 || codepoint > 0x10FFFF) {
+                *dest++ = REPLACEMENT_CHAR;
+            } else {
+                // Convert to surrogate pair.
+                codepoint -= 0x10000;
+                unsigned short highSurrogate = 0xD800 | ((codepoint >> 10) & 0x3FF);
+                unsigned short lowSurrogate  = 0xDC00 | (codepoint & 0x3FF);
+                *dest++ = highSurrogate;
+                *dest++ = lowSurrogate;
+            }
+        } else {
+            // Bytes above 0xF7 are invalid in modern UTF-8.
+            *dest++ = REPLACEMENT_CHAR;
+        }
+    }
+
+    BBString *str = bbStringFromShorts(buffer, dest - buffer);
+    free(buffer);
+    return str;
 }
 
 BBString *bbStringToString( BBString *t ){
@@ -367,7 +423,7 @@ BBString *bbStringToString( BBString *t ){
 
 int bbStringCompare( BBString *x,BBString *y ){
 	int k,n,sz;
-	if (x->clas != &bbStringClass || y->clas != &bbStringClass) return -1; // only compare strings with strings
+	if (x->clas != (BBClass*)&bbStringClass || y->clas != (BBClass*)&bbStringClass) return -1; // only compare strings with strings
 
 	sz=x->length<y->length ? x->length : y->length;
 	if (x->length == y->length && x->hash) {
@@ -853,13 +909,37 @@ BBString *bbStringToLower( BBString *str ){
 	int n = 0;
 	
 	while (n < str->length) {
-		int c=str->buf[n];
-		// ascii upper or other unicode char
-		if (c >= 192 || (c>='A' && c<='Z')) {
-			break;
-		}
-		++n;
-	}
+        int c = str->buf[n];
+        if (c < 192) {
+            // ASCII character
+            if (c >= 'A' && c <= 'Z') {
+                // Found an uppercase ASCII character
+                break;
+            }
+        } else {
+            // Unicode character
+            // Check if the character is an uppercase Unicode character
+            int lo = 0, hi = (3828 / 4) - 1; // sizeof(bbToLowerData) = 3828
+            int is_upper = 0;
+            while (lo <= hi) {
+                int mid = (lo + hi) / 2;
+                int upper = bbToLowerData[mid * 2];
+                if (c < upper) {
+                    hi = mid - 1;
+                } else if (c > upper) {
+                    lo = mid + 1;
+                } else {
+                    // Found an uppercase Unicode character
+                    is_upper = 1;
+                    break;
+                }
+            }
+            if (is_upper) {
+                break;
+            }
+        }
+        ++n;
+    }
 	
 	if (n == str->length) {
 		return str;
@@ -900,13 +980,37 @@ BBString *bbStringToUpper( BBString *str ){
 	int n = 0;
 	
 	while (n < str->length) {
-		int c=str->buf[n];
-		// ascii lower or other unicode char
-		if (c >= 181 || (c>='a' && c<='z')) {
-			break;
-		}
-		++n;
-	}
+        int c = str->buf[n];
+        if (c < 181) {
+            // ASCII character
+            if (c >= 'a' && c <= 'z') {
+                // Found a lowercase ASCII character
+                break;
+            }
+        } else {
+            // Unicode character
+            // Check if the character is a lowercase Unicode character
+            int lo = 0, hi = (3860 / 4) - 1; // sizeof(bbToUpperData) = 3860
+            int is_lower = 0;
+            while (lo <= hi) {
+                int mid = (lo + hi) / 2;
+                int lower = bbToUpperData[mid * 2];
+                if (c < lower) {
+                    hi = mid - 1;
+                } else if (c > lower) {
+                    lo = mid + 1;
+                } else {
+                    // Found a lowercase Unicode character
+                    is_lower = 1;
+                    break;
+                }
+            }
+            if (is_lower) {
+                break;
+            }
+        }
+        ++n;
+    }
 	
 	if (n == str->length) {
 		return str;
@@ -989,7 +1093,7 @@ unsigned char *bbStringToUTF8StringBuffer( BBString *str, unsigned char * buf, s
 				++i;
 			}
 		}
-		int n = q - buf;
+		size_t n = q - buf;
 		if( c<0x80 ){
 			if (buflen <= n+1) break;
 			*q++=c;

blitz.mod/blitz_string.h

@@ -104,6 +104,16 @@ struct BBClass_String{
 	BBString* (*bbStringFromLongInt)( BBLONGINT n );
 	BBULONGINT (*bbStringToULongInt)( BBString *t );
 	BBString* (*bbStringFromULongInt)( BBULONGINT n );
+
+	int (*bbStringToDoubleEx)( BBString *str, double *val, int start, int end, BBULONG format, BBString *sep );
+	int (*bbStringToFloatEx)( BBString *str, float *val, int start, int end, BBULONG format, BBString *sep );
+	int (*bbStringToIntEx)( BBString *str, int *val, int start, int end, BBULONG format, int base );
+	int (*bbStringToUIntEx)( BBString *str, unsigned int *val, int start, int end, BBULONG format, int base );
+	int (*bbStringToLongEx)( BBString *str, BBInt64 *val, int start, int end, BBULONG format, int base );
+	int (*bbStringToULongEx)( BBString *str, BBUInt64 *val, int start, int end, BBULONG format, int base );
+	int (*bbStringToSizeTEx)( BBString *str, BBSIZET *val, int start, int end, BBULONG format, int base );
+	int (*bbStringToLongIntEx)( BBString *str, BBLONGINT *val, int start, int end, BBULONG format, int base );
+	int (*bbStringToULongIntEx)( BBString *str, BBULONGINT *val, int start, int end, BBULONG format, int base );
 };
 
 extern	struct BBClass_String bbStringClass;
@@ -162,6 +172,16 @@ unsigned char* bbStringToUTF8String( BBString *str );
 BBLONGINT bbStringToLongInt( BBString *str );
 BBULONGINT bbStringToULongInt( BBString *str );
 
+int	bbStringToDoubleEx( BBString *str, double *val, int start, int end, BBULONG format, BBString *sep );
+int	bbStringToFloatEx( BBString *str, float *val, int start, int end, BBULONG format, BBString *sep );
+int	bbStringToIntEx( BBString *str, int *val, int start, int end, BBULONG format, int base );
+int	bbStringToUIntEx( BBString *str, unsigned int *val, int start, int end, BBULONG format, int base );
+int	bbStringToLongEx( BBString *str, BBInt64 *val, int start, int end, BBULONG format, int base );
+int	bbStringToULongEx( BBString *str, BBUInt64 *val, int start, int end, BBULONG format, int base );
+int	bbStringToSizeTEx( BBString *str, BBSIZET *val, int start, int end, BBULONG format, int base );
+int	bbStringToLongIntEx( BBString *str, BBLONGINT *val, int start, int end, BBULONG format, int base );
+int	bbStringToULongIntEx( BBString *str, BBULONGINT *val, int start, int end, BBULONG format, int base );
+
 BBUINT* bbStringToUTF32String( BBString *str );
 BBString* bbStringFromUTF32String( const BBUINT *p );
 BBString* bbStringFromUTF32Bytes( const BBUINT *p, int n );
@@ -188,7 +208,7 @@ inline BBULONG bbStringHash( BBString * x ) {
 }
 
 inline int bbStringEquals( BBString *x,BBString *y ){
-	if (x->clas != &bbStringClass || y->clas != &bbStringClass) return 0; // only strings with strings
+	if (x->clas != (BBClass *)&bbStringClass || y->clas != (BBClass *)&bbStringClass) return 0; // only strings with strings
 
 	if (x->length-y->length != 0) return 0;
 	if (x->hash != 0 ) {

blitz.mod/blitz_string_ex.cpp

@@ -0,0 +1,272 @@
+/*
+   Copyright 2024 Bruce A Henderson
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+*/
+#include "fast_float/fast_float.h"
+#include "blitz_debug.h"
+#ifdef _WIN32
+#if defined(_WIN64)
+ typedef __int64 LONG_PTR; 
+ typedef unsigned __int64 UINT_PTR;
+#else
+ typedef long LONG_PTR;
+ typedef unsigned int UINT_PTR;
+#endif
+typedef UINT_PTR WPARAM;
+typedef LONG_PTR LPARAM;
+#endif
+#include "blitz_string.h"
+
+// extracts a double from a string, from the range startPos to endPos
+// endPos of -1 means the end of the string
+// returns 0 if the string is not a valid double, or the position of the first character after the double otherwise
+int bbStringToDoubleEx( BBString *str, double * val, int startPos, int endPos, BBULONG format, BBString* sep ) {
+    if ( startPos < 0 || startPos >= str->length || endPos < -1 || endPos > str->length ) {
+        return 0;
+    }
+    if (endPos == -1) {
+        endPos = str->length;
+    }
+    const char16_t * start = (char16_t*)str->buf;
+    const char16_t * end = start + str->length;
+    const char16_t * p = start + startPos;
+    const char16_t * e = start + endPos;
+    const char16_t sepChar = sep->length > 0 ? sep->buf[0] : '.';
+    double result;
+
+    if ( sepChar != 0 && sepChar != '.' ) {
+        fast_float::parse_options_t<char16_t> options{static_cast<fast_float::chars_format>(format), sepChar};
+        fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars_advanced(p, e, result, options);
+        if (res.ptr != nullptr) {
+            *val = result;
+            return res.ptr - start;
+        }
+    }
+    else {
+        fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars(p, e, result, static_cast<fast_float::chars_format>(format));
+        if (res.ptr != nullptr) {
+            *val = result;
+            return res.ptr - start;
+        }
+    }
+    return 0;
+}
+
+// extracts a float from a string, from the range startPos to endPos
+// endPos of -1 means the end of the string
+// returns 0 if the string is not a valid float, or the position of the first character after the float otherwise
+int bbStringToFloatEx( BBString *str, float * val, int startPos, int endPos, BBULONG format, BBString* sep ) {
+    if ( startPos < 0 || startPos >= str->length || endPos < -1 || endPos > str->length ) {
+        return 0;
+    }
+    if (endPos == -1) {
+        endPos = str->length;
+    }
+    const char16_t * start = (char16_t*)str->buf;
+    const char16_t * end = start + str->length;
+    const char16_t * p = start + startPos;
+    const char16_t * e = start + endPos;
+    const char16_t sepChar = sep->length > 0 ? sep->buf[0] : '.';
+    float result;
+
+    if ( sepChar != 0 && sepChar != '.' ) {
+        fast_float::parse_options_t<char16_t> options{static_cast<fast_float::chars_format>(format), sepChar};
+        fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars_advanced(p, e, result, options);
+        if (res.ptr != nullptr) {
+            *val = result;
+            return res.ptr - start;
+        }
+    }
+    else {
+        fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars(p, e, result, static_cast<fast_float::chars_format>(format));
+        if (res.ptr != nullptr) {
+            *val = result;
+            return res.ptr - start;
+        }
+    }
+    return 0;
+}
+
+// extracts a int from a string, from the range startPos to endPos
+// endPos of -1 means the end of the string
+// returns 0 if the string is not a valid int, or the position of the first character after the int otherwise
+int bbStringToIntEx( BBString *str, int * val, int startPos, int endPos, BBULONG format, int base ) {
+    if ( startPos < 0 || startPos >= str->length || endPos < -1 || endPos > str->length ) {
+        return 0;
+    }
+    if (endPos == -1) {
+        endPos = str->length;
+    }
+    const char16_t * start = (char16_t*)str->buf;
+    const char16_t * end = start + str->length;
+    const char16_t * p = start + startPos;
+    const char16_t * e = start + endPos;
+    int result;
+
+    fast_float::parse_options_t<char16_t> options{static_cast<fast_float::chars_format>(format), '.', base};
+    fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars_advanced(p, e, result, options);
+    if (res.ptr != nullptr) {
+        *val = result;
+        return res.ptr - start;
+    }
+    return 0;
+}
+
+// extracts a UInt from a string, from the range startPos to endPos
+// endPos of -1 means the end of the string
+// returns 0 if the string is not a valid UInt, or the position of the first character after the UInt otherwise
+int bbStringToUIntEx( BBString *str, unsigned int * val, int startPos, int endPos, BBULONG format, int base ) {
+    if ( startPos < 0 || startPos >= str->length || endPos < -1 || endPos > str->length ) {
+        return 0;
+    }
+    if (endPos == -1) {
+        endPos = str->length;
+    }
+    const char16_t * start = (char16_t*)str->buf;
+    const char16_t * end = start + str->length;
+    const char16_t * p = start + startPos;
+    const char16_t * e = start + endPos;
+    unsigned int result;
+
+    fast_float::parse_options_t<char16_t> options{static_cast<fast_float::chars_format>(format), '.', base};
+    fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars_advanced(p, e, result, options);
+    if (res.ptr != nullptr) {
+        *val = result;
+        return res.ptr - start;
+    }
+    return 0;
+}
+
+// extracts a Long from a string, from the range startPos to endPos
+// endPos of -1 means the end of the string
+// returns 0 if the string is not a valid Long, or the position of the first character after the Long otherwise
+int bbStringToLongEx( BBString *str, BBInt64 * val, int startPos, int endPos, BBULONG format, int base ) {
+    if ( startPos < 0 || startPos >= str->length || endPos < -1 || endPos > str->length ) {
+        return 0;
+    }
+    if (endPos == -1) {
+        endPos = str->length;
+    }
+    const char16_t * start = (char16_t*)str->buf;
+    const char16_t * end = start + str->length;
+    const char16_t * p = start + startPos;
+    const char16_t * e = start + endPos;
+    BBInt64 result;
+
+    fast_float::parse_options_t<char16_t> options{static_cast<fast_float::chars_format>(format), '.', base};
+    fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars_advanced(p, e, result, options);
+    if (res.ptr != nullptr) {
+        *val = result;
+        return res.ptr - start;
+    }
+    return 0;
+}
+
+// extracts a ULong from a string, from the range startPos to endPos
+// endPos of -1 means the end of the string
+// returns 0 if the string is not a valid ULong, or the position of the first character after the ULong otherwise
+int bbStringToULongEx( BBString *str, BBUInt64 * val, int startPos, int endPos, BBULONG format, int base ) {
+    if ( startPos < 0 || startPos >= str->length || endPos < -1 || endPos > str->length ) {
+        return 0;
+    }
+    if (endPos == -1) {
+        endPos = str->length;
+    }
+    const char16_t * start = (char16_t*)str->buf;
+    const char16_t * end = start + str->length;
+    const char16_t * p = start + startPos;
+    const char16_t * e = start + endPos;
+    BBUInt64 result;
+
+    fast_float::parse_options_t<char16_t> options{static_cast<fast_float::chars_format>(format), '.', base};
+    fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars_advanced(p, e, result, options);
+    if (res.ptr != nullptr) {
+        *val = result;
+        return res.ptr - start;
+    }
+    return 0;
+}
+
+// extracts a Size_T from a string, from the range startPos to endPos
+// endPos of -1 means the end of the string
+// returns 0 if the string is not a valid Size_T, or the position of the first character after the Size_T otherwise
+int bbStringToSizeTEx( BBString *str, BBSIZET * val, int startPos, int endPos, BBULONG format, int base ) {
+    if ( startPos < 0 || startPos >= str->length || endPos < -1 || endPos > str->length ) {
+        return 0;
+    }
+    if (endPos == -1) {
+        endPos = str->length;
+    }
+    const char16_t * start = (char16_t*)str->buf;
+    const char16_t * end = start + str->length;
+    const char16_t * p = start + startPos;
+    const char16_t * e = start + endPos;
+    BBSIZET result;
+
+    fast_float::parse_options_t<char16_t> options{static_cast<fast_float::chars_format>(format), '.', base};
+    fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars_advanced(p, e, result, options);
+    if (res.ptr != nullptr) {
+        *val = result;
+        return res.ptr - start;
+    }
+    return 0;
+}
+
+// extracts a LongInt from a string, from the range startPos to endPos
+// endPos of -1 means the end of the string
+// returns -1 if the string is not a valid LongInt, or the position of the first character after the LongInt otherwise
+int bbStringToLongIntEx( BBString *str, BBLONGINT * val, int startPos, int endPos, BBULONG format, int base ) {
+    if ( startPos < 0 || startPos >= str->length || endPos < -1 || endPos > str->length ) {
+        return 0;
+    }
+    if (endPos == -1) {
+        endPos = str->length;
+    }
+    const char16_t * start = (char16_t*)str->buf;
+    const char16_t * end = start + str->length;
+    const char16_t * p = start + startPos;
+    const char16_t * e = start + endPos;
+    BBLONGINT result;
+
+    fast_float::parse_options_t<char16_t> options{static_cast<fast_float::chars_format>(format), '.', base};
+    fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars_advanced(p, e, result, options);
+    if (res.ptr != nullptr) {
+        *val = result;
+        return res.ptr - start;
+    }
+    return 0;
+}
+
+// extracts a ULongInt from a string, from the range startPos to endPos
+// endPos of -1 means the end of the string
+// returns 0 if the string is not a valid ULongInt, or the position of the first character after the ULongInt otherwise
+int bbStringToULongIntEx( BBString *str, BBULONGINT * val, int startPos, int endPos, BBULONG format, int base ) {
+    if ( startPos < 0 || startPos >= str->length || endPos < -1 || endPos > str->length ) {
+        return 0;
+    }
+    if (endPos == -1) {
+        endPos = str->length;
+    }
+    const char16_t * start = (char16_t*)str->buf;
+    const char16_t * end = start + str->length;
+    const char16_t * p = start + startPos;
+    const char16_t * e = start + endPos;
+    BBULONGINT result;
+
+    fast_float::parse_options_t<char16_t> options{static_cast<fast_float::chars_format>(format), '.', base};
+    fast_float::from_chars_result_t<char16_t> res = fast_float::from_chars_advanced(p, e, result, options);
+    if (res.ptr != nullptr) {
+        *val = result;
+        return res.ptr - start;
+    }
+    return 0;
+}

blitz.mod/blitz_thread.c

@@ -186,7 +186,7 @@ int bbThreadResume( BBThread *thread ){
 	return ResumeThread( thread->handle );
 }
 
-BBThread *bbThreadRegister( DWORD id ) {
+BBThread *bbThreadRegister( bb_thread_t id ) {
 
 	GC_call_with_stack_base(bbRegisterGCThread, NULL);
 
@@ -364,7 +364,7 @@ static void *threadProc( void *p ){
 
 	GC_call_with_stack_base(bbRegisterGCThread, NULL);
 
-	BBThread *thread=p;
+	BBThread *thread=(BBThread *)p;
 	
 	pthread_setspecific( curThreadTls,thread );
 	
@@ -406,7 +406,7 @@ BBThread *bbThreadCreate( BBThreadProc proc,BBObject *data ){
 	return 0;
 }
 
-BBThread *bbThreadRegister( void * thd ) {
+BBThread *bbThreadRegister( bb_thread_t thd ) {
 
 	GC_call_with_stack_base(bbRegisterGCThread, NULL);
 
@@ -444,7 +444,7 @@ void bbThreadDetach( BBThread *thread ){
 BBObject *bbThreadWait( BBThread *thread ){
 	BBObject *p=0;
 	thread->detached=1;
-	pthread_join( thread->handle,&p );
+	pthread_join( thread->handle,(void**)&p );
 	return p;
 }
 

blitz.mod/blitz_thread.h

@@ -9,6 +9,7 @@ extern "C"{
 #ifdef _WIN32
 
 #include <windows.h>
+typedef DWORD bb_thread_t;
 typedef CRITICAL_SECTION bb_mutex_t;
 #define bb_mutex_init(MUTPTR) (InitializeCriticalSection(MUTPTR),1)
 #define bb_mutex_destroy(MUTPTR) DeleteCriticalSection(MUTPTR)
@@ -37,6 +38,7 @@ typedef HANDLE bb_sem_t;
 #include<switch/kernel/semaphore.h>
 #include <threads.h>
 
+typedef thrd_t bb_thread_t;
 typedef mtx_t bb_mutex_t;
 #define bb_mutex_init(MUTPTR) (mtx_init(MUTPTR,mtx_recursive),1)
 #define bb_mutex_destroy(MUTPTR)
@@ -53,6 +55,7 @@ typedef Semaphore bb_sem_t;
 #else
 
 #include <pthread.h>
+typedef pthread_t bb_thread_t;
 typedef pthread_mutex_t bb_mutex_t;
 extern pthread_mutexattr_t _bb_mutexattr;
 #define bb_mutex_init(MUTPTR) (pthread_mutex_init((MUTPTR),&_bb_mutexattr)>=0)
@@ -107,11 +110,9 @@ struct BBThread{
 #ifdef _WIN32
 	BBObject * result;
 	HANDLE handle;
-	DWORD id;
-#elif __SWITCH__
-	thrd_t handle;
+	bb_thread_t id;
 #else
-	pthread_t handle;
+	bb_thread_t handle;
 #endif
 };
 
@@ -134,11 +135,7 @@ BBObject*		bbThreadGetData( int index );
 int			bbAtomicCAS( volatile int *target,int oldVal,int newVal );
 int			bbAtomicAdd( volatile int *target,int incr );
 
-#ifdef _WIN32
-BBThread *bbThreadRegister( DWORD id );
-#else
-BBThread *bbThreadRegister( void * thd );
-#endif
+BBThread *bbThreadRegister( bb_thread_t id );
 void bbThreadUnregister( BBThread * thread );
 
 

blitz.mod/fast_float/fast_float.h

@@ -0,0 +1,4025 @@
+// fast_float by Daniel Lemire
+// fast_float by João Paulo Magalhaes
+//
+//
+// with contributions from Eugene Golushkov
+// with contributions from Maksim Kita
+// with contributions from Marcin Wojdyr
+// with contributions from Neal Richardson
+// with contributions from Tim Paine
+// with contributions from Fabio Pellacini
+// with contributions from Lénárd Szolnoki
+// with contributions from Jan Pharago
+// with contributions from Maya Warrier
+// with contributions from Taha Khokhar
+// with contributions from Anders Dalvander
+//
+//
+// Licensed under the Apache License, Version 2.0, or the
+// MIT License or the Boost License. This file may not be copied,
+// modified, or distributed except according to those terms.
+//
+// MIT License Notice
+//
+//    MIT License
+//
+//    Copyright (c) 2021 The fast_float authors
+//
+//    Permission is hereby granted, free of charge, to any
+//    person obtaining a copy of this software and associated
+//    documentation files (the "Software"), to deal in the
+//    Software without restriction, including without
+//    limitation the rights to use, copy, modify, merge,
+//    publish, distribute, sublicense, and/or sell copies of
+//    the Software, and to permit persons to whom the Software
+//    is furnished to do so, subject to the following
+//    conditions:
+//
+//    The above copyright notice and this permission notice
+//    shall be included in all copies or substantial portions
+//    of the Software.
+//
+//    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF
+//    ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED
+//    TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A
+//    PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT
+//    SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+//    CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+//    OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR
+//    IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+//    DEALINGS IN THE SOFTWARE.
+//
+// Apache License (Version 2.0) Notice
+//
+//    Copyright 2021 The fast_float authors
+//    Licensed under the Apache License, Version 2.0 (the "License");
+//    you may not use this file except in compliance with the License.
+//    You may obtain a copy of the License at
+//
+//    http://www.apache.org/licenses/LICENSE-2.0
+//
+//    Unless required by applicable law or agreed to in writing, software
+//    distributed under the License is distributed on an "AS IS" BASIS,
+//    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+//    See the License for the specific language governing permissions and
+//
+// BOOST License Notice
+//
+//    Boost Software License - Version 1.0 - August 17th, 2003
+//
+//    Permission is hereby granted, free of charge, to any person or organization
+//    obtaining a copy of the software and accompanying documentation covered by
+//    this license (the "Software") to use, reproduce, display, distribute,
+//    execute, and transmit the Software, and to prepare derivative works of the
+//    Software, and to permit third-parties to whom the Software is furnished to
+//    do so, all subject to the following:
+//
+//    The copyright notices in the Software and this entire statement, including
+//    the above license grant, this restriction and the following disclaimer,
+//    must be included in all copies of the Software, in whole or in part, and
+//    all derivative works of the Software, unless such copies or derivative
+//    works are solely in the form of machine-executable object code generated by
+//    a source language processor.
+//
+//    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+//    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+//    FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT
+//    SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE
+//    FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE,
+//    ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+//    DEALINGS IN THE SOFTWARE.
+//
+
+#ifndef FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H
+#define FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H
+
+#ifdef __has_include
+#if __has_include(<version>)
+#include <version>
+#endif
+#endif
+
+// Testing for https://wg21.link/N3652, adopted in C++14
+#if __cpp_constexpr >= 201304
+#define FASTFLOAT_CONSTEXPR14 constexpr
+#else
+#define FASTFLOAT_CONSTEXPR14
+#endif
+
+#if defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L
+#define FASTFLOAT_HAS_BIT_CAST 1
+#else
+#define FASTFLOAT_HAS_BIT_CAST 0
+#endif
+
+#if defined(__cpp_lib_is_constant_evaluated) &&                                \
+    __cpp_lib_is_constant_evaluated >= 201811L
+#define FASTFLOAT_HAS_IS_CONSTANT_EVALUATED 1
+#else
+#define FASTFLOAT_HAS_IS_CONSTANT_EVALUATED 0
+#endif
+
+// Testing for relevant C++20 constexpr library features
+#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED && FASTFLOAT_HAS_BIT_CAST &&           \
+    __cpp_lib_constexpr_algorithms >= 201806L /*For std::copy and std::fill*/
+#define FASTFLOAT_CONSTEXPR20 constexpr
+#define FASTFLOAT_IS_CONSTEXPR 1
+#else
+#define FASTFLOAT_CONSTEXPR20
+#define FASTFLOAT_IS_CONSTEXPR 0
+#endif
+
+#if __cplusplus >= 201703L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L)
+#define FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE 0
+#else
+#define FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE 1
+#endif
+
+#endif // FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H
+
+#ifndef FASTFLOAT_FLOAT_COMMON_H
+#define FASTFLOAT_FLOAT_COMMON_H
+
+#include <cfloat>
+#include <cstdint>
+#include <cassert>
+#include <cstring>
+#include <type_traits>
+#include <system_error>
+#ifdef __has_include
+#if __has_include(<stdfloat>) && (__cplusplus > 202002L || _MSVC_LANG > 202002L)
+#include <stdfloat>
+#endif
+#endif
+
+namespace fast_float {
+
+enum class chars_format : uint64_t;
+
+namespace detail {
+constexpr chars_format basic_json_fmt = chars_format(1 << 5);
+constexpr chars_format basic_fortran_fmt = chars_format(1 << 6);
+} // namespace detail
+
+enum class chars_format : uint64_t {
+  scientific = 1 << 0,
+  fixed = 1 << 2,
+  hex = 1 << 3,
+  no_infnan = 1 << 4,
+  // RFC 8259: https://datatracker.ietf.org/doc/html/rfc8259#section-6
+  json = uint64_t(detail::basic_json_fmt) | fixed | scientific | no_infnan,
+  // Extension of RFC 8259 where, e.g., "inf" and "nan" are allowed.
+  json_or_infnan = uint64_t(detail::basic_json_fmt) | fixed | scientific,
+  fortran = uint64_t(detail::basic_fortran_fmt) | fixed | scientific,
+  general = fixed | scientific,
+  allow_leading_plus = 1 << 7,
+  skip_white_space = 1 << 8,
+};
+
+template <typename UC> struct from_chars_result_t {
+  UC const *ptr;
+  std::errc ec;
+};
+using from_chars_result = from_chars_result_t<char>;
+
+template <typename UC> struct parse_options_t {
+  constexpr explicit parse_options_t(chars_format fmt = chars_format::general,
+                                     UC dot = UC('.'), int b = 10)
+      : format(fmt), decimal_point(dot), base(b) {}
+
+  /** Which number formats are accepted */
+  chars_format format;
+  /** The character used as decimal point */
+  UC decimal_point;
+  /** The base used for integers */
+  int base;
+};
+using parse_options = parse_options_t<char>;
+
+} // namespace fast_float
+
+#if FASTFLOAT_HAS_BIT_CAST
+#include <bit>
+#endif
+
+#if (defined(__x86_64) || defined(__x86_64__) || defined(_M_X64) ||            \
+     defined(__amd64) || defined(__aarch64__) || defined(_M_ARM64) ||          \
+     defined(__MINGW64__) || defined(__s390x__) ||                             \
+     (defined(__ppc64__) || defined(__PPC64__) || defined(__ppc64le__) ||      \
+      defined(__PPC64LE__)) ||                                                 \
+     defined(__loongarch64))
+#define FASTFLOAT_64BIT 1
+#elif (defined(__i386) || defined(__i386__) || defined(_M_IX86) ||             \
+       defined(__arm__) || defined(_M_ARM) || defined(__ppc__) ||              \
+       defined(__MINGW32__) || defined(__EMSCRIPTEN__))
+#define FASTFLOAT_32BIT 1
+#else
+  // Need to check incrementally, since SIZE_MAX is a size_t, avoid overflow.
+// We can never tell the register width, but the SIZE_MAX is a good
+// approximation. UINTPTR_MAX and INTPTR_MAX are optional, so avoid them for max
+// portability.
+#if SIZE_MAX == 0xffff
+#error Unknown platform (16-bit, unsupported)
+#elif SIZE_MAX == 0xffffffff
+#define FASTFLOAT_32BIT 1
+#elif SIZE_MAX == 0xffffffffffffffff
+#define FASTFLOAT_64BIT 1
+#else
+#error Unknown platform (not 32-bit, not 64-bit?)
+#endif
+#endif
+
+#if ((defined(_WIN32) || defined(_WIN64)) && !defined(__clang__)) ||           \
+    (defined(_M_ARM64) && !defined(__MINGW32__))
+#include <intrin.h>
+#endif
+
+#if defined(_MSC_VER) && !defined(__clang__)
+#define FASTFLOAT_VISUAL_STUDIO 1
+#endif
+
+#if defined __BYTE_ORDER__ && defined __ORDER_BIG_ENDIAN__
+#define FASTFLOAT_IS_BIG_ENDIAN (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#elif defined _WIN32
+#define FASTFLOAT_IS_BIG_ENDIAN 0
+#else
+#if defined(__APPLE__) || defined(__FreeBSD__)
+#include <machine/endian.h>
+#elif defined(sun) || defined(__sun)
+#include <sys/byteorder.h>
+#elif defined(__MVS__)
+#include <sys/endian.h>
+#else
+#ifdef __has_include
+#if __has_include(<endian.h>)
+#include <endian.h>
+#endif //__has_include(<endian.h>)
+#endif //__has_include
+#endif
+#
+#ifndef __BYTE_ORDER__
+// safe choice
+#define FASTFLOAT_IS_BIG_ENDIAN 0
+#endif
+#
+#ifndef __ORDER_LITTLE_ENDIAN__
+// safe choice
+#define FASTFLOAT_IS_BIG_ENDIAN 0
+#endif
+#
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+#define FASTFLOAT_IS_BIG_ENDIAN 0
+#else
+#define FASTFLOAT_IS_BIG_ENDIAN 1
+#endif
+#endif
+
+#if defined(__SSE2__) || (defined(FASTFLOAT_VISUAL_STUDIO) &&                  \
+                          (defined(_M_AMD64) || defined(_M_X64) ||             \
+                           (defined(_M_IX86_FP) && _M_IX86_FP == 2)))
+#define FASTFLOAT_SSE2 1
+#endif
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+#define FASTFLOAT_NEON 1
+#endif
+
+#if defined(FASTFLOAT_SSE2) || defined(FASTFLOAT_NEON)
+#define FASTFLOAT_HAS_SIMD 1
+#endif
+
+#if defined(__GNUC__)
+// disable -Wcast-align=strict (GCC only)
+#define FASTFLOAT_SIMD_DISABLE_WARNINGS                                        \
+  _Pragma("GCC diagnostic push")                                               \
+      _Pragma("GCC diagnostic ignored \"-Wcast-align\"")
+#else
+#define FASTFLOAT_SIMD_DISABLE_WARNINGS
+#endif
+
+#if defined(__GNUC__)
+#define FASTFLOAT_SIMD_RESTORE_WARNINGS _Pragma("GCC diagnostic pop")
+#else
+#define FASTFLOAT_SIMD_RESTORE_WARNINGS
+#endif
+
+#ifdef FASTFLOAT_VISUAL_STUDIO
+#define fastfloat_really_inline __forceinline
+#else
+#define fastfloat_really_inline inline __attribute__((always_inline))
+#endif
+
+#ifndef FASTFLOAT_ASSERT
+#define FASTFLOAT_ASSERT(x)                                                    \
+  { ((void)(x)); }
+#endif
+
+#ifndef FASTFLOAT_DEBUG_ASSERT
+#define FASTFLOAT_DEBUG_ASSERT(x)                                              \
+  { ((void)(x)); }
+#endif
+
+// rust style `try!()` macro, or `?` operator
+#define FASTFLOAT_TRY(x)                                                       \
+  {                                                                            \
+    if (!(x))                                                                  \
+      return false;                                                            \
+  }
+
+#define FASTFLOAT_ENABLE_IF(...)                                               \
+  typename std::enable_if<(__VA_ARGS__), int>::type
+
+namespace fast_float {
+
+fastfloat_really_inline constexpr bool cpp20_and_in_constexpr() {
+#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED
+  return std::is_constant_evaluated();
+#else
+  return false;
+#endif
+}
+
+template <typename T>
+fastfloat_really_inline constexpr bool is_supported_float_type() {
+  return std::is_same<T, float>::value || std::is_same<T, double>::value
+#if __STDCPP_FLOAT32_T__
+         || std::is_same<T, std::float32_t>::value
+#endif
+#if __STDCPP_FLOAT64_T__
+         || std::is_same<T, std::float64_t>::value
+#endif
+      ;
+}
+
+template <typename UC>
+fastfloat_really_inline constexpr bool is_supported_char_type() {
+  return std::is_same<UC, char>::value || std::is_same<UC, wchar_t>::value ||
+         std::is_same<UC, char16_t>::value || std::is_same<UC, char32_t>::value;
+}
+
+// Compares two ASCII strings in a case insensitive manner.
+template <typename UC>
+inline FASTFLOAT_CONSTEXPR14 bool
+fastfloat_strncasecmp(UC const *actual_mixedcase, UC const *expected_lowercase,
+                      size_t length) {
+  for (size_t i = 0; i < length; ++i) {
+    UC const actual = actual_mixedcase[i];
+    if ((actual < 256 ? actual | 32 : actual) != expected_lowercase[i]) {
+      return false;
+    }
+  }
+  return true;
+}
+
+#ifndef FLT_EVAL_METHOD
+#error "FLT_EVAL_METHOD should be defined, please include cfloat."
+#endif
+
+// a pointer and a length to a contiguous block of memory
+template <typename T> struct span {
+  const T *ptr;
+  size_t length;
+  constexpr span(const T *_ptr, size_t _length) : ptr(_ptr), length(_length) {}
+  constexpr span() : ptr(nullptr), length(0) {}
+
+  constexpr size_t len() const noexcept { return length; }
+
+  FASTFLOAT_CONSTEXPR14 const T &operator[](size_t index) const noexcept {
+    FASTFLOAT_DEBUG_ASSERT(index < length);
+    return ptr[index];
+  }
+};
+
+struct value128 {
+  uint64_t low;
+  uint64_t high;
+  constexpr value128(uint64_t _low, uint64_t _high) : low(_low), high(_high) {}
+  constexpr value128() : low(0), high(0) {}
+};
+
+/* Helper C++14 constexpr generic implementation of leading_zeroes */
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 int
+leading_zeroes_generic(uint64_t input_num, int last_bit = 0) {
+  if (input_num & uint64_t(0xffffffff00000000)) {
+    input_num >>= 32;
+    last_bit |= 32;
+  }
+  if (input_num & uint64_t(0xffff0000)) {
+    input_num >>= 16;
+    last_bit |= 16;
+  }
+  if (input_num & uint64_t(0xff00)) {
+    input_num >>= 8;
+    last_bit |= 8;
+  }
+  if (input_num & uint64_t(0xf0)) {
+    input_num >>= 4;
+    last_bit |= 4;
+  }
+  if (input_num & uint64_t(0xc)) {
+    input_num >>= 2;
+    last_bit |= 2;
+  }
+  if (input_num & uint64_t(0x2)) { /* input_num >>=  1; */
+    last_bit |= 1;
+  }
+  return 63 - last_bit;
+}
+
+/* result might be undefined when input_num is zero */
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 int
+leading_zeroes(uint64_t input_num) {
+  assert(input_num > 0);
+  if (cpp20_and_in_constexpr()) {
+    return leading_zeroes_generic(input_num);
+  }
+#ifdef FASTFLOAT_VISUAL_STUDIO
+#if defined(_M_X64) || defined(_M_ARM64)
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  _BitScanReverse64(&leading_zero, input_num);
+  return (int)(63 - leading_zero);
+#else
+  return leading_zeroes_generic(input_num);
+#endif
+#else
+  return __builtin_clzll(input_num);
+#endif
+}
+
+// slow emulation routine for 32-bit
+fastfloat_really_inline constexpr uint64_t emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t
+umul128_generic(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = (uint64_t)(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + (uint64_t)(lo < bd);
+  return lo;
+}
+
+#ifdef FASTFLOAT_32BIT
+
+// slow emulation routine for 32-bit
+#if !defined(__MINGW64__)
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t _umul128(uint64_t ab,
+                                                                uint64_t cd,
+                                                                uint64_t *hi) {
+  return umul128_generic(ab, cd, hi);
+}
+#endif // !__MINGW64__
+
+#endif // FASTFLOAT_32BIT
+
+// compute 64-bit a*b
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 value128
+full_multiplication(uint64_t a, uint64_t b) {
+  if (cpp20_and_in_constexpr()) {
+    value128 answer;
+    answer.low = umul128_generic(a, b, &answer.high);
+    return answer;
+  }
+  value128 answer;
+#if defined(_M_ARM64) && !defined(__MINGW32__)
+  // ARM64 has native support for 64-bit multiplications, no need to emulate
+  // But MinGW on ARM64 doesn't have native support for 64-bit multiplications
+  answer.high = __umulh(a, b);
+  answer.low = a * b;
+#elif defined(FASTFLOAT_32BIT) ||                                              \
+    (defined(_WIN64) && !defined(__clang__) && !defined(_M_ARM64))
+  answer.low = _umul128(a, b, &answer.high); // _umul128 not available on ARM64
+#elif defined(FASTFLOAT_64BIT) && defined(__SIZEOF_INT128__)
+  __uint128_t r = ((__uint128_t)a) * b;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#else
+  answer.low = umul128_generic(a, b, &answer.high);
+#endif
+  return answer;
+}
+
+struct adjusted_mantissa {
+  uint64_t mantissa{0};
+  int32_t power2{0}; // a negative value indicates an invalid result
+  adjusted_mantissa() = default;
+  constexpr bool operator==(const adjusted_mantissa &o) const {
+    return mantissa == o.mantissa && power2 == o.power2;
+  }
+  constexpr bool operator!=(const adjusted_mantissa &o) const {
+    return mantissa != o.mantissa || power2 != o.power2;
+  }
+};
+
+// Bias so we can get the real exponent with an invalid adjusted_mantissa.
+constexpr static int32_t invalid_am_bias = -0x8000;
+
+// used for binary_format_lookup_tables<T>::max_mantissa
+constexpr uint64_t constant_55555 = 5 * 5 * 5 * 5 * 5;
+
+template <typename T, typename U = void> struct binary_format_lookup_tables;
+
+template <typename T> struct binary_format : binary_format_lookup_tables<T> {
+  using equiv_uint =
+      typename std::conditional<sizeof(T) == 4, uint32_t, uint64_t>::type;
+
+  static inline constexpr int mantissa_explicit_bits();
+  static inline constexpr int minimum_exponent();
+  static inline constexpr int infinite_power();
+  static inline constexpr int sign_index();
+  static inline constexpr int
+  min_exponent_fast_path(); // used when fegetround() == FE_TONEAREST
+  static inline constexpr int max_exponent_fast_path();
+  static inline constexpr int max_exponent_round_to_even();
+  static inline constexpr int min_exponent_round_to_even();
+  static inline constexpr uint64_t max_mantissa_fast_path(int64_t power);
+  static inline constexpr uint64_t
+  max_mantissa_fast_path(); // used when fegetround() == FE_TONEAREST
+  static inline constexpr int largest_power_of_ten();
+  static inline constexpr int smallest_power_of_ten();
+  static inline constexpr T exact_power_of_ten(int64_t power);
+  static inline constexpr size_t max_digits();
+  static inline constexpr equiv_uint exponent_mask();
+  static inline constexpr equiv_uint mantissa_mask();
+  static inline constexpr equiv_uint hidden_bit_mask();
+};
+
+template <typename U> struct binary_format_lookup_tables<double, U> {
+  static constexpr double powers_of_ten[] = {
+      1e0,  1e1,  1e2,  1e3,  1e4,  1e5,  1e6,  1e7,  1e8,  1e9,  1e10, 1e11,
+      1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19, 1e20, 1e21, 1e22};
+
+  // Largest integer value v so that (5**index * v) <= 1<<53.
+  // 0x20000000000000 == 1 << 53
+  static constexpr uint64_t max_mantissa[] = {
+      0x20000000000000,
+      0x20000000000000 / 5,
+      0x20000000000000 / (5 * 5),
+      0x20000000000000 / (5 * 5 * 5),
+      0x20000000000000 / (5 * 5 * 5 * 5),
+      0x20000000000000 / (constant_55555),
+      0x20000000000000 / (constant_55555 * 5),
+      0x20000000000000 / (constant_55555 * 5 * 5),
+      0x20000000000000 / (constant_55555 * 5 * 5 * 5),
+      0x20000000000000 / (constant_55555 * 5 * 5 * 5 * 5),
+      0x20000000000000 / (constant_55555 * constant_55555),
+      0x20000000000000 / (constant_55555 * constant_55555 * 5),
+      0x20000000000000 / (constant_55555 * constant_55555 * 5 * 5),
+      0x20000000000000 / (constant_55555 * constant_55555 * 5 * 5 * 5),
+      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555),
+      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 * 5),
+      0x20000000000000 /
+          (constant_55555 * constant_55555 * constant_55555 * 5 * 5),
+      0x20000000000000 /
+          (constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5),
+      0x20000000000000 /
+          (constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5 * 5),
+      0x20000000000000 /
+          (constant_55555 * constant_55555 * constant_55555 * constant_55555),
+      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
+                          constant_55555 * 5),
+      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
+                          constant_55555 * 5 * 5),
+      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
+                          constant_55555 * 5 * 5 * 5),
+      0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
+                          constant_55555 * 5 * 5 * 5 * 5)};
+};
+
+#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
+
+template <typename U>
+constexpr double binary_format_lookup_tables<double, U>::powers_of_ten[];
+
+template <typename U>
+constexpr uint64_t binary_format_lookup_tables<double, U>::max_mantissa[];
+
+#endif
+
+template <typename U> struct binary_format_lookup_tables<float, U> {
+  static constexpr float powers_of_ten[] = {1e0f, 1e1f, 1e2f, 1e3f, 1e4f, 1e5f,
+                                            1e6f, 1e7f, 1e8f, 1e9f, 1e10f};
+
+  // Largest integer value v so that (5**index * v) <= 1<<24.
+  // 0x1000000 == 1<<24
+  static constexpr uint64_t max_mantissa[] = {
+      0x1000000,
+      0x1000000 / 5,
+      0x1000000 / (5 * 5),
+      0x1000000 / (5 * 5 * 5),
+      0x1000000 / (5 * 5 * 5 * 5),
+      0x1000000 / (constant_55555),
+      0x1000000 / (constant_55555 * 5),
+      0x1000000 / (constant_55555 * 5 * 5),
+      0x1000000 / (constant_55555 * 5 * 5 * 5),
+      0x1000000 / (constant_55555 * 5 * 5 * 5 * 5),
+      0x1000000 / (constant_55555 * constant_55555),
+      0x1000000 / (constant_55555 * constant_55555 * 5)};
+};
+
+#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
+
+template <typename U>
+constexpr float binary_format_lookup_tables<float, U>::powers_of_ten[];
+
+template <typename U>
+constexpr uint64_t binary_format_lookup_tables<float, U>::max_mantissa[];
+
+#endif
+
+template <>
+inline constexpr int binary_format<double>::min_exponent_fast_path() {
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  return 0;
+#else
+  return -22;
+#endif
+}
+
+template <>
+inline constexpr int binary_format<float>::min_exponent_fast_path() {
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  return 0;
+#else
+  return -10;
+#endif
+}
+
+template <>
+inline constexpr int binary_format<double>::mantissa_explicit_bits() {
+  return 52;
+}
+template <>
+inline constexpr int binary_format<float>::mantissa_explicit_bits() {
+  return 23;
+}
+
+template <>
+inline constexpr int binary_format<double>::max_exponent_round_to_even() {
+  return 23;
+}
+
+template <>
+inline constexpr int binary_format<float>::max_exponent_round_to_even() {
+  return 10;
+}
+
+template <>
+inline constexpr int binary_format<double>::min_exponent_round_to_even() {
+  return -4;
+}
+
+template <>
+inline constexpr int binary_format<float>::min_exponent_round_to_even() {
+  return -17;
+}
+
+template <> inline constexpr int binary_format<double>::minimum_exponent() {
+  return -1023;
+}
+template <> inline constexpr int binary_format<float>::minimum_exponent() {
+  return -127;
+}
+
+template <> inline constexpr int binary_format<double>::infinite_power() {
+  return 0x7FF;
+}
+template <> inline constexpr int binary_format<float>::infinite_power() {
+  return 0xFF;
+}
+
+template <> inline constexpr int binary_format<double>::sign_index() {
+  return 63;
+}
+template <> inline constexpr int binary_format<float>::sign_index() {
+  return 31;
+}
+
+template <>
+inline constexpr int binary_format<double>::max_exponent_fast_path() {
+  return 22;
+}
+template <>
+inline constexpr int binary_format<float>::max_exponent_fast_path() {
+  return 10;
+}
+
+template <>
+inline constexpr uint64_t binary_format<double>::max_mantissa_fast_path() {
+  return uint64_t(2) << mantissa_explicit_bits();
+}
+template <>
+inline constexpr uint64_t
+binary_format<double>::max_mantissa_fast_path(int64_t power) {
+  // caller is responsible to ensure that
+  // power >= 0 && power <= 22
+  //
+  // Work around clang bug https://godbolt.org/z/zedh7rrhc
+  return (void)max_mantissa[0], max_mantissa[power];
+}
+template <>
+inline constexpr uint64_t binary_format<float>::max_mantissa_fast_path() {
+  return uint64_t(2) << mantissa_explicit_bits();
+}
+template <>
+inline constexpr uint64_t
+binary_format<float>::max_mantissa_fast_path(int64_t power) {
+  // caller is responsible to ensure that
+  // power >= 0 && power <= 10
+  //
+  // Work around clang bug https://godbolt.org/z/zedh7rrhc
+  return (void)max_mantissa[0], max_mantissa[power];
+}
+
+template <>
+inline constexpr double
+binary_format<double>::exact_power_of_ten(int64_t power) {
+  // Work around clang bug https://godbolt.org/z/zedh7rrhc
+  return (void)powers_of_ten[0], powers_of_ten[power];
+}
+template <>
+inline constexpr float binary_format<float>::exact_power_of_ten(int64_t power) {
+  // Work around clang bug https://godbolt.org/z/zedh7rrhc
+  return (void)powers_of_ten[0], powers_of_ten[power];
+}
+
+template <> inline constexpr int binary_format<double>::largest_power_of_ten() {
+  return 308;
+}
+template <> inline constexpr int binary_format<float>::largest_power_of_ten() {
+  return 38;
+}
+
+template <>
+inline constexpr int binary_format<double>::smallest_power_of_ten() {
+  return -342;
+}
+template <> inline constexpr int binary_format<float>::smallest_power_of_ten() {
+  return -64;
+}
+
+template <> inline constexpr size_t binary_format<double>::max_digits() {
+  return 769;
+}
+template <> inline constexpr size_t binary_format<float>::max_digits() {
+  return 114;
+}
+
+template <>
+inline constexpr binary_format<float>::equiv_uint
+binary_format<float>::exponent_mask() {
+  return 0x7F800000;
+}
+template <>
+inline constexpr binary_format<double>::equiv_uint
+binary_format<double>::exponent_mask() {
+  return 0x7FF0000000000000;
+}
+
+template <>
+inline constexpr binary_format<float>::equiv_uint
+binary_format<float>::mantissa_mask() {
+  return 0x007FFFFF;
+}
+template <>
+inline constexpr binary_format<double>::equiv_uint
+binary_format<double>::mantissa_mask() {
+  return 0x000FFFFFFFFFFFFF;
+}
+
+template <>
+inline constexpr binary_format<float>::equiv_uint
+binary_format<float>::hidden_bit_mask() {
+  return 0x00800000;
+}
+template <>
+inline constexpr binary_format<double>::equiv_uint
+binary_format<double>::hidden_bit_mask() {
+  return 0x0010000000000000;
+}
+
+template <typename T>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
+to_float(bool negative, adjusted_mantissa am, T &value) {
+  using fastfloat_uint = typename binary_format<T>::equiv_uint;
+  fastfloat_uint word = (fastfloat_uint)am.mantissa;
+  word |= fastfloat_uint(am.power2)
+          << binary_format<T>::mantissa_explicit_bits();
+  word |= fastfloat_uint(negative) << binary_format<T>::sign_index();
+#if FASTFLOAT_HAS_BIT_CAST
+  value = std::bit_cast<T>(word);
+#else
+  ::memcpy(&value, &word, sizeof(T));
+#endif
+}
+
+template <typename = void> struct space_lut {
+  static constexpr bool value[] = {
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+};
+
+#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
+
+template <typename T> constexpr bool space_lut<T>::value[];
+
+#endif
+
+template <typename UC> constexpr bool is_space(UC c) {
+  return c < 256 && space_lut<>::value[uint8_t(c)];
+}
+
+template <typename UC> static constexpr uint64_t int_cmp_zeros() {
+  static_assert((sizeof(UC) == 1) || (sizeof(UC) == 2) || (sizeof(UC) == 4),
+                "Unsupported character size");
+  return (sizeof(UC) == 1) ? 0x3030303030303030
+         : (sizeof(UC) == 2)
+             ? (uint64_t(UC('0')) << 48 | uint64_t(UC('0')) << 32 |
+                uint64_t(UC('0')) << 16 | UC('0'))
+             : (uint64_t(UC('0')) << 32 | UC('0'));
+}
+template <typename UC> static constexpr int int_cmp_len() {
+  return sizeof(uint64_t) / sizeof(UC);
+}
+template <typename UC> static constexpr UC const *str_const_nan() {
+  return nullptr;
+}
+template <> constexpr char const *str_const_nan<char>() { return "nan"; }
+template <> constexpr wchar_t const *str_const_nan<wchar_t>() { return L"nan"; }
+template <> constexpr char16_t const *str_const_nan<char16_t>() {
+  return u"nan";
+}
+template <> constexpr char32_t const *str_const_nan<char32_t>() {
+  return U"nan";
+}
+template <typename UC> static constexpr UC const *str_const_inf() {
+  return nullptr;
+}
+template <> constexpr char const *str_const_inf<char>() { return "infinity"; }
+template <> constexpr wchar_t const *str_const_inf<wchar_t>() {
+  return L"infinity";
+}
+template <> constexpr char16_t const *str_const_inf<char16_t>() {
+  return u"infinity";
+}
+template <> constexpr char32_t const *str_const_inf<char32_t>() {
+  return U"infinity";
+}
+
+template <typename = void> struct int_luts {
+  static constexpr uint8_t chdigit[] = {
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   255, 255,
+      255, 255, 255, 255, 255, 10,  11,  12,  13,  14,  15,  16,  17,  18,  19,
+      20,  21,  22,  23,  24,  25,  26,  27,  28,  29,  30,  31,  32,  33,  34,
+      35,  255, 255, 255, 255, 255, 255, 10,  11,  12,  13,  14,  15,  16,  17,
+      18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,  30,  31,  32,
+      33,  34,  35,  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+      255};
+
+  static constexpr size_t maxdigits_u64[] = {
+      64, 41, 32, 28, 25, 23, 22, 21, 20, 19, 18, 18, 17, 17, 16, 16, 16, 16,
+      15, 15, 15, 15, 14, 14, 14, 14, 14, 14, 14, 13, 13, 13, 13, 13, 13};
+
+  static constexpr uint64_t min_safe_u64[] = {
+      9223372036854775808ull,  12157665459056928801ull, 4611686018427387904,
+      7450580596923828125,     4738381338321616896,     3909821048582988049,
+      9223372036854775808ull,  12157665459056928801ull, 10000000000000000000ull,
+      5559917313492231481,     2218611106740436992,     8650415919381337933,
+      2177953337809371136,     6568408355712890625,     1152921504606846976,
+      2862423051509815793,     6746640616477458432,     15181127029874798299ull,
+      1638400000000000000,     3243919932521508681,     6221821273427820544,
+      11592836324538749809ull, 876488338465357824,      1490116119384765625,
+      2481152873203736576,     4052555153018976267,     6502111422497947648,
+      10260628712958602189ull, 15943230000000000000ull, 787662783788549761,
+      1152921504606846976,     1667889514952984961,     2386420683693101056,
+      3379220508056640625,     4738381338321616896};
+};
+
+#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
+
+template <typename T> constexpr uint8_t int_luts<T>::chdigit[];
+
+template <typename T> constexpr size_t int_luts<T>::maxdigits_u64[];
+
+template <typename T> constexpr uint64_t int_luts<T>::min_safe_u64[];
+
+#endif
+
+template <typename UC>
+fastfloat_really_inline constexpr uint8_t ch_to_digit(UC c) {
+  return int_luts<>::chdigit[static_cast<unsigned char>(c)];
+}
+
+fastfloat_really_inline constexpr size_t max_digits_u64(int base) {
+  return int_luts<>::maxdigits_u64[base - 2];
+}
+
+// If a u64 is exactly max_digits_u64() in length, this is
+// the value below which it has definitely overflowed.
+fastfloat_really_inline constexpr uint64_t min_safe_u64(int base) {
+  return int_luts<>::min_safe_u64[base - 2];
+}
+
+constexpr chars_format operator~(chars_format rhs) noexcept {
+  using int_type = std::underlying_type<chars_format>::type;
+  return static_cast<chars_format>(~static_cast<int_type>(rhs));
+}
+
+constexpr chars_format operator&(chars_format lhs, chars_format rhs) noexcept {
+  using int_type = std::underlying_type<chars_format>::type;
+  return static_cast<chars_format>(static_cast<int_type>(lhs) &
+                                   static_cast<int_type>(rhs));
+}
+
+constexpr chars_format operator|(chars_format lhs, chars_format rhs) noexcept {
+  using int_type = std::underlying_type<chars_format>::type;
+  return static_cast<chars_format>(static_cast<int_type>(lhs) |
+                                   static_cast<int_type>(rhs));
+}
+
+constexpr chars_format operator^(chars_format lhs, chars_format rhs) noexcept {
+  using int_type = std::underlying_type<chars_format>::type;
+  return static_cast<chars_format>(static_cast<int_type>(lhs) ^
+                                   static_cast<int_type>(rhs));
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 chars_format &
+operator&=(chars_format &lhs, chars_format rhs) noexcept {
+  return lhs = (lhs & rhs);
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 chars_format &
+operator|=(chars_format &lhs, chars_format rhs) noexcept {
+  return lhs = (lhs | rhs);
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 chars_format &
+operator^=(chars_format &lhs, chars_format rhs) noexcept {
+  return lhs = (lhs ^ rhs);
+}
+
+namespace detail {
+// adjust for deprecated feature macros
+constexpr chars_format adjust_for_feature_macros(chars_format fmt) {
+  return fmt
+#ifdef FASTFLOAT_ALLOWS_LEADING_PLUS
+         | chars_format::allow_leading_plus
+#endif
+#ifdef FASTFLOAT_SKIP_WHITE_SPACE
+         | chars_format::skip_white_space
+#endif
+      ;
+}
+} // namespace detail
+
+} // namespace fast_float
+
+#endif
+
+
+#ifndef FASTFLOAT_FAST_FLOAT_H
+#define FASTFLOAT_FAST_FLOAT_H
+
+
+namespace fast_float {
+/**
+ * This function parses the character sequence [first,last) for a number. It
+ * parses floating-point numbers expecting a locale-indepent format equivalent
+ * to what is used by std::strtod in the default ("C") locale. The resulting
+ * floating-point value is the closest floating-point values (using either float
+ * or double), using the "round to even" convention for values that would
+ * otherwise fall right in-between two values. That is, we provide exact parsing
+ * according to the IEEE standard.
+ *
+ * Given a successful parse, the pointer (`ptr`) in the returned value is set to
+ * point right after the parsed number, and the `value` referenced is set to the
+ * parsed value. In case of error, the returned `ec` contains a representative
+ * error, otherwise the default (`std::errc()`) value is stored.
+ *
+ * The implementation does not throw and does not allocate memory (e.g., with
+ * `new` or `malloc`).
+ *
+ * Like the C++17 standard, the `fast_float::from_chars` functions take an
+ * optional last argument of the type `fast_float::chars_format`. It is a bitset
+ * value: we check whether `fmt & fast_float::chars_format::fixed` and `fmt &
+ * fast_float::chars_format::scientific` are set to determine whether we allow
+ * the fixed point and scientific notation respectively. The default is
+ * `fast_float::chars_format::general` which allows both `fixed` and
+ * `scientific`.
+ */
+template <typename T, typename UC = char,
+          typename = FASTFLOAT_ENABLE_IF(is_supported_float_type<T>())>
+FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+from_chars(UC const *first, UC const *last, T &value,
+           chars_format fmt = chars_format::general) noexcept;
+
+/**
+ * Like from_chars, but accepts an `options` argument to govern number parsing.
+ * Both for floating-point types and integer types.
+ */
+template <typename T, typename UC = char>
+FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+from_chars_advanced(UC const *first, UC const *last, T &value,
+                    parse_options_t<UC> options) noexcept;
+
+/**
+ * from_chars for integer types.
+ */
+template <typename T, typename UC = char,
+          typename = FASTFLOAT_ENABLE_IF(!is_supported_float_type<T>())>
+FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+from_chars(UC const *first, UC const *last, T &value, int base = 10) noexcept;
+
+} // namespace fast_float
+#endif // FASTFLOAT_FAST_FLOAT_H
+
+#ifndef FASTFLOAT_ASCII_NUMBER_H
+#define FASTFLOAT_ASCII_NUMBER_H
+
+#include <cctype>
+#include <cstdint>
+#include <cstring>
+#include <iterator>
+#include <limits>
+#include <type_traits>
+
+
+#ifdef FASTFLOAT_SSE2
+#include <emmintrin.h>
+#endif
+
+#ifdef FASTFLOAT_NEON
+#include <arm_neon.h>
+#endif
+
+namespace fast_float {
+
+template <typename UC> fastfloat_really_inline constexpr bool has_simd_opt() {
+#ifdef FASTFLOAT_HAS_SIMD
+  return std::is_same<UC, char16_t>::value;
+#else
+  return false;
+#endif
+}
+
+// Next function can be micro-optimized, but compilers are entirely
+// able to optimize it well.
+template <typename UC>
+fastfloat_really_inline constexpr bool is_integer(UC c) noexcept {
+  return !(c > UC('9') || c < UC('0'));
+}
+
+fastfloat_really_inline constexpr uint64_t byteswap(uint64_t val) {
+  return (val & 0xFF00000000000000) >> 56 | (val & 0x00FF000000000000) >> 40 |
+         (val & 0x0000FF0000000000) >> 24 | (val & 0x000000FF00000000) >> 8 |
+         (val & 0x00000000FF000000) << 8 | (val & 0x0000000000FF0000) << 24 |
+         (val & 0x000000000000FF00) << 40 | (val & 0x00000000000000FF) << 56;
+}
+
+// Read 8 UC into a u64. Truncates UC if not char.
+template <typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
+read8_to_u64(const UC *chars) {
+  if (cpp20_and_in_constexpr() || !std::is_same<UC, char>::value) {
+    uint64_t val = 0;
+    for (int i = 0; i < 8; ++i) {
+      val |= uint64_t(uint8_t(*chars)) << (i * 8);
+      ++chars;
+    }
+    return val;
+  }
+  uint64_t val;
+  ::memcpy(&val, chars, sizeof(uint64_t));
+#if FASTFLOAT_IS_BIG_ENDIAN == 1
+  // Need to read as-if the number was in little-endian order.
+  val = byteswap(val);
+#endif
+  return val;
+}
+
+#ifdef FASTFLOAT_SSE2
+
+fastfloat_really_inline uint64_t simd_read8_to_u64(const __m128i data) {
+  FASTFLOAT_SIMD_DISABLE_WARNINGS
+  const __m128i packed = _mm_packus_epi16(data, data);
+#ifdef FASTFLOAT_64BIT
+  return uint64_t(_mm_cvtsi128_si64(packed));
+#else
+  uint64_t value;
+  // Visual Studio + older versions of GCC don't support _mm_storeu_si64
+  _mm_storel_epi64(reinterpret_cast<__m128i *>(&value), packed);
+  return value;
+#endif
+  FASTFLOAT_SIMD_RESTORE_WARNINGS
+}
+
+fastfloat_really_inline uint64_t simd_read8_to_u64(const char16_t *chars) {
+  FASTFLOAT_SIMD_DISABLE_WARNINGS
+  return simd_read8_to_u64(
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(chars)));
+  FASTFLOAT_SIMD_RESTORE_WARNINGS
+}
+
+#elif defined(FASTFLOAT_NEON)
+
+fastfloat_really_inline uint64_t simd_read8_to_u64(const uint16x8_t data) {
+  FASTFLOAT_SIMD_DISABLE_WARNINGS
+  uint8x8_t utf8_packed = vmovn_u16(data);
+  return vget_lane_u64(vreinterpret_u64_u8(utf8_packed), 0);
+  FASTFLOAT_SIMD_RESTORE_WARNINGS
+}
+
+fastfloat_really_inline uint64_t simd_read8_to_u64(const char16_t *chars) {
+  FASTFLOAT_SIMD_DISABLE_WARNINGS
+  return simd_read8_to_u64(
+      vld1q_u16(reinterpret_cast<const uint16_t *>(chars)));
+  FASTFLOAT_SIMD_RESTORE_WARNINGS
+}
+
+#endif // FASTFLOAT_SSE2
+
+// MSVC SFINAE is broken pre-VS2017
+#if defined(_MSC_VER) && _MSC_VER <= 1900
+template <typename UC>
+#else
+template <typename UC, FASTFLOAT_ENABLE_IF(!has_simd_opt<UC>()) = 0>
+#endif
+// dummy for compile
+uint64_t simd_read8_to_u64(UC const *) {
+  return 0;
+}
+
+// credit  @aqrit
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint32_t
+parse_eight_digits_unrolled(uint64_t val) {
+  const uint64_t mask = 0x000000FF000000FF;
+  const uint64_t mul1 = 0x000F424000000064; // 100 + (1000000ULL << 32)
+  const uint64_t mul2 = 0x0000271000000001; // 1 + (10000ULL << 32)
+  val -= 0x3030303030303030;
+  val = (val * 10) + (val >> 8); // val = (val * 2561) >> 8;
+  val = (((val & mask) * mul1) + (((val >> 16) & mask) * mul2)) >> 32;
+  return uint32_t(val);
+}
+
+// Call this if chars are definitely 8 digits.
+template <typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint32_t
+parse_eight_digits_unrolled(UC const *chars) noexcept {
+  if (cpp20_and_in_constexpr() || !has_simd_opt<UC>()) {
+    return parse_eight_digits_unrolled(read8_to_u64(chars)); // truncation okay
+  }
+  return parse_eight_digits_unrolled(simd_read8_to_u64(chars));
+}
+
+// credit @aqrit
+fastfloat_really_inline constexpr bool
+is_made_of_eight_digits_fast(uint64_t val) noexcept {
+  return !((((val + 0x4646464646464646) | (val - 0x3030303030303030)) &
+            0x8080808080808080));
+}
+
+#ifdef FASTFLOAT_HAS_SIMD
+
+// Call this if chars might not be 8 digits.
+// Using this style (instead of is_made_of_eight_digits_fast() then
+// parse_eight_digits_unrolled()) ensures we don't load SIMD registers twice.
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
+simd_parse_if_eight_digits_unrolled(const char16_t *chars,
+                                    uint64_t &i) noexcept {
+  if (cpp20_and_in_constexpr()) {
+    return false;
+  }
+#ifdef FASTFLOAT_SSE2
+  FASTFLOAT_SIMD_DISABLE_WARNINGS
+  const __m128i data =
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(chars));
+
+  // (x - '0') <= 9
+  // http://0x80.pl/articles/simd-parsing-int-sequences.html
+  const __m128i t0 = _mm_add_epi16(data, _mm_set1_epi16(32720));
+  const __m128i t1 = _mm_cmpgt_epi16(t0, _mm_set1_epi16(-32759));
+
+  if (_mm_movemask_epi8(t1) == 0) {
+    i = i * 100000000 + parse_eight_digits_unrolled(simd_read8_to_u64(data));
+    return true;
+  } else
+    return false;
+  FASTFLOAT_SIMD_RESTORE_WARNINGS
+#elif defined(FASTFLOAT_NEON)
+  FASTFLOAT_SIMD_DISABLE_WARNINGS
+  const uint16x8_t data = vld1q_u16(reinterpret_cast<const uint16_t *>(chars));
+
+  // (x - '0') <= 9
+  // http://0x80.pl/articles/simd-parsing-int-sequences.html
+  const uint16x8_t t0 = vsubq_u16(data, vmovq_n_u16('0'));
+  const uint16x8_t mask = vcltq_u16(t0, vmovq_n_u16('9' - '0' + 1));
+
+  if (vminvq_u16(mask) == 0xFFFF) {
+    i = i * 100000000 + parse_eight_digits_unrolled(simd_read8_to_u64(data));
+    return true;
+  } else
+    return false;
+  FASTFLOAT_SIMD_RESTORE_WARNINGS
+#else
+  (void)chars;
+  (void)i;
+  return false;
+#endif // FASTFLOAT_SSE2
+}
+
+#endif // FASTFLOAT_HAS_SIMD
+
+// MSVC SFINAE is broken pre-VS2017
+#if defined(_MSC_VER) && _MSC_VER <= 1900
+template <typename UC>
+#else
+template <typename UC, FASTFLOAT_ENABLE_IF(!has_simd_opt<UC>()) = 0>
+#endif
+// dummy for compile
+bool simd_parse_if_eight_digits_unrolled(UC const *, uint64_t &) {
+  return 0;
+}
+
+template <typename UC, FASTFLOAT_ENABLE_IF(!std::is_same<UC, char>::value) = 0>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
+loop_parse_if_eight_digits(const UC *&p, const UC *const pend, uint64_t &i) {
+  if (!has_simd_opt<UC>()) {
+    return;
+  }
+  while ((std::distance(p, pend) >= 8) &&
+         simd_parse_if_eight_digits_unrolled(
+             p, i)) { // in rare cases, this will overflow, but that's ok
+    p += 8;
+  }
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
+loop_parse_if_eight_digits(const char *&p, const char *const pend,
+                           uint64_t &i) {
+  // optimizes better than parse_if_eight_digits_unrolled() for UC = char.
+  while ((std::distance(p, pend) >= 8) &&
+         is_made_of_eight_digits_fast(read8_to_u64(p))) {
+    i = i * 100000000 +
+        parse_eight_digits_unrolled(read8_to_u64(
+            p)); // in rare cases, this will overflow, but that's ok
+    p += 8;
+  }
+}
+
+enum class parse_error {
+  no_error,
+  // [JSON-only] The minus sign must be followed by an integer.
+  missing_integer_after_sign,
+  // A sign must be followed by an integer or dot.
+  missing_integer_or_dot_after_sign,
+  // [JSON-only] The integer part must not have leading zeros.
+  leading_zeros_in_integer_part,
+  // [JSON-only] The integer part must have at least one digit.
+  no_digits_in_integer_part,
+  // [JSON-only] If there is a decimal point, there must be digits in the
+  // fractional part.
+  no_digits_in_fractional_part,
+  // The mantissa must have at least one digit.
+  no_digits_in_mantissa,
+  // Scientific notation requires an exponential part.
+  missing_exponential_part,
+};
+
+template <typename UC> struct parsed_number_string_t {
+  int64_t exponent{0};
+  uint64_t mantissa{0};
+  UC const *lastmatch{nullptr};
+  bool negative{false};
+  bool valid{false};
+  bool too_many_digits{false};
+  // contains the range of the significant digits
+  span<const UC> integer{};  // non-nullable
+  span<const UC> fraction{}; // nullable
+  parse_error error{parse_error::no_error};
+};
+
+using byte_span = span<const char>;
+using parsed_number_string = parsed_number_string_t<char>;
+
+template <typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 parsed_number_string_t<UC>
+report_parse_error(UC const *p, parse_error error) {
+  parsed_number_string_t<UC> answer;
+  answer.valid = false;
+  answer.lastmatch = p;
+  answer.error = error;
+  return answer;
+}
+
+// Assuming that you use no more than 19 digits, this will
+// parse an ASCII string.
+template <typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 parsed_number_string_t<UC>
+parse_number_string(UC const *p, UC const *pend,
+                    parse_options_t<UC> options) noexcept {
+  chars_format const fmt = detail::adjust_for_feature_macros(options.format);
+  UC const decimal_point = options.decimal_point;
+
+  parsed_number_string_t<UC> answer;
+  answer.valid = false;
+  answer.too_many_digits = false;
+  // assume p < pend, so dereference without checks;
+  answer.negative = (*p == UC('-'));
+  // C++17 20.19.3.(7.1) explicitly forbids '+' sign here
+  if ((*p == UC('-')) ||
+      (uint64_t(fmt & chars_format::allow_leading_plus) &&
+       !uint64_t(fmt & detail::basic_json_fmt) && *p == UC('+'))) {
+    ++p;
+    if (p == pend) {
+      return report_parse_error<UC>(
+          p, parse_error::missing_integer_or_dot_after_sign);
+    }
+    if (uint64_t(fmt & detail::basic_json_fmt)) {
+      if (!is_integer(*p)) { // a sign must be followed by an integer
+        return report_parse_error<UC>(p,
+                                      parse_error::missing_integer_after_sign);
+      }
+    } else {
+      if (!is_integer(*p) &&
+          (*p !=
+           decimal_point)) { // a sign must be followed by an integer or the dot
+        return report_parse_error<UC>(
+            p, parse_error::missing_integer_or_dot_after_sign);
+      }
+    }
+  }
+  UC const *const start_digits = p;
+
+  uint64_t i = 0; // an unsigned int avoids signed overflows (which are bad)
+
+  while ((p != pend) && is_integer(*p)) {
+    // a multiplication by 10 is cheaper than an arbitrary integer
+    // multiplication
+    i = 10 * i +
+        uint64_t(*p -
+                 UC('0')); // might overflow, we will handle the overflow later
+    ++p;
+  }
+  UC const *const end_of_integer_part = p;
+  int64_t digit_count = int64_t(end_of_integer_part - start_digits);
+  answer.integer = span<const UC>(start_digits, size_t(digit_count));
+  if (uint64_t(fmt & detail::basic_json_fmt)) {
+    // at least 1 digit in integer part, without leading zeros
+    if (digit_count == 0) {
+      return report_parse_error<UC>(p, parse_error::no_digits_in_integer_part);
+    }
+    if ((start_digits[0] == UC('0') && digit_count > 1)) {
+      return report_parse_error<UC>(start_digits,
+                                    parse_error::leading_zeros_in_integer_part);
+    }
+  }
+
+  int64_t exponent = 0;
+  const bool has_decimal_point = (p != pend) && (*p == decimal_point);
+  if (has_decimal_point) {
+    ++p;
+    UC const *before = p;
+    // can occur at most twice without overflowing, but let it occur more, since
+    // for integers with many digits, digit parsing is the primary bottleneck.
+    loop_parse_if_eight_digits(p, pend, i);
+
+    while ((p != pend) && is_integer(*p)) {
+      uint8_t digit = uint8_t(*p - UC('0'));
+      ++p;
+      i = i * 10 + digit; // in rare cases, this will overflow, but that's ok
+    }
+    exponent = before - p;
+    answer.fraction = span<const UC>(before, size_t(p - before));
+    digit_count -= exponent;
+  }
+  if (uint64_t(fmt & detail::basic_json_fmt)) {
+    // at least 1 digit in fractional part
+    if (has_decimal_point && exponent == 0) {
+      return report_parse_error<UC>(p,
+                                    parse_error::no_digits_in_fractional_part);
+    }
+  } else if (digit_count ==
+             0) { // we must have encountered at least one integer!
+    return report_parse_error<UC>(p, parse_error::no_digits_in_mantissa);
+  }
+  int64_t exp_number = 0; // explicit exponential part
+  if ((uint64_t(fmt & chars_format::scientific) && (p != pend) &&
+       ((UC('e') == *p) || (UC('E') == *p))) ||
+      (uint64_t(fmt & detail::basic_fortran_fmt) && (p != pend) &&
+       ((UC('+') == *p) || (UC('-') == *p) || (UC('d') == *p) ||
+        (UC('D') == *p)))) {
+    UC const *location_of_e = p;
+    if ((UC('e') == *p) || (UC('E') == *p) || (UC('d') == *p) ||
+        (UC('D') == *p)) {
+      ++p;
+    }
+    bool neg_exp = false;
+    if ((p != pend) && (UC('-') == *p)) {
+      neg_exp = true;
+      ++p;
+    } else if ((p != pend) &&
+               (UC('+') ==
+                *p)) { // '+' on exponent is allowed by C++17 20.19.3.(7.1)
+      ++p;
+    }
+    if ((p == pend) || !is_integer(*p)) {
+      if (!uint64_t(fmt & chars_format::fixed)) {
+        // The exponential part is invalid for scientific notation, so it must
+        // be a trailing token for fixed notation. However, fixed notation is
+        // disabled, so report a scientific notation error.
+        return report_parse_error<UC>(p, parse_error::missing_exponential_part);
+      }
+      // Otherwise, we will be ignoring the 'e'.
+      p = location_of_e;
+    } else {
+      while ((p != pend) && is_integer(*p)) {
+        uint8_t digit = uint8_t(*p - UC('0'));
+        if (exp_number < 0x10000000) {
+          exp_number = 10 * exp_number + digit;
+        }
+        ++p;
+      }
+      if (neg_exp) {
+        exp_number = -exp_number;
+      }
+      exponent += exp_number;
+    }
+  } else {
+    // If it scientific and not fixed, we have to bail out.
+    if (uint64_t(fmt & chars_format::scientific) &&
+        !uint64_t(fmt & chars_format::fixed)) {
+      return report_parse_error<UC>(p, parse_error::missing_exponential_part);
+    }
+  }
+  answer.lastmatch = p;
+  answer.valid = true;
+
+  // If we frequently had to deal with long strings of digits,
+  // we could extend our code by using a 128-bit integer instead
+  // of a 64-bit integer. However, this is uncommon.
+  //
+  // We can deal with up to 19 digits.
+  if (digit_count > 19) { // this is uncommon
+    // It is possible that the integer had an overflow.
+    // We have to handle the case where we have 0.0000somenumber.
+    // We need to be mindful of the case where we only have zeroes...
+    // E.g., 0.000000000...000.
+    UC const *start = start_digits;
+    while ((start != pend) && (*start == UC('0') || *start == decimal_point)) {
+      if (*start == UC('0')) {
+        digit_count--;
+      }
+      start++;
+    }
+
+    if (digit_count > 19) {
+      answer.too_many_digits = true;
+      // Let us start again, this time, avoiding overflows.
+      // We don't need to check if is_integer, since we use the
+      // pre-tokenized spans from above.
+      i = 0;
+      p = answer.integer.ptr;
+      UC const *int_end = p + answer.integer.len();
+      const uint64_t minimal_nineteen_digit_integer{1000000000000000000};
+      while ((i < minimal_nineteen_digit_integer) && (p != int_end)) {
+        i = i * 10 + uint64_t(*p - UC('0'));
+        ++p;
+      }
+      if (i >= minimal_nineteen_digit_integer) { // We have a big integers
+        exponent = end_of_integer_part - p + exp_number;
+      } else { // We have a value with a fractional component.
+        p = answer.fraction.ptr;
+        UC const *frac_end = p + answer.fraction.len();
+        while ((i < minimal_nineteen_digit_integer) && (p != frac_end)) {
+          i = i * 10 + uint64_t(*p - UC('0'));
+          ++p;
+        }
+        exponent = answer.fraction.ptr - p + exp_number;
+      }
+      // We have now corrected both exponent and i, to a truncated value
+    }
+  }
+  answer.exponent = exponent;
+  answer.mantissa = i;
+  return answer;
+}
+
+template <typename T, typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+parse_int_string(UC const *p, UC const *pend, T &value,
+                 parse_options_t<UC> options) {
+  chars_format const fmt = detail::adjust_for_feature_macros(options.format);
+  int const base = options.base;
+
+  from_chars_result_t<UC> answer;
+
+  UC const *const first = p;
+
+  bool const negative = (*p == UC('-'));
+  if (!std::is_signed<T>::value && negative) {
+    answer.ec = std::errc::invalid_argument;
+    answer.ptr = first;
+    return answer;
+  }
+  if ((*p == UC('-')) ||
+      (uint64_t(fmt & chars_format::allow_leading_plus) && (*p == UC('+')))) {
+    ++p;
+  }
+
+  UC const *const start_num = p;
+
+  while (p != pend && *p == UC('0')) {
+    ++p;
+  }
+
+  const bool has_leading_zeros = p > start_num;
+
+  UC const *const start_digits = p;
+
+  uint64_t i = 0;
+  if (base == 10) {
+    loop_parse_if_eight_digits(p, pend, i); // use SIMD if possible
+  }
+  while (p != pend) {
+    uint8_t digit = ch_to_digit(*p);
+    if (digit >= base) {
+      break;
+    }
+    i = uint64_t(base) * i + digit; // might overflow, check this later
+    p++;
+  }
+
+  size_t digit_count = size_t(p - start_digits);
+
+  if (digit_count == 0) {
+    if (has_leading_zeros) {
+      value = 0;
+      answer.ec = std::errc();
+      answer.ptr = p;
+    } else {
+      answer.ec = std::errc::invalid_argument;
+      answer.ptr = first;
+    }
+    return answer;
+  }
+
+  answer.ptr = p;
+
+  // check u64 overflow
+  size_t max_digits = max_digits_u64(base);
+  if (digit_count > max_digits) {
+    answer.ec = std::errc::result_out_of_range;
+    return answer;
+  }
+  // this check can be eliminated for all other types, but they will all require
+  // a max_digits(base) equivalent
+  if (digit_count == max_digits && i < min_safe_u64(base)) {
+    answer.ec = std::errc::result_out_of_range;
+    return answer;
+  }
+
+  // check other types overflow
+  if (!std::is_same<T, uint64_t>::value) {
+    if (i > uint64_t(std::numeric_limits<T>::max()) + uint64_t(negative)) {
+      answer.ec = std::errc::result_out_of_range;
+      return answer;
+    }
+  }
+
+  if (negative) {
+#ifdef FASTFLOAT_VISUAL_STUDIO
+#pragma warning(push)
+#pragma warning(disable : 4146)
+#endif
+    // this weird workaround is required because:
+    // - converting unsigned to signed when its value is greater than signed max
+    // is UB pre-C++23.
+    // - reinterpret_casting (~i + 1) would work, but it is not constexpr
+    // this is always optimized into a neg instruction (note: T is an integer
+    // type)
+    value = T(-std::numeric_limits<T>::max() -
+              T(i - uint64_t(std::numeric_limits<T>::max())));
+#ifdef FASTFLOAT_VISUAL_STUDIO
+#pragma warning(pop)
+#endif
+  } else {
+    value = T(i);
+  }
+
+  answer.ec = std::errc();
+  return answer;
+}
+
+} // namespace fast_float
+
+#endif
+
+#ifndef FASTFLOAT_FAST_TABLE_H
+#define FASTFLOAT_FAST_TABLE_H
+
+#include <cstdint>
+
+namespace fast_float {
+
+/**
+ * When mapping numbers from decimal to binary,
+ * we go from w * 10^q to m * 2^p but we have
+ * 10^q = 5^q * 2^q, so effectively
+ * we are trying to match
+ * w * 2^q * 5^q to m * 2^p. Thus the powers of two
+ * are not a concern since they can be represented
+ * exactly using the binary notation, only the powers of five
+ * affect the binary significand.
+ */
+
+/**
+ * The smallest non-zero float (binary64) is 2^-1074.
+ * We take as input numbers of the form w x 10^q where w < 2^64.
+ * We have that w * 10^-343  <  2^(64-344) 5^-343 < 2^-1076.
+ * However, we have that
+ * (2^64-1) * 10^-342 =  (2^64-1) * 2^-342 * 5^-342 > 2^-1074.
+ * Thus it is possible for a number of the form w * 10^-342 where
+ * w is a 64-bit value to be a non-zero floating-point number.
+ *********
+ * Any number of form w * 10^309 where w>= 1 is going to be
+ * infinite in binary64 so we never need to worry about powers
+ * of 5 greater than 308.
+ */
+template <class unused = void> struct powers_template {
+
+  constexpr static int smallest_power_of_five =
+      binary_format<double>::smallest_power_of_ten();
+  constexpr static int largest_power_of_five =
+      binary_format<double>::largest_power_of_ten();
+  constexpr static int number_of_entries =
+      2 * (largest_power_of_five - smallest_power_of_five + 1);
+  // Powers of five from 5^-342 all the way to 5^308 rounded toward one.
+  constexpr static uint64_t power_of_five_128[number_of_entries] = {
+      0xeef453d6923bd65a, 0x113faa2906a13b3f,
+      0x9558b4661b6565f8, 0x4ac7ca59a424c507,
+      0xbaaee17fa23ebf76, 0x5d79bcf00d2df649,
+      0xe95a99df8ace6f53, 0xf4d82c2c107973dc,
+      0x91d8a02bb6c10594, 0x79071b9b8a4be869,
+      0xb64ec836a47146f9, 0x9748e2826cdee284,
+      0xe3e27a444d8d98b7, 0xfd1b1b2308169b25,
+      0x8e6d8c6ab0787f72, 0xfe30f0f5e50e20f7,
+      0xb208ef855c969f4f, 0xbdbd2d335e51a935,
+      0xde8b2b66b3bc4723, 0xad2c788035e61382,
+      0x8b16fb203055ac76, 0x4c3bcb5021afcc31,
+      0xaddcb9e83c6b1793, 0xdf4abe242a1bbf3d,
+      0xd953e8624b85dd78, 0xd71d6dad34a2af0d,
+      0x87d4713d6f33aa6b, 0x8672648c40e5ad68,
+      0xa9c98d8ccb009506, 0x680efdaf511f18c2,
+      0xd43bf0effdc0ba48, 0x212bd1b2566def2,
+      0x84a57695fe98746d, 0x14bb630f7604b57,
+      0xa5ced43b7e3e9188, 0x419ea3bd35385e2d,
+      0xcf42894a5dce35ea, 0x52064cac828675b9,
+      0x818995ce7aa0e1b2, 0x7343efebd1940993,
+      0xa1ebfb4219491a1f, 0x1014ebe6c5f90bf8,
+      0xca66fa129f9b60a6, 0xd41a26e077774ef6,
+      0xfd00b897478238d0, 0x8920b098955522b4,
+      0x9e20735e8cb16382, 0x55b46e5f5d5535b0,
+      0xc5a890362fddbc62, 0xeb2189f734aa831d,
+      0xf712b443bbd52b7b, 0xa5e9ec7501d523e4,
+      0x9a6bb0aa55653b2d, 0x47b233c92125366e,
+      0xc1069cd4eabe89f8, 0x999ec0bb696e840a,
+      0xf148440a256e2c76, 0xc00670ea43ca250d,
+      0x96cd2a865764dbca, 0x380406926a5e5728,
+      0xbc807527ed3e12bc, 0xc605083704f5ecf2,
+      0xeba09271e88d976b, 0xf7864a44c633682e,
+      0x93445b8731587ea3, 0x7ab3ee6afbe0211d,
+      0xb8157268fdae9e4c, 0x5960ea05bad82964,
+      0xe61acf033d1a45df, 0x6fb92487298e33bd,
+      0x8fd0c16206306bab, 0xa5d3b6d479f8e056,
+      0xb3c4f1ba87bc8696, 0x8f48a4899877186c,
+      0xe0b62e2929aba83c, 0x331acdabfe94de87,
+      0x8c71dcd9ba0b4925, 0x9ff0c08b7f1d0b14,
+      0xaf8e5410288e1b6f, 0x7ecf0ae5ee44dd9,
+      0xdb71e91432b1a24a, 0xc9e82cd9f69d6150,
+      0x892731ac9faf056e, 0xbe311c083a225cd2,
+      0xab70fe17c79ac6ca, 0x6dbd630a48aaf406,
+      0xd64d3d9db981787d, 0x92cbbccdad5b108,
+      0x85f0468293f0eb4e, 0x25bbf56008c58ea5,
+      0xa76c582338ed2621, 0xaf2af2b80af6f24e,
+      0xd1476e2c07286faa, 0x1af5af660db4aee1,
+      0x82cca4db847945ca, 0x50d98d9fc890ed4d,
+      0xa37fce126597973c, 0xe50ff107bab528a0,
+      0xcc5fc196fefd7d0c, 0x1e53ed49a96272c8,
+      0xff77b1fcbebcdc4f, 0x25e8e89c13bb0f7a,
+      0x9faacf3df73609b1, 0x77b191618c54e9ac,
+      0xc795830d75038c1d, 0xd59df5b9ef6a2417,
+      0xf97ae3d0d2446f25, 0x4b0573286b44ad1d,
+      0x9becce62836ac577, 0x4ee367f9430aec32,
+      0xc2e801fb244576d5, 0x229c41f793cda73f,
+      0xf3a20279ed56d48a, 0x6b43527578c1110f,
+      0x9845418c345644d6, 0x830a13896b78aaa9,
+      0xbe5691ef416bd60c, 0x23cc986bc656d553,
+      0xedec366b11c6cb8f, 0x2cbfbe86b7ec8aa8,
+      0x94b3a202eb1c3f39, 0x7bf7d71432f3d6a9,
+      0xb9e08a83a5e34f07, 0xdaf5ccd93fb0cc53,
+      0xe858ad248f5c22c9, 0xd1b3400f8f9cff68,
+      0x91376c36d99995be, 0x23100809b9c21fa1,
+      0xb58547448ffffb2d, 0xabd40a0c2832a78a,
+      0xe2e69915b3fff9f9, 0x16c90c8f323f516c,
+      0x8dd01fad907ffc3b, 0xae3da7d97f6792e3,
+      0xb1442798f49ffb4a, 0x99cd11cfdf41779c,
+      0xdd95317f31c7fa1d, 0x40405643d711d583,
+      0x8a7d3eef7f1cfc52, 0x482835ea666b2572,
+      0xad1c8eab5ee43b66, 0xda3243650005eecf,
+      0xd863b256369d4a40, 0x90bed43e40076a82,
+      0x873e4f75e2224e68, 0x5a7744a6e804a291,
+      0xa90de3535aaae202, 0x711515d0a205cb36,
+      0xd3515c2831559a83, 0xd5a5b44ca873e03,
+      0x8412d9991ed58091, 0xe858790afe9486c2,
+      0xa5178fff668ae0b6, 0x626e974dbe39a872,
+      0xce5d73ff402d98e3, 0xfb0a3d212dc8128f,
+      0x80fa687f881c7f8e, 0x7ce66634bc9d0b99,
+      0xa139029f6a239f72, 0x1c1fffc1ebc44e80,
+      0xc987434744ac874e, 0xa327ffb266b56220,
+      0xfbe9141915d7a922, 0x4bf1ff9f0062baa8,
+      0x9d71ac8fada6c9b5, 0x6f773fc3603db4a9,
+      0xc4ce17b399107c22, 0xcb550fb4384d21d3,
+      0xf6019da07f549b2b, 0x7e2a53a146606a48,
+      0x99c102844f94e0fb, 0x2eda7444cbfc426d,
+      0xc0314325637a1939, 0xfa911155fefb5308,
+      0xf03d93eebc589f88, 0x793555ab7eba27ca,
+      0x96267c7535b763b5, 0x4bc1558b2f3458de,
+      0xbbb01b9283253ca2, 0x9eb1aaedfb016f16,
+      0xea9c227723ee8bcb, 0x465e15a979c1cadc,
+      0x92a1958a7675175f, 0xbfacd89ec191ec9,
+      0xb749faed14125d36, 0xcef980ec671f667b,
+      0xe51c79a85916f484, 0x82b7e12780e7401a,
+      0x8f31cc0937ae58d2, 0xd1b2ecb8b0908810,
+      0xb2fe3f0b8599ef07, 0x861fa7e6dcb4aa15,
+      0xdfbdcece67006ac9, 0x67a791e093e1d49a,
+      0x8bd6a141006042bd, 0xe0c8bb2c5c6d24e0,
+      0xaecc49914078536d, 0x58fae9f773886e18,
+      0xda7f5bf590966848, 0xaf39a475506a899e,
+      0x888f99797a5e012d, 0x6d8406c952429603,
+      0xaab37fd7d8f58178, 0xc8e5087ba6d33b83,
+      0xd5605fcdcf32e1d6, 0xfb1e4a9a90880a64,
+      0x855c3be0a17fcd26, 0x5cf2eea09a55067f,
+      0xa6b34ad8c9dfc06f, 0xf42faa48c0ea481e,
+      0xd0601d8efc57b08b, 0xf13b94daf124da26,
+      0x823c12795db6ce57, 0x76c53d08d6b70858,
+      0xa2cb1717b52481ed, 0x54768c4b0c64ca6e,
+      0xcb7ddcdda26da268, 0xa9942f5dcf7dfd09,
+      0xfe5d54150b090b02, 0xd3f93b35435d7c4c,
+      0x9efa548d26e5a6e1, 0xc47bc5014a1a6daf,
+      0xc6b8e9b0709f109a, 0x359ab6419ca1091b,
+      0xf867241c8cc6d4c0, 0xc30163d203c94b62,
+      0x9b407691d7fc44f8, 0x79e0de63425dcf1d,
+      0xc21094364dfb5636, 0x985915fc12f542e4,
+      0xf294b943e17a2bc4, 0x3e6f5b7b17b2939d,
+      0x979cf3ca6cec5b5a, 0xa705992ceecf9c42,
+      0xbd8430bd08277231, 0x50c6ff782a838353,
+      0xece53cec4a314ebd, 0xa4f8bf5635246428,
+      0x940f4613ae5ed136, 0x871b7795e136be99,
+      0xb913179899f68584, 0x28e2557b59846e3f,
+      0xe757dd7ec07426e5, 0x331aeada2fe589cf,
+      0x9096ea6f3848984f, 0x3ff0d2c85def7621,
+      0xb4bca50b065abe63, 0xfed077a756b53a9,
+      0xe1ebce4dc7f16dfb, 0xd3e8495912c62894,
+      0x8d3360f09cf6e4bd, 0x64712dd7abbbd95c,
+      0xb080392cc4349dec, 0xbd8d794d96aacfb3,
+      0xdca04777f541c567, 0xecf0d7a0fc5583a0,
+      0x89e42caaf9491b60, 0xf41686c49db57244,
+      0xac5d37d5b79b6239, 0x311c2875c522ced5,
+      0xd77485cb25823ac7, 0x7d633293366b828b,
+      0x86a8d39ef77164bc, 0xae5dff9c02033197,
+      0xa8530886b54dbdeb, 0xd9f57f830283fdfc,
+      0xd267caa862a12d66, 0xd072df63c324fd7b,
+      0x8380dea93da4bc60, 0x4247cb9e59f71e6d,
+      0xa46116538d0deb78, 0x52d9be85f074e608,
+      0xcd795be870516656, 0x67902e276c921f8b,
+      0x806bd9714632dff6, 0xba1cd8a3db53b6,
+      0xa086cfcd97bf97f3, 0x80e8a40eccd228a4,
+      0xc8a883c0fdaf7df0, 0x6122cd128006b2cd,
+      0xfad2a4b13d1b5d6c, 0x796b805720085f81,
+      0x9cc3a6eec6311a63, 0xcbe3303674053bb0,
+      0xc3f490aa77bd60fc, 0xbedbfc4411068a9c,
+      0xf4f1b4d515acb93b, 0xee92fb5515482d44,
+      0x991711052d8bf3c5, 0x751bdd152d4d1c4a,
+      0xbf5cd54678eef0b6, 0xd262d45a78a0635d,
+      0xef340a98172aace4, 0x86fb897116c87c34,
+      0x9580869f0e7aac0e, 0xd45d35e6ae3d4da0,
+      0xbae0a846d2195712, 0x8974836059cca109,
+      0xe998d258869facd7, 0x2bd1a438703fc94b,
+      0x91ff83775423cc06, 0x7b6306a34627ddcf,
+      0xb67f6455292cbf08, 0x1a3bc84c17b1d542,
+      0xe41f3d6a7377eeca, 0x20caba5f1d9e4a93,
+      0x8e938662882af53e, 0x547eb47b7282ee9c,
+      0xb23867fb2a35b28d, 0xe99e619a4f23aa43,
+      0xdec681f9f4c31f31, 0x6405fa00e2ec94d4,
+      0x8b3c113c38f9f37e, 0xde83bc408dd3dd04,
+      0xae0b158b4738705e, 0x9624ab50b148d445,
+      0xd98ddaee19068c76, 0x3badd624dd9b0957,
+      0x87f8a8d4cfa417c9, 0xe54ca5d70a80e5d6,
+      0xa9f6d30a038d1dbc, 0x5e9fcf4ccd211f4c,
+      0xd47487cc8470652b, 0x7647c3200069671f,
+      0x84c8d4dfd2c63f3b, 0x29ecd9f40041e073,
+      0xa5fb0a17c777cf09, 0xf468107100525890,
+      0xcf79cc9db955c2cc, 0x7182148d4066eeb4,
+      0x81ac1fe293d599bf, 0xc6f14cd848405530,
+      0xa21727db38cb002f, 0xb8ada00e5a506a7c,
+      0xca9cf1d206fdc03b, 0xa6d90811f0e4851c,
+      0xfd442e4688bd304a, 0x908f4a166d1da663,
+      0x9e4a9cec15763e2e, 0x9a598e4e043287fe,
+      0xc5dd44271ad3cdba, 0x40eff1e1853f29fd,
+      0xf7549530e188c128, 0xd12bee59e68ef47c,
+      0x9a94dd3e8cf578b9, 0x82bb74f8301958ce,
+      0xc13a148e3032d6e7, 0xe36a52363c1faf01,
+      0xf18899b1bc3f8ca1, 0xdc44e6c3cb279ac1,
+      0x96f5600f15a7b7e5, 0x29ab103a5ef8c0b9,
+      0xbcb2b812db11a5de, 0x7415d448f6b6f0e7,
+      0xebdf661791d60f56, 0x111b495b3464ad21,
+      0x936b9fcebb25c995, 0xcab10dd900beec34,
+      0xb84687c269ef3bfb, 0x3d5d514f40eea742,
+      0xe65829b3046b0afa, 0xcb4a5a3112a5112,
+      0x8ff71a0fe2c2e6dc, 0x47f0e785eaba72ab,
+      0xb3f4e093db73a093, 0x59ed216765690f56,
+      0xe0f218b8d25088b8, 0x306869c13ec3532c,
+      0x8c974f7383725573, 0x1e414218c73a13fb,
+      0xafbd2350644eeacf, 0xe5d1929ef90898fa,
+      0xdbac6c247d62a583, 0xdf45f746b74abf39,
+      0x894bc396ce5da772, 0x6b8bba8c328eb783,
+      0xab9eb47c81f5114f, 0x66ea92f3f326564,
+      0xd686619ba27255a2, 0xc80a537b0efefebd,
+      0x8613fd0145877585, 0xbd06742ce95f5f36,
+      0xa798fc4196e952e7, 0x2c48113823b73704,
+      0xd17f3b51fca3a7a0, 0xf75a15862ca504c5,
+      0x82ef85133de648c4, 0x9a984d73dbe722fb,
+      0xa3ab66580d5fdaf5, 0xc13e60d0d2e0ebba,
+      0xcc963fee10b7d1b3, 0x318df905079926a8,
+      0xffbbcfe994e5c61f, 0xfdf17746497f7052,
+      0x9fd561f1fd0f9bd3, 0xfeb6ea8bedefa633,
+      0xc7caba6e7c5382c8, 0xfe64a52ee96b8fc0,
+      0xf9bd690a1b68637b, 0x3dfdce7aa3c673b0,
+      0x9c1661a651213e2d, 0x6bea10ca65c084e,
+      0xc31bfa0fe5698db8, 0x486e494fcff30a62,
+      0xf3e2f893dec3f126, 0x5a89dba3c3efccfa,
+      0x986ddb5c6b3a76b7, 0xf89629465a75e01c,
+      0xbe89523386091465, 0xf6bbb397f1135823,
+      0xee2ba6c0678b597f, 0x746aa07ded582e2c,
+      0x94db483840b717ef, 0xa8c2a44eb4571cdc,
+      0xba121a4650e4ddeb, 0x92f34d62616ce413,
+      0xe896a0d7e51e1566, 0x77b020baf9c81d17,
+      0x915e2486ef32cd60, 0xace1474dc1d122e,
+      0xb5b5ada8aaff80b8, 0xd819992132456ba,
+      0xe3231912d5bf60e6, 0x10e1fff697ed6c69,
+      0x8df5efabc5979c8f, 0xca8d3ffa1ef463c1,
+      0xb1736b96b6fd83b3, 0xbd308ff8a6b17cb2,
+      0xddd0467c64bce4a0, 0xac7cb3f6d05ddbde,
+      0x8aa22c0dbef60ee4, 0x6bcdf07a423aa96b,
+      0xad4ab7112eb3929d, 0x86c16c98d2c953c6,
+      0xd89d64d57a607744, 0xe871c7bf077ba8b7,
+      0x87625f056c7c4a8b, 0x11471cd764ad4972,
+      0xa93af6c6c79b5d2d, 0xd598e40d3dd89bcf,
+      0xd389b47879823479, 0x4aff1d108d4ec2c3,
+      0x843610cb4bf160cb, 0xcedf722a585139ba,
+      0xa54394fe1eedb8fe, 0xc2974eb4ee658828,
+      0xce947a3da6a9273e, 0x733d226229feea32,
+      0x811ccc668829b887, 0x806357d5a3f525f,
+      0xa163ff802a3426a8, 0xca07c2dcb0cf26f7,
+      0xc9bcff6034c13052, 0xfc89b393dd02f0b5,
+      0xfc2c3f3841f17c67, 0xbbac2078d443ace2,
+      0x9d9ba7832936edc0, 0xd54b944b84aa4c0d,
+      0xc5029163f384a931, 0xa9e795e65d4df11,
+      0xf64335bcf065d37d, 0x4d4617b5ff4a16d5,
+      0x99ea0196163fa42e, 0x504bced1bf8e4e45,
+      0xc06481fb9bcf8d39, 0xe45ec2862f71e1d6,
+      0xf07da27a82c37088, 0x5d767327bb4e5a4c,
+      0x964e858c91ba2655, 0x3a6a07f8d510f86f,
+      0xbbe226efb628afea, 0x890489f70a55368b,
+      0xeadab0aba3b2dbe5, 0x2b45ac74ccea842e,
+      0x92c8ae6b464fc96f, 0x3b0b8bc90012929d,
+      0xb77ada0617e3bbcb, 0x9ce6ebb40173744,
+      0xe55990879ddcaabd, 0xcc420a6a101d0515,
+      0x8f57fa54c2a9eab6, 0x9fa946824a12232d,
+      0xb32df8e9f3546564, 0x47939822dc96abf9,
+      0xdff9772470297ebd, 0x59787e2b93bc56f7,
+      0x8bfbea76c619ef36, 0x57eb4edb3c55b65a,
+      0xaefae51477a06b03, 0xede622920b6b23f1,
+      0xdab99e59958885c4, 0xe95fab368e45eced,
+      0x88b402f7fd75539b, 0x11dbcb0218ebb414,
+      0xaae103b5fcd2a881, 0xd652bdc29f26a119,
+      0xd59944a37c0752a2, 0x4be76d3346f0495f,
+      0x857fcae62d8493a5, 0x6f70a4400c562ddb,
+      0xa6dfbd9fb8e5b88e, 0xcb4ccd500f6bb952,
+      0xd097ad07a71f26b2, 0x7e2000a41346a7a7,
+      0x825ecc24c873782f, 0x8ed400668c0c28c8,
+      0xa2f67f2dfa90563b, 0x728900802f0f32fa,
+      0xcbb41ef979346bca, 0x4f2b40a03ad2ffb9,
+      0xfea126b7d78186bc, 0xe2f610c84987bfa8,
+      0x9f24b832e6b0f436, 0xdd9ca7d2df4d7c9,
+      0xc6ede63fa05d3143, 0x91503d1c79720dbb,
+      0xf8a95fcf88747d94, 0x75a44c6397ce912a,
+      0x9b69dbe1b548ce7c, 0xc986afbe3ee11aba,
+      0xc24452da229b021b, 0xfbe85badce996168,
+      0xf2d56790ab41c2a2, 0xfae27299423fb9c3,
+      0x97c560ba6b0919a5, 0xdccd879fc967d41a,
+      0xbdb6b8e905cb600f, 0x5400e987bbc1c920,
+      0xed246723473e3813, 0x290123e9aab23b68,
+      0x9436c0760c86e30b, 0xf9a0b6720aaf6521,
+      0xb94470938fa89bce, 0xf808e40e8d5b3e69,
+      0xe7958cb87392c2c2, 0xb60b1d1230b20e04,
+      0x90bd77f3483bb9b9, 0xb1c6f22b5e6f48c2,
+      0xb4ecd5f01a4aa828, 0x1e38aeb6360b1af3,
+      0xe2280b6c20dd5232, 0x25c6da63c38de1b0,
+      0x8d590723948a535f, 0x579c487e5a38ad0e,
+      0xb0af48ec79ace837, 0x2d835a9df0c6d851,
+      0xdcdb1b2798182244, 0xf8e431456cf88e65,
+      0x8a08f0f8bf0f156b, 0x1b8e9ecb641b58ff,
+      0xac8b2d36eed2dac5, 0xe272467e3d222f3f,
+      0xd7adf884aa879177, 0x5b0ed81dcc6abb0f,
+      0x86ccbb52ea94baea, 0x98e947129fc2b4e9,
+      0xa87fea27a539e9a5, 0x3f2398d747b36224,
+      0xd29fe4b18e88640e, 0x8eec7f0d19a03aad,
+      0x83a3eeeef9153e89, 0x1953cf68300424ac,
+      0xa48ceaaab75a8e2b, 0x5fa8c3423c052dd7,
+      0xcdb02555653131b6, 0x3792f412cb06794d,
+      0x808e17555f3ebf11, 0xe2bbd88bbee40bd0,
+      0xa0b19d2ab70e6ed6, 0x5b6aceaeae9d0ec4,
+      0xc8de047564d20a8b, 0xf245825a5a445275,
+      0xfb158592be068d2e, 0xeed6e2f0f0d56712,
+      0x9ced737bb6c4183d, 0x55464dd69685606b,
+      0xc428d05aa4751e4c, 0xaa97e14c3c26b886,
+      0xf53304714d9265df, 0xd53dd99f4b3066a8,
+      0x993fe2c6d07b7fab, 0xe546a8038efe4029,
+      0xbf8fdb78849a5f96, 0xde98520472bdd033,
+      0xef73d256a5c0f77c, 0x963e66858f6d4440,
+      0x95a8637627989aad, 0xdde7001379a44aa8,
+      0xbb127c53b17ec159, 0x5560c018580d5d52,
+      0xe9d71b689dde71af, 0xaab8f01e6e10b4a6,
+      0x9226712162ab070d, 0xcab3961304ca70e8,
+      0xb6b00d69bb55c8d1, 0x3d607b97c5fd0d22,
+      0xe45c10c42a2b3b05, 0x8cb89a7db77c506a,
+      0x8eb98a7a9a5b04e3, 0x77f3608e92adb242,
+      0xb267ed1940f1c61c, 0x55f038b237591ed3,
+      0xdf01e85f912e37a3, 0x6b6c46dec52f6688,
+      0x8b61313bbabce2c6, 0x2323ac4b3b3da015,
+      0xae397d8aa96c1b77, 0xabec975e0a0d081a,
+      0xd9c7dced53c72255, 0x96e7bd358c904a21,
+      0x881cea14545c7575, 0x7e50d64177da2e54,
+      0xaa242499697392d2, 0xdde50bd1d5d0b9e9,
+      0xd4ad2dbfc3d07787, 0x955e4ec64b44e864,
+      0x84ec3c97da624ab4, 0xbd5af13bef0b113e,
+      0xa6274bbdd0fadd61, 0xecb1ad8aeacdd58e,
+      0xcfb11ead453994ba, 0x67de18eda5814af2,
+      0x81ceb32c4b43fcf4, 0x80eacf948770ced7,
+      0xa2425ff75e14fc31, 0xa1258379a94d028d,
+      0xcad2f7f5359a3b3e, 0x96ee45813a04330,
+      0xfd87b5f28300ca0d, 0x8bca9d6e188853fc,
+      0x9e74d1b791e07e48, 0x775ea264cf55347e,
+      0xc612062576589dda, 0x95364afe032a819e,
+      0xf79687aed3eec551, 0x3a83ddbd83f52205,
+      0x9abe14cd44753b52, 0xc4926a9672793543,
+      0xc16d9a0095928a27, 0x75b7053c0f178294,
+      0xf1c90080baf72cb1, 0x5324c68b12dd6339,
+      0x971da05074da7bee, 0xd3f6fc16ebca5e04,
+      0xbce5086492111aea, 0x88f4bb1ca6bcf585,
+      0xec1e4a7db69561a5, 0x2b31e9e3d06c32e6,
+      0x9392ee8e921d5d07, 0x3aff322e62439fd0,
+      0xb877aa3236a4b449, 0x9befeb9fad487c3,
+      0xe69594bec44de15b, 0x4c2ebe687989a9b4,
+      0x901d7cf73ab0acd9, 0xf9d37014bf60a11,
+      0xb424dc35095cd80f, 0x538484c19ef38c95,
+      0xe12e13424bb40e13, 0x2865a5f206b06fba,
+      0x8cbccc096f5088cb, 0xf93f87b7442e45d4,
+      0xafebff0bcb24aafe, 0xf78f69a51539d749,
+      0xdbe6fecebdedd5be, 0xb573440e5a884d1c,
+      0x89705f4136b4a597, 0x31680a88f8953031,
+      0xabcc77118461cefc, 0xfdc20d2b36ba7c3e,
+      0xd6bf94d5e57a42bc, 0x3d32907604691b4d,
+      0x8637bd05af6c69b5, 0xa63f9a49c2c1b110,
+      0xa7c5ac471b478423, 0xfcf80dc33721d54,
+      0xd1b71758e219652b, 0xd3c36113404ea4a9,
+      0x83126e978d4fdf3b, 0x645a1cac083126ea,
+      0xa3d70a3d70a3d70a, 0x3d70a3d70a3d70a4,
+      0xcccccccccccccccc, 0xcccccccccccccccd,
+      0x8000000000000000, 0x0,
+      0xa000000000000000, 0x0,
+      0xc800000000000000, 0x0,
+      0xfa00000000000000, 0x0,
+      0x9c40000000000000, 0x0,
+      0xc350000000000000, 0x0,
+      0xf424000000000000, 0x0,
+      0x9896800000000000, 0x0,
+      0xbebc200000000000, 0x0,
+      0xee6b280000000000, 0x0,
+      0x9502f90000000000, 0x0,
+      0xba43b74000000000, 0x0,
+      0xe8d4a51000000000, 0x0,
+      0x9184e72a00000000, 0x0,
+      0xb5e620f480000000, 0x0,
+      0xe35fa931a0000000, 0x0,
+      0x8e1bc9bf04000000, 0x0,
+      0xb1a2bc2ec5000000, 0x0,
+      0xde0b6b3a76400000, 0x0,
+      0x8ac7230489e80000, 0x0,
+      0xad78ebc5ac620000, 0x0,
+      0xd8d726b7177a8000, 0x0,
+      0x878678326eac9000, 0x0,
+      0xa968163f0a57b400, 0x0,
+      0xd3c21bcecceda100, 0x0,
+      0x84595161401484a0, 0x0,
+      0xa56fa5b99019a5c8, 0x0,
+      0xcecb8f27f4200f3a, 0x0,
+      0x813f3978f8940984, 0x4000000000000000,
+      0xa18f07d736b90be5, 0x5000000000000000,
+      0xc9f2c9cd04674ede, 0xa400000000000000,
+      0xfc6f7c4045812296, 0x4d00000000000000,
+      0x9dc5ada82b70b59d, 0xf020000000000000,
+      0xc5371912364ce305, 0x6c28000000000000,
+      0xf684df56c3e01bc6, 0xc732000000000000,
+      0x9a130b963a6c115c, 0x3c7f400000000000,
+      0xc097ce7bc90715b3, 0x4b9f100000000000,
+      0xf0bdc21abb48db20, 0x1e86d40000000000,
+      0x96769950b50d88f4, 0x1314448000000000,
+      0xbc143fa4e250eb31, 0x17d955a000000000,
+      0xeb194f8e1ae525fd, 0x5dcfab0800000000,
+      0x92efd1b8d0cf37be, 0x5aa1cae500000000,
+      0xb7abc627050305ad, 0xf14a3d9e40000000,
+      0xe596b7b0c643c719, 0x6d9ccd05d0000000,
+      0x8f7e32ce7bea5c6f, 0xe4820023a2000000,
+      0xb35dbf821ae4f38b, 0xdda2802c8a800000,
+      0xe0352f62a19e306e, 0xd50b2037ad200000,
+      0x8c213d9da502de45, 0x4526f422cc340000,
+      0xaf298d050e4395d6, 0x9670b12b7f410000,
+      0xdaf3f04651d47b4c, 0x3c0cdd765f114000,
+      0x88d8762bf324cd0f, 0xa5880a69fb6ac800,
+      0xab0e93b6efee0053, 0x8eea0d047a457a00,
+      0xd5d238a4abe98068, 0x72a4904598d6d880,
+      0x85a36366eb71f041, 0x47a6da2b7f864750,
+      0xa70c3c40a64e6c51, 0x999090b65f67d924,
+      0xd0cf4b50cfe20765, 0xfff4b4e3f741cf6d,
+      0x82818f1281ed449f, 0xbff8f10e7a8921a4,
+      0xa321f2d7226895c7, 0xaff72d52192b6a0d,
+      0xcbea6f8ceb02bb39, 0x9bf4f8a69f764490,
+      0xfee50b7025c36a08, 0x2f236d04753d5b4,
+      0x9f4f2726179a2245, 0x1d762422c946590,
+      0xc722f0ef9d80aad6, 0x424d3ad2b7b97ef5,
+      0xf8ebad2b84e0d58b, 0xd2e0898765a7deb2,
+      0x9b934c3b330c8577, 0x63cc55f49f88eb2f,
+      0xc2781f49ffcfa6d5, 0x3cbf6b71c76b25fb,
+      0xf316271c7fc3908a, 0x8bef464e3945ef7a,
+      0x97edd871cfda3a56, 0x97758bf0e3cbb5ac,
+      0xbde94e8e43d0c8ec, 0x3d52eeed1cbea317,
+      0xed63a231d4c4fb27, 0x4ca7aaa863ee4bdd,
+      0x945e455f24fb1cf8, 0x8fe8caa93e74ef6a,
+      0xb975d6b6ee39e436, 0xb3e2fd538e122b44,
+      0xe7d34c64a9c85d44, 0x60dbbca87196b616,
+      0x90e40fbeea1d3a4a, 0xbc8955e946fe31cd,
+      0xb51d13aea4a488dd, 0x6babab6398bdbe41,
+      0xe264589a4dcdab14, 0xc696963c7eed2dd1,
+      0x8d7eb76070a08aec, 0xfc1e1de5cf543ca2,
+      0xb0de65388cc8ada8, 0x3b25a55f43294bcb,
+      0xdd15fe86affad912, 0x49ef0eb713f39ebe,
+      0x8a2dbf142dfcc7ab, 0x6e3569326c784337,
+      0xacb92ed9397bf996, 0x49c2c37f07965404,
+      0xd7e77a8f87daf7fb, 0xdc33745ec97be906,
+      0x86f0ac99b4e8dafd, 0x69a028bb3ded71a3,
+      0xa8acd7c0222311bc, 0xc40832ea0d68ce0c,
+      0xd2d80db02aabd62b, 0xf50a3fa490c30190,
+      0x83c7088e1aab65db, 0x792667c6da79e0fa,
+      0xa4b8cab1a1563f52, 0x577001b891185938,
+      0xcde6fd5e09abcf26, 0xed4c0226b55e6f86,
+      0x80b05e5ac60b6178, 0x544f8158315b05b4,
+      0xa0dc75f1778e39d6, 0x696361ae3db1c721,
+      0xc913936dd571c84c, 0x3bc3a19cd1e38e9,
+      0xfb5878494ace3a5f, 0x4ab48a04065c723,
+      0x9d174b2dcec0e47b, 0x62eb0d64283f9c76,
+      0xc45d1df942711d9a, 0x3ba5d0bd324f8394,
+      0xf5746577930d6500, 0xca8f44ec7ee36479,
+      0x9968bf6abbe85f20, 0x7e998b13cf4e1ecb,
+      0xbfc2ef456ae276e8, 0x9e3fedd8c321a67e,
+      0xefb3ab16c59b14a2, 0xc5cfe94ef3ea101e,
+      0x95d04aee3b80ece5, 0xbba1f1d158724a12,
+      0xbb445da9ca61281f, 0x2a8a6e45ae8edc97,
+      0xea1575143cf97226, 0xf52d09d71a3293bd,
+      0x924d692ca61be758, 0x593c2626705f9c56,
+      0xb6e0c377cfa2e12e, 0x6f8b2fb00c77836c,
+      0xe498f455c38b997a, 0xb6dfb9c0f956447,
+      0x8edf98b59a373fec, 0x4724bd4189bd5eac,
+      0xb2977ee300c50fe7, 0x58edec91ec2cb657,
+      0xdf3d5e9bc0f653e1, 0x2f2967b66737e3ed,
+      0x8b865b215899f46c, 0xbd79e0d20082ee74,
+      0xae67f1e9aec07187, 0xecd8590680a3aa11,
+      0xda01ee641a708de9, 0xe80e6f4820cc9495,
+      0x884134fe908658b2, 0x3109058d147fdcdd,
+      0xaa51823e34a7eede, 0xbd4b46f0599fd415,
+      0xd4e5e2cdc1d1ea96, 0x6c9e18ac7007c91a,
+      0x850fadc09923329e, 0x3e2cf6bc604ddb0,
+      0xa6539930bf6bff45, 0x84db8346b786151c,
+      0xcfe87f7cef46ff16, 0xe612641865679a63,
+      0x81f14fae158c5f6e, 0x4fcb7e8f3f60c07e,
+      0xa26da3999aef7749, 0xe3be5e330f38f09d,
+      0xcb090c8001ab551c, 0x5cadf5bfd3072cc5,
+      0xfdcb4fa002162a63, 0x73d9732fc7c8f7f6,
+      0x9e9f11c4014dda7e, 0x2867e7fddcdd9afa,
+      0xc646d63501a1511d, 0xb281e1fd541501b8,
+      0xf7d88bc24209a565, 0x1f225a7ca91a4226,
+      0x9ae757596946075f, 0x3375788de9b06958,
+      0xc1a12d2fc3978937, 0x52d6b1641c83ae,
+      0xf209787bb47d6b84, 0xc0678c5dbd23a49a,
+      0x9745eb4d50ce6332, 0xf840b7ba963646e0,
+      0xbd176620a501fbff, 0xb650e5a93bc3d898,
+      0xec5d3fa8ce427aff, 0xa3e51f138ab4cebe,
+      0x93ba47c980e98cdf, 0xc66f336c36b10137,
+      0xb8a8d9bbe123f017, 0xb80b0047445d4184,
+      0xe6d3102ad96cec1d, 0xa60dc059157491e5,
+      0x9043ea1ac7e41392, 0x87c89837ad68db2f,
+      0xb454e4a179dd1877, 0x29babe4598c311fb,
+      0xe16a1dc9d8545e94, 0xf4296dd6fef3d67a,
+      0x8ce2529e2734bb1d, 0x1899e4a65f58660c,
+      0xb01ae745b101e9e4, 0x5ec05dcff72e7f8f,
+      0xdc21a1171d42645d, 0x76707543f4fa1f73,
+      0x899504ae72497eba, 0x6a06494a791c53a8,
+      0xabfa45da0edbde69, 0x487db9d17636892,
+      0xd6f8d7509292d603, 0x45a9d2845d3c42b6,
+      0x865b86925b9bc5c2, 0xb8a2392ba45a9b2,
+      0xa7f26836f282b732, 0x8e6cac7768d7141e,
+      0xd1ef0244af2364ff, 0x3207d795430cd926,
+      0x8335616aed761f1f, 0x7f44e6bd49e807b8,
+      0xa402b9c5a8d3a6e7, 0x5f16206c9c6209a6,
+      0xcd036837130890a1, 0x36dba887c37a8c0f,
+      0x802221226be55a64, 0xc2494954da2c9789,
+      0xa02aa96b06deb0fd, 0xf2db9baa10b7bd6c,
+      0xc83553c5c8965d3d, 0x6f92829494e5acc7,
+      0xfa42a8b73abbf48c, 0xcb772339ba1f17f9,
+      0x9c69a97284b578d7, 0xff2a760414536efb,
+      0xc38413cf25e2d70d, 0xfef5138519684aba,
+      0xf46518c2ef5b8cd1, 0x7eb258665fc25d69,
+      0x98bf2f79d5993802, 0xef2f773ffbd97a61,
+      0xbeeefb584aff8603, 0xaafb550ffacfd8fa,
+      0xeeaaba2e5dbf6784, 0x95ba2a53f983cf38,
+      0x952ab45cfa97a0b2, 0xdd945a747bf26183,
+      0xba756174393d88df, 0x94f971119aeef9e4,
+      0xe912b9d1478ceb17, 0x7a37cd5601aab85d,
+      0x91abb422ccb812ee, 0xac62e055c10ab33a,
+      0xb616a12b7fe617aa, 0x577b986b314d6009,
+      0xe39c49765fdf9d94, 0xed5a7e85fda0b80b,
+      0x8e41ade9fbebc27d, 0x14588f13be847307,
+      0xb1d219647ae6b31c, 0x596eb2d8ae258fc8,
+      0xde469fbd99a05fe3, 0x6fca5f8ed9aef3bb,
+      0x8aec23d680043bee, 0x25de7bb9480d5854,
+      0xada72ccc20054ae9, 0xaf561aa79a10ae6a,
+      0xd910f7ff28069da4, 0x1b2ba1518094da04,
+      0x87aa9aff79042286, 0x90fb44d2f05d0842,
+      0xa99541bf57452b28, 0x353a1607ac744a53,
+      0xd3fa922f2d1675f2, 0x42889b8997915ce8,
+      0x847c9b5d7c2e09b7, 0x69956135febada11,
+      0xa59bc234db398c25, 0x43fab9837e699095,
+      0xcf02b2c21207ef2e, 0x94f967e45e03f4bb,
+      0x8161afb94b44f57d, 0x1d1be0eebac278f5,
+      0xa1ba1ba79e1632dc, 0x6462d92a69731732,
+      0xca28a291859bbf93, 0x7d7b8f7503cfdcfe,
+      0xfcb2cb35e702af78, 0x5cda735244c3d43e,
+      0x9defbf01b061adab, 0x3a0888136afa64a7,
+      0xc56baec21c7a1916, 0x88aaa1845b8fdd0,
+      0xf6c69a72a3989f5b, 0x8aad549e57273d45,
+      0x9a3c2087a63f6399, 0x36ac54e2f678864b,
+      0xc0cb28a98fcf3c7f, 0x84576a1bb416a7dd,
+      0xf0fdf2d3f3c30b9f, 0x656d44a2a11c51d5,
+      0x969eb7c47859e743, 0x9f644ae5a4b1b325,
+      0xbc4665b596706114, 0x873d5d9f0dde1fee,
+      0xeb57ff22fc0c7959, 0xa90cb506d155a7ea,
+      0x9316ff75dd87cbd8, 0x9a7f12442d588f2,
+      0xb7dcbf5354e9bece, 0xc11ed6d538aeb2f,
+      0xe5d3ef282a242e81, 0x8f1668c8a86da5fa,
+      0x8fa475791a569d10, 0xf96e017d694487bc,
+      0xb38d92d760ec4455, 0x37c981dcc395a9ac,
+      0xe070f78d3927556a, 0x85bbe253f47b1417,
+      0x8c469ab843b89562, 0x93956d7478ccec8e,
+      0xaf58416654a6babb, 0x387ac8d1970027b2,
+      0xdb2e51bfe9d0696a, 0x6997b05fcc0319e,
+      0x88fcf317f22241e2, 0x441fece3bdf81f03,
+      0xab3c2fddeeaad25a, 0xd527e81cad7626c3,
+      0xd60b3bd56a5586f1, 0x8a71e223d8d3b074,
+      0x85c7056562757456, 0xf6872d5667844e49,
+      0xa738c6bebb12d16c, 0xb428f8ac016561db,
+      0xd106f86e69d785c7, 0xe13336d701beba52,
+      0x82a45b450226b39c, 0xecc0024661173473,
+      0xa34d721642b06084, 0x27f002d7f95d0190,
+      0xcc20ce9bd35c78a5, 0x31ec038df7b441f4,
+      0xff290242c83396ce, 0x7e67047175a15271,
+      0x9f79a169bd203e41, 0xf0062c6e984d386,
+      0xc75809c42c684dd1, 0x52c07b78a3e60868,
+      0xf92e0c3537826145, 0xa7709a56ccdf8a82,
+      0x9bbcc7a142b17ccb, 0x88a66076400bb691,
+      0xc2abf989935ddbfe, 0x6acff893d00ea435,
+      0xf356f7ebf83552fe, 0x583f6b8c4124d43,
+      0x98165af37b2153de, 0xc3727a337a8b704a,
+      0xbe1bf1b059e9a8d6, 0x744f18c0592e4c5c,
+      0xeda2ee1c7064130c, 0x1162def06f79df73,
+      0x9485d4d1c63e8be7, 0x8addcb5645ac2ba8,
+      0xb9a74a0637ce2ee1, 0x6d953e2bd7173692,
+      0xe8111c87c5c1ba99, 0xc8fa8db6ccdd0437,
+      0x910ab1d4db9914a0, 0x1d9c9892400a22a2,
+      0xb54d5e4a127f59c8, 0x2503beb6d00cab4b,
+      0xe2a0b5dc971f303a, 0x2e44ae64840fd61d,
+      0x8da471a9de737e24, 0x5ceaecfed289e5d2,
+      0xb10d8e1456105dad, 0x7425a83e872c5f47,
+      0xdd50f1996b947518, 0xd12f124e28f77719,
+      0x8a5296ffe33cc92f, 0x82bd6b70d99aaa6f,
+      0xace73cbfdc0bfb7b, 0x636cc64d1001550b,
+      0xd8210befd30efa5a, 0x3c47f7e05401aa4e,
+      0x8714a775e3e95c78, 0x65acfaec34810a71,
+      0xa8d9d1535ce3b396, 0x7f1839a741a14d0d,
+      0xd31045a8341ca07c, 0x1ede48111209a050,
+      0x83ea2b892091e44d, 0x934aed0aab460432,
+      0xa4e4b66b68b65d60, 0xf81da84d5617853f,
+      0xce1de40642e3f4b9, 0x36251260ab9d668e,
+      0x80d2ae83e9ce78f3, 0xc1d72b7c6b426019,
+      0xa1075a24e4421730, 0xb24cf65b8612f81f,
+      0xc94930ae1d529cfc, 0xdee033f26797b627,
+      0xfb9b7cd9a4a7443c, 0x169840ef017da3b1,
+      0x9d412e0806e88aa5, 0x8e1f289560ee864e,
+      0xc491798a08a2ad4e, 0xf1a6f2bab92a27e2,
+      0xf5b5d7ec8acb58a2, 0xae10af696774b1db,
+      0x9991a6f3d6bf1765, 0xacca6da1e0a8ef29,
+      0xbff610b0cc6edd3f, 0x17fd090a58d32af3,
+      0xeff394dcff8a948e, 0xddfc4b4cef07f5b0,
+      0x95f83d0a1fb69cd9, 0x4abdaf101564f98e,
+      0xbb764c4ca7a4440f, 0x9d6d1ad41abe37f1,
+      0xea53df5fd18d5513, 0x84c86189216dc5ed,
+      0x92746b9be2f8552c, 0x32fd3cf5b4e49bb4,
+      0xb7118682dbb66a77, 0x3fbc8c33221dc2a1,
+      0xe4d5e82392a40515, 0xfabaf3feaa5334a,
+      0x8f05b1163ba6832d, 0x29cb4d87f2a7400e,
+      0xb2c71d5bca9023f8, 0x743e20e9ef511012,
+      0xdf78e4b2bd342cf6, 0x914da9246b255416,
+      0x8bab8eefb6409c1a, 0x1ad089b6c2f7548e,
+      0xae9672aba3d0c320, 0xa184ac2473b529b1,
+      0xda3c0f568cc4f3e8, 0xc9e5d72d90a2741e,
+      0x8865899617fb1871, 0x7e2fa67c7a658892,
+      0xaa7eebfb9df9de8d, 0xddbb901b98feeab7,
+      0xd51ea6fa85785631, 0x552a74227f3ea565,
+      0x8533285c936b35de, 0xd53a88958f87275f,
+      0xa67ff273b8460356, 0x8a892abaf368f137,
+      0xd01fef10a657842c, 0x2d2b7569b0432d85,
+      0x8213f56a67f6b29b, 0x9c3b29620e29fc73,
+      0xa298f2c501f45f42, 0x8349f3ba91b47b8f,
+      0xcb3f2f7642717713, 0x241c70a936219a73,
+      0xfe0efb53d30dd4d7, 0xed238cd383aa0110,
+      0x9ec95d1463e8a506, 0xf4363804324a40aa,
+      0xc67bb4597ce2ce48, 0xb143c6053edcd0d5,
+      0xf81aa16fdc1b81da, 0xdd94b7868e94050a,
+      0x9b10a4e5e9913128, 0xca7cf2b4191c8326,
+      0xc1d4ce1f63f57d72, 0xfd1c2f611f63a3f0,
+      0xf24a01a73cf2dccf, 0xbc633b39673c8cec,
+      0x976e41088617ca01, 0xd5be0503e085d813,
+      0xbd49d14aa79dbc82, 0x4b2d8644d8a74e18,
+      0xec9c459d51852ba2, 0xddf8e7d60ed1219e,
+      0x93e1ab8252f33b45, 0xcabb90e5c942b503,
+      0xb8da1662e7b00a17, 0x3d6a751f3b936243,
+      0xe7109bfba19c0c9d, 0xcc512670a783ad4,
+      0x906a617d450187e2, 0x27fb2b80668b24c5,
+      0xb484f9dc9641e9da, 0xb1f9f660802dedf6,
+      0xe1a63853bbd26451, 0x5e7873f8a0396973,
+      0x8d07e33455637eb2, 0xdb0b487b6423e1e8,
+      0xb049dc016abc5e5f, 0x91ce1a9a3d2cda62,
+      0xdc5c5301c56b75f7, 0x7641a140cc7810fb,
+      0x89b9b3e11b6329ba, 0xa9e904c87fcb0a9d,
+      0xac2820d9623bf429, 0x546345fa9fbdcd44,
+      0xd732290fbacaf133, 0xa97c177947ad4095,
+      0x867f59a9d4bed6c0, 0x49ed8eabcccc485d,
+      0xa81f301449ee8c70, 0x5c68f256bfff5a74,
+      0xd226fc195c6a2f8c, 0x73832eec6fff3111,
+      0x83585d8fd9c25db7, 0xc831fd53c5ff7eab,
+      0xa42e74f3d032f525, 0xba3e7ca8b77f5e55,
+      0xcd3a1230c43fb26f, 0x28ce1bd2e55f35eb,
+      0x80444b5e7aa7cf85, 0x7980d163cf5b81b3,
+      0xa0555e361951c366, 0xd7e105bcc332621f,
+      0xc86ab5c39fa63440, 0x8dd9472bf3fefaa7,
+      0xfa856334878fc150, 0xb14f98f6f0feb951,
+      0x9c935e00d4b9d8d2, 0x6ed1bf9a569f33d3,
+      0xc3b8358109e84f07, 0xa862f80ec4700c8,
+      0xf4a642e14c6262c8, 0xcd27bb612758c0fa,
+      0x98e7e9cccfbd7dbd, 0x8038d51cb897789c,
+      0xbf21e44003acdd2c, 0xe0470a63e6bd56c3,
+      0xeeea5d5004981478, 0x1858ccfce06cac74,
+      0x95527a5202df0ccb, 0xf37801e0c43ebc8,
+      0xbaa718e68396cffd, 0xd30560258f54e6ba,
+      0xe950df20247c83fd, 0x47c6b82ef32a2069,
+      0x91d28b7416cdd27e, 0x4cdc331d57fa5441,
+      0xb6472e511c81471d, 0xe0133fe4adf8e952,
+      0xe3d8f9e563a198e5, 0x58180fddd97723a6,
+      0x8e679c2f5e44ff8f, 0x570f09eaa7ea7648,
+  };
+};
+
+#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
+
+template <class unused>
+constexpr uint64_t
+    powers_template<unused>::power_of_five_128[number_of_entries];
+
+#endif
+
+using powers = powers_template<>;
+
+} // namespace fast_float
+
+#endif
+
+#ifndef FASTFLOAT_DECIMAL_TO_BINARY_H
+#define FASTFLOAT_DECIMAL_TO_BINARY_H
+
+#include <cfloat>
+#include <cinttypes>
+#include <cmath>
+#include <cstdint>
+#include <cstdlib>
+#include <cstring>
+
+namespace fast_float {
+
+// This will compute or rather approximate w * 5**q and return a pair of 64-bit
+// words approximating the result, with the "high" part corresponding to the
+// most significant bits and the low part corresponding to the least significant
+// bits.
+//
+template <int bit_precision>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 value128
+compute_product_approximation(int64_t q, uint64_t w) {
+  const int index = 2 * int(q - powers::smallest_power_of_five);
+  // For small values of q, e.g., q in [0,27], the answer is always exact
+  // because The line value128 firstproduct = full_multiplication(w,
+  // power_of_five_128[index]); gives the exact answer.
+  value128 firstproduct =
+      full_multiplication(w, powers::power_of_five_128[index]);
+  static_assert((bit_precision >= 0) && (bit_precision <= 64),
+                " precision should  be in (0,64]");
+  constexpr uint64_t precision_mask =
+      (bit_precision < 64) ? (uint64_t(0xFFFFFFFFFFFFFFFF) >> bit_precision)
+                           : uint64_t(0xFFFFFFFFFFFFFFFF);
+  if ((firstproduct.high & precision_mask) ==
+      precision_mask) { // could further guard with  (lower + w < lower)
+    // regarding the second product, we only need secondproduct.high, but our
+    // expectation is that the compiler will optimize this extra work away if
+    // needed.
+    value128 secondproduct =
+        full_multiplication(w, powers::power_of_five_128[index + 1]);
+    firstproduct.low += secondproduct.high;
+    if (secondproduct.high > firstproduct.low) {
+      firstproduct.high++;
+    }
+  }
+  return firstproduct;
+}
+
+namespace detail {
+/**
+ * For q in (0,350), we have that
+ *  f = (((152170 + 65536) * q ) >> 16);
+ * is equal to
+ *   floor(p) + q
+ * where
+ *   p = log(5**q)/log(2) = q * log(5)/log(2)
+ *
+ * For negative values of q in (-400,0), we have that
+ *  f = (((152170 + 65536) * q ) >> 16);
+ * is equal to
+ *   -ceil(p) + q
+ * where
+ *   p = log(5**-q)/log(2) = -q * log(5)/log(2)
+ */
+constexpr fastfloat_really_inline int32_t power(int32_t q) noexcept {
+  return (((152170 + 65536) * q) >> 16) + 63;
+}
+} // namespace detail
+
+// create an adjusted mantissa, biased by the invalid power2
+// for significant digits already multiplied by 10 ** q.
+template <typename binary>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 adjusted_mantissa
+compute_error_scaled(int64_t q, uint64_t w, int lz) noexcept {
+  int hilz = int(w >> 63) ^ 1;
+  adjusted_mantissa answer;
+  answer.mantissa = w << hilz;
+  int bias = binary::mantissa_explicit_bits() - binary::minimum_exponent();
+  answer.power2 = int32_t(detail::power(int32_t(q)) + bias - hilz - lz - 62 +
+                          invalid_am_bias);
+  return answer;
+}
+
+// w * 10 ** q, without rounding the representation up.
+// the power2 in the exponent will be adjusted by invalid_am_bias.
+template <typename binary>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
+compute_error(int64_t q, uint64_t w) noexcept {
+  int lz = leading_zeroes(w);
+  w <<= lz;
+  value128 product =
+      compute_product_approximation<binary::mantissa_explicit_bits() + 3>(q, w);
+  return compute_error_scaled<binary>(q, product.high, lz);
+}
+
+// w * 10 ** q
+// The returned value should be a valid ieee64 number that simply need to be
+// packed. However, in some very rare cases, the computation will fail. In such
+// cases, we return an adjusted_mantissa with a negative power of 2: the caller
+// should recompute in such cases.
+template <typename binary>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
+compute_float(int64_t q, uint64_t w) noexcept {
+  adjusted_mantissa answer;
+  if ((w == 0) || (q < binary::smallest_power_of_ten())) {
+    answer.power2 = 0;
+    answer.mantissa = 0;
+    // result should be zero
+    return answer;
+  }
+  if (q > binary::largest_power_of_ten()) {
+    // we want to get infinity:
+    answer.power2 = binary::infinite_power();
+    answer.mantissa = 0;
+    return answer;
+  }
+  // At this point in time q is in [powers::smallest_power_of_five,
+  // powers::largest_power_of_five].
+
+  // We want the most significant bit of i to be 1. Shift if needed.
+  int lz = leading_zeroes(w);
+  w <<= lz;
+
+  // The required precision is binary::mantissa_explicit_bits() + 3 because
+  // 1. We need the implicit bit
+  // 2. We need an extra bit for rounding purposes
+  // 3. We might lose a bit due to the "upperbit" routine (result too small,
+  // requiring a shift)
+
+  value128 product =
+      compute_product_approximation<binary::mantissa_explicit_bits() + 3>(q, w);
+  // The computed 'product' is always sufficient.
+  // Mathematical proof:
+  // Noble Mushtak and Daniel Lemire, Fast Number Parsing Without Fallback (to
+  // appear) See script/mushtak_lemire.py
+
+  // The "compute_product_approximation" function can be slightly slower than a
+  // branchless approach: value128 product = compute_product(q, w); but in
+  // practice, we can win big with the compute_product_approximation if its
+  // additional branch is easily predicted. Which is best is data specific.
+  int upperbit = int(product.high >> 63);
+  int shift = upperbit + 64 - binary::mantissa_explicit_bits() - 3;
+
+  answer.mantissa = product.high >> shift;
+
+  answer.power2 = int32_t(detail::power(int32_t(q)) + upperbit - lz -
+                          binary::minimum_exponent());
+  if (answer.power2 <= 0) { // we have a subnormal?
+    // Here have that answer.power2 <= 0 so -answer.power2 >= 0
+    if (-answer.power2 + 1 >=
+        64) { // if we have more than 64 bits below the minimum exponent, you
+              // have a zero for sure.
+      answer.power2 = 0;
+      answer.mantissa = 0;
+      // result should be zero
+      return answer;
+    }
+    // next line is safe because -answer.power2 + 1 < 64
+    answer.mantissa >>= -answer.power2 + 1;
+    // Thankfully, we can't have both "round-to-even" and subnormals because
+    // "round-to-even" only occurs for powers close to 0.
+    answer.mantissa += (answer.mantissa & 1); // round up
+    answer.mantissa >>= 1;
+    // There is a weird scenario where we don't have a subnormal but just.
+    // Suppose we start with 2.2250738585072013e-308, we end up
+    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
+    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
+    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
+    // subnormal, but we can only know this after rounding.
+    // So we only declare a subnormal if we are smaller than the threshold.
+    answer.power2 =
+        (answer.mantissa < (uint64_t(1) << binary::mantissa_explicit_bits()))
+            ? 0
+            : 1;
+    return answer;
+  }
+
+  // usually, we round *up*, but if we fall right in between and and we have an
+  // even basis, we need to round down
+  // We are only concerned with the cases where 5**q fits in single 64-bit word.
+  if ((product.low <= 1) && (q >= binary::min_exponent_round_to_even()) &&
+      (q <= binary::max_exponent_round_to_even()) &&
+      ((answer.mantissa & 3) == 1)) { // we may fall between two floats!
+    // To be in-between two floats we need that in doing
+    //   answer.mantissa = product.high >> (upperbit + 64 -
+    //   binary::mantissa_explicit_bits() - 3);
+    // ... we dropped out only zeroes. But if this happened, then we can go
+    // back!!!
+    if ((answer.mantissa << shift) == product.high) {
+      answer.mantissa &= ~uint64_t(1); // flip it so that we do not round up
+    }
+  }
+
+  answer.mantissa += (answer.mantissa & 1); // round up
+  answer.mantissa >>= 1;
+  if (answer.mantissa >= (uint64_t(2) << binary::mantissa_explicit_bits())) {
+    answer.mantissa = (uint64_t(1) << binary::mantissa_explicit_bits());
+    answer.power2++; // undo previous addition
+  }
+
+  answer.mantissa &= ~(uint64_t(1) << binary::mantissa_explicit_bits());
+  if (answer.power2 >= binary::infinite_power()) { // infinity
+    answer.power2 = binary::infinite_power();
+    answer.mantissa = 0;
+  }
+  return answer;
+}
+
+} // namespace fast_float
+
+#endif
+
+#ifndef FASTFLOAT_BIGINT_H
+#define FASTFLOAT_BIGINT_H
+
+#include <algorithm>
+#include <cstdint>
+#include <climits>
+#include <cstring>
+
+
+namespace fast_float {
+
+// the limb width: we want efficient multiplication of double the bits in
+// limb, or for 64-bit limbs, at least 64-bit multiplication where we can
+// extract the high and low parts efficiently. this is every 64-bit
+// architecture except for sparc, which emulates 128-bit multiplication.
+// we might have platforms where `CHAR_BIT` is not 8, so let's avoid
+// doing `8 * sizeof(limb)`.
+#if defined(FASTFLOAT_64BIT) && !defined(__sparc)
+#define FASTFLOAT_64BIT_LIMB 1
+typedef uint64_t limb;
+constexpr size_t limb_bits = 64;
+#else
+#define FASTFLOAT_32BIT_LIMB
+typedef uint32_t limb;
+constexpr size_t limb_bits = 32;
+#endif
+
+typedef span<limb> limb_span;
+
+// number of bits in a bigint. this needs to be at least the number
+// of bits required to store the largest bigint, which is
+// `log2(10**(digits + max_exp))`, or `log2(10**(767 + 342))`, or
+// ~3600 bits, so we round to 4000.
+constexpr size_t bigint_bits = 4000;
+constexpr size_t bigint_limbs = bigint_bits / limb_bits;
+
+// vector-like type that is allocated on the stack. the entire
+// buffer is pre-allocated, and only the length changes.
+template <uint16_t size> struct stackvec {
+  limb data[size];
+  // we never need more than 150 limbs
+  uint16_t length{0};
+
+  stackvec() = default;
+  stackvec(const stackvec &) = delete;
+  stackvec &operator=(const stackvec &) = delete;
+  stackvec(stackvec &&) = delete;
+  stackvec &operator=(stackvec &&other) = delete;
+
+  // create stack vector from existing limb span.
+  FASTFLOAT_CONSTEXPR20 stackvec(limb_span s) {
+    FASTFLOAT_ASSERT(try_extend(s));
+  }
+
+  FASTFLOAT_CONSTEXPR14 limb &operator[](size_t index) noexcept {
+    FASTFLOAT_DEBUG_ASSERT(index < length);
+    return data[index];
+  }
+  FASTFLOAT_CONSTEXPR14 const limb &operator[](size_t index) const noexcept {
+    FASTFLOAT_DEBUG_ASSERT(index < length);
+    return data[index];
+  }
+  // index from the end of the container
+  FASTFLOAT_CONSTEXPR14 const limb &rindex(size_t index) const noexcept {
+    FASTFLOAT_DEBUG_ASSERT(index < length);
+    size_t rindex = length - index - 1;
+    return data[rindex];
+  }
+
+  // set the length, without bounds checking.
+  FASTFLOAT_CONSTEXPR14 void set_len(size_t len) noexcept {
+    length = uint16_t(len);
+  }
+  constexpr size_t len() const noexcept { return length; }
+  constexpr bool is_empty() const noexcept { return length == 0; }
+  constexpr size_t capacity() const noexcept { return size; }
+  // append item to vector, without bounds checking
+  FASTFLOAT_CONSTEXPR14 void push_unchecked(limb value) noexcept {
+    data[length] = value;
+    length++;
+  }
+  // append item to vector, returning if item was added
+  FASTFLOAT_CONSTEXPR14 bool try_push(limb value) noexcept {
+    if (len() < capacity()) {
+      push_unchecked(value);
+      return true;
+    } else {
+      return false;
+    }
+  }
+  // add items to the vector, from a span, without bounds checking
+  FASTFLOAT_CONSTEXPR20 void extend_unchecked(limb_span s) noexcept {
+    limb *ptr = data + length;
+    std::copy_n(s.ptr, s.len(), ptr);
+    set_len(len() + s.len());
+  }
+  // try to add items to the vector, returning if items were added
+  FASTFLOAT_CONSTEXPR20 bool try_extend(limb_span s) noexcept {
+    if (len() + s.len() <= capacity()) {
+      extend_unchecked(s);
+      return true;
+    } else {
+      return false;
+    }
+  }
+  // resize the vector, without bounds checking
+  // if the new size is longer than the vector, assign value to each
+  // appended item.
+  FASTFLOAT_CONSTEXPR20
+  void resize_unchecked(size_t new_len, limb value) noexcept {
+    if (new_len > len()) {
+      size_t count = new_len - len();
+      limb *first = data + len();
+      limb *last = first + count;
+      ::std::fill(first, last, value);
+      set_len(new_len);
+    } else {
+      set_len(new_len);
+    }
+  }
+  // try to resize the vector, returning if the vector was resized.
+  FASTFLOAT_CONSTEXPR20 bool try_resize(size_t new_len, limb value) noexcept {
+    if (new_len > capacity()) {
+      return false;
+    } else {
+      resize_unchecked(new_len, value);
+      return true;
+    }
+  }
+  // check if any limbs are non-zero after the given index.
+  // this needs to be done in reverse order, since the index
+  // is relative to the most significant limbs.
+  FASTFLOAT_CONSTEXPR14 bool nonzero(size_t index) const noexcept {
+    while (index < len()) {
+      if (rindex(index) != 0) {
+        return true;
+      }
+      index++;
+    }
+    return false;
+  }
+  // normalize the big integer, so most-significant zero limbs are removed.
+  FASTFLOAT_CONSTEXPR14 void normalize() noexcept {
+    while (len() > 0 && rindex(0) == 0) {
+      length--;
+    }
+  }
+};
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t
+empty_hi64(bool &truncated) noexcept {
+  truncated = false;
+  return 0;
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
+uint64_hi64(uint64_t r0, bool &truncated) noexcept {
+  truncated = false;
+  int shl = leading_zeroes(r0);
+  return r0 << shl;
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
+uint64_hi64(uint64_t r0, uint64_t r1, bool &truncated) noexcept {
+  int shl = leading_zeroes(r0);
+  if (shl == 0) {
+    truncated = r1 != 0;
+    return r0;
+  } else {
+    int shr = 64 - shl;
+    truncated = (r1 << shl) != 0;
+    return (r0 << shl) | (r1 >> shr);
+  }
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
+uint32_hi64(uint32_t r0, bool &truncated) noexcept {
+  return uint64_hi64(r0, truncated);
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
+uint32_hi64(uint32_t r0, uint32_t r1, bool &truncated) noexcept {
+  uint64_t x0 = r0;
+  uint64_t x1 = r1;
+  return uint64_hi64((x0 << 32) | x1, truncated);
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
+uint32_hi64(uint32_t r0, uint32_t r1, uint32_t r2, bool &truncated) noexcept {
+  uint64_t x0 = r0;
+  uint64_t x1 = r1;
+  uint64_t x2 = r2;
+  return uint64_hi64(x0, (x1 << 32) | x2, truncated);
+}
+
+// add two small integers, checking for overflow.
+// we want an efficient operation. for msvc, where
+// we don't have built-in intrinsics, this is still
+// pretty fast.
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 limb
+scalar_add(limb x, limb y, bool &overflow) noexcept {
+  limb z;
+// gcc and clang
+#if defined(__has_builtin)
+#if __has_builtin(__builtin_add_overflow)
+  if (!cpp20_and_in_constexpr()) {
+    overflow = __builtin_add_overflow(x, y, &z);
+    return z;
+  }
+#endif
+#endif
+
+  // generic, this still optimizes correctly on MSVC.
+  z = x + y;
+  overflow = z < x;
+  return z;
+}
+
+// multiply two small integers, getting both the high and low bits.
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 limb
+scalar_mul(limb x, limb y, limb &carry) noexcept {
+#ifdef FASTFLOAT_64BIT_LIMB
+#if defined(__SIZEOF_INT128__)
+  // GCC and clang both define it as an extension.
+  __uint128_t z = __uint128_t(x) * __uint128_t(y) + __uint128_t(carry);
+  carry = limb(z >> limb_bits);
+  return limb(z);
+#else
+  // fallback, no native 128-bit integer multiplication with carry.
+  // on msvc, this optimizes identically, somehow.
+  value128 z = full_multiplication(x, y);
+  bool overflow;
+  z.low = scalar_add(z.low, carry, overflow);
+  z.high += uint64_t(overflow); // cannot overflow
+  carry = z.high;
+  return z.low;
+#endif
+#else
+  uint64_t z = uint64_t(x) * uint64_t(y) + uint64_t(carry);
+  carry = limb(z >> limb_bits);
+  return limb(z);
+#endif
+}
+
+// add scalar value to bigint starting from offset.
+// used in grade school multiplication
+template <uint16_t size>
+inline FASTFLOAT_CONSTEXPR20 bool small_add_from(stackvec<size> &vec, limb y,
+                                                 size_t start) noexcept {
+  size_t index = start;
+  limb carry = y;
+  bool overflow;
+  while (carry != 0 && index < vec.len()) {
+    vec[index] = scalar_add(vec[index], carry, overflow);
+    carry = limb(overflow);
+    index += 1;
+  }
+  if (carry != 0) {
+    FASTFLOAT_TRY(vec.try_push(carry));
+  }
+  return true;
+}
+
+// add scalar value to bigint.
+template <uint16_t size>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
+small_add(stackvec<size> &vec, limb y) noexcept {
+  return small_add_from(vec, y, 0);
+}
+
+// multiply bigint by scalar value.
+template <uint16_t size>
+inline FASTFLOAT_CONSTEXPR20 bool small_mul(stackvec<size> &vec,
+                                            limb y) noexcept {
+  limb carry = 0;
+  for (size_t index = 0; index < vec.len(); index++) {
+    vec[index] = scalar_mul(vec[index], y, carry);
+  }
+  if (carry != 0) {
+    FASTFLOAT_TRY(vec.try_push(carry));
+  }
+  return true;
+}
+
+// add bigint to bigint starting from index.
+// used in grade school multiplication
+template <uint16_t size>
+FASTFLOAT_CONSTEXPR20 bool large_add_from(stackvec<size> &x, limb_span y,
+                                          size_t start) noexcept {
+  // the effective x buffer is from `xstart..x.len()`, so exit early
+  // if we can't get that current range.
+  if (x.len() < start || y.len() > x.len() - start) {
+    FASTFLOAT_TRY(x.try_resize(y.len() + start, 0));
+  }
+
+  bool carry = false;
+  for (size_t index = 0; index < y.len(); index++) {
+    limb xi = x[index + start];
+    limb yi = y[index];
+    bool c1 = false;
+    bool c2 = false;
+    xi = scalar_add(xi, yi, c1);
+    if (carry) {
+      xi = scalar_add(xi, 1, c2);
+    }
+    x[index + start] = xi;
+    carry = c1 | c2;
+  }
+
+  // handle overflow
+  if (carry) {
+    FASTFLOAT_TRY(small_add_from(x, 1, y.len() + start));
+  }
+  return true;
+}
+
+// add bigint to bigint.
+template <uint16_t size>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
+large_add_from(stackvec<size> &x, limb_span y) noexcept {
+  return large_add_from(x, y, 0);
+}
+
+// grade-school multiplication algorithm
+template <uint16_t size>
+FASTFLOAT_CONSTEXPR20 bool long_mul(stackvec<size> &x, limb_span y) noexcept {
+  limb_span xs = limb_span(x.data, x.len());
+  stackvec<size> z(xs);
+  limb_span zs = limb_span(z.data, z.len());
+
+  if (y.len() != 0) {
+    limb y0 = y[0];
+    FASTFLOAT_TRY(small_mul(x, y0));
+    for (size_t index = 1; index < y.len(); index++) {
+      limb yi = y[index];
+      stackvec<size> zi;
+      if (yi != 0) {
+        // re-use the same buffer throughout
+        zi.set_len(0);
+        FASTFLOAT_TRY(zi.try_extend(zs));
+        FASTFLOAT_TRY(small_mul(zi, yi));
+        limb_span zis = limb_span(zi.data, zi.len());
+        FASTFLOAT_TRY(large_add_from(x, zis, index));
+      }
+    }
+  }
+
+  x.normalize();
+  return true;
+}
+
+// grade-school multiplication algorithm
+template <uint16_t size>
+FASTFLOAT_CONSTEXPR20 bool large_mul(stackvec<size> &x, limb_span y) noexcept {
+  if (y.len() == 1) {
+    FASTFLOAT_TRY(small_mul(x, y[0]));
+  } else {
+    FASTFLOAT_TRY(long_mul(x, y));
+  }
+  return true;
+}
+
+template <typename = void> struct pow5_tables {
+  static constexpr uint32_t large_step = 135;
+  static constexpr uint64_t small_power_of_5[] = {
+      1UL,
+      5UL,
+      25UL,
+      125UL,
+      625UL,
+      3125UL,
+      15625UL,
+      78125UL,
+      390625UL,
+      1953125UL,
+      9765625UL,
+      48828125UL,
+      244140625UL,
+      1220703125UL,
+      6103515625UL,
+      30517578125UL,
+      152587890625UL,
+      762939453125UL,
+      3814697265625UL,
+      19073486328125UL,
+      95367431640625UL,
+      476837158203125UL,
+      2384185791015625UL,
+      11920928955078125UL,
+      59604644775390625UL,
+      298023223876953125UL,
+      1490116119384765625UL,
+      7450580596923828125UL,
+  };
+#ifdef FASTFLOAT_64BIT_LIMB
+  constexpr static limb large_power_of_5[] = {
+      1414648277510068013UL, 9180637584431281687UL, 4539964771860779200UL,
+      10482974169319127550UL, 198276706040285095UL};
+#else
+  constexpr static limb large_power_of_5[] = {
+      4279965485U, 329373468U,  4020270615U, 2137533757U, 4287402176U,
+      1057042919U, 1071430142U, 2440757623U, 381945767U,  46164893U};
+#endif
+};
+
+#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
+
+template <typename T> constexpr uint32_t pow5_tables<T>::large_step;
+
+template <typename T> constexpr uint64_t pow5_tables<T>::small_power_of_5[];
+
+template <typename T> constexpr limb pow5_tables<T>::large_power_of_5[];
+
+#endif
+
+// big integer type. implements a small subset of big integer
+// arithmetic, using simple algorithms since asymptotically
+// faster algorithms are slower for a small number of limbs.
+// all operations assume the big-integer is normalized.
+struct bigint : pow5_tables<> {
+  // storage of the limbs, in little-endian order.
+  stackvec<bigint_limbs> vec;
+
+  FASTFLOAT_CONSTEXPR20 bigint() : vec() {}
+  bigint(const bigint &) = delete;
+  bigint &operator=(const bigint &) = delete;
+  bigint(bigint &&) = delete;
+  bigint &operator=(bigint &&other) = delete;
+
+  FASTFLOAT_CONSTEXPR20 bigint(uint64_t value) : vec() {
+#ifdef FASTFLOAT_64BIT_LIMB
+    vec.push_unchecked(value);
+#else
+    vec.push_unchecked(uint32_t(value));
+    vec.push_unchecked(uint32_t(value >> 32));
+#endif
+    vec.normalize();
+  }
+
+  // get the high 64 bits from the vector, and if bits were truncated.
+  // this is to get the significant digits for the float.
+  FASTFLOAT_CONSTEXPR20 uint64_t hi64(bool &truncated) const noexcept {
+#ifdef FASTFLOAT_64BIT_LIMB
+    if (vec.len() == 0) {
+      return empty_hi64(truncated);
+    } else if (vec.len() == 1) {
+      return uint64_hi64(vec.rindex(0), truncated);
+    } else {
+      uint64_t result = uint64_hi64(vec.rindex(0), vec.rindex(1), truncated);
+      truncated |= vec.nonzero(2);
+      return result;
+    }
+#else
+    if (vec.len() == 0) {
+      return empty_hi64(truncated);
+    } else if (vec.len() == 1) {
+      return uint32_hi64(vec.rindex(0), truncated);
+    } else if (vec.len() == 2) {
+      return uint32_hi64(vec.rindex(0), vec.rindex(1), truncated);
+    } else {
+      uint64_t result =
+          uint32_hi64(vec.rindex(0), vec.rindex(1), vec.rindex(2), truncated);
+      truncated |= vec.nonzero(3);
+      return result;
+    }
+#endif
+  }
+
+  // compare two big integers, returning the large value.
+  // assumes both are normalized. if the return value is
+  // negative, other is larger, if the return value is
+  // positive, this is larger, otherwise they are equal.
+  // the limbs are stored in little-endian order, so we
+  // must compare the limbs in ever order.
+  FASTFLOAT_CONSTEXPR20 int compare(const bigint &other) const noexcept {
+    if (vec.len() > other.vec.len()) {
+      return 1;
+    } else if (vec.len() < other.vec.len()) {
+      return -1;
+    } else {
+      for (size_t index = vec.len(); index > 0; index--) {
+        limb xi = vec[index - 1];
+        limb yi = other.vec[index - 1];
+        if (xi > yi) {
+          return 1;
+        } else if (xi < yi) {
+          return -1;
+        }
+      }
+      return 0;
+    }
+  }
+
+  // shift left each limb n bits, carrying over to the new limb
+  // returns true if we were able to shift all the digits.
+  FASTFLOAT_CONSTEXPR20 bool shl_bits(size_t n) noexcept {
+    // Internally, for each item, we shift left by n, and add the previous
+    // right shifted limb-bits.
+    // For example, we transform (for u8) shifted left 2, to:
+    //      b10100100 b01000010
+    //      b10 b10010001 b00001000
+    FASTFLOAT_DEBUG_ASSERT(n != 0);
+    FASTFLOAT_DEBUG_ASSERT(n < sizeof(limb) * 8);
+
+    size_t shl = n;
+    size_t shr = limb_bits - shl;
+    limb prev = 0;
+    for (size_t index = 0; index < vec.len(); index++) {
+      limb xi = vec[index];
+      vec[index] = (xi << shl) | (prev >> shr);
+      prev = xi;
+    }
+
+    limb carry = prev >> shr;
+    if (carry != 0) {
+      return vec.try_push(carry);
+    }
+    return true;
+  }
+
+  // move the limbs left by `n` limbs.
+  FASTFLOAT_CONSTEXPR20 bool shl_limbs(size_t n) noexcept {
+    FASTFLOAT_DEBUG_ASSERT(n != 0);
+    if (n + vec.len() > vec.capacity()) {
+      return false;
+    } else if (!vec.is_empty()) {
+      // move limbs
+      limb *dst = vec.data + n;
+      const limb *src = vec.data;
+      std::copy_backward(src, src + vec.len(), dst + vec.len());
+      // fill in empty limbs
+      limb *first = vec.data;
+      limb *last = first + n;
+      ::std::fill(first, last, 0);
+      vec.set_len(n + vec.len());
+      return true;
+    } else {
+      return true;
+    }
+  }
+
+  // move the limbs left by `n` bits.
+  FASTFLOAT_CONSTEXPR20 bool shl(size_t n) noexcept {
+    size_t rem = n % limb_bits;
+    size_t div = n / limb_bits;
+    if (rem != 0) {
+      FASTFLOAT_TRY(shl_bits(rem));
+    }
+    if (div != 0) {
+      FASTFLOAT_TRY(shl_limbs(div));
+    }
+    return true;
+  }
+
+  // get the number of leading zeros in the bigint.
+  FASTFLOAT_CONSTEXPR20 int ctlz() const noexcept {
+    if (vec.is_empty()) {
+      return 0;
+    } else {
+#ifdef FASTFLOAT_64BIT_LIMB
+      return leading_zeroes(vec.rindex(0));
+#else
+      // no use defining a specialized leading_zeroes for a 32-bit type.
+      uint64_t r0 = vec.rindex(0);
+      return leading_zeroes(r0 << 32);
+#endif
+    }
+  }
+
+  // get the number of bits in the bigint.
+  FASTFLOAT_CONSTEXPR20 int bit_length() const noexcept {
+    int lz = ctlz();
+    return int(limb_bits * vec.len()) - lz;
+  }
+
+  FASTFLOAT_CONSTEXPR20 bool mul(limb y) noexcept { return small_mul(vec, y); }
+
+  FASTFLOAT_CONSTEXPR20 bool add(limb y) noexcept { return small_add(vec, y); }
+
+  // multiply as if by 2 raised to a power.
+  FASTFLOAT_CONSTEXPR20 bool pow2(uint32_t exp) noexcept { return shl(exp); }
+
+  // multiply as if by 5 raised to a power.
+  FASTFLOAT_CONSTEXPR20 bool pow5(uint32_t exp) noexcept {
+    // multiply by a power of 5
+    size_t large_length = sizeof(large_power_of_5) / sizeof(limb);
+    limb_span large = limb_span(large_power_of_5, large_length);
+    while (exp >= large_step) {
+      FASTFLOAT_TRY(large_mul(vec, large));
+      exp -= large_step;
+    }
+#ifdef FASTFLOAT_64BIT_LIMB
+    uint32_t small_step = 27;
+    limb max_native = 7450580596923828125UL;
+#else
+    uint32_t small_step = 13;
+    limb max_native = 1220703125U;
+#endif
+    while (exp >= small_step) {
+      FASTFLOAT_TRY(small_mul(vec, max_native));
+      exp -= small_step;
+    }
+    if (exp != 0) {
+      // Work around clang bug https://godbolt.org/z/zedh7rrhc
+      // This is similar to https://github.com/llvm/llvm-project/issues/47746,
+      // except the workaround described there don't work here
+      FASTFLOAT_TRY(small_mul(
+          vec, limb(((void)small_power_of_5[0], small_power_of_5[exp]))));
+    }
+
+    return true;
+  }
+
+  // multiply as if by 10 raised to a power.
+  FASTFLOAT_CONSTEXPR20 bool pow10(uint32_t exp) noexcept {
+    FASTFLOAT_TRY(pow5(exp));
+    return pow2(exp);
+  }
+};
+
+} // namespace fast_float
+
+#endif
+
+#ifndef FASTFLOAT_DIGIT_COMPARISON_H
+#define FASTFLOAT_DIGIT_COMPARISON_H
+
+#include <algorithm>
+#include <cstdint>
+#include <cstring>
+#include <iterator>
+
+
+namespace fast_float {
+
+// 1e0 to 1e19
+constexpr static uint64_t powers_of_ten_uint64[] = {1UL,
+                                                    10UL,
+                                                    100UL,
+                                                    1000UL,
+                                                    10000UL,
+                                                    100000UL,
+                                                    1000000UL,
+                                                    10000000UL,
+                                                    100000000UL,
+                                                    1000000000UL,
+                                                    10000000000UL,
+                                                    100000000000UL,
+                                                    1000000000000UL,
+                                                    10000000000000UL,
+                                                    100000000000000UL,
+                                                    1000000000000000UL,
+                                                    10000000000000000UL,
+                                                    100000000000000000UL,
+                                                    1000000000000000000UL,
+                                                    10000000000000000000UL};
+
+// calculate the exponent, in scientific notation, of the number.
+// this algorithm is not even close to optimized, but it has no practical
+// effect on performance: in order to have a faster algorithm, we'd need
+// to slow down performance for faster algorithms, and this is still fast.
+template <typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 int32_t
+scientific_exponent(parsed_number_string_t<UC> &num) noexcept {
+  uint64_t mantissa = num.mantissa;
+  int32_t exponent = int32_t(num.exponent);
+  while (mantissa >= 10000) {
+    mantissa /= 10000;
+    exponent += 4;
+  }
+  while (mantissa >= 100) {
+    mantissa /= 100;
+    exponent += 2;
+  }
+  while (mantissa >= 10) {
+    mantissa /= 10;
+    exponent += 1;
+  }
+  return exponent;
+}
+
+// this converts a native floating-point number to an extended-precision float.
+template <typename T>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
+to_extended(T value) noexcept {
+  using equiv_uint = typename binary_format<T>::equiv_uint;
+  constexpr equiv_uint exponent_mask = binary_format<T>::exponent_mask();
+  constexpr equiv_uint mantissa_mask = binary_format<T>::mantissa_mask();
+  constexpr equiv_uint hidden_bit_mask = binary_format<T>::hidden_bit_mask();
+
+  adjusted_mantissa am;
+  int32_t bias = binary_format<T>::mantissa_explicit_bits() -
+                 binary_format<T>::minimum_exponent();
+  equiv_uint bits;
+#if FASTFLOAT_HAS_BIT_CAST
+  bits = std::bit_cast<equiv_uint>(value);
+#else
+  ::memcpy(&bits, &value, sizeof(T));
+#endif
+  if ((bits & exponent_mask) == 0) {
+    // denormal
+    am.power2 = 1 - bias;
+    am.mantissa = bits & mantissa_mask;
+  } else {
+    // normal
+    am.power2 = int32_t((bits & exponent_mask) >>
+                        binary_format<T>::mantissa_explicit_bits());
+    am.power2 -= bias;
+    am.mantissa = (bits & mantissa_mask) | hidden_bit_mask;
+  }
+
+  return am;
+}
+
+// get the extended precision value of the halfway point between b and b+u.
+// we are given a native float that represents b, so we need to adjust it
+// halfway between b and b+u.
+template <typename T>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
+to_extended_halfway(T value) noexcept {
+  adjusted_mantissa am = to_extended(value);
+  am.mantissa <<= 1;
+  am.mantissa += 1;
+  am.power2 -= 1;
+  return am;
+}
+
+// round an extended-precision float to the nearest machine float.
+template <typename T, typename callback>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void round(adjusted_mantissa &am,
+                                                         callback cb) noexcept {
+  int32_t mantissa_shift = 64 - binary_format<T>::mantissa_explicit_bits() - 1;
+  if (-am.power2 >= mantissa_shift) {
+    // have a denormal float
+    int32_t shift = -am.power2 + 1;
+    cb(am, std::min<int32_t>(shift, 64));
+    // check for round-up: if rounding-nearest carried us to the hidden bit.
+    am.power2 = (am.mantissa <
+                 (uint64_t(1) << binary_format<T>::mantissa_explicit_bits()))
+                    ? 0
+                    : 1;
+    return;
+  }
+
+  // have a normal float, use the default shift.
+  cb(am, mantissa_shift);
+
+  // check for carry
+  if (am.mantissa >=
+      (uint64_t(2) << binary_format<T>::mantissa_explicit_bits())) {
+    am.mantissa = (uint64_t(1) << binary_format<T>::mantissa_explicit_bits());
+    am.power2++;
+  }
+
+  // check for infinite: we could have carried to an infinite power
+  am.mantissa &= ~(uint64_t(1) << binary_format<T>::mantissa_explicit_bits());
+  if (am.power2 >= binary_format<T>::infinite_power()) {
+    am.power2 = binary_format<T>::infinite_power();
+    am.mantissa = 0;
+  }
+}
+
+template <typename callback>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void
+round_nearest_tie_even(adjusted_mantissa &am, int32_t shift,
+                       callback cb) noexcept {
+  const uint64_t mask = (shift == 64) ? UINT64_MAX : (uint64_t(1) << shift) - 1;
+  const uint64_t halfway = (shift == 0) ? 0 : uint64_t(1) << (shift - 1);
+  uint64_t truncated_bits = am.mantissa & mask;
+  bool is_above = truncated_bits > halfway;
+  bool is_halfway = truncated_bits == halfway;
+
+  // shift digits into position
+  if (shift == 64) {
+    am.mantissa = 0;
+  } else {
+    am.mantissa >>= shift;
+  }
+  am.power2 += shift;
+
+  bool is_odd = (am.mantissa & 1) == 1;
+  am.mantissa += uint64_t(cb(is_odd, is_halfway, is_above));
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void
+round_down(adjusted_mantissa &am, int32_t shift) noexcept {
+  if (shift == 64) {
+    am.mantissa = 0;
+  } else {
+    am.mantissa >>= shift;
+  }
+  am.power2 += shift;
+}
+template <typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
+skip_zeros(UC const *&first, UC const *last) noexcept {
+  uint64_t val;
+  while (!cpp20_and_in_constexpr() &&
+         std::distance(first, last) >= int_cmp_len<UC>()) {
+    ::memcpy(&val, first, sizeof(uint64_t));
+    if (val != int_cmp_zeros<UC>()) {
+      break;
+    }
+    first += int_cmp_len<UC>();
+  }
+  while (first != last) {
+    if (*first != UC('0')) {
+      break;
+    }
+    first++;
+  }
+}
+
+// determine if any non-zero digits were truncated.
+// all characters must be valid digits.
+template <typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
+is_truncated(UC const *first, UC const *last) noexcept {
+  // do 8-bit optimizations, can just compare to 8 literal 0s.
+  uint64_t val;
+  while (!cpp20_and_in_constexpr() &&
+         std::distance(first, last) >= int_cmp_len<UC>()) {
+    ::memcpy(&val, first, sizeof(uint64_t));
+    if (val != int_cmp_zeros<UC>()) {
+      return true;
+    }
+    first += int_cmp_len<UC>();
+  }
+  while (first != last) {
+    if (*first != UC('0')) {
+      return true;
+    }
+    ++first;
+  }
+  return false;
+}
+template <typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
+is_truncated(span<const UC> s) noexcept {
+  return is_truncated(s.ptr, s.ptr + s.len());
+}
+
+template <typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
+parse_eight_digits(const UC *&p, limb &value, size_t &counter,
+                   size_t &count) noexcept {
+  value = value * 100000000 + parse_eight_digits_unrolled(p);
+  p += 8;
+  counter += 8;
+  count += 8;
+}
+
+template <typename UC>
+fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void
+parse_one_digit(UC const *&p, limb &value, size_t &counter,
+                size_t &count) noexcept {
+  value = value * 10 + limb(*p - UC('0'));
+  p++;
+  counter++;
+  count++;
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
+add_native(bigint &big, limb power, limb value) noexcept {
+  big.mul(power);
+  big.add(value);
+}
+
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
+round_up_bigint(bigint &big, size_t &count) noexcept {
+  // need to round-up the digits, but need to avoid rounding
+  // ....9999 to ...10000, which could cause a false halfway point.
+  add_native(big, 10, 1);
+  count++;
+}
+
+// parse the significant digits into a big integer
+template <typename UC>
+inline FASTFLOAT_CONSTEXPR20 void
+parse_mantissa(bigint &result, parsed_number_string_t<UC> &num,
+               size_t max_digits, size_t &digits) noexcept {
+  // try to minimize the number of big integer and scalar multiplication.
+  // therefore, try to parse 8 digits at a time, and multiply by the largest
+  // scalar value (9 or 19 digits) for each step.
+  size_t counter = 0;
+  digits = 0;
+  limb value = 0;
+#ifdef FASTFLOAT_64BIT_LIMB
+  size_t step = 19;
+#else
+  size_t step = 9;
+#endif
+
+  // process all integer digits.
+  UC const *p = num.integer.ptr;
+  UC const *pend = p + num.integer.len();
+  skip_zeros(p, pend);
+  // process all digits, in increments of step per loop
+  while (p != pend) {
+    while ((std::distance(p, pend) >= 8) && (step - counter >= 8) &&
+           (max_digits - digits >= 8)) {
+      parse_eight_digits(p, value, counter, digits);
+    }
+    while (counter < step && p != pend && digits < max_digits) {
+      parse_one_digit(p, value, counter, digits);
+    }
+    if (digits == max_digits) {
+      // add the temporary value, then check if we've truncated any digits
+      add_native(result, limb(powers_of_ten_uint64[counter]), value);
+      bool truncated = is_truncated(p, pend);
+      if (num.fraction.ptr != nullptr) {
+        truncated |= is_truncated(num.fraction);
+      }
+      if (truncated) {
+        round_up_bigint(result, digits);
+      }
+      return;
+    } else {
+      add_native(result, limb(powers_of_ten_uint64[counter]), value);
+      counter = 0;
+      value = 0;
+    }
+  }
+
+  // add our fraction digits, if they're available.
+  if (num.fraction.ptr != nullptr) {
+    p = num.fraction.ptr;
+    pend = p + num.fraction.len();
+    if (digits == 0) {
+      skip_zeros(p, pend);
+    }
+    // process all digits, in increments of step per loop
+    while (p != pend) {
+      while ((std::distance(p, pend) >= 8) && (step - counter >= 8) &&
+             (max_digits - digits >= 8)) {
+        parse_eight_digits(p, value, counter, digits);
+      }
+      while (counter < step && p != pend && digits < max_digits) {
+        parse_one_digit(p, value, counter, digits);
+      }
+      if (digits == max_digits) {
+        // add the temporary value, then check if we've truncated any digits
+        add_native(result, limb(powers_of_ten_uint64[counter]), value);
+        bool truncated = is_truncated(p, pend);
+        if (truncated) {
+          round_up_bigint(result, digits);
+        }
+        return;
+      } else {
+        add_native(result, limb(powers_of_ten_uint64[counter]), value);
+        counter = 0;
+        value = 0;
+      }
+    }
+  }
+
+  if (counter != 0) {
+    add_native(result, limb(powers_of_ten_uint64[counter]), value);
+  }
+}
+
+template <typename T>
+inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
+positive_digit_comp(bigint &bigmant, int32_t exponent) noexcept {
+  FASTFLOAT_ASSERT(bigmant.pow10(uint32_t(exponent)));
+  adjusted_mantissa answer;
+  bool truncated;
+  answer.mantissa = bigmant.hi64(truncated);
+  int bias = binary_format<T>::mantissa_explicit_bits() -
+             binary_format<T>::minimum_exponent();
+  answer.power2 = bigmant.bit_length() - 64 + bias;
+
+  round<T>(answer, [truncated](adjusted_mantissa &a, int32_t shift) {
+    round_nearest_tie_even(
+        a, shift,
+        [truncated](bool is_odd, bool is_halfway, bool is_above) -> bool {
+          return is_above || (is_halfway && truncated) ||
+                 (is_odd && is_halfway);
+        });
+  });
+
+  return answer;
+}
+
+// the scaling here is quite simple: we have, for the real digits `m * 10^e`,
+// and for the theoretical digits `n * 2^f`. Since `e` is always negative,
+// to scale them identically, we do `n * 2^f * 5^-f`, so we now have `m * 2^e`.
+// we then need to scale by `2^(f- e)`, and then the two significant digits
+// are of the same magnitude.
+template <typename T>
+inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa negative_digit_comp(
+    bigint &bigmant, adjusted_mantissa am, int32_t exponent) noexcept {
+  bigint &real_digits = bigmant;
+  int32_t real_exp = exponent;
+
+  // get the value of `b`, rounded down, and get a bigint representation of b+h
+  adjusted_mantissa am_b = am;
+  // gcc7 buf: use a lambda to remove the noexcept qualifier bug with
+  // -Wnoexcept-type.
+  round<T>(am_b,
+           [](adjusted_mantissa &a, int32_t shift) { round_down(a, shift); });
+  T b;
+  to_float(false, am_b, b);
+  adjusted_mantissa theor = to_extended_halfway(b);
+  bigint theor_digits(theor.mantissa);
+  int32_t theor_exp = theor.power2;
+
+  // scale real digits and theor digits to be same power.
+  int32_t pow2_exp = theor_exp - real_exp;
+  uint32_t pow5_exp = uint32_t(-real_exp);
+  if (pow5_exp != 0) {
+    FASTFLOAT_ASSERT(theor_digits.pow5(pow5_exp));
+  }
+  if (pow2_exp > 0) {
+    FASTFLOAT_ASSERT(theor_digits.pow2(uint32_t(pow2_exp)));
+  } else if (pow2_exp < 0) {
+    FASTFLOAT_ASSERT(real_digits.pow2(uint32_t(-pow2_exp)));
+  }
+
+  // compare digits, and use it to director rounding
+  int ord = real_digits.compare(theor_digits);
+  adjusted_mantissa answer = am;
+  round<T>(answer, [ord](adjusted_mantissa &a, int32_t shift) {
+    round_nearest_tie_even(
+        a, shift, [ord](bool is_odd, bool _, bool __) -> bool {
+          (void)_;  // not needed, since we've done our comparison
+          (void)__; // not needed, since we've done our comparison
+          if (ord > 0) {
+            return true;
+          } else if (ord < 0) {
+            return false;
+          } else {
+            return is_odd;
+          }
+        });
+  });
+
+  return answer;
+}
+
+// parse the significant digits as a big integer to unambiguously round the
+// the significant digits. here, we are trying to determine how to round
+// an extended float representation close to `b+h`, halfway between `b`
+// (the float rounded-down) and `b+u`, the next positive float. this
+// algorithm is always correct, and uses one of two approaches. when
+// the exponent is positive relative to the significant digits (such as
+// 1234), we create a big-integer representation, get the high 64-bits,
+// determine if any lower bits are truncated, and use that to direct
+// rounding. in case of a negative exponent relative to the significant
+// digits (such as 1.2345), we create a theoretical representation of
+// `b` as a big-integer type, scaled to the same binary exponent as
+// the actual digits. we then compare the big integer representations
+// of both, and use that to direct rounding.
+template <typename T, typename UC>
+inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
+digit_comp(parsed_number_string_t<UC> &num, adjusted_mantissa am) noexcept {
+  // remove the invalid exponent bias
+  am.power2 -= invalid_am_bias;
+
+  int32_t sci_exp = scientific_exponent(num);
+  size_t max_digits = binary_format<T>::max_digits();
+  size_t digits = 0;
+  bigint bigmant;
+  parse_mantissa(bigmant, num, max_digits, digits);
+  // can't underflow, since digits is at most max_digits.
+  int32_t exponent = sci_exp + 1 - int32_t(digits);
+  if (exponent >= 0) {
+    return positive_digit_comp<T>(bigmant, exponent);
+  } else {
+    return negative_digit_comp<T>(bigmant, am, exponent);
+  }
+}
+
+} // namespace fast_float
+
+#endif
+
+#ifndef FASTFLOAT_PARSE_NUMBER_H
+#define FASTFLOAT_PARSE_NUMBER_H
+
+
+#include <cmath>
+#include <cstring>
+#include <limits>
+#include <system_error>
+namespace fast_float {
+
+namespace detail {
+/**
+ * Special case +inf, -inf, nan, infinity, -infinity.
+ * The case comparisons could be made much faster given that we know that the
+ * strings a null-free and fixed.
+ **/
+template <typename T, typename UC>
+from_chars_result_t<UC>
+    FASTFLOAT_CONSTEXPR14 parse_infnan(UC const *first, UC const *last,
+                                       T &value, chars_format fmt) noexcept {
+  from_chars_result_t<UC> answer{};
+  answer.ptr = first;
+  answer.ec = std::errc(); // be optimistic
+  // assume first < last, so dereference without checks;
+  bool const minusSign = (*first == UC('-'));
+  // C++17 20.19.3.(7.1) explicitly forbids '+' sign here
+  if ((*first == UC('-')) ||
+      (uint64_t(fmt & chars_format::allow_leading_plus) &&
+       (*first == UC('+')))) {
+    ++first;
+  }
+  if (last - first >= 3) {
+    if (fastfloat_strncasecmp(first, str_const_nan<UC>(), 3)) {
+      answer.ptr = (first += 3);
+      value = minusSign ? -std::numeric_limits<T>::quiet_NaN()
+                        : std::numeric_limits<T>::quiet_NaN();
+      // Check for possible nan(n-char-seq-opt), C++17 20.19.3.7,
+      // C11 7.20.1.3.3. At least MSVC produces nan(ind) and nan(snan).
+      if (first != last && *first == UC('(')) {
+        for (UC const *ptr = first + 1; ptr != last; ++ptr) {
+          if (*ptr == UC(')')) {
+            answer.ptr = ptr + 1; // valid nan(n-char-seq-opt)
+            break;
+          } else if (!((UC('a') <= *ptr && *ptr <= UC('z')) ||
+                       (UC('A') <= *ptr && *ptr <= UC('Z')) ||
+                       (UC('0') <= *ptr && *ptr <= UC('9')) || *ptr == UC('_')))
+            break; // forbidden char, not nan(n-char-seq-opt)
+        }
+      }
+      return answer;
+    }
+    if (fastfloat_strncasecmp(first, str_const_inf<UC>(), 3)) {
+      if ((last - first >= 8) &&
+          fastfloat_strncasecmp(first + 3, str_const_inf<UC>() + 3, 5)) {
+        answer.ptr = first + 8;
+      } else {
+        answer.ptr = first + 3;
+      }
+      value = minusSign ? -std::numeric_limits<T>::infinity()
+                        : std::numeric_limits<T>::infinity();
+      return answer;
+    }
+  }
+  answer.ec = std::errc::invalid_argument;
+  return answer;
+}
+
+/**
+ * Returns true if the floating-pointing rounding mode is to 'nearest'.
+ * It is the default on most system. This function is meant to be inexpensive.
+ * Credit : @mwalcott3
+ */
+fastfloat_really_inline bool rounds_to_nearest() noexcept {
+  // https://lemire.me/blog/2020/06/26/gcc-not-nearest/
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  return false;
+#endif
+  // See
+  // A fast function to check your floating-point rounding mode
+  // https://lemire.me/blog/2022/11/16/a-fast-function-to-check-your-floating-point-rounding-mode/
+  //
+  // This function is meant to be equivalent to :
+  // prior: #include <cfenv>
+  //  return fegetround() == FE_TONEAREST;
+  // However, it is expected to be much faster than the fegetround()
+  // function call.
+  //
+  // The volatile keyword prevents the compiler from computing the function
+  // at compile-time.
+  // There might be other ways to prevent compile-time optimizations (e.g.,
+  // asm). The value does not need to be std::numeric_limits<float>::min(), any
+  // small value so that 1 + x should round to 1 would do (after accounting for
+  // excess precision, as in 387 instructions).
+  static volatile float fmin = std::numeric_limits<float>::min();
+  float fmini = fmin; // we copy it so that it gets loaded at most once.
+//
+// Explanation:
+// Only when fegetround() == FE_TONEAREST do we have that
+// fmin + 1.0f == 1.0f - fmin.
+//
+// FE_UPWARD:
+//  fmin + 1.0f > 1
+//  1.0f - fmin == 1
+//
+// FE_DOWNWARD or  FE_TOWARDZERO:
+//  fmin + 1.0f == 1
+//  1.0f - fmin < 1
+//
+// Note: This may fail to be accurate if fast-math has been
+// enabled, as rounding conventions may not apply.
+#ifdef FASTFLOAT_VISUAL_STUDIO
+#pragma warning(push)
+//  todo: is there a VS warning?
+//  see
+//  https://stackoverflow.com/questions/46079446/is-there-a-warning-for-floating-point-equality-checking-in-visual-studio-2013
+#elif defined(__clang__)
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wfloat-equal"
+#elif defined(__GNUC__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+  return (fmini + 1.0f == 1.0f - fmini);
+#ifdef FASTFLOAT_VISUAL_STUDIO
+#pragma warning(pop)
+#elif defined(__clang__)
+#pragma clang diagnostic pop
+#elif defined(__GNUC__)
+#pragma GCC diagnostic pop
+#endif
+}
+
+} // namespace detail
+
+template <typename T> struct from_chars_caller {
+  template <typename UC>
+  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
+  call(UC const *first, UC const *last, T &value,
+       parse_options_t<UC> options) noexcept {
+    return from_chars_advanced(first, last, value, options);
+  }
+};
+
+#if __STDCPP_FLOAT32_T__ == 1
+template <> struct from_chars_caller<std::float32_t> {
+  template <typename UC>
+  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
+  call(UC const *first, UC const *last, std::float32_t &value,
+       parse_options_t<UC> options) noexcept {
+    // if std::float32_t is defined, and we are in C++23 mode; macro set for
+    // float32; set value to float due to equivalence between float and
+    // float32_t
+    float val;
+    auto ret = from_chars_advanced(first, last, val, options);
+    value = val;
+    return ret;
+  }
+};
+#endif
+
+#if __STDCPP_FLOAT64_T__ == 1
+template <> struct from_chars_caller<std::float64_t> {
+  template <typename UC>
+  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
+  call(UC const *first, UC const *last, std::float64_t &value,
+       parse_options_t<UC> options) noexcept {
+    // if std::float64_t is defined, and we are in C++23 mode; macro set for
+    // float64; set value as double due to equivalence between double and
+    // float64_t
+    double val;
+    auto ret = from_chars_advanced(first, last, val, options);
+    value = val;
+    return ret;
+  }
+};
+#endif
+
+template <typename T, typename UC, typename>
+FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+from_chars(UC const *first, UC const *last, T &value,
+           chars_format fmt /*= chars_format::general*/) noexcept {
+  return from_chars_caller<T>::call(first, last, value,
+                                    parse_options_t<UC>(fmt));
+}
+
+/**
+ * This function overload takes parsed_number_string_t structure that is created
+ * and populated either by from_chars_advanced function taking chars range and
+ * parsing options or other parsing custom function implemented by user.
+ */
+template <typename T, typename UC>
+FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+from_chars_advanced(parsed_number_string_t<UC> &pns, T &value) noexcept {
+
+  static_assert(is_supported_float_type<T>(),
+                "only some floating-point types are supported");
+  static_assert(is_supported_char_type<UC>(),
+                "only char, wchar_t, char16_t and char32_t are supported");
+
+  from_chars_result_t<UC> answer;
+
+  answer.ec = std::errc(); // be optimistic
+  answer.ptr = pns.lastmatch;
+  // The implementation of the Clinger's fast path is convoluted because
+  // we want round-to-nearest in all cases, irrespective of the rounding mode
+  // selected on the thread.
+  // We proceed optimistically, assuming that detail::rounds_to_nearest()
+  // returns true.
+  if (binary_format<T>::min_exponent_fast_path() <= pns.exponent &&
+      pns.exponent <= binary_format<T>::max_exponent_fast_path() &&
+      !pns.too_many_digits) {
+    // Unfortunately, the conventional Clinger's fast path is only possible
+    // when the system rounds to the nearest float.
+    //
+    // We expect the next branch to almost always be selected.
+    // We could check it first (before the previous branch), but
+    // there might be performance advantages at having the check
+    // be last.
+    if (!cpp20_and_in_constexpr() && detail::rounds_to_nearest()) {
+      // We have that fegetround() == FE_TONEAREST.
+      // Next is Clinger's fast path.
+      if (pns.mantissa <= binary_format<T>::max_mantissa_fast_path()) {
+        value = T(pns.mantissa);
+        if (pns.exponent < 0) {
+          value = value / binary_format<T>::exact_power_of_ten(-pns.exponent);
+        } else {
+          value = value * binary_format<T>::exact_power_of_ten(pns.exponent);
+        }
+        if (pns.negative) {
+          value = -value;
+        }
+        return answer;
+      }
+    } else {
+      // We do not have that fegetround() == FE_TONEAREST.
+      // Next is a modified Clinger's fast path, inspired by Jakub Jelínek's
+      // proposal
+      if (pns.exponent >= 0 &&
+          pns.mantissa <=
+              binary_format<T>::max_mantissa_fast_path(pns.exponent)) {
+#if defined(__clang__) || defined(FASTFLOAT_32BIT)
+        // Clang may map 0 to -0.0 when fegetround() == FE_DOWNWARD
+        if (pns.mantissa == 0) {
+          value = pns.negative ? T(-0.) : T(0.);
+          return answer;
+        }
+#endif
+        value = T(pns.mantissa) *
+                binary_format<T>::exact_power_of_ten(pns.exponent);
+        if (pns.negative) {
+          value = -value;
+        }
+        return answer;
+      }
+    }
+  }
+  adjusted_mantissa am =
+      compute_float<binary_format<T>>(pns.exponent, pns.mantissa);
+  if (pns.too_many_digits && am.power2 >= 0) {
+    if (am != compute_float<binary_format<T>>(pns.exponent, pns.mantissa + 1)) {
+      am = compute_error<binary_format<T>>(pns.exponent, pns.mantissa);
+    }
+  }
+  // If we called compute_float<binary_format<T>>(pns.exponent, pns.mantissa)
+  // and we have an invalid power (am.power2 < 0), then we need to go the long
+  // way around again. This is very uncommon.
+  if (am.power2 < 0) {
+    am = digit_comp<T>(pns, am);
+  }
+  to_float(pns.negative, am, value);
+  // Test for over/underflow.
+  if ((pns.mantissa != 0 && am.mantissa == 0 && am.power2 == 0) ||
+      am.power2 == binary_format<T>::infinite_power()) {
+    answer.ec = std::errc::result_out_of_range;
+  }
+  return answer;
+}
+
+template <typename T, typename UC>
+FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+from_chars_float_advanced(UC const *first, UC const *last, T &value,
+                          parse_options_t<UC> options) noexcept {
+
+  static_assert(is_supported_float_type<T>(),
+                "only some floating-point types are supported");
+  static_assert(is_supported_char_type<UC>(),
+                "only char, wchar_t, char16_t and char32_t are supported");
+
+  chars_format const fmt = detail::adjust_for_feature_macros(options.format);
+
+  from_chars_result_t<UC> answer;
+  if (uint64_t(fmt & chars_format::skip_white_space)) {
+    while ((first != last) && fast_float::is_space(*first)) {
+      first++;
+    }
+  }
+  if (first == last) {
+    answer.ec = std::errc::invalid_argument;
+    answer.ptr = first;
+    return answer;
+  }
+  parsed_number_string_t<UC> pns =
+      parse_number_string<UC>(first, last, options);
+  if (!pns.valid) {
+    if (uint64_t(fmt & chars_format::no_infnan)) {
+      answer.ec = std::errc::invalid_argument;
+      answer.ptr = first;
+      return answer;
+    } else {
+      return detail::parse_infnan(first, last, value, fmt);
+    }
+  }
+
+  // call overload that takes parsed_number_string_t directly.
+  return from_chars_advanced(pns, value);
+}
+
+template <typename T, typename UC, typename>
+FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+from_chars(UC const *first, UC const *last, T &value, int base) noexcept {
+
+  static_assert(std::is_integral<T>::value, "only integer types are supported");
+  static_assert(is_supported_char_type<UC>(),
+                "only char, wchar_t, char16_t and char32_t are supported");
+
+  parse_options_t<UC> options;
+  options.base = base;
+  return from_chars_advanced(first, last, value, options);
+}
+
+template <typename T, typename UC>
+FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+from_chars_int_advanced(UC const *first, UC const *last, T &value,
+                        parse_options_t<UC> options) noexcept {
+
+  static_assert(std::is_integral<T>::value, "only integer types are supported");
+  static_assert(is_supported_char_type<UC>(),
+                "only char, wchar_t, char16_t and char32_t are supported");
+
+  chars_format const fmt = detail::adjust_for_feature_macros(options.format);
+  int const base = options.base;
+
+  from_chars_result_t<UC> answer;
+  if (uint64_t(fmt & chars_format::skip_white_space)) {
+    while ((first != last) && fast_float::is_space(*first)) {
+      first++;
+    }
+  }
+  if (first == last || base < 2 || base > 36) {
+    answer.ec = std::errc::invalid_argument;
+    answer.ptr = first;
+    return answer;
+  }
+
+  return parse_int_string(first, last, value, options);
+}
+
+template <bool> struct from_chars_advanced_caller {
+  template <typename T, typename UC>
+  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
+  call(UC const *first, UC const *last, T &value,
+       parse_options_t<UC> options) noexcept {
+    return from_chars_float_advanced(first, last, value, options);
+  }
+};
+
+template <> struct from_chars_advanced_caller<false> {
+  template <typename T, typename UC>
+  FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
+  call(UC const *first, UC const *last, T &value,
+       parse_options_t<UC> options) noexcept {
+    return from_chars_int_advanced(first, last, value, options);
+  }
+};
+
+template <typename T, typename UC>
+FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
+from_chars_advanced(UC const *first, UC const *last, T &value,
+                    parse_options_t<UC> options) noexcept {
+  return from_chars_advanced_caller<is_supported_float_type<T>()>::call(
+      first, last, value, options);
+}
+
+} // namespace fast_float
+
+#endif
+

blitz.mod/hash/xxh3.h

@@ -1,2320 +0,0 @@
-/*
- * xxHash - Extremely Fast Hash algorithm
- * Development source file for `xxh3`
- * Copyright (C) 2019-2020 Yann Collet
- *
- * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are
- * met:
- *
- *    * Redistributions of source code must retain the above copyright
- *      notice, this list of conditions and the following disclaimer.
- *    * Redistributions in binary form must reproduce the above
- *      copyright notice, this list of conditions and the following disclaimer
- *      in the documentation and/or other materials provided with the
- *      distribution.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
- * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
- * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
- * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
- * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
- * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
- * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
- * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
- * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
- * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- * You can contact the author at:
- *   - xxHash homepage: https://www.xxhash.com
- *   - xxHash source repository: https://github.com/Cyan4973/xxHash
- */
-
-/*
- * Note: This file is separated for development purposes.
- * It will be integrated into `xxhash.h` when development stage is completed.
- *
- * Credit: most of the work on vectorial and asm variants comes from @easyaspi314
- */
-
-#ifndef XXH3_H_1397135465
-#define XXH3_H_1397135465
-
-/* ===   Dependencies   === */
-#ifndef XXHASH_H_5627135585666179
-/* special: when including `xxh3.h` directly, turn on XXH_INLINE_ALL */
-#  undef XXH_INLINE_ALL   /* avoid redefinition */
-#  define XXH_INLINE_ALL
-#endif
-#include "xxhash.h"
-
-
-/* ===   Compiler specifics   === */
-
-#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* >= C99 */
-#  define XXH_RESTRICT   restrict
-#else
-/* Note: it might be useful to define __restrict or __restrict__ for some C++ compilers */
-#  define XXH_RESTRICT   /* disable */
-#endif
-
-#if (defined(__GNUC__) && (__GNUC__ >= 3))  \
-  || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
-  || defined(__clang__)
-#    define XXH_likely(x) __builtin_expect(x, 1)
-#    define XXH_unlikely(x) __builtin_expect(x, 0)
-#else
-#    define XXH_likely(x) (x)
-#    define XXH_unlikely(x) (x)
-#endif
-
-#if defined(__GNUC__)
-#  if defined(__AVX2__)
-#    include <immintrin.h>
-#  elif defined(__SSE2__)
-#    include <emmintrin.h>
-#  elif defined(__ARM_NEON__) || defined(__ARM_NEON)
-#    define inline __inline__  /* clang bug */
-#    include <arm_neon.h>
-#    undef inline
-#  endif
-#elif defined(_MSC_VER)
-#  include <intrin.h>
-#endif
-
-/*
- * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
- * remaining a true 64-bit/128-bit hash function.
- *
- * This is done by prioritizing a subset of 64-bit operations that can be
- * emulated without too many steps on the average 32-bit machine.
- *
- * For example, these two lines seem similar, and run equally fast on 64-bit:
- *
- *   xxh_u64 x;
- *   x ^= (x >> 47); // good
- *   x ^= (x >> 13); // bad
- *
- * However, to a 32-bit machine, there is a major difference.
- *
- * x ^= (x >> 47) looks like this:
- *
- *   x.lo ^= (x.hi >> (47 - 32));
- *
- * while x ^= (x >> 13) looks like this:
- *
- *   // note: funnel shifts are not usually cheap.
- *   x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
- *   x.hi ^= (x.hi >> 13);
- *
- * The first one is significantly faster than the second, simply because the
- * shift is larger than 32. This means:
- *  - All the bits we need are in the upper 32 bits, so we can ignore the lower
- *    32 bits in the shift.
- *  - The shift result will always fit in the lower 32 bits, and therefore,
- *    we can ignore the upper 32 bits in the xor.
- *
- * Thanks to this optimization, XXH3 only requires these features to be efficient:
- *
- *  - Usable unaligned access
- *  - A 32-bit or 64-bit ALU
- *      - If 32-bit, a decent ADC instruction
- *  - A 32 or 64-bit multiply with a 64-bit result
- *  - For the 128-bit variant, a decent byteswap helps short inputs.
- *
- * The first two are already required by XXH32, and almost all 32-bit and 64-bit
- * platforms which can run XXH32 can run XXH3 efficiently.
- *
- * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
- * notable exception.
- *
- * First of all, Thumb-1 lacks support for the UMULL instruction which
- * performs the important long multiply. This means numerous __aeabi_lmul
- * calls.
- *
- * Second of all, the 8 functional registers are just not enough.
- * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
- * Lo registers, and this shuffling results in thousands more MOVs than A32.
- *
- * A32 and T32 don't have this limitation. They can access all 14 registers,
- * do a 32->64 multiply with UMULL, and the flexible operand allowing free
- * shifts is helpful, too.
- *
- * Therefore, we do a quick sanity check.
- *
- * If compiling Thumb-1 for a target which supports ARM instructions, we will
- * emit a warning, as it is not a "sane" platform to compile for.
- *
- * Usually, if this happens, it is because of an accident and you probably need
- * to specify -march, as you likely meant to compile for a newer architecture.
- */
-#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
-#   warning "XXH3 is highly inefficient without ARM or Thumb-2."
-#endif
-
-/* ==========================================
- * Vectorization detection
- * ========================================== */
-#define XXH_SCALAR 0 /* Portable scalar version */
-#define XXH_SSE2   1 /* SSE2 for Pentium 4 and all x86_64 */
-#define XXH_AVX2   2 /* AVX2 for Haswell and Bulldozer */
-#define XXH_NEON   3 /* NEON for most ARMv7-A and all AArch64 */
-#define XXH_VSX    4 /* VSX and ZVector for POWER8/z13 */
-
-#ifndef XXH_VECTOR    /* can be defined on command line */
-#  if defined(__AVX2__)
-#    define XXH_VECTOR XXH_AVX2
-#  elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
-#    define XXH_VECTOR XXH_SSE2
-#  elif defined(__GNUC__) /* msvc support maybe later */ \
-  && (defined(__ARM_NEON__) || defined(__ARM_NEON)) \
-  && (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \
-    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__))
-#    define XXH_VECTOR XXH_NEON
-#  elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
-     || (defined(__s390x__) && defined(__VEC__)) \
-     && defined(__GNUC__) /* TODO: IBM XL */
-#    define XXH_VECTOR XXH_VSX
-#  else
-#    define XXH_VECTOR XXH_SCALAR
-#  endif
-#endif
-
-/*
- * Controls the alignment of the accumulator.
- * This is for compatibility with aligned vector loads, which are usually faster.
- */
-#ifndef XXH_ACC_ALIGN
-#  if XXH_VECTOR == XXH_SCALAR  /* scalar */
-#     define XXH_ACC_ALIGN 8
-#  elif XXH_VECTOR == XXH_SSE2  /* sse2 */
-#     define XXH_ACC_ALIGN 16
-#  elif XXH_VECTOR == XXH_AVX2  /* avx2 */
-#     define XXH_ACC_ALIGN 32
-#  elif XXH_VECTOR == XXH_NEON  /* neon */
-#     define XXH_ACC_ALIGN 16
-#  elif XXH_VECTOR == XXH_VSX   /* vsx */
-#     define XXH_ACC_ALIGN 16
-#  endif
-#endif
-
-/*
- * UGLY HACK:
- * GCC usually generates the best code with -O3 for xxHash.
- *
- * However, when targeting AVX2, it is overzealous in its unrolling resulting
- * in code roughly 3/4 the speed of Clang.
- *
- * There are other issues, such as GCC splitting _mm256_loadu_si256 into
- * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
- * only applies to Sandy and Ivy Bridge... which don't even support AVX2.
- *
- * That is why when compiling the AVX2 version, it is recommended to use either
- *   -O2 -mavx2 -march=haswell
- * or
- *   -O2 -mavx2 -mno-avx256-split-unaligned-load
- * for decent performance, or to use Clang instead.
- *
- * Fortunately, we can control the first one with a pragma that forces GCC into
- * -O2, but the other one we can't control without "failed to inline always
- * inline function due to target mismatch" warnings.
- */
-#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
-  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
-  && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
-#  pragma GCC push_options
-#  pragma GCC optimize("-O2")
-#endif
-
-
-#if XXH_VECTOR == XXH_NEON
-/*
- * NEON's setup for vmlal_u32 is a little more complicated than it is on
- * SSE2, AVX2, and VSX.
- *
- * While PMULUDQ and VMULEUW both perform a mask, VMLAL.U32 performs an upcast.
- *
- * To do the same operation, the 128-bit 'Q' register needs to be split into
- * two 64-bit 'D' registers, performing this operation::
- *
- *   [                a                 |                 b                ]
- *            |              '---------. .--------'                |
- *            |                         x                          |
- *            |              .---------' '--------.                |
- *   [ a & 0xFFFFFFFF | b & 0xFFFFFFFF ],[    a >> 32     |     b >> 32    ]
- *
- * Due to significant changes in aarch64, the fastest method for aarch64 is
- * completely different than the fastest method for ARMv7-A.
- *
- * ARMv7-A treats D registers as unions overlaying Q registers, so modifying
- * D11 will modify the high half of Q5. This is similar to how modifying AH
- * will only affect bits 8-15 of AX on x86.
- *
- * VZIP takes two registers, and puts even lanes in one register and odd lanes
- * in the other.
- *
- * On ARMv7-A, this strangely modifies both parameters in place instead of
- * taking the usual 3-operand form.
- *
- * Therefore, if we want to do this, we can simply use a D-form VZIP.32 on the
- * lower and upper halves of the Q register to end up with the high and low
- * halves where we want - all in one instruction.
- *
- *   vzip.32   d10, d11       @ d10 = { d10[0], d11[0] }; d11 = { d10[1], d11[1] }
- *
- * Unfortunately we need inline assembly for this: Instructions modifying two
- * registers at once is not possible in GCC or Clang's IR, and they have to
- * create a copy.
- *
- * aarch64 requires a different approach.
- *
- * In order to make it easier to write a decent compiler for aarch64, many
- * quirks were removed, such as conditional execution.
- *
- * NEON was also affected by this.
- *
- * aarch64 cannot access the high bits of a Q-form register, and writes to a
- * D-form register zero the high bits, similar to how writes to W-form scalar
- * registers (or DWORD registers on x86_64) work.
- *
- * The formerly free vget_high intrinsics now require a vext (with a few
- * exceptions)
- *
- * Additionally, VZIP was replaced by ZIP1 and ZIP2, which are the equivalent
- * of PUNPCKL* and PUNPCKH* in SSE, respectively, in order to only modify one
- * operand.
- *
- * The equivalent of the VZIP.32 on the lower and upper halves would be this
- * mess:
- *
- *   ext     v2.4s, v0.4s, v0.4s, #2 // v2 = { v0[2], v0[3], v0[0], v0[1] }
- *   zip1    v1.2s, v0.2s, v2.2s     // v1 = { v0[0], v2[0] }
- *   zip2    v0.2s, v0.2s, v1.2s     // v0 = { v0[1], v2[1] }
- *
- * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 (SHRN):
- *
- *   shrn    v1.2s, v0.2d, #32  // v1 = (uint32x2_t)(v0 >> 32);
- *   xtn     v0.2s, v0.2d       // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF);
- *
- * This is available on ARMv7-A, but is less efficient than a single VZIP.32.
- */
-
-/*
- * Function-like macro:
- * void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t &outHi)
- * {
- *     outLo = (uint32x2_t)(in & 0xFFFFFFFF);
- *     outHi = (uint32x2_t)(in >> 32);
- *     in = UNDEFINED;
- * }
- */
-# if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \
-   && defined(__GNUC__) \
-   && !defined(__aarch64__) && !defined(__arm64__)
-#  define XXH_SPLIT_IN_PLACE(in, outLo, outHi)                                              \
-    do {                                                                                    \
-      /* Undocumented GCC/Clang operand modifier: %e0 = lower D half, %f0 = upper D half */ \
-      /* https://github.com/gcc-mirror/gcc/blob/38cf91e5/gcc/config/arm/arm.c#L22486 */     \
-      /* https://github.com/llvm-mirror/llvm/blob/2c4ca683/lib/Target/ARM/ARMAsmPrinter.cpp#L399 */ \
-      __asm__("vzip.32  %e0, %f0" : "+w" (in));                                             \
-      (outLo) = vget_low_u32 (vreinterpretq_u32_u64(in));                                   \
-      (outHi) = vget_high_u32(vreinterpretq_u32_u64(in));                                   \
-   } while (0)
-# else
-#  define XXH_SPLIT_IN_PLACE(in, outLo, outHi)                                            \
-    do {                                                                                  \
-      (outLo) = vmovn_u64    (in);                                                        \
-      (outHi) = vshrn_n_u64  ((in), 32);                                                  \
-    } while (0)
-# endif
-#endif  /* XXH_VECTOR == XXH_NEON */
-
-/*
- * VSX and Z Vector helpers.
- *
- * This is very messy, and any pull requests to clean this up are welcome.
- *
- * There are a lot of problems with supporting VSX and s390x, due to
- * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
- */
-#if XXH_VECTOR == XXH_VSX
-#  if defined(__s390x__)
-#    include <s390intrin.h>
-#  else
-#    include <altivec.h>
-#  endif
-
-#  undef vector /* Undo the pollution */
-
-typedef __vector unsigned long long xxh_u64x2;
-typedef __vector unsigned char xxh_u8x16;
-typedef __vector unsigned xxh_u32x4;
-
-# ifndef XXH_VSX_BE
-#  if defined(__BIG_ENDIAN__) \
-  || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
-#    define XXH_VSX_BE 1
-#  elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
-#    warning "-maltivec=be is not recommended. Please use native endianness."
-#    define XXH_VSX_BE 1
-#  else
-#    define XXH_VSX_BE 0
-#  endif
-# endif /* !defined(XXH_VSX_BE) */
-
-# if XXH_VSX_BE
-/* A wrapper for POWER9's vec_revb. */
-#  if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
-#    define XXH_vec_revb vec_revb
-#  else
-XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
-{
-    xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
-                                  0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
-    return vec_perm(val, val, vByteSwap);
-}
-#  endif
-# endif /* XXH_VSX_BE */
-
-/*
- * Performs an unaligned load and byte swaps it on big endian.
- */
-XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
-{
-    xxh_u64x2 ret;
-    memcpy(&ret, ptr, sizeof(xxh_u64x2));
-# if XXH_VSX_BE
-    ret = XXH_vec_revb(ret);
-# endif
-    return ret;
-}
-
-/*
- * vec_mulo and vec_mule are very problematic intrinsics on PowerPC
- *
- * These intrinsics weren't added until GCC 8, despite existing for a while,
- * and they are endian dependent. Also, their meaning swap depending on version.
- * */
-# if defined(__s390x__)
- /* s390x is always big endian, no issue on this platform */
-#  define XXH_vec_mulo vec_mulo
-#  define XXH_vec_mule vec_mule
-# elif defined(__clang__) && __has_builtin(__builtin_altivec_vmuleuw)
-/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
-#  define XXH_vec_mulo __builtin_altivec_vmulouw
-#  define XXH_vec_mule __builtin_altivec_vmuleuw
-# else
-/* gcc needs inline assembly */
-/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
-XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
-{
-    xxh_u64x2 result;
-    __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
-    return result;
-}
-XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
-{
-    xxh_u64x2 result;
-    __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
-    return result;
-}
-# endif /* XXH_vec_mulo, XXH_vec_mule */
-#endif /* XXH_VECTOR == XXH_VSX */
-
-
-/* prefetch
- * can be disabled, by declaring XXH_NO_PREFETCH build macro */
-#if defined(XXH_NO_PREFETCH)
-#  define XXH_PREFETCH(ptr)  (void)(ptr)  /* disabled */
-#else
-#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86))  /* _mm_prefetch() is not defined outside of x86/x64 */
-#    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
-#    define XXH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
-#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
-#    define XXH_PREFETCH(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
-#  else
-#    define XXH_PREFETCH(ptr) (void)(ptr)  /* disabled */
-#  endif
-#endif  /* XXH_NO_PREFETCH */
-
-
-/* ==========================================
- * XXH3 default settings
- * ========================================== */
-
-#define XXH_SECRET_DEFAULT_SIZE 192   /* minimum XXH3_SECRET_SIZE_MIN */
-
-#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
-#  error "default keyset is not large enough"
-#endif
-
-/* Pseudorandom secret taken directly from FARSH */
-XXH_ALIGN(64) static const xxh_u8 kSecret[XXH_SECRET_DEFAULT_SIZE] = {
-    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
-    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
-    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
-    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
-    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
-    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
-    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
-    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
-
-    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
-    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
-    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
-    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
-};
-
-/*
- * Calculates a 32-bit to 64-bit long multiply.
- *
- * Wraps __emulu on MSVC x86 because it tends to call __allmul when it doesn't
- * need to (but it shouldn't need to anyways, it is about 7 instructions to do
- * a 64x64 multiply...). Since we know that this will _always_ emit MULL, we
- * use that instead of the normal method.
- *
- * If you are compiling for platforms like Thumb-1 and don't have a better option,
- * you may also want to write your own long multiply routine here.
- *
- * XXH_FORCE_INLINE xxh_u64 XXH_mult32to64(xxh_u64 x, xxh_u64 y)
- * {
- *    return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
- * }
- */
-#if defined(_MSC_VER) && defined(_M_IX86)
-#    include <intrin.h>
-#    define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
-#else
-/*
- * Downcast + upcast is usually better than masking on older compilers like
- * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
- *
- * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
- * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
- */
-#    define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
-#endif
-
-/*
- * Calculates a 64->128-bit long multiply.
- *
- * Uses __uint128_t and _umul128 if available, otherwise uses a scalar version.
- */
-static XXH128_hash_t
-XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
-{
-    /*
-     * GCC/Clang __uint128_t method.
-     *
-     * On most 64-bit targets, GCC and Clang define a __uint128_t type.
-     * This is usually the best way as it usually uses a native long 64-bit
-     * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
-     *
-     * Usually.
-     *
-     * Despite being a 32-bit platform, Clang (and emscripten) define this type
-     * despite not having the arithmetic for it. This results in a laggy
-     * compiler builtin call which calculates a full 128-bit multiply.
-     * In that case it is best to use the portable one.
-     * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
-     */
-#if defined(__GNUC__) && !defined(__wasm__) \
-    && defined(__SIZEOF_INT128__) \
-    || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
-
-    __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
-    XXH128_hash_t r128;
-    r128.low64  = (xxh_u64)(product);
-    r128.high64 = (xxh_u64)(product >> 64);
-    return r128;
-
-    /*
-     * MSVC for x64's _umul128 method.
-     *
-     * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
-     *
-     * This compiles to single operand MUL on x64.
-     */
-#elif defined(_M_X64) || defined(_M_IA64)
-
-#ifndef _MSC_VER
-#   pragma intrinsic(_umul128)
-#endif
-    xxh_u64 product_high;
-    xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
-    XXH128_hash_t r128;
-    r128.low64  = product_low;
-    r128.high64 = product_high;
-    return r128;
-
-#else
-    /*
-     * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
-     *
-     * This is a fast and simple grade school multiply, which is shown below
-     * with base 10 arithmetic instead of base 0x100000000.
-     *
-     *           9 3 // D2 lhs = 93
-     *         x 7 5 // D2 rhs = 75
-     *     ----------
-     *           1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
-     *         4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
-     *         2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
-     *     + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
-     *     ---------
-     *         2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
-     *     + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
-     *     ---------
-     *       6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
-     *
-     * The reasons for adding the products like this are:
-     *  1. It avoids manual carry tracking. Just like how
-     *     (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
-     *     This avoids a lot of complexity.
-     *
-     *  2. It hints for, and on Clang, compiles to, the powerful UMAAL
-     *     instruction available in ARM's Digital Signal Processing extension
-     *     in 32-bit ARMv6 and later, which is shown below:
-     *
-     *         void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
-     *         {
-     *             xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
-     *             *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
-     *             *RdHi = (xxh_u32)(product >> 32);
-     *         }
-     *
-     *     This instruction was designed for efficient long multiplication, and
-     *     allows this to be calculated in only 4 instructions at speeds
-     *     comparable to some 64-bit ALUs.
-     *
-     *  3. It isn't terrible on other platforms. Usually this will be a couple
-     *     of 32-bit ADD/ADCs.
-     */
-
-    /* First calculate all of the cross products. */
-    xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
-    xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32,        rhs & 0xFFFFFFFF);
-    xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
-    xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32,        rhs >> 32);
-
-    /* Now add the products together. These will never overflow. */
-    xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
-    xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32)        + hi_hi;
-    xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
-
-    XXH128_hash_t r128;
-    r128.low64  = lower;
-    r128.high64 = upper;
-    return r128;
-#endif
-}
-
-/*
- * Does a 64-bit to 128-bit multiply, then XOR folds it.
- *
- * The reason for the separate function is to prevent passing too many structs
- * around by value. This will hopefully inline the multiply, but we don't force it.
- */
-static xxh_u64
-XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
-{
-    XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
-    return product.low64 ^ product.high64;
-}
-
-/* Seems to produce slightly better code on GCC for some reason. */
-XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
-{
-    XXH_ASSERT(0 <= shift && shift < 64);
-    return v64 ^ (v64 >> shift);
-}
-
-/*
- * We don't need to (or want to) mix as much as XXH64.
- *
- * Short hashes are more evenly distributed, so it isn't necessary.
- */
-static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
-{
-    h64 = XXH_xorshift64(h64, 37);
-    h64 *= 0x165667919E3779F9ULL;
-    h64 = XXH_xorshift64(h64, 32);
-    return h64;
-}
-
-
-/* ==========================================
- * Short keys
- * ==========================================
- * One of the shortcomings of XXH32 and XXH64 was that their performance was
- * sub-optimal on short lengths. It used an iterative algorithm which strongly
- * favored lengths that were a multiple of 4 or 8.
- *
- * Instead of iterating over individual inputs, we use a set of single shot
- * functions which piece together a range of lengths and operate in constant time.
- *
- * Additionally, the number of multiplies has been significantly reduced. This
- * reduces latency, especially when emulating 64-bit multiplies on 32-bit.
- *
- * Depending on the platform, this may or may not be faster than XXH32, but it
- * is almost guaranteed to be faster than XXH64.
- */
-
-/*
- * At very short lengths, there isn't enough input to fully hide secrets, or use
- * the entire secret.
- *
- * There is also only a limited amount of mixing we can do before significantly
- * impacting performance.
- *
- * Therefore, we use different sections of the secret and always mix two secret
- * samples with an XOR. This should have no effect on performance on the
- * seedless or withSeed variants because everything _should_ be constant folded
- * by modern compilers.
- *
- * The XOR mixing hides individual parts of the secret and increases entropy.
- *
- * This adds an extra layer of strength for custom secrets.
- */
-XXH_FORCE_INLINE XXH64_hash_t
-XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
-{
-    XXH_ASSERT(input != NULL);
-    XXH_ASSERT(1 <= len && len <= 3);
-    XXH_ASSERT(secret != NULL);
-    /*
-     * len = 1: combined = { input[0], 0x01, input[0], input[0] }
-     * len = 2: combined = { input[1], 0x02, input[0], input[1] }
-     * len = 3: combined = { input[2], 0x03, input[0], input[1] }
-     */
-    {   xxh_u8 const c1 = input[0];
-        xxh_u8 const c2 = input[len >> 1];
-        xxh_u8 const c3 = input[len - 1];
-        xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2  << 24)
-                               | ((xxh_u32)c3 <<  0) | ((xxh_u32)len << 8);
-        xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
-        xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
-        xxh_u64 const mixed = keyed * PRIME64_1;
-        return XXH3_avalanche(mixed);
-    }
-}
-
-XXH_FORCE_INLINE XXH64_hash_t
-XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
-{
-    XXH_ASSERT(input != NULL);
-    XXH_ASSERT(secret != NULL);
-    XXH_ASSERT(4 <= len && len < 8);
-    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
-    {   xxh_u32 const input1 = XXH_readLE32(input);
-        xxh_u32 const input2 = XXH_readLE32(input + len - 4);
-        xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
-        xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
-        xxh_u64 x = input64 ^ bitflip;
-        /* this mix is inspired by Pelle Evensen's rrmxmx */
-        x ^= XXH_rotl64(x, 49) ^ XXH_rotl64(x, 24);
-        x *= 0x9FB21C651E98DF25ULL;
-        x ^= (x >> 35) + len ;
-        x *= 0x9FB21C651E98DF25ULL;
-        return XXH_xorshift64(x, 28);
-    }
-}
-
-XXH_FORCE_INLINE XXH64_hash_t
-XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
-{
-    XXH_ASSERT(input != NULL);
-    XXH_ASSERT(secret != NULL);
-    XXH_ASSERT(8 <= len && len <= 16);
-    {   xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
-        xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
-        xxh_u64 const input_lo = XXH_readLE64(input)           ^ bitflip1;
-        xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
-        xxh_u64 const acc = len
-                          + XXH_swap64(input_lo) + input_hi
-                          + XXH3_mul128_fold64(input_lo, input_hi);
-        return XXH3_avalanche(acc);
-    }
-}
-
-XXH_FORCE_INLINE XXH64_hash_t
-XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
-{
-    XXH_ASSERT(len <= 16);
-    {   if (XXH_likely(len >  8)) return XXH3_len_9to16_64b(input, len, secret, seed);
-        if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
-        if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
-        return XXH3_avalanche((PRIME64_1 + seed) ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
-    }
-}
-
-/*
- * DISCLAIMER: There are known *seed-dependent* multicollisions here due to
- * multiplication by zero, affecting hashes of lengths 17 to 240.
- *
- * However, they are very unlikely.
- *
- * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
- * unseeded non-cryptographic hashes, it does not attempt to defend itself
- * against specially crafted inputs, only random inputs.
- *
- * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
- * cancelling out the secret is taken an arbitrary number of times (addressed
- * in XXH3_accumulate_512), this collision is very unlikely with random inputs
- * and/or proper seeding:
- *
- * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
- * function that is only called up to 16 times per hash with up to 240 bytes of
- * input.
- *
- * This is not too bad for a non-cryptographic hash function, especially with
- * only 64 bit outputs.
- *
- * The 128-bit variant (which trades some speed for strength) is NOT affected
- * by this, although it is always a good idea to use a proper seed if you care
- * about strength.
- */
-XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
-                                     const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
-{
-#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
-  && defined(__i386__) && defined(__SSE2__)  /* x86 + SSE2 */ \
-  && !defined(XXH_ENABLE_AUTOVECTORIZE)      /* Define to disable like XXH32 hack */
-    /*
-     * UGLY HACK:
-     * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
-     * slower code.
-     *
-     * By forcing seed64 into a register, we disrupt the cost model and
-     * cause it to scalarize. See `XXH32_round()`
-     *
-     * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
-     * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
-     * GCC 9.2, despite both emitting scalar code.
-     *
-     * GCC generates much better scalar code than Clang for the rest of XXH3,
-     * which is why finding a more optimal codepath is an interest.
-     */
-    __asm__ ("" : "+r" (seed64));
-#endif
-    {   xxh_u64 const input_lo = XXH_readLE64(input);
-        xxh_u64 const input_hi = XXH_readLE64(input+8);
-        return XXH3_mul128_fold64(
-            input_lo ^ (XXH_readLE64(secret)   + seed64),
-            input_hi ^ (XXH_readLE64(secret+8) - seed64)
-        );
-    }
-}
-
-/* For mid range keys, XXH3 uses a Mum-hash variant. */
-XXH_FORCE_INLINE XXH64_hash_t
-XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
-                     const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
-                     XXH64_hash_t seed)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
-    XXH_ASSERT(16 < len && len <= 128);
-
-    {   xxh_u64 acc = len * PRIME64_1;
-        if (len > 32) {
-            if (len > 64) {
-                if (len > 96) {
-                    acc += XXH3_mix16B(input+48, secret+96, seed);
-                    acc += XXH3_mix16B(input+len-64, secret+112, seed);
-                }
-                acc += XXH3_mix16B(input+32, secret+64, seed);
-                acc += XXH3_mix16B(input+len-48, secret+80, seed);
-            }
-            acc += XXH3_mix16B(input+16, secret+32, seed);
-            acc += XXH3_mix16B(input+len-32, secret+48, seed);
-        }
-        acc += XXH3_mix16B(input+0, secret+0, seed);
-        acc += XXH3_mix16B(input+len-16, secret+16, seed);
-
-        return XXH3_avalanche(acc);
-    }
-}
-
-#define XXH3_MIDSIZE_MAX 240
-
-XXH_NO_INLINE XXH64_hash_t
-XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
-                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
-                      XXH64_hash_t seed)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
-    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
-
-    #define XXH3_MIDSIZE_STARTOFFSET 3
-    #define XXH3_MIDSIZE_LASTOFFSET  17
-
-    {   xxh_u64 acc = len * PRIME64_1;
-        int const nbRounds = (int)len / 16;
-        int i;
-        for (i=0; i<8; i++) {
-            acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
-        }
-        acc = XXH3_avalanche(acc);
-        XXH_ASSERT(nbRounds >= 8);
-#if defined(__clang__)                                /* Clang */ \
-    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
-    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
-        /*
-         * UGLY HACK:
-         * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
-         * In everywhere else, it uses scalar code.
-         *
-         * For 64->128-bit multiplies, even if the NEON was 100% optimal, it
-         * would still be slower than UMAAL (see XXH_mult64to128).
-         *
-         * Unfortunately, Clang doesn't handle the long multiplies properly and
-         * converts them to the nonexistent "vmulq_u64" intrinsic, which is then
-         * scalarized into an ugly mess of VMOV.32 instructions.
-         *
-         * This mess is difficult to avoid without turning autovectorization
-         * off completely, but they are usually relatively minor and/or not
-         * worth it to fix.
-         *
-         * This loop is the easiest to fix, as unlike XXH32, this pragma
-         * _actually works_ because it is a loop vectorization instead of an
-         * SLP vectorization.
-         */
-        #pragma clang loop vectorize(disable)
-#endif
-        for (i=8 ; i < nbRounds; i++) {
-            acc += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
-        }
-        /* last bytes */
-        acc += XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
-        return XXH3_avalanche(acc);
-    }
-}
-
-
-/* ===    Long Keys    === */
-
-#define STRIPE_LEN 64
-#define XXH_SECRET_CONSUME_RATE 8   /* nb of secret bytes consumed at each accumulation */
-#define ACC_NB (STRIPE_LEN / sizeof(xxh_u64))
-
-typedef enum { XXH3_acc_64bits, XXH3_acc_128bits } XXH3_accWidth_e;
-
-/*
- * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
- *
- * It is a hardened version of UMAC, based off of FARSH's implementation.
- *
- * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
- * implementations, and it is ridiculously fast.
- *
- * We harden it by mixing the original input to the accumulators as well as the product.
- *
- * This means that in the (relatively likely) case of a multiply by zero, the
- * original input is preserved.
- *
- * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
- * cross-pollination, as otherwise the upper and lower halves would be
- * essentially independent.
- *
- * This doesn't matter on 64-bit hashes since they all get merged together in
- * the end, so we skip the extra step.
- *
- * Both XXH3_64bits and XXH3_128bits use this subroutine.
- */
-XXH_FORCE_INLINE void
-XXH3_accumulate_512(      void* XXH_RESTRICT acc,
-                    const void* XXH_RESTRICT input,
-                    const void* XXH_RESTRICT secret,
-                    XXH3_accWidth_e accWidth)
-{
-#if (XXH_VECTOR == XXH_AVX2)
-
-    XXH_ASSERT((((size_t)acc) & 31) == 0);
-    {   XXH_ALIGN(32) __m256i* const xacc    =       (__m256i *) acc;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm256_loadu_si256 requires  a const __m256i * pointer for some reason. */
-        const         __m256i* const xinput  = (const __m256i *) input;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
-        const         __m256i* const xsecret = (const __m256i *) secret;
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) {
-            /* data_vec    = xinput[i]; */
-            __m256i const data_vec    = _mm256_loadu_si256    (xinput+i);
-            /* key_vec     = xsecret[i]; */
-            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
-            /* data_key    = data_vec ^ key_vec; */
-            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
-            /* data_key_lo = data_key >> 32; */
-            __m256i const data_key_lo = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
-            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
-            __m256i const product     = _mm256_mul_epu32     (data_key, data_key_lo);
-            if (accWidth == XXH3_acc_128bits) {
-                /* xacc[i] += swap(data_vec); */
-                __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
-                __m256i const sum       = _mm256_add_epi64(xacc[i], data_swap);
-                /* xacc[i] += product; */
-                xacc[i] = _mm256_add_epi64(product, sum);
-            } else {  /* XXH3_acc_64bits */
-                /* xacc[i] += data_vec; */
-                __m256i const sum = _mm256_add_epi64(xacc[i], data_vec);
-                /* xacc[i] += product; */
-                xacc[i] = _mm256_add_epi64(product, sum);
-            }
-    }   }
-
-#elif (XXH_VECTOR == XXH_SSE2)
-
-    /* SSE2 is just a half-scale version of the AVX2 version. */
-    XXH_ASSERT((((size_t)acc) & 15) == 0);
-    {   XXH_ALIGN(16) __m128i* const xacc    =       (__m128i *) acc;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
-        const         __m128i* const xinput  = (const __m128i *) input;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
-        const         __m128i* const xsecret = (const __m128i *) secret;
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) {
-            /* data_vec    = xinput[i]; */
-            __m128i const data_vec    = _mm_loadu_si128   (xinput+i);
-            /* key_vec     = xsecret[i]; */
-            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
-            /* data_key    = data_vec ^ key_vec; */
-            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
-            /* data_key_lo = data_key >> 32; */
-            __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
-            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
-            __m128i const product     = _mm_mul_epu32     (data_key, data_key_lo);
-            if (accWidth == XXH3_acc_128bits) {
-                /* xacc[i] += swap(data_vec); */
-                __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
-                __m128i const sum       = _mm_add_epi64(xacc[i], data_swap);
-                /* xacc[i] += product; */
-                xacc[i] = _mm_add_epi64(product, sum);
-            } else {  /* XXH3_acc_64bits */
-                /* xacc[i] += data_vec; */
-                __m128i const sum = _mm_add_epi64(xacc[i], data_vec);
-                /* xacc[i] += product; */
-                xacc[i] = _mm_add_epi64(product, sum);
-            }
-    }   }
-
-#elif (XXH_VECTOR == XXH_NEON)
-
-    XXH_ASSERT((((size_t)acc) & 15) == 0);
-    {
-        XXH_ALIGN(16) uint64x2_t* const xacc = (uint64x2_t *) acc;
-        /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
-        uint8_t const* const xinput = (const uint8_t *) input;
-        uint8_t const* const xsecret  = (const uint8_t *) secret;
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN / sizeof(uint64x2_t); i++) {
-            /* data_vec = xinput[i]; */
-            uint8x16_t data_vec    = vld1q_u8(xinput  + (i * 16));
-            /* key_vec  = xsecret[i];  */
-            uint8x16_t key_vec     = vld1q_u8(xsecret + (i * 16));
-            uint64x2_t data_key;
-            uint32x2_t data_key_lo, data_key_hi;
-            if (accWidth == XXH3_acc_64bits) {
-                /* xacc[i] += data_vec; */
-                xacc[i] = vaddq_u64 (xacc[i], vreinterpretq_u64_u8(data_vec));
-            } else {  /* XXH3_acc_128bits */
-                /* xacc[i] += swap(data_vec); */
-                uint64x2_t const data64  = vreinterpretq_u64_u8(data_vec);
-                uint64x2_t const swapped = vextq_u64(data64, data64, 1);
-                xacc[i] = vaddq_u64 (xacc[i], swapped);
-            }
-            /* data_key = data_vec ^ key_vec; */
-            data_key = vreinterpretq_u64_u8(veorq_u8(data_vec, key_vec));
-            /* data_key_lo = (uint32x2_t) (data_key & 0xFFFFFFFF);
-             * data_key_hi = (uint32x2_t) (data_key >> 32);
-             * data_key = UNDEFINED; */
-            XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi);
-            /* xacc[i] += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */
-            xacc[i] = vmlal_u32 (xacc[i], data_key_lo, data_key_hi);
-
-        }
-    }
-
-#elif (XXH_VECTOR == XXH_VSX)
-          xxh_u64x2* const xacc     =       (xxh_u64x2*) acc;    /* presumed aligned */
-    xxh_u64x2 const* const xinput   = (xxh_u64x2 const*) input;   /* no alignment restriction */
-    xxh_u64x2 const* const xsecret  = (xxh_u64x2 const*) secret;    /* no alignment restriction */
-    xxh_u64x2 const v32 = { 32, 32 };
-    size_t i;
-    for (i = 0; i < STRIPE_LEN / sizeof(xxh_u64x2); i++) {
-        /* data_vec = xinput[i]; */
-        xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + i);
-        /* key_vec = xsecret[i]; */
-        xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + i);
-        xxh_u64x2 const data_key = data_vec ^ key_vec;
-        /* shuffled = (data_key << 32) | (data_key >> 32); */
-        xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32);
-        /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */
-        xxh_u64x2 const product  = XXH_vec_mulo((xxh_u32x4)data_key, shuffled);
-        xacc[i] += product;
-
-        if (accWidth == XXH3_acc_64bits) {
-            xacc[i] += data_vec;
-        } else {  /* XXH3_acc_128bits */
-            /* swap high and low halves */
-#ifdef __s390x__
-            xxh_u64x2 const data_swapped = vec_permi(data_vec, data_vec, 2);
-#else
-            xxh_u64x2 const data_swapped = vec_xxpermdi(data_vec, data_vec, 2);
-#endif
-            xacc[i] += data_swapped;
-        }
-    }
-
-#else   /* scalar variant of Accumulator - universal */
-
-    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */
-    const xxh_u8* const xinput  = (const xxh_u8*) input;  /* no alignment restriction */
-    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
-    size_t i;
-    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
-    for (i=0; i < ACC_NB; i++) {
-        xxh_u64 const data_val = XXH_readLE64(xinput + 8*i);
-        xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + i*8);
-
-        if (accWidth == XXH3_acc_64bits) {
-            xacc[i] += data_val;
-        } else {
-            xacc[i ^ 1] += data_val; /* swap adjacent lanes */
-        }
-        xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32);
-    }
-#endif
-}
-
-/*
- * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
- *
- * Multiplication isn't perfect, as explained by Google in HighwayHash:
- *
- *  // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to
- *  // varying degrees. In descending order of goodness, bytes
- *  // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32.
- *  // As expected, the upper and lower bytes are much worse.
- *
- * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291
- *
- * Since our algorithm uses a pseudorandom secret to add some variance into the
- * mix, we don't need to (or want to) mix as often or as much as HighwayHash does.
- *
- * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid
- * extraction.
- *
- * Both XXH3_64bits and XXH3_128bits use this subroutine.
- */
-XXH_FORCE_INLINE void
-XXH3_scrambleAcc(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
-{
-#if (XXH_VECTOR == XXH_AVX2)
-
-    XXH_ASSERT((((size_t)acc) & 31) == 0);
-    {   XXH_ALIGN(32) __m256i* const xacc = (__m256i*) acc;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
-        const         __m256i* const xsecret = (const __m256i *) secret;
-        const __m256i prime32 = _mm256_set1_epi32((int)PRIME32_1);
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) {
-            /* xacc[i] ^= (xacc[i] >> 47) */
-            __m256i const acc_vec     = xacc[i];
-            __m256i const shifted     = _mm256_srli_epi64    (acc_vec, 47);
-            __m256i const data_vec    = _mm256_xor_si256     (acc_vec, shifted);
-            /* xacc[i] ^= xsecret; */
-            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
-            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
-
-            /* xacc[i] *= PRIME32_1; */
-            __m256i const data_key_hi = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
-            __m256i const prod_lo     = _mm256_mul_epu32     (data_key, prime32);
-            __m256i const prod_hi     = _mm256_mul_epu32     (data_key_hi, prime32);
-            xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
-        }
-    }
-
-#elif (XXH_VECTOR == XXH_SSE2)
-
-    XXH_ASSERT((((size_t)acc) & 15) == 0);
-    {   XXH_ALIGN(16) __m128i* const xacc = (__m128i*) acc;
-        /* Unaligned. This is mainly for pointer arithmetic, and because
-         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
-        const         __m128i* const xsecret = (const __m128i *) secret;
-        const __m128i prime32 = _mm_set1_epi32((int)PRIME32_1);
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) {
-            /* xacc[i] ^= (xacc[i] >> 47) */
-            __m128i const acc_vec     = xacc[i];
-            __m128i const shifted     = _mm_srli_epi64    (acc_vec, 47);
-            __m128i const data_vec    = _mm_xor_si128     (acc_vec, shifted);
-            /* xacc[i] ^= xsecret[i]; */
-            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
-            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
-
-            /* xacc[i] *= PRIME32_1; */
-            __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
-            __m128i const prod_lo     = _mm_mul_epu32     (data_key, prime32);
-            __m128i const prod_hi     = _mm_mul_epu32     (data_key_hi, prime32);
-            xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
-        }
-    }
-
-#elif (XXH_VECTOR == XXH_NEON)
-
-    XXH_ASSERT((((size_t)acc) & 15) == 0);
-
-    {   uint64x2_t* xacc       = (uint64x2_t*) acc;
-        uint8_t const* xsecret = (uint8_t const*) secret;
-        uint32x2_t prime       = vdup_n_u32 (PRIME32_1);
-
-        size_t i;
-        for (i=0; i < STRIPE_LEN/sizeof(uint64x2_t); i++) {
-            /* xacc[i] ^= (xacc[i] >> 47); */
-            uint64x2_t acc_vec  = xacc[i];
-            uint64x2_t shifted  = vshrq_n_u64 (acc_vec, 47);
-            uint64x2_t data_vec = veorq_u64   (acc_vec, shifted);
-
-            /* xacc[i] ^= xsecret[i]; */
-            uint8x16_t key_vec  = vld1q_u8(xsecret + (i * 16));
-            uint64x2_t data_key = veorq_u64(data_vec, vreinterpretq_u64_u8(key_vec));
-
-            /* xacc[i] *= PRIME32_1 */
-            uint32x2_t data_key_lo, data_key_hi;
-            /* data_key_lo = (uint32x2_t) (xacc[i] & 0xFFFFFFFF);
-             * data_key_hi = (uint32x2_t) (xacc[i] >> 32);
-             * xacc[i] = UNDEFINED; */
-            XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi);
-            {   /*
-                 * prod_hi = (data_key >> 32) * PRIME32_1;
-                 *
-                 * Avoid vmul_u32 + vshll_n_u32 since Clang 6 and 7 will
-                 * incorrectly "optimize" this:
-                 *   tmp     = vmul_u32(vmovn_u64(a), vmovn_u64(b));
-                 *   shifted = vshll_n_u32(tmp, 32);
-                 * to this:
-                 *   tmp     = "vmulq_u64"(a, b); // no such thing!
-                 *   shifted = vshlq_n_u64(tmp, 32);
-                 *
-                 * However, unlike SSE, Clang lacks a 64-bit multiply routine
-                 * for NEON, and it scalarizes two 64-bit multiplies instead.
-                 *
-                 * vmull_u32 has the same timing as vmul_u32, and it avoids
-                 * this bug completely.
-                 * See https://bugs.llvm.org/show_bug.cgi?id=39967
-                 */
-                uint64x2_t prod_hi = vmull_u32 (data_key_hi, prime);
-                /* xacc[i] = prod_hi << 32; */
-                xacc[i] = vshlq_n_u64(prod_hi, 32);
-                /* xacc[i] += (prod_hi & 0xFFFFFFFF) * PRIME32_1; */
-                xacc[i] = vmlal_u32(xacc[i], data_key_lo, prime);
-            }
-    }   }
-
-#elif (XXH_VECTOR == XXH_VSX)
-
-    XXH_ASSERT((((size_t)acc) & 15) == 0);
-
-    {         xxh_u64x2* const xacc    =       (xxh_u64x2*) acc;
-        const xxh_u64x2* const xsecret = (const xxh_u64x2*) secret;
-        /* constants */
-        xxh_u64x2 const v32  = { 32, 32 };
-        xxh_u64x2 const v47 = { 47, 47 };
-        xxh_u32x4 const prime = { PRIME32_1, PRIME32_1, PRIME32_1, PRIME32_1 };
-        size_t i;
-        for (i = 0; i < STRIPE_LEN / sizeof(xxh_u64x2); i++) {
-            /* xacc[i] ^= (xacc[i] >> 47); */
-            xxh_u64x2 const acc_vec  = xacc[i];
-            xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47);
-
-            /* xacc[i] ^= xsecret[i]; */
-            xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + i);
-            xxh_u64x2 const data_key = data_vec ^ key_vec;
-
-            /* xacc[i] *= PRIME32_1 */
-            /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF);  */
-            xxh_u64x2 const prod_even  = XXH_vec_mule((xxh_u32x4)data_key, prime);
-            /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32);  */
-            xxh_u64x2 const prod_odd  = XXH_vec_mulo((xxh_u32x4)data_key, prime);
-            xacc[i] = prod_odd + (prod_even << v32);
-    }   }
-
-#else   /* scalar variant of Scrambler - universal */
-
-    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc;   /* presumed aligned */
-    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
-    size_t i;
-    XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
-    for (i=0; i < ACC_NB; i++) {
-        xxh_u64 const key64 = XXH_readLE64(xsecret + 8*i);
-        xxh_u64 acc64 = xacc[i];
-        acc64 = XXH_xorshift64(acc64, 47);
-        acc64 ^= key64;
-        acc64 *= PRIME32_1;
-        xacc[i] = acc64;
-    }
-
-#endif
-}
-
-#define XXH_PREFETCH_DIST 384
-
-/*
- * XXH3_accumulate()
- * Loops over XXH3_accumulate_512().
- * Assumption: nbStripes will not overflow the secret size
- */
-XXH_FORCE_INLINE void
-XXH3_accumulate(     xxh_u64* XXH_RESTRICT acc,
-                const xxh_u8* XXH_RESTRICT input,
-                const xxh_u8* XXH_RESTRICT secret,
-                      size_t nbStripes,
-                      XXH3_accWidth_e accWidth)
-{
-    size_t n;
-    for (n = 0; n < nbStripes; n++ ) {
-        const xxh_u8* const in = input + n*STRIPE_LEN;
-        XXH_PREFETCH(in + XXH_PREFETCH_DIST);
-        XXH3_accumulate_512(acc,
-                            in,
-                            secret + n*XXH_SECRET_CONSUME_RATE,
-                            accWidth);
-    }
-}
-
-XXH_FORCE_INLINE void
-XXH3_hashLong_internal_loop( xxh_u64* XXH_RESTRICT acc,
-                      const xxh_u8* XXH_RESTRICT input, size_t len,
-                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
-                            XXH3_accWidth_e accWidth)
-{
-    size_t const nb_rounds = (secretSize - STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
-    size_t const block_len = STRIPE_LEN * nb_rounds;
-    size_t const nb_blocks = len / block_len;
-
-    size_t n;
-
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
-
-    for (n = 0; n < nb_blocks; n++) {
-        XXH3_accumulate(acc, input + n*block_len, secret, nb_rounds, accWidth);
-        XXH3_scrambleAcc(acc, secret + secretSize - STRIPE_LEN);
-    }
-
-    /* last partial block */
-    XXH_ASSERT(len > STRIPE_LEN);
-    {   size_t const nbStripes = (len - (block_len * nb_blocks)) / STRIPE_LEN;
-        XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE));
-        XXH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, accWidth);
-
-        /* last stripe */
-        if (len & (STRIPE_LEN - 1)) {
-            const xxh_u8* const p = input + len - STRIPE_LEN;
-            /* Do not align on 8, so that the secret is different from the scrambler */
-#define XXH_SECRET_LASTACC_START 7
-            XXH3_accumulate_512(acc, p, secret + secretSize - STRIPE_LEN - XXH_SECRET_LASTACC_START, accWidth);
-    }   }
-}
-
-XXH_FORCE_INLINE xxh_u64
-XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret)
-{
-    return XXH3_mul128_fold64(
-               acc[0] ^ XXH_readLE64(secret),
-               acc[1] ^ XXH_readLE64(secret+8) );
-}
-
-static XXH64_hash_t
-XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start)
-{
-    xxh_u64 result64 = start;
-    size_t i = 0;
-
-    for (i = 0; i < 4; i++) {
-        result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i);
-#if defined(__clang__)                                /* Clang */ \
-    && (defined(__arm__) || defined(__thumb__))       /* ARMv7 */ \
-    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */  \
-    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
-        /*
-         * UGLY HACK:
-         * Prevent autovectorization on Clang ARMv7-a. Exact same problem as
-         * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b.
-         * XXH3_64bits, len == 256, Snapdragon 835:
-         *   without hack: 2063.7 MB/s
-         *   with hack:    2560.7 MB/s
-         */
-        __asm__("" : "+r" (result64));
-#endif
-    }
-
-    return XXH3_avalanche(result64);
-}
-
-#define XXH3_INIT_ACC { PRIME32_3, PRIME64_1, PRIME64_2, PRIME64_3, \
-                        PRIME64_4, PRIME32_2, PRIME64_5, PRIME32_1 }
-
-XXH_FORCE_INLINE XXH64_hash_t
-XXH3_hashLong_64b_internal(const xxh_u8* XXH_RESTRICT input, size_t len,
-                           const xxh_u8* XXH_RESTRICT secret, size_t secretSize)
-{
-    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[ACC_NB] = XXH3_INIT_ACC;
-
-    XXH3_hashLong_internal_loop(acc, input, len, secret, secretSize, XXH3_acc_64bits);
-
-    /* converge into final hash */
-    XXH_STATIC_ASSERT(sizeof(acc) == 64);
-    /* do not align on 8, so that the secret is different from the accumulator */
-#define XXH_SECRET_MERGEACCS_START 11
-    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
-    return XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * PRIME64_1);
-}
-
-XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
-{
-    if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
-    memcpy(dst, &v64, sizeof(v64));
-}
-
-/* XXH3_initCustomSecret() :
- * destination `customSecret` is presumed allocated and same size as `kSecret`.
- */
-XXH_FORCE_INLINE void XXH3_initCustomSecret(xxh_u8* XXH_RESTRICT customSecret, xxh_u64 seed64)
-{
-    int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
-    int i;
-    /*
-     * We need a separate pointer for the hack below.
-     * Any decent compiler will optimize this out otherwise.
-     */
-    const xxh_u8 *kSecretPtr = kSecret;
-
-    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
-
-#if defined(__clang__) && defined(__aarch64__)
-    /*
-     * UGLY HACK:
-     * Clang generates a bunch of MOV/MOVK pairs for aarch64, and they are
-     * placed sequentially, in order, at the top of the unrolled loop.
-     *
-     * While MOVK is great for generating constants (2 cycles for a 64-bit
-     * constant compared to 4 cycles for LDR), long MOVK chains stall the
-     * integer pipelines:
-     *   I   L   S
-     * MOVK
-     * MOVK
-     * MOVK
-     * MOVK
-     * ADD
-     * SUB      STR
-     *          STR
-     * By forcing loads from memory (as the asm line causes Clang to assume
-     * that kSecretPtr has been changed), the pipelines are used more efficiently:
-     *   I   L   S
-     *      LDR
-     *  ADD LDR
-     *  SUB     STR
-     *          STR
-     * XXH3_64bits_withSeed, len == 256, Snapdragon 835
-     *   without hack: 2654.4 MB/s
-     *   with hack:    3202.9 MB/s
-     */
-    __asm__("" : "+r" (kSecretPtr));
-#endif
-    /*
-     * Note: in debug mode, this overrides the asm optimization
-     * and Clang will emit MOVK chains again.
-     */
-    XXH_ASSERT(kSecretPtr == kSecret);
-
-    for (i=0; i < nbRounds; i++) {
-        /*
-         * The asm hack causes Clang to assume that kSecretPtr aliases with
-         * customSecret, and on aarch64, this prevented LDP from merging two
-         * loads together for free. Putting the loads together before the stores
-         * properly generates LDP.
-         */
-        xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i)     + seed64;
-        xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
-        XXH_writeLE64(customSecret + 16*i,     lo);
-        XXH_writeLE64(customSecret + 16*i + 8, hi);
-    }
-}
-
-
-/*
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH64_hash_t
-XXH3_hashLong_64b_defaultSecret(const xxh_u8* XXH_RESTRICT input, size_t len)
-{
-    return XXH3_hashLong_64b_internal(input, len, kSecret, sizeof(kSecret));
-}
-
-/*
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH64_hash_t
-XXH3_hashLong_64b_withSecret(const xxh_u8* XXH_RESTRICT input, size_t len,
-                             const xxh_u8* XXH_RESTRICT secret, size_t secretSize)
-{
-    return XXH3_hashLong_64b_internal(input, len, secret, secretSize);
-}
-
-/*
- * XXH3_hashLong_64b_withSeed():
- * Generate a custom key based on alteration of default kSecret with the seed,
- * and then use this key for long mode hashing.
- *
- * This operation is decently fast but nonetheless costs a little bit of time.
- * Try to avoid it whenever possible (typically when seed==0).
- *
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH64_hash_t
-XXH3_hashLong_64b_withSeed(const xxh_u8* input, size_t len, XXH64_hash_t seed)
-{
-    XXH_ALIGN(8) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
-    if (seed==0) return XXH3_hashLong_64b_defaultSecret(input, len);
-    XXH3_initCustomSecret(secret, seed);
-    return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret));
-}
-
-/* ===   Public entry point   === */
-
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* input, size_t len)
-{
-    if (len <= 16)
-        return XXH3_len_0to16_64b((const xxh_u8*)input, len, kSecret, 0);
-    if (len <= 128)
-        return XXH3_len_17to128_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0);
-    if (len <= XXH3_MIDSIZE_MAX)
-         return XXH3_len_129to240_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0);
-    return XXH3_hashLong_64b_defaultSecret((const xxh_u8*)input, len);
-}
-
-XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
-    /*
-     * If an action is to be taken if `secret` conditions are not respected,
-     * it should be done here.
-     * For now, it's a contract pre-condition.
-     * Adding a check and a branch here would cost performance at every hash.
-     */
-    if (len <= 16)
-        return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, 0);
-    if (len <= 128)
-        return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0);
-    if (len <= XXH3_MIDSIZE_MAX)
-        return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0);
-    return XXH3_hashLong_64b_withSecret((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize);
-}
-
-XXH_PUBLIC_API XXH64_hash_t
-XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
-{
-    if (len <= 16)
-        return XXH3_len_0to16_64b((const xxh_u8*)input, len, kSecret, seed);
-    if (len <= 128)
-        return XXH3_len_17to128_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed);
-    if (len <= XXH3_MIDSIZE_MAX)
-        return XXH3_len_129to240_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed);
-    return XXH3_hashLong_64b_withSeed((const xxh_u8*)input, len, seed);
-}
-
-/* ===   XXH3 streaming   === */
-
-
-/*
- * Malloc's a pointer that is always aligned to align.
- *
- * This must be freed with `XXH_alignedFree()`.
- *
- * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte
- * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2
- * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON.
- *
- * This underalignment previously caused a rather obvious crash which went
- * completely unnoticed due to XXH3_createState() not actually being tested.
- * Credit to RedSpah for noticing this bug.
- *
- * The alignment is done manually: Functions like posix_memalign or _mm_malloc
- * are avoided: To maintain portability, we would have to write a fallback
- * like this anyways, and besides, testing for the existence of library
- * functions without relying on external build tools is impossible.
- *
- * The method is simple: Overallocate, manually align, and store the offset
- * to the original behind the returned pointer.
- *
- * Align must be a power of 2 and 8 <= align <= 128.
- */
-static void* XXH_alignedMalloc(size_t s, size_t align)
-{
-    XXH_ASSERT(align <= 128 && align >= 8); /* range check */
-    XXH_ASSERT((align & (align-1)) == 0);   /* power of 2 */
-    XXH_ASSERT(s != 0 && s < (s + align));  /* empty/overflow */
-    {   /* Overallocate to make room for manual realignment and an offset byte */
-        xxh_u8* base = (xxh_u8*)XXH_malloc(s + align);
-        if (base != NULL) {
-            /*
-             * Get the offset needed to align this pointer.
-             *
-             * Even if the returned pointer is aligned, there will always be
-             * at least one byte to store the offset to the original pointer.
-             */
-            size_t offset = align - ((size_t)base & (align - 1)); /* base % align */
-            /* Add the offset for the now-aligned pointer */
-            xxh_u8* ptr = base + offset;
-
-            XXH_ASSERT((size_t)ptr % align == 0);
-
-            /* Store the offset immediately before the returned pointer. */
-            ptr[-1] = (xxh_u8)offset;
-            return ptr;
-        }
-        return NULL;
-    }
-}
-/*
- * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass
- * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout.
- */
-static void XXH_alignedFree(void* p)
-{
-    if (p != NULL) {
-        xxh_u8* ptr = (xxh_u8*)p;
-        /* Get the offset byte we added in XXH_malloc. */
-        xxh_u8 offset = ptr[-1];
-        /* Free the original malloc'd pointer */
-        xxh_u8* base = ptr - offset;
-        XXH_free(base);
-    }
-}
-XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
-{
-    return (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
-}
-
-XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
-{
-    XXH_alignedFree(statePtr);
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API void
-XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state)
-{
-    memcpy(dst_state, src_state, sizeof(*dst_state));
-}
-
-static void
-XXH3_64bits_reset_internal(XXH3_state_t* statePtr,
-                           XXH64_hash_t seed,
-                           const xxh_u8* secret, size_t secretSize)
-{
-    XXH_ASSERT(statePtr != NULL);
-    memset(statePtr, 0, sizeof(*statePtr));
-    statePtr->acc[0] = PRIME32_3;
-    statePtr->acc[1] = PRIME64_1;
-    statePtr->acc[2] = PRIME64_2;
-    statePtr->acc[3] = PRIME64_3;
-    statePtr->acc[4] = PRIME64_4;
-    statePtr->acc[5] = PRIME32_2;
-    statePtr->acc[6] = PRIME64_5;
-    statePtr->acc[7] = PRIME32_1;
-    statePtr->seed = seed;
-    XXH_ASSERT(secret != NULL);
-    statePtr->secret = secret;
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
-    statePtr->secretLimit = (XXH32_hash_t)(secretSize - STRIPE_LEN);
-    statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset(XXH3_state_t* statePtr)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_64bits_reset_internal(statePtr, 0, kSecret, XXH_SECRET_DEFAULT_SIZE);
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_64bits_reset_internal(statePtr, 0, (const xxh_u8*)secret, secretSize);
-    if (secret == NULL) return XXH_ERROR;
-    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_64bits_reset_internal(statePtr, seed, kSecret, XXH_SECRET_DEFAULT_SIZE);
-    XXH3_initCustomSecret(statePtr->customSecret, seed);
-    statePtr->secret = statePtr->customSecret;
-    return XXH_OK;
-}
-
-XXH_FORCE_INLINE void
-XXH3_consumeStripes( xxh_u64* acc,
-                    XXH32_hash_t* nbStripesSoFarPtr, XXH32_hash_t nbStripesPerBlock,
-                    const xxh_u8* input, size_t totalStripes,
-                    const xxh_u8* secret, size_t secretLimit,
-                    XXH3_accWidth_e accWidth)
-{
-    XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock);
-    if (nbStripesPerBlock - *nbStripesSoFarPtr <= totalStripes) {
-        /* need a scrambling operation */
-        size_t const nbStripes = nbStripesPerBlock - *nbStripesSoFarPtr;
-        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes, accWidth);
-        XXH3_scrambleAcc(acc, secret + secretLimit);
-        XXH3_accumulate(acc, input + nbStripes * STRIPE_LEN, secret, totalStripes - nbStripes, accWidth);
-        *nbStripesSoFarPtr = (XXH32_hash_t)(totalStripes - nbStripes);
-    } else {
-        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, totalStripes, accWidth);
-        *nbStripesSoFarPtr += (XXH32_hash_t)totalStripes;
-    }
-}
-
-/*
- * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
- */
-XXH_FORCE_INLINE XXH_errorcode
-XXH3_update(XXH3_state_t* state, const xxh_u8* input, size_t len, XXH3_accWidth_e accWidth)
-{
-    if (input==NULL)
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-        return XXH_OK;
-#else
-        return XXH_ERROR;
-#endif
-
-    {   const xxh_u8* const bEnd = input + len;
-
-        state->totalLen += len;
-
-        if (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) {  /* fill in tmp buffer */
-            XXH_memcpy(state->buffer + state->bufferedSize, input, len);
-            state->bufferedSize += (XXH32_hash_t)len;
-            return XXH_OK;
-        }
-        /* input is now > XXH3_INTERNALBUFFER_SIZE */
-
-        #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / STRIPE_LEN)
-        XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % STRIPE_LEN == 0);   /* clean multiple */
-
-        /*
-         * There is some input left inside the internal buffer.
-         * Fill it, then consume it.
-         */
-        if (state->bufferedSize) {
-            size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize;
-            XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize);
-            input += loadSize;
-            XXH3_consumeStripes(state->acc,
-                               &state->nbStripesSoFar, state->nbStripesPerBlock,
-                                state->buffer, XXH3_INTERNALBUFFER_STRIPES,
-                                state->secret, state->secretLimit,
-                                accWidth);
-            state->bufferedSize = 0;
-        }
-
-        /* Consume input by full buffer quantities */
-        if (input+XXH3_INTERNALBUFFER_SIZE <= bEnd) {
-            const xxh_u8* const limit = bEnd - XXH3_INTERNALBUFFER_SIZE;
-            do {
-                XXH3_consumeStripes(state->acc,
-                                   &state->nbStripesSoFar, state->nbStripesPerBlock,
-                                    input, XXH3_INTERNALBUFFER_STRIPES,
-                                    state->secret, state->secretLimit,
-                                    accWidth);
-                input += XXH3_INTERNALBUFFER_SIZE;
-            } while (input<=limit);
-        }
-
-        if (input < bEnd) { /* Some remaining input: buffer it */
-            XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
-            state->bufferedSize = (XXH32_hash_t)(bEnd-input);
-        }
-    }
-
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_64bits_update(XXH3_state_t* state, const void* input, size_t len)
-{
-    return XXH3_update(state, (const xxh_u8*)input, len, XXH3_acc_64bits);
-}
-
-
-XXH_FORCE_INLINE void
-XXH3_digest_long (XXH64_hash_t* acc, const XXH3_state_t* state, XXH3_accWidth_e accWidth)
-{
-    /*
-     * Digest on a local copy. This way, the state remains unaltered, and it can
-     * continue ingesting more input afterwards.
-     */
-    memcpy(acc, state->acc, sizeof(state->acc));
-    if (state->bufferedSize >= STRIPE_LEN) {
-        size_t const totalNbStripes = state->bufferedSize / STRIPE_LEN;
-        XXH32_hash_t nbStripesSoFar = state->nbStripesSoFar;
-        XXH3_consumeStripes(acc,
-                           &nbStripesSoFar, state->nbStripesPerBlock,
-                            state->buffer, totalNbStripes,
-                            state->secret, state->secretLimit,
-                            accWidth);
-        if (state->bufferedSize % STRIPE_LEN) {  /* one last partial stripe */
-            XXH3_accumulate_512(acc,
-                                state->buffer + state->bufferedSize - STRIPE_LEN,
-                                state->secret + state->secretLimit - XXH_SECRET_LASTACC_START,
-                                accWidth);
-        }
-    } else {  /* bufferedSize < STRIPE_LEN */
-        if (state->bufferedSize) { /* one last stripe */
-            xxh_u8 lastStripe[STRIPE_LEN];
-            size_t const catchupSize = STRIPE_LEN - state->bufferedSize;
-            memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
-            memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
-            XXH3_accumulate_512(acc,
-                                lastStripe,
-                                state->secret + state->secretLimit - XXH_SECRET_LASTACC_START,
-                                accWidth);
-    }   }
-}
-
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state)
-{
-    if (state->totalLen > XXH3_MIDSIZE_MAX) {
-        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[ACC_NB];
-        XXH3_digest_long(acc, state, XXH3_acc_64bits);
-        return XXH3_mergeAccs(acc,
-                              state->secret + XXH_SECRET_MERGEACCS_START,
-                              (xxh_u64)state->totalLen * PRIME64_1);
-    }
-    /* len <= XXH3_MIDSIZE_MAX: short code */
-    if (state->seed)
-        return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
-    return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
-                                  state->secret, state->secretLimit + STRIPE_LEN);
-}
-
-/* ==========================================
- * XXH3 128 bits (a.k.a XXH128)
- * ==========================================
- * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant,
- * even without counting the significantly larger output size.
- *
- * For example, extra steps are taken to avoid the seed-dependent collisions
- * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B).
- *
- * This strength naturally comes at the cost of some speed, especially on short
- * lengths. Note that longer hashes are about as fast as the 64-bit version
- * due to it using only a slight modification of the 64-bit loop.
- *
- * XXH128 is also more oriented towards 64-bit machines. It is still extremely
- * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
- */
-
-XXH_FORCE_INLINE XXH128_hash_t
-XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
-{
-    /* A doubled version of 1to3_64b with different constants. */
-    XXH_ASSERT(input != NULL);
-    XXH_ASSERT(1 <= len && len <= 3);
-    XXH_ASSERT(secret != NULL);
-    /*
-     * len = 1: combinedl = { input[0], 0x01, input[0], input[0] }
-     * len = 2: combinedl = { input[1], 0x02, input[0], input[1] }
-     * len = 3: combinedl = { input[2], 0x03, input[0], input[1] }
-     */
-    {   xxh_u8 const c1 = input[0];
-        xxh_u8 const c2 = input[len >> 1];
-        xxh_u8 const c3 = input[len - 1];
-        xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24)
-                                | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
-        xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13);
-        xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
-        xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed;
-        xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl;
-        xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph;
-        xxh_u64 const mixedl = keyed_lo * PRIME64_1;
-        xxh_u64 const mixedh = keyed_hi * PRIME64_5;
-        XXH128_hash_t h128;
-        h128.low64  = XXH3_avalanche(mixedl);
-        h128.high64 = XXH3_avalanche(mixedh);
-        return h128;
-    }
-}
-
-XXH_FORCE_INLINE XXH128_hash_t
-XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
-{
-    XXH_ASSERT(input != NULL);
-    XXH_ASSERT(secret != NULL);
-    XXH_ASSERT(4 <= len && len <= 8);
-    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
-    {   xxh_u32 const input_lo = XXH_readLE32(input);
-        xxh_u32 const input_hi = XXH_readLE32(input + len - 4);
-        xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32);
-        xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed;
-        xxh_u64 const keyed = input_64 ^ bitflip;
-
-        /* Shift len to the left to ensure it is even, this avoids even multiplies. */
-        XXH128_hash_t m128 = XXH_mult64to128(keyed, PRIME64_1 + (len << 2));
-
-        m128.high64 += (m128.low64 << 1);
-        m128.low64  ^= (m128.high64 >> 3);
-
-        m128.low64   = XXH_xorshift64(m128.low64, 35);
-        m128.low64  *= 0x9FB21C651E98DF25ULL;
-        m128.low64   = XXH_xorshift64(m128.low64, 28);
-        m128.high64  = XXH3_avalanche(m128.high64);
-        return m128;
-    }
-}
-
-XXH_FORCE_INLINE XXH128_hash_t
-XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
-{
-    XXH_ASSERT(input != NULL);
-    XXH_ASSERT(secret != NULL);
-    XXH_ASSERT(9 <= len && len <= 16);
-    {   xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed;
-        xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed;
-        xxh_u64 const input_lo = XXH_readLE64(input);
-        xxh_u64       input_hi = XXH_readLE64(input + len - 8);
-        XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, PRIME64_1);
-        /*
-         * Put len in the middle of m128 to ensure that the length gets mixed to
-         * both the low and high bits in the 128x64 multiply below.
-         */
-        m128.low64 += (xxh_u64)(len - 1) << 54;
-        input_hi   ^= bitfliph;
-        /*
-         * Add the high 32 bits of input_hi to the high 32 bits of m128, then
-         * add the long product of the low 32 bits of input_hi and PRIME32_2 to
-         * the high 64 bits of m128.
-         *
-         * The best approach to this operation is different on 32-bit and 64-bit.
-         */
-        if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */
-            /*
-             * 32-bit optimized version, which is more readable.
-             *
-             * On 32-bit, it removes an ADC and delays a dependency between the two
-             * halves of m128.high64, but it generates an extra mask on 64-bit.
-             */
-            m128.high64 += (input_hi & 0xFFFFFFFF00000000) + XXH_mult32to64((xxh_u32)input_hi, PRIME32_2);
-        } else {
-            /*
-             * 64-bit optimized (albeit more confusing) version.
-             *
-             * Uses some properties of addition and multiplication to remove the mask:
-             *
-             * Let:
-             *    a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF)
-             *    b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000)
-             *    c = PRIME32_2
-             *
-             *    a + (b * c)
-             * Inverse Property: x + y - x == y
-             *    a + (b * (1 + c - 1))
-             * Distributive Property: x * (y + z) == (x * y) + (x * z)
-             *    a + (b * 1) + (b * (c - 1))
-             * Identity Property: x * 1 == x
-             *    a + b + (b * (c - 1))
-             *
-             * Substitute a, b, and c:
-             *    input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (PRIME32_2 - 1))
-             *
-             * Since input_hi.hi + input_hi.lo == input_hi, we get this:
-             *    input_hi + ((xxh_u64)input_hi.lo * (PRIME32_2 - 1))
-             */
-            m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, PRIME32_2 - 1);
-        }
-        /* m128 ^= XXH_swap64(m128 >> 64); */
-        m128.low64  ^= XXH_swap64(m128.high64);
-
-        {   /* 128x64 multiply: h128 = m128 * PRIME64_2; */
-            XXH128_hash_t h128 = XXH_mult64to128(m128.low64, PRIME64_2);
-            h128.high64 += m128.high64 * PRIME64_2;
-
-            h128.low64   = XXH3_avalanche(h128.low64);
-            h128.high64  = XXH3_avalanche(h128.high64);
-            return h128;
-    }   }
-}
-
-/*
- * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
- */
-XXH_FORCE_INLINE XXH128_hash_t
-XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
-{
-    XXH_ASSERT(len <= 16);
-    {   if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed);
-        if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed);
-        if (len) return XXH3_len_1to3_128b(input, len, secret, seed);
-        {   XXH128_hash_t h128;
-            xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72);
-            xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88);
-            h128.low64 = XXH3_avalanche((PRIME64_1 + seed) ^ bitflipl);
-            h128.high64 = XXH3_avalanche((PRIME64_2 - seed) ^ bitfliph);
-            return h128;
-    }   }
-}
-
-/*
- * A bit slower than XXH3_mix16B, but handles multiply by zero better.
- */
-XXH_FORCE_INLINE XXH128_hash_t
-XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
-              const xxh_u8* secret, XXH64_hash_t seed)
-{
-    acc.low64  += XXH3_mix16B (input_1, secret+0, seed);
-    acc.low64  ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8);
-    acc.high64 += XXH3_mix16B (input_2, secret+16, seed);
-    acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8);
-    return acc;
-}
-
-
-XXH_FORCE_INLINE XXH128_hash_t
-XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
-                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
-                      XXH64_hash_t seed)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
-    XXH_ASSERT(16 < len && len <= 128);
-
-    {   XXH128_hash_t acc;
-        acc.low64 = len * PRIME64_1;
-        acc.high64 = 0;
-        if (len > 32) {
-            if (len > 64) {
-                if (len > 96) {
-                    acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed);
-                }
-                acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed);
-            }
-            acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed);
-        }
-        acc = XXH128_mix32B(acc, input, input+len-16, secret, seed);
-        {   XXH128_hash_t h128;
-            h128.low64  = acc.low64 + acc.high64;
-            h128.high64 = (acc.low64    * PRIME64_1)
-                        + (acc.high64   * PRIME64_4)
-                        + ((len - seed) * PRIME64_2);
-            h128.low64  = XXH3_avalanche(h128.low64);
-            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
-            return h128;
-        }
-    }
-}
-
-XXH_NO_INLINE XXH128_hash_t
-XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
-                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
-                       XXH64_hash_t seed)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
-    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
-
-    {   XXH128_hash_t acc;
-        int const nbRounds = (int)len / 32;
-        int i;
-        acc.low64 = len * PRIME64_1;
-        acc.high64 = 0;
-        for (i=0; i<4; i++) {
-            acc = XXH128_mix32B(acc,
-                                input  + (32 * i),
-                                input  + (32 * i) + 16,
-                                secret + (32 * i),
-                                seed);
-        }
-        acc.low64 = XXH3_avalanche(acc.low64);
-        acc.high64 = XXH3_avalanche(acc.high64);
-        XXH_ASSERT(nbRounds >= 4);
-        for (i=4 ; i < nbRounds; i++) {
-            acc = XXH128_mix32B(acc,
-                                input + (32 * i),
-                                input + (32 * i) + 16,
-                                secret + XXH3_MIDSIZE_STARTOFFSET + (32 * (i - 4)),
-                                seed);
-        }
-        /* last bytes */
-        acc = XXH128_mix32B(acc,
-                            input + len - 16,
-                            input + len - 32,
-                            secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
-                            0ULL - seed);
-
-        {   XXH128_hash_t h128;
-            h128.low64  = acc.low64 + acc.high64;
-            h128.high64 = (acc.low64    * PRIME64_1)
-                        + (acc.high64   * PRIME64_4)
-                        + ((len - seed) * PRIME64_2);
-            h128.low64  = XXH3_avalanche(h128.low64);
-            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
-            return h128;
-        }
-    }
-}
-
-XXH_FORCE_INLINE XXH128_hash_t
-XXH3_hashLong_128b_internal(const xxh_u8* XXH_RESTRICT input, size_t len,
-                            const xxh_u8* XXH_RESTRICT secret, size_t secretSize)
-{
-    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[ACC_NB] = XXH3_INIT_ACC;
-
-    XXH3_hashLong_internal_loop(acc, input, len, secret, secretSize, XXH3_acc_128bits);
-
-    /* converge into final hash */
-    XXH_STATIC_ASSERT(sizeof(acc) == 64);
-    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
-    {   XXH128_hash_t h128;
-        h128.low64  = XXH3_mergeAccs(acc,
-                                     secret + XXH_SECRET_MERGEACCS_START,
-                                     (xxh_u64)len * PRIME64_1);
-        h128.high64 = XXH3_mergeAccs(acc,
-                                     secret + secretSize
-                                            - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
-                                     ~((xxh_u64)len * PRIME64_2));
-        return h128;
-    }
-}
-
-/*
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH128_hash_t
-XXH3_hashLong_128b_defaultSecret(const xxh_u8* input, size_t len)
-{
-    return XXH3_hashLong_128b_internal(input, len, kSecret, sizeof(kSecret));
-}
-
-/*
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH128_hash_t
-XXH3_hashLong_128b_withSecret(const xxh_u8* input, size_t len,
-                              const xxh_u8* secret, size_t secretSize)
-{
-    return XXH3_hashLong_128b_internal(input, len, secret, secretSize);
-}
-
-/*
- * It's important for performance that XXH3_hashLong is not inlined. Not sure
- * why (uop cache maybe?), but the difference is large and easily measurable.
- */
-XXH_NO_INLINE XXH128_hash_t
-XXH3_hashLong_128b_withSeed(const xxh_u8* input, size_t len, XXH64_hash_t seed)
-{
-    XXH_ALIGN(8) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
-    if (seed == 0) return XXH3_hashLong_128b_defaultSecret(input, len);
-    XXH3_initCustomSecret(secret, seed);
-    return XXH3_hashLong_128b_internal(input, len, secret, sizeof(secret));
-}
-
-
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* input, size_t len)
-{
-    if (len <= 16)
-        return XXH3_len_0to16_128b((const xxh_u8*)input, len, kSecret, 0);
-    if (len <= 128)
-        return XXH3_len_17to128_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0);
-    if (len <= XXH3_MIDSIZE_MAX)
-        return XXH3_len_129to240_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0);
-    return XXH3_hashLong_128b_defaultSecret((const xxh_u8*)input, len);
-}
-
-XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
-{
-    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
-    /*
-     * If an action is to be taken if `secret` conditions are not respected,
-     * it should be done here.
-     * For now, it's a contract pre-condition.
-     * Adding a check and a branch here would cost performance at every hash.
-     */
-    if (len <= 16)
-        return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, 0);
-    if (len <= 128)
-        return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0);
-    if (len <= XXH3_MIDSIZE_MAX)
-        return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0);
-    return XXH3_hashLong_128b_withSecret((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize);
-}
-
-XXH_PUBLIC_API XXH128_hash_t
-XXH3_128bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
-{
-    if (len <= 16)
-        return XXH3_len_0to16_128b((const xxh_u8*)input, len, kSecret, seed);
-    if (len <= 128)
-         return XXH3_len_17to128_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed);
-    if (len <= XXH3_MIDSIZE_MAX)
-         return XXH3_len_129to240_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed);
-    return XXH3_hashLong_128b_withSeed((const xxh_u8*)input, len, seed);
-}
-
-XXH_PUBLIC_API XXH128_hash_t
-XXH128(const void* input, size_t len, XXH64_hash_t seed)
-{
-    return XXH3_128bits_withSeed(input, len, seed);
-}
-
-
-/* ===   XXH3 128-bit streaming   === */
-
-/*
- * All the functions are actually the same as for 64-bit streaming variant.
- * The only difference is the finalizatiom routine.
- */
-
-static void
-XXH3_128bits_reset_internal(XXH3_state_t* statePtr,
-                            XXH64_hash_t seed,
-                            const xxh_u8* secret, size_t secretSize)
-{
-    XXH3_64bits_reset_internal(statePtr, seed, secret, secretSize);
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset(XXH3_state_t* statePtr)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_128bits_reset_internal(statePtr, 0, kSecret, XXH_SECRET_DEFAULT_SIZE);
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_128bits_reset_internal(statePtr, 0, (const xxh_u8*)secret, secretSize);
-    if (secret == NULL) return XXH_ERROR;
-    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed)
-{
-    if (statePtr == NULL) return XXH_ERROR;
-    XXH3_128bits_reset_internal(statePtr, seed, kSecret, XXH_SECRET_DEFAULT_SIZE);
-    XXH3_initCustomSecret(statePtr->customSecret, seed);
-    statePtr->secret = statePtr->customSecret;
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode
-XXH3_128bits_update(XXH3_state_t* state, const void* input, size_t len)
-{
-    return XXH3_update(state, (const xxh_u8*)input, len, XXH3_acc_128bits);
-}
-
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* state)
-{
-    if (state->totalLen > XXH3_MIDSIZE_MAX) {
-        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[ACC_NB];
-        XXH3_digest_long(acc, state, XXH3_acc_128bits);
-        XXH_ASSERT(state->secretLimit + STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
-        {   XXH128_hash_t h128;
-            h128.low64  = XXH3_mergeAccs(acc,
-                                         state->secret + XXH_SECRET_MERGEACCS_START,
-                                         (xxh_u64)state->totalLen * PRIME64_1);
-            h128.high64 = XXH3_mergeAccs(acc,
-                                         state->secret + state->secretLimit + STRIPE_LEN
-                                                       - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
-                                         ~((xxh_u64)state->totalLen * PRIME64_2));
-            return h128;
-        }
-    }
-    /* len <= XXH3_MIDSIZE_MAX : short code */
-    if (state->seed)
-        return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
-    return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen),
-                                   state->secret, state->secretLimit + STRIPE_LEN);
-}
-
-/* 128-bit utility functions */
-
-#include <string.h>   /* memcmp, memcpy */
-
-/* return : 1 is equal, 0 if different */
-XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
-{
-    /* note : XXH128_hash_t is compact, it has no padding byte */
-    return !(memcmp(&h1, &h2, sizeof(h1)));
-}
-
-/* This prototype is compatible with stdlib's qsort().
- * return : >0 if *h128_1  > *h128_2
- *          <0 if *h128_1  < *h128_2
- *          =0 if *h128_1 == *h128_2  */
-XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2)
-{
-    XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
-    XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2;
-    int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64);
-    /* note : bets that, in most cases, hash values are different */
-    if (hcmp) return hcmp;
-    return (h1.low64 > h2.low64) - (h2.low64 > h1.low64);
-}
-
-
-/*======   Canonical representation   ======*/
-XXH_PUBLIC_API void
-XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash)
-{
-    XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
-    if (XXH_CPU_LITTLE_ENDIAN) {
-        hash.high64 = XXH_swap64(hash.high64);
-        hash.low64  = XXH_swap64(hash.low64);
-    }
-    memcpy(dst, &hash.high64, sizeof(hash.high64));
-    memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
-}
-
-XXH_PUBLIC_API XXH128_hash_t
-XXH128_hashFromCanonical(const XXH128_canonical_t* src)
-{
-    XXH128_hash_t h;
-    h.high64 = XXH_readBE64(src);
-    h.low64  = XXH_readBE64(src->digest + 8);
-    return h;
-}
-
-/* Pop our optimization override from above */
-#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
-  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
-  && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
-#  pragma GCC pop_options
-#endif
-
-#endif  /* XXH3_H_1397135465 */

blitz.mod/hash/xxhash.h

@@ -1,7 +1,7 @@
 /*
  * xxHash - Extremely Fast Hash algorithm
  * Header File
- * Copyright (C) 2012-2020 Yann Collet
+ * Copyright (C) 2012-2023 Yann Collet
  *
  * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
  *
@@ -33,43 +33,213 @@
  *   - xxHash source repository: https://github.com/Cyan4973/xxHash
  */
 
-/* TODO: update */
-/* Notice extracted from xxHash homepage:
-
-xxHash is an extremely fast hash algorithm, running at RAM speed limits.
-It also successfully passes all tests from the SMHasher suite.
-
-Comparison (single thread, Windows Seven 32 bits, using SMHasher on a Core 2 Duo @3GHz)
-
-Name            Speed       Q.Score   Author
-xxHash          5.4 GB/s     10
-CrapWow         3.2 GB/s      2       Andrew
-MumurHash 3a    2.7 GB/s     10       Austin Appleby
-SpookyHash      2.0 GB/s     10       Bob Jenkins
-SBox            1.4 GB/s      9       Bret Mulvey
-Lookup3         1.2 GB/s      9       Bob Jenkins
-SuperFastHash   1.2 GB/s      1       Paul Hsieh
-CityHash64      1.05 GB/s    10       Pike & Alakuijala
-FNV             0.55 GB/s     5       Fowler, Noll, Vo
-CRC32           0.43 GB/s     9
-MD5-32          0.33 GB/s    10       Ronald L. Rivest
-SHA1-32         0.28 GB/s    10
-
-Q.Score is a measure of quality of the hash function.
-It depends on successfully passing SMHasher test set.
-10 is a perfect score.
-
-Note: SMHasher's CRC32 implementation is not the fastest one.
-Other speed-oriented implementations can be faster,
-especially in combination with PCLMUL instruction:
-https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html?showComment=1552696407071#c3490092340461170735
-
-A 64-bit version, named XXH64, is available since r35.
-It offers much better speed, but for 64-bit applications only.
-Name     Speed on 64 bits    Speed on 32 bits
-XXH64       13.8 GB/s            1.9 GB/s
-XXH32        6.8 GB/s            6.0 GB/s
-*/
+/*!
+ * @mainpage xxHash
+ *
+ * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed
+ * limits.
+ *
+ * It is proposed in four flavors, in three families:
+ * 1. @ref XXH32_family
+ *   - Classic 32-bit hash function. Simple, compact, and runs on almost all
+ *     32-bit and 64-bit systems.
+ * 2. @ref XXH64_family
+ *   - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most
+ *     64-bit systems (but _not_ 32-bit systems).
+ * 3. @ref XXH3_family
+ *   - Modern 64-bit and 128-bit hash function family which features improved
+ *     strength and performance across the board, especially on smaller data.
+ *     It benefits greatly from SIMD and 64-bit without requiring it.
+ *
+ * Benchmarks
+ * ---
+ * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04.
+ * The open source benchmark program is compiled with clang v10.0 using -O3 flag.
+ *
+ * | Hash Name            | ISA ext | Width | Large Data Speed | Small Data Velocity |
+ * | -------------------- | ------- | ----: | ---------------: | ------------------: |
+ * | XXH3_64bits()        | @b AVX2 |    64 |        59.4 GB/s |               133.1 |
+ * | MeowHash             | AES-NI  |   128 |        58.2 GB/s |                52.5 |
+ * | XXH3_128bits()       | @b AVX2 |   128 |        57.9 GB/s |               118.1 |
+ * | CLHash               | PCLMUL  |    64 |        37.1 GB/s |                58.1 |
+ * | XXH3_64bits()        | @b SSE2 |    64 |        31.5 GB/s |               133.1 |
+ * | XXH3_128bits()       | @b SSE2 |   128 |        29.6 GB/s |               118.1 |
+ * | RAM sequential read  |         |   N/A |        28.0 GB/s |                 N/A |
+ * | ahash                | AES-NI  |    64 |        22.5 GB/s |               107.2 |
+ * | City64               |         |    64 |        22.0 GB/s |                76.6 |
+ * | T1ha2                |         |    64 |        22.0 GB/s |                99.0 |
+ * | City128              |         |   128 |        21.7 GB/s |                57.7 |
+ * | FarmHash             | AES-NI  |    64 |        21.3 GB/s |                71.9 |
+ * | XXH64()              |         |    64 |        19.4 GB/s |                71.0 |
+ * | SpookyHash           |         |    64 |        19.3 GB/s |                53.2 |
+ * | Mum                  |         |    64 |        18.0 GB/s |                67.0 |
+ * | CRC32C               | SSE4.2  |    32 |        13.0 GB/s |                57.9 |
+ * | XXH32()              |         |    32 |         9.7 GB/s |                71.9 |
+ * | City32               |         |    32 |         9.1 GB/s |                66.0 |
+ * | Blake3*              | @b AVX2 |   256 |         4.4 GB/s |                 8.1 |
+ * | Murmur3              |         |    32 |         3.9 GB/s |                56.1 |
+ * | SipHash*             |         |    64 |         3.0 GB/s |                43.2 |
+ * | Blake3*              | @b SSE2 |   256 |         2.4 GB/s |                 8.1 |
+ * | HighwayHash          |         |    64 |         1.4 GB/s |                 6.0 |
+ * | FNV64                |         |    64 |         1.2 GB/s |                62.7 |
+ * | Blake2*              |         |   256 |         1.1 GB/s |                 5.1 |
+ * | SHA1*                |         |   160 |         0.8 GB/s |                 5.6 |
+ * | MD5*                 |         |   128 |         0.6 GB/s |                 7.8 |
+ * @note
+ *   - Hashes which require a specific ISA extension are noted. SSE2 is also noted,
+ *     even though it is mandatory on x64.
+ *   - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic
+ *     by modern standards.
+ *   - Small data velocity is a rough average of algorithm's efficiency for small
+ *     data. For more accurate information, see the wiki.
+ *   - More benchmarks and strength tests are found on the wiki:
+ *         https://github.com/Cyan4973/xxHash/wiki
+ *
+ * Usage
+ * ------
+ * All xxHash variants use a similar API. Changing the algorithm is a trivial
+ * substitution.
+ *
+ * @pre
+ *    For functions which take an input and length parameter, the following
+ *    requirements are assumed:
+ *    - The range from [`input`, `input + length`) is valid, readable memory.
+ *      - The only exception is if the `length` is `0`, `input` may be `NULL`.
+ *    - For C++, the objects must have the *TriviallyCopyable* property, as the
+ *      functions access bytes directly as if it was an array of `unsigned char`.
+ *
+ * @anchor single_shot_example
+ * **Single Shot**
+ *
+ * These functions are stateless functions which hash a contiguous block of memory,
+ * immediately returning the result. They are the easiest and usually the fastest
+ * option.
+ *
+ * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits()
+ *
+ * @code{.c}
+ *   #include <string.h>
+ *   #include "xxhash.h"
+ *
+ *   // Example for a function which hashes a null terminated string with XXH32().
+ *   XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed)
+ *   {
+ *       // NULL pointers are only valid if the length is zero
+ *       size_t length = (string == NULL) ? 0 : strlen(string);
+ *       return XXH32(string, length, seed);
+ *   }
+ * @endcode
+ *
+ *
+ * @anchor streaming_example
+ * **Streaming**
+ *
+ * These groups of functions allow incremental hashing of unknown size, even
+ * more than what would fit in a size_t.
+ *
+ * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset()
+ *
+ * @code{.c}
+ *   #include <stdio.h>
+ *   #include <assert.h>
+ *   #include "xxhash.h"
+ *   // Example for a function which hashes a FILE incrementally with XXH3_64bits().
+ *   XXH64_hash_t hashFile(FILE* f)
+ *   {
+ *       // Allocate a state struct. Do not just use malloc() or new.
+ *       XXH3_state_t* state = XXH3_createState();
+ *       assert(state != NULL && "Out of memory!");
+ *       // Reset the state to start a new hashing session.
+ *       XXH3_64bits_reset(state);
+ *       char buffer[4096];
+ *       size_t count;
+ *       // Read the file in chunks
+ *       while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) {
+ *           // Run update() as many times as necessary to process the data
+ *           XXH3_64bits_update(state, buffer, count);
+ *       }
+ *       // Retrieve the finalized hash. This will not change the state.
+ *       XXH64_hash_t result = XXH3_64bits_digest(state);
+ *       // Free the state. Do not use free().
+ *       XXH3_freeState(state);
+ *       return result;
+ *   }
+ * @endcode
+ *
+ * Streaming functions generate the xxHash value from an incremental input.
+ * This method is slower than single-call functions, due to state management.
+ * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
+ *
+ * An XXH state must first be allocated using `XXH*_createState()`.
+ *
+ * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
+ *
+ * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
+ *
+ * The function returns an error code, with 0 meaning OK, and any other value
+ * meaning there is an error.
+ *
+ * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
+ * This function returns the nn-bits hash as an int or long long.
+ *
+ * It's still possible to continue inserting input into the hash state after a
+ * digest, and generate new hash values later on by invoking `XXH*_digest()`.
+ *
+ * When done, release the state using `XXH*_freeState()`.
+ *
+ *
+ * @anchor canonical_representation_example
+ * **Canonical Representation**
+ *
+ * The default return values from XXH functions are unsigned 32, 64 and 128 bit
+ * integers.
+ * This the simplest and fastest format for further post-processing.
+ *
+ * However, this leaves open the question of what is the order on the byte level,
+ * since little and big endian conventions will store the same number differently.
+ *
+ * The canonical representation settles this issue by mandating big-endian
+ * convention, the same convention as human-readable numbers (large digits first).
+ *
+ * When writing hash values to storage, sending them over a network, or printing
+ * them, it's highly recommended to use the canonical representation to ensure
+ * portability across a wider range of systems, present and future.
+ *
+ * The following functions allow transformation of hash values to and from
+ * canonical format.
+ *
+ * XXH32_canonicalFromHash(), XXH32_hashFromCanonical(),
+ * XXH64_canonicalFromHash(), XXH64_hashFromCanonical(),
+ * XXH128_canonicalFromHash(), XXH128_hashFromCanonical(),
+ *
+ * @code{.c}
+ *   #include <stdio.h>
+ *   #include "xxhash.h"
+ *
+ *   // Example for a function which prints XXH32_hash_t in human readable format
+ *   void printXxh32(XXH32_hash_t hash)
+ *   {
+ *       XXH32_canonical_t cano;
+ *       XXH32_canonicalFromHash(&cano, hash);
+ *       size_t i;
+ *       for(i = 0; i < sizeof(cano.digest); ++i) {
+ *           printf("%02x", cano.digest[i]);
+ *       }
+ *       printf("\n");
+ *   }
+ *
+ *   // Example for a function which converts XXH32_canonical_t to XXH32_hash_t
+ *   XXH32_hash_t convertCanonicalToXxh32(XXH32_canonical_t cano)
+ *   {
+ *       XXH32_hash_t hash = XXH32_hashFromCanonical(&cano);
+ *       return hash;
+ *   }
+ * @endcode
+ *
+ *
+ * @file xxhash.h
+ * xxHash prototypes and implementation
+ */
 
 #if defined (__cplusplus)
 extern "C" {
@@ -79,21 +249,80 @@ extern "C" {
  *  INLINE mode
  ******************************/
 /*!
- * XXH_INLINE_ALL (and XXH_PRIVATE_API)
+ * @defgroup public Public API
+ * Contains details on the public xxHash functions.
+ * @{
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Gives access to internal state declaration, required for static allocation.
+ *
+ * Incompatible with dynamic linking, due to risks of ABI changes.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_STATIC_LINKING_ONLY
+ *     #include "xxhash.h"
+ * @endcode
+ */
+#  define XXH_STATIC_LINKING_ONLY
+/* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */
+
+/*!
+ * @brief Gives access to internal definitions.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_STATIC_LINKING_ONLY
+ *     #define XXH_IMPLEMENTATION
+ *     #include "xxhash.h"
+ * @endcode
+ */
+#  define XXH_IMPLEMENTATION
+/* Do not undef XXH_IMPLEMENTATION for Doxygen */
+
+/*!
+ * @brief Exposes the implementation and marks all functions as `inline`.
+ *
  * Use these build macros to inline xxhash into the target unit.
  * Inlining improves performance on small inputs, especially when the length is
  * expressed as a compile-time constant:
  *
- *      https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
+ *  https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
  *
  * It also keeps xxHash symbols private to the unit, so they are not exported.
  *
  * Usage:
+ * @code{.c}
  *     #define XXH_INLINE_ALL
  *     #include "xxhash.h"
- *
+ * @endcode
  * Do not compile and link xxhash.o as a separate object, as it is not useful.
  */
+#  define XXH_INLINE_ALL
+#  undef XXH_INLINE_ALL
+/*!
+ * @brief Exposes the implementation without marking functions as inline.
+ */
+#  define XXH_PRIVATE_API
+#  undef XXH_PRIVATE_API
+/*!
+ * @brief Emulate a namespace by transparently prefixing all symbols.
+ *
+ * If you want to include _and expose_ xxHash functions from within your own
+ * library, but also want to avoid symbol collisions with other libraries which
+ * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix
+ * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE
+ * (therefore, avoid empty or numeric values).
+ *
+ * Note that no change is required within the calling program as long as it
+ * includes `xxhash.h`: Regular symbol names will be automatically translated
+ * by this header.
+ */
+#  define XXH_NAMESPACE /* YOUR NAME HERE */
+#  undef XXH_NAMESPACE
+#endif
+
 #if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
     && !defined(XXH_INLINE_ALL_31684351384)
    /* this section should be traversed only once */
@@ -104,7 +333,7 @@ extern "C" {
    /* make all functions private */
 #  undef XXH_PUBLIC_API
 #  if defined(__GNUC__)
-#    define XXH_PUBLIC_API static __inline __attribute__((unused))
+#    define XXH_PUBLIC_API static __inline __attribute__((__unused__))
 #  elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
 #    define XXH_PUBLIC_API static inline
 #  elif defined(_MSC_VER)
@@ -116,29 +345,80 @@ extern "C" {
 
    /*
     * This part deals with the special case where a unit wants to inline xxHash,
-    * but "xxhash.h" has previously been included without XXH_INLINE_ALL, such
-    * as part of some previously included *.h header file.
+    * but "xxhash.h" has previously been included without XXH_INLINE_ALL,
+    * such as part of some previously included *.h header file.
     * Without further action, the new include would just be ignored,
     * and functions would effectively _not_ be inlined (silent failure).
     * The following macros solve this situation by prefixing all inlined names,
     * avoiding naming collision with previous inclusions.
     */
-#  ifdef XXH_NAMESPACE
-#    error "XXH_INLINE_ALL with XXH_NAMESPACE is not supported"
-     /*
-      * Note: Alternative: #undef all symbols (it's a pretty large list).
-      * Without #error: it compiles, but functions are actually not inlined.
-      */
-#  endif
+   /* Before that, we unconditionally #undef all symbols,
+    * in case they were already defined with XXH_NAMESPACE.
+    * They will then be redefined for XXH_INLINE_ALL
+    */
+#  undef XXH_versionNumber
+    /* XXH32 */
+#  undef XXH32
+#  undef XXH32_createState
+#  undef XXH32_freeState
+#  undef XXH32_reset
+#  undef XXH32_update
+#  undef XXH32_digest
+#  undef XXH32_copyState
+#  undef XXH32_canonicalFromHash
+#  undef XXH32_hashFromCanonical
+    /* XXH64 */
+#  undef XXH64
+#  undef XXH64_createState
+#  undef XXH64_freeState
+#  undef XXH64_reset
+#  undef XXH64_update
+#  undef XXH64_digest
+#  undef XXH64_copyState
+#  undef XXH64_canonicalFromHash
+#  undef XXH64_hashFromCanonical
+    /* XXH3_64bits */
+#  undef XXH3_64bits
+#  undef XXH3_64bits_withSecret
+#  undef XXH3_64bits_withSeed
+#  undef XXH3_64bits_withSecretandSeed
+#  undef XXH3_createState
+#  undef XXH3_freeState
+#  undef XXH3_copyState
+#  undef XXH3_64bits_reset
+#  undef XXH3_64bits_reset_withSeed
+#  undef XXH3_64bits_reset_withSecret
+#  undef XXH3_64bits_update
+#  undef XXH3_64bits_digest
+#  undef XXH3_generateSecret
+    /* XXH3_128bits */
+#  undef XXH128
+#  undef XXH3_128bits
+#  undef XXH3_128bits_withSeed
+#  undef XXH3_128bits_withSecret
+#  undef XXH3_128bits_reset
+#  undef XXH3_128bits_reset_withSeed
+#  undef XXH3_128bits_reset_withSecret
+#  undef XXH3_128bits_reset_withSecretandSeed
+#  undef XXH3_128bits_update
+#  undef XXH3_128bits_digest
+#  undef XXH128_isEqual
+#  undef XXH128_cmp
+#  undef XXH128_canonicalFromHash
+#  undef XXH128_hashFromCanonical
+    /* Finally, free the namespace itself */
+#  undef XXH_NAMESPACE
+
+    /* employ the namespace for XXH_INLINE_ALL */
 #  define XXH_NAMESPACE XXH_INLINE_
    /*
-    * Some identifiers (enums, type names) are not symbols, but they must
-    * still be renamed to avoid redeclaration.
+    * Some identifiers (enums, type names) are not symbols,
+    * but they must nonetheless be renamed to avoid redeclaration.
     * Alternative solution: do not redeclare them.
-    * However, this requires some #ifdefs, and is a more dispersed action.
-    * Meanwhile, renaming can be achieved in a single block
+    * However, this requires some #ifdefs, and has a more dispersed impact.
+    * Meanwhile, renaming can be achieved in a single place.
     */
-#  define XXH_IPREF(Id)   XXH_INLINE_ ## Id
+#  define XXH_IPREF(Id)   XXH_NAMESPACE ## Id
 #  define XXH_OK XXH_IPREF(XXH_OK)
 #  define XXH_ERROR XXH_IPREF(XXH_ERROR)
 #  define XXH_errorcode XXH_IPREF(XXH_errorcode)
@@ -157,17 +437,15 @@ extern "C" {
 #  undef XXHASH_H_STATIC_13879238742
 #endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */
 
-
-
 /* ****************************************************************
  *  Stable API
  *****************************************************************/
 #ifndef XXHASH_H_5627135585666179
 #define XXHASH_H_5627135585666179 1
 
-/* specific declaration modes for Windows */
+/*! @brief Marks a global symbol. */
 #if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
-#  if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+#  if defined(_WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
 #    ifdef XXH_EXPORT
 #      define XXH_PUBLIC_API __declspec(dllexport)
 #    elif XXH_IMPORT
@@ -178,23 +456,11 @@ extern "C" {
 #  endif
 #endif
 
-/*!
- * XXH_NAMESPACE, aka Namespace Emulation:
- *
- * If you want to include _and expose_ xxHash functions from within your own
- * library, but also want to avoid symbol collisions with other libraries which
- * may also include xxHash, you can use XXH_NAMESPACE to automatically prefix
- * any public symbol from xxhash library with the value of XXH_NAMESPACE
- * (therefore, avoid empty or numeric values).
- *
- * Note that no change is required within the calling program as long as it
- * includes `xxhash.h`: Regular symbol names will be automatically translated
- * by this header.
- */
 #ifdef XXH_NAMESPACE
 #  define XXH_CAT(A,B) A##B
 #  define XXH_NAME2(A,B) XXH_CAT(A,B)
 #  define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber)
+/* XXH32 */
 #  define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32)
 #  define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState)
 #  define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState)
@@ -204,6 +470,7 @@ extern "C" {
 #  define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState)
 #  define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash)
 #  define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical)
+/* XXH64 */
 #  define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64)
 #  define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState)
 #  define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState)
@@ -213,1753 +480,6609 @@ extern "C" {
 #  define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState)
 #  define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash)
 #  define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical)
+/* XXH3_64bits */
+#  define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
+#  define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
+#  define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
+#  define XXH3_64bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecretandSeed)
+#  define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
+#  define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
+#  define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
+#  define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
+#  define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
+#  define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
+#  define XXH3_64bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecretandSeed)
+#  define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
+#  define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
+#  define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret)
+#  define XXH3_generateSecret_fromSeed XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret_fromSeed)
+/* XXH3_128bits */
+#  define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
+#  define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
+#  define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
+#  define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
+#  define XXH3_128bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecretandSeed)
+#  define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
+#  define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
+#  define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
+#  define XXH3_128bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecretandSeed)
+#  define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
+#  define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
+#  define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
+#  define XXH128_cmp     XXH_NAME2(XXH_NAMESPACE, XXH128_cmp)
+#  define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash)
+#  define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical)
+#endif
+
+
+/* *************************************
+*  Compiler specifics
+***************************************/
+
+/* specific declaration modes for Windows */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+#  if defined(_WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+#    ifdef XXH_EXPORT
+#      define XXH_PUBLIC_API __declspec(dllexport)
+#    elif XXH_IMPORT
+#      define XXH_PUBLIC_API __declspec(dllimport)
+#    endif
+#  else
+#    define XXH_PUBLIC_API   /* do nothing */
+#  endif
 #endif
 
+#if defined (__GNUC__)
+# define XXH_CONSTF  __attribute__((__const__))
+# define XXH_PUREF   __attribute__((__pure__))
+# define XXH_MALLOCF __attribute__((__malloc__))
+#else
+# define XXH_CONSTF  /* disable */
+# define XXH_PUREF
+# define XXH_MALLOCF
+#endif
 
 /* *************************************
 *  Version
 ***************************************/
 #define XXH_VERSION_MAJOR    0
-#define XXH_VERSION_MINOR    7
-#define XXH_VERSION_RELEASE  4
+#define XXH_VERSION_MINOR    8
+#define XXH_VERSION_RELEASE  3
+/*! @brief Version number, encoded as two digits each */
 #define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
-XXH_PUBLIC_API unsigned XXH_versionNumber (void);
+
+/*!
+ * @brief Obtains the xxHash version.
+ *
+ * This is mostly useful when xxHash is compiled as a shared library,
+ * since the returned value comes from the library, as opposed to header file.
+ *
+ * @return @ref XXH_VERSION_NUMBER of the invoked library.
+ */
+XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void);
 
 
 /* ****************************
-*  Definitions
+*  Common basic types
 ******************************/
 #include <stddef.h>   /* size_t */
-typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
+/*!
+ * @brief Exit code for the streaming API.
+ */
+typedef enum {
+    XXH_OK = 0, /*!< OK */
+    XXH_ERROR   /*!< Error */
+} XXH_errorcode;
 
 
 /*-**********************************************************************
 *  32-bit hash
 ************************************************************************/
-#if !defined (__VMS) \
+#if defined(XXH_DOXYGEN) /* Don't show <stdint.h> include */
+/*!
+ * @brief An unsigned 32-bit integer.
+ *
+ * Not necessarily defined to `uint32_t` but functionally equivalent.
+ */
+typedef uint32_t XXH32_hash_t;
+
+#elif !defined (__VMS) \
   && (defined (__cplusplus) \
   || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-#   include <stdint.h>
+#   ifdef _AIX
+#     include <inttypes.h>
+#   else
+#     include <stdint.h>
+#   endif
     typedef uint32_t XXH32_hash_t;
+
 #else
 #   include <limits.h>
 #   if UINT_MAX == 0xFFFFFFFFUL
       typedef unsigned int XXH32_hash_t;
+#   elif ULONG_MAX == 0xFFFFFFFFUL
+      typedef unsigned long XXH32_hash_t;
 #   else
-#     if ULONG_MAX == 0xFFFFFFFFUL
-        typedef unsigned long XXH32_hash_t;
-#     else
-#       error "unsupported platform: need a 32-bit type"
-#     endif
+#     error "unsupported platform: need a 32-bit type"
 #   endif
 #endif
 
 /*!
- * XXH32():
- *  Calculate the 32-bit hash of sequence "length" bytes stored at memory address "input".
- *  The memory between input & input+length must be valid (allocated and read-accessible).
- *  "seed" can be used to alter the result predictably.
- *  Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s
+ * @}
+ *
+ * @defgroup XXH32_family XXH32 family
+ * @ingroup public
+ * Contains functions used in the classic 32-bit xxHash algorithm.
  *
- * Note: XXH3 provides competitive speed for both 32-bit and 64-bit systems,
- * and offers true 64/128 bit hash results. It provides a superior level of
- * dispersion, and greatly reduces the risks of collisions.
+ * @note
+ *   XXH32 is useful for older platforms, with no or poor 64-bit performance.
+ *   Note that the @ref XXH3_family provides competitive speed for both 32-bit
+ *   and 64-bit systems, and offers true 64/128 bit hash results.
+ *
+ * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families
+ * @see @ref XXH32_impl for implementation details
+ * @{
  */
-XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
 
-/*******   Streaming   *******/
+/*!
+ * @brief Calculates the 32-bit hash of @p input using xxHash32.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 32-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 32-bit xxHash32 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
 
-/*
- * Streaming functions generate the xxHash value from an incrememtal input.
- * This method is slower than single-call functions, due to state management.
- * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
+#ifndef XXH_NO_STREAM
+/*!
+ * @typedef struct XXH32_state_s XXH32_state_t
+ * @brief The opaque state struct for the XXH32 streaming API.
  *
- * An XXH state must first be allocated using `XXH*_createState()`.
+ * @see XXH32_state_s for details.
+ * @see @ref streaming_example "Streaming Example"
+ */
+typedef struct XXH32_state_s XXH32_state_t;
+
+/*!
+ * @brief Allocates an @ref XXH32_state_t.
  *
- * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
+ * @return An allocated pointer of @ref XXH32_state_t on success.
+ * @return `NULL` on failure.
  *
- * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
+ * @note Must be freed with XXH32_freeState().
  *
- * The function returns an error code, with 0 meaning OK, and any other value
- * meaning there is an error.
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void);
+/*!
+ * @brief Frees an @ref XXH32_state_t.
  *
- * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
- * This function returns the nn-bits hash as an int or long long.
+ * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState().
  *
- * It's still possible to continue inserting input into the hash state after a
- * digest, and generate new hash values later on by invoking `XXH*_digest()`.
+ * @return @ref XXH_OK.
+ *
+ * @note @p statePtr must be allocated with XXH32_createState().
+ *
+ * @see @ref streaming_example "Streaming Example"
  *
- * When done, release the state using `XXH*_freeState()`.
  */
-
-typedef struct XXH32_state_s XXH32_state_t;   /* incomplete type */
-XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void);
 XXH_PUBLIC_API XXH_errorcode  XXH32_freeState(XXH32_state_t* statePtr);
+/*!
+ * @brief Copies one @ref XXH32_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
 XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);
 
+/*!
+ * @brief Resets an @ref XXH32_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 32-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note This function resets and seeds a state. Call it before @ref XXH32_update().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
 XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, XXH32_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH32_hash_t  XXH32_digest (const XXH32_state_t* statePtr);
 
-/*******   Canonical representation   *******/
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH32_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
 
-/*
- * The default return values from XXH functions are unsigned 32 and 64 bit
- * integers.
- * This the simplest and fastest format for further post-processing.
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH32_state_t.
  *
- * However, this leaves open the question of what is the order on the byte level,
- * since little and big endian conventions will store the same number differently.
+ * @param statePtr The state struct to calculate the hash from.
  *
- * The canonical representation settles this issue by mandating big-endian
- * convention, the same convention as human-readable numbers (large digits first).
+ * @pre
+ *  @p statePtr must not be `NULL`.
  *
- * When writing hash values to storage, sending them over a network, or printing
- * them, it's highly recommended to use the canonical representation to ensure
- * portability across a wider range of systems, present and future.
+ * @return The calculated 32-bit xxHash32 value from that state.
  *
- * The following functions allow transformation of hash values to and from
- * canonical format.
+ * @note
+ *   Calling XXH32_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @see @ref streaming_example "Streaming Example"
  */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
 
-typedef struct { unsigned char digest[4]; } XXH32_canonical_t;
-XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
-XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+/*******   Canonical representation   *******/
 
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH32_hash_t.
+ */
+typedef struct {
+    unsigned char digest[4]; /*!< Hash bytes, big endian */
+} XXH32_canonical_t;
 
-#ifndef XXH_NO_LONG_LONG
-/*-**********************************************************************
-*  64-bit hash
-************************************************************************/
-#if !defined (__VMS) \
-  && (defined (__cplusplus) \
-  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-#   include <stdint.h>
-    typedef uint64_t XXH64_hash_t;
-#else
-    /* the following type must have a width of 64-bit */
-    typedef unsigned long long XXH64_hash_t;
-#endif
+/*!
+ * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t.
+ *
+ * @param dst  The @ref XXH32_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH32_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
 
 /*!
- * XXH64():
- * Returns the 64-bit hash of sequence of length @length stored at memory
- * address @input.
- * @seed can be used to alter the result predictably.
+ * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t.
+ *
+ * @param src The @ref XXH32_canonical_t to convert.
  *
- * This function usually runs faster on 64-bit systems, but slower on 32-bit
- * systems (see benchmark).
+ * @pre
+ *   @p src must not be `NULL`.
  *
- * Note: XXH3 provides competitive speed for both 32-bit and 64-bit systems,
- * and offers true 64/128 bit hash results. It provides a superior level of
- * dispersion, and greatly reduces the risks of collisions.
+ * @return The converted hash.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
  */
-XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t length, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
 
-/*******   Streaming   *******/
+
+/*! @cond Doxygen ignores this part */
+#ifdef __has_attribute
+# define XXH_HAS_ATTRIBUTE(x) __has_attribute(x)
+#else
+# define XXH_HAS_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * C23 __STDC_VERSION__ number hasn't been specified yet. For now
+ * leave as `201711L` (C17 + 1).
+ * TODO: Update to correct value when its been specified.
+ */
+#define XXH_C23_VN 201711L
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/* C-language Attributes are added in C23. */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute)
+# define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
+#else
+# define XXH_HAS_C_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+#if defined(__cplusplus) && defined(__has_cpp_attribute)
+# define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+# define XXH_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute
+ * introduced in CPP17 and C23.
+ * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough
+ * C23   : https://en.cppreference.com/w/c/language/attributes/fallthrough
+ */
+#if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough)
+# define XXH_FALLTHROUGH [[fallthrough]]
+#elif XXH_HAS_ATTRIBUTE(__fallthrough__)
+# define XXH_FALLTHROUGH __attribute__ ((__fallthrough__))
+#else
+# define XXH_FALLTHROUGH /* fallthrough */
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * Define XXH_NOESCAPE for annotated pointers in public API.
+ * https://clang.llvm.org/docs/AttributeReference.html#noescape
+ * As of writing this, only supported by clang.
+ */
+#if XXH_HAS_ATTRIBUTE(noescape)
+# define XXH_NOESCAPE __attribute__((__noescape__))
+#else
+# define XXH_NOESCAPE
+#endif
+/*! @endcond */
+
+
+/*!
+ * @}
+ * @ingroup public
+ * @{
+ */
+
+#ifndef XXH_NO_LONG_LONG
+/*-**********************************************************************
+*  64-bit hash
+************************************************************************/
+#if defined(XXH_DOXYGEN) /* don't include <stdint.h> */
+/*!
+ * @brief An unsigned 64-bit integer.
+ *
+ * Not necessarily defined to `uint64_t` but functionally equivalent.
+ */
+typedef uint64_t XXH64_hash_t;
+#elif !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   ifdef _AIX
+#     include <inttypes.h>
+#   else
+#     include <stdint.h>
+#   endif
+   typedef uint64_t XXH64_hash_t;
+#else
+#  include <limits.h>
+#  if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL
+     /* LP64 ABI says uint64_t is unsigned long */
+     typedef unsigned long XXH64_hash_t;
+#  else
+     /* the following type must have a width of 64-bit */
+     typedef unsigned long long XXH64_hash_t;
+#  endif
+#endif
+
+/*!
+ * @}
+ *
+ * @defgroup XXH64_family XXH64 family
+ * @ingroup public
+ * @{
+ * Contains functions used in the classic 64-bit xxHash algorithm.
+ *
+ * @note
+ *   XXH3 provides competitive speed for both 32-bit and 64-bit systems,
+ *   and offers true 64/128 bit hash results.
+ *   It provides better speed for systems with vector processing capabilities.
+ */
+
+/*!
+ * @brief Calculates the 64-bit hash of @p input using xxHash64.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit xxHash64 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
+
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*!
+ * @brief The opaque state struct for the XXH64 streaming API.
+ *
+ * @see XXH64_state_s for details.
+ * @see @ref streaming_example "Streaming Example"
+ */
 typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
-XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void);
+
+/*!
+ * @brief Allocates an @ref XXH64_state_t.
+ *
+ * @return An allocated pointer of @ref XXH64_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH64_freeState().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);
+
+/*!
+ * @brief Frees an @ref XXH64_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note @p statePtr must be allocated with XXH64_createState().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
 XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dst_state, const XXH64_state_t* src_state);
 
-XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH64_state_t* statePtr, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH64_hash_t  XXH64_digest (const XXH64_state_t* statePtr);
+/*!
+ * @brief Copies one @ref XXH64_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH64_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note This function resets and seeds a state. Call it before @ref XXH64_update().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH64_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH64_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated 64-bit xxHash64 value from that state.
+ *
+ * @note
+ *   Calling XXH64_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+/*******   Canonical representation   *******/
+
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH64_hash_t.
+ */
+typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
+
+/*!
+ * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t.
+ *
+ * @param dst The @ref XXH64_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH64_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t.
+ *
+ * @param src The @ref XXH64_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src);
+
+#ifndef XXH_NO_XXH3
+
+/*!
+ * @}
+ * ************************************************************************
+ * @defgroup XXH3_family XXH3 family
+ * @ingroup public
+ * @{
+ *
+ * XXH3 is a more recent hash algorithm featuring:
+ *  - Improved speed for both small and large inputs
+ *  - True 64-bit and 128-bit outputs
+ *  - SIMD acceleration
+ *  - Improved 32-bit viability
+ *
+ * Speed analysis methodology is explained here:
+ *
+ *    https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ *
+ * Compared to XXH64, expect XXH3 to run approximately
+ * ~2x faster on large inputs and >3x faster on small ones,
+ * exact differences vary depending on platform.
+ *
+ * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
+ * but does not require it.
+ * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3
+ * at competitive speeds, even without vector support. Further details are
+ * explained in the implementation.
+ *
+ * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD
+ * implementations for many common platforms:
+ *   - AVX512
+ *   - AVX2
+ *   - SSE2
+ *   - ARM NEON
+ *   - WebAssembly SIMD128
+ *   - POWER8 VSX
+ *   - s390x ZVector
+ * This can be controlled via the @ref XXH_VECTOR macro, but it automatically
+ * selects the best version according to predefined macros. For the x86 family, an
+ * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c.
+ *
+ * XXH3 implementation is portable:
+ * it has a generic C90 formulation that can be compiled on any platform,
+ * all implementations generate exactly the same hash value on all platforms.
+ * Starting from v0.8.0, it's also labelled "stable", meaning that
+ * any future version will also generate the same hash value.
+ *
+ * XXH3 offers 2 variants, _64bits and _128bits.
+ *
+ * When only 64 bits are needed, prefer invoking the _64bits variant, as it
+ * reduces the amount of mixing, resulting in faster speed on small inputs.
+ * It's also generally simpler to manipulate a scalar return type than a struct.
+ *
+ * The API supports one-shot hashing, streaming mode, and custom secrets.
+ */
+/*-**********************************************************************
+*  XXH3 64-bit variant
+************************************************************************/
+
+/*!
+ * @brief Calculates 64-bit unseeded variant of XXH3 hash of @p input.
+ *
+ * @param input  The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @note
+ *   This is equivalent to @ref XXH3_64bits_withSeed() with a seed of `0`, however
+ *   it may have slightly better performance due to constant propagation of the
+ *   defaults.
+ *
+ * @see
+ *    XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Calculates 64-bit seeded variant of XXH3 hash of @p input.
+ *
+ * @param input  The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed   The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @note
+ *    seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the @p seed value.
+ *
+ * While this operation is decently fast, note that it's not completely free.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
+
+/*!
+ * The bare minimum size for a custom secret.
+ *
+ * @see
+ *  XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(),
+ *  XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret().
+ */
+#define XXH3_SECRET_SIZE_MIN 136
+
+/*!
+ * @brief Calculates 64-bit variant of XXH3 with a custom "secret".
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @pre
+ *   The memory between @p data and @p data + @p len must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p data may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
+ * However, the quality of the secret impacts the dispersion of the hash algorithm.
+ * Therefore, the secret _must_ look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever in doubt about the "randomness" of the blob of bytes,
+ * consider employing @ref XXH3_generateSecret() instead (see below).
+ * It will generate a proper high entropy secret derived from the blob of bytes.
+ * Another advantage of using XXH3_generateSecret() is that
+ * it guarantees that all bits within the initial blob of bytes
+ * will impact every bit of the output.
+ * This is not necessarily the case when using the blob of bytes directly
+ * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ */
+
+/*!
+ * @brief The opaque state struct for the XXH3 streaming API.
+ *
+ * @see XXH3_state_s for details.
+ * @see @ref streaming_example "Streaming Example"
+ */
+typedef struct XXH3_state_s XXH3_state_t;
+XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
+
+/*!
+ * @brief Copies one @ref XXH3_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret with default parameters.
+ *   - Call this function before @ref XXH3_64bits_update().
+ *   - Digest will be equivalent to `XXH3_64bits()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed     The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret from `seed`.
+ *   - Call this function before @ref XXH3_64bits_update().
+ *   - Digest will be equivalent to `XXH3_64bits_withSeed()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   `secret` is referenced, it _must outlive_ the hash streaming session.
+ *
+ * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 64-bit hash value from that state.
+ *
+ * @note
+ *   Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t  XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/* note : canonical representation of XXH3 is the same as XXH64
+ * since they both produce XXH64_hash_t values */
+
+
+/*-**********************************************************************
+*  XXH3 128-bit variant
+************************************************************************/
+
+/*!
+ * @brief The return value from 128-bit hashes.
+ *
+ * Stored in little endian order, although the fields themselves are in native
+ * endianness.
+ */
+typedef struct {
+    XXH64_hash_t low64;   /*!< `value & 0xFFFFFFFFFFFFFFFF` */
+    XXH64_hash_t high64;  /*!< `value >> 64` */
+} XXH128_hash_t;
+
+/*!
+ * @brief Calculates 128-bit unseeded variant of XXH3 of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p length bytes in size.
+ * @param len  The length of @p data, in bytes.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead
+ * for shorter inputs.
+ *
+ * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of `0`, however
+ * it may have slightly better performance due to constant propagation of the
+ * defaults.
+ *
+ * @see XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len);
+/*! @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p length bytes in size.
+ * @param len  The length of @p data, in bytes.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * @note
+ *    seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the @p seed value.
+ *
+ * While this operation is decently fast, note that it's not completely free.
+ *
+ * @see XXH3_128bits(), XXH3_128bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+/*!
+ * @brief Calculates 128-bit variant of XXH3 with a custom "secret".
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
+ * However, the quality of the secret impacts the dispersion of the hash algorithm.
+ * Therefore, the secret _must_ look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever in doubt about the "randomness" of the blob of bytes,
+ * consider employing @ref XXH3_generateSecret() instead (see below).
+ * It will generate a proper high entropy secret derived from the blob of bytes.
+ * Another advantage of using XXH3_generateSecret() is that
+ * it guarantees that all bits within the initial blob of bytes
+ * will impact every bit of the output.
+ * This is not necessarily the case when using the blob of bytes directly
+ * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ *
+ * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits().
+ * Use already declared XXH3_createState() and XXH3_freeState().
+ *
+ * All reset and streaming functions have same meaning as their 64-bit counterpart.
+ */
+
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret with default parameters.
+ *   - Call it before @ref XXH3_128bits_update().
+ *   - Digest will be equivalent to `XXH3_128bits()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed     The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret from `seed`.
+ *   - Call it before @ref XXH3_128bits_update().
+ *   - Digest will be equivalent to `XXH3_128bits_withSeed()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   The state struct to reset.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * `secret` is referenced, it _must outlive_ the hash streaming session.
+ * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * Call this to incrementally consume blocks of data.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 128-bit hash value from that state.
+ *
+ * @note
+ *   Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/* Following helper functions make it possible to compare XXH128_hast_t values.
+ * Since XXH128_hash_t is a structure, this capability is not offered by the language.
+ * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
+
+/*!
+ * @brief Check equality of two XXH128_hash_t values
+ *
+ * @param h1 The 128-bit hash value.
+ * @param h2 Another 128-bit hash value.
+ *
+ * @return `1` if `h1` and `h2` are equal.
+ * @return `0` if they are not.
+ */
+XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+
+/*!
+ * @brief Compares two @ref XXH128_hash_t
+ *
+ * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
+ *
+ * @param h128_1 Left-hand side value
+ * @param h128_2 Right-hand side value
+ *
+ * @return >0 if @p h128_1  > @p h128_2
+ * @return =0 if @p h128_1 == @p h128_2
+ * @return <0 if @p h128_1  < @p h128_2
+ */
+XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2);
+
+
+/*******   Canonical representation   *******/
+typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
+
+
+/*!
+ * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t.
+ *
+ * @param dst  The @ref XXH128_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH128_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t.
+ *
+ * @param src The @ref XXH128_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);
+
+
+#endif  /* !XXH_NO_XXH3 */
+#endif  /* XXH_NO_LONG_LONG */
+
+/*!
+ * @}
+ */
+#endif /* XXHASH_H_5627135585666179 */
+
+
+
+#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742)
+#define XXHASH_H_STATIC_13879238742
+/* ****************************************************************************
+ * This section contains declarations which are not guaranteed to remain stable.
+ * They may change in future versions, becoming incompatible with a different
+ * version of the library.
+ * These declarations should only be used with static linking.
+ * Never use them in association with dynamic linking!
+ ***************************************************************************** */
+
+/*
+ * These definitions are only present to allow static allocation
+ * of XXH states, on stack or in a struct, for example.
+ * Never **ever** access their members directly.
+ */
+
+/*!
+ * @internal
+ * @brief Structure for XXH32 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH32_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH64_state_s, XXH3_state_s
+ */
+struct XXH32_state_s {
+   XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */
+   XXH32_hash_t large_len;    /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */
+   XXH32_hash_t v[4];         /*!< Accumulator lanes */
+   XXH32_hash_t mem32[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */
+   XXH32_hash_t memsize;      /*!< Amount of data in @ref mem32 */
+   XXH32_hash_t reserved;     /*!< Reserved field. Do not read nor write to it. */
+};   /* typedef'd to XXH32_state_t */
+
+
+#ifndef XXH_NO_LONG_LONG  /* defined when there is no 64-bit support */
+
+/*!
+ * @internal
+ * @brief Structure for XXH64 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH64_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH32_state_s, XXH3_state_s
+ */
+struct XXH64_state_s {
+   XXH64_hash_t total_len;    /*!< Total length hashed. This is always 64-bit. */
+   XXH64_hash_t v[4];         /*!< Accumulator lanes */
+   XXH64_hash_t mem64[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */
+   XXH32_hash_t memsize;      /*!< Amount of data in @ref mem64 */
+   XXH32_hash_t reserved32;   /*!< Reserved field, needed for padding anyways*/
+   XXH64_hash_t reserved64;   /*!< Reserved field. Do not read or write to it. */
+};   /* typedef'd to XXH64_state_t */
+
+#ifndef XXH_NO_XXH3
+
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */
+#  define XXH_ALIGN(n)      _Alignas(n)
+#elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */
+/* In C++ alignas() is a keyword */
+#  define XXH_ALIGN(n)      alignas(n)
+#elif defined(__GNUC__)
+#  define XXH_ALIGN(n)      __attribute__ ((aligned(n)))
+#elif defined(_MSC_VER)
+#  define XXH_ALIGN(n)      __declspec(align(n))
+#else
+#  define XXH_ALIGN(n)   /* disabled */
+#endif
+
+/* Old GCC versions only accept the attribute after the type in structures. */
+#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L))   /* C11+ */ \
+    && ! (defined(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \
+    && defined(__GNUC__)
+#   define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
+#else
+#   define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
+#endif
+
+/*!
+ * @brief The size of the internal XXH3 buffer.
+ *
+ * This is the optimal update size for incremental hashing.
+ *
+ * @see XXH3_64b_update(), XXH3_128b_update().
+ */
+#define XXH3_INTERNALBUFFER_SIZE 256
+
+/*!
+ * @internal
+ * @brief Default size of the secret buffer (and @ref XXH3_kSecret).
+ *
+ * This is the size used in @ref XXH3_kSecret and the seeded functions.
+ *
+ * Not to be confused with @ref XXH3_SECRET_SIZE_MIN.
+ */
+#define XXH3_SECRET_DEFAULT_SIZE 192
+
+/*!
+ * @internal
+ * @brief Structure for XXH3 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined.
+ * Otherwise it is an opaque type.
+ * Never use this definition in combination with dynamic library.
+ * This allows fields to safely be changed in the future.
+ *
+ * @note ** This structure has a strict alignment requirement of 64 bytes!! **
+ * Do not allocate this with `malloc()` or `new`,
+ * it will not be sufficiently aligned.
+ * Use @ref XXH3_createState() and @ref XXH3_freeState(), or stack allocation.
+ *
+ * Typedef'd to @ref XXH3_state_t.
+ * Do never access the members of this struct directly.
+ *
+ * @see XXH3_INITSTATE() for stack initialization.
+ * @see XXH3_createState(), XXH3_freeState().
+ * @see XXH32_state_s, XXH64_state_s
+ */
+struct XXH3_state_s {
+   XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
+       /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */
+   XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
+       /*!< Used to store a custom secret generated from a seed. */
+   XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
+       /*!< The internal buffer. @see XXH32_state_s::mem32 */
+   XXH32_hash_t bufferedSize;
+       /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */
+   XXH32_hash_t useSeed;
+       /*!< Reserved field. Needed for padding on 64-bit. */
+   size_t nbStripesSoFar;
+       /*!< Number or stripes processed. */
+   XXH64_hash_t totalLen;
+       /*!< Total length hashed. 64-bit even on 32-bit targets. */
+   size_t nbStripesPerBlock;
+       /*!< Number of stripes per block. */
+   size_t secretLimit;
+       /*!< Size of @ref customSecret or @ref extSecret */
+   XXH64_hash_t seed;
+       /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */
+   XXH64_hash_t reserved64;
+       /*!< Reserved field. */
+   const unsigned char* extSecret;
+       /*!< Reference to an external secret for the _withSecret variants, NULL
+        *   for other variants. */
+   /* note: there may be some padding at the end due to alignment on 64 bytes */
+}; /* typedef'd to XXH3_state_t */
+
+#undef XXH_ALIGN_MEMBER
+
+/*!
+ * @brief Initializes a stack-allocated `XXH3_state_s`.
+ *
+ * When the @ref XXH3_state_t structure is merely emplaced on stack,
+ * it should be initialized with XXH3_INITSTATE() or a memset()
+ * in case its first reset uses XXH3_NNbits_reset_withSeed().
+ * This init can be omitted if the first reset uses default or _withSecret mode.
+ * This operation isn't necessary when the state is created with XXH3_createState().
+ * Note that this doesn't prepare the state for a streaming operation,
+ * it's still necessary to use XXH3_NNbits_reset*() afterwards.
+ */
+#define XXH3_INITSTATE(XXH3_state_ptr)                       \
+    do {                                                     \
+        XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \
+        tmp_xxh3_state_ptr->seed = 0;                        \
+        tmp_xxh3_state_ptr->extSecret = NULL;                \
+    } while(0)
+
+
+/*!
+ * @brief Calculates the 128-bit hash of @p data using XXH3.
+ *
+ * @param data The block of data to be hashed, at least @p len bytes in size.
+ * @param len  The length of @p data, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p data and @p data + @p len must be valid,
+ *   readable, contiguous memory. However, if @p len is `0`, @p data may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 128-bit XXH3 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+
+
+/* ===   Experimental API   === */
+/* Symbols defined below must be considered tied to a specific library version. */
+
+/*!
+ * @brief Derive a high-entropy secret from any user-defined content, named customSeed.
+ *
+ * @param secretBuffer    A writable buffer for derived high-entropy secret data.
+ * @param secretSize      Size of secretBuffer, in bytes.  Must be >= XXH3_SECRET_SIZE_MIN.
+ * @param customSeed      A user-defined content.
+ * @param customSeedSize  Size of customSeed, in bytes.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * The generated secret can be used in combination with `*_withSecret()` functions.
+ * The `_withSecret()` variants are useful to provide a higher level of protection
+ * than 64-bit seed, as it becomes much more difficult for an external actor to
+ * guess how to impact the calculation logic.
+ *
+ * The function accepts as input a custom seed of any length and any content,
+ * and derives from it a high-entropy secret of length @p secretSize into an
+ * already allocated buffer @p secretBuffer.
+ *
+ * The generated secret can then be used with any `*_withSecret()` variant.
+ * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(),
+ * @ref XXH3_128bits_reset_withSecret() and @ref XXH3_64bits_reset_withSecret()
+ * are part of this list. They all accept a `secret` parameter
+ * which must be large enough for implementation reasons (>= @ref XXH3_SECRET_SIZE_MIN)
+ * _and_ feature very high entropy (consist of random-looking bytes).
+ * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can
+ * be employed to ensure proper quality.
+ *
+ * @p customSeed can be anything. It can have any size, even small ones,
+ * and its content can be anything, even "poor entropy" sources such as a bunch
+ * of zeroes. The resulting `secret` will nonetheless provide all required qualities.
+ *
+ * @pre
+ *   - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN
+ *   - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
+ *
+ * Example code:
+ * @code{.c}
+ *    #include <stdio.h>
+ *    #include <stdlib.h>
+ *    #include <string.h>
+ *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ *    #include "xxhash.h"
+ *    // Hashes argv[2] using the entropy from argv[1].
+ *    int main(int argc, char* argv[])
+ *    {
+ *        char secret[XXH3_SECRET_SIZE_MIN];
+ *        if (argv != 3) { return 1; }
+ *        XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1]));
+ *        XXH64_hash_t h = XXH3_64bits_withSecret(
+ *             argv[2], strlen(argv[2]),
+ *             secret, sizeof(secret)
+ *        );
+ *        printf("%016llx\n", (unsigned long long) h);
+ *    }
+ * @endcode
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize);
+
+/*!
+ * @brief Generate the same secret as the _withSeed() variants.
+ *
+ * @param secretBuffer A writable buffer of @ref XXH3_SECRET_DEFAULT_SIZE bytes
+ * @param seed         The 64-bit seed to alter the hash result predictably.
+ *
+ * The generated secret can be used in combination with
+ *`*_withSecret()` and `_withSecretandSeed()` variants.
+ *
+ * Example C++ `std::string` hash class:
+ * @code{.cpp}
+ *    #include <string>
+ *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ *    #include "xxhash.h"
+ *    // Slow, seeds each time
+ *    class HashSlow {
+ *        XXH64_hash_t seed;
+ *    public:
+ *        HashSlow(XXH64_hash_t s) : seed{s} {}
+ *        size_t operator()(const std::string& x) const {
+ *            return size_t{XXH3_64bits_withSeed(x.c_str(), x.length(), seed)};
+ *        }
+ *    };
+ *    // Fast, caches the seeded secret for future uses.
+ *    class HashFast {
+ *        unsigned char secret[XXH3_SECRET_DEFAULT_SIZE];
+ *    public:
+ *        HashFast(XXH64_hash_t s) {
+ *            XXH3_generateSecret_fromSeed(secret, seed);
+ *        }
+ *        size_t operator()(const std::string& x) const {
+ *            return size_t{
+ *                XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret))
+ *            };
+ *        }
+ *    };
+ * @endcode
+ */
+XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed);
+
+/*!
+ * @brief Maximum size of "short" key in bytes.
+ */
+#define XXH3_MIDSIZE_MAX 240
+
+/*!
+ * @brief Calculates 64/128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed       The 64-bit seed to alter the hash result predictably.
+ *
+ * These variants generate hash values using either:
+ * - @p seed for "short" keys (< @ref XXH3_MIDSIZE_MAX = 240 bytes)
+ * - @p secret for "large" keys (>= @ref XXH3_MIDSIZE_MAX).
+ *
+ * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`.
+ * `_withSeed()` has to generate the secret on the fly for "large" keys.
+ * It's fast, but can be perceptible for "not so large" keys (< 1 KB).
+ * `_withSecret()` has to generate the masks on the fly for "small" keys,
+ * which requires more instructions than _withSeed() variants.
+ * Therefore, _withSecretandSeed variant combines the best of both worlds.
+ *
+ * When @p secret has been generated by XXH3_generateSecret_fromSeed(),
+ * this variant produces *exactly* the same results as `_withSeed()` variant,
+ * hence offering only a pure speed benefit on "large" input,
+ * by skipping the need to regenerate the secret for every large input.
+ *
+ * Another usage scenario is to hash the secret to a 64-bit hash value,
+ * for example with XXH3_64bits(), which then becomes the seed,
+ * and then employ both the seed and the secret in _withSecretandSeed().
+ * On top of speed, an added benefit is that each bit in the secret
+ * has a 50% chance to swap each bit in the output, via its impact to the seed.
+ *
+ * This is not guaranteed when using the secret directly in "small data" scenarios,
+ * because only portions of the secret are employed for small data.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* data, size_t len,
+                              XXH_NOESCAPE const void* secret, size_t secretSize,
+                              XXH64_hash_t seed);
+
+/*!
+ * @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data       The memory segment to be hashed, at least @p len bytes in size.
+ * @param length     The length of @p data, in bytes.
+ * @param secret     The secret used to alter hash result predictably.
+ * @param secretSize The length of @p secret, in bytes (must be >= XXH3_SECRET_SIZE_MIN)
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed(): contract is the same.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length,
+                               XXH_NOESCAPE const void* secret, size_t secretSize,
+                               XXH64_hash_t seed64);
+
+#ifndef XXH_NO_STREAM
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed(). Contract is identical.
+ */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+                                    XXH_NOESCAPE const void* secret, size_t secretSize,
+                                    XXH64_hash_t seed64);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed(). Contract is identical.
+ *
+ * Note: there was a bug in an earlier version of this function (<= v0.8.2)
+ * that would make it generate an incorrect hash value
+ * when @p seed == 0 and @p length < XXH3_MIDSIZE_MAX
+ * and @p secret is different from XXH3_generateSecret_fromSeed().
+ * As stated in the contract, the correct hash result must be
+ * the same as XXH3_128bits_withSeed() when @p length <= XXH3_MIDSIZE_MAX.
+ * Results generated by this older version are wrong, hence not comparable.
+ */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+                                     XXH_NOESCAPE const void* secret, size_t secretSize,
+                                     XXH64_hash_t seed64);
+
+#endif /* !XXH_NO_STREAM */
+
+#endif  /* !XXH_NO_XXH3 */
+#endif  /* XXH_NO_LONG_LONG */
+#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
+#  define XXH_IMPLEMENTATION
+#endif
+
+#endif  /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */
+
+
+/* ======================================================================== */
+/* ======================================================================== */
+/* ======================================================================== */
+
+
+/*-**********************************************************************
+ * xxHash implementation
+ *-**********************************************************************
+ * xxHash's implementation used to be hosted inside xxhash.c.
+ *
+ * However, inlining requires implementation to be visible to the compiler,
+ * hence be included alongside the header.
+ * Previously, implementation was hosted inside xxhash.c,
+ * which was then #included when inlining was activated.
+ * This construction created issues with a few build and install systems,
+ * as it required xxhash.c to be stored in /include directory.
+ *
+ * xxHash implementation is now directly integrated within xxhash.h.
+ * As a consequence, xxhash.c is no longer needed in /include.
+ *
+ * xxhash.c is still available and is still useful.
+ * In a "normal" setup, when xxhash is not inlined,
+ * xxhash.h only exposes the prototypes and public symbols,
+ * while xxhash.c can be built into an object file xxhash.o
+ * which can then be linked into the final binary.
+ ************************************************************************/
+
+#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
+   || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
+#  define XXH_IMPLEM_13a8737387
+
+/* *************************************
+*  Tuning parameters
+***************************************/
+
+/*!
+ * @defgroup tuning Tuning parameters
+ * @{
+ *
+ * Various macros to control xxHash's behavior.
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Define this to disable 64-bit code.
+ *
+ * Useful if only using the @ref XXH32_family and you have a strict C90 compiler.
+ */
+#  define XXH_NO_LONG_LONG
+#  undef XXH_NO_LONG_LONG /* don't actually */
+/*!
+ * @brief Controls how unaligned memory is accessed.
+ *
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is
+ * safe and portable.
+ *
+ * Unfortunately, on some target/compiler combinations, the generated assembly
+ * is sub-optimal.
+ *
+ * The below switch allow selection of a different access method
+ * in the search for improved performance.
+ *
+ * @par Possible options:
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy`
+ *   @par
+ *     Use `memcpy()`. Safe and portable. Note that most modern compilers will
+ *     eliminate the function call and treat it as an unaligned access.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))`
+ *   @par
+ *     Depends on compiler extensions and is therefore not portable.
+ *     This method is safe _if_ your compiler supports it,
+ *     and *generally* as fast or faster than `memcpy`.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast
+ *  @par
+ *     Casts directly and dereferences. This method doesn't depend on the
+ *     compiler, but it violates the C standard as it directly dereferences an
+ *     unaligned pointer. It can generate buggy code on targets which do not
+ *     support unaligned memory accesses, but in some circumstances, it's the
+ *     only known way to get the most performance.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift
+ *  @par
+ *     Also portable. This can generate the best code on old compilers which don't
+ *     inline small `memcpy()` calls, and it might also be faster on big-endian
+ *     systems which lack a native byteswap instruction. However, some compilers
+ *     will emit literal byteshifts even if the target supports unaligned access.
+ *
+ *
+ * @warning
+ *   Methods 1 and 2 rely on implementation-defined behavior. Use these with
+ *   care, as what works on one compiler/platform/optimization level may cause
+ *   another to read garbage data or even crash.
+ *
+ * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
+ *
+ * Prefer these methods in priority order (0 > 3 > 1 > 2)
+ */
+#  define XXH_FORCE_MEMORY_ACCESS 0
+
+/*!
+ * @def XXH_SIZE_OPT
+ * @brief Controls how much xxHash optimizes for size.
+ *
+ * xxHash, when compiled, tends to result in a rather large binary size. This
+ * is mostly due to heavy usage to forced inlining and constant folding of the
+ * @ref XXH3_family to increase performance.
+ *
+ * However, some developers prefer size over speed. This option can
+ * significantly reduce the size of the generated code. When using the `-Os`
+ * or `-Oz` options on GCC or Clang, this is defined to 1 by default,
+ * otherwise it is defined to 0.
+ *
+ * Most of these size optimizations can be controlled manually.
+ *
+ * This is a number from 0-2.
+ *  - `XXH_SIZE_OPT` == 0: Default. xxHash makes no size optimizations. Speed
+ *    comes first.
+ *  - `XXH_SIZE_OPT` == 1: Default for `-Os` and `-Oz`. xxHash is more
+ *    conservative and disables hacks that increase code size. It implies the
+ *    options @ref XXH_NO_INLINE_HINTS == 1, @ref XXH_FORCE_ALIGN_CHECK == 0,
+ *    and @ref XXH3_NEON_LANES == 8 if they are not already defined.
+ *  - `XXH_SIZE_OPT` == 2: xxHash tries to make itself as small as possible.
+ *    Performance may cry. For example, the single shot functions just use the
+ *    streaming API.
+ */
+#  define XXH_SIZE_OPT 0
+
+/*!
+ * @def XXH_FORCE_ALIGN_CHECK
+ * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32()
+ * and XXH64() only).
+ *
+ * This is an important performance trick for architectures without decent
+ * unaligned memory access performance.
+ *
+ * It checks for input alignment, and when conditions are met, uses a "fast
+ * path" employing direct 32-bit/64-bit reads, resulting in _dramatically
+ * faster_ read speed.
+ *
+ * The check costs one initial branch per hash, which is generally negligible,
+ * but not zero.
+ *
+ * Moreover, it's not useful to generate an additional code path if memory
+ * access uses the same instruction for both aligned and unaligned
+ * addresses (e.g. x86 and aarch64).
+ *
+ * In these cases, the alignment check can be removed by setting this macro to 0.
+ * Then the code will always use unaligned memory access.
+ * Align check is automatically disabled on x86, x64, ARM64, and some ARM chips
+ * which are platforms known to offer good unaligned memory accesses performance.
+ *
+ * It is also disabled by default when @ref XXH_SIZE_OPT >= 1.
+ *
+ * This option does not affect XXH3 (only XXH32 and XXH64).
+ */
+#  define XXH_FORCE_ALIGN_CHECK 0
+
+/*!
+ * @def XXH_NO_INLINE_HINTS
+ * @brief When non-zero, sets all functions to `static`.
+ *
+ * By default, xxHash tries to force the compiler to inline almost all internal
+ * functions.
+ *
+ * This can usually improve performance due to reduced jumping and improved
+ * constant folding, but significantly increases the size of the binary which
+ * might not be favorable.
+ *
+ * Additionally, sometimes the forced inlining can be detrimental to performance,
+ * depending on the architecture.
+ *
+ * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
+ * compiler full control on whether to inline or not.
+ *
+ * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if
+ * @ref XXH_SIZE_OPT >= 1, this will automatically be defined.
+ */
+#  define XXH_NO_INLINE_HINTS 0
+
+/*!
+ * @def XXH3_INLINE_SECRET
+ * @brief Determines whether to inline the XXH3 withSecret code.
+ *
+ * When the secret size is known, the compiler can improve the performance
+ * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret().
+ *
+ * However, if the secret size is not known, it doesn't have any benefit. This
+ * happens when xxHash is compiled into a global symbol. Therefore, if
+ * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0.
+ *
+ * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers
+ * that are *sometimes* force inline on -Og, and it is impossible to automatically
+ * detect this optimization level.
+ */
+#  define XXH3_INLINE_SECRET 0
+
+/*!
+ * @def XXH32_ENDJMP
+ * @brief Whether to use a jump for `XXH32_finalize`.
+ *
+ * For performance, `XXH32_finalize` uses multiple branches in the finalizer.
+ * This is generally preferable for performance,
+ * but depending on exact architecture, a jmp may be preferable.
+ *
+ * This setting is only possibly making a difference for very small inputs.
+ */
+#  define XXH32_ENDJMP 0
+
+/*!
+ * @internal
+ * @brief Redefines old internal names.
+ *
+ * For compatibility with code that uses xxHash's internals before the names
+ * were changed to improve namespacing. There is no other reason to use this.
+ */
+#  define XXH_OLD_NAMES
+#  undef XXH_OLD_NAMES /* don't actually use, it is ugly. */
+
+/*!
+ * @def XXH_NO_STREAM
+ * @brief Disables the streaming API.
+ *
+ * When xxHash is not inlined and the streaming functions are not used, disabling
+ * the streaming functions can improve code size significantly, especially with
+ * the @ref XXH3_family which tends to make constant folded copies of itself.
+ */
+#  define XXH_NO_STREAM
+#  undef XXH_NO_STREAM /* don't actually */
+#endif /* XXH_DOXYGEN */
+/*!
+ * @}
+ */
+
+#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+   /* prefer __packed__ structures (method 1) for GCC
+    * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy
+    * which for some reason does unaligned loads. */
+#  if defined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED))
+#    define XXH_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+#ifndef XXH_SIZE_OPT
+   /* default to 1 for -Os or -Oz */
+#  if (defined(__GNUC__) || defined(__clang__)) && defined(__OPTIMIZE_SIZE__)
+#    define XXH_SIZE_OPT 1
+#  else
+#    define XXH_SIZE_OPT 0
+#  endif
+#endif
+
+#ifndef XXH_FORCE_ALIGN_CHECK  /* can be defined externally */
+   /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */
+#  if XXH_SIZE_OPT >= 1 || \
+      defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \
+   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64)    || defined(_M_ARM) /* visual */
+#    define XXH_FORCE_ALIGN_CHECK 0
+#  else
+#    define XXH_FORCE_ALIGN_CHECK 1
+#  endif
+#endif
+
+#ifndef XXH_NO_INLINE_HINTS
+#  if XXH_SIZE_OPT >= 1 || defined(__NO_INLINE__)  /* -O0, -fno-inline */
+#    define XXH_NO_INLINE_HINTS 1
+#  else
+#    define XXH_NO_INLINE_HINTS 0
+#  endif
+#endif
+
+#ifndef XXH3_INLINE_SECRET
+#  if (defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12) \
+     || !defined(XXH_INLINE_ALL)
+#    define XXH3_INLINE_SECRET 0
+#  else
+#    define XXH3_INLINE_SECRET 1
+#  endif
+#endif
+
+#ifndef XXH32_ENDJMP
+/* generally preferable for performance */
+#  define XXH32_ENDJMP 0
+#endif
+
+/*!
+ * @defgroup impl Implementation
+ * @{
+ */
+
+
+/* *************************************
+*  Includes & Memory related functions
+***************************************/
+#if defined(XXH_NO_STREAM)
+/* nothing */
+#elif defined(XXH_NO_STDLIB)
+
+/* When requesting to disable any mention of stdlib,
+ * the library loses the ability to invoked malloc / free.
+ * In practice, it means that functions like `XXH*_createState()`
+ * will always fail, and return NULL.
+ * This flag is useful in situations where
+ * xxhash.h is integrated into some kernel, embedded or limited environment
+ * without access to dynamic allocation.
+ */
+
+static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; }
+static void XXH_free(void* p) { (void)p; }
+
+#else
+
+/*
+ * Modify the local functions below should you wish to use
+ * different memory routines for malloc() and free()
+ */
+#include <stdlib.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than malloc().
+ */
+static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than free().
+ */
+static void XXH_free(void* p) { free(p); }
+
+#endif  /* XXH_NO_STDLIB */
+
+#include <string.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than memcpy().
+ */
+static void* XXH_memcpy(void* dest, const void* src, size_t size)
+{
+    return memcpy(dest,src,size);
+}
+
+#include <limits.h>   /* ULLONG_MAX */
+
+
+/* *************************************
+*  Compiler Specific Options
+***************************************/
+#ifdef _MSC_VER /* Visual Studio warning fix */
+#  pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
+#endif
+
+#if XXH_NO_INLINE_HINTS  /* disable inlining hints */
+#  if defined(__GNUC__) || defined(__clang__)
+#    define XXH_FORCE_INLINE static __attribute__((__unused__))
+#  else
+#    define XXH_FORCE_INLINE static
+#  endif
+#  define XXH_NO_INLINE static
+/* enable inlining hints */
+#elif defined(__GNUC__) || defined(__clang__)
+#  define XXH_FORCE_INLINE static __inline__ __attribute__((__always_inline__, __unused__))
+#  define XXH_NO_INLINE static __attribute__((__noinline__))
+#elif defined(_MSC_VER)  /* Visual Studio */
+#  define XXH_FORCE_INLINE static __forceinline
+#  define XXH_NO_INLINE static __declspec(noinline)
+#elif defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))   /* C99 */
+#  define XXH_FORCE_INLINE static inline
+#  define XXH_NO_INLINE static
+#else
+#  define XXH_FORCE_INLINE static
+#  define XXH_NO_INLINE static
+#endif
+
+#if XXH3_INLINE_SECRET
+#  define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE
+#else
+#  define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE
+#endif
+
+
+/* *************************************
+*  Debug
+***************************************/
+/*!
+ * @ingroup tuning
+ * @def XXH_DEBUGLEVEL
+ * @brief Sets the debugging level.
+ *
+ * XXH_DEBUGLEVEL is expected to be defined externally, typically via the
+ * compiler's command line options. The value must be a number.
+ */
+#ifndef XXH_DEBUGLEVEL
+#  ifdef DEBUGLEVEL /* backwards compat */
+#    define XXH_DEBUGLEVEL DEBUGLEVEL
+#  else
+#    define XXH_DEBUGLEVEL 0
+#  endif
+#endif
+
+#if (XXH_DEBUGLEVEL>=1)
+#  include <assert.h>   /* note: can still be disabled with NDEBUG */
+#  define XXH_ASSERT(c)   assert(c)
+#else
+#  if defined(__INTEL_COMPILER)
+#    define XXH_ASSERT(c)   XXH_ASSUME((unsigned char) (c))
+#  else
+#    define XXH_ASSERT(c)   XXH_ASSUME(c)
+#  endif
+#endif
+
+/* note: use after variable declarations */
+#ifndef XXH_STATIC_ASSERT
+#  if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)    /* C11 */
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0)
+#  elif defined(__cplusplus) && (__cplusplus >= 201103L)            /* C++11 */
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0)
+#  else
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { struct xxh_sa { char x[(c) ? 1 : -1]; }; } while(0)
+#  endif
+#  define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c),#c)
+#endif
+
+/*!
+ * @internal
+ * @def XXH_COMPILER_GUARD(var)
+ * @brief Used to prevent unwanted optimizations for @p var.
+ *
+ * It uses an empty GCC inline assembly statement with a register constraint
+ * which forces @p var into a general purpose register (eg eax, ebx, ecx
+ * on x86) and marks it as modified.
+ *
+ * This is used in a few places to avoid unwanted autovectorization (e.g.
+ * XXH32_round()). All vectorization we want is explicit via intrinsics,
+ * and _usually_ isn't wanted elsewhere.
+ *
+ * We also use it to prevent unwanted constant folding for AArch64 in
+ * XXH3_initCustomSecret_scalar().
+ */
+#if defined(__GNUC__) || defined(__clang__)
+#  define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var))
+#else
+#  define XXH_COMPILER_GUARD(var) ((void)0)
+#endif
+
+/* Specifically for NEON vectors which use the "w" constraint, on
+ * Clang. */
+#if defined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__)
+#  define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var))
+#else
+#  define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0)
+#endif
+
+/* *************************************
+*  Basic Types
+***************************************/
+#if !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   ifdef _AIX
+#     include <inttypes.h>
+#   else
+#     include <stdint.h>
+#   endif
+    typedef uint8_t xxh_u8;
+#else
+    typedef unsigned char xxh_u8;
+#endif
+typedef XXH32_hash_t xxh_u32;
+
+#ifdef XXH_OLD_NAMES
+#  warning "XXH_OLD_NAMES is planned to be removed starting v0.9. If the program depends on it, consider moving away from it by employing newer type names directly"
+#  define BYTE xxh_u8
+#  define U8   xxh_u8
+#  define U32  xxh_u32
+#endif
+
+/* ***   Memory access   *** */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_read32(const void* ptr)
+ * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit native endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit little endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readBE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit big endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit big endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32_align(const void* ptr, XXH_alignment align)
+ * @brief Like @ref XXH_readLE32(), but has an option for aligned reads.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ * Note that when @ref XXH_FORCE_ALIGN_CHECK == 0, the @p align parameter is
+ * always @ref XXH_alignment::XXH_unaligned.
+ *
+ * @param ptr The pointer to read from.
+ * @param align Whether @p ptr is aligned.
+ * @pre
+ *   If @p align == @ref XXH_alignment::XXH_aligned, @p ptr must be 4 byte
+ *   aligned.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE32 and XXH_readBE32.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/*
+ * Force direct memory access. Only works on CPU which support unaligned memory
+ * access in hardware.
+ */
+static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; } __attribute__((__packed__)) unalign;
+#endif
+static xxh_u32 XXH_read32(const void* ptr)
+{
+    typedef __attribute__((__aligned__(1))) xxh_u32 xxh_unalign32;
+    return *((const xxh_unalign32*)ptr);
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ */
+static xxh_u32 XXH_read32(const void* memPtr)
+{
+    xxh_u32 val;
+    XXH_memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+
+/* ***   Endianness   *** */
+
+/*!
+ * @ingroup tuning
+ * @def XXH_CPU_LITTLE_ENDIAN
+ * @brief Whether the target is little endian.
+ *
+ * Defined to 1 if the target is little endian, or 0 if it is big endian.
+ * It can be defined externally, for example on the compiler command line.
+ *
+ * If it is not defined,
+ * a runtime check (which is usually constant folded) is used instead.
+ *
+ * @note
+ *   This is not necessarily defined to an integer constant.
+ *
+ * @see XXH_isLittleEndian() for the runtime check.
+ */
+#ifndef XXH_CPU_LITTLE_ENDIAN
+/*
+ * Try to detect endianness automatically, to avoid the nonstandard behavior
+ * in `XXH_isLittleEndian()`
+ */
+#  if defined(_WIN32) /* Windows is always little endian */ \
+     || defined(__LITTLE_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#    define XXH_CPU_LITTLE_ENDIAN 1
+#  elif defined(__BIG_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXH_CPU_LITTLE_ENDIAN 0
+#  else
+/*!
+ * @internal
+ * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN.
+ *
+ * Most compilers will constant fold this.
+ */
+static int XXH_isLittleEndian(void)
+{
+    /*
+     * Portable and well-defined behavior.
+     * Don't use static: it is detrimental to performance.
+     */
+    const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };
+    return one.c[0];
+}
+#   define XXH_CPU_LITTLE_ENDIAN   XXH_isLittleEndian()
+#  endif
+#endif
+
+
+
+
+/* ****************************************
+*  Compiler-specific Functions and Macros
+******************************************/
+#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+
+#ifdef __has_builtin
+#  define XXH_HAS_BUILTIN(x) __has_builtin(x)
+#else
+#  define XXH_HAS_BUILTIN(x) 0
+#endif
+
+
+
+/*
+ * C23 and future versions have standard "unreachable()".
+ * Once it has been implemented reliably we can add it as an
+ * additional case:
+ *
+ * ```
+ * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN)
+ * #  include <stddef.h>
+ * #  ifdef unreachable
+ * #    define XXH_UNREACHABLE() unreachable()
+ * #  endif
+ * #endif
+ * ```
+ *
+ * Note C++23 also has std::unreachable() which can be detected
+ * as follows:
+ * ```
+ * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L)
+ * #  include <utility>
+ * #  define XXH_UNREACHABLE() std::unreachable()
+ * #endif
+ * ```
+ * NB: `__cpp_lib_unreachable` is defined in the `<version>` header.
+ * We don't use that as including `<utility>` in `extern "C"` blocks
+ * doesn't work on GCC12
+ */
+
+#if XXH_HAS_BUILTIN(__builtin_unreachable)
+#  define XXH_UNREACHABLE() __builtin_unreachable()
+
+#elif defined(_MSC_VER)
+#  define XXH_UNREACHABLE() __assume(0)
+
+#else
+#  define XXH_UNREACHABLE()
+#endif
+
+#if XXH_HAS_BUILTIN(__builtin_assume)
+#  define XXH_ASSUME(c) __builtin_assume(c)
+#else
+#  define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); }
+#endif
+
+/*!
+ * @internal
+ * @def XXH_rotl32(x,r)
+ * @brief 32-bit rotate left.
+ *
+ * @param x The 32-bit integer to be rotated.
+ * @param r The number of bits to rotate.
+ * @pre
+ *   @p r > 0 && @p r < 32
+ * @note
+ *   @p x and @p r may be evaluated multiple times.
+ * @return The rotated result.
+ */
+#if !defined(NO_CLANG_BUILTIN) && XXH_HAS_BUILTIN(__builtin_rotateleft32) \
+                               && XXH_HAS_BUILTIN(__builtin_rotateleft64)
+#  define XXH_rotl32 __builtin_rotateleft32
+#  define XXH_rotl64 __builtin_rotateleft64
+/* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */
+#elif defined(_MSC_VER)
+#  define XXH_rotl32(x,r) _rotl(x,r)
+#  define XXH_rotl64(x,r) _rotl64(x,r)
+#else
+#  define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
+#  define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
+#endif
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_swap32(xxh_u32 x)
+ * @brief A 32-bit byteswap.
+ *
+ * @param x The 32-bit integer to byteswap.
+ * @return @p x, byteswapped.
+ */
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap32 _byteswap_ulong
+#elif XXH_GCC_VERSION >= 403
+#  define XXH_swap32 __builtin_bswap32
+#else
+static xxh_u32 XXH_swap32 (xxh_u32 x)
+{
+    return  ((x << 24) & 0xff000000 ) |
+            ((x <<  8) & 0x00ff0000 ) |
+            ((x >>  8) & 0x0000ff00 ) |
+            ((x >> 24) & 0x000000ff );
+}
+#endif
+
+
+/* ***************************
+*  Memory reads
+*****************************/
+
+/*!
+ * @internal
+ * @brief Enum to indicate whether a pointer is aligned.
+ */
+typedef enum {
+    XXH_aligned,  /*!< Aligned */
+    XXH_unaligned /*!< Possibly unaligned */
+} XXH_alignment;
+
+/*
+ * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load.
+ *
+ * This is ideal for older compilers which don't inline memcpy.
+ */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[0]
+         | ((xxh_u32)bytePtr[1] << 8)
+         | ((xxh_u32)bytePtr[2] << 16)
+         | ((xxh_u32)bytePtr[3] << 24);
+}
+
+XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[3]
+         | ((xxh_u32)bytePtr[2] << 8)
+         | ((xxh_u32)bytePtr[1] << 16)
+         | ((xxh_u32)bytePtr[0] << 24);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
+}
+
+static xxh_u32 XXH_readBE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u32
+XXH_readLE32_align(const void* ptr, XXH_alignment align)
+{
+    if (align==XXH_unaligned) {
+        return XXH_readLE32(ptr);
+    } else {
+        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr);
+    }
+}
+
+
+/* *************************************
+*  Misc
+***************************************/
+/*! @ingroup public */
+XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
+
+
+/* *******************************************************************
+*  32-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @defgroup XXH32_impl XXH32 implementation
+ * @ingroup impl
+ *
+ * Details on the XXH32 implementation.
+ * @{
+ */
+ /* #define instead of static const, to be used as initializers */
+#define XXH_PRIME32_1  0x9E3779B1U  /*!< 0b10011110001101110111100110110001 */
+#define XXH_PRIME32_2  0x85EBCA77U  /*!< 0b10000101111010111100101001110111 */
+#define XXH_PRIME32_3  0xC2B2AE3DU  /*!< 0b11000010101100101010111000111101 */
+#define XXH_PRIME32_4  0x27D4EB2FU  /*!< 0b00100111110101001110101100101111 */
+#define XXH_PRIME32_5  0x165667B1U  /*!< 0b00010110010101100110011110110001 */
+
+#ifdef XXH_OLD_NAMES
+#  define PRIME32_1 XXH_PRIME32_1
+#  define PRIME32_2 XXH_PRIME32_2
+#  define PRIME32_3 XXH_PRIME32_3
+#  define PRIME32_4 XXH_PRIME32_4
+#  define PRIME32_5 XXH_PRIME32_5
+#endif
+
+/*!
+ * @internal
+ * @brief Normal stripe processing routine.
+ *
+ * This shuffles the bits so that any bit from @p input impacts several bits in
+ * @p acc.
+ *
+ * @param acc The accumulator lane.
+ * @param input The stripe of input to mix.
+ * @return The mixed accumulator lane.
+ */
+static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
+{
+    acc += input * XXH_PRIME32_2;
+    acc  = XXH_rotl32(acc, 13);
+    acc *= XXH_PRIME32_1;
+#if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    /*
+     * UGLY HACK:
+     * A compiler fence is used to prevent GCC and Clang from
+     * autovectorizing the XXH32 loop (pragmas and attributes don't work for some
+     * reason) without globally disabling SSE4.1.
+     *
+     * The reason we want to avoid vectorization is because despite working on
+     * 4 integers at a time, there are multiple factors slowing XXH32 down on
+     * SSE4:
+     * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on
+     *   newer chips!) making it slightly slower to multiply four integers at
+     *   once compared to four integers independently. Even when pmulld was
+     *   fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE
+     *   just to multiply unless doing a long operation.
+     *
+     * - Four instructions are required to rotate,
+     *      movqda tmp,  v // not required with VEX encoding
+     *      pslld  tmp, 13 // tmp <<= 13
+     *      psrld  v,   19 // x >>= 19
+     *      por    v,  tmp // x |= tmp
+     *   compared to one for scalar:
+     *      roll   v, 13    // reliably fast across the board
+     *      shldl  v, v, 13 // Sandy Bridge and later prefer this for some reason
+     *
+     * - Instruction level parallelism is actually more beneficial here because
+     *   the SIMD actually serializes this operation: While v1 is rotating, v2
+     *   can load data, while v3 can multiply. SSE forces them to operate
+     *   together.
+     *
+     * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing
+     * the loop. NEON is only faster on the A53, and with the newer cores, it is less
+     * than half the speed.
+     *
+     * Additionally, this is used on WASM SIMD128 because it JITs to the same
+     * SIMD instructions and has the same issue.
+     */
+    XXH_COMPILER_GUARD(acc);
+#endif
+    return acc;
+}
+
+/*!
+ * @internal
+ * @brief Mixes all bits to finalize the hash.
+ *
+ * The final mix ensures that all input bits have a chance to impact any bit in
+ * the output digest, resulting in an unbiased distribution.
+ *
+ * @param hash The hash to avalanche.
+ * @return The avalanched hash.
+ */
+static xxh_u32 XXH32_avalanche(xxh_u32 hash)
+{
+    hash ^= hash >> 15;
+    hash *= XXH_PRIME32_2;
+    hash ^= hash >> 13;
+    hash *= XXH_PRIME32_3;
+    hash ^= hash >> 16;
+    return hash;
+}
+
+#define XXH_get32bits(p) XXH_readLE32_align(p, align)
+
+/*!
+ * @internal
+ * @brief Processes the last 0-15 bytes of @p ptr.
+ *
+ * There may be up to 15 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param hash The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 16.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash.
+ * @see XXH64_finalize().
+ */
+static XXH_PUREF xxh_u32
+XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+#define XXH_PROCESS1 do {                             \
+    hash += (*ptr++) * XXH_PRIME32_5;                 \
+    hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1;      \
+} while (0)
+
+#define XXH_PROCESS4 do {                             \
+    hash += XXH_get32bits(ptr) * XXH_PRIME32_3;       \
+    ptr += 4;                                         \
+    hash  = XXH_rotl32(hash, 17) * XXH_PRIME32_4;     \
+} while (0)
+
+    if (ptr==NULL) XXH_ASSERT(len == 0);
+
+    /* Compact rerolled version; generally faster */
+    if (!XXH32_ENDJMP) {
+        len &= 15;
+        while (len >= 4) {
+            XXH_PROCESS4;
+            len -= 4;
+        }
+        while (len > 0) {
+            XXH_PROCESS1;
+            --len;
+        }
+        return XXH32_avalanche(hash);
+    } else {
+         switch(len&15) /* or switch(bEnd - p) */ {
+           case 12:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 8:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 4:       XXH_PROCESS4;
+                         return XXH32_avalanche(hash);
+
+           case 13:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 9:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 5:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(hash);
+
+           case 14:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 10:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 6:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(hash);
+
+           case 15:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 11:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 7:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 3:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 2:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 1:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 0:       return XXH32_avalanche(hash);
+        }
+        XXH_ASSERT(0);
+        return hash;   /* reaching this point is deemed impossible */
+    }
+}
+
+#ifdef XXH_OLD_NAMES
+#  define PROCESS1 XXH_PROCESS1
+#  define PROCESS4 XXH_PROCESS4
+#else
+#  undef XXH_PROCESS1
+#  undef XXH_PROCESS4
+#endif
+
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH32().
+ *
+ * @param input , len , seed Directly passed from @ref XXH32().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u32
+XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
+{
+    xxh_u32 h32;
+
+    if (input==NULL) XXH_ASSERT(len == 0);
+
+    if (len>=16) {
+        const xxh_u8* const bEnd = input + len;
+        const xxh_u8* const limit = bEnd - 15;
+        xxh_u32 v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+        xxh_u32 v2 = seed + XXH_PRIME32_2;
+        xxh_u32 v3 = seed + 0;
+        xxh_u32 v4 = seed - XXH_PRIME32_1;
+
+        do {
+            v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4;
+            v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4;
+            v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4;
+            v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4;
+        } while (input < limit);
+
+        h32 = XXH_rotl32(v1, 1)  + XXH_rotl32(v2, 7)
+            + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
+    } else {
+        h32  = seed + XXH_PRIME32_5;
+    }
+
+    h32 += (xxh_u32)len;
+
+    return XXH32_finalize(h32, input, len&15, align);
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
+{
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH32_state_t state;
+    XXH32_reset(&state, seed);
+    XXH32_update(&state, (const xxh_u8*)input, len);
+    return XXH32_digest(&state);
+#else
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
+            return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+    }   }
+
+    return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+#endif
+}
+
+
+
+/*******   Hash streaming   *******/
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
+{
+    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
+}
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
+{
+    XXH_memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
+{
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    statePtr->v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+    statePtr->v[1] = seed + XXH_PRIME32_2;
+    statePtr->v[2] = seed + 0;
+    statePtr->v[3] = seed - XXH_PRIME32_1;
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH32_update(XXH32_state_t* state, const void* input, size_t len)
+{
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
+
+    {   const xxh_u8* p = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = p + len;
+
+        state->total_len_32 += (XXH32_hash_t)len;
+        state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16));
+
+        if (state->memsize + len < 16)  {   /* fill in tmp buffer */
+            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len);
+            state->memsize += (XXH32_hash_t)len;
+            return XXH_OK;
+        }
+
+        if (state->memsize) {   /* some data left from previous update */
+            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize);
+            {   const xxh_u32* p32 = state->mem32;
+                state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p32)); p32++;
+                state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p32)); p32++;
+                state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p32)); p32++;
+                state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p32));
+            }
+            p += 16-state->memsize;
+            state->memsize = 0;
+        }
+
+        if (p <= bEnd-16) {
+            const xxh_u8* const limit = bEnd - 16;
+
+            do {
+                state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p)); p+=4;
+                state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p)); p+=4;
+                state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p)); p+=4;
+                state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p)); p+=4;
+            } while (p<=limit);
+
+        }
+
+        if (p < bEnd) {
+            XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
+{
+    xxh_u32 h32;
+
+    if (state->large_len) {
+        h32 = XXH_rotl32(state->v[0], 1)
+            + XXH_rotl32(state->v[1], 7)
+            + XXH_rotl32(state->v[2], 12)
+            + XXH_rotl32(state->v[3], 18);
+    } else {
+        h32 = state->v[2] /* == seed */ + XXH_PRIME32_5;
+    }
+
+    h32 += state->total_len_32;
+
+    return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
+}
+#endif /* !XXH_NO_STREAM */
+
+/*******   Canonical representation   *******/
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
+    XXH_memcpy(dst, &hash, sizeof(*dst));
+}
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
+{
+    return XXH_readBE32(src);
+}
+
+
+#ifndef XXH_NO_LONG_LONG
+
+/* *******************************************************************
+*  64-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @ingroup impl
+ * @{
+ */
+/*******   Memory access   *******/
+
+typedef XXH64_hash_t xxh_u64;
+
+#ifdef XXH_OLD_NAMES
+#  define U64 xxh_u64
+#endif
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE64 and XXH_readBE64.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    return *(const xxh_u64*) memPtr;
+}
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((__packed__)) unalign64;
+#endif
+static xxh_u64 XXH_read64(const void* ptr)
+{
+    typedef __attribute__((__aligned__(1))) xxh_u64 xxh_unalign64;
+    return *((const xxh_unalign64*)ptr);
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    xxh_u64 val;
+    XXH_memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap64 _byteswap_uint64
+#elif XXH_GCC_VERSION >= 403
+#  define XXH_swap64 __builtin_bswap64
+#else
+static xxh_u64 XXH_swap64(xxh_u64 x)
+{
+    return  ((x << 56) & 0xff00000000000000ULL) |
+            ((x << 40) & 0x00ff000000000000ULL) |
+            ((x << 24) & 0x0000ff0000000000ULL) |
+            ((x << 8)  & 0x000000ff00000000ULL) |
+            ((x >> 8)  & 0x00000000ff000000ULL) |
+            ((x >> 24) & 0x0000000000ff0000ULL) |
+            ((x >> 40) & 0x000000000000ff00ULL) |
+            ((x >> 56) & 0x00000000000000ffULL);
+}
+#endif
+
+
+/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[0]
+         | ((xxh_u64)bytePtr[1] << 8)
+         | ((xxh_u64)bytePtr[2] << 16)
+         | ((xxh_u64)bytePtr[3] << 24)
+         | ((xxh_u64)bytePtr[4] << 32)
+         | ((xxh_u64)bytePtr[5] << 40)
+         | ((xxh_u64)bytePtr[6] << 48)
+         | ((xxh_u64)bytePtr[7] << 56);
+}
+
+XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[7]
+         | ((xxh_u64)bytePtr[6] << 8)
+         | ((xxh_u64)bytePtr[5] << 16)
+         | ((xxh_u64)bytePtr[4] << 24)
+         | ((xxh_u64)bytePtr[3] << 32)
+         | ((xxh_u64)bytePtr[2] << 40)
+         | ((xxh_u64)bytePtr[1] << 48)
+         | ((xxh_u64)bytePtr[0] << 56);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
+}
+
+static xxh_u64 XXH_readBE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u64
+XXH_readLE64_align(const void* ptr, XXH_alignment align)
+{
+    if (align==XXH_unaligned)
+        return XXH_readLE64(ptr);
+    else
+        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr);
+}
+
+
+/*******   xxh64   *******/
+/*!
+ * @}
+ * @defgroup XXH64_impl XXH64 implementation
+ * @ingroup impl
+ *
+ * Details on the XXH64 implementation.
+ * @{
+ */
+/* #define rather that static const, to be used as initializers */
+#define XXH_PRIME64_1  0x9E3779B185EBCA87ULL  /*!< 0b1001111000110111011110011011000110000101111010111100101010000111 */
+#define XXH_PRIME64_2  0xC2B2AE3D27D4EB4FULL  /*!< 0b1100001010110010101011100011110100100111110101001110101101001111 */
+#define XXH_PRIME64_3  0x165667B19E3779F9ULL  /*!< 0b0001011001010110011001111011000110011110001101110111100111111001 */
+#define XXH_PRIME64_4  0x85EBCA77C2B2AE63ULL  /*!< 0b1000010111101011110010100111011111000010101100101010111001100011 */
+#define XXH_PRIME64_5  0x27D4EB2F165667C5ULL  /*!< 0b0010011111010100111010110010111100010110010101100110011111000101 */
+
+#ifdef XXH_OLD_NAMES
+#  define PRIME64_1 XXH_PRIME64_1
+#  define PRIME64_2 XXH_PRIME64_2
+#  define PRIME64_3 XXH_PRIME64_3
+#  define PRIME64_4 XXH_PRIME64_4
+#  define PRIME64_5 XXH_PRIME64_5
+#endif
+
+/*! @copydoc XXH32_round */
+static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
+{
+    acc += input * XXH_PRIME64_2;
+    acc  = XXH_rotl64(acc, 31);
+    acc *= XXH_PRIME64_1;
+#if (defined(__AVX512F__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    /*
+     * DISABLE AUTOVECTORIZATION:
+     * A compiler fence is used to prevent GCC and Clang from
+     * autovectorizing the XXH64 loop (pragmas and attributes don't work for some
+     * reason) without globally disabling AVX512.
+     *
+     * Autovectorization of XXH64 tends to be detrimental,
+     * though the exact outcome may change depending on exact cpu and compiler version.
+     * For information, it has been reported as detrimental for Skylake-X,
+     * but possibly beneficial for Zen4.
+     *
+     * The default is to disable auto-vectorization,
+     * but you can select to enable it instead using `XXH_ENABLE_AUTOVECTORIZE` build variable.
+     */
+    XXH_COMPILER_GUARD(acc);
+#endif
+    return acc;
+}
+
+static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
+{
+    val  = XXH64_round(0, val);
+    acc ^= val;
+    acc  = acc * XXH_PRIME64_1 + XXH_PRIME64_4;
+    return acc;
+}
+
+/*! @copydoc XXH32_avalanche */
+static xxh_u64 XXH64_avalanche(xxh_u64 hash)
+{
+    hash ^= hash >> 33;
+    hash *= XXH_PRIME64_2;
+    hash ^= hash >> 29;
+    hash *= XXH_PRIME64_3;
+    hash ^= hash >> 32;
+    return hash;
+}
+
+
+#define XXH_get64bits(p) XXH_readLE64_align(p, align)
+
+/*!
+ * @internal
+ * @brief Processes the last 0-31 bytes of @p ptr.
+ *
+ * There may be up to 31 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param hash The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 32.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash
+ * @see XXH32_finalize().
+ */
+static XXH_PUREF xxh_u64
+XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+    if (ptr==NULL) XXH_ASSERT(len == 0);
+    len &= 31;
+    while (len >= 8) {
+        xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
+        ptr += 8;
+        hash ^= k1;
+        hash  = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
+        len -= 8;
+    }
+    if (len >= 4) {
+        hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
+        ptr += 4;
+        hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
+        len -= 4;
+    }
+    while (len > 0) {
+        hash ^= (*ptr++) * XXH_PRIME64_5;
+        hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
+        --len;
+    }
+    return  XXH64_avalanche(hash);
+}
+
+#ifdef XXH_OLD_NAMES
+#  define PROCESS1_64 XXH_PROCESS1_64
+#  define PROCESS4_64 XXH_PROCESS4_64
+#  define PROCESS8_64 XXH_PROCESS8_64
+#else
+#  undef XXH_PROCESS1_64
+#  undef XXH_PROCESS4_64
+#  undef XXH_PROCESS8_64
+#endif
+
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH64().
+ *
+ * @param input , len , seed Directly passed from @ref XXH64().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u64
+XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
+{
+    xxh_u64 h64;
+    if (input==NULL) XXH_ASSERT(len == 0);
+
+    if (len>=32) {
+        const xxh_u8* const bEnd = input + len;
+        const xxh_u8* const limit = bEnd - 31;
+        xxh_u64 v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+        xxh_u64 v2 = seed + XXH_PRIME64_2;
+        xxh_u64 v3 = seed + 0;
+        xxh_u64 v4 = seed - XXH_PRIME64_1;
+
+        do {
+            v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8;
+            v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8;
+            v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8;
+            v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8;
+        } while (input<limit);
+
+        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
+        h64 = XXH64_mergeRound(h64, v1);
+        h64 = XXH64_mergeRound(h64, v2);
+        h64 = XXH64_mergeRound(h64, v3);
+        h64 = XXH64_mergeRound(h64, v4);
+
+    } else {
+        h64  = seed + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) len;
+
+    return XXH64_finalize(h64, input, len, align);
+}
+
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64 (XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH64_state_t state;
+    XXH64_reset(&state, seed);
+    XXH64_update(&state, (const xxh_u8*)input, len);
+    return XXH64_digest(&state);
+#else
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
+            return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+    }   }
+
+    return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+
+#endif
+}
+
+/*******   Hash Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH64_family*/
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
+{
+    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
+}
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState)
+{
+    XXH_memcpy(dstState, srcState, sizeof(*dstState));
+}
 
-/*******   Canonical representation   *******/
-typedef struct { unsigned char digest[8]; } XXH64_canonical_t;
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash);
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src);
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed)
+{
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    statePtr->v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+    statePtr->v[1] = seed + XXH_PRIME64_2;
+    statePtr->v[2] = seed + 0;
+    statePtr->v[3] = seed - XXH_PRIME64_1;
+    return XXH_OK;
+}
 
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
 
-#endif  /* XXH_NO_LONG_LONG */
+    {   const xxh_u8* p = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = p + len;
 
-#endif /* XXHASH_H_5627135585666179 */
+        state->total_len += len;
 
+        if (state->memsize + len < 32) {  /* fill in tmp buffer */
+            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len);
+            state->memsize += (xxh_u32)len;
+            return XXH_OK;
+        }
 
+        if (state->memsize) {   /* tmp buffer is full */
+            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize);
+            state->v[0] = XXH64_round(state->v[0], XXH_readLE64(state->mem64+0));
+            state->v[1] = XXH64_round(state->v[1], XXH_readLE64(state->mem64+1));
+            state->v[2] = XXH64_round(state->v[2], XXH_readLE64(state->mem64+2));
+            state->v[3] = XXH64_round(state->v[3], XXH_readLE64(state->mem64+3));
+            p += 32 - state->memsize;
+            state->memsize = 0;
+        }
 
-#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742)
-#define XXHASH_H_STATIC_13879238742
-/* ****************************************************************************
- * This section contains declarations which are not guaranteed to remain stable.
- * They may change in future versions, becoming incompatible with a different
- * version of the library.
- * These declarations should only be used with static linking.
- * Never use them in association with dynamic linking!
- ***************************************************************************** */
+        if (p+32 <= bEnd) {
+            const xxh_u8* const limit = bEnd - 32;
 
-/*
- * These definitions are only present to allow static allocation of an XXH
- * state, for example, on the stack or in a struct.
- * Never **ever** access members directly.
- */
+            do {
+                state->v[0] = XXH64_round(state->v[0], XXH_readLE64(p)); p+=8;
+                state->v[1] = XXH64_round(state->v[1], XXH_readLE64(p)); p+=8;
+                state->v[2] = XXH64_round(state->v[2], XXH_readLE64(p)); p+=8;
+                state->v[3] = XXH64_round(state->v[3], XXH_readLE64(p)); p+=8;
+            } while (p<=limit);
 
-struct XXH32_state_s {
-   XXH32_hash_t total_len_32;
-   XXH32_hash_t large_len;
-   XXH32_hash_t v1;
-   XXH32_hash_t v2;
-   XXH32_hash_t v3;
-   XXH32_hash_t v4;
-   XXH32_hash_t mem32[4];
-   XXH32_hash_t memsize;
-   XXH32_hash_t reserved;   /* never read nor write, might be removed in a future version */
-};   /* typedef'd to XXH32_state_t */
+        }
 
+        if (p < bEnd) {
+            XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
 
-#ifndef XXH_NO_LONG_LONG  /* defined when there is no 64-bit support */
+    return XXH_OK;
+}
 
-struct XXH64_state_s {
-   XXH64_hash_t total_len;
-   XXH64_hash_t v1;
-   XXH64_hash_t v2;
-   XXH64_hash_t v3;
-   XXH64_hash_t v4;
-   XXH64_hash_t mem64[4];
-   XXH32_hash_t memsize;
-   XXH32_hash_t reserved32;  /* required for padding anyway */
-   XXH64_hash_t reserved64;  /* never read nor write, might be removed in a future version */
-};   /* typedef'd to XXH64_state_t */
 
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state)
+{
+    xxh_u64 h64;
 
-/*-**********************************************************************
+    if (state->total_len >= 32) {
+        h64 = XXH_rotl64(state->v[0], 1) + XXH_rotl64(state->v[1], 7) + XXH_rotl64(state->v[2], 12) + XXH_rotl64(state->v[3], 18);
+        h64 = XXH64_mergeRound(h64, state->v[0]);
+        h64 = XXH64_mergeRound(h64, state->v[1]);
+        h64 = XXH64_mergeRound(h64, state->v[2]);
+        h64 = XXH64_mergeRound(h64, state->v[3]);
+    } else {
+        h64  = state->v[2] /*seed*/ + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) state->total_len;
+
+    return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
+}
+#endif /* !XXH_NO_STREAM */
+
+/******* Canonical representation   *******/
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
+    XXH_memcpy(dst, &hash, sizeof(*dst));
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src)
+{
+    return XXH_readBE64(src);
+}
+
+#ifndef XXH_NO_XXH3
+
+/* *********************************************************************
 *  XXH3
-*  New experimental hash
-************************************************************************/
+*  New generation hash designed for speed on small keys and vectorization
+************************************************************************ */
+/*!
+ * @}
+ * @defgroup XXH3_impl XXH3 implementation
+ * @ingroup impl
+ * @{
+ */
 
-/* ************************************************************************
- * XXH3 is a new hash algorithm featuring:
- *  - Improved speed for both small and large inputs
- *  - True 64-bit and 128-bit outputs
- *  - SIMD acceleration
- *  - Improved 32-bit viability
+/* ===   Compiler specifics   === */
+
+#if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */
+#  define XXH_RESTRICT   /* disable */
+#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* >= C99 */
+#  define XXH_RESTRICT   restrict
+#elif (defined (__GNUC__) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))) \
+   || (defined (__clang__)) \
+   || (defined (_MSC_VER) && (_MSC_VER >= 1400)) \
+   || (defined (__INTEL_COMPILER) && (__INTEL_COMPILER >= 1300))
+/*
+ * There are a LOT more compilers that recognize __restrict but this
+ * covers the major ones.
+ */
+#  define XXH_RESTRICT   __restrict
+#else
+#  define XXH_RESTRICT   /* disable */
+#endif
+
+#if (defined(__GNUC__) && (__GNUC__ >= 3))  \
+  || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
+  || defined(__clang__)
+#    define XXH_likely(x) __builtin_expect(x, 1)
+#    define XXH_unlikely(x) __builtin_expect(x, 0)
+#else
+#    define XXH_likely(x) (x)
+#    define XXH_unlikely(x) (x)
+#endif
+
+#ifndef XXH_HAS_INCLUDE
+#  ifdef __has_include
+/*
+ * Not defined as XXH_HAS_INCLUDE(x) (function-like) because
+ * this causes segfaults in Apple Clang 4.2 (on Mac OS X 10.7 Lion)
+ */
+#    define XXH_HAS_INCLUDE __has_include
+#  else
+#    define XXH_HAS_INCLUDE(x) 0
+#  endif
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+#  if defined(__ARM_FEATURE_SVE)
+#    include <arm_sve.h>
+#  endif
+#  if defined(__ARM_NEON__) || defined(__ARM_NEON) \
+   || (defined(_M_ARM) && _M_ARM >= 7) \
+   || defined(_M_ARM64) || defined(_M_ARM64EC) \
+   || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* WASM SIMD128 via SIMDe */
+#    define inline __inline__  /* circumvent a clang bug */
+#    include <arm_neon.h>
+#    undef inline
+#  elif defined(__AVX2__)
+#    include <immintrin.h>
+#  elif defined(__SSE2__)
+#    include <emmintrin.h>
+#  endif
+#endif
+
+#if defined(_MSC_VER)
+#  include <intrin.h>
+#endif
+
+/*
+ * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
+ * remaining a true 64-bit/128-bit hash function.
  *
- * Speed analysis methodology is explained here:
+ * This is done by prioritizing a subset of 64-bit operations that can be
+ * emulated without too many steps on the average 32-bit machine.
  *
- *    https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ * For example, these two lines seem similar, and run equally fast on 64-bit:
  *
- * In general, expect XXH3 to run about ~2x faster on large inputs and >3x
- * faster on small ones compared to XXH64, though exact differences depend on
- * the platform.
+ *   xxh_u64 x;
+ *   x ^= (x >> 47); // good
+ *   x ^= (x >> 13); // bad
  *
- * The algorithm is portable: Like XXH32 and XXH64, it generates the same hash
- * on all platforms.
+ * However, to a 32-bit machine, there is a major difference.
  *
- * It benefits greatly from SIMD and 64-bit arithmetic, but does not require it.
+ * x ^= (x >> 47) looks like this:
  *
- * Almost all 32-bit and 64-bit targets that can run XXH32 smoothly can run
- * XXH3 at competitive speeds, even if XXH64 runs slowly. Further details are
- * explained in the implementation.
+ *   x.lo ^= (x.hi >> (47 - 32));
  *
- * Optimized implementations are provided for AVX2, SSE2, NEON, POWER8, ZVector,
- * and scalar targets. This can be controlled with the XXH_VECTOR macro.
+ * while x ^= (x >> 13) looks like this:
  *
- * XXH3 offers 2 variants, _64bits and _128bits.
- * When only 64 bits are needed, prefer calling the _64bits variant, as it
- * reduces the amount of mixing, resulting in faster speed on small inputs.
+ *   // note: funnel shifts are not usually cheap.
+ *   x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
+ *   x.hi ^= (x.hi >> 13);
  *
- * It's also generally simpler to manipulate a scalar return type than a struct.
+ * The first one is significantly faster than the second, simply because the
+ * shift is larger than 32. This means:
+ *  - All the bits we need are in the upper 32 bits, so we can ignore the lower
+ *    32 bits in the shift.
+ *  - The shift result will always fit in the lower 32 bits, and therefore,
+ *    we can ignore the upper 32 bits in the xor.
  *
- * The 128-bit version adds additional strength, but it is slightly slower.
+ * Thanks to this optimization, XXH3 only requires these features to be efficient:
  *
- * The XXH3 algorithm is still in development.
- * The results it produces may still change in future versions.
+ *  - Usable unaligned access
+ *  - A 32-bit or 64-bit ALU
+ *      - If 32-bit, a decent ADC instruction
+ *  - A 32 or 64-bit multiply with a 64-bit result
+ *  - For the 128-bit variant, a decent byteswap helps short inputs.
  *
- * Results produced by v0.7.x are not comparable with results from v0.7.y.
- * However, the API is completely stable, and it can safely be used for
- * ephemeral data (local sessions).
+ * The first two are already required by XXH32, and almost all 32-bit and 64-bit
+ * platforms which can run XXH32 can run XXH3 efficiently.
  *
- * Avoid storing values in long-term storage until the algorithm is finalized.
+ * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
+ * notable exception.
  *
- * Since v0.7.3, XXH3 has reached "release candidate" status, meaning that, if
- * everything remains fine, its current format will be "frozen" and become the
- * final one.
+ * First of all, Thumb-1 lacks support for the UMULL instruction which
+ * performs the important long multiply. This means numerous __aeabi_lmul
+ * calls.
  *
- * After which, return values of XXH3 and XXH128 will no longer change in
- * future versions.
+ * Second of all, the 8 functional registers are just not enough.
+ * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
+ * Lo registers, and this shuffling results in thousands more MOVs than A32.
  *
- * XXH3's return values will be officially finalized upon reaching v0.8.0.
+ * A32 and T32 don't have this limitation. They can access all 14 registers,
+ * do a 32->64 multiply with UMULL, and the flexible operand allowing free
+ * shifts is helpful, too.
  *
- * The API supports one-shot hashing, streaming mode, and custom secrets.
+ * Therefore, we do a quick sanity check.
+ *
+ * If compiling Thumb-1 for a target which supports ARM instructions, we will
+ * emit a warning, as it is not a "sane" platform to compile for.
+ *
+ * Usually, if this happens, it is because of an accident and you probably need
+ * to specify -march, as you likely meant to compile for a newer architecture.
+ *
+ * Credit: large sections of the vectorial and asm source code paths
+ *         have been contributed by @easyaspi314
  */
+#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
+#   warning "XXH3 is highly inefficient without ARM or Thumb-2."
+#endif
 
-#ifdef XXH_NAMESPACE
-#  define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
-#  define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
-#  define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
+/* ==========================================
+ * Vectorization detection
+ * ========================================== */
 
-#  define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
-#  define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
-#  define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
+#ifdef XXH_DOXYGEN
+/*!
+ * @ingroup tuning
+ * @brief Overrides the vectorization implementation chosen for XXH3.
+ *
+ * Can be defined to 0 to disable SIMD or any of the values mentioned in
+ * @ref XXH_VECTOR_TYPE.
+ *
+ * If this is not defined, it uses predefined macros to determine the best
+ * implementation.
+ */
+#  define XXH_VECTOR XXH_SCALAR
+/*!
+ * @ingroup tuning
+ * @brief Possible values for @ref XXH_VECTOR.
+ *
+ * Note that these are actually implemented as macros.
+ *
+ * If this is not defined, it is detected automatically.
+ * internal macro XXH_X86DISPATCH overrides this.
+ */
+enum XXH_VECTOR_TYPE /* fake enum */ {
+    XXH_SCALAR = 0,  /*!< Portable scalar version */
+    XXH_SSE2   = 1,  /*!<
+                      * SSE2 for Pentium 4, Opteron, all x86_64.
+                      *
+                      * @note SSE2 is also guaranteed on Windows 10, macOS, and
+                      * Android x86.
+                      */
+    XXH_AVX2   = 2,  /*!< AVX2 for Haswell and Bulldozer */
+    XXH_AVX512 = 3,  /*!< AVX512 for Skylake and Icelake */
+    XXH_NEON   = 4,  /*!<
+                       * NEON for most ARMv7-A, all AArch64, and WASM SIMD128
+                       * via the SIMDeverywhere polyfill provided with the
+                       * Emscripten SDK.
+                       */
+    XXH_VSX    = 5,  /*!< VSX and ZVector for POWER8/z13 (64-bit) */
+    XXH_SVE    = 6,  /*!< SVE for some ARMv8-A and ARMv9-A */
+};
+/*!
+ * @ingroup tuning
+ * @brief Selects the minimum alignment for XXH3's accumulators.
+ *
+ * When using SIMD, this should match the alignment required for said vector
+ * type, so, for example, 32 for AVX2.
+ *
+ * Default: Auto detected.
+ */
+#  define XXH_ACC_ALIGN 8
+#endif
 
-#  define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
-#  define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
-#  define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
-#  define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
-#  define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
+/* Actual definition */
+#ifndef XXH_DOXYGEN
+#  define XXH_SCALAR 0
+#  define XXH_SSE2   1
+#  define XXH_AVX2   2
+#  define XXH_AVX512 3
+#  define XXH_NEON   4
+#  define XXH_VSX    5
+#  define XXH_SVE    6
+#endif
+
+#ifndef XXH_VECTOR    /* can be defined on command line */
+#  if defined(__ARM_FEATURE_SVE)
+#    define XXH_VECTOR XXH_SVE
+#  elif ( \
+        defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \
+     || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \
+     || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* wasm simd128 via SIMDe */ \
+   ) && ( \
+        defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \
+    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
+   )
+#    define XXH_VECTOR XXH_NEON
+#  elif defined(__AVX512F__)
+#    define XXH_VECTOR XXH_AVX512
+#  elif defined(__AVX2__)
+#    define XXH_VECTOR XXH_AVX2
+#  elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
+#    define XXH_VECTOR XXH_SSE2
+#  elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
+     || (defined(__s390x__) && defined(__VEC__)) \
+     && defined(__GNUC__) /* TODO: IBM XL */
+#    define XXH_VECTOR XXH_VSX
+#  else
+#    define XXH_VECTOR XXH_SCALAR
+#  endif
 #endif
 
-/* XXH3_64bits():
- * default 64-bit variant, using default secret and default seed of 0.
- * It's the fastest variant. */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* data, size_t len);
+/* __ARM_FEATURE_SVE is only supported by GCC & Clang. */
+#if (XXH_VECTOR == XXH_SVE) && !defined(__ARM_FEATURE_SVE)
+#  ifdef _MSC_VER
+#    pragma warning(once : 4606)
+#  else
+#    warning "__ARM_FEATURE_SVE isn't supported. Use SCALAR instead."
+#  endif
+#  undef XXH_VECTOR
+#  define XXH_VECTOR XXH_SCALAR
+#endif
 
 /*
- * XXH3_64bits_withSecret():
- * It's possible to provide any blob of bytes as a "secret" to generate the hash.
- * This makes it more difficult for an external actor to prepare an intentional
- * collision.
- * The secret *must* be large enough (>= XXH3_SECRET_SIZE_MIN).
- * It should consist of random bytes.
- * Avoid trivial sequences, such as repeating sequences and especially '\0',
- * as this can cancel out itself.
- * Failure to respect these conditions will result in a poor quality hash.
+ * Controls the alignment of the accumulator,
+ * for compatibility with aligned vector loads, which are usually faster.
  */
-#define XXH3_SECRET_SIZE_MIN 136
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+#ifndef XXH_ACC_ALIGN
+#  if defined(XXH_X86DISPATCH)
+#     define XXH_ACC_ALIGN 64  /* for compatibility with avx512 */
+#  elif XXH_VECTOR == XXH_SCALAR  /* scalar */
+#     define XXH_ACC_ALIGN 8
+#  elif XXH_VECTOR == XXH_SSE2  /* sse2 */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_AVX2  /* avx2 */
+#     define XXH_ACC_ALIGN 32
+#  elif XXH_VECTOR == XXH_NEON  /* neon */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_VSX   /* vsx */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_AVX512  /* avx512 */
+#     define XXH_ACC_ALIGN 64
+#  elif XXH_VECTOR == XXH_SVE   /* sve */
+#     define XXH_ACC_ALIGN 64
+#  endif
+#endif
+
+#if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \
+    || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512
+#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#elif XXH_VECTOR == XXH_SVE
+#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#else
+#  define XXH_SEC_ALIGN 8
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+#  define XXH_ALIASING __attribute__((__may_alias__))
+#else
+#  define XXH_ALIASING /* nothing */
+#endif
 
 /*
- * XXH3_64bits_withSeed():
- * This variant generates a custom secret on the fly based on the default
- * secret, altered using the `seed` value.
- * While this operation is decently fast, note that it's not completely free.
- * Note: seed==0 produces the same results as XXH3_64bits().
+ * UGLY HACK:
+ * GCC usually generates the best code with -O3 for xxHash.
+ *
+ * However, when targeting AVX2, it is overzealous in its unrolling resulting
+ * in code roughly 3/4 the speed of Clang.
+ *
+ * There are other issues, such as GCC splitting _mm256_loadu_si256 into
+ * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
+ * only applies to Sandy and Ivy Bridge... which don't even support AVX2.
+ *
+ * That is why when compiling the AVX2 version, it is recommended to use either
+ *   -O2 -mavx2 -march=haswell
+ * or
+ *   -O2 -mavx2 -mno-avx256-split-unaligned-load
+ * for decent performance, or to use Clang instead.
+ *
+ * Fortunately, we can control the first one with a pragma that forces GCC into
+ * -O2, but the other one we can't control without "failed to inline always
+ * inline function due to target mismatch" warnings.
  */
-XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
+#  pragma GCC push_options
+#  pragma GCC optimize("-O2")
+#endif
 
+#if XXH_VECTOR == XXH_NEON
 
-/* streaming 64-bit */
+/*
+ * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3
+ * optimizes out the entire hashLong loop because of the aliasing violation.
+ *
+ * However, GCC is also inefficient at load-store optimization with vld1q/vst1q,
+ * so the only option is to mark it as aliasing.
+ */
+typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING;
 
-#if defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)   /* C11+ */
-#  include <stdalign.h>
-#  define XXH_ALIGN(n)      alignas(n)
-#elif defined(__GNUC__)
-#  define XXH_ALIGN(n)      __attribute__ ((aligned(n)))
-#elif defined(_MSC_VER)
-#  define XXH_ALIGN(n)      __declspec(align(n))
+/*!
+ * @internal
+ * @brief `vld1q_u64` but faster and alignment-safe.
+ *
+ * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only
+ * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86).
+ *
+ * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it
+ * prohibits load-store optimizations. Therefore, a direct dereference is used.
+ *
+ * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe
+ * unaligned load.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__)
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) /* silence -Wcast-align */
+{
+    return *(xxh_aliasing_uint64x2_t const *)ptr;
+}
 #else
-#  define XXH_ALIGN(n)   /* disabled */
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
+{
+    return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr));
+}
 #endif
 
-/* Old GCC versions only accept the attribute after the type in structures. */
-#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L))   /* C11+ */ \
-    && defined(__GNUC__)
-#   define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
+/*!
+ * @internal
+ * @brief `vmlal_u32` on low and high halves of a vector.
+ *
+ * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with
+ * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32`
+ * with `vmlal_u32`.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    /* Inline assembly is the only way */
+    __asm__("umlal   %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs));
+    return acc;
+}
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    /* This intrinsic works as expected */
+    return vmlal_high_u32(acc, lhs, rhs);
+}
 #else
-#   define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
+/* Portable intrinsic versions */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs));
+}
+/*! @copydoc XXH_vmlal_low_u32
+ * Assume the compiler converts this to vmlal_high_u32 on aarch64 */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs));
+}
 #endif
 
-typedef struct XXH3_state_s XXH3_state_t;
-
-#define XXH3_SECRET_DEFAULT_SIZE 192   /* minimum XXH3_SECRET_SIZE_MIN */
-#define XXH3_INTERNALBUFFER_SIZE 256
-struct XXH3_state_s {
-   XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
-  /* used to store a custom secret generated from the seed. Makes state larger.
-   * Design might change */
-   XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
-   XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
-   XXH32_hash_t bufferedSize;
-   XXH32_hash_t nbStripesPerBlock;
-   XXH32_hash_t nbStripesSoFar;
-   XXH32_hash_t secretLimit;
-   XXH32_hash_t reserved32;
-   XXH32_hash_t reserved32_2;
-   XXH64_hash_t totalLen;
-   XXH64_hash_t seed;
-   XXH64_hash_t reserved64;
-   /* note: there is some padding after due to alignment on 64 bytes */
-   const unsigned char* secret;
-};   /* typedef'd to XXH3_state_t */
-
-#undef XXH_ALIGN_MEMBER
+/*!
+ * @ingroup tuning
+ * @brief Controls the NEON to scalar ratio for XXH3
+ *
+ * This can be set to 2, 4, 6, or 8.
+ *
+ * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used.
+ *
+ * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but only 2 of those
+ * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU
+ * bandwidth.
+ *
+ * This is even more noticeable on the more advanced cores like the Cortex-A76 which
+ * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once.
+ *
+ * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes
+ * and 2 scalar lanes, which is chosen by default.
+ *
+ * This does not apply to Apple processors or 32-bit processors, which run better with
+ * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes.
+ *
+ * This change benefits CPUs with large micro-op buffers without negatively affecting
+ * most other CPUs:
+ *
+ *  | Chipset               | Dispatch type       | NEON only | 6:2 hybrid | Diff. |
+ *  |:----------------------|:--------------------|----------:|-----------:|------:|
+ *  | Snapdragon 730 (A76)  | 2 NEON/8 micro-ops  |  8.8 GB/s |  10.1 GB/s |  ~16% |
+ *  | Snapdragon 835 (A73)  | 2 NEON/3 micro-ops  |  5.1 GB/s |   5.3 GB/s |   ~5% |
+ *  | Marvell PXA1928 (A53) | In-order dual-issue |  1.9 GB/s |   1.9 GB/s |    0% |
+ *  | Apple M1              | 4 NEON/8 micro-ops  | 37.3 GB/s |  36.1 GB/s |  ~-3% |
+ *
+ * It also seems to fix some bad codegen on GCC, making it almost as fast as clang.
+ *
+ * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning
+ * it effectively becomes worse 4.
+ *
+ * @see XXH3_accumulate_512_neon()
+ */
+# ifndef XXH3_NEON_LANES
+#  if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
+   && !defined(__APPLE__) && XXH_SIZE_OPT <= 0
+#   define XXH3_NEON_LANES 6
+#  else
+#   define XXH3_NEON_LANES XXH_ACC_NB
+#  endif
+# endif
+#endif  /* XXH_VECTOR == XXH_NEON */
 
 /*
- * Streaming requires state maintenance.
- * This operation costs memory and CPU.
- * As a consequence, streaming is slower than one-shot hashing.
- * For better performance, prefer one-shot functions whenever possible.
+ * VSX and Z Vector helpers.
+ *
+ * This is very messy, and any pull requests to clean this up are welcome.
+ *
+ * There are a lot of problems with supporting VSX and s390x, due to
+ * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
  */
-XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void);
-XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
-XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state);
+#if XXH_VECTOR == XXH_VSX
+/* Annoyingly, these headers _may_ define three macros: `bool`, `vector`,
+ * and `pixel`. This is a problem for obvious reasons.
+ *
+ * These keywords are unnecessary; the spec literally says they are
+ * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd
+ * after including the header.
+ *
+ * We use pragma push_macro/pop_macro to keep the namespace clean. */
+#  pragma push_macro("bool")
+#  pragma push_macro("vector")
+#  pragma push_macro("pixel")
+/* silence potential macro redefined warnings */
+#  undef bool
+#  undef vector
+#  undef pixel
+
+#  if defined(__s390x__)
+#    include <s390intrin.h>
+#  else
+#    include <altivec.h>
+#  endif
+
+/* Restore the original macro values, if applicable. */
+#  pragma pop_macro("pixel")
+#  pragma pop_macro("vector")
+#  pragma pop_macro("bool")
 
+typedef __vector unsigned long long xxh_u64x2;
+typedef __vector unsigned char xxh_u8x16;
+typedef __vector unsigned xxh_u32x4;
 
 /*
- * XXH3_64bits_reset():
- * Initialize with the default parameters.
- * The result will be equivalent to `XXH3_64bits()`.
+ * UGLY HACK: Similar to aarch64 macOS GCC, s390x GCC has the same aliasing issue.
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t* statePtr);
-/*
- * XXH3_64bits_reset_withSeed():
- * Generate a custom secret from `seed`, and store it into `statePtr`.
- * digest will be equivalent to `XXH3_64bits_withSeed()`.
+typedef xxh_u64x2 xxh_aliasing_u64x2 XXH_ALIASING;
+
+# ifndef XXH_VSX_BE
+#  if defined(__BIG_ENDIAN__) \
+  || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXH_VSX_BE 1
+#  elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
+#    warning "-maltivec=be is not recommended. Please use native endianness."
+#    define XXH_VSX_BE 1
+#  else
+#    define XXH_VSX_BE 0
+#  endif
+# endif /* !defined(XXH_VSX_BE) */
+
+# if XXH_VSX_BE
+#  if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
+#    define XXH_vec_revb vec_revb
+#  else
+/*!
+ * A polyfill for POWER9's vec_revb().
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
-/*
- * XXH3_64bits_reset_withSecret():
- * `secret` is referenced, and must outlive the hash streaming session, so
- * be careful when using stack arrays.
- * `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`.
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
+{
+    xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
+                                  0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
+    return vec_perm(val, val, vByteSwap);
+}
+#  endif
+# endif /* XXH_VSX_BE */
+
+/*!
+ * Performs an unaligned vector load and byte swaps it on big endian.
  */
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
+{
+    xxh_u64x2 ret;
+    XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2));
+# if XXH_VSX_BE
+    ret = XXH_vec_revb(ret);
+# endif
+    return ret;
+}
 
-XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_digest (const XXH3_state_t* statePtr);
+/*
+ * vec_mulo and vec_mule are very problematic intrinsics on PowerPC
+ *
+ * These intrinsics weren't added until GCC 8, despite existing for a while,
+ * and they are endian dependent. Also, their meaning swap depending on version.
+ * */
+# if defined(__s390x__)
+ /* s390x is always big endian, no issue on this platform */
+#  define XXH_vec_mulo vec_mulo
+#  define XXH_vec_mule vec_mule
+# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) && !defined(__ibmxl__)
+/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
+ /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */
+#  define XXH_vec_mulo __builtin_altivec_vmulouw
+#  define XXH_vec_mule __builtin_altivec_vmuleuw
+# else
+/* gcc needs inline assembly */
+/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
+{
+    xxh_u64x2 result;
+    __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
+{
+    xxh_u64x2 result;
+    __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+# endif /* XXH_vec_mulo, XXH_vec_mule */
+#endif /* XXH_VECTOR == XXH_VSX */
+
+#if XXH_VECTOR == XXH_SVE
+#define ACCRND(acc, offset) \
+do { \
+    svuint64_t input_vec = svld1_u64(mask, xinput + offset);         \
+    svuint64_t secret_vec = svld1_u64(mask, xsecret + offset);       \
+    svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec);     \
+    svuint64_t swapped = svtbl_u64(input_vec, kSwap);                \
+    svuint64_t mixed_lo = svextw_u64_x(mask, mixed);                 \
+    svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32);            \
+    svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \
+    acc = svadd_u64_x(mask, acc, mul);                               \
+} while (0)
+#endif /* XXH_VECTOR == XXH_SVE */
 
+/* prefetch
+ * can be disabled, by declaring XXH_NO_PREFETCH build macro */
+#if defined(XXH_NO_PREFETCH)
+#  define XXH_PREFETCH(ptr)  (void)(ptr)  /* disabled */
+#else
+#  if XXH_SIZE_OPT >= 1
+#    define XXH_PREFETCH(ptr) (void)(ptr)
+#  elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
+#    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
+#    define XXH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
+#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
+#    define XXH_PREFETCH(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
+#  else
+#    define XXH_PREFETCH(ptr) (void)(ptr)  /* disabled */
+#  endif
+#endif  /* XXH_NO_PREFETCH */
 
-/* 128-bit */
 
-#ifdef XXH_NAMESPACE
-#  define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
-#  define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
-#  define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
-#  define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
+/* ==========================================
+ * XXH3 default settings
+ * ========================================== */
 
-#  define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
-#  define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
-#  define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
-#  define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
-#  define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
+#define XXH_SECRET_DEFAULT_SIZE 192   /* minimum XXH3_SECRET_SIZE_MIN */
 
-#  define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
-#  define XXH128_cmp     XXH_NAME2(XXH_NAMESPACE, XXH128_cmp)
-#  define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash)
-#  define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical)
+#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
+#  error "default keyset is not large enough"
 #endif
 
-typedef struct {
-    XXH64_hash_t low64;
-    XXH64_hash_t high64;
-} XXH128_hash_t;
-
-XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* data, size_t len);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);  /* == XXH128() */
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
-
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t* statePtr);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
-
-XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* statePtr);
-
-
-/* Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
+/*! Pseudorandom secret taken directly from FARSH. */
+XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = {
+    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
+    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
+    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
+    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
+    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
+    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
+    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
+    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
+    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
+    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
+    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
+    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
+};
+
+static const xxh_u64 PRIME_MX1 = 0x165667919E3779F9ULL;  /*!< 0b0001011001010110011001111001000110011110001101110111100111111001 */
+static const xxh_u64 PRIME_MX2 = 0x9FB21C651E98DF25ULL;  /*!< 0b1001111110110010000111000110010100011110100110001101111100100101 */
+
+#ifdef XXH_OLD_NAMES
+#  define kSecret XXH3_kSecret
+#endif
 
+#ifdef XXH_DOXYGEN
 /*!
- * XXH128_isEqual():
- * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
+ * @brief Calculates a 32-bit to 64-bit long multiply.
+ *
+ * Implemented as a macro.
+ *
+ * Wraps `__emulu` on MSVC x86 because it tends to call `__allmul` when it doesn't
+ * need to (but it shouldn't need to anyways, it is about 7 instructions to do
+ * a 64x64 multiply...). Since we know that this will _always_ emit `MULL`, we
+ * use that instead of the normal method.
+ *
+ * If you are compiling for platforms like Thumb-1 and don't have a better option,
+ * you may also want to write your own long multiply routine here.
+ *
+ * @param x, y Numbers to be multiplied
+ * @return 64-bit product of the low 32 bits of @p x and @p y.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64(xxh_u64 x, xxh_u64 y)
+{
+   return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
+}
+#elif defined(_MSC_VER) && defined(_M_IX86)
+#    define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
+#else
+/*
+ * Downcast + upcast is usually better than masking on older compilers like
+ * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
+ *
+ * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
+ * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
  */
-XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+#    define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
+#endif
 
 /*!
- * XXH128_cmp():
+ * @brief Calculates a 64->128-bit long multiply.
  *
- * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
+ * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar
+ * version.
  *
- * return: >0 if *h128_1  > *h128_2
- *         <0 if *h128_1  < *h128_2
- *         =0 if *h128_1 == *h128_2
+ * @param lhs , rhs The 64-bit integers to be multiplied
+ * @return The 128-bit result represented in an @ref XXH128_hash_t.
  */
-XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2);
-
-
-/*******   Canonical representation   *******/
-typedef struct { unsigned char digest[16]; } XXH128_canonical_t;
-XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash);
-XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* src);
+static XXH128_hash_t
+XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
+{
+    /*
+     * GCC/Clang __uint128_t method.
+     *
+     * On most 64-bit targets, GCC and Clang define a __uint128_t type.
+     * This is usually the best way as it usually uses a native long 64-bit
+     * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
+     *
+     * Usually.
+     *
+     * Despite being a 32-bit platform, Clang (and emscripten) define this type
+     * despite not having the arithmetic for it. This results in a laggy
+     * compiler builtin call which calculates a full 128-bit multiply.
+     * In that case it is best to use the portable one.
+     * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
+     */
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \
+    && defined(__SIZEOF_INT128__) \
+    || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
 
+    __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
+    XXH128_hash_t r128;
+    r128.low64  = (xxh_u64)(product);
+    r128.high64 = (xxh_u64)(product >> 64);
+    return r128;
 
-#endif  /* XXH_NO_LONG_LONG */
+    /*
+     * MSVC for x64's _umul128 method.
+     *
+     * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
+     *
+     * This compiles to single operand MUL on x64.
+     */
+#elif (defined(_M_X64) || defined(_M_IA64)) && !defined(_M_ARM64EC)
 
-#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
-#  define XXH_IMPLEMENTATION
+#ifndef _MSC_VER
+#   pragma intrinsic(_umul128)
 #endif
+    xxh_u64 product_high;
+    xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
+    XXH128_hash_t r128;
+    r128.low64  = product_low;
+    r128.high64 = product_high;
+    return r128;
 
-#endif  /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */
+    /*
+     * MSVC for ARM64's __umulh method.
+     *
+     * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method.
+     */
+#elif defined(_M_ARM64) || defined(_M_ARM64EC)
 
+#ifndef _MSC_VER
+#   pragma intrinsic(__umulh)
+#endif
+    XXH128_hash_t r128;
+    r128.low64  = lhs * rhs;
+    r128.high64 = __umulh(lhs, rhs);
+    return r128;
 
-/* ======================================================================== */
-/* ======================================================================== */
-/* ======================================================================== */
+#else
+    /*
+     * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
+     *
+     * This is a fast and simple grade school multiply, which is shown below
+     * with base 10 arithmetic instead of base 0x100000000.
+     *
+     *           9 3 // D2 lhs = 93
+     *         x 7 5 // D2 rhs = 75
+     *     ----------
+     *           1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
+     *         4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
+     *         2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
+     *     + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
+     *     ---------
+     *         2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
+     *     + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
+     *     ---------
+     *       6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
+     *
+     * The reasons for adding the products like this are:
+     *  1. It avoids manual carry tracking. Just like how
+     *     (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
+     *     This avoids a lot of complexity.
+     *
+     *  2. It hints for, and on Clang, compiles to, the powerful UMAAL
+     *     instruction available in ARM's Digital Signal Processing extension
+     *     in 32-bit ARMv6 and later, which is shown below:
+     *
+     *         void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
+     *         {
+     *             xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
+     *             *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
+     *             *RdHi = (xxh_u32)(product >> 32);
+     *         }
+     *
+     *     This instruction was designed for efficient long multiplication, and
+     *     allows this to be calculated in only 4 instructions at speeds
+     *     comparable to some 64-bit ALUs.
+     *
+     *  3. It isn't terrible on other platforms. Usually this will be a couple
+     *     of 32-bit ADD/ADCs.
+     */
 
+    /* First calculate all of the cross products. */
+    xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
+    xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32,        rhs & 0xFFFFFFFF);
+    xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
+    xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32,        rhs >> 32);
+
+    /* Now add the products together. These will never overflow. */
+    xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
+    xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32)        + hi_hi;
+    xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
+
+    XXH128_hash_t r128;
+    r128.low64  = lower;
+    r128.high64 = upper;
+    return r128;
+#endif
+}
 
-/*-**********************************************************************
- * xxHash implementation
- *-**********************************************************************
- * xxHash's implementation used to be found in xxhash.c.
- *
- * However, code inlining requires the implementation to be visible to the
- * compiler, usually within the header.
+/*!
+ * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it.
  *
- * As a workaround, xxhash.c used to be included within xxhash.h. This caused
- * some issues with some build systems, especially ones which treat .c files
- * as source files.
+ * The reason for the separate function is to prevent passing too many structs
+ * around by value. This will hopefully inline the multiply, but we don't force it.
  *
- * Therefore, the implementation is now directly integrated within xxhash.h.
- * Another small advantage is that xxhash.c is no longer needed in /include.
- ************************************************************************/
+ * @param lhs , rhs The 64-bit integers to multiply
+ * @return The low 64 bits of the product XOR'd by the high 64 bits.
+ * @see XXH_mult64to128()
+ */
+static xxh_u64
+XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
+{
+    XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
+    return product.low64 ^ product.high64;
+}
 
-#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
-   || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
-#  define XXH_IMPLEM_13a8737387
+/*! Seems to produce slightly better code on GCC for some reason. */
+XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
+{
+    XXH_ASSERT(0 <= shift && shift < 64);
+    return v64 ^ (v64 >> shift);
+}
 
-/* *************************************
-*  Tuning parameters
-***************************************/
-/*!
- * XXH_FORCE_MEMORY_ACCESS:
- * By default, access to unaligned memory is controlled by `memcpy()`, which is
- * safe and portable.
- *
- * Unfortunately, on some target/compiler combinations, the generated assembly
- * is sub-optimal.
- *
- * The below switch allow to select a different access method for improved
- * performance.
- * Method 0 (default):
- *     Use `memcpy()`. Safe and portable.
- * Method 1:
- *     `__attribute__((packed))` statement. It depends on compiler extensions
- *     and is therefore not portable.
- *     This method is safe if your compiler supports it, and *generally* as
- *     fast or faster than `memcpy`.
- * Method 2:
- *     Direct access via cast. This method doesn't depend on the compiler but
- *     violates the C standard.
- *     It can generate buggy code on targets which do not support unaligned
- *     memory accesses.
- *     But in some circumstances, it's the only known way to get the most
- *     performance (ie GCC + ARMv6)
- * Method 3:
- *     Byteshift. This can generate the best code on old compilers which don't
- *     inline small `memcpy()` calls, and it might also be faster on big-endian
- *     systems which lack a native byteswap instruction.
- * See https://stackoverflow.com/a/32095106/646947 for details.
- * Prefer these methods in priority order (0 > 1 > 2 > 3)
+/*
+ * This is a fast avalanche stage,
+ * suitable when input bits are already partially mixed
  */
-#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if !defined(__clang__) && defined(__GNUC__) && defined(__ARM_FEATURE_UNALIGNED) && defined(__ARM_ARCH) && (__ARM_ARCH == 6)
-#    define XXH_FORCE_MEMORY_ACCESS 2
-#  elif !defined(__clang__) && ((defined(__INTEL_COMPILER) && !defined(_WIN32)) || \
-  (defined(__GNUC__) && (defined(__ARM_ARCH) && __ARM_ARCH >= 7)))
-#    define XXH_FORCE_MEMORY_ACCESS 1
-#  endif
-#endif
+static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
+{
+    h64 = XXH_xorshift64(h64, 37);
+    h64 *= PRIME_MX1;
+    h64 = XXH_xorshift64(h64, 32);
+    return h64;
+}
 
-/*!
- *XXH_ACCEPT_NULL_INPUT_POINTER:
- * If the input pointer is NULL, xxHash's default behavior is to dereference it,
- * triggering a segfault.
- * When this macro is enabled, xxHash actively checks the input for a null pointer.
- * If it is, the result for null input pointers is the same as a zero-length input.
+/*
+ * This is a stronger avalanche,
+ * inspired by Pelle Evensen's rrmxmx
+ * preferable when input has not been previously mixed
  */
-#ifndef XXH_ACCEPT_NULL_INPUT_POINTER   /* can be defined externally */
-#  define XXH_ACCEPT_NULL_INPUT_POINTER 0
-#endif
+static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
+{
+    /* this mix is inspired by Pelle Evensen's rrmxmx */
+    h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24);
+    h64 *= PRIME_MX2;
+    h64 ^= (h64 >> 35) + len ;
+    h64 *= PRIME_MX2;
+    return XXH_xorshift64(h64, 28);
+}
 
-/*!
- * XXH_FORCE_ALIGN_CHECK:
- * This is a minor performance trick, only useful with lots of very small keys.
- * It means: check for aligned/unaligned input.
- * The check costs one initial branch per hash;
- * Set it to 0 when the input is guaranteed to be aligned or when alignment
- * doesn't matter for performance.
+
+/* ==========================================
+ * Short keys
+ * ==========================================
+ * One of the shortcomings of XXH32 and XXH64 was that their performance was
+ * sub-optimal on short lengths. It used an iterative algorithm which strongly
+ * favored lengths that were a multiple of 4 or 8.
  *
- * This option does not affect XXH3.
+ * Instead of iterating over individual inputs, we use a set of single shot
+ * functions which piece together a range of lengths and operate in constant time.
+ *
+ * Additionally, the number of multiplies has been significantly reduced. This
+ * reduces latency, especially when emulating 64-bit multiplies on 32-bit.
+ *
+ * Depending on the platform, this may or may not be faster than XXH32, but it
+ * is almost guaranteed to be faster than XXH64.
  */
-#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */
-#  if defined(__i386) || defined(_M_IX86) || defined(__x86_64__) || defined(_M_X64)
-#    define XXH_FORCE_ALIGN_CHECK 0
-#  else
-#    define XXH_FORCE_ALIGN_CHECK 1
-#  endif
-#endif
 
-/*!
- * XXH_NO_INLINE_HINTS:
- *
- * By default, xxHash tries to force the compiler to inline almost all internal
- * functions.
+/*
+ * At very short lengths, there isn't enough input to fully hide secrets, or use
+ * the entire secret.
  *
- * This can usually improve performance due to reduced jumping and improved
- * constant folding, but significantly increases the size of the binary which
- * might not be favorable.
+ * There is also only a limited amount of mixing we can do before significantly
+ * impacting performance.
  *
- * Additionally, sometimes the forced inlining can be detrimental to performance,
- * depending on the architecture.
+ * Therefore, we use different sections of the secret and always mix two secret
+ * samples with an XOR. This should have no effect on performance on the
+ * seedless or withSeed variants because everything _should_ be constant folded
+ * by modern compilers.
  *
- * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
- * compiler full control on whether to inline or not.
+ * The XOR mixing hides individual parts of the secret and increases entropy.
  *
- * When not optimizing (-O0), optimizing for size (-Os, -Oz), or using
- * -fno-inline with GCC or Clang, this will automatically be defined.
+ * This adds an extra layer of strength for custom secrets.
  */
-#ifndef XXH_NO_INLINE_HINTS
-#  if defined(__OPTIMIZE_SIZE__) /* -Os, -Oz */ \
-   || defined(__NO_INLINE__)     /* -O0, -fno-inline */
-#    define XXH_NO_INLINE_HINTS 1
-#  else
-#    define XXH_NO_INLINE_HINTS 0
-#  endif
-#endif
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(1 <= len && len <= 3);
+    XXH_ASSERT(secret != NULL);
+    /*
+     * len = 1: combined = { input[0], 0x01, input[0], input[0] }
+     * len = 2: combined = { input[1], 0x02, input[0], input[1] }
+     * len = 3: combined = { input[2], 0x03, input[0], input[1] }
+     */
+    {   xxh_u8  const c1 = input[0];
+        xxh_u8  const c2 = input[len >> 1];
+        xxh_u8  const c3 = input[len - 1];
+        xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2  << 24)
+                               | ((xxh_u32)c3 <<  0) | ((xxh_u32)len << 8);
+        xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+        xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
+        return XXH64_avalanche(keyed);
+    }
+}
 
-/*!
- * XXH_REROLL:
- * Whether to reroll XXH32_finalize, and XXH64_finalize,
- * instead of using an unrolled jump table/if statement loop.
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(4 <= len && len <= 8);
+    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+    {   xxh_u32 const input1 = XXH_readLE32(input);
+        xxh_u32 const input2 = XXH_readLE32(input + len - 4);
+        xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
+        xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
+        xxh_u64 const keyed = input64 ^ bitflip;
+        return XXH3_rrmxmx(keyed, len);
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
+        xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
+        xxh_u64 const input_lo = XXH_readLE64(input)           ^ bitflip1;
+        xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
+        xxh_u64 const acc = len
+                          + XXH_swap64(input_lo) + input_hi
+                          + XXH3_mul128_fold64(input_lo, input_hi);
+        return XXH3_avalanche(acc);
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(len <= 16);
+    {   if (XXH_likely(len >  8)) return XXH3_len_9to16_64b(input, len, secret, seed);
+        if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
+        if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
+        return XXH64_avalanche(seed ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
+    }
+}
+
+/*
+ * DISCLAIMER: There are known *seed-dependent* multicollisions here due to
+ * multiplication by zero, affecting hashes of lengths 17 to 240.
+ *
+ * However, they are very unlikely.
+ *
+ * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
+ * unseeded non-cryptographic hashes, it does not attempt to defend itself
+ * against specially crafted inputs, only random inputs.
+ *
+ * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
+ * cancelling out the secret is taken an arbitrary number of times (addressed
+ * in XXH3_accumulate_512), this collision is very unlikely with random inputs
+ * and/or proper seeding:
+ *
+ * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
+ * function that is only called up to 16 times per hash with up to 240 bytes of
+ * input.
  *
- * This is automatically defined on -Os/-Oz on GCC and Clang.
+ * This is not too bad for a non-cryptographic hash function, especially with
+ * only 64 bit outputs.
+ *
+ * The 128-bit variant (which trades some speed for strength) is NOT affected
+ * by this, although it is always a good idea to use a proper seed if you care
+ * about strength.
  */
-#ifndef XXH_REROLL
-#  if defined(__OPTIMIZE_SIZE__)
-#    define XXH_REROLL 1
-#  else
-#    define XXH_REROLL 0
-#  endif
+XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
+                                     const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
+{
+#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__i386__) && defined(__SSE2__)  /* x86 + SSE2 */ \
+  && !defined(XXH_ENABLE_AUTOVECTORIZE)      /* Define to disable like XXH32 hack */
+    /*
+     * UGLY HACK:
+     * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
+     * slower code.
+     *
+     * By forcing seed64 into a register, we disrupt the cost model and
+     * cause it to scalarize. See `XXH32_round()`
+     *
+     * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
+     * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
+     * GCC 9.2, despite both emitting scalar code.
+     *
+     * GCC generates much better scalar code than Clang for the rest of XXH3,
+     * which is why finding a more optimal codepath is an interest.
+     */
+    XXH_COMPILER_GUARD(seed64);
 #endif
+    {   xxh_u64 const input_lo = XXH_readLE64(input);
+        xxh_u64 const input_hi = XXH_readLE64(input+8);
+        return XXH3_mul128_fold64(
+            input_lo ^ (XXH_readLE64(secret)   + seed64),
+            input_hi ^ (XXH_readLE64(secret+8) - seed64)
+        );
+    }
+}
 
+/* For mid range keys, XXH3 uses a Mum-hash variant. */
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                     const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                     XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(16 < len && len <= 128);
 
-/* *************************************
-*  Includes & Memory related functions
-***************************************/
-/*!
- * Modify the local functions below should you wish to use some other memory
- * routines for malloc() and free()
- */
-#include <stdlib.h>
-
-static void* XXH_malloc(size_t s) { return malloc(s); }
-static void XXH_free(void* p) { free(p); }
+    {   xxh_u64 acc = len * XXH_PRIME64_1;
+#if XXH_SIZE_OPT >= 1
+        /* Smaller and cleaner, but slightly slower. */
+        unsigned int i = (unsigned int)(len - 1) / 32;
+        do {
+            acc += XXH3_mix16B(input+16 * i, secret+32*i, seed);
+            acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed);
+        } while (i-- != 0);
+#else
+        if (len > 32) {
+            if (len > 64) {
+                if (len > 96) {
+                    acc += XXH3_mix16B(input+48, secret+96, seed);
+                    acc += XXH3_mix16B(input+len-64, secret+112, seed);
+                }
+                acc += XXH3_mix16B(input+32, secret+64, seed);
+                acc += XXH3_mix16B(input+len-48, secret+80, seed);
+            }
+            acc += XXH3_mix16B(input+16, secret+32, seed);
+            acc += XXH3_mix16B(input+len-32, secret+48, seed);
+        }
+        acc += XXH3_mix16B(input+0, secret+0, seed);
+        acc += XXH3_mix16B(input+len-16, secret+16, seed);
+#endif
+        return XXH3_avalanche(acc);
+    }
+}
 
-/*! and for memcpy() */
-#include <string.h>
-static void* XXH_memcpy(void* dest, const void* src, size_t size)
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                      XXH64_hash_t seed)
 {
-    return memcpy(dest,src,size);
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+    #define XXH3_MIDSIZE_STARTOFFSET 3
+    #define XXH3_MIDSIZE_LASTOFFSET  17
+
+    {   xxh_u64 acc = len * XXH_PRIME64_1;
+        xxh_u64 acc_end;
+        unsigned int const nbRounds = (unsigned int)len / 16;
+        unsigned int i;
+        XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+        for (i=0; i<8; i++) {
+            acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
+        }
+        /* last bytes */
+        acc_end = XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
+        XXH_ASSERT(nbRounds >= 8);
+        acc = XXH3_avalanche(acc);
+#if defined(__clang__)                                /* Clang */ \
+    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
+    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
+        /*
+         * UGLY HACK:
+         * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
+         * In everywhere else, it uses scalar code.
+         *
+         * For 64->128-bit multiplies, even if the NEON was 100% optimal, it
+         * would still be slower than UMAAL (see XXH_mult64to128).
+         *
+         * Unfortunately, Clang doesn't handle the long multiplies properly and
+         * converts them to the nonexistent "vmulq_u64" intrinsic, which is then
+         * scalarized into an ugly mess of VMOV.32 instructions.
+         *
+         * This mess is difficult to avoid without turning autovectorization
+         * off completely, but they are usually relatively minor and/or not
+         * worth it to fix.
+         *
+         * This loop is the easiest to fix, as unlike XXH32, this pragma
+         * _actually works_ because it is a loop vectorization instead of an
+         * SLP vectorization.
+         */
+        #pragma clang loop vectorize(disable)
+#endif
+        for (i=8 ; i < nbRounds; i++) {
+            /*
+             * Prevents clang for unrolling the acc loop and interleaving with this one.
+             */
+            XXH_COMPILER_GUARD(acc);
+            acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
+        }
+        return XXH3_avalanche(acc + acc_end);
+    }
 }
 
-#include <limits.h>   /* ULLONG_MAX */
 
+/* =======     Long Keys     ======= */
 
-/* *************************************
-*  Compiler Specific Options
-***************************************/
-#ifdef _MSC_VER /* Visual Studio warning fix */
-#  pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
+#define XXH_STRIPE_LEN 64
+#define XXH_SECRET_CONSUME_RATE 8   /* nb of secret bytes consumed at each accumulation */
+#define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64))
+
+#ifdef XXH_OLD_NAMES
+#  define STRIPE_LEN XXH_STRIPE_LEN
+#  define ACC_NB XXH_ACC_NB
 #endif
 
-#if XXH_NO_INLINE_HINTS /* disable inlining hints */
-#  define XXH_FORCE_INLINE static
-#  define XXH_NO_INLINE static
-#elif defined(_MSC_VER)    /* Visual Studio */
-#  define XXH_FORCE_INLINE static __forceinline
-#  define XXH_NO_INLINE static __declspec(noinline)
-#else
-#  if defined (__cplusplus) \
-    || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
-#    ifdef __GNUC__
-#      define XXH_FORCE_INLINE static inline __attribute__((always_inline))
-#      define XXH_NO_INLINE static __attribute__((noinline))
+#ifndef XXH_PREFETCH_DIST
+#  ifdef __clang__
+#    define XXH_PREFETCH_DIST 320
+#  else
+#    if (XXH_VECTOR == XXH_AVX512)
+#      define XXH_PREFETCH_DIST 512
 #    else
-#      define XXH_FORCE_INLINE static inline
-#      define XXH_NO_INLINE static
+#      define XXH_PREFETCH_DIST 384
 #    endif
-#  else
-#    define XXH_FORCE_INLINE static
-#    define XXH_NO_INLINE static
-#  endif /* __STDC_VERSION__ */
-#endif
-
+#  endif  /* __clang__ */
+#endif  /* XXH_PREFETCH_DIST */
 
-
-/* *************************************
-*  Debug
-***************************************/
 /*
- * DEBUGLEVEL is expected to be defined externally, typically via the compiler's
- * command line options. The value must be a number.
+ * These macros are to generate an XXH3_accumulate() function.
+ * The two arguments select the name suffix and target attribute.
+ *
+ * The name of this symbol is XXH3_accumulate_<name>() and it calls
+ * XXH3_accumulate_512_<name>().
+ *
+ * It may be useful to hand implement this function if the compiler fails to
+ * optimize the inline function.
  */
-#ifndef DEBUGLEVEL
-#  define DEBUGLEVEL 0
-#endif
-
-#if (DEBUGLEVEL>=1)
-#  include <assert.h>   /* note: can still be disabled with NDEBUG */
-#  define XXH_ASSERT(c)   assert(c)
-#else
-#  define XXH_ASSERT(c)   ((void)0)
-#endif
+#define XXH3_ACCUMULATE_TEMPLATE(name)                      \
+void                                                        \
+XXH3_accumulate_##name(xxh_u64* XXH_RESTRICT acc,           \
+                       const xxh_u8* XXH_RESTRICT input,    \
+                       const xxh_u8* XXH_RESTRICT secret,   \
+                       size_t nbStripes)                    \
+{                                                           \
+    size_t n;                                               \
+    for (n = 0; n < nbStripes; n++ ) {                      \
+        const xxh_u8* const in = input + n*XXH_STRIPE_LEN;  \
+        XXH_PREFETCH(in + XXH_PREFETCH_DIST);               \
+        XXH3_accumulate_512_##name(                         \
+                 acc,                                       \
+                 in,                                        \
+                 secret + n*XXH_SECRET_CONSUME_RATE);       \
+    }                                                       \
+}
 
-/* note: use after variable declarations */
-#define XXH_STATIC_ASSERT(c)  do { enum { XXH_sa = 1/(int)(!!(c)) }; } while (0)
 
+XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
+{
+    if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
+    XXH_memcpy(dst, &v64, sizeof(v64));
+}
 
-/* *************************************
-*  Basic Types
-***************************************/
+/* Several intrinsic functions below are supposed to accept __int64 as argument,
+ * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ .
+ * However, several environments do not define __int64 type,
+ * requiring a workaround.
+ */
 #if !defined (__VMS) \
- && (defined (__cplusplus) \
- || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-# include <stdint.h>
-  typedef uint8_t  xxh_u8;
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+    typedef int64_t xxh_i64;
 #else
-  typedef unsigned char      xxh_u8;
-#endif
-typedef XXH32_hash_t xxh_u32;
-
+    /* the following type must have a width of 64-bit */
+    typedef long long xxh_i64;
+#endif
 
-/* ***   Memory access   *** */
 
-#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
 /*
- * Manual byteshift. Best for old compilers which don't inline memcpy.
- * We actually directly use XXH_readLE32 and XXH_readBE32.
+ * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
+ *
+ * It is a hardened version of UMAC, based off of FARSH's implementation.
+ *
+ * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
+ * implementations, and it is ridiculously fast.
+ *
+ * We harden it by mixing the original input to the accumulators as well as the product.
+ *
+ * This means that in the (relatively likely) case of a multiply by zero, the
+ * original input is preserved.
+ *
+ * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
+ * cross-pollination, as otherwise the upper and lower halves would be
+ * essentially independent.
+ *
+ * This doesn't matter on 64-bit hashes since they all get merged together in
+ * the end, so we skip the extra step.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
  */
-#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
 
-/*
- * Force direct memory access. Only works on CPU which support unaligned memory
- * access in hardware.
- */
-static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }
+#if (XXH_VECTOR == XXH_AVX512) \
+     || (defined(XXH_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0)
 
-#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+#ifndef XXH_TARGET_AVX512
+# define XXH_TARGET_AVX512  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
+                     const void* XXH_RESTRICT input,
+                     const void* XXH_RESTRICT secret)
+{
+    __m512i* const xacc = (__m512i *) acc;
+    XXH_ASSERT((((size_t)acc) & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+
+    {
+        /* data_vec    = input[0]; */
+        __m512i const data_vec    = _mm512_loadu_si512   (input);
+        /* key_vec     = secret[0]; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        /* data_key    = data_vec ^ key_vec; */
+        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
+        /* data_key_lo = data_key >> 32; */
+        __m512i const data_key_lo = _mm512_srli_epi64 (data_key, 32);
+        /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+        __m512i const product     = _mm512_mul_epu32     (data_key, data_key_lo);
+        /* xacc[0] += swap(data_vec); */
+        __m512i const data_swap = _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2));
+        __m512i const sum       = _mm512_add_epi64(*xacc, data_swap);
+        /* xacc[0] += product; */
+        *xacc = _mm512_add_epi64(product, sum);
+    }
+}
+XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512)
 
 /*
- * __pack instructions are safer but compiler specific, hence potentially
- * problematic for some compilers.
+ * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
  *
- * Currently only defined for GCC and ICC.
+ * Multiplication isn't perfect, as explained by Google in HighwayHash:
+ *
+ *  // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to
+ *  // varying degrees. In descending order of goodness, bytes
+ *  // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32.
+ *  // As expected, the upper and lower bytes are much worse.
+ *
+ * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291
+ *
+ * Since our algorithm uses a pseudorandom secret to add some variance into the
+ * mix, we don't need to (or want to) mix as often or as much as HighwayHash does.
+ *
+ * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid
+ * extraction.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
  */
-typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
-static xxh_u32 XXH_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
 
-#else
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+    {   __m512i* const xacc = (__m512i*) acc;
+        const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1);
+
+        /* xacc[0] ^= (xacc[0] >> 47) */
+        __m512i const acc_vec     = *xacc;
+        __m512i const shifted     = _mm512_srli_epi64    (acc_vec, 47);
+        /* xacc[0] ^= secret; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        __m512i const data_key    = _mm512_ternarylogic_epi32(key_vec, acc_vec, shifted, 0x96 /* key_vec ^ acc_vec ^ shifted */);
+
+        /* xacc[0] *= XXH_PRIME32_1; */
+        __m512i const data_key_hi = _mm512_srli_epi64 (data_key, 32);
+        __m512i const prod_lo     = _mm512_mul_epu32     (data_key, prime32);
+        __m512i const prod_hi     = _mm512_mul_epu32     (data_key_hi, prime32);
+        *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32));
+    }
+}
 
-/*
- * Portable and safe solution. Generally efficient.
- * see: https://stackoverflow.com/a/32095106/646947
- */
-static xxh_u32 XXH_read32(const void* memPtr)
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
 {
-    xxh_u32 val;
-    memcpy(&val, memPtr, sizeof(val));
-    return val;
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64);
+    XXH_ASSERT(((size_t)customSecret & 63) == 0);
+    (void)(&XXH_writeLE64);
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i);
+        __m512i const seed_pos = _mm512_set1_epi64((xxh_i64)seed64);
+        __m512i const seed     = _mm512_mask_sub_epi64(seed_pos, 0xAA, _mm512_set1_epi8(0), seed_pos);
+
+        const __m512i* const src  = (const __m512i*) ((const void*) XXH3_kSecret);
+              __m512i* const dest = (      __m512i*) customSecret;
+        int i;
+        XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dest & 63) == 0);
+        for (i=0; i < nbRounds; ++i) {
+            dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed);
+    }   }
 }
 
-#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+#endif
 
+#if (XXH_VECTOR == XXH_AVX2) \
+    || (defined(XXH_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0)
 
-/* ***   Endianess   *** */
-typedef enum { XXH_bigEndian=0, XXH_littleEndian=1 } XXH_endianess;
+#ifndef XXH_TARGET_AVX2
+# define XXH_TARGET_AVX2  /* disable attribute target */
+#endif
 
-/*!
- * XXH_CPU_LITTLE_ENDIAN:
- * Defined to 1 if the target is little endian, or 0 if it is big endian.
- * It can be defined externally, for example on the compiler command line.
- *
- * If it is not defined, a runtime check (which is usually constant folded)
- * is used instead.
- */
-#ifndef XXH_CPU_LITTLE_ENDIAN
-/*
- * Try to detect endianness automatically, to avoid the nonstandard behavior
- * in `XXH_isLittleEndian()`
- */
-#  if defined(_WIN32) /* Windows is always little endian */ \
-     || defined(__LITTLE_ENDIAN__) \
-     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
-#    define XXH_CPU_LITTLE_ENDIAN 1
-#  elif defined(__BIG_ENDIAN__) \
-     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
-#    define XXH_CPU_LITTLE_ENDIAN 0
-#  else
-static int XXH_isLittleEndian(void)
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
 {
-    /*
-     * Nonstandard, but well-defined behavior in practice.
-     * Don't use static: it is detrimental to performance.
-     */
-    const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };
-    return one.c[0];
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   __m256i* const xacc    =       (__m256i *) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires  a const __m256i * pointer for some reason. */
+        const         __m256i* const xinput  = (const __m256i *) input;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+        const         __m256i* const xsecret = (const __m256i *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+            /* data_vec    = xinput[i]; */
+            __m256i const data_vec    = _mm256_loadu_si256    (xinput+i);
+            /* key_vec     = xsecret[i]; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            /* data_key    = data_vec ^ key_vec; */
+            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
+            /* data_key_lo = data_key >> 32; */
+            __m256i const data_key_lo = _mm256_srli_epi64 (data_key, 32);
+            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+            __m256i const product     = _mm256_mul_epu32     (data_key, data_key_lo);
+            /* xacc[i] += swap(data_vec); */
+            __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
+            __m256i const sum       = _mm256_add_epi64(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = _mm256_add_epi64(product, sum);
+    }   }
 }
-#   define XXH_CPU_LITTLE_ENDIAN   XXH_isLittleEndian()
-#  endif
-#endif
+XXH_FORCE_INLINE XXH_TARGET_AVX2 XXH3_ACCUMULATE_TEMPLATE(avx2)
 
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   __m256i* const xacc = (__m256i*) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+        const         __m256i* const xsecret = (const __m256i *) secret;
+        const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m256i const acc_vec     = xacc[i];
+            __m256i const shifted     = _mm256_srli_epi64    (acc_vec, 47);
+            __m256i const data_vec    = _mm256_xor_si256     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_1; */
+            __m256i const data_key_hi = _mm256_srli_epi64 (data_key, 32);
+            __m256i const prod_lo     = _mm256_mul_epu32     (data_key, prime32);
+            __m256i const prod_hi     = _mm256_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
+        }
+    }
+}
 
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0);
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6);
+    XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64);
+    (void)(&XXH_writeLE64);
+    XXH_PREFETCH(customSecret);
+    {   __m256i const seed = _mm256_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64);
+
+        const __m256i* const src  = (const __m256i*) ((const void*) XXH3_kSecret);
+              __m256i*       dest = (      __m256i*) customSecret;
+
+#       if defined(__GNUC__) || defined(__clang__)
+        /*
+         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+         *   - do not extract the secret from sse registers in the internal loop
+         *   - use less common registers, and avoid pushing these reg into stack
+         */
+        XXH_COMPILER_GUARD(dest);
+#       endif
+        XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dest & 31) == 0);
+
+        /* GCC -O2 need unroll loop manually */
+        dest[0] = _mm256_add_epi64(_mm256_load_si256(src+0), seed);
+        dest[1] = _mm256_add_epi64(_mm256_load_si256(src+1), seed);
+        dest[2] = _mm256_add_epi64(_mm256_load_si256(src+2), seed);
+        dest[3] = _mm256_add_epi64(_mm256_load_si256(src+3), seed);
+        dest[4] = _mm256_add_epi64(_mm256_load_si256(src+4), seed);
+        dest[5] = _mm256_add_epi64(_mm256_load_si256(src+5), seed);
+    }
+}
 
+#endif
 
-/* ****************************************
-*  Compiler-specific Functions and Macros
-******************************************/
-#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+/* x86dispatch always generates SSE2 */
+#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH)
 
-#ifndef __has_builtin
-#  define __has_builtin(x) 0
+#ifndef XXH_TARGET_SSE2
+# define XXH_TARGET_SSE2  /* disable attribute target */
 #endif
 
-#if !defined(NO_CLANG_BUILTIN) && __has_builtin(__builtin_rotateleft32) \
-                               && __has_builtin(__builtin_rotateleft64)
-#  define XXH_rotl32 __builtin_rotateleft32
-#  define XXH_rotl64 __builtin_rotateleft64
-/* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */
-#elif defined(_MSC_VER)
-#  define XXH_rotl32(x,r) _rotl(x,r)
-#  define XXH_rotl64(x,r) _rotl64(x,r)
-#else
-#  define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
-#  define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
-#endif
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    /* SSE2 is just a half-scale version of the AVX2 version. */
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   __m128i* const xacc    =       (__m128i *) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xinput  = (const __m128i *) input;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xsecret = (const __m128i *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+            /* data_vec    = xinput[i]; */
+            __m128i const data_vec    = _mm_loadu_si128   (xinput+i);
+            /* key_vec     = xsecret[i]; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            /* data_key    = data_vec ^ key_vec; */
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+            /* data_key_lo = data_key >> 32; */
+            __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+            __m128i const product     = _mm_mul_epu32     (data_key, data_key_lo);
+            /* xacc[i] += swap(data_vec); */
+            __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
+            __m128i const sum       = _mm_add_epi64(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = _mm_add_epi64(product, sum);
+    }   }
+}
+XXH_FORCE_INLINE XXH_TARGET_SSE2 XXH3_ACCUMULATE_TEMPLATE(sse2)
 
-#if defined(_MSC_VER)     /* Visual Studio */
-#  define XXH_swap32 _byteswap_ulong
-#elif XXH_GCC_VERSION >= 403
-#  define XXH_swap32 __builtin_bswap32
-#else
-static xxh_u32 XXH_swap32 (xxh_u32 x)
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 {
-    return  ((x << 24) & 0xff000000 ) |
-            ((x <<  8) & 0x00ff0000 ) |
-            ((x >>  8) & 0x0000ff00 ) |
-            ((x >> 24) & 0x000000ff );
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   __m128i* const xacc = (__m128i*) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xsecret = (const __m128i *) secret;
+        const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m128i const acc_vec     = xacc[i];
+            __m128i const shifted     = _mm_srli_epi64    (acc_vec, 47);
+            __m128i const data_vec    = _mm_xor_si128     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret[i]; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_1; */
+            __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __m128i const prod_lo     = _mm_mul_epu32     (data_key, prime32);
+            __m128i const prod_hi     = _mm_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+    (void)(&XXH_writeLE64);
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i);
+
+#       if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900
+        /* MSVC 32bit mode does not support _mm_set_epi64x before 2015 */
+        XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, (xxh_i64)(0U - seed64) };
+        __m128i const seed = _mm_load_si128((__m128i const*)seed64x2);
+#       else
+        __m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64);
+#       endif
+        int i;
+
+        const void* const src16 = XXH3_kSecret;
+        __m128i* dst16 = (__m128i*) customSecret;
+#       if defined(__GNUC__) || defined(__clang__)
+        /*
+         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+         *   - do not extract the secret from sse registers in the internal loop
+         *   - use less common registers, and avoid pushing these reg into stack
+         */
+        XXH_COMPILER_GUARD(dst16);
+#       endif
+        XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dst16 & 15) == 0);
+
+        for (i=0; i < nbRounds; ++i) {
+            dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed);
+    }   }
 }
+
 #endif
 
+#if (XXH_VECTOR == XXH_NEON)
 
-/* ***************************
-*  Memory reads
-*****************************/
-typedef enum { XXH_aligned, XXH_unaligned } XXH_alignment;
+/* forward declarations for the scalar routines */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc, void const* XXH_RESTRICT input,
+                 void const* XXH_RESTRICT secret, size_t lane);
 
-/*
- * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load.
+XXH_FORCE_INLINE void
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+                         void const* XXH_RESTRICT secret, size_t lane);
+
+/*!
+ * @internal
+ * @brief The bulk processing loop for NEON and WASM SIMD128.
  *
- * This is ideal for older compilers which don't inline memcpy.
+ * The NEON code path is actually partially scalar when running on AArch64. This
+ * is to optimize the pipelining and can have up to 15% speedup depending on the
+ * CPU, and it also mitigates some GCC codegen issues.
+ *
+ * @see XXH3_NEON_LANES for configuring this and details about this optimization.
+ *
+ * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit
+ * integers instead of the other platforms which mask full 64-bit vectors,
+ * so the setup is more complicated than just shifting right.
+ *
+ * Additionally, there is an optimization for 4 lanes at once noted below.
+ *
+ * Since, as stated, the most optimal amount of lanes for Cortexes is 6,
+ * there needs to be *three* versions of the accumulate operation used
+ * for the remaining 2 lanes.
+ *
+ * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap
+ * nearly perfectly.
  */
-#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
 
-XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr)
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
 {
-    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
-    return bytePtr[0]
-         | ((xxh_u32)bytePtr[1] << 8)
-         | ((xxh_u32)bytePtr[2] << 16)
-         | ((xxh_u32)bytePtr[3] << 24);
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    XXH_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0);
+    {   /* GCC for darwin arm64 does not like aliasing here */
+        xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc;
+        /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
+        uint8_t const* xinput = (const uint8_t *) input;
+        uint8_t const* xsecret  = (const uint8_t *) secret;
+
+        size_t i;
+#ifdef __wasm_simd128__
+        /*
+         * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret
+         * is constant propagated, which results in it converting it to this
+         * inside the loop:
+         *
+         *    a = v128.load(XXH3_kSecret +  0 + $secret_offset, offset = 0)
+         *    b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0)
+         *    ...
+         *
+         * This requires a full 32-bit address immediate (and therefore a 6 byte
+         * instruction) as well as an add for each offset.
+         *
+         * Putting an asm guard prevents it from folding (at the cost of losing
+         * the alignment hint), and uses the free offset in `v128.load` instead
+         * of adding secret_offset each time which overall reduces code size by
+         * about a kilobyte and improves performance.
+         */
+        XXH_COMPILER_GUARD(xsecret);
+#endif
+        /* Scalar lanes use the normal scalarRound routine */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarRound(acc, input, secret, i);
+        }
+        i = 0;
+        /* 4 NEON lanes at a time. */
+        for (; i+1 < XXH3_NEON_LANES / 2; i+=2) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput  + (i * 16));
+            uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput  + ((i+1) * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec_1  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t key_vec_2  = XXH_vld1q_u64(xsecret + ((i+1) * 16));
+            /* data_swap = swap(data_vec) */
+            uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1);
+            uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1);
+            uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2);
+
+            /*
+             * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a
+             * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to
+             * get one vector with the low 32 bits of each lane, and one vector
+             * with the high 32 bits of each lane.
+             *
+             * The intrinsic returns a double vector because the original ARMv7-a
+             * instruction modified both arguments in place. AArch64 and SIMD128 emit
+             * two instructions from this intrinsic.
+             *
+             *  [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ]
+             *  [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
+             */
+            uint32x4x2_t unzipped = vuzpq_u32(
+                vreinterpretq_u32_u64(data_key_1),
+                vreinterpretq_u32_u64(data_key_2)
+            );
+            /* data_key_lo = data_key & 0xFFFFFFFF */
+            uint32x4_t data_key_lo = unzipped.val[0];
+            /* data_key_hi = data_key >> 32 */
+            uint32x4_t data_key_hi = unzipped.val[1];
+            /*
+             * Then, we can split the vectors horizontally and multiply which, as for most
+             * widening intrinsics, have a variant that works on both high half vectors
+             * for free on AArch64. A similar instruction is available on SIMD128.
+             *
+             * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
+             */
+            uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi);
+            uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi);
+            /*
+             * Clang reorders
+             *    a += b * c;     // umlal   swap.2d, dkl.2s, dkh.2s
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             * to
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             *    c += b * c;     // umlal   acc.2d, dkl.2s, dkh.2s
+             *
+             * While it would make sense in theory since the addition is faster,
+             * for reasons likely related to umlal being limited to certain NEON
+             * pipelines, this is worse. A compiler guard fixes this.
+             */
+            XXH_COMPILER_GUARD_CLANG_NEON(sum_1);
+            XXH_COMPILER_GUARD_CLANG_NEON(sum_2);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i]   = vaddq_u64(xacc[i], sum_1);
+            xacc[i+1] = vaddq_u64(xacc[i+1], sum_2);
+        }
+        /* Operate on the remaining NEON lanes 2 at a time. */
+        for (; i < XXH3_NEON_LANES / 2; i++) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec = XXH_vld1q_u64(xinput  + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            /* acc_vec_2 = swap(data_vec) */
+            uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+            /* For two lanes, just use VMOVN and VSHRN. */
+            /* data_key_lo = data_key & 0xFFFFFFFF; */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* data_key_hi = data_key >> 32; */
+            uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32);
+            /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */
+            uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi);
+            /* Same Clang workaround as before */
+            XXH_COMPILER_GUARD_CLANG_NEON(sum);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i] = vaddq_u64 (xacc[i], sum);
+        }
+    }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)
 
-XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr)
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 {
-    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
-    return bytePtr[3]
-         | ((xxh_u32)bytePtr[2] << 8)
-         | ((xxh_u32)bytePtr[1] << 16)
-         | ((xxh_u32)bytePtr[0] << 24);
-}
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   xxh_aliasing_uint64x2_t* xacc       = (xxh_aliasing_uint64x2_t*) acc;
+        uint8_t const* xsecret = (uint8_t const*) secret;
+
+        size_t i;
+        /* WASM uses operator overloads and doesn't need these. */
+#ifndef __wasm_simd128__
+        /* { prime32_1, prime32_1 } */
+        uint32x2_t const kPrimeLo = vdup_n_u32(XXH_PRIME32_1);
+        /* { 0, prime32_1, 0, prime32_1 } */
+        uint32x4_t const kPrimeHi = vreinterpretq_u32_u64(vdupq_n_u64((xxh_u64)XXH_PRIME32_1 << 32));
+#endif
 
+        /* AArch64 uses both scalar and neon at the same time */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarScrambleRound(acc, secret, i);
+        }
+        for (i=0; i < XXH3_NEON_LANES / 2; i++) {
+            /* xacc[i] ^= (xacc[i] >> 47); */
+            uint64x2_t acc_vec  = xacc[i];
+            uint64x2_t shifted  = vshrq_n_u64(acc_vec, 47);
+            uint64x2_t data_vec = veorq_u64(acc_vec, shifted);
+
+            /* xacc[i] ^= xsecret[i]; */
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+            /* xacc[i] *= XXH_PRIME32_1 */
+#ifdef __wasm_simd128__
+            /* SIMD128 has multiply by u64x2, use it instead of expanding and scalarizing */
+            xacc[i] = data_key * XXH_PRIME32_1;
 #else
-XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr)
+            /*
+             * Expanded version with portable NEON intrinsics
+             *
+             *    lo(x) * lo(y) + (hi(x) * lo(y) << 32)
+             *
+             * prod_hi = hi(data_key) * lo(prime) << 32
+             *
+             * Since we only need 32 bits of this multiply a trick can be used, reinterpreting the vector
+             * as a uint32x4_t and multiplying by { 0, prime, 0, prime } to cancel out the unwanted bits
+             * and avoid the shift.
+             */
+            uint32x4_t prod_hi = vmulq_u32 (vreinterpretq_u32_u64(data_key), kPrimeHi);
+            /* Extract low bits for vmlal_u32  */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* xacc[i] = prod_hi + lo(data_key) * XXH_PRIME32_1; */
+            xacc[i] = vmlal_u32(vreinterpretq_u64_u32(prod_hi), data_key_lo, kPrimeLo);
+#endif
+        }
+    }
+}
+#endif
+
+#if (XXH_VECTOR == XXH_VSX)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
 {
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
+    /* presumed aligned */
+    xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
+    xxh_u8 const* const xinput   = (xxh_u8 const*) input;   /* no alignment restriction */
+    xxh_u8 const* const xsecret  = (xxh_u8 const*) secret;    /* no alignment restriction */
+    xxh_u64x2 const v32 = { 32, 32 };
+    size_t i;
+    for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+        /* data_vec = xinput[i]; */
+        xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + 16*i);
+        /* key_vec = xsecret[i]; */
+        xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
+        xxh_u64x2 const data_key = data_vec ^ key_vec;
+        /* shuffled = (data_key << 32) | (data_key >> 32); */
+        xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32);
+        /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */
+        xxh_u64x2 const product  = XXH_vec_mulo((xxh_u32x4)data_key, shuffled);
+        /* acc_vec = xacc[i]; */
+        xxh_u64x2 acc_vec        = xacc[i];
+        acc_vec += product;
+
+        /* swap high and low halves */
+#ifdef __s390x__
+        acc_vec += vec_permi(data_vec, data_vec, 2);
+#else
+        acc_vec += vec_xxpermdi(data_vec, data_vec, 2);
+#endif
+        xacc[i] = acc_vec;
+    }
 }
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(vsx)
 
-static xxh_u32 XXH_readBE32(const void* ptr)
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 {
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
+        const xxh_u8* const xsecret = (const xxh_u8*) secret;
+        /* constants */
+        xxh_u64x2 const v32  = { 32, 32 };
+        xxh_u64x2 const v47 = { 47, 47 };
+        xxh_u32x4 const prime = { XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1 };
+        size_t i;
+        for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47); */
+            xxh_u64x2 const acc_vec  = xacc[i];
+            xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47);
+
+            /* xacc[i] ^= xsecret[i]; */
+            xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
+            xxh_u64x2 const data_key = data_vec ^ key_vec;
+
+            /* xacc[i] *= XXH_PRIME32_1 */
+            /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF);  */
+            xxh_u64x2 const prod_even  = XXH_vec_mule((xxh_u32x4)data_key, prime);
+            /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32);  */
+            xxh_u64x2 const prod_odd  = XXH_vec_mulo((xxh_u32x4)data_key, prime);
+            xacc[i] = prod_odd + (prod_even << v32);
+    }   }
 }
+
 #endif
 
-XXH_FORCE_INLINE xxh_u32
-XXH_readLE32_align(const void* ptr, XXH_alignment align)
+#if (XXH_VECTOR == XXH_SVE)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_sve( void* XXH_RESTRICT acc,
+                   const void* XXH_RESTRICT input,
+                   const void* XXH_RESTRICT secret)
 {
-    if (align==XXH_unaligned) {
-        return XXH_readLE32(ptr);
+    uint64_t *xacc = (uint64_t *)acc;
+    const uint64_t *xinput = (const uint64_t *)(const void *)input;
+    const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
+    svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
+    uint64_t element_count = svcntd();
+    if (element_count >= 8) {
+        svbool_t mask = svptrue_pat_b64(SV_VL8);
+        svuint64_t vacc = svld1_u64(mask, xacc);
+        ACCRND(vacc, 0);
+        svst1_u64(mask, xacc, vacc);
+    } else if (element_count == 2) {   /* sve128 */
+        svbool_t mask = svptrue_pat_b64(SV_VL2);
+        svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+        svuint64_t acc1 = svld1_u64(mask, xacc + 2);
+        svuint64_t acc2 = svld1_u64(mask, xacc + 4);
+        svuint64_t acc3 = svld1_u64(mask, xacc + 6);
+        ACCRND(acc0, 0);
+        ACCRND(acc1, 2);
+        ACCRND(acc2, 4);
+        ACCRND(acc3, 6);
+        svst1_u64(mask, xacc + 0, acc0);
+        svst1_u64(mask, xacc + 2, acc1);
+        svst1_u64(mask, xacc + 4, acc2);
+        svst1_u64(mask, xacc + 6, acc3);
     } else {
-        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr);
+        svbool_t mask = svptrue_pat_b64(SV_VL4);
+        svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+        svuint64_t acc1 = svld1_u64(mask, xacc + 4);
+        ACCRND(acc0, 0);
+        ACCRND(acc1, 4);
+        svst1_u64(mask, xacc + 0, acc0);
+        svst1_u64(mask, xacc + 4, acc1);
     }
 }
 
+XXH_FORCE_INLINE void
+XXH3_accumulate_sve(xxh_u64* XXH_RESTRICT acc,
+               const xxh_u8* XXH_RESTRICT input,
+               const xxh_u8* XXH_RESTRICT secret,
+               size_t nbStripes)
+{
+    if (nbStripes != 0) {
+        uint64_t *xacc = (uint64_t *)acc;
+        const uint64_t *xinput = (const uint64_t *)(const void *)input;
+        const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
+        svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
+        uint64_t element_count = svcntd();
+        if (element_count >= 8) {
+            svbool_t mask = svptrue_pat_b64(SV_VL8);
+            svuint64_t vacc = svld1_u64(mask, xacc + 0);
+            do {
+                /* svprfd(svbool_t, void *, enum svfprop); */
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(vacc, 0);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, vacc);
+        } else if (element_count == 2) { /* sve128 */
+            svbool_t mask = svptrue_pat_b64(SV_VL2);
+            svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+            svuint64_t acc1 = svld1_u64(mask, xacc + 2);
+            svuint64_t acc2 = svld1_u64(mask, xacc + 4);
+            svuint64_t acc3 = svld1_u64(mask, xacc + 6);
+            do {
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(acc0, 0);
+                ACCRND(acc1, 2);
+                ACCRND(acc2, 4);
+                ACCRND(acc3, 6);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, acc0);
+           svst1_u64(mask, xacc + 2, acc1);
+           svst1_u64(mask, xacc + 4, acc2);
+           svst1_u64(mask, xacc + 6, acc3);
+        } else {
+            svbool_t mask = svptrue_pat_b64(SV_VL4);
+            svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+            svuint64_t acc1 = svld1_u64(mask, xacc + 4);
+            do {
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(acc0, 0);
+                ACCRND(acc1, 4);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, acc0);
+           svst1_u64(mask, xacc + 4, acc1);
+       }
+    }
+}
 
-/* *************************************
-*  Misc
-***************************************/
-XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
-
+#endif
 
-/* *******************************************************************
-*  32-bit hash functions
-*********************************************************************/
-static const xxh_u32 PRIME32_1 = 0x9E3779B1U;   /* 0b10011110001101110111100110110001 */
-static const xxh_u32 PRIME32_2 = 0x85EBCA77U;   /* 0b10000101111010111100101001110111 */
-static const xxh_u32 PRIME32_3 = 0xC2B2AE3DU;   /* 0b11000010101100101010111000111101 */
-static const xxh_u32 PRIME32_4 = 0x27D4EB2FU;   /* 0b00100111110101001110101100101111 */
-static const xxh_u32 PRIME32_5 = 0x165667B1U;   /* 0b00010110010101100110011110110001 */
+/* scalar variants - universal */
 
-static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
+#if defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__))
+/*
+ * In XXH3_scalarRound(), GCC and Clang have a similar codegen issue, where they
+ * emit an excess mask and a full 64-bit multiply-add (MADD X-form).
+ *
+ * While this might not seem like much, as AArch64 is a 64-bit architecture, only
+ * big Cortex designs have a full 64-bit multiplier.
+ *
+ * On the little cores, the smaller 32-bit multiplier is used, and full 64-bit
+ * multiplies expand to 2-3 multiplies in microcode. This has a major penalty
+ * of up to 4 latency cycles and 2 stall cycles in the multiply pipeline.
+ *
+ * Thankfully, AArch64 still provides the 32-bit long multiply-add (UMADDL) which does
+ * not have this penalty and does the mask automatically.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
 {
-    acc += input * PRIME32_2;
-    acc  = XXH_rotl32(acc, 13);
-    acc *= PRIME32_1;
-#if defined(__GNUC__) && defined(__SSE4_1__) && !defined(XXH_ENABLE_AUTOVECTORIZE)
-    /*
-     * UGLY HACK:
-     * This inline assembly hack forces acc into a normal register. This is the
-     * only thing that prevents GCC and Clang from autovectorizing the XXH32
-     * loop (pragmas and attributes don't work for some resason) without globally
-     * disabling SSE4.1.
-     *
-     * The reason we want to avoid vectorization is because despite working on
-     * 4 integers at a time, there are multiple factors slowing XXH32 down on
-     * SSE4:
-     * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on
-     *   newer chips!) making it slightly slower to multiply four integers at
-     *   once compared to four integers independently. Even when pmulld was
-     *   fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE
-     *   just to multiply unless doing a long operation.
-     *
-     * - Four instructions are required to rotate,
-     *      movqda tmp,  v // not required with VEX encoding
-     *      pslld  tmp, 13 // tmp <<= 13
-     *      psrld  v,   19 // x >>= 19
-     *      por    v,  tmp // x |= tmp
-     *   compared to one for scalar:
-     *      roll   v, 13    // reliably fast across the board
-     *      shldl  v, v, 13 // Sandy Bridge and later prefer this for some reason
-     *
-     * - Instruction level parallelism is actually more beneficial here because
-     *   the SIMD actually serializes this operation: While v1 is rotating, v2
-     *   can load data, while v3 can multiply. SSE forces them to operate
-     *   together.
-     *
-     * How this hack works:
-     * __asm__(""       // Declare an assembly block but don't declare any instructions
-     *          :       // However, as an Input/Output Operand,
-     *          "+r"    // constrain a read/write operand (+) as a general purpose register (r).
-     *          (acc)   // and set acc as the operand
-     * );
-     *
-     * Because of the 'r', the compiler has promised that seed will be in a
-     * general purpose register and the '+' says that it will be 'read/write',
-     * so it has to assume it has changed. It is like volatile without all the
-     * loads and stores.
-     *
-     * Since the argument has to be in a normal register (not an SSE register),
-     * each time XXH32_round is called, it is impossible to vectorize.
-     */
-    __asm__("" : "+r" (acc));
-#endif
-    return acc;
+    xxh_u64 ret;
+    /* note: %x = 64-bit register, %w = 32-bit register */
+    __asm__("umaddl %x0, %w1, %w2, %x3" : "=r" (ret) : "r" (lhs), "r" (rhs), "r" (acc));
+    return ret;
 }
-
-/* mix all bits */
-static xxh_u32 XXH32_avalanche(xxh_u32 h32)
+#else
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
 {
-    h32 ^= h32 >> 15;
-    h32 *= PRIME32_2;
-    h32 ^= h32 >> 13;
-    h32 *= PRIME32_3;
-    h32 ^= h32 >> 16;
-    return(h32);
+    return XXH_mult32to64((xxh_u32)lhs, (xxh_u32)rhs) + acc;
 }
+#endif
 
-#define XXH_get32bits(p) XXH_readLE32_align(p, align)
-
-static xxh_u32
-XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
+/*!
+ * @internal
+ * @brief Scalar round for @ref XXH3_accumulate_512_scalar().
+ *
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
+ */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc,
+                 void const* XXH_RESTRICT input,
+                 void const* XXH_RESTRICT secret,
+                 size_t lane)
 {
-#define PROCESS1 do {                           \
-    h32 += (*ptr++) * PRIME32_5;                \
-    h32 = XXH_rotl32(h32, 11) * PRIME32_1;      \
-} while (0)
+    xxh_u64* xacc = (xxh_u64*) acc;
+    xxh_u8 const* xinput  = (xxh_u8 const*) input;
+    xxh_u8 const* xsecret = (xxh_u8 const*) secret;
+    XXH_ASSERT(lane < XXH_ACC_NB);
+    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
+    {
+        xxh_u64 const data_val = XXH_readLE64(xinput + lane * 8);
+        xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + lane * 8);
+        xacc[lane ^ 1] += data_val; /* swap adjacent lanes */
+        xacc[lane] = XXH_mult32to64_add64(data_key /* & 0xFFFFFFFF */, data_key >> 32, xacc[lane]);
+    }
+}
 
-#define PROCESS4 do {                           \
-    h32 += XXH_get32bits(ptr) * PRIME32_3;      \
-    ptr += 4;                                   \
-    h32  = XXH_rotl32(h32, 17) * PRIME32_4;     \
-} while (0)
+/*!
+ * @internal
+ * @brief Processes a 64 byte block of data using the scalar path.
+ */
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc,
+                     const void* XXH_RESTRICT input,
+                     const void* XXH_RESTRICT secret)
+{
+    size_t i;
+    /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */
+#if defined(__GNUC__) && !defined(__clang__) \
+  && (defined(__arm__) || defined(__thumb2__)) \
+  && defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \
+  && XXH_SIZE_OPT <= 0
+#  pragma GCC unroll 8
+#endif
+    for (i=0; i < XXH_ACC_NB; i++) {
+        XXH3_scalarRound(acc, input, secret, i);
+    }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar)
 
-    /* Compact rerolled version */
-    if (XXH_REROLL) {
-        len &= 15;
-        while (len >= 4) {
-            PROCESS4;
-            len -= 4;
-        }
-        while (len > 0) {
-            PROCESS1;
-            --len;
-        }
-        return XXH32_avalanche(h32);
-    } else {
-         switch(len&15) /* or switch(bEnd - p) */ {
-           case 12:      PROCESS4;
-                         /* fallthrough */
-           case 8:       PROCESS4;
-                         /* fallthrough */
-           case 4:       PROCESS4;
-                         return XXH32_avalanche(h32);
-
-           case 13:      PROCESS4;
-                         /* fallthrough */
-           case 9:       PROCESS4;
-                         /* fallthrough */
-           case 5:       PROCESS4;
-                         PROCESS1;
-                         return XXH32_avalanche(h32);
-
-           case 14:      PROCESS4;
-                         /* fallthrough */
-           case 10:      PROCESS4;
-                         /* fallthrough */
-           case 6:       PROCESS4;
-                         PROCESS1;
-                         PROCESS1;
-                         return XXH32_avalanche(h32);
-
-           case 15:      PROCESS4;
-                         /* fallthrough */
-           case 11:      PROCESS4;
-                         /* fallthrough */
-           case 7:       PROCESS4;
-                         /* fallthrough */
-           case 3:       PROCESS1;
-                         /* fallthrough */
-           case 2:       PROCESS1;
-                         /* fallthrough */
-           case 1:       PROCESS1;
-                         /* fallthrough */
-           case 0:       return XXH32_avalanche(h32);
-        }
-        XXH_ASSERT(0);
-        return h32;   /* reaching this point is deemed impossible */
+/*!
+ * @internal
+ * @brief Scalar scramble step for @ref XXH3_scrambleAcc_scalar().
+ *
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
+ */
+XXH_FORCE_INLINE void
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+                         void const* XXH_RESTRICT secret,
+                         size_t lane)
+{
+    xxh_u64* const xacc = (xxh_u64*) acc;   /* presumed aligned */
+    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
+    XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
+    XXH_ASSERT(lane < XXH_ACC_NB);
+    {
+        xxh_u64 const key64 = XXH_readLE64(xsecret + lane * 8);
+        xxh_u64 acc64 = xacc[lane];
+        acc64 = XXH_xorshift64(acc64, 47);
+        acc64 ^= key64;
+        acc64 *= XXH_PRIME32_1;
+        xacc[lane] = acc64;
     }
 }
 
-XXH_FORCE_INLINE xxh_u32
-XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
+/*!
+ * @internal
+ * @brief Scrambles the accumulators after a large chunk has been read
+ */
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
 {
-    const xxh_u8* bEnd = input + len;
-    xxh_u32 h32;
-
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-    if (input==NULL) {
-        len=0;
-        bEnd=input=(const xxh_u8*)(size_t)16;
+    size_t i;
+    for (i=0; i < XXH_ACC_NB; i++) {
+        XXH3_scalarScrambleRound(acc, secret, i);
     }
+}
+
+XXH_FORCE_INLINE void
+XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    /*
+     * We need a separate pointer for the hack below,
+     * which requires a non-const pointer.
+     * Any decent compiler will optimize this out otherwise.
+     */
+    const xxh_u8* kSecretPtr = XXH3_kSecret;
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+
+#if defined(__GNUC__) && defined(__aarch64__)
+    /*
+     * UGLY HACK:
+     * GCC and Clang generate a bunch of MOV/MOVK pairs for aarch64, and they are
+     * placed sequentially, in order, at the top of the unrolled loop.
+     *
+     * While MOVK is great for generating constants (2 cycles for a 64-bit
+     * constant compared to 4 cycles for LDR), it fights for bandwidth with
+     * the arithmetic instructions.
+     *
+     *   I   L   S
+     * MOVK
+     * MOVK
+     * MOVK
+     * MOVK
+     * ADD
+     * SUB      STR
+     *          STR
+     * By forcing loads from memory (as the asm line causes the compiler to assume
+     * that XXH3_kSecretPtr has been changed), the pipelines are used more
+     * efficiently:
+     *   I   L   S
+     *      LDR
+     *  ADD LDR
+     *  SUB     STR
+     *          STR
+     *
+     * See XXH3_NEON_LANES for details on the pipsline.
+     *
+     * XXH3_64bits_withSeed, len == 256, Snapdragon 835
+     *   without hack: 2654.4 MB/s
+     *   with hack:    3202.9 MB/s
+     */
+    XXH_COMPILER_GUARD(kSecretPtr);
 #endif
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
+        int i;
+        for (i=0; i < nbRounds; i++) {
+            /*
+             * The asm hack causes the compiler to assume that kSecretPtr aliases with
+             * customSecret, and on aarch64, this prevented LDP from merging two
+             * loads together for free. Putting the loads together before the stores
+             * properly generates LDP.
+             */
+            xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i)     + seed64;
+            xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
+            XXH_writeLE64((xxh_u8*)customSecret + 16*i,     lo);
+            XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi);
+    }   }
+}
 
-    if (len>=16) {
-        const xxh_u8* const limit = bEnd - 15;
-        xxh_u32 v1 = seed + PRIME32_1 + PRIME32_2;
-        xxh_u32 v2 = seed + PRIME32_2;
-        xxh_u32 v3 = seed + 0;
-        xxh_u32 v4 = seed - PRIME32_1;
 
-        do {
-            v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4;
-            v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4;
-            v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4;
-            v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4;
-        } while (input < limit);
+typedef void (*XXH3_f_accumulate)(xxh_u64* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, size_t);
+typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*);
+typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
 
-        h32 = XXH_rotl32(v1, 1)  + XXH_rotl32(v2, 7)
-            + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
-    } else {
-        h32  = seed + PRIME32_5;
-    }
 
-    h32 += (xxh_u32)len;
+#if (XXH_VECTOR == XXH_AVX512)
 
-    return XXH32_finalize(h32, input, len&15, align);
-}
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx512
+#define XXH3_accumulate     XXH3_accumulate_avx512
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_avx512
+#define XXH3_initCustomSecret XXH3_initCustomSecret_avx512
 
+#elif (XXH_VECTOR == XXH_AVX2)
 
-XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
-{
-#if 0
-    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
-    XXH32_state_t state;
-    XXH32_reset(&state, seed);
-    XXH32_update(&state, (const xxh_u8*)input, len);
-    return XXH32_digest(&state);
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx2
+#define XXH3_accumulate     XXH3_accumulate_avx2
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_avx2
+#define XXH3_initCustomSecret XXH3_initCustomSecret_avx2
 
-#else
+#elif (XXH_VECTOR == XXH_SSE2)
 
-    if (XXH_FORCE_ALIGN_CHECK) {
-        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
-            return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
-    }   }
+#define XXH3_accumulate_512 XXH3_accumulate_512_sse2
+#define XXH3_accumulate     XXH3_accumulate_sse2
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_sse2
+#define XXH3_initCustomSecret XXH3_initCustomSecret_sse2
 
-    return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
-#endif
-}
+#elif (XXH_VECTOR == XXH_NEON)
 
+#define XXH3_accumulate_512 XXH3_accumulate_512_neon
+#define XXH3_accumulate     XXH3_accumulate_neon
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_neon
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 
+#elif (XXH_VECTOR == XXH_VSX)
 
-/*******   Hash streaming   *******/
+#define XXH3_accumulate_512 XXH3_accumulate_512_vsx
+#define XXH3_accumulate     XXH3_accumulate_vsx
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_vsx
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 
-XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
-{
-    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
-}
-XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
-{
-    XXH_free(statePtr);
-    return XXH_OK;
-}
+#elif (XXH_VECTOR == XXH_SVE)
+#define XXH3_accumulate_512 XXH3_accumulate_512_sve
+#define XXH3_accumulate     XXH3_accumulate_sve
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 
-XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
-{
-    memcpy(dstState, srcState, sizeof(*dstState));
-}
+#else /* scalar */
 
-XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
-{
-    XXH32_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
-    memset(&state, 0, sizeof(state));
-    state.v1 = seed + PRIME32_1 + PRIME32_2;
-    state.v2 = seed + PRIME32_2;
-    state.v3 = seed + 0;
-    state.v4 = seed - PRIME32_1;
-    /* do not write into reserved, planned to be removed in a future version */
-    memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved));
-    return XXH_OK;
-}
+#define XXH3_accumulate_512 XXH3_accumulate_512_scalar
+#define XXH3_accumulate     XXH3_accumulate_scalar
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 
+#endif
 
-XXH_PUBLIC_API XXH_errorcode
-XXH32_update(XXH32_state_t* state, const void* input, size_t len)
-{
-    if (input==NULL)
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-        return XXH_OK;
-#else
-        return XXH_ERROR;
+#if XXH_SIZE_OPT >= 1 /* don't do SIMD for initialization */
+#  undef XXH3_initCustomSecret
+#  define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
 #endif
 
-    {   const xxh_u8* p = (const xxh_u8*)input;
-        const xxh_u8* const bEnd = p + len;
+XXH_FORCE_INLINE void
+XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
+                      const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                            XXH3_f_accumulate f_acc,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
+    size_t const block_len = XXH_STRIPE_LEN * nbStripesPerBlock;
+    size_t const nb_blocks = (len - 1) / block_len;
 
-        state->total_len_32 += (XXH32_hash_t)len;
-        state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16));
+    size_t n;
 
-        if (state->memsize + len < 16)  {   /* fill in tmp buffer */
-            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len);
-            state->memsize += (XXH32_hash_t)len;
-            return XXH_OK;
-        }
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
 
-        if (state->memsize) {   /* some data left from previous update */
-            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize);
-            {   const xxh_u32* p32 = state->mem32;
-                state->v1 = XXH32_round(state->v1, XXH_readLE32(p32)); p32++;
-                state->v2 = XXH32_round(state->v2, XXH_readLE32(p32)); p32++;
-                state->v3 = XXH32_round(state->v3, XXH_readLE32(p32)); p32++;
-                state->v4 = XXH32_round(state->v4, XXH_readLE32(p32));
-            }
-            p += 16-state->memsize;
-            state->memsize = 0;
-        }
+    for (n = 0; n < nb_blocks; n++) {
+        f_acc(acc, input + n*block_len, secret, nbStripesPerBlock);
+        f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN);
+    }
 
-        if (p <= bEnd-16) {
-            const xxh_u8* const limit = bEnd - 16;
-            xxh_u32 v1 = state->v1;
-            xxh_u32 v2 = state->v2;
-            xxh_u32 v3 = state->v3;
-            xxh_u32 v4 = state->v4;
+    /* last partial block */
+    XXH_ASSERT(len > XXH_STRIPE_LEN);
+    {   size_t const nbStripes = ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN;
+        XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE));
+        f_acc(acc, input + nb_blocks*block_len, secret, nbStripes);
 
-            do {
-                v1 = XXH32_round(v1, XXH_readLE32(p)); p+=4;
-                v2 = XXH32_round(v2, XXH_readLE32(p)); p+=4;
-                v3 = XXH32_round(v3, XXH_readLE32(p)); p+=4;
-                v4 = XXH32_round(v4, XXH_readLE32(p)); p+=4;
-            } while (p<=limit);
+        /* last stripe */
+        {   const xxh_u8* const p = input + len - XXH_STRIPE_LEN;
+#define XXH_SECRET_LASTACC_START 7  /* not aligned on 8, last secret is different from acc & scrambler */
+            XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
+    }   }
+}
 
-            state->v1 = v1;
-            state->v2 = v2;
-            state->v3 = v3;
-            state->v4 = v4;
-        }
+XXH_FORCE_INLINE xxh_u64
+XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret)
+{
+    return XXH3_mul128_fold64(
+               acc[0] ^ XXH_readLE64(secret),
+               acc[1] ^ XXH_readLE64(secret+8) );
+}
 
-        if (p < bEnd) {
-            XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
-            state->memsize = (unsigned)(bEnd-p);
-        }
+static XXH64_hash_t
+XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start)
+{
+    xxh_u64 result64 = start;
+    size_t i = 0;
+
+    for (i = 0; i < 4; i++) {
+        result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i);
+#if defined(__clang__)                                /* Clang */ \
+    && (defined(__arm__) || defined(__thumb__))       /* ARMv7 */ \
+    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */  \
+    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
+        /*
+         * UGLY HACK:
+         * Prevent autovectorization on Clang ARMv7-a. Exact same problem as
+         * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b.
+         * XXH3_64bits, len == 256, Snapdragon 835:
+         *   without hack: 2063.7 MB/s
+         *   with hack:    2560.7 MB/s
+         */
+        XXH_COMPILER_GUARD(result64);
+#endif
     }
 
-    return XXH_OK;
+    return XXH3_avalanche(result64);
 }
 
+#define XXH3_INIT_ACC { XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, \
+                        XXH_PRIME64_4, XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 }
 
-XXH_PUBLIC_API XXH32_hash_t XXH32_digest (const XXH32_state_t* state)
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
+                           const void* XXH_RESTRICT secret, size_t secretSize,
+                           XXH3_f_accumulate f_acc,
+                           XXH3_f_scrambleAcc f_scramble)
 {
-    xxh_u32 h32;
-
-    if (state->large_len) {
-        h32 = XXH_rotl32(state->v1, 1)
-            + XXH_rotl32(state->v2, 7)
-            + XXH_rotl32(state->v3, 12)
-            + XXH_rotl32(state->v4, 18);
-    } else {
-        h32 = state->v3 /* == seed */ + PRIME32_5;
-    }
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
 
-    h32 += state->total_len_32;
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc, f_scramble);
 
-    return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
+    /* converge into final hash */
+    XXH_STATIC_ASSERT(sizeof(acc) == 64);
+    /* do not align on 8, so that the secret is different from the accumulator */
+#define XXH_SECRET_MERGEACCS_START 11
+    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+    return XXH3_mergeAccs(acc, (const xxh_u8*)secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * XXH_PRIME64_1);
 }
 
+/*
+ * It's important for performance to transmit secret's size (when it's static)
+ * so that the compiler can properly optimize the vectorized loop.
+ * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set.
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
+ */
+XXH3_WITH_SECRET_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                             XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64;
+    return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate, XXH3_scrambleAcc);
+}
 
-/*******   Canonical representation   *******/
+/*
+ * It's preferable for performance that XXH3_hashLong is not inlined,
+ * as it results in a smaller function for small data, easier to the instruction cache.
+ * Note that inside this no_inline function, we do inline the internal loop,
+ * and provide a statically defined secret size to allow optimization of vector loop.
+ */
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
+XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
+                          XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate, XXH3_scrambleAcc);
+}
 
 /*
- * The default return values from XXH functions are unsigned 32 and 64 bit
- * integers.
- *
- * The canonical representation uses big endian convention, the same convention
- * as human-readable numbers (large digits first).
+ * XXH3_hashLong_64b_withSeed():
+ * Generate a custom key based on alteration of default XXH3_kSecret with the seed,
+ * and then use this key for long mode hashing.
  *
- * This way, hash values can be written into a file or buffer, remaining
- * comparable across different systems.
+ * This operation is decently fast but nonetheless costs a little bit of time.
+ * Try to avoid it whenever possible (typically when seed==0).
  *
- * The following functions allow transformation of hash values to and from their
- * canonical format.
+ * It's important for performance that XXH3_hashLong is not inlined. Not sure
+ * why (uop cache maybe?), but the difference is large and easily measurable.
  */
-XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
+                                    XXH64_hash_t seed,
+                                    XXH3_f_accumulate f_acc,
+                                    XXH3_f_scrambleAcc f_scramble,
+                                    XXH3_f_initCustomSecret f_initSec)
 {
-    XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
-    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
-    memcpy(dst, &hash, sizeof(*dst));
+#if XXH_SIZE_OPT <= 0
+    if (seed == 0)
+        return XXH3_hashLong_64b_internal(input, len,
+                                          XXH3_kSecret, sizeof(XXH3_kSecret),
+                                          f_acc, f_scramble);
+#endif
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed);
+        return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret),
+                                          f_acc, f_scramble);
+    }
 }
 
-XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed(const void* XXH_RESTRICT input, size_t len,
+                           XXH64_hash_t seed, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
 {
-    return XXH_readBE32(src);
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_withSeed_internal(input, len, seed,
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
 }
 
 
-#ifndef XXH_NO_LONG_LONG
+typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t,
+                                          XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t);
 
-/* *******************************************************************
-*  64-bit hash functions
-*********************************************************************/
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len,
+                     XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+                     XXH3_hashLong64_f f_hashLong)
+{
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+    /*
+     * If an action is to be taken if `secretLen` condition is not respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash.
+     * Also, note that function signature doesn't offer room to return an error.
+     */
+    if (len <= 16)
+        return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+    if (len <= 128)
+        return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    return f_hashLong(input, len, seed64, (const xxh_u8*)secret, secretLen);
+}
 
-/*******   Memory access   *******/
 
-typedef XXH64_hash_t xxh_u64;
+/* ===   Public entry point   === */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length)
+{
+    return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecret(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    return XXH3_64bits_internal(input, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed)
+{
+    return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
+}
+
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    if (length <= XXH3_MIDSIZE_MAX)
+        return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_64b_withSecret(input, length, seed, (const xxh_u8*)secret, secretSize);
+}
 
 
+/* ===   XXH3 streaming   === */
+#ifndef XXH_NO_STREAM
+/*
+ * Malloc's a pointer that is always aligned to @align.
+ *
+ * This must be freed with `XXH_alignedFree()`.
+ *
+ * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte
+ * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2
+ * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON.
+ *
+ * This underalignment previously caused a rather obvious crash which went
+ * completely unnoticed due to XXH3_createState() not actually being tested.
+ * Credit to RedSpah for noticing this bug.
+ *
+ * The alignment is done manually: Functions like posix_memalign or _mm_malloc
+ * are avoided: To maintain portability, we would have to write a fallback
+ * like this anyways, and besides, testing for the existence of library
+ * functions without relying on external build tools is impossible.
+ *
+ * The method is simple: Overallocate, manually align, and store the offset
+ * to the original behind the returned pointer.
+ *
+ * Align must be a power of 2 and 8 <= align <= 128.
+ */
+static XXH_MALLOCF void* XXH_alignedMalloc(size_t s, size_t align)
+{
+    XXH_ASSERT(align <= 128 && align >= 8); /* range check */
+    XXH_ASSERT((align & (align-1)) == 0);   /* power of 2 */
+    XXH_ASSERT(s != 0 && s < (s + align));  /* empty/overflow */
+    {   /* Overallocate to make room for manual realignment and an offset byte */
+        xxh_u8* base = (xxh_u8*)XXH_malloc(s + align);
+        if (base != NULL) {
+            /*
+             * Get the offset needed to align this pointer.
+             *
+             * Even if the returned pointer is aligned, there will always be
+             * at least one byte to store the offset to the original pointer.
+             */
+            size_t offset = align - ((size_t)base & (align - 1)); /* base % align */
+            /* Add the offset for the now-aligned pointer */
+            xxh_u8* ptr = base + offset;
+
+            XXH_ASSERT((size_t)ptr % align == 0);
+
+            /* Store the offset immediately before the returned pointer. */
+            ptr[-1] = (xxh_u8)offset;
+            return ptr;
+        }
+        return NULL;
+    }
+}
+/*
+ * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass
+ * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout.
+ */
+static void XXH_alignedFree(void* p)
+{
+    if (p != NULL) {
+        xxh_u8* ptr = (xxh_u8*)p;
+        /* Get the offset byte we added in XXH_malloc. */
+        xxh_u8 offset = ptr[-1];
+        /* Free the original malloc'd pointer */
+        xxh_u8* base = ptr - offset;
+        XXH_free(base);
+    }
+}
+/*! @ingroup XXH3_family */
 /*!
- * XXH_REROLL_XXH64:
- * Whether to reroll the XXH64_finalize() loop.
+ * @brief Allocate an @ref XXH3_state_t.
  *
- * Just like XXH32, we can unroll the XXH64_finalize() loop. This can be a
- * performance gain on 64-bit hosts, as only one jump is required.
+ * @return An allocated pointer of @ref XXH3_state_t on success.
+ * @return `NULL` on failure.
  *
- * However, on 32-bit hosts, because arithmetic needs to be done with two 32-bit
- * registers, and 64-bit arithmetic needs to be simulated, it isn't beneficial
- * to unroll. The code becomes ridiculously large (the largest function in the
- * binary on i386!), and rerolling it saves anywhere from 3kB to 20kB. It is
- * also slightly faster because it fits into cache better and is more likely
- * to be inlined by the compiler.
+ * @note Must be freed with XXH3_freeState().
  *
- * If XXH_REROLL is defined, this is ignored and the loop is always rerolled.
+ * @see @ref streaming_example "Streaming Example"
  */
-#ifndef XXH_REROLL_XXH64
-#  if (defined(__ILP32__) || defined(_ILP32)) /* ILP32 is often defined on 32-bit GCC family */ \
-   || !(defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) /* x86-64 */ \
-     || defined(_M_ARM64) || defined(__aarch64__) || defined(__arm64__) /* aarch64 */ \
-     || defined(__PPC64__) || defined(__PPC64LE__) || defined(__ppc64__) || defined(__powerpc64__) /* ppc64 */ \
-     || defined(__mips64__) || defined(__mips64)) /* mips64 */ \
-   || (!defined(SIZE_MAX) || SIZE_MAX < ULLONG_MAX) /* check limits */
-#    define XXH_REROLL_XXH64 1
-#  else
-#    define XXH_REROLL_XXH64 0
-#  endif
-#endif /* !defined(XXH_REROLL_XXH64) */
+XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
+{
+    XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
+    if (state==NULL) return NULL;
+    XXH3_INITSTATE(state);
+    return state;
+}
 
-#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
-/*
- * Manual byteshift. Best for old compilers which don't inline memcpy.
- * We actually directly use XXH_readLE64 and XXH_readBE64.
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Frees an @ref XXH3_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note Must be allocated with XXH3_createState().
+ *
+ * @see @ref streaming_example "Streaming Example"
  */
-#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
+{
+    XXH_alignedFree(statePtr);
+    return XXH_OK;
+}
 
-/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
-static xxh_u64 XXH_read64(const void* memPtr) { return *(const xxh_u64*) memPtr; }
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state)
+{
+    XXH_memcpy(dst_state, src_state, sizeof(*dst_state));
+}
 
-#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+static void
+XXH3_reset_internal(XXH3_state_t* statePtr,
+                    XXH64_hash_t seed,
+                    const void* secret, size_t secretSize)
+{
+    size_t const initStart = offsetof(XXH3_state_t, bufferedSize);
+    size_t const initLength = offsetof(XXH3_state_t, nbStripesPerBlock) - initStart;
+    XXH_ASSERT(offsetof(XXH3_state_t, nbStripesPerBlock) > initStart);
+    XXH_ASSERT(statePtr != NULL);
+    /* set members from bufferedSize to nbStripesPerBlock (excluded) to 0 */
+    memset((char*)statePtr + initStart, 0, initLength);
+    statePtr->acc[0] = XXH_PRIME32_3;
+    statePtr->acc[1] = XXH_PRIME64_1;
+    statePtr->acc[2] = XXH_PRIME64_2;
+    statePtr->acc[3] = XXH_PRIME64_3;
+    statePtr->acc[4] = XXH_PRIME64_4;
+    statePtr->acc[5] = XXH_PRIME32_2;
+    statePtr->acc[6] = XXH_PRIME64_5;
+    statePtr->acc[7] = XXH_PRIME32_1;
+    statePtr->seed = seed;
+    statePtr->useSeed = (seed != 0);
+    statePtr->extSecret = (const unsigned char*)secret;
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+    statePtr->secretLimit = secretSize - XXH_STRIPE_LEN;
+    statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
+}
 
-/*
- * __pack instructions are safer, but compiler specific, hence potentially
- * problematic for some compilers.
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, secret, secretSize);
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    if (seed==0) return XXH3_64bits_reset(statePtr);
+    if ((seed != statePtr->seed) || (statePtr->extSecret != NULL))
+        XXH3_initCustomSecret(statePtr->customSecret, seed);
+    XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed64)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, seed64, secret, secretSize);
+    statePtr->useSeed = 1; /* always, even if seed64==0 */
+    return XXH_OK;
+}
+
+/*!
+ * @internal
+ * @brief Processes a large input for XXH3_update() and XXH3_digest_long().
  *
- * Currently only defined for GCC and ICC.
+ * Unlike XXH3_hashLong_internal_loop(), this can process data that overlaps a block.
+ *
+ * @param acc                Pointer to the 8 accumulator lanes
+ * @param nbStripesSoFarPtr  In/out pointer to the number of leftover stripes in the block*
+ * @param nbStripesPerBlock  Number of stripes in a block
+ * @param input              Input pointer
+ * @param nbStripes          Number of stripes to process
+ * @param secret             Secret pointer
+ * @param secretLimit        Offset of the last block in @p secret
+ * @param f_acc              Pointer to an XXH3_accumulate implementation
+ * @param f_scramble         Pointer to an XXH3_scrambleAcc implementation
+ * @return                   Pointer past the end of @p input after processing
  */
-typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
-static xxh_u64 XXH_read64(const void* ptr) { return ((const unalign64*)ptr)->u64; }
+XXH_FORCE_INLINE const xxh_u8 *
+XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
+                    size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock,
+                    const xxh_u8* XXH_RESTRICT input, size_t nbStripes,
+                    const xxh_u8* XXH_RESTRICT secret, size_t secretLimit,
+                    XXH3_f_accumulate f_acc,
+                    XXH3_f_scrambleAcc f_scramble)
+{
+    const xxh_u8* initialSecret = secret + *nbStripesSoFarPtr * XXH_SECRET_CONSUME_RATE;
+    /* Process full blocks */
+    if (nbStripes >= (nbStripesPerBlock - *nbStripesSoFarPtr)) {
+        /* Process the initial partial block... */
+        size_t nbStripesThisIter = nbStripesPerBlock - *nbStripesSoFarPtr;
 
-#else
+        do {
+            /* Accumulate and scramble */
+            f_acc(acc, input, initialSecret, nbStripesThisIter);
+            f_scramble(acc, secret + secretLimit);
+            input += nbStripesThisIter * XXH_STRIPE_LEN;
+            nbStripes -= nbStripesThisIter;
+            /* Then continue the loop with the full block size */
+            nbStripesThisIter = nbStripesPerBlock;
+            initialSecret = secret;
+        } while (nbStripes >= nbStripesPerBlock);
+        *nbStripesSoFarPtr = 0;
+    }
+    /* Process a partial block */
+    if (nbStripes > 0) {
+        f_acc(acc, input, initialSecret, nbStripes);
+        input += nbStripes * XXH_STRIPE_LEN;
+        *nbStripesSoFarPtr += nbStripes;
+    }
+    /* Return end pointer */
+    return input;
+}
 
+#ifndef XXH3_STREAM_USE_STACK
+# if XXH_SIZE_OPT <= 0 && !defined(__clang__) /* clang doesn't need additional stack space */
+#   define XXH3_STREAM_USE_STACK 1
+# endif
+#endif
 /*
- * Portable and safe solution. Generally efficient.
- * see: https://stackoverflow.com/a/32095106/646947
+ * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
  */
-static xxh_u64 XXH_read64(const void* memPtr)
+XXH_FORCE_INLINE XXH_errorcode
+XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
+            const xxh_u8* XXH_RESTRICT input, size_t len,
+            XXH3_f_accumulate f_acc,
+            XXH3_f_scrambleAcc f_scramble)
 {
-    xxh_u64 val;
-    memcpy(&val, memPtr, sizeof(val));
-    return val;
-}
-
-#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
 
-#if defined(_MSC_VER)     /* Visual Studio */
-#  define XXH_swap64 _byteswap_uint64
-#elif XXH_GCC_VERSION >= 403
-#  define XXH_swap64 __builtin_bswap64
+    XXH_ASSERT(state != NULL);
+    {   const xxh_u8* const bEnd = input + len;
+        const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+        /* For some reason, gcc and MSVC seem to suffer greatly
+         * when operating accumulators directly into state.
+         * Operating into stack space seems to enable proper optimization.
+         * clang, on the other hand, doesn't seem to need this trick */
+        XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8];
+        XXH_memcpy(acc, state->acc, sizeof(acc));
 #else
-static xxh_u64 XXH_swap64 (xxh_u64 x)
-{
-    return  ((x << 56) & 0xff00000000000000ULL) |
-            ((x << 40) & 0x00ff000000000000ULL) |
-            ((x << 24) & 0x0000ff0000000000ULL) |
-            ((x << 8)  & 0x000000ff00000000ULL) |
-            ((x >> 8)  & 0x00000000ff000000ULL) |
-            ((x >> 24) & 0x0000000000ff0000ULL) |
-            ((x >> 40) & 0x000000000000ff00ULL) |
-            ((x >> 56) & 0x00000000000000ffULL);
-}
+        xxh_u64* XXH_RESTRICT const acc = state->acc;
 #endif
+        state->totalLen += len;
+        XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE);
 
+        /* small input : just fill in tmp buffer */
+        if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) {
+            XXH_memcpy(state->buffer + state->bufferedSize, input, len);
+            state->bufferedSize += (XXH32_hash_t)len;
+            return XXH_OK;
+        }
 
-/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */
-#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+        /* total input is now > XXH3_INTERNALBUFFER_SIZE */
+        #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN)
+        XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == 0);   /* clean multiple */
+
+        /*
+         * Internal buffer is partially filled (always, except at beginning)
+         * Complete it, then consume it.
+         */
+        if (state->bufferedSize) {
+            size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize;
+            XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize);
+            input += loadSize;
+            XXH3_consumeStripes(acc,
+                               &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                state->buffer, XXH3_INTERNALBUFFER_STRIPES,
+                                secret, state->secretLimit,
+                                f_acc, f_scramble);
+            state->bufferedSize = 0;
+        }
+        XXH_ASSERT(input < bEnd);
+        if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) {
+            size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN;
+            input = XXH3_consumeStripes(acc,
+                                       &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                       input, nbStripes,
+                                       secret, state->secretLimit,
+                                       f_acc, f_scramble);
+            XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
 
-XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr)
-{
-    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
-    return bytePtr[0]
-         | ((xxh_u64)bytePtr[1] << 8)
-         | ((xxh_u64)bytePtr[2] << 16)
-         | ((xxh_u64)bytePtr[3] << 24)
-         | ((xxh_u64)bytePtr[4] << 32)
-         | ((xxh_u64)bytePtr[5] << 40)
-         | ((xxh_u64)bytePtr[6] << 48)
-         | ((xxh_u64)bytePtr[7] << 56);
-}
+        }
+        /* Some remaining input (always) : buffer it */
+        XXH_ASSERT(input < bEnd);
+        XXH_ASSERT(bEnd - input <= XXH3_INTERNALBUFFER_SIZE);
+        XXH_ASSERT(state->bufferedSize == 0);
+        XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
+        state->bufferedSize = (XXH32_hash_t)(bEnd-input);
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+        /* save stack accumulators into state */
+        XXH_memcpy(state->acc, acc, sizeof(acc));
+#endif
+    }
 
-XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr)
-{
-    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
-    return bytePtr[7]
-         | ((xxh_u64)bytePtr[6] << 8)
-         | ((xxh_u64)bytePtr[5] << 16)
-         | ((xxh_u64)bytePtr[4] << 24)
-         | ((xxh_u64)bytePtr[3] << 32)
-         | ((xxh_u64)bytePtr[2] << 40)
-         | ((xxh_u64)bytePtr[1] << 48)
-         | ((xxh_u64)bytePtr[0] << 56);
+    return XXH_OK;
 }
 
-#else
-XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
 {
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
+    return XXH3_update(state, (const xxh_u8*)input, len,
+                       XXH3_accumulate, XXH3_scrambleAcc);
 }
 
-static xxh_u64 XXH_readBE64(const void* ptr)
+
+XXH_FORCE_INLINE void
+XXH3_digest_long (XXH64_hash_t* acc,
+                  const XXH3_state_t* state,
+                  const unsigned char* secret)
 {
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
+    xxh_u8 lastStripe[XXH_STRIPE_LEN];
+    const xxh_u8* lastStripePtr;
+
+    /*
+     * Digest on a local copy. This way, the state remains unaltered, and it can
+     * continue ingesting more input afterwards.
+     */
+    XXH_memcpy(acc, state->acc, sizeof(state->acc));
+    if (state->bufferedSize >= XXH_STRIPE_LEN) {
+        /* Consume remaining stripes then point to remaining data in buffer */
+        size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN;
+        size_t nbStripesSoFar = state->nbStripesSoFar;
+        XXH3_consumeStripes(acc,
+                           &nbStripesSoFar, state->nbStripesPerBlock,
+                            state->buffer, nbStripes,
+                            secret, state->secretLimit,
+                            XXH3_accumulate, XXH3_scrambleAcc);
+        lastStripePtr = state->buffer + state->bufferedSize - XXH_STRIPE_LEN;
+    } else {  /* bufferedSize < XXH_STRIPE_LEN */
+        /* Copy to temp buffer */
+        size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize;
+        XXH_ASSERT(state->bufferedSize > 0);  /* there is always some input buffered */
+        XXH_memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
+        XXH_memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
+        lastStripePtr = lastStripe;
+    }
+    /* Last stripe */
+    XXH3_accumulate_512(acc,
+                        lastStripePtr,
+                        secret + state->secretLimit - XXH_SECRET_LASTACC_START);
 }
-#endif
 
-XXH_FORCE_INLINE xxh_u64
-XXH_readLE64_align(const void* ptr, XXH_alignment align)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
 {
-    if (align==XXH_unaligned)
-        return XXH_readLE64(ptr);
-    else
-        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr);
+    const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+    if (state->totalLen > XXH3_MIDSIZE_MAX) {
+        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+        XXH3_digest_long(acc, state, secret);
+        return XXH3_mergeAccs(acc,
+                              secret + XXH_SECRET_MERGEACCS_START,
+                              (xxh_u64)state->totalLen * XXH_PRIME64_1);
+    }
+    /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */
+    if (state->useSeed)
+        return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+    return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
+                                  secret, state->secretLimit + XXH_STRIPE_LEN);
 }
+#endif /* !XXH_NO_STREAM */
 
 
-/*******   xxh64   *******/
+/* ==========================================
+ * XXH3 128 bits (a.k.a XXH128)
+ * ==========================================
+ * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant,
+ * even without counting the significantly larger output size.
+ *
+ * For example, extra steps are taken to avoid the seed-dependent collisions
+ * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B).
+ *
+ * This strength naturally comes at the cost of some speed, especially on short
+ * lengths. Note that longer hashes are about as fast as the 64-bit version
+ * due to it using only a slight modification of the 64-bit loop.
+ *
+ * XXH128 is also more oriented towards 64-bit machines. It is still extremely
+ * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
+ */
 
-static const xxh_u64 PRIME64_1 = 0x9E3779B185EBCA87ULL;   /* 0b1001111000110111011110011011000110000101111010111100101010000111 */
-static const xxh_u64 PRIME64_2 = 0xC2B2AE3D27D4EB4FULL;   /* 0b1100001010110010101011100011110100100111110101001110101101001111 */
-static const xxh_u64 PRIME64_3 = 0x165667B19E3779F9ULL;   /* 0b0001011001010110011001111011000110011110001101110111100111111001 */
-static const xxh_u64 PRIME64_4 = 0x85EBCA77C2B2AE63ULL;   /* 0b1000010111101011110010100111011111000010101100101010111001100011 */
-static const xxh_u64 PRIME64_5 = 0x27D4EB2F165667C5ULL;   /* 0b0010011111010100111010110010111100010110010101100110011111000101 */
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    /* A doubled version of 1to3_64b with different constants. */
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(1 <= len && len <= 3);
+    XXH_ASSERT(secret != NULL);
+    /*
+     * len = 1: combinedl = { input[0], 0x01, input[0], input[0] }
+     * len = 2: combinedl = { input[1], 0x02, input[0], input[1] }
+     * len = 3: combinedl = { input[2], 0x03, input[0], input[1] }
+     */
+    {   xxh_u8 const c1 = input[0];
+        xxh_u8 const c2 = input[len >> 1];
+        xxh_u8 const c3 = input[len - 1];
+        xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24)
+                                | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
+        xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13);
+        xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+        xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed;
+        xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl;
+        xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph;
+        XXH128_hash_t h128;
+        h128.low64  = XXH64_avalanche(keyed_lo);
+        h128.high64 = XXH64_avalanche(keyed_hi);
+        return h128;
+    }
+}
 
-static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
-    acc += input * PRIME64_2;
-    acc  = XXH_rotl64(acc, 31);
-    acc *= PRIME64_1;
-    return acc;
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(4 <= len && len <= 8);
+    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+    {   xxh_u32 const input_lo = XXH_readLE32(input);
+        xxh_u32 const input_hi = XXH_readLE32(input + len - 4);
+        xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32);
+        xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed;
+        xxh_u64 const keyed = input_64 ^ bitflip;
+
+        /* Shift len to the left to ensure it is even, this avoids even multiplies. */
+        XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2));
+
+        m128.high64 += (m128.low64 << 1);
+        m128.low64  ^= (m128.high64 >> 3);
+
+        m128.low64   = XXH_xorshift64(m128.low64, 35);
+        m128.low64  *= PRIME_MX2;
+        m128.low64   = XXH_xorshift64(m128.low64, 28);
+        m128.high64  = XXH3_avalanche(m128.high64);
+        return m128;
+    }
 }
 
-static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
-    val  = XXH64_round(0, val);
-    acc ^= val;
-    acc  = acc * PRIME64_1 + PRIME64_4;
-    return acc;
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed;
+        xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed;
+        xxh_u64 const input_lo = XXH_readLE64(input);
+        xxh_u64       input_hi = XXH_readLE64(input + len - 8);
+        XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1);
+        /*
+         * Put len in the middle of m128 to ensure that the length gets mixed to
+         * both the low and high bits in the 128x64 multiply below.
+         */
+        m128.low64 += (xxh_u64)(len - 1) << 54;
+        input_hi   ^= bitfliph;
+        /*
+         * Add the high 32 bits of input_hi to the high 32 bits of m128, then
+         * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to
+         * the high 64 bits of m128.
+         *
+         * The best approach to this operation is different on 32-bit and 64-bit.
+         */
+        if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */
+            /*
+             * 32-bit optimized version, which is more readable.
+             *
+             * On 32-bit, it removes an ADC and delays a dependency between the two
+             * halves of m128.high64, but it generates an extra mask on 64-bit.
+             */
+            m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2);
+        } else {
+            /*
+             * 64-bit optimized (albeit more confusing) version.
+             *
+             * Uses some properties of addition and multiplication to remove the mask:
+             *
+             * Let:
+             *    a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF)
+             *    b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000)
+             *    c = XXH_PRIME32_2
+             *
+             *    a + (b * c)
+             * Inverse Property: x + y - x == y
+             *    a + (b * (1 + c - 1))
+             * Distributive Property: x * (y + z) == (x * y) + (x * z)
+             *    a + (b * 1) + (b * (c - 1))
+             * Identity Property: x * 1 == x
+             *    a + b + (b * (c - 1))
+             *
+             * Substitute a, b, and c:
+             *    input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+             *
+             * Since input_hi.hi + input_hi.lo == input_hi, we get this:
+             *    input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+             */
+            m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1);
+        }
+        /* m128 ^= XXH_swap64(m128 >> 64); */
+        m128.low64  ^= XXH_swap64(m128.high64);
+
+        {   /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */
+            XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2);
+            h128.high64 += m128.high64 * XXH_PRIME64_2;
+
+            h128.low64   = XXH3_avalanche(h128.low64);
+            h128.high64  = XXH3_avalanche(h128.high64);
+            return h128;
+    }   }
 }
 
-static xxh_u64 XXH64_avalanche(xxh_u64 h64)
+/*
+ * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
+ */
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
 {
-    h64 ^= h64 >> 33;
-    h64 *= PRIME64_2;
-    h64 ^= h64 >> 29;
-    h64 *= PRIME64_3;
-    h64 ^= h64 >> 32;
-    return h64;
+    XXH_ASSERT(len <= 16);
+    {   if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed);
+        if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed);
+        if (len) return XXH3_len_1to3_128b(input, len, secret, seed);
+        {   XXH128_hash_t h128;
+            xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72);
+            xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88);
+            h128.low64 = XXH64_avalanche(seed ^ bitflipl);
+            h128.high64 = XXH64_avalanche( seed ^ bitfliph);
+            return h128;
+    }   }
 }
 
+/*
+ * A bit slower than XXH3_mix16B, but handles multiply by zero better.
+ */
+XXH_FORCE_INLINE XXH128_hash_t
+XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
+              const xxh_u8* secret, XXH64_hash_t seed)
+{
+    acc.low64  += XXH3_mix16B (input_1, secret+0, seed);
+    acc.low64  ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8);
+    acc.high64 += XXH3_mix16B (input_2, secret+16, seed);
+    acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8);
+    return acc;
+}
 
-#define XXH_get64bits(p) XXH_readLE64_align(p, align)
 
-static xxh_u64
-XXH64_finalize(xxh_u64 h64, const xxh_u8* ptr, size_t len, XXH_alignment align)
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                      XXH64_hash_t seed)
 {
-#define PROCESS1_64 do {                                   \
-    h64 ^= (*ptr++) * PRIME64_5;                           \
-    h64 = XXH_rotl64(h64, 11) * PRIME64_1;                 \
-} while (0)
-
-#define PROCESS4_64 do {                                   \
-    h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * PRIME64_1;      \
-    ptr += 4;                                              \
-    h64 = XXH_rotl64(h64, 23) * PRIME64_2 + PRIME64_3;     \
-} while (0)
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(16 < len && len <= 128);
 
-#define PROCESS8_64 do {                                   \
-    xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr)); \
-    ptr += 8;                                              \
-    h64 ^= k1;                                             \
-    h64  = XXH_rotl64(h64,27) * PRIME64_1 + PRIME64_4;     \
-} while (0)
+    {   XXH128_hash_t acc;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
 
-    /* Rerolled version for 32-bit targets is faster and much smaller. */
-    if (XXH_REROLL || XXH_REROLL_XXH64) {
-        len &= 31;
-        while (len >= 8) {
-            PROCESS8_64;
-            len -= 8;
-        }
-        if (len >= 4) {
-            PROCESS4_64;
-            len -= 4;
+#if XXH_SIZE_OPT >= 1
+        {
+            /* Smaller, but slightly slower. */
+            unsigned int i = (unsigned int)(len - 1) / 32;
+            do {
+                acc = XXH128_mix32B(acc, input+16*i, input+len-16*(i+1), secret+32*i, seed);
+            } while (i-- != 0);
         }
-        while (len > 0) {
-            PROCESS1_64;
-            --len;
+#else
+        if (len > 32) {
+            if (len > 64) {
+                if (len > 96) {
+                    acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed);
+                }
+                acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed);
+            }
+            acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed);
         }
-         return  XXH64_avalanche(h64);
-    } else {
-        switch(len & 31) {
-           case 24: PROCESS8_64;
-                         /* fallthrough */
-           case 16: PROCESS8_64;
-                         /* fallthrough */
-           case  8: PROCESS8_64;
-                    return XXH64_avalanche(h64);
-
-           case 28: PROCESS8_64;
-                         /* fallthrough */
-           case 20: PROCESS8_64;
-                         /* fallthrough */
-           case 12: PROCESS8_64;
-                         /* fallthrough */
-           case  4: PROCESS4_64;
-                    return XXH64_avalanche(h64);
-
-           case 25: PROCESS8_64;
-                         /* fallthrough */
-           case 17: PROCESS8_64;
-                         /* fallthrough */
-           case  9: PROCESS8_64;
-                    PROCESS1_64;
-                    return XXH64_avalanche(h64);
-
-           case 29: PROCESS8_64;
-                         /* fallthrough */
-           case 21: PROCESS8_64;
-                         /* fallthrough */
-           case 13: PROCESS8_64;
-                         /* fallthrough */
-           case  5: PROCESS4_64;
-                    PROCESS1_64;
-                    return XXH64_avalanche(h64);
-
-           case 26: PROCESS8_64;
-                         /* fallthrough */
-           case 18: PROCESS8_64;
-                         /* fallthrough */
-           case 10: PROCESS8_64;
-                    PROCESS1_64;
-                    PROCESS1_64;
-                    return XXH64_avalanche(h64);
-
-           case 30: PROCESS8_64;
-                         /* fallthrough */
-           case 22: PROCESS8_64;
-                         /* fallthrough */
-           case 14: PROCESS8_64;
-                         /* fallthrough */
-           case  6: PROCESS4_64;
-                    PROCESS1_64;
-                    PROCESS1_64;
-                    return XXH64_avalanche(h64);
-
-           case 27: PROCESS8_64;
-                         /* fallthrough */
-           case 19: PROCESS8_64;
-                         /* fallthrough */
-           case 11: PROCESS8_64;
-                    PROCESS1_64;
-                    PROCESS1_64;
-                    PROCESS1_64;
-                    return XXH64_avalanche(h64);
-
-           case 31: PROCESS8_64;
-                         /* fallthrough */
-           case 23: PROCESS8_64;
-                         /* fallthrough */
-           case 15: PROCESS8_64;
-                         /* fallthrough */
-           case  7: PROCESS4_64;
-                         /* fallthrough */
-           case  3: PROCESS1_64;
-                         /* fallthrough */
-           case  2: PROCESS1_64;
-                         /* fallthrough */
-           case  1: PROCESS1_64;
-                         /* fallthrough */
-           case  0: return XXH64_avalanche(h64);
+        acc = XXH128_mix32B(acc, input, input+len-16, secret, seed);
+#endif
+        {   XXH128_hash_t h128;
+            h128.low64  = acc.low64 + acc.high64;
+            h128.high64 = (acc.low64    * XXH_PRIME64_1)
+                        + (acc.high64   * XXH_PRIME64_4)
+                        + ((len - seed) * XXH_PRIME64_2);
+            h128.low64  = XXH3_avalanche(h128.low64);
+            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+            return h128;
         }
     }
-    /* impossible to reach */
-    XXH_ASSERT(0);
-    return 0;  /* unreachable, but some compilers complain without it */
 }
 
-XXH_FORCE_INLINE xxh_u64
-XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                       XXH64_hash_t seed)
 {
-    const xxh_u8* bEnd = input + len;
-    xxh_u64 h64;
-
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-    if (input==NULL) {
-        len=0;
-        bEnd=input=(const xxh_u8*)(size_t)32;
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+    {   XXH128_hash_t acc;
+        unsigned i;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
+        /*
+         *  We set as `i` as offset + 32. We do this so that unchanged
+         * `len` can be used as upper bound. This reaches a sweet spot
+         * where both x86 and aarch64 get simple agen and good codegen
+         * for the loop.
+         */
+        for (i = 32; i < 160; i += 32) {
+            acc = XXH128_mix32B(acc,
+                                input  + i - 32,
+                                input  + i - 16,
+                                secret + i - 32,
+                                seed);
+        }
+        acc.low64 = XXH3_avalanche(acc.low64);
+        acc.high64 = XXH3_avalanche(acc.high64);
+        /*
+         * NB: `i <= len` will duplicate the last 32-bytes if
+         * len % 32 was zero. This is an unfortunate necessity to keep
+         * the hash result stable.
+         */
+        for (i=160; i <= len; i += 32) {
+            acc = XXH128_mix32B(acc,
+                                input + i - 32,
+                                input + i - 16,
+                                secret + XXH3_MIDSIZE_STARTOFFSET + i - 160,
+                                seed);
+        }
+        /* last bytes */
+        acc = XXH128_mix32B(acc,
+                            input + len - 16,
+                            input + len - 32,
+                            secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
+                            (XXH64_hash_t)0 - seed);
+
+        {   XXH128_hash_t h128;
+            h128.low64  = acc.low64 + acc.high64;
+            h128.high64 = (acc.low64    * XXH_PRIME64_1)
+                        + (acc.high64   * XXH_PRIME64_4)
+                        + ((len - seed) * XXH_PRIME64_2);
+            h128.low64  = XXH3_avalanche(h128.low64);
+            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+            return h128;
+        }
     }
-#endif
+}
 
-    if (len>=32) {
-        const xxh_u8* const limit = bEnd - 32;
-        xxh_u64 v1 = seed + PRIME64_1 + PRIME64_2;
-        xxh_u64 v2 = seed + PRIME64_2;
-        xxh_u64 v3 = seed + 0;
-        xxh_u64 v4 = seed - PRIME64_1;
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
+                            const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                            XXH3_f_accumulate f_acc,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc, f_scramble);
+
+    /* converge into final hash */
+    XXH_STATIC_ASSERT(sizeof(acc) == 64);
+    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+    {   XXH128_hash_t h128;
+        h128.low64  = XXH3_mergeAccs(acc,
+                                     secret + XXH_SECRET_MERGEACCS_START,
+                                     (xxh_u64)len * XXH_PRIME64_1);
+        h128.high64 = XXH3_mergeAccs(acc,
+                                     secret + secretSize
+                                            - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+                                     ~((xxh_u64)len * XXH_PRIME64_2));
+        return h128;
+    }
+}
 
-        do {
-            v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8;
-            v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8;
-            v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8;
-            v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8;
-        } while (input<=limit);
+/*
+ * It's important for performance that XXH3_hashLong() is not inlined.
+ */
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
+XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len,
+                           XXH64_hash_t seed64,
+                           const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret),
+                                       XXH3_accumulate, XXH3_scrambleAcc);
+}
 
-        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
-        h64 = XXH64_mergeRound(h64, v1);
-        h64 = XXH64_mergeRound(h64, v2);
-        h64 = XXH64_mergeRound(h64, v3);
-        h64 = XXH64_mergeRound(h64, v4);
+/*
+ * It's important for performance to pass @p secretLen (when it's static)
+ * to the compiler, so that it can properly optimize the vectorized loop.
+ *
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
+ */
+XXH3_WITH_SECRET_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                              XXH64_hash_t seed64,
+                              const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64;
+    return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen,
+                                       XXH3_accumulate, XXH3_scrambleAcc);
+}
 
-    } else {
-        h64  = seed + PRIME64_5;
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed_internal(const void* XXH_RESTRICT input, size_t len,
+                                XXH64_hash_t seed64,
+                                XXH3_f_accumulate f_acc,
+                                XXH3_f_scrambleAcc f_scramble,
+                                XXH3_f_initCustomSecret f_initSec)
+{
+    if (seed64 == 0)
+        return XXH3_hashLong_128b_internal(input, len,
+                                           XXH3_kSecret, sizeof(XXH3_kSecret),
+                                           f_acc, f_scramble);
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed64);
+        return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, sizeof(secret),
+                                           f_acc, f_scramble);
     }
+}
 
-    h64 += (xxh_u64) len;
-
-    return XXH64_finalize(h64, input, len, align);
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed(const void* input, size_t len,
+                            XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_withSeed_internal(input, len, seed64,
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
 }
 
+typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t,
+                                            XXH64_hash_t, const void* XXH_RESTRICT, size_t);
 
-XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t len, XXH64_hash_t seed)
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_128bits_internal(const void* input, size_t len,
+                      XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+                      XXH3_hashLong128_f f_hl128)
 {
-#if 0
-    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
-    XXH64_state_t state;
-    XXH64_reset(&state, seed);
-    XXH64_update(&state, (const xxh_u8*)input, len);
-    return XXH64_digest(&state);
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+    /*
+     * If an action is to be taken if `secret` conditions are not respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash.
+     */
+    if (len <= 16)
+        return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+    if (len <= 128)
+        return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    return f_hl128(input, len, seed64, secret, secretLen);
+}
 
-#else
 
-    if (XXH_FORCE_ALIGN_CHECK) {
-        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
-            return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
-    }   }
+/* ===   Public XXH128 API   === */
 
-    return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* input, size_t len)
+{
+    return XXH3_128bits_internal(input, len, 0,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_default);
+}
 
-#endif
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecret(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    return XXH3_128bits_internal(input, len, 0,
+                                 (const xxh_u8*)secret, secretSize,
+                                 XXH3_hashLong_128b_withSecret);
 }
 
-/*******   Hash Streaming   *******/
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSeed(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_128bits_internal(input, len, seed,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_withSeed);
+}
 
-XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
 {
-    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_128bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_128b_withSecret(input, len, seed, secret, secretSize);
 }
-XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
 {
-    XXH_free(statePtr);
-    return XXH_OK;
+    return XXH3_128bits_withSeed(input, len, seed);
 }
 
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dstState, const XXH64_state_t* srcState)
+
+/* ===   XXH3 128-bit streaming   === */
+#ifndef XXH_NO_STREAM
+/*
+ * All initialization and update functions are identical to 64-bit streaming variant.
+ * The only difference is the finalization routine.
+ */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
 {
-    memcpy(dstState, srcState, sizeof(*dstState));
+    return XXH3_64bits_reset(statePtr);
 }
 
-XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t seed)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
 {
-    XXH64_state_t state;   /* use a local state to memcpy() in order to avoid strict-aliasing warnings */
-    memset(&state, 0, sizeof(state));
-    state.v1 = seed + PRIME64_1 + PRIME64_2;
-    state.v2 = seed + PRIME64_2;
-    state.v3 = seed + 0;
-    state.v4 = seed - PRIME64_1;
-     /* do not write into reserved64, might be removed in a future version */
-    memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved64));
-    return XXH_OK;
+    return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize);
 }
 
+/*! @ingroup XXH3_family */
 XXH_PUBLIC_API XXH_errorcode
-XXH64_update (XXH64_state_t* state, const void* input, size_t len)
+XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
 {
-    if (input==NULL)
-#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1)
-        return XXH_OK;
-#else
-        return XXH_ERROR;
-#endif
+    return XXH3_64bits_reset_withSeed(statePtr, seed);
+}
 
-    {   const xxh_u8* p = (const xxh_u8*)input;
-        const xxh_u8* const bEnd = p + len;
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    return XXH3_64bits_reset_withSecretandSeed(statePtr, secret, secretSize, seed);
+}
 
-        state->total_len += len;
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+    return XXH3_64bits_update(state, input, len);
+}
 
-        if (state->memsize + len < 32) {  /* fill in tmp buffer */
-            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len);
-            state->memsize += (xxh_u32)len;
-            return XXH_OK;
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
+{
+    const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+    if (state->totalLen > XXH3_MIDSIZE_MAX) {
+        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+        XXH3_digest_long(acc, state, secret);
+        XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+        {   XXH128_hash_t h128;
+            h128.low64  = XXH3_mergeAccs(acc,
+                                         secret + XXH_SECRET_MERGEACCS_START,
+                                         (xxh_u64)state->totalLen * XXH_PRIME64_1);
+            h128.high64 = XXH3_mergeAccs(acc,
+                                         secret + state->secretLimit + XXH_STRIPE_LEN
+                                                - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+                                         ~((xxh_u64)state->totalLen * XXH_PRIME64_2));
+            return h128;
         }
+    }
+    /* len <= XXH3_MIDSIZE_MAX : short code */
+    if (state->useSeed)
+        return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+    return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen),
+                                   secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+#endif /* !XXH_NO_STREAM */
+/* 128-bit utility functions */
 
-        if (state->memsize) {   /* tmp buffer is full */
-            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize);
-            state->v1 = XXH64_round(state->v1, XXH_readLE64(state->mem64+0));
-            state->v2 = XXH64_round(state->v2, XXH_readLE64(state->mem64+1));
-            state->v3 = XXH64_round(state->v3, XXH_readLE64(state->mem64+2));
-            state->v4 = XXH64_round(state->v4, XXH_readLE64(state->mem64+3));
-            p += 32-state->memsize;
-            state->memsize = 0;
-        }
+#include <string.h>   /* memcmp, memcpy */
 
-        if (p+32 <= bEnd) {
-            const xxh_u8* const limit = bEnd - 32;
-            xxh_u64 v1 = state->v1;
-            xxh_u64 v2 = state->v2;
-            xxh_u64 v3 = state->v3;
-            xxh_u64 v4 = state->v4;
+/* return : 1 is equal, 0 if different */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
+{
+    /* note : XXH128_hash_t is compact, it has no padding byte */
+    return !(memcmp(&h1, &h2, sizeof(h1)));
+}
 
-            do {
-                v1 = XXH64_round(v1, XXH_readLE64(p)); p+=8;
-                v2 = XXH64_round(v2, XXH_readLE64(p)); p+=8;
-                v3 = XXH64_round(v3, XXH_readLE64(p)); p+=8;
-                v4 = XXH64_round(v4, XXH_readLE64(p)); p+=8;
-            } while (p<=limit);
+/* This prototype is compatible with stdlib's qsort().
+ * @return : >0 if *h128_1  > *h128_2
+ *           <0 if *h128_1  < *h128_2
+ *           =0 if *h128_1 == *h128_2  */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2)
+{
+    XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
+    XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2;
+    int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64);
+    /* note : bets that, in most cases, hash values are different */
+    if (hcmp) return hcmp;
+    return (h1.low64 > h2.low64) - (h2.low64 > h1.low64);
+}
 
-            state->v1 = v1;
-            state->v2 = v2;
-            state->v3 = v3;
-            state->v4 = v4;
-        }
 
-        if (p < bEnd) {
-            XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
-            state->memsize = (unsigned)(bEnd-p);
-        }
+/*======   Canonical representation   ======*/
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) {
+        hash.high64 = XXH_swap64(hash.high64);
+        hash.low64  = XXH_swap64(hash.low64);
     }
-
-    return XXH_OK;
+    XXH_memcpy(dst, &hash.high64, sizeof(hash.high64));
+    XXH_memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
 }
 
-
-XXH_PUBLIC_API XXH64_hash_t XXH64_digest (const XXH64_state_t* state)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src)
 {
-    xxh_u64 h64;
+    XXH128_hash_t h;
+    h.high64 = XXH_readBE64(src);
+    h.low64  = XXH_readBE64(src->digest + 8);
+    return h;
+}
 
-    if (state->total_len >= 32) {
-        xxh_u64 const v1 = state->v1;
-        xxh_u64 const v2 = state->v2;
-        xxh_u64 const v3 = state->v3;
-        xxh_u64 const v4 = state->v4;
 
-        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
-        h64 = XXH64_mergeRound(h64, v1);
-        h64 = XXH64_mergeRound(h64, v2);
-        h64 = XXH64_mergeRound(h64, v3);
-        h64 = XXH64_mergeRound(h64, v4);
-    } else {
-        h64  = state->v3 /*seed*/ + PRIME64_5;
-    }
 
-    h64 += (xxh_u64) state->total_len;
+/* ==========================================
+ * Secret generators
+ * ==========================================
+ */
+#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))
 
-    return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
+XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128)
+{
+    XXH_writeLE64( dst, XXH_readLE64(dst) ^ h128.low64 );
+    XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 );
 }
 
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize)
+{
+#if (XXH_DEBUGLEVEL >= 1)
+    XXH_ASSERT(secretBuffer != NULL);
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+#else
+    /* production mode, assert() are disabled */
+    if (secretBuffer == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+#endif
 
-/******* Canonical representation   *******/
+    if (customSeedSize == 0) {
+        customSeed = XXH3_kSecret;
+        customSeedSize = XXH_SECRET_DEFAULT_SIZE;
+    }
+#if (XXH_DEBUGLEVEL >= 1)
+    XXH_ASSERT(customSeed != NULL);
+#else
+    if (customSeed == NULL) return XXH_ERROR;
+#endif
 
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash)
-{
-    XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
-    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
-    memcpy(dst, &hash, sizeof(*dst));
+    /* Fill secretBuffer with a copy of customSeed - repeat as needed */
+    {   size_t pos = 0;
+        while (pos < secretSize) {
+            size_t const toCopy = XXH_MIN((secretSize - pos), customSeedSize);
+            memcpy((char*)secretBuffer + pos, customSeed, toCopy);
+            pos += toCopy;
+    }   }
+
+    {   size_t const nbSeg16 = secretSize / 16;
+        size_t n;
+        XXH128_canonical_t scrambler;
+        XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0));
+        for (n=0; n<nbSeg16; n++) {
+            XXH128_hash_t const h128 = XXH128(&scrambler, sizeof(scrambler), n);
+            XXH3_combine16((char*)secretBuffer + n*16, h128);
+        }
+        /* last segment */
+        XXH3_combine16((char*)secretBuffer + secretSize - 16, XXH128_hashFromCanonical(&scrambler));
+    }
+    return XXH_OK;
 }
 
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src)
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed)
 {
-    return XXH_readBE64(src);
+    XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+    XXH3_initCustomSecret(secret, seed);
+    XXH_ASSERT(secretBuffer != NULL);
+    memcpy(secretBuffer, secret, XXH_SECRET_DEFAULT_SIZE);
 }
 
 
 
-/* *********************************************************************
-*  XXH3
-*  New generation hash designed for speed on small keys and vectorization
-************************************************************************ */
-
-#include "xxh3.h"
-
+/* Pop our optimization override from above */
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
+#  pragma GCC pop_options
+#endif
 
 #endif  /* XXH_NO_LONG_LONG */
 
+#endif  /* XXH_NO_XXH3 */
 
+/*!
+ * @}
+ */
 #endif  /* XXH_IMPLEMENTATION */
 
 
 #if defined (__cplusplus)
-}
+} /* extern "C" */
 #endif

blitz.mod/tests/test.bmx

@@ -10,6 +10,15 @@ Type TStringTest Extends TTest
 	Field bigUnicode:UInt[] = [$10300, $10301, $10302, $10303, $10304, $10305, 0]
 	Field unicode:Int[] = [1055, 1088, 1080, 1074, 1077, 1090]
 	Field utf8:Byte[] = [208, 159, 209, 128, 208, 184, 208, 178, 208, 181, 209, 130, 0]
+
+	Const HELLO_UPPER:String = "HELLO"
+	Const HELLO_LOWER:String = "hello"
+	Const UMLAUT_UPPER:String = "123ÄÖÜABC"
+	Const UMLAUT_LOWER:String = "123äöüabc"
+	Const ARABIC_UPPER:String = "123كلمة"
+	Const ARABIC_LOWER:String = "123كلمة"
+	Const CYRILLIC_UPPER:String = "123БУДИНОК"
+	Const CYRILLIC_LOWER:String = "123будинок"
 	
 	Method setup() { before }
 	End Method
@@ -24,4 +33,411 @@ Type TStringTest Extends TTest
 		Next
 	End Method
 
+	Method testToUTF8StringBuffer() { test }
+		Local s:String = "1234567890"
+		Local buf:Byte Ptr = StackAlloc(50)
+
+		Local length:size_t = 10
+		s.ToUTF8StringBuffer(buf, length)
+
+	End Method
+
+	Method testASCIIToLower() { test }
+		Local s:String = HELLO_UPPER
+		assertEquals(HELLO_LOWER, s.ToLower())
+
+		Local obj:Object = HELLO_LOWER
+		Local obj1:Object = HELLO_LOWER.ToLower()
+
+		assertTrue(obj = obj1, "Already lowercase ASCII strings should return the same object")
+
+	End Method
+
+	Method testASCIIToUpper() { test }
+		Local s:String = HELLO_LOWER
+		assertEquals(HELLO_UPPER, s.ToUpper())
+
+		Local obj:Object = HELLO_UPPER
+		Local obj1:Object = HELLO_UPPER.ToUpper()
+
+		assertTrue(obj = obj1, "Already uppercase ASCII strings should return the same object")
+
+	End Method
+
+	Method testUnicodeToLower() { test }
+
+		Local s:String = UMLAUT_UPPER
+		assertEquals(UMLAUT_LOWER, s.ToLower())
+
+		Local obj:Object = UMLAUT_LOWER
+		Local obj1:Object = UMLAUT_LOWER.ToLower()
+
+		assertTrue(obj = obj1, "Already lowercase Unicode strings should return the same object")
+
+	End Method
+
+	Method testUnicodeToUpper() { test }
+
+		Local s:String = UMLAUT_LOWER
+		assertEquals(UMLAUT_UPPER, s.ToUpper())
+
+		Local obj:Object = UMLAUT_UPPER
+		Local obj1:Object = UMLAUT_UPPER.ToUpper()
+
+		assertTrue(obj = obj1, "Already uppercase Unicode strings should return the same object")
+
+	End Method
+
+	Method testArabicToLower() { test }
+
+		Local s:String = ARABIC_UPPER
+		assertEquals(ARABIC_LOWER, s.ToLower(), "Arabic lower case")
+
+		Local obj:Object = ARABIC_LOWER
+		Local obj1:Object = ARABIC_LOWER.ToLower()
+
+		assertTrue(obj = obj1, "Already lowercase Arabic strings should return the same object")
+
+	End Method
+
+	Method testArabicToUpper() { test }
+
+		Local s:String = ARABIC_LOWER
+		assertEquals(ARABIC_UPPER, s.ToUpper(), "Arabic upper case")
+
+		Local obj:Object = ARABIC_UPPER
+		Local obj1:Object = ARABIC_UPPER.ToUpper()
+
+		assertTrue(obj = obj1, "Already uppercase Arabic strings should return the same object")
+
+	End Method
+
+	Method testArabicUpperToLower() { test }
+
+		Local s:String = ARABIC_UPPER
+		assertEquals(ARABIC_UPPER, s.ToLower(), "Arabic lower case and upper case should be the same")
+
+		Local obj:Object = ARABIC_UPPER
+		Local obj1:Object = ARABIC_UPPER.ToLower()
+
+		assertTrue(obj = obj1, "Uppercase Arabic strings should return the same object when lowered")
+
+	End Method
+
+	Method testArabicLowerToUpper() { test }
+
+		Local s:String = ARABIC_LOWER
+		assertEquals(ARABIC_LOWER, s.ToUpper(), "Arabic upper case and lower case should be the same")
+
+		Local obj:Object = ARABIC_LOWER
+		Local obj1:Object = ARABIC_LOWER.ToUpper()
+
+		assertTrue(obj = obj1, "Lowercase Arabic strings should return the same object when uppered")
+
+	End Method
+
+	Method testCyrillicToLower() { test }
+
+		Local s:String = CYRILLIC_UPPER
+		assertEquals(CYRILLIC_LOWER, s.ToLower(), "Cyrillic lower case")
+
+		Local obj:Object = CYRILLIC_LOWER
+		Local obj1:Object = CYRILLIC_LOWER.ToLower()
+
+		assertTrue(obj = obj1, "Already lowercase Cyrillic strings should return the same object")
+
+	End Method
+
+	Method testCyrrilicToUpper() { test }
+
+		Local s:String = CYRILLIC_LOWER
+		assertEquals(CYRILLIC_UPPER, s.ToUpper(), "Cyrillic upper case")
+
+		Local obj:Object = CYRILLIC_UPPER
+		Local obj1:Object = CYRILLIC_UPPER.ToUpper()
+
+		assertTrue(obj = obj1, "Already uppercase Cyrillic strings should return the same object")
+
+	End Method
+
+End Type
+
+Struct STestStruct
+	Field a:Int
+	Field c:Float
+	Field d:Double
+	Field b:ULong
+End Struct
+
+Type TStructArrayTest Extends TTest
+
+	Method testStructArray() { test }
+
+		Local arr:STestStruct[] = New STestStruct[10]
+		
+		For Local i:Int = 0 Until 10
+			arr[i].a = i
+			arr[i].b = i * i
+		Next
+
+		For Local i:Int = 0 Until 10
+			assertEquals(i, arr[i].a)
+			assertEquals(i * i, arr[i].b)
+		Next
+	End Method
+
+	Method testStructArraySlice() { test }
+
+		Local arr:STestStruct[] = New STestStruct[10]
+		
+		For Local i:Int = 0 Until 10
+			arr[i].a = i
+			arr[i].b = i * i
+		Next
+
+		Local slice:STestStruct[] = arr[2..5]
+
+		assertEquals(3, slice.Length)
+		assertEquals(2, slice[0].a)
+		assertEquals(3, slice[1].a)
+		assertEquals(4, slice[2].a)
+	End Method
+
+End Type
+
+Type TStringToDoubleExTest Extends TTest
+
+	Method testToDoubleEx() { test }
+		Local val:Double
+		Local s:String = "123.456"
+		assertEquals(7, s.ToDoubleEx(val))
+		assertEquals(123.456, val, 0.0001)
+	End Method
+
+	Method testToDoubleExMulti() { test }
+		Local val:Double
+		Local s:String = "1,2,3,4,5,6,7,8,9,10"
+
+		Local start:Int = 0
+		For Local i:Int = 0 Until 10
+			start = s.ToDoubleEx(val, start) + 1
+
+			assertFalse(start = 1)
+			assertEquals(i + 1, val, 0.0001)
+		Next
+	End Method
+
+	Method testToDoubleExMultiTab() { test }
+		Local val:Double
+		Local s:String = "1~t2~t3~t4~t5~t6~t7~t8~t9~t10"
+
+		Local start:Int = 0
+		For Local i:Int = 0 Until 10
+			start = s.ToDoubleEx(val, start,,CHARSFORMAT_SKIPWHITESPACE)
+			assertFalse(start = 0)
+			assertEquals(i + 1, val, 0.0001)
+		Next
+	End Method
+
+	Method testLeadingWhitespace() { test }
+		Local val:Double
+		Local s:String = "  ~t123.456"
+		assertEquals(10, s.ToDoubleEx(val,,,CHARSFORMAT_SKIPWHITESPACE))
+		assertEquals(123.456, val, 0.0001)
+	End Method
+
+	Method testToDoubleExCommaSeparator() { test }
+		Local val:Double
+		Local s:String = "123,456"
+		assertEquals(7, s.ToDoubleEx(val,,,,","))
+		assertEquals(123.456, val, 0.0001)
+	End Method
+
+End Type
+
+Type TStringToFloatExTest Extends TTest
+
+	Method testToFloatEx() { test }
+		Local val:Float
+		Local s:String = "123.456"
+		assertEquals(7, s.ToFloatEx(val))
+		assertEquals(123.456, val, 0.0001)
+	End Method
+
+	Method testToFloatExMulti() { test }
+		Local val:Float
+		Local s:String = "1,2,3,4,5,6,7,8,9,10"
+
+		Local start:Int = 0
+		For Local i:Int = 0 Until 10
+			start = s.ToFloatEx(val, start) + 1
+
+			assertFalse(start = 1)
+			assertEquals(i + 1, val, 0.0001)
+		Next
+	End Method
+
+	Method testLeadingWhitespace() { test }
+		Local val:Float
+		Local s:String = "  ~t123.456"
+		assertEquals(10, s.ToFloatEx(val,,,CHARSFORMAT_SKIPWHITESPACE))
+		assertEquals(123.456, val, 0.0001)
+	End Method
+
+	Method testToFloatExCommaSeparator() { test }
+		Local val:Float
+		Local s:String = "123,456"
+		assertEquals(7, s.ToFloatEx(val,,,,","))
+		assertEquals(123.456, val, 0.0001)
+	End Method
+
+End Type
+
+Type TStringToIntExTest Extends TTest
+
+	Method testToIntEx() { test }
+		Local val:Int
+		Local s:String = "123456"
+		assertEquals(6, s.ToIntEx(val))
+		assertEquals(123456, val)
+	End Method
+
+	Method testToIntExMulti() { test }
+		Local val:Int
+		Local s:String = "1,2,3,4,5,6,7,8,9,10"
+
+		Local start:Int = 0
+		For Local i:Int = 0 Until 10
+			start = s.ToIntEx(val, start) + 1
+
+			assertFalse(start = 1)
+			assertEquals(i + 1, val)
+		Next
+	End Method
+
+	Method testLeadingWhitespace() { test }
+		Local val:Int
+		Local s:String = "  ~t123456"
+		assertEquals(9, s.ToIntEx(val,,,CHARSFORMAT_SKIPWHITESPACE))
+		assertEquals(123456, val)
+	End Method
+
+	Method testHex() { test }
+		Local val:Int
+		Local s:String = "abc001"
+		assertEquals(6, s.ToIntEx(val,,,,16))
+		assertEquals(11255809, val)
+	End Method
+
+End Type
+
+Type TStringFromUTF8BytesTest Extends TTest
+
+    ' Test valid ASCII conversion.
+    Method testASCII() { test }
+        Local data:Byte[] = [72, 101, 108, 108, 111] ' "Hello"
+        Local text:String = String.FromUTF8Bytes(data, data.Length)
+        assertEquals("Hello", text)
+    End Method
+
+    ' Test conversion of a 2-byte UTF-8 sequence (e.g. ©: U+00A9).
+    Method testTwoByteSequence() { test }
+        ' © U+00A9: UTF-8: $C2, $A9.
+        Local data:Byte[] = [$C2, $A9]
+        Local text:String = String.FromUTF8Bytes(data, data.Length)
+        assertEquals(Chr($00A9), text)
+    End Method
+
+    ' Test conversion of a 3-byte UTF-8 sequence (e.g. €: U+20AC).
+    Method testThreeByteSequence() { test }
+        ' € U+20AC: UTF-8: $E2, $82, $AC.
+        Local data:Byte[] = [$E2, $82, $AC]
+        Local text:String = String.FromUTF8Bytes(data, data.Length)
+        assertEquals(Chr($20AC), text)
+    End Method
+
+    ' Test conversion of a 4-byte UTF-8 sequence (e.g. U+1F600: grinning face emoji).
+    Method testFourByteSequence() { test }
+        ' Grinning Face U+1F600: UTF-8: $F0, $9F, $98, $80.
+        Local data:Byte[] = [$F0, $9F, $98, $80]
+        Local text:String = String.FromUTF8Bytes(data, data.Length)
+        ' Expected string in UTF-16: surrogate pair (high: $D83D, low: $DE00).
+        Local expected:String = Chr($D83D) + Chr($DE00)
+        assertEquals(expected, text)
+    End Method
+
+    ' Test an incomplete sequence (missing continuation bytes).
+    Method testIncompleteSequence() { test }
+        ' Incomplete 3-byte sequence: [$E2, $82] missing the final byte.
+        Local data:Byte[] = [$E2, $82]
+        Local text:String = String.FromUTF8Bytes(data, data.Length)
+        ' Expect a replacement character.
+        assertEquals(Chr($FFFD), text)
+    End Method
+
+    ' Test an invalid continuation byte following a valid starter.
+    Method testInvalidContinuation() { test }
+        ' [$C2, $20]: $20 is not a valid continuation byte.
+        Local data:Byte[] = [$C2, $20]
+        Local text:String = String.FromUTF8Bytes(data, data.Length)
+        assertEquals(Chr($FFFD), text)
+    End Method
+
+    ' Test a stray continuation byte.
+    Method testStrayContinuation() { test }
+        ' A single continuation byte $80 is invalid.
+        Local data:Byte[] = [$80]
+        Local text:String = String.FromUTF8Bytes(data, data.Length)
+        assertEquals(Chr($FFFD), text)
+    End Method
+
+    ' Test a mix of valid and invalid sequences.
+    Method testMixedValidInvalid() { test }
+        ' "A" ($41), stray continuation ($80), then "B" ($42).
+        Local data:Byte[] = [65, $80, 66]
+        Local text:String = String.FromUTF8Bytes(data, data.Length)
+        Local expected:String = Chr(65) + Chr($FFFD) + Chr(66)
+        assertEquals(expected, text)
+    End Method
+
+    ' Test overlong encoding.
+    Method testOverlongEncoding() { test }
+        ' Overlong encoding for NUL: [$C0, $80] should be rejected.
+        Local data:Byte[] = [$C0, $80]
+        Local text:String = String.FromUTF8Bytes(data, data.Length)
+        assertEquals(Chr($FFFD), text)
+    End Method
+
+    ' Test a UTF-8 sequence encoding a surrogate half (e.g. U+D800).
+    Method testSurrogateHalf() { test }
+        ' U+D800 encoded in UTF-8: [$ED, $A0, $80].
+        Local data:Byte[] = [$ED, $A0, $80]
+        Local text:String = String.FromUTF8Bytes(data, data.Length)
+        assertEquals(Chr($FFFD), text)
+    End Method
+
+	' Test conversion of Russian "hello" ("привет").
+	Method testRussianHello() { test }
+		' "привет": [$D0, $BF, $D1, $80, $D0, $B8, $D0, $B2, $D0, $B5, $D1, $82]
+		Local data:Byte[] = [$D0, $BF, $D1, $80, $D0, $B8, $D0, $B2, $D0, $B5, $D1, $82]
+		Local text:String = String.FromUTF8Bytes(data, data.Length)
+		assertEquals("привет", text)
+	End Method
+
+	' Test conversion of Japanese "hello" ("こんにちは").
+	Method testJapaneseHello() { test }
+		' "こんにちは": [$E3, $81, $93, $E3, $82, $93, $E3, $81, $AB, $E3, $81, $A1, $E3, $81, $AF]
+		Local data:Byte[] = [$E3, $81, $93, $E3, $82, $93, $E3, $81, $AB, $E3, $81, $A1, $E3, $81, $AF]
+		Local text:String = String.FromUTF8Bytes(data, data.Length)
+		assertEquals("こんにちは", text)
+	End Method
+
+	' Test conversion of Chinese "hello" ("你好").
+	Method testChineseHello() { test }
+		' "你好": [$E4, $BD, $A0, $E5, $A5, $BD]
+		Local data:Byte[] = [$E4, $BD, $A0, $E5, $A5, $BD]
+		Local text:String = String.FromUTF8Bytes(data, data.Length)
+		assertEquals("你好", text)
+	End Method
+
 End Type

bytebuffer.mod/bytebuffer.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2020 Bruce A Henderson
+' Copyright (c) 2024 Bruce A Henderson
 ' 
 ' This software is provided 'as-is', without any express or implied
 ' warranty. In no event will the authors be held liable for any damages
@@ -23,15 +23,16 @@ bbdoc: Byte Buffer
 End Rem
 Module BRL.ByteBuffer
 
-ModuleInfo "Version: 1.01"
+ModuleInfo "Version: 1.02"
 ModuleInfo "License: zlib/libpng"
-ModuleInfo "Copyright: 2020 Bruce A Henderson"
+ModuleInfo "Copyright: 2024 Bruce A Henderson"
 
+ModuleInfo "History: 1.02"
+ModuleInfo "History: Added Slice() and Compact() methods"
 ModuleInfo "History: 1.01"
 ModuleInfo "History: Added GetBytes() and PutBytes()"
 ModuleInfo "History: 1.00 Initial Release"
 
-
 Import "glue.c"
 
 Rem
@@ -57,6 +58,13 @@ Protected
 		_size = size
 		_limit = size
 	End Method
+
+	Method New(mark:Int, position:Int, limit:Int, capacity:Int)
+		_mark = mark
+		_position = position
+		_limit = limit
+		_size = capacity
+	End Method
 	
 Public
 	Rem
@@ -330,15 +338,29 @@ Type TByteBuffer Extends TBuffer
 
 	Rem
 	bbdoc: Returns a sliced #TByteBuffer that shares its content with this one.
-	about: TODO
+	about: The new buffer's position, limit, and mark are independent of this buffer.
 	End Rem
 	Method Slice:TByteBuffer() Abstract
-	
+
+	Rem
+	bbdoc: Returns a sliced #TByteBuffer that shares its content with this one.
+	about: The new buffer's position, limit, and mark are independent of this buffer.
+	End Rem
+	Method Slice:TByteBuffer(length:Int) Abstract
+
 	Rem
 	bbdoc: Creates a duplicate #TByteBuffer that shares its content with this one.
 	End Rem
 	Method Duplicate:TByteBuffer() Abstract
 	
+	Rem
+	bbdoc: Compacts this #TByteBuffer.
+	End Rem
+	Method Compact:TByteBuffer() Abstract
+
+	Method BytePtr:Byte Ptr() Abstract
+	Method Offset:Int() Abstract
+
 End Type
 
 Rem
@@ -489,7 +511,7 @@ Type TBytePtrBuffer Extends TByteBuffer
 		
 		Local pos:Int = _position + _offset
 		MemCopy(dst, _data + pos, Size_T(length))
-		
+
 		_position = newPosition
 	End Method
 
@@ -699,10 +721,44 @@ Type TBytePtrBuffer Extends TByteBuffer
 		Return New TBytePtrBuffer(_data, remaining(), _offset + _position, _readOnly)
 	End Method
 
+	Method Slice:TByteBuffer(length:Int) Override
+		If length > remaining() Then
+			Throw New TBufferOverflowException
+		End If
+
+		Return New TBytePtrBuffer(_data, length, _offset + _position, _readOnly)
+	End Method
+
 	Method Duplicate:TByteBuffer() Override
 		Return Copy(Self, _mark, _readOnly)
 	End Method
 
+	Method Compact:TByteBuffer() Override
+		Local remaining:Int
+		If _position <= _limit Then
+			remaining = _limit - _position
+		End If
+
+		MemCopy(_data + _offset, _data + _offset + _position, Size_T(remaining))
+        Position(remaining)
+        Limit(_size)
+        _mark = -1
+
+		Return Self
+	End Method
+
+	Method BytePtr:Byte Ptr()
+		If _readOnly Then
+            Throw New TReadOnlyBufferException()
+		End If
+
+		Return _data
+	End Method
+
+	Method Offset:Int()
+		Return _offset
+	End Method
+
 Private
 	Function Copy:TBytePtrBuffer(buffer:TBytePtrBuffer, mark:Int, isReadOnly:Int)
 		Local bufCopy:TBytePtrBuffer = New TBytePtrBuffer(buffer._data, buffer._size, buffer._offset, isReadOnly)
@@ -746,14 +802,27 @@ Private
 
 Public
 	Method Slice:TByteBuffer() Override
-		Return New TByteArrayBuffer(_data, remaining(), _offset + _position, _readOnly)
+		Return New TByteArrayBuffer(_array, remaining(), _offset + _position, _readOnly)
+	End Method
+
+	Method Slice:TByteBuffer(length:Int) Override
+		If length > remaining() Then
+			Throw New TBufferOverflowException
+		End If
+		Return New TByteArrayBuffer(_array, length, _offset + _position, _readOnly)
 	End Method
 
 	Method Duplicate:TByteBuffer() Override
 		Return Copy(Self, _mark, _readOnly)
 	End Method
 
-Private
+	Method BytePtr:Byte Ptr()
+        If _readOnly Then
+            Throw New TReadOnlyBufferException()
+		End If
+
+		Return _array
+	End Method
 
 End Type
 
@@ -767,6 +836,8 @@ End Type
 Type TReadOnlyBufferException Extends TBlitzException
 End Type
 
+Private
+
 Extern
 	Function bmx_bytebuffer_intbitstofloat:Float(value:Int)
 	Function bmx_bytebuffer_floattointbits:Int(value:Float)

collections.mod/collections.bmx

@@ -5,11 +5,13 @@ bbdoc: Generic Collections.
 End Rem
 Module BRL.Collections
 
-ModuleInfo "Version: 1.00"
+ModuleInfo "Version: 1.01"
 ModuleInfo "Author: Bruce A Henderson"
 ModuleInfo "License: zlib/libpng"
-ModuleInfo "Copyright: 2019 Bruce A Henderson"
+ModuleInfo "Copyright: 2019-2024 Bruce A Henderson"
 
+ModuleInfo "History: 1.01"
+ModuleInfo "History: Added TBlockingQueue"
 ModuleInfo "History: 1.00"
 ModuleInfo "History: Initial Release"
 

collections.mod/examples/blockingqueue_01.bmx

@@ -0,0 +1,38 @@
+'
+' Demonstrates how to use a blocking queue to synchronize threads.
+'
+SuperStrict
+
+Framework Brl.StandardIO
+Import Brl.Threads
+Import Brl.Collections
+
+
+Function Producer:Object(data:Object)
+	Local queue:TBlockingQueue<Int> = TBlockingQueue<Int>(data)
+
+    For Local i:Int = 1 To 10
+        Print "Producing " + i
+        queue.Enqueue(i)
+        Delay 500 ' Simulate work
+    Next
+End Function
+
+Function Consumer:Object(data:Object)
+	Local queue:TBlockingQueue<Int> = TBlockingQueue<Int>(data)
+
+    For Local i:Int = 1 To 10
+        Local item:Int = queue.Dequeue()
+        Print "Consuming " + item
+        Delay 1000 ' Simulate work
+    Next
+End Function
+
+Local queue:TBlockingQueue<Int> = New TBlockingQueue<Int>(5)
+Local producerThread:TThread = CreateThread(Producer, queue)
+Local consumerThread:TThread = CreateThread(Consumer, queue)
+
+WaitThread(producerThread)
+WaitThread(consumerThread)
+
+Print "All tasks are done."

collections.mod/examples/blockingqueue_02.bmx

@@ -0,0 +1,46 @@
+'
+' Demonstrates how to use a blocking queue to synchronize threads.
+'
+SuperStrict
+
+Framework Brl.StandardIO
+Import Brl.Threads
+Import Brl.Collections
+
+
+Function Producer:Object(data:Object)
+	Local queue:TBlockingQueue<Int> = TBlockingQueue<Int>(data)
+
+    For Local i:Int = 1 To 10
+        Try
+            Print "Producing " + i
+            queue.Enqueue(i, 100, ETimeUnit.Milliseconds) ' 100 milliseconds timeout
+            Delay 100 ' Simulate work
+        Catch ex:TTimeoutException
+            Print "Enqueue timed out: " + ex.ToString()
+        End Try
+    Next
+End Function
+
+Function Consumer:Object(data:Object)
+	Local queue:TBlockingQueue<Int> = TBlockingQueue<Int>(data)
+
+    For Local i:Int = 1 To 10
+        Try
+            Local item:Int = queue.Dequeue(1500, ETimeUnit.Milliseconds) ' 1.5 second timeout
+            Print "Consuming " + item
+            Delay 1000 ' Simulate work
+        Catch ex:TTimeoutException
+            Print "Dequeue timed out: " + ex.ToString()
+        End Try
+    Next
+End Function
+
+Local queue:TBlockingQueue<Int> = New TBlockingQueue<Int>(5)
+Local producerThread:TThread = CreateThread(Producer, queue)
+Local consumerThread:TThread = CreateThread(Consumer, queue)
+
+WaitThread(producerThread)
+WaitThread(consumerThread)
+
+Print "All tasks are done."

collections.mod/queue.bmx

@@ -1,6 +1,10 @@
 SuperStrict
 
 Import "collection.bmx"
+?threaded
+Import BRL.threads
+Import BRL.Time
+?
 
 Rem
 bbdoc: A first-in, first-out (FIFO) collection of elements.
@@ -312,3 +316,310 @@ Type TQueueIterator<T> Implements IIterator<T>
 		Return index <> queue.tail
 	End Method
 End Type
+
+?threaded
+
+Rem
+bbdoc: A thread-safe first-in, first-out (FIFO) collection of elements.
+about: Implements a queue as a circular array. Elements stored in a #TBlockingQueue are inserted at one end and removed from the other.
+Use a #TBlockingQueue if you need to access the information in the same order that it is stored in the collection and you need to ensure that the collection is thread-safe.
+A call to #Dequeue will block if the queue is empty. A call to #Enqueue will block if the queue is full.
+The capacity of a #TBlockingQueue is the number of elements the #TBlockingQueue can hold. Once the queue is full, any attempt to add an element will block until space is available.
+End Rem
+Type TBlockingQueue<T> Extends TQueue<T>
+
+	Private
+		Field lock:TMutex
+		Field notEmpty:TCondVar
+		Field notFull:TCondVar
+	Public
+
+	Method New(capacity:Int = 16)
+		Super.New(capacity)
+		lock = TMutex.Create()
+		notEmpty = TCondVar.Create()
+		notFull = TCondVar.Create()
+	End Method
+	
+	Method Enqueue(element:T)
+		lock.Lock()
+		While full
+			notFull.Wait(lock)
+		Wend
+		Super.Enqueue(element)
+		notEmpty.Signal()
+		lock.Unlock()
+	End Method
+	
+	Rem
+	bbdoc: Adds an element to the end of the #TBlockingQueue, waiting up to the specified wait time if necessary for space to become available
+	about: If the queue is full, the operation will block until space becomes available or the specified timeout elapses.
+	Throws a #TTimeoutException if the operation times out.
+	End Rem
+	Method Enqueue(element:T, timeout:ULong, unit:ETimeUnit = ETimeUnit.Milliseconds)
+		Local timeoutMs:ULong = TimeUnitToMillis(timeout, unit)
+	
+		Local startTime:ULong = CurrentUnixTime()
+		lock.Lock()
+		While full
+			Local now:ULong = CurrentUnixTime()
+			If timeout > 0 And now - startTime >= timeoutMs
+				lock.Unlock()
+				Throw New TTimeoutException("The operation timed out after " + timeoutMs + "ms")
+			End If
+			notFull.TimedWait(lock, Int(timeoutMs - (now - startTime)))
+		Wend
+		Super.Enqueue(element)
+		notEmpty.Signal()
+		lock.Unlock()
+	End Method
+	
+	Rem
+	bbdoc: Removes and returns the element at the beginning of the #TBlockingQueue, waiting up to the specified wait time if necessary for an element to become available.
+	about: If the queue is empty, the operation will block until an element becomes available or the specified timeout elapses.
+	Throws a #TTimeoutException if the operation times out.
+	End Rem
+	Method Dequeue:T(timeout:ULong, unit:ETimeUnit = ETimeUnit.Milliseconds)
+		Local timeoutMs:ULong = TimeUnitToMillis(timeout, unit)
+	
+		Local startTime:Long = CurrentUnixTime()
+		lock.Lock()
+		While IsEmpty()
+			Local now:ULong = CurrentUnixTime()
+			If timeout > 0 And now - startTime >= timeoutMs
+				lock.Unlock()
+				Throw New TTimeoutException("The operation timed out after " + timeoutMs + "ms")
+			End If
+			notEmpty.TimedWait(lock, Int(timeoutMs - (now - startTime)))
+		Wend
+		Local element:T = Super.Dequeue()
+		notFull.Signal()
+		lock.Unlock()
+		Return element
+	End Method
+
+	Method Dequeue:T()
+		lock.Lock()
+		While IsEmpty()
+			notEmpty.Wait(lock)
+		Wend
+		Local element:T = Super.Dequeue()
+		notFull.Signal()
+		lock.Unlock()
+		Return element
+	End Method
+	
+	Method TryDequeue:Int(value:T Var)
+		lock.Lock()
+		If IsEmpty()
+			lock.Unlock()
+			Return False
+		End If
+		value = Super.Dequeue()
+		notFull.Signal()
+		lock.Unlock()
+		Return True
+	End Method
+	
+	Method TryPeek:Int(value:T Var)
+		lock.Lock()
+		If IsEmpty()
+			lock.Unlock()
+			Return False
+		End If
+		value = data[head]
+		lock.Unlock()
+		Return True
+	End Method
+	
+	Method Clear()
+		lock.Lock()
+		Super.Clear()
+		notFull.Signal()
+		lock.Unlock()
+	End Method
+	
+	Method TrimExcess()
+		' noop since a blocking queue does not grow beyond its initial capacity
+	End Method
+	
+	Method Resize()
+		lock.Lock()
+		Super.Resize()
+		notFull.Signal()
+		lock.Unlock()
+	End Method
+	
+End Type
+
+Rem
+bbdoc: A thread-safe first-in, first-out (FIFO) collection of elements that supports the concept of tasks.
+about: When a task is complete, the task should call the #TaskDone method to signal that the task is done.
+End Rem
+Type TBlockingTaskQueue<T> Extends TQueue<T>
+
+	Private
+		Field lock:TMutex
+		Field notEmpty:TCondVar
+		Field notFull:TCondVar
+		Field allTasksDone:TCondVar
+		Field taskLock:TMutex
+		Field unfinishedTasks:Int
+	Public
+
+	Method New(capacity:Int = 16)
+		Super.New(capacity)
+		lock = TMutex.Create()
+		notEmpty = TCondVar.Create()
+		notFull = TCondVar.Create()
+		allTasksDone = TCondVar.Create()
+		taskLock = TMutex.Create()
+		unfinishedTasks = 0
+	End Method
+	
+	Method Enqueue(element:T)
+		lock.Lock()
+		While full
+			notFull.Wait(lock)
+		Wend
+		Super.Enqueue(element)
+		taskLock.Lock()
+		unfinishedTasks :+ 1
+		taskLock.Unlock()
+		notEmpty.Signal()
+		lock.Unlock()
+	End Method
+	
+	Rem
+	bbdoc: Adds an element to the end of the #TBlockingTaskQueue, waiting up to the specified wait time if necessary for space to become available
+	about: If the queue is full, the operation will block until space becomes available or the specified timeout elapses.
+	Throws a #TTimeoutException if the operation times out.
+	End Rem
+	Method Enqueue(element:T, timeout:ULong, unit:ETimeUnit = ETimeUnit.Milliseconds)
+		Local timeoutMs:ULong = TimeUnitToMillis(timeout, unit)
+	
+		Local startTime:ULong = CurrentUnixTime()
+		lock.Lock()
+		While full
+			Local now:ULong = CurrentUnixTime()
+			If timeout > 0 And now - startTime >= timeoutMs
+				lock.Unlock()
+				Throw New TTimeoutException("The operation timed out after " + timeoutMs + "ms")
+			End If
+			notFull.TimedWait(lock, Int(timeoutMs - (now - startTime)))
+		Wend
+		Super.Enqueue(element)
+		taskLock.Lock()
+		unfinishedTasks :+ 1
+		taskLock.Unlock()
+		notEmpty.Signal()
+		lock.Unlock()
+	End Method
+	
+	Rem
+	bbdoc: Removes and returns the element at the beginning of the #TBlockingTaskQueue, waiting up to the specified wait time if necessary for an element to become available.
+	about: If the queue is empty, the operation will block until an element becomes available or the specified timeout elapses.
+	Throws a #TTimeoutException if the operation times out.
+	End Rem
+	Method Dequeue:T(timeout:ULong, unit:ETimeUnit = ETimeUnit.Milliseconds)
+		Local timeoutMs:ULong = TimeUnitToMillis(timeout, unit)
+	
+		Local startTime:Long = CurrentUnixTime()
+		lock.Lock()
+		While IsEmpty()
+			Local now:ULong = CurrentUnixTime()
+			If timeout > 0 And now - startTime >= timeoutMs
+				lock.Unlock()
+				Throw New TTimeoutException("The operation timed out after " + timeoutMs + "ms")
+			End If
+			notEmpty.TimedWait(lock, Int(timeoutMs - (now - startTime)))
+		Wend
+		Local element:T = Super.Dequeue()
+		notFull.Signal()
+		lock.Unlock()
+		Return element
+	End Method
+
+	Method Dequeue:T()
+		lock.Lock()
+		While IsEmpty()
+			notEmpty.Wait(lock)
+		Wend
+		Local element:T = Super.Dequeue()
+		notFull.Signal()
+		lock.Unlock()
+		Return element
+	End Method
+	
+	Method TryDequeue:Int(value:T Var)
+		lock.Lock()
+		If IsEmpty()
+			lock.Unlock()
+			Return False
+		End If
+		value = Super.Dequeue()
+		notFull.Signal()
+		lock.Unlock()
+		Return True
+	End Method
+	
+	Method TryPeek:Int(value:T Var)
+		lock.Lock()
+		If IsEmpty()
+			lock.Unlock()
+			Return False
+		End If
+		value = data[head]
+		lock.Unlock()
+		Return True
+	End Method
+	
+	Method Clear()
+		lock.Lock()
+		Super.Clear()
+		taskLock.Lock()
+		unfinishedTasks = 0
+		allTasksDone.Signal()
+		taskLock.Unlock()
+		notFull.Signal()
+		lock.Unlock()
+	End Method
+	
+	Method TrimExcess()
+		' noop since a blocking queue does not grow beyond its initial capacity
+	End Method
+	
+	Method Resize()
+		lock.Lock()
+		Super.Resize()
+		notFull.Signal()
+		lock.Unlock()
+	End Method
+
+	Rem
+	bbdoc: Signals that a task is done.
+	End Rem
+	Method TaskDone()
+		taskLock.Lock()
+		If unfinishedTasks > 0 Then
+			unfinishedTasks :- 1
+			If unfinishedTasks = 0 Then
+				allTasksDone.Signal()
+			End If
+		End If
+		taskLock.Unlock()
+	End Method
+
+	Rem
+	bbdoc: Waits until all tasks are done.
+	End Rem
+	Method Join()
+		taskLock.Lock()
+		While unfinishedTasks > 0
+			allTasksDone.Wait(taskLock)
+		Wend
+		taskLock.Unlock()
+	End Method
+	
+End Type
+?

collections.mod/set.bmx

@@ -335,7 +335,7 @@ Type TSet<T> Implements ISet<T>
 		End If
 
 		If Not size Then
-			Union(other)
+			UnionOf(other)
 			Return
 		End If
 		
@@ -374,7 +374,7 @@ Type TSet<T> Implements ISet<T>
 	Rem
 	bbdoc: Modifies the current #TSet so that it contains all elements that are present in either the current #TSet or the specified #IIterable.
 	End Rem
-	Method Union(other:IIterable<T>)
+	Method UnionOf(other:IIterable<T>)
 		If Not other Then
 			Throw New TArgumentNullException("other")
 		End If

font.mod/font.bmx

@@ -14,9 +14,30 @@ ModuleInfo "History: Module is now SuperStrict"
 ModuleInfo "History: 1.05 Release"
 ModuleInfo "History: Modified interface for improved unicode support"
 
-Const BOLDFONT:Int=1
-Const ITALICFONT:Int=2
-Const SMOOTHFONT:Int=4
+Const BOLDFONT:Int=  $001
+Const ITALICFONT:Int=$002
+Const SMOOTHFONT:Int=$004
+
+Const SMALLCAPSFONT:Int=              $0000100
+Const ALLSMALLCAPSFONT:Int=           $0000200
+Const LIGATURESFONT:Int=              $0000400
+Const DISCRETIONARYLIGATURESFONT:Int= $0000800
+Const OLDSTYLEFIGURESFONT:Int=        $0001000
+Const TABULARFIGURESFONT:Int=         $0002000
+Const FRACTIONSFONT:Int=              $0004000
+Const SUPERSCRIPTFONT:Int=            $0008000
+Const SUBSCRIPTFONT:Int=              $0010000
+Const SWASHESFONT:Int=                $0020000
+Const STYLISTICALTERNATESFONT:Int=    $0040000
+Const CONTEXTUALALTERNATESFONT:Int=   $0080000
+Const HISTORICALFORMSFONT:Int=        $0100000
+Const DENOMINATORSFONT:Int=           $0200000
+Const NUMERATORFONT:Int=              $0400000
+Const LININGFIGURESFONT:Int=          $0800000
+Const SCIENTIFICINFERIORSFONT:Int=    $1000000
+Const PROPORTIONALFIGURESFONT:Int=    $2000000
+Const KERNFONT:Int=                   $4000000
+Const ZEROFONT:Int=                   $8000000
 
 Type TGlyph
 	
@@ -24,6 +45,7 @@ Type TGlyph
 
 	Method Advance:Float() Abstract
 	Method GetRect( x:Int Var,y:Int Var,width:Int Var,height:Int Var ) Abstract
+	Method Index:Int() Abstract
 
 End Type
 
@@ -34,13 +56,14 @@ Type TFont
 	Method CountGlyphs:Int() Abstract
 	Method CharToGlyph:Int( char:Int ) Abstract
 	Method LoadGlyph:TGlyph( index:Int ) Abstract
+	Method LoadGlyphs:TGlyph[]( text:String ) Abstract
 
 End Type
 
 Type TFontLoader
 	Field _succ:TFontLoader
 
-	Method LoadFont:TFont( url:Object,size:Int,style:Int ) Abstract
+	Method LoadFont:TFont( url:Object,size:Float,style:Int ) Abstract
 
 End Type
 
@@ -56,7 +79,7 @@ Function AddFontLoader( loader:TFontLoader )
 	_loaders=loader
 End Function
 
-Function LoadFont:TFont( url:Object,size:Int,style:Int=SMOOTHFONT )
+Function LoadFont:TFont( url:Object,size:Float,style:Int=SMOOTHFONT )
 
 	Local loader:TFontLoader=_loaders
 	

freetypefont.mod/freetypefont.bmx

@@ -49,6 +49,7 @@ Type TFreeTypeGlyph Extends TGlyph
 
 	Field _pixmap:TPixmap
 	Field _advance:Float,_x:Int,_y:Int,_w:Int,_h:Int
+	Field _index:Int
 	
 	Method Pixels:TPixmap() Override
 		If _pixmap Return _pixmap
@@ -67,6 +68,10 @@ Type TFreeTypeGlyph Extends TGlyph
 		h=_h
 	End Method
 
+	Method Index:Int() Override
+		Return _index
+	End Method
+
 End Type
 
 Type TFreeTypeFont Extends BRL.Font.TFont
@@ -94,7 +99,7 @@ Type TFreeTypeFont Extends BRL.Font.TFont
 	End Method
 	
 	Method CharToGlyph:Int( char:Int ) Override
-		Return FT_Get_Char_Index( _ft_face,char )-1
+		Return FT_Get_Char_Index( _ft_face,ULongInt(char) )-1
 	End Method
 
 	Method FamilyName:String()
@@ -111,9 +116,10 @@ Type TFreeTypeFont Extends BRL.Font.TFont
 		If glyph Return glyph
 
 		glyph=New TFreeTypeGlyph
+		glyph._index=index
 		_glyphs[index]=glyph
 		
-		If FT_Load_Glyph( _ft_face,index+1,FT_LOAD_RENDER ) Return glyph
+		If FT_Load_Glyph( _ft_face,UInt(index+1),FT_LOAD_RENDER ) Return glyph
 			
 		Local _slot:Byte Ptr = bmx_freetype_Face_glyph(_ft_face)
 
@@ -162,16 +168,43 @@ Type TFreeTypeFont Extends BRL.Font.TFont
 		Return glyph
 
 	End Method
+
+	Method LoadGlyphs:TGlyph[]( text:String )
+		Throw "Not supported"
+	End Method
+	
+	Function Load:TFreeTypeFont( src:Object,size:Float,style:Int )
+
+		Local buf:Byte[]
+				
+		Local ft_face:Byte Ptr = LoadFace(src, size, style, buf)
+
+		If Not ft_face Then
+			Return Null
+		End If
+		
+		Local ft_size:Byte Ptr = bmx_freetype_Face_size(ft_face)
+		
+		Local font:TFreeTypeFont=New TFreeTypeFont
+		font._ft_face=ft_face
+		font._style=style
+		font._height=bmx_freetype_Size_height(ft_size) Sar 6
+		font._ascend=bmx_freetype_Size_ascend(ft_size) Sar 6
+		font._descend=bmx_freetype_Size_descend(ft_size) Sar 6
+		font._glyphs=New TFreeTypeGlyph[bmx_freetype_Face_numglyphs(ft_face)]
+		font._buf=buf
+		
+		Return font
 	
-	Function Load:TFreeTypeFont( src:Object,size:Int,style:Int )
+	End Function
+
+	Function LoadFace:Byte Ptr( src:Object,size:Float,style:Int, buf:Byte[] Var )
 
 		Global ft_lib:Byte Ptr
 		
 		If Not ft_lib
 			If FT_Init_FreeType( Varptr ft_lib ) Return Null
 		EndIf
-
-		Local buf:Byte[]
 				
 		Local ft_face:Byte Ptr
 
@@ -196,7 +229,7 @@ Type TFreeTypeFont Extends BRL.Font.TFont
 			
 			buf = data
 			
-			If FT_New_Memory_Face( ft_lib, buf, buf.length, 0, Varptr ft_face )
+			If FT_New_Memory_Face( ft_lib, buf, LongInt(buf.length), 0, Varptr ft_face )
 				Return Null
 			EndIf
 
@@ -211,7 +244,7 @@ Type TFreeTypeFont Extends BRL.Font.TFont
 			MemCopy(buf, data, TBank(src).Size())
 			TBank(src).UnLock()
 			
-			If FT_New_Memory_Face( ft_lib, buf, buf.length, 0, Varptr ft_face )
+			If FT_New_Memory_Face( ft_lib, buf, LongInt(buf.length), 0, Varptr ft_face )
 				Return Null
 			EndIf
 
@@ -223,7 +256,7 @@ Type TFreeTypeFont Extends BRL.Font.TFont
 
 				If Not buf.length Return Null
 
-				If FT_New_Memory_Face( ft_lib,buf,buf.length,0,Varptr ft_face )
+				If FT_New_Memory_Face( ft_lib,buf,LongInt(buf.length),0,Varptr ft_face )
 					Return Null
 				EndIf
 			Else
@@ -232,36 +265,33 @@ Type TFreeTypeFont Extends BRL.Font.TFont
 		Else
 			Return Null
 		End If
-		
+
+
+		' Freetype's char height is "FreeType 26.6 Fixed-Point"
+		' -> 26 bit for the integer part, 6 bit for the decimal places
+		' -> "64" equals to 1.0 in floating point numbers
+		' So the module only support fractional sizes with a detail of 1/64
 		While size
-			If Not FT_Set_Pixel_Sizes( ft_face,0,size ) Exit
-			size:-1
+			'default DPI is 72 (setting 0,0 will use that)
+			If Not FT_Set_Char_Size( ft_face, 0, LongInt(size * 64), 0,0 ) Then Exit
+			' If it failed, ensure to use only integer sizes now
+			' (eg. a bitmap font was tried to get loaded) 
+			' First try will be the next integer (10.5 -> 10)
+			size = Ceil(size) - 1
 		Wend
 		If Not size 
 			FT_Done_Face ft_face
 			Return Null
 		EndIf
-		
-		Local ft_size:Byte Ptr = bmx_freetype_Face_size(ft_face)
-		
-		Local font:TFreeTypeFont=New TFreeTypeFont
-		font._ft_face=ft_face
-		font._style=style
-		font._height=bmx_freetype_Size_height(ft_size) Sar 6
-		font._ascend=bmx_freetype_Size_ascend(ft_size) Sar 6
-		font._descend=bmx_freetype_Size_descend(ft_size) Sar 6
-		font._glyphs=New TFreeTypeGlyph[bmx_freetype_Face_numglyphs(ft_face)]
-		font._buf=buf
-		
-		Return font
-	
+
+		Return ft_face
 	End Function
 
 End Type
 
 Type TFreeTypeFontLoader Extends TFontLoader
 
-	Method LoadFont:TFreeTypeFont( url:Object,size:Int,style:Int ) Override
+	Method LoadFont:TFreeTypeFont( url:Object,size:Float,style:Int ) Override
 	
 		Return TFreeTypeFont.Load( url,size,style )
 	

glmax2d.mod/glmax2d.bmx

@@ -193,6 +193,43 @@ Function AdjustTexSize( width:Int Var, height:Int Var )
 	Forever
 End Function
 
+
+
+Global dead_FBOs:TDynamicArray = New TDynamicArray(32)
+Global dead_FBO_seq:Int
+
+'Enqueues a FBO for deletion, to prevent releasing framebuffers on wrong thread.
+Function DeleteFBO( FBO:Int,seq:Int )
+	If seq<>dead_FBO_seq Return
+
+	dead_FBOs.AddLast(FBO)
+End Function
+
+Function CreateFBO:Int(TextureName:Int )
+	Local FrameBufferObject:Int
+	glGenFramebuffers(1, Varptr FrameBufferObject)
+	glBindFramebuffer(GL_FRAMEBUFFER, FrameBufferObject)
+	glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, TextureName, 0)
+
+	' Flush dead FBOs, this ensures to delete FBOs from within the
+	' main thread, while Delete() of image frames can happen from subthread
+	' too.
+	' This also means, it only deletes FBOs if a new is created!
+	If dead_FBO_seq = GraphicsSeq
+		Local deadFBO:Int = dead_FBOs.RemoveLast()
+		While deadFBO <> $FFFFFFFF
+			glDeleteFramebuffers(1, Varptr deadFBO) ' gl ignores 0
+
+			deadFBO = dead_FBOs.RemoveLast()
+		Wend
+	EndIf
+
+	dead_FBO_seq = GraphicsSeq
+
+	Return FrameBufferObject
+End Function
+
+
 Type TDynamicArray
 
 	Private
@@ -371,10 +408,9 @@ Type TGLRenderImageFrame Extends TGLImageFrame
 	
 	Function Create:TGLRenderImageFrame(width:UInt, height:UInt, flags:Int)
 		' Need this to enable frame buffer objects - glGenFramebuffers
-		Global GlewIsInitialised:Int = False
-		If Not GlewIsInitialised
+		If Not glewIsInit
 			GlewInit()
-			GlewIsInitialised = True
+			glewIsInit = True
 		EndIf
 		
 		' store so that we can restore once the fbo is created
@@ -383,7 +419,10 @@ Type TGLRenderImageFrame Extends TGLImageFrame
 		
 		Local TextureName:Int
 		glGenTextures(1, Varptr TextureName)
-		glBindTexture(GL_TEXTURE_2D, TextureName)
+		' inform engine about TextureName being GL_TEXTURE_2D target 
+		' do not just call glBindTexture directly!
+		BindTex(TextureName)
+		'glBindTexture(GL_TEXTURE_2D, TextureName)
 		glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, Null)
 		
 		If flags & FILTEREDIMAGE
@@ -397,11 +436,7 @@ Type TGLRenderImageFrame Extends TGLImageFrame
 		glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE)
 		glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE)
 		
-		Local FrameBufferObject:Int
-		glGenFramebuffers(1, Varptr FrameBufferObject)
-		glBindFramebuffer(GL_FRAMEBUFFER, FrameBufferObject)
-		glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, TextureName, 0)
-		
+		Local FrameBufferObject:Int = CreateFBO(TextureName)
 		Local RenderTarget:TGLRenderImageFrame = New TGLRenderImageFrame
 		RenderTarget.name = TextureName
 		RenderTarget.FBO = FrameBufferObject
@@ -422,11 +457,13 @@ Type TGLRenderImageFrame Extends TGLImageFrame
 	
 Private
 	Method Delete()
-		'remove framebuffer if used
-		if FBO <> 0
-			glDeleteFramebuffers(1, Varptr FBO) ' gl ignores 0
-		EndIf
-	EndMethod
+		If Not seq Then Return
+		If Not FBO Then Return
+
+		'delete FBO deferred
+		DeleteFBO( FBO, seq )
+		FBO = 0
+	End Method
 
 	Method New()
 	EndMethod
@@ -486,6 +523,13 @@ Type TGLMax2DDriver Extends TMax2DDriver
 		glMatrixMode GL_MODELVIEW
 		glLoadIdentity
 		glViewport 0,0,gw,gh
+
+		' Need this to enable "glBlendFuncSeparate" (required for
+		' alpha blending on non-opaque backgrounds like render images)
+		If Not glewIsInit
+			GlewInit()
+			glewIsInit = True
+		EndIf
 		
 		' Create default back buffer render image - the FBO will be value 0 which is the default for the existing backbuffer
 		Local BackBufferRenderImageFrame:TGLRenderImageFrame = New TGLRenderImageFrame
@@ -528,7 +572,11 @@ Type TGLMax2DDriver Extends TMax2DDriver
 			glDisable( GL_ALPHA_TEST )
 		Case ALPHABLEND
 			glEnable( GL_BLEND )
-			glBlendFunc( GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA )
+			' simple alphablend:
+			'glBlendFunc( GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA )
+			' more advanced blend function allows blending on a non-opaque
+			' "background" (eg. render image)
+			glBlendFuncSeparate(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA, GL_ONE, GL_ONE_MINUS_SRC_ALPHA)
 			glDisable( GL_ALPHA_TEST )
 		Case LIGHTBLEND
 			glEnable( GL_BLEND )
@@ -716,8 +764,6 @@ Type TGLMax2DDriver Extends TMax2DDriver
 	EndMethod
 	
 Private
-	Field _glewIsInitialised:Int = False
-
 	Method SetMatrixAndViewportToCurrentRenderImage()
 		glMatrixMode(GL_PROJECTION)
 		glLoadIdentity()

io.mod/glue.c

@@ -213,10 +213,8 @@ int bmx_PHYSFS_setRoot(BBString * archive, BBString * subdir) {
 	char sbuf[1024];
 	size_t slen = 1024;
 
-	char * sd = 0;
 	if (subdir != &bbEmptyString) {
 		bbStringToUTF8StringBuffer(subdir, sbuf, &slen);
-		sd = &sbuf;
 	}
 
 	return PHYSFS_setRoot(abuf, sbuf);

linkedlist.mod/linkedlist.bmx

@@ -66,9 +66,9 @@ Type TLink
 	bbdoc: Removes the link from the List.
 	End Rem
 	Method Remove()
-		_value=Null
-		_succ._pred=_pred
 		_pred._succ=_succ
+		_succ._pred=_pred
+		_value=Null
 	End Method
 
 End Type

map.mod/intmap.bmx

@@ -8,7 +8,7 @@ Extern
 	Function bmx_map_intmap_insert(key:Int, value:Object, root:SavlRoot Ptr Ptr)
 	Function bmx_map_intmap_contains:Int(key:Int, root:SavlRoot Ptr)
 	Function bmx_map_intmap_valueforkey:Object(key:Int, root:SavlRoot Ptr)
-	Function bmx_map_intmap_remove:Int(key:Int, root:SavlRoot Ptr)
+	Function bmx_map_intmap_remove:Int(key:Int, root:SavlRoot Ptr Ptr)
 	Function bmx_map_intmap_firstnode:SIntMapNode Ptr(root:SavlRoot Ptr)
 	Function bmx_map_intmap_nextnode:SIntMapNode Ptr(node:SIntMapNode Ptr)
 	Function bmx_map_intmap_key:Int(node:SIntMapNode Ptr)

map.mod/ptrmap.bmx

@@ -177,7 +177,7 @@ Type TPtrMap
 	about: If the map does not contain @key, a #Null object is returned.
 	End Rem
 	Method Operator[]:Object(key:Byte Ptr)
-		Return bmx_map_ptrmap_valueforkey(key, Varptr _root)
+		Return bmx_map_ptrmap_valueforkey(key, _root)
 	End Method
 	
 	Rem

math.mod/bbMath.h

@@ -70,6 +70,13 @@ inline double bbRound( double x ){
 inline double bbTrunc( double x ){
 	return trunc( x );
 }
+inline double bbDeg2Rad( double x ){
+	return x * DEG_TO_RAD;
+}
+inline double bbRad2Deg( double x ){
+	return x * RAD_TO_DEG;
+}
+
 
 #define RAD_TO_DEGF 57.2957795
 #define DEG_TO_RADF 0.0174532
@@ -128,6 +135,12 @@ inline float bbRoundf( float x ){
 inline float bbTruncf( float x ){
 	return truncf( x );
 }
+inline float bbDeg2Radf( float x ){
+	return x * DEG_TO_RADF;
+}
+inline float bbRad2Degf( float x ){
+	return x * RAD_TO_DEGF;
+}
 #else
 int bbIsNan( double x );
 int bbIsInf( double x );
@@ -149,6 +162,8 @@ double bbLog10( double x );
 double bbCeil( double x );
 double bbRound( double x );
 double bbTrunc( double x );
+double bbDeg2Rad( double x );
+double bbRad2Deg( double x );
 
 int bbIsNanf( float x );
 int bbIsInff( float x );
@@ -170,4 +185,6 @@ float bbLog10f( float x );
 float bbCeilf( float x );
 float bbRoundf( float x );
 float bbTruncf( float x );
+float bbDeg2Radf( float x );
+float bbRad2Degf( float x );
 #endif

math.mod/math.bmx

@@ -6,12 +6,14 @@ bbdoc: Math/Math
 End Rem
 Module BRL.Math
 
-ModuleInfo "Version: 1.08"
+ModuleInfo "Version: 1.09"
 ModuleInfo "Author: Mark Sibly"
 ModuleInfo "License: zlib/libpng"
 ModuleInfo "Copyright: Blitz Research Ltd"
 ModuleInfo "Modserver: BRL"
 
+ModuleInfo "History: 1.09"
+ModuleInfo "History: Added Deg2Rad, Rad2Deg functions."
 ModuleInfo "History: 1.08"
 ModuleInfo "History: Inlined math functions."
 ModuleInfo "History: 1.07"
@@ -128,6 +130,18 @@ bbdoc: Nearest integral not greater in magnitude than @x.
 End Rem
 Function Trunc:Double( x:Double )="double bbTrunc(double)!"
 
+Rem
+bbdoc: Converts a degree value @x of type #Double to its equivalent radiant value.
+about: Takes a degree value and converts it to radiants using the formula: `radiant = degree × (π / 180)`.
+End Rem
+Function Deg2Rad:Double( x:Double )="double bbDeg2Rad(double)!"
+
+Rem
+bbdoc: Converts a radiant value @x of type #Double to its equivalent degree value.
+about: Takes a radiant value and converts it to degrees using the formula: `degree = radiant × (180 / π)`.
+End Rem
+Function Rad2Deg:Double( x:Double )="double bbRad2Deg(double)!"
+
 
 Rem
 bbdoc: Square root of @x
@@ -219,4 +233,16 @@ bbdoc: Nearest integral not greater in magnitude than @x.
 End Rem
 Function TruncF:Float( x:Float )="float bbTruncf(float)!"
 
+Rem
+bbdoc: Converts a degree value @x of type #Float to its equivalent radiant value.
+about: Takes a degree value and converts it to radiants using the formula: `radiant = degree × (π / 180)`.
+End Rem
+Function Deg2RadF:Float( x:Float )="float bbDeg2Radf(float)!"
+
+Rem
+bbdoc: Converts a radiant value @x of type #Float to its equivalent degree value.
+about: Takes a radiant value and converts it to degrees using the formula: `degree = radiant × (180 / π)`.
+End Rem
+Function Rad2DegF:Float( x:Float )="float bbRad2Degf(float)!"
+
 End Extern

math.mod/math.c

@@ -25,6 +25,8 @@ extern double bbLog10( double x );
 extern double bbCeil( double x );
 extern double bbRound( double x );
 extern double bbTrunc( double x );
+extern double bbDeg2Rad( double x );
+extern double bbRad2Deg( double x );
 
 extern int bbIsNanf( float x );
 extern int bbIsInff( float x );
@@ -46,6 +48,8 @@ extern float bbLog10f( float x );
 extern float bbCeilf( float x );
 extern float bbRoundf( float x );
 extern float bbTruncf( float x );
+extern float bbDeg2Radf( float x );
+extern float bbRad2Degf( float x );
 
 #else
 
@@ -109,6 +113,12 @@ double bbRound( double x ){
 double bbTrunc( double x ){
 	return trunc( x );
 }
+double bbDeg2Rad( double x ){
+	return x * DEG_TO_RAD;
+}
+double bbRad2Deg( double x ){
+	return x * RAD_TO_DEG;
+}
 
 #define RAD_TO_DEGF RAD_TO_DEG
 #define DEG_TO_RADF DEG_TO_RAD
@@ -186,4 +196,10 @@ float bbRoundf( float x ){
 float bbTruncf( float x ){
 	return truncf( x );
 }
+float bbDeg2Radf( float x ){
+	return x * DEG_TO_RADF;
+}
+float bbRad2Degf( float x ){
+	return x * RAD_TO_DEGF;
+}
 #endif

max2d.mod/imagefont.bmx

@@ -34,6 +34,7 @@ Type TImageFont
 	Field _src_font:TFont
 	Field _glyphs:TImageGlyph[]
 	Field _imageFlags:Int
+	Field _style:Int
 
 	Method Style:Int()
 		If _src_font Return _src_font.Style()
@@ -78,29 +79,82 @@ Type TImageFont
 		Return glyph
 		
 	End Method
-	
-	Method Draw( text:String,x:Float,y:Float,ix:Float,iy:Float,jx:Float,jy:Float )
+
+	Method LoadGlyphs:TImageGlyph[]( text:String )
+
+		Local src_glyph:TGlyph[]=_src_font.LoadGlyphs( text )
+
+		Local glyphs:TImageGlyph[]=New TImageGlyph[text.length]
 
 		For Local i:Int=0 Until text.length
+			Local src:TGlyph=src_glyph[i]
+
+			Local glyph:TImageGlyph=New TImageGlyph
+			glyphs[i]=glyph
+
+			If src Then
+				Local index:Int = src.Index()
+				Local cachedGlyph:TImageGlyph = _glyphs[index]
+
+				If cachedGlyph Then
+					glyph._image = cachedGlyph._image
+				End If
 		
-			Local n:Int=CharToGlyph( text[i] )
-			If n<0 Continue
-			
-			Local glyph:TImageGlyph=LoadGlyph(n)
-			Local image:TImage=glyph._image
+				glyph._advance=src.Advance()
+				src.GetRect glyph._x,glyph._y,glyph._w,glyph._h
+				If Not glyph._image
+					Local pixmap:TPixmap=TPixmap( src.Pixels() )
+					If pixmap Then
+						glyph._image=TImage.Load( pixmap.Copy(),_imageFlags,0,0,0 )
+					End If
+				End If
+			End If
+		Next
+
+		Return glyphs
+	End Method
+	
+	Method Draw( text:String,x:Float,y:Float,ix:Float,iy:Float,jx:Float,jy:Float )
+
+		If Not (_style & KERNFONT) Then
+			For Local i:Int=0 Until text.length
 			
-			If image
-				Local frame:TImageFrame=image.Frame(0)
-				If frame
-					Local tx:Float=glyph._x*ix+glyph._y*iy
-					Local ty:Float=glyph._x*jx+glyph._y*jy			
-					frame.Draw 0,0,image.width,image.height,x+tx,y+ty,0,0,image.width,image.height
+				Local n:Int=CharToGlyph( text[i] )
+				If n<0 Continue
+				
+				Local glyph:TImageGlyph=LoadGlyph(n)
+				Local image:TImage=glyph._image
+				
+				If image
+					Local frame:TImageFrame=image.Frame(0)
+					If frame
+						Local tx:Float=glyph._x*ix+glyph._y*iy
+						Local ty:Float=glyph._x*jx+glyph._y*jy			
+						frame.Draw 0,0,image.width,image.height,x+tx,y+ty,0,0,image.width,image.height
+					EndIf
 				EndIf
-			EndIf
-			
-			x:+glyph._advance*ix
-			y:+glyph._advance*jx
-		Next
+				
+				x:+glyph._advance*ix
+				y:+glyph._advance*jx
+			Next
+		Else
+			Local glyphs:TImageGlyph[] = LoadGlyphs( text )
+
+			For Local i:Int=0 Until glyphs.length
+				Local glyph:TImageGlyph=glyphs[i]
+				Local image:TImage=glyph._image
+				If image
+					Local frame:TImageFrame=image.Frame(0)
+					If frame
+						Local tx:Float=glyph._x*ix+glyph._y*iy
+						Local ty:Float=glyph._x*jx+glyph._y*jy			
+						frame.Draw 0,0,image.width,image.height,x+tx,y+ty,0,0,image.width,image.height
+					EndIf
+				EndIf
+				x:+glyph._advance*ix
+				y:+glyph._advance*jx
+			Next
+		End If
 		
 	End Method
 	
@@ -112,6 +166,7 @@ Type TImageFont
 		Local font:TImageFont=New TImageFont
 		font._src_font=src
 		font._glyphs=New TImageGlyph[src.CountGlyphs()]
+		font._style=style
 		If style & SMOOTHFONT font._imageFlags=FILTEREDIMAGE|MIPMAPPEDIMAGE
 		
 		Return font

rectpacker.mod/examples/example_01.bmx

@@ -0,0 +1,24 @@
+SuperStrict
+
+Framework BRL.standardio
+Import BRL.RectPacker
+
+Local packer:TRectPacker = New TRectPacker
+
+packer.Add(32, 32, 0)
+packer.Add(64, 64, 1)
+packer.Add(128, 128, 2)
+packer.Add(256, 256, 3)
+packer.Add(512, 512, 4)
+packer.Add(1024, 1024, 5)
+
+Local sheets:TPackedSheet[] = packer.Pack()
+
+For Local i:Int = 0 Until sheets.Length
+	Local sheet:TPackedSheet = sheets[i]
+	Print "Sheet: " + i + " : " + sheet.width + " " + sheet.height
+	For Local j:Int = 0 Until sheet.rects.Length
+		Local rect:SPackedRect = sheet.rects[j]
+		Print "  Rect: " + j + " " + rect.x + " " + rect.y + " " + rect.width + " " + rect.height
+	Next
+Next

rectpacker.mod/glue.cpp

@@ -0,0 +1,82 @@
+/*
+  Copyright (c) 2024-2025 Bruce A Henderson
+  
+  This software is provided 'as-is', without any express or implied
+  warranty. In no event will the authors be held liable for any damages
+  arising from the use of this software.
+  
+  Permission is granted to anyone to use this software for any purpose,
+  including commercial applications, and to alter it and redistribute it
+  freely, subject to the following restrictions:
+  
+  1. The origin of this software must not be misrepresented; you must not
+     claim that you wrote the original software. If you use this software
+     in a product, an acknowledgment in the product documentation would be
+     appreciated but is not required.
+  2. Altered source versions must be plainly marked as such, and must not be
+     misrepresented as being the original software.
+  3. This notice may not be removed or altered from any source distribution.
+*/ 
+#include "rect_pack.h"
+
+#include "brl.mod/blitz.mod/blitz.h"
+
+extern "C" {
+
+    void brl_rectpacker_TRectPacker__GetSize(BBObject * packer, int index, int * width, int * height, int * id);
+    BBArray * brl_rectpacker_TRectPacker__NewSheetArray(int size);
+    void brl_rectpacker_TRectPacker__SetSheet(BBArray * sheets, int index, BBObject * sheet);
+    BBObject * brl_rectpacker_TPackedSheet__Create(int width, int height, int size);
+    void brl_rectpacker_TPackedSheet__SetRect(BBObject * sheet, int index, int id, int x, int y, int width, int height, int rotated);
+
+    BBArray * bmx_rectpacker_pack(BBObject * packer, int packingMethod, int maxSheets, int powerOfTwo, int square, int allowRotate, int alignWidth, int borderPadding, int sheetPadding, int overAllocate, int minWidth, int minHeight, int maxWidth, int maxHeight, int count);
+}
+
+BBArray * bmx_rectpacker_pack(BBObject * packer, int packingMethod, int maxSheets, int powerOfTwo, int square, int allowRotate, int alignWidth, int borderPadding, int sheetPadding, int overAllocate, int minWidth, int minHeight, int maxWidth, int maxHeight, int count) {
+    rect_pack::Settings settings;
+    settings.method = static_cast<rect_pack::Method>(packingMethod);
+    settings.max_sheets = maxSheets;
+    settings.power_of_two = static_cast<bool>(powerOfTwo);
+    settings.square = static_cast<bool>(square);
+    settings.allow_rotate = static_cast<bool>(allowRotate);
+    settings.align_width = static_cast<bool>(alignWidth);
+    settings.border_padding = sheetPadding;
+    settings.over_allocate = overAllocate;
+    settings.min_width = minWidth;
+    settings.min_height = minHeight;
+    settings.max_width = maxWidth;
+    settings.max_height = maxHeight;
+
+    std::vector<rect_pack::Size> sizes;
+
+    for (int i = 0; i < count; i++) {
+        rect_pack::Size s;
+        brl_rectpacker_TRectPacker__GetSize(packer, i, &s.width, &s.height, &s.id);
+        if ( borderPadding > 0 ) {
+            s.width += borderPadding * 2;
+            s.height += borderPadding * 2;
+        }
+        sizes.push_back(s);
+    }
+
+    std::vector<rect_pack::Sheet> sheets = rect_pack::pack(settings, sizes);
+
+    BBArray * result = brl_rectpacker_TRectPacker__NewSheetArray(sheets.size());
+
+    for (int i = 0; i < sheets.size(); i++) {
+        BBObject * sheet = brl_rectpacker_TPackedSheet__Create(sheets[i].width, sheets[i].height, sheets[i].rects.size());
+        for (int j = 0; j < sheets[i].rects.size(); j++) {
+            rect_pack::Rect r = sheets[i].rects[j];
+            if ( borderPadding > 0 ) {
+                r.x += borderPadding;
+                r.y += borderPadding;
+                r.width -= borderPadding * 2;
+                r.height -= borderPadding * 2;
+            }
+            brl_rectpacker_TPackedSheet__SetRect(sheet, j, r.id, r.x, r.y, r.width, r.height, r.rotated);
+        }
+        brl_rectpacker_TRectPacker__SetSheet(result, i, sheet);
+    }
+
+    return result;
+}

rectpacker.mod/rect_pack/LICENSE

@@ -0,0 +1,24 @@
+This is free and unencumbered software released into the public domain.
+
+Anyone is free to copy, modify, publish, use, compile, sell, or
+distribute this software, either in source code form or as a compiled
+binary, for any purpose, commercial or non-commercial, and by any
+means.
+
+In jurisdictions that recognize copyright laws, the author or authors
+of this software dedicate any and all copyright interest in the
+software to the public domain. We make this dedication for the benefit
+of the public at large and to the detriment of our heirs and
+successors. We intend this dedication to be an overt act of
+relinquishment in perpetuity of all present and future rights to this
+software under copyright law.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+OTHER DEALINGS IN THE SOFTWARE.
+
+For more information, please refer to <https://unlicense.org>

rectpacker.mod/rect_pack/MaxRectsBinPack.cpp

@@ -0,0 +1,583 @@
+/** @file MaxRectsBinPack.cpp
+	@author Jukka Jylänki
+
+	@brief Implements different bin packer algorithms that use the MAXRECTS data structure.
+
+	This work is released to Public Domain, do whatever you want with it.
+*/
+#include <algorithm>
+#include <utility>
+#include <iostream>
+#include <limits>
+
+#include <cassert>
+#include <cstring>
+#include <cmath>
+
+#include "MaxRectsBinPack.h"
+
+#define RBP_ENABLE_OPTIMIZATIONS
+#define RBP_REVERSE_ORDER
+
+namespace rbp {
+
+#if defined(RBP_ENABLE_OPTIMIZATIONS)
+// order of elements after erased element is not stable
+template<typename C, typename It>
+void erase_unstable(C& container, const It& it) {
+	std::swap(*it, container.back());
+	container.pop_back();
+}
+#endif
+
+using namespace std;
+
+bool IsContainedIn(const Rect &a, const Rect &b)
+{
+	return a.x >= b.x && a.y >= b.y
+		&& a.x+a.width <= b.x+b.width
+		&& a.y+a.height <= b.y+b.height;
+}
+
+MaxRectsBinPack::MaxRectsBinPack()
+:binWidth(0),
+binHeight(0)
+{
+}
+
+MaxRectsBinPack::MaxRectsBinPack(int width, int height, bool allowFlip)
+{
+	Init(width, height, allowFlip);
+}
+
+void MaxRectsBinPack::Init(int width, int height, bool allowFlip)
+{
+	binAllowFlip = allowFlip;
+	binWidth = width;
+	binHeight = height;
+
+	Rect n;
+	n.x = 0;
+	n.y = 0;
+	n.width = width;
+	n.height = height;
+
+	usedRectangles.clear();
+
+	freeRectangles.clear();
+	freeRectangles.push_back(n);
+}
+
+Rect MaxRectsBinPack::Insert(int width, int height, FreeRectChoiceHeuristic method)
+{
+	Rect newNode;
+	// Unused in this function. We don't need to know the score after finding the position.
+	int score1 = std::numeric_limits<int>::max();
+	int score2 = std::numeric_limits<int>::max();
+	switch(method)
+	{
+		case RectBestShortSideFit: newNode = FindPositionForNewNodeBestShortSideFit(width, height, score1, score2); break;
+		case RectBottomLeftRule: newNode = FindPositionForNewNodeBottomLeft(width, height, score1, score2); break;
+		case RectContactPointRule: newNode = FindPositionForNewNodeContactPoint(width, height, score1); break;
+		case RectBestLongSideFit: newNode = FindPositionForNewNodeBestLongSideFit(width, height, score2, score1); break;
+		case RectBestAreaFit: newNode = FindPositionForNewNodeBestAreaFit(width, height, score1, score2); break;
+	}
+		
+	if (newNode.height == 0)
+		return newNode;
+
+	size_t numRectanglesToProcess = freeRectangles.size();
+	for(size_t i = 0; i < numRectanglesToProcess; ++i)
+	{
+		if (SplitFreeNode(freeRectangles[i], newNode))
+		{
+#if defined(RBP_ENABLE_OPTIMIZATIONS)
+			erase_unstable(freeRectangles, freeRectangles.begin() + i);
+#else
+			freeRectangles.erase(freeRectangles.begin() + i);
+#endif
+			--i;
+			--numRectanglesToProcess;
+		}
+	}
+
+	PruneFreeList();
+
+	usedRectangles.push_back(newNode);
+	return newNode;
+}
+
+void MaxRectsBinPack::Insert(std::vector<RectSize> &rects, std::vector<Rect> &dst, FreeRectChoiceHeuristic method)
+{
+	dst.clear();
+
+	while(rects.size() > 0)
+	{
+		int bestScore1 = std::numeric_limits<int>::max();
+		int bestScore2 = std::numeric_limits<int>::max();
+		int bestRectIndex = -1;
+		Rect bestNode;
+
+#if defined(RBP_REVERSE_ORDER)
+		for(int i = static_cast<int>(rects.size()) - 1; i >= 0; --i)
+#else
+		for(size_t i = 0; i < rects.size(); ++i)
+#endif
+		{
+			int score1;
+			int score2;
+			Rect newNode = ScoreRect(rects[i].width, rects[i].height, method, score1, score2);
+			newNode.id = rects[i].id;
+
+			if (score1 < bestScore1 || (score1 == bestScore1 && score2 < bestScore2))
+			{
+				bestScore1 = score1;
+				bestScore2 = score2;
+				bestNode = newNode;
+				bestRectIndex = i;
+			}
+		}
+
+		if (bestRectIndex == -1)
+			return;
+
+		PlaceRect(bestNode);
+		dst.push_back(bestNode);
+
+#if defined(RBP_ENABLE_OPTIMIZATIONS)
+		erase_unstable(rects, rects.begin() + bestRectIndex);
+#else
+		rects.erase(rects.begin() + bestRectIndex);
+#endif
+	}
+}
+
+void MaxRectsBinPack::PlaceRect(const Rect &node)
+{
+	size_t numRectanglesToProcess = freeRectangles.size();
+	for(size_t i = 0; i < numRectanglesToProcess; ++i)
+	{
+		if (SplitFreeNode(freeRectangles[i], node))
+		{
+#if defined(RBP_ENABLE_OPTIMIZATIONS)
+			const auto current = freeRectangles.begin() + i;
+			const auto last = freeRectangles.begin() + numRectanglesToProcess - 1;
+			std::swap(*current, *last);
+			erase_unstable(freeRectangles, last);
+#else
+			freeRectangles.erase(freeRectangles.begin() + i);
+#endif
+			--i;
+			--numRectanglesToProcess;
+		}
+	}
+
+	PruneFreeList();
+
+	usedRectangles.push_back(node);
+}
+
+Rect MaxRectsBinPack::ScoreRect(int width, int height, FreeRectChoiceHeuristic method, int &score1, int &score2) const
+{
+	Rect newNode;
+	score1 = std::numeric_limits<int>::max();
+	score2 = std::numeric_limits<int>::max();
+	switch(method)
+	{
+	case RectBestShortSideFit: newNode = FindPositionForNewNodeBestShortSideFit(width, height, score1, score2); break;
+	case RectBottomLeftRule: newNode = FindPositionForNewNodeBottomLeft(width, height, score1, score2); break;
+	case RectContactPointRule: newNode = FindPositionForNewNodeContactPoint(width, height, score1); 
+		score1 = -score1; // Reverse since we are minimizing, but for contact point score bigger is better.
+		break;
+	case RectBestLongSideFit: newNode = FindPositionForNewNodeBestLongSideFit(width, height, score2, score1); break;
+	case RectBestAreaFit: newNode = FindPositionForNewNodeBestAreaFit(width, height, score1, score2); break;
+	}
+
+	// Cannot fit the current rectangle.
+	if (newNode.height == 0)
+	{
+		score1 = std::numeric_limits<int>::max();
+		score2 = std::numeric_limits<int>::max();
+	}
+
+	return newNode;
+}
+
+/// Computes the ratio of used surface area.
+float MaxRectsBinPack::Occupancy() const
+{
+	unsigned long usedSurfaceArea = 0;
+	for(size_t i = 0; i < usedRectangles.size(); ++i)
+		usedSurfaceArea += usedRectangles[i].width * usedRectangles[i].height;
+
+	return (float)usedSurfaceArea / (binWidth * binHeight);
+}
+
+std::pair<int, int> MaxRectsBinPack::BottomRight() const
+{
+	int x = 0;
+	int y = 0;
+	for(size_t i = 0; i < usedRectangles.size(); ++i) {
+		x = std::max(x, usedRectangles[i].x + usedRectangles[i].width);
+		y = std::max(y, usedRectangles[i].y + usedRectangles[i].height);
+  }
+	return { x, y };
+}
+
+Rect MaxRectsBinPack::FindPositionForNewNodeBottomLeft(int width, int height, int &bestY, int &bestX) const
+{
+	Rect bestNode;
+	memset(&bestNode, 0, sizeof(Rect));
+
+	bestY = std::numeric_limits<int>::max();
+	bestX = std::numeric_limits<int>::max();
+
+	for(size_t i = 0; i < freeRectangles.size(); ++i)
+	{
+		// Try to place the rectangle in upright (non-flipped) orientation.
+		if (freeRectangles[i].width >= width && freeRectangles[i].height >= height)
+		{
+			int topSideY = freeRectangles[i].y + height;
+			if (topSideY < bestY || (topSideY == bestY && freeRectangles[i].x < bestX))
+			{
+				bestNode.x = freeRectangles[i].x;
+				bestNode.y = freeRectangles[i].y;
+				bestNode.width = width;
+				bestNode.height = height;
+				bestY = topSideY;
+				bestX = freeRectangles[i].x;
+			}
+		}
+		if (binAllowFlip && freeRectangles[i].width >= height && freeRectangles[i].height >= width)
+		{
+			int topSideY = freeRectangles[i].y + width;
+			if (topSideY < bestY || (topSideY == bestY && freeRectangles[i].x < bestX))
+			{
+				bestNode.x = freeRectangles[i].x;
+				bestNode.y = freeRectangles[i].y;
+				bestNode.width = height;
+				bestNode.height = width;
+				bestY = topSideY;
+				bestX = freeRectangles[i].x;
+			}
+		}
+	}
+	return bestNode;
+}
+
+Rect MaxRectsBinPack::FindPositionForNewNodeBestShortSideFit(int width, int height, 
+	int &bestShortSideFit, int &bestLongSideFit) const
+{
+	Rect bestNode;
+	memset(&bestNode, 0, sizeof(Rect));
+
+	bestShortSideFit = std::numeric_limits<int>::max();
+	bestLongSideFit = std::numeric_limits<int>::max();
+
+	for(size_t i = 0; i < freeRectangles.size(); ++i)
+	{
+		// Try to place the rectangle in upright (non-flipped) orientation.
+		if (freeRectangles[i].width >= width && freeRectangles[i].height >= height)
+		{
+			int leftoverHoriz = abs(freeRectangles[i].width - width);
+			int leftoverVert = abs(freeRectangles[i].height - height);
+			int shortSideFit = min(leftoverHoriz, leftoverVert);
+			int longSideFit = max(leftoverHoriz, leftoverVert);
+
+			if (shortSideFit < bestShortSideFit || (shortSideFit == bestShortSideFit && longSideFit < bestLongSideFit))
+			{
+				bestNode.x = freeRectangles[i].x;
+				bestNode.y = freeRectangles[i].y;
+				bestNode.width = width;
+				bestNode.height = height;
+				bestShortSideFit = shortSideFit;
+				bestLongSideFit = longSideFit;
+			}
+		}
+
+		if (binAllowFlip && freeRectangles[i].width >= height && freeRectangles[i].height >= width)
+		{
+			int flippedLeftoverHoriz = abs(freeRectangles[i].width - height);
+			int flippedLeftoverVert = abs(freeRectangles[i].height - width);
+			int flippedShortSideFit = min(flippedLeftoverHoriz, flippedLeftoverVert);
+			int flippedLongSideFit = max(flippedLeftoverHoriz, flippedLeftoverVert);
+
+			if (flippedShortSideFit < bestShortSideFit || (flippedShortSideFit == bestShortSideFit && flippedLongSideFit < bestLongSideFit))
+			{
+				bestNode.x = freeRectangles[i].x;
+				bestNode.y = freeRectangles[i].y;
+				bestNode.width = height;
+				bestNode.height = width;
+				bestShortSideFit = flippedShortSideFit;
+				bestLongSideFit = flippedLongSideFit;
+			}
+		}
+	}
+	return bestNode;
+}
+
+Rect MaxRectsBinPack::FindPositionForNewNodeBestLongSideFit(int width, int height, 
+	int &bestShortSideFit, int &bestLongSideFit) const
+{
+	Rect bestNode;
+	memset(&bestNode, 0, sizeof(Rect));
+
+	bestShortSideFit = std::numeric_limits<int>::max();
+	bestLongSideFit = std::numeric_limits<int>::max();
+
+	for(size_t i = 0; i < freeRectangles.size(); ++i)
+	{
+		// Try to place the rectangle in upright (non-flipped) orientation.
+		if (freeRectangles[i].width >= width && freeRectangles[i].height >= height)
+		{
+			int leftoverHoriz = abs(freeRectangles[i].width - width);
+			int leftoverVert = abs(freeRectangles[i].height - height);
+			int shortSideFit = min(leftoverHoriz, leftoverVert);
+			int longSideFit = max(leftoverHoriz, leftoverVert);
+
+			if (longSideFit < bestLongSideFit || (longSideFit == bestLongSideFit && shortSideFit < bestShortSideFit))
+			{
+				bestNode.x = freeRectangles[i].x;
+				bestNode.y = freeRectangles[i].y;
+				bestNode.width = width;
+				bestNode.height = height;
+				bestShortSideFit = shortSideFit;
+				bestLongSideFit = longSideFit;
+			}
+		}
+
+		if (binAllowFlip && freeRectangles[i].width >= height && freeRectangles[i].height >= width)
+		{
+			int leftoverHoriz = abs(freeRectangles[i].width - height);
+			int leftoverVert = abs(freeRectangles[i].height - width);
+			int shortSideFit = min(leftoverHoriz, leftoverVert);
+			int longSideFit = max(leftoverHoriz, leftoverVert);
+
+			if (longSideFit < bestLongSideFit || (longSideFit == bestLongSideFit && shortSideFit < bestShortSideFit))
+			{
+				bestNode.x = freeRectangles[i].x;
+				bestNode.y = freeRectangles[i].y;
+				bestNode.width = height;
+				bestNode.height = width;
+				bestShortSideFit = shortSideFit;
+				bestLongSideFit = longSideFit;
+			}
+		}
+	}
+	return bestNode;
+}
+
+Rect MaxRectsBinPack::FindPositionForNewNodeBestAreaFit(int width, int height, 
+	int &bestAreaFit, int &bestShortSideFit) const
+{
+	Rect bestNode;
+	memset(&bestNode, 0, sizeof(Rect));
+
+	bestAreaFit = std::numeric_limits<int>::max();
+	bestShortSideFit = std::numeric_limits<int>::max();
+
+	for(size_t i = 0; i < freeRectangles.size(); ++i)
+	{
+		int areaFit = freeRectangles[i].width * freeRectangles[i].height - width * height;
+
+		// Try to place the rectangle in upright (non-flipped) orientation.
+		if (freeRectangles[i].width >= width && freeRectangles[i].height >= height)
+		{
+			int leftoverHoriz = abs(freeRectangles[i].width - width);
+			int leftoverVert = abs(freeRectangles[i].height - height);
+			int shortSideFit = min(leftoverHoriz, leftoverVert);
+
+			if (areaFit < bestAreaFit || (areaFit == bestAreaFit && shortSideFit < bestShortSideFit))
+			{
+				bestNode.x = freeRectangles[i].x;
+				bestNode.y = freeRectangles[i].y;
+				bestNode.width = width;
+				bestNode.height = height;
+				bestShortSideFit = shortSideFit;
+				bestAreaFit = areaFit;
+			}
+		}
+
+		if (binAllowFlip && freeRectangles[i].width >= height && freeRectangles[i].height >= width)
+		{
+			int leftoverHoriz = abs(freeRectangles[i].width - height);
+			int leftoverVert = abs(freeRectangles[i].height - width);
+			int shortSideFit = min(leftoverHoriz, leftoverVert);
+
+			if (areaFit < bestAreaFit || (areaFit == bestAreaFit && shortSideFit < bestShortSideFit))
+			{
+				bestNode.x = freeRectangles[i].x;
+				bestNode.y = freeRectangles[i].y;
+				bestNode.width = height;
+				bestNode.height = width;
+				bestShortSideFit = shortSideFit;
+				bestAreaFit = areaFit;
+			}
+		}
+	}
+	return bestNode;
+}
+
+/// Returns 0 if the two intervals i1 and i2 are disjoint, or the length of their overlap otherwise.
+int CommonIntervalLength(int i1start, int i1end, int i2start, int i2end)
+{
+	if (i1end < i2start || i2end < i1start)
+		return 0;
+	return min(i1end, i2end) - max(i1start, i2start);
+}
+
+int MaxRectsBinPack::ContactPointScoreNode(int x, int y, int width, int height) const
+{
+	int score = 0;
+
+	if (x == 0 || x + width == binWidth)
+		score += height;
+	if (y == 0 || y + height == binHeight)
+		score += width;
+
+	for(size_t i = 0; i < usedRectangles.size(); ++i)
+	{
+		if (usedRectangles[i].x == x + width || usedRectangles[i].x + usedRectangles[i].width == x)
+			score += CommonIntervalLength(usedRectangles[i].y, usedRectangles[i].y + usedRectangles[i].height, y, y + height);
+		if (usedRectangles[i].y == y + height || usedRectangles[i].y + usedRectangles[i].height == y)
+			score += CommonIntervalLength(usedRectangles[i].x, usedRectangles[i].x + usedRectangles[i].width, x, x + width);
+	}
+	return score;
+}
+
+Rect MaxRectsBinPack::FindPositionForNewNodeContactPoint(int width, int height, int &bestContactScore) const
+{
+	Rect bestNode;
+	memset(&bestNode, 0, sizeof(Rect));
+
+	bestContactScore = -1;
+
+	for(size_t i = 0; i < freeRectangles.size(); ++i)
+	{
+		// Try to place the rectangle in upright (non-flipped) orientation.
+		if (freeRectangles[i].width >= width && freeRectangles[i].height >= height)
+		{
+			int score = ContactPointScoreNode(freeRectangles[i].x, freeRectangles[i].y, width, height);
+			if (score > bestContactScore)
+			{
+				bestNode.x = freeRectangles[i].x;
+				bestNode.y = freeRectangles[i].y;
+				bestNode.width = width;
+				bestNode.height = height;
+				bestContactScore = score;
+			}
+		}
+		if (binAllowFlip && freeRectangles[i].width >= height && freeRectangles[i].height >= width)
+		{
+			int score = ContactPointScoreNode(freeRectangles[i].x, freeRectangles[i].y, height, width);
+			if (score > bestContactScore)
+			{
+				bestNode.x = freeRectangles[i].x;
+				bestNode.y = freeRectangles[i].y;
+				bestNode.width = height;
+				bestNode.height = width;
+				bestContactScore = score;
+			}
+		}
+	}
+	return bestNode;
+}
+
+bool MaxRectsBinPack::SplitFreeNode(Rect freeNode, const Rect &usedNode)
+{
+	// Test with SAT if the rectangles even intersect.
+	if (usedNode.x >= freeNode.x + freeNode.width || usedNode.x + usedNode.width <= freeNode.x ||
+		usedNode.y >= freeNode.y + freeNode.height || usedNode.y + usedNode.height <= freeNode.y)
+		return false;
+
+	if (usedNode.x < freeNode.x + freeNode.width && usedNode.x + usedNode.width > freeNode.x)
+	{
+		// New node at the top side of the used node.
+		if (usedNode.y > freeNode.y && usedNode.y < freeNode.y + freeNode.height)
+		{
+			Rect newNode = freeNode;
+			newNode.height = usedNode.y - newNode.y;
+			freeRectangles.push_back(newNode);
+		}
+
+		// New node at the bottom side of the used node.
+		if (usedNode.y + usedNode.height < freeNode.y + freeNode.height)
+		{
+			Rect newNode = freeNode;
+			newNode.y = usedNode.y + usedNode.height;
+			newNode.height = freeNode.y + freeNode.height - (usedNode.y + usedNode.height);
+			freeRectangles.push_back(newNode);
+		}
+	}
+
+	if (usedNode.y < freeNode.y + freeNode.height && usedNode.y + usedNode.height > freeNode.y)
+	{
+		// New node at the left side of the used node.
+		if (usedNode.x > freeNode.x && usedNode.x < freeNode.x + freeNode.width)
+		{
+			Rect newNode = freeNode;
+			newNode.width = usedNode.x - newNode.x;
+			freeRectangles.push_back(newNode);
+		}
+
+		// New node at the right side of the used node.
+		if (usedNode.x + usedNode.width < freeNode.x + freeNode.width)
+		{
+			Rect newNode = freeNode;
+			newNode.x = usedNode.x + usedNode.width;
+			newNode.width = freeNode.x + freeNode.width - (usedNode.x + usedNode.width);
+			freeRectangles.push_back(newNode);
+		}
+	}
+
+	return true;
+}
+
+void MaxRectsBinPack::PruneFreeList()
+{
+	/* 
+	///  Would be nice to do something like this, to avoid a Theta(n^2) loop through each pair.
+	///  But unfortunately it doesn't quite cut it, since we also want to detect containment. 
+	///  Perhaps there's another way to do this faster than Theta(n^2).
+
+	if (freeRectangles.size() > 0)
+		clb::sort::QuickSort(&freeRectangles[0], freeRectangles.size(), NodeSortCmp);
+
+	for(size_t i = 0; i < freeRectangles.size()-1; ++i)
+		if (freeRectangles[i].x == freeRectangles[i+1].x &&
+		    freeRectangles[i].y == freeRectangles[i+1].y &&
+		    freeRectangles[i].width == freeRectangles[i+1].width &&
+		    freeRectangles[i].height == freeRectangles[i+1].height)
+		{
+			freeRectangles.erase(freeRectangles.begin() + i);
+			--i;
+		}
+	*/
+
+	/// Go through each pair and remove any rectangle that is redundant.
+	for(size_t i = 0; i < freeRectangles.size(); ++i)
+		for(size_t j = i+1; j < freeRectangles.size(); ++j)
+		{
+			if (IsContainedIn(freeRectangles[i], freeRectangles[j]))
+			{
+#if defined(RBP_ENABLE_OPTIMIZATIONS)
+				erase_unstable(freeRectangles, freeRectangles.begin()+i);
+#else
+				freeRectangles.erase(freeRectangles.begin()+i);
+#endif
+				--i;
+				break;
+			}
+			if (IsContainedIn(freeRectangles[j], freeRectangles[i]))
+			{
+#if defined(RBP_ENABLE_OPTIMIZATIONS)
+				erase_unstable(freeRectangles, freeRectangles.begin()+j);
+#else
+				freeRectangles.erase(freeRectangles.begin()+j);
+#endif
+				--j;
+			}
+		}
+}
+
+}

rectpacker.mod/rect_pack/MaxRectsBinPack.h

@@ -0,0 +1,107 @@
+/** @file MaxRectsBinPack.h
+	@author Jukka Jylänki
+
+	@brief Implements different bin packer algorithms that use the MAXRECTS data structure.
+
+	This work is released to Public Domain, do whatever you want with it.
+*/
+#pragma once
+
+#include <vector>
+#include <utility>
+
+namespace rbp {
+
+struct RectSize
+{
+	int width;
+	int height;
+
+	int id;
+};
+
+struct Rect
+{
+	int x;
+	int y;
+	int width;
+	int height;
+
+	int id;
+};
+
+/** MaxRectsBinPack implements the MAXRECTS data structure and different bin packing algorithms that 
+	use this structure. */
+class MaxRectsBinPack
+{
+public:
+	/// Instantiates a bin of size (0,0). Call Init to create a new bin.
+	MaxRectsBinPack();
+
+	/// Instantiates a bin of the given size.
+	/// @param allowFlip Specifies whether the packing algorithm is allowed to rotate the input rectangles by 90 degrees to consider a better placement.
+	MaxRectsBinPack(int width, int height, bool allowFlip = true);
+
+	/// (Re)initializes the packer to an empty bin of width x height units. Call whenever
+	/// you need to restart with a new bin.
+	void Init(int width, int height, bool allowFlip = true);
+
+	/// Specifies the different heuristic rules that can be used when deciding where to place a new rectangle.
+	enum FreeRectChoiceHeuristic
+	{
+		RectBestShortSideFit, ///< -BSSF: Positions the rectangle against the short side of a free rectangle into which it fits the best.
+		RectBestLongSideFit, ///< -BLSF: Positions the rectangle against the long side of a free rectangle into which it fits the best.
+		RectBestAreaFit, ///< -BAF: Positions the rectangle into the smallest free rect into which it fits.
+		RectBottomLeftRule, ///< -BL: Does the Tetris placement.
+		RectContactPointRule ///< -CP: Choosest the placement where the rectangle touches other rects as much as possible.
+	};
+
+	/// Inserts the given list of rectangles in an offline/batch mode, possibly rotated.
+	/// @param rects The list of rectangles to insert. This vector will be destroyed in the process.
+	/// @param dst [out] This list will contain the packed rectangles. The indices will not correspond to that of rects.
+	/// @param method The rectangle placement rule to use when packing.
+	void Insert(std::vector<RectSize> &rects, std::vector<Rect> &dst, FreeRectChoiceHeuristic method);
+
+	/// Inserts a single rectangle into the bin, possibly rotated.
+	Rect Insert(int width, int height, FreeRectChoiceHeuristic method);
+
+	/// Computes the ratio of used surface area to the total bin area.
+	float Occupancy() const;
+
+  std::pair<int, int> BottomRight() const;
+
+private:
+	int binWidth;
+	int binHeight;
+
+	bool binAllowFlip;
+
+	std::vector<Rect> usedRectangles;
+	std::vector<Rect> freeRectangles;
+
+	/// Computes the placement score for placing the given rectangle with the given method.
+	/// @param score1 [out] The primary placement score will be outputted here.
+	/// @param score2 [out] The secondary placement score will be outputted here. This isu sed to break ties.
+	/// @return This struct identifies where the rectangle would be placed if it were placed.
+	Rect ScoreRect(int width, int height, FreeRectChoiceHeuristic method, int &score1, int &score2) const;
+
+	/// Places the given rectangle into the bin.
+	void PlaceRect(const Rect &node);
+
+	/// Computes the placement score for the -CP variant.
+	int ContactPointScoreNode(int x, int y, int width, int height) const;
+
+	Rect FindPositionForNewNodeBottomLeft(int width, int height, int &bestY, int &bestX) const;
+	Rect FindPositionForNewNodeBestShortSideFit(int width, int height, int &bestShortSideFit, int &bestLongSideFit) const;
+	Rect FindPositionForNewNodeBestLongSideFit(int width, int height, int &bestShortSideFit, int &bestLongSideFit) const;
+	Rect FindPositionForNewNodeBestAreaFit(int width, int height, int &bestAreaFit, int &bestShortSideFit) const;
+	Rect FindPositionForNewNodeContactPoint(int width, int height, int &contactScore) const;
+
+	/// @return True if the free node was split.
+	bool SplitFreeNode(Rect freeNode, const Rect &usedNode);
+
+	/// Goes through the free rectangle list and removes any redundant entries.
+	void PruneFreeList();
+};
+
+}

rectpacker.mod/rect_pack/README.md

@@ -0,0 +1,70 @@
+# Multi-sheet rectangle packing
+
+A C++17 library for packing rectangles to one or more sprite sheets/atlases with optional constraints.
+
+It is part of the [spright](https://github.com/houmain/spright) project and utilizing [Sean T. Barrett's Skyline](https://github.com/nothings/stb) and  [Jukka Jylänki's MaxRects](https://github.com/juj/RectangleBinPack) packing algorithm implementations.
+
+Simply pass your sheet constraints and the rectangle sizes to the _pack_ function. It will return one or more sheets with rectangle positions. The _id_ can be used to correlate in- and output (there are no rectangles for sizes which did not fit).
+
+For now the header may serve as documentation:
+
+## rect_pack.h
+
+```cpp
+#include <vector>
+
+namespace rect_pack {
+
+enum class Method {
+  Best,
+  Best_Skyline,
+  Best_MaxRects,
+  Skyline_BottomLeft,
+  Skyline_BestFit,
+  MaxRects_BestShortSideFit,
+  MaxRects_BestLongSideFit,
+  MaxRects_BestAreaFit,
+  MaxRects_BottomLeftRule,
+  MaxRects_ContactPointRule
+};
+
+struct Size {
+  int id;
+  int width;
+  int height;
+};
+
+struct Rect {
+  int id;
+  int x;
+  int y;
+  int width;
+  int height;
+  bool rotated;
+};
+
+struct Sheet {
+  int width;
+  int height;
+  std::vector<Rect> rects;
+};
+
+struct Settings {
+  Method method;
+  int max_sheets;
+  bool power_of_two;
+  bool square;
+  bool allow_rotate;
+  int align_width;
+  int border_padding;
+  int over_allocate;
+  int min_width;
+  int min_height;
+  int max_width;
+  int max_height;
+};
+
+std::vector<Sheet> pack(Settings settings, std::vector<Size> sizes);
+
+} // namespace
+```

rectpacker.mod/rect_pack/rect_pack.cpp

@@ -0,0 +1,596 @@
+
+#define STBRP_LARGE_RECTS
+#include "rect_pack.h"
+#include "MaxRectsBinPack.h"
+#include "stb_rect_pack.h"
+#include <optional>
+#include <algorithm>
+#include <cmath>
+#include <cassert>
+
+namespace rect_pack {
+
+namespace {
+  const auto first_Skyline_method = Method::Skyline_BottomLeft;
+  const auto last_Skyline_method = Method::Skyline_BestFit;
+  const auto first_MaxRects_method = Method::MaxRects_BestShortSideFit;
+  const auto last_MaxRects_method = Method::MaxRects_ContactPointRule;
+
+  int floor(int v, int q) { return (v / q) * q; };
+  int ceil(int v, int q) { return ((v + q - 1) / q) * q; };
+  int sqrt(int a) { return static_cast<int>(std::sqrt(a)); }
+  int div_ceil(int a, int b) { return (b > 0 ? (a + b - 1) / b : -1); }
+
+  int ceil_to_pot(int value) {
+    for (auto pot = 1; ; pot <<= 1)
+      if (pot >= value)
+        return pot;
+  }
+
+  int floor_to_pot(int value) {
+    for (auto pot = 1; ; pot <<= 1)
+      if (pot > value)
+        return (pot >> 1);
+  }
+
+  bool is_stb_method(Method method) {
+    const auto first = static_cast<int>(first_Skyline_method);
+    const auto last = static_cast<int>(last_Skyline_method);
+    const auto index = static_cast<int>(method);
+    return (index >= first && index <= last);
+  }
+
+  bool is_rbp_method(Method method) {
+    const auto first = static_cast<int>(first_MaxRects_method);
+    const auto last = static_cast<int>(last_MaxRects_method);
+    const auto index = static_cast<int>(method);
+    return (index >= first && index <= last);
+  }
+
+  int to_stb_method(Method method) {
+    assert(is_stb_method(method));
+    return static_cast<int>(method) - static_cast<int>(first_Skyline_method);
+  }
+
+  rbp::MaxRectsBinPack::FreeRectChoiceHeuristic to_rbp_method(Method method) {
+    assert(is_rbp_method(method));
+    return static_cast<rbp::MaxRectsBinPack::FreeRectChoiceHeuristic>(
+      static_cast<int>(method) - static_cast<int>(first_MaxRects_method));
+  }
+
+  std::vector<Method> get_concrete_methods(Method settings_method) {
+    auto methods = std::vector<Method>();
+    const auto add_skyline_methods = [&]() {
+      methods.insert(end(methods), {
+        Method::Skyline_BottomLeft,
+        Method::Skyline_BestFit
+      });
+    };
+    const auto add_maxrect_methods = [&]() {
+      methods.insert(end(methods), {
+        Method::MaxRects_BestShortSideFit,
+        Method::MaxRects_BestLongSideFit,
+        Method::MaxRects_BestAreaFit,
+        Method::MaxRects_BottomLeftRule,
+        // do not automatically try costy contact point rule
+        // Method::MaxRects_ContactPointRule,
+      });
+    };
+    switch (settings_method) {
+      case Method::Best:
+        add_skyline_methods();
+        add_maxrect_methods();
+        break;
+
+      case Method::Best_Skyline:
+        add_skyline_methods();
+        break;
+
+      case Method::Best_MaxRects:
+        add_maxrect_methods();
+        break;
+
+      default:
+        methods.push_back(settings_method);
+        break;
+    }
+    return methods;
+  }
+
+  bool can_fit(const Settings& settings, int width, int height) {
+    return ((width <= settings.max_width &&
+             height <= settings.max_height) ||
+             (settings.allow_rotate &&
+              width <= settings.max_height &&
+              height <= settings.max_width));
+  }
+
+  void apply_padding(const Settings& settings, int& width, int& height, bool indent) {
+    const auto dir = (indent ? 1 : -1);
+    width -= dir * settings.border_padding * 2;
+    height -= dir * settings.border_padding * 2;
+    width += dir * settings.over_allocate;
+    height += dir * settings.over_allocate;
+  }
+
+  bool correct_settings(Settings& settings, std::vector<Size>& sizes) {
+    // clamp max to far less than numeric_limits<int>::max() to prevent overflow
+    const auto size_limit = 1'000'000'000;
+    if (settings.max_width <= 0 || settings.max_width > size_limit)
+      settings.max_width = size_limit;
+    if (settings.max_height <= 0 || settings.max_height > size_limit)
+      settings.max_height = size_limit;
+
+    if (settings.min_width < 0 ||
+        settings.min_height < 0 ||
+        settings.min_width > settings.max_width ||
+        settings.min_height > settings.max_height)
+      return false;
+
+    // immediately apply padding and over allocation, only relevant for power-of-two and alignment constraint
+    apply_padding(settings, settings.min_width, settings.min_height, true);
+    apply_padding(settings, settings.max_width, settings.max_height, true);
+
+    auto max_rect_width = 0;
+    auto max_rect_height = 0;
+    for (auto it = begin(sizes); it != end(sizes); )
+      if (it->width <= 0 ||
+          it->height <= 0 ||
+          !can_fit(settings, it->width, it->height)) {
+        it = sizes.erase(it);
+      }
+      else {
+        if (settings.allow_rotate && 
+            it->height > it->width && 
+            it->height <= settings.max_width &&
+            it->width <= settings.max_height) {
+          max_rect_width = std::max(max_rect_width, it->height);
+          max_rect_height = std::max(max_rect_height, it->width);
+        }
+        else {
+          max_rect_width = std::max(max_rect_width, it->width);
+          max_rect_height = std::max(max_rect_height, it->height);
+        }
+        ++it;
+      }
+
+    settings.min_width = std::max(settings.min_width, max_rect_width);
+    settings.min_height = std::max(settings.min_height, max_rect_height);
+
+    // clamp min to max and still pack the sprites which fit
+    settings.min_width = std::min(settings.min_width, settings.max_width);
+    settings.min_height = std::min(settings.min_height, settings.max_height);
+    return true;
+  }
+
+  struct Run {
+    Method method;
+    int width;
+    int height;
+    std::vector<Sheet> sheets;
+    int total_area;
+  };
+
+  void correct_size(const Settings& settings, int& width, int& height) {
+    width = std::max(width, settings.min_width);
+    height = std::max(height, settings.min_height);
+    apply_padding(settings, width, height, false);
+
+    if (settings.power_of_two) {
+      width = ceil_to_pot(width);
+      height = ceil_to_pot(height);
+    }
+
+    if (settings.align_width)
+      width = ceil(width, settings.align_width);
+
+    if (settings.square)
+      width = height = std::max(width, height);
+
+    apply_padding(settings, width, height, true);
+    width = std::min(width, settings.max_width);
+    height = std::min(height, settings.max_height);
+    apply_padding(settings, width, height, false);
+
+    if (settings.power_of_two) {
+      width = floor_to_pot(width);
+      height = floor_to_pot(height);
+    }
+
+    if (settings.align_width)
+      width = floor(width, settings.align_width);
+
+    if (settings.square)
+      width = height = std::min(width, height);
+
+    apply_padding(settings, width, height, true);
+  }
+
+  bool is_better_than(const Run& a, const Run& b, bool a_incomplete = false) {
+    if (a_incomplete) {
+      if (b.sheets.size() <= a.sheets.size())
+        return false;
+    }
+    else {
+      if (a.sheets.size() < b.sheets.size())
+        return true;
+      if (b.sheets.size() < a.sheets.size())
+        return false;
+    }
+    return (a.total_area < b.total_area);
+  }
+
+  int get_perfect_area(const std::vector<Size>& sizes) {
+    auto area = 0;
+    for (const auto& size : sizes)
+      area += size.width * size.height;
+    return area;
+  }
+
+  std::pair<int, int> get_run_size(const Settings& settings, int area) {
+    auto width = sqrt(area);
+    auto height = div_ceil(area, width);
+    if (width < settings.min_width || width > settings.max_width) {
+      width = std::clamp(width, settings.min_width, settings.max_width);
+      height = div_ceil(area, width);
+    }
+    else if (height < settings.min_height || height > settings.max_height) {
+      height = std::clamp(height, settings.min_height, settings.max_height);
+      width = div_ceil(area, height);
+    }
+    correct_size(settings, width, height);
+    return { width, height };
+  }
+
+  std::pair<int, int> get_initial_run_size(const Settings& settings, int perfect_area) {
+    return get_run_size(settings, perfect_area * 5 / 4);
+  }
+
+  enum class OptimizationStage {
+    first_run,
+    minimize_sheet_count,
+    shrink_square,
+    shrink_width_fast,
+    shrink_height_fast,
+    shrink_width_slow,
+    shrink_height_slow,
+    end
+  };
+
+  struct OptimizationState {
+    const int perfect_area;
+    int width;
+    int height;
+    OptimizationStage stage;
+    int iteration;
+  };
+
+  bool advance(OptimizationStage& stage) {
+    if (stage == OptimizationStage::end)
+      return false;
+    stage = static_cast<OptimizationStage>(static_cast<int>(stage) + 1);
+    return true;
+  }
+
+  // returns true when stage should be kept, false to advance
+  bool optimize_stage(OptimizationState& state,
+      const Settings& pack_settings, const Run& best_run) {
+
+    switch (state.stage) {
+      case OptimizationStage::first_run:
+      case OptimizationStage::end:
+        return false;
+
+      case OptimizationStage::minimize_sheet_count: {
+        if (best_run.sheets.size() <= 1 ||
+            state.iteration > 5)
+          return false;
+
+        const auto& last_sheet = best_run.sheets.back();
+        auto area = last_sheet.width * last_sheet.height;
+        for (auto i = 0; area > 0; ++i) {
+          if (state.width == pack_settings.max_width &&
+              state.height == pack_settings.max_height)
+            break;
+          if (state.height == pack_settings.max_height ||
+              (state.width < pack_settings.max_width && i % 2)) {
+            ++state.width;
+            area -= state.height;
+          }
+          else {
+            ++state.height;
+            area -= state.width;
+          }
+        }
+        return true;
+      }
+
+      case OptimizationStage::shrink_square: {
+        if (state.width != best_run.width ||
+            state.height != best_run.height ||
+            state.iteration > 5)
+          return false;
+
+        const auto [width, height] = get_run_size(pack_settings, state.perfect_area);
+        state.width = (state.width + width) / 2;
+        state.height = (state.height + height) / 2;
+        return true;
+      }
+
+      case OptimizationStage::shrink_width_fast:
+      case OptimizationStage::shrink_height_fast:
+      case OptimizationStage::shrink_width_slow:
+      case OptimizationStage::shrink_height_slow: {
+        if (state.iteration > 5)
+          return false;
+
+        const auto [width, height] = get_run_size(pack_settings, state.perfect_area);
+        switch (state.stage) {
+          default:
+          case OptimizationStage::shrink_width_fast:
+            if (state.width > width + 4)
+              state.width = (state.width + width) / 2;
+            break;
+          case OptimizationStage::shrink_height_fast:
+            if (state.height > height + 4)
+              state.height = (state.height + height) / 2;
+            break;
+          case OptimizationStage::shrink_width_slow:
+            if (state.width > width)
+              --state.width;
+            break;
+          case OptimizationStage::shrink_height_slow:
+            if (state.height > height)
+              --state.height;
+            break;
+        }
+        return true;
+      }
+    }
+    return false;
+  }
+
+  bool optimize_run_settings(OptimizationState& state,
+      const Settings& pack_settings, const Run& best_run) {
+
+    const auto previous_state = state;
+    for (;;) {
+      if (!optimize_stage(state, pack_settings, best_run))
+        if (advance(state.stage)) {
+          state.width = best_run.width;
+          state.height = best_run.height;
+          state.iteration = 0;
+          continue;
+        }
+
+      if (state.stage == OptimizationStage::end)
+        return false;
+
+      ++state.iteration;
+
+      auto width = state.width;
+      auto height = state.height;
+      correct_size(pack_settings, width, height);
+      if (width != previous_state.width ||
+          height != previous_state.height) {
+        state.width = width;
+        state.height = height;
+        return true;
+      }
+    }
+  }
+
+  template<typename T>
+  void copy_vector(const T& source, T& dest) {
+    dest.resize(source.size());
+    std::copy(begin(source), end(source), begin(dest));
+  }
+
+  struct RbpState {
+    rbp::MaxRectsBinPack max_rects;
+    std::vector<rbp::Rect> rects;
+    std::vector<rbp::RectSize> rect_sizes;
+    std::vector<rbp::RectSize> run_rect_sizes;
+  };
+
+  RbpState init_rbp_state(const std::vector<Size>& sizes) {
+    auto rbp = RbpState();
+    rbp.rects.reserve(sizes.size());
+    rbp.rect_sizes.reserve(sizes.size());
+    for (const auto& size : sizes)
+      rbp.rect_sizes.push_back({ size.width, size.height,
+        static_cast<int>(rbp.rect_sizes.size()) });
+
+    // to preserve order of identical rects (RBP_REVERSE_ORDER is also defined)
+    std::reverse(begin(rbp.rect_sizes), end(rbp.rect_sizes));
+    return rbp;
+  }
+
+  bool run_rbp_method(RbpState& rbp, const Settings& settings, Run& run,
+      const std::optional<Run>& best_run, const std::vector<Size>& sizes) {
+    copy_vector(rbp.rect_sizes, rbp.run_rect_sizes);
+    auto cancelled = false;
+    while (!rbp.run_rect_sizes.empty()) {
+      rbp.rects.clear();
+      rbp.max_rects.Init(run.width, run.height, settings.allow_rotate);
+      rbp.max_rects.Insert(rbp.run_rect_sizes, rbp.rects, to_rbp_method(run.method));
+      auto [width, height] = rbp.max_rects.BottomRight();
+
+      correct_size(settings, width, height);
+      run.total_area += width * height;
+
+      apply_padding(settings, width, height, false);
+      auto& sheet = run.sheets.emplace_back(Sheet{ width, height, { } });
+
+      // cancel when not making any progress
+      if (rbp.rects.empty())
+        return false;
+
+      // cancel when already worse than best run
+      const auto done = rbp.run_rect_sizes.empty();
+      if (best_run && !is_better_than(run, *best_run, !done)) {
+        cancelled = true;
+        break;
+      }
+
+      sheet.rects.reserve(rbp.rects.size());
+      for (auto& rbp_rect : rbp.rects) {
+        const auto& size = sizes[static_cast<size_t>(rbp_rect.id)];
+        sheet.rects.push_back({
+          size.id,
+          rbp_rect.x + settings.border_padding,
+          rbp_rect.y + settings.border_padding,
+          rbp_rect.width,
+          rbp_rect.height,
+          (rbp_rect.width != size.width)
+        });
+      }
+    }
+    return !cancelled;
+  }
+
+  struct StbState {
+    stbrp_context context{ };
+    std::vector<stbrp_node> nodes;
+    std::vector<stbrp_rect> rects;
+    std::vector<stbrp_rect> run_rects;
+  };
+
+  StbState init_stb_state(const Settings& settings, const std::vector<Size>& sizes) {
+    auto stb = StbState{ };
+    stb.rects.reserve(sizes.size());
+    stb.run_rects.reserve(sizes.size());
+    for (const auto& size : sizes)
+      stb.rects.push_back({ static_cast<int>(stb.rects.size()),
+        size.width, size.height, 0, 0, false });
+
+    if (settings.allow_rotate)
+      for (auto& rect : stb.rects)
+        if (rect.w > settings.max_width || rect.h > settings.max_height)
+          std::swap(rect.w, rect.h);
+
+    return stb;
+  }
+
+  bool run_stb_method(StbState& stb, const Settings& settings, Run& run,
+      const std::optional<Run>& best_run, const std::vector<Size>& sizes) {
+    copy_vector(stb.rects, stb.run_rects);
+    stb.nodes.resize(std::max(stb.nodes.size(), static_cast<size_t>(run.width)));
+
+    auto cancelled = false;
+    while (!stb.run_rects.empty()) {
+      stbrp_init_target(&stb.context, run.width, run.height,
+        stb.nodes.data(), static_cast<int>(stb.nodes.size()));
+      stbrp_setup_heuristic(&stb.context, to_stb_method(run.method));
+
+      [[maybe_unused]] const auto all_packed =
+        (stbrp_pack_rects(&stb.context, stb.run_rects.data(),
+          static_cast<int>(stb.run_rects.size())) == 1);
+
+      auto width = 0;
+      auto height = 0;
+      auto rects = std::vector<Rect>();
+      rects.reserve(stb.run_rects.size());
+      stb.run_rects.erase(std::remove_if(begin(stb.run_rects), end(stb.run_rects),
+        [&](const stbrp_rect& stb_rect) {
+          if (!stb_rect.was_packed)
+            return false;
+
+          width = std::max(width, stb_rect.x + stb_rect.w);
+          height = std::max(height, stb_rect.y + stb_rect.h);
+
+          const auto& size = sizes[static_cast<size_t>(stb_rect.id)];
+          rects.push_back({
+            size.id,
+            stb_rect.x + settings.border_padding,
+            stb_rect.y + settings.border_padding,
+            stb_rect.w, stb_rect.h,
+            (stb_rect.w != size.width)
+          });
+          return true;
+        }), end(stb.run_rects));
+
+      correct_size(settings, width, height);
+      run.total_area += width * height;
+
+      apply_padding(settings, width, height, false);
+      const auto& sheet = run.sheets.emplace_back(Sheet{ width, height, std::move(rects) });
+      const auto done = stb.run_rects.empty();
+      if (sheet.rects.empty() ||
+          (best_run && !is_better_than(run, *best_run, !done))) {
+        cancelled = true;
+        break;
+      }
+    }
+    return !cancelled;
+  }
+} // namespace
+
+std::vector<Sheet> pack(Settings settings, std::vector<Size> sizes) {
+  if (!correct_settings(settings, sizes))
+    return { };
+
+  if (sizes.empty())
+    return { };
+
+  auto stb_state = std::optional<StbState>();
+  if (settings.method == Method::Best ||
+      settings.method == Method::Best_Skyline ||
+      is_stb_method(settings.method))
+    stb_state.emplace(init_stb_state(settings, sizes));
+
+  auto rbp_state = std::optional<RbpState>();
+  if (settings.method == Method::Best ||
+      settings.method == Method::Best_MaxRects ||
+      is_rbp_method(settings.method))
+    rbp_state.emplace(init_rbp_state(sizes));
+
+  const auto perfect_area = get_perfect_area(sizes);
+  const auto target_area = perfect_area + perfect_area / 100;
+  const auto [initial_width, initial_height] = get_initial_run_size(settings, perfect_area);
+
+  auto total_best_run = std::optional<Run>{ };
+  const auto methods = get_concrete_methods(settings.method);
+  for (const auto& method : methods) {
+    auto best_run = std::optional<Run>{ };
+    auto state = OptimizationState{
+      perfect_area,
+      initial_width,
+      initial_height,
+      OptimizationStage::first_run,
+      0,
+    };
+    for (;;) {
+      if (best_run.has_value() &&
+          best_run->sheets.size() == 1 &&
+          best_run->total_area <= target_area)
+        break;
+
+      auto run = Run{ method, state.width, state.height, { }, 0 };
+      const auto succeeded = is_rbp_method(run.method) ?
+        run_rbp_method(*rbp_state, settings, run, best_run, sizes) :
+        run_stb_method(*stb_state, settings, run, best_run, sizes);
+
+      if (succeeded && (!best_run || is_better_than(run, *best_run)))
+        best_run = std::move(run);
+
+      if (!best_run.has_value() ||
+          !optimize_run_settings(state, settings, *best_run))
+        break;
+    }
+    if (best_run && (!total_best_run || is_better_than(*best_run, *total_best_run))) {
+      total_best_run = std::move(best_run);
+    }
+  }
+
+  if (!total_best_run)
+    return { };
+
+  if (settings.max_sheets &&
+      settings.max_sheets < static_cast<int>(total_best_run->sheets.size()))
+    total_best_run->sheets.resize(static_cast<size_t>(settings.max_sheets));
+
+  return std::move(total_best_run->sheets);
+}
+
+} // namespace

rectpacker.mod/rect_pack/rect_pack.h

@@ -0,0 +1,109 @@
+// rect_pack.h - public domain - rectangle packing
+// Albert Kalchmair 2021
+//
+// Useful for e.g. packing rectangular textures into one or multiple atlases.
+//
+// LICENSE
+//
+//   See end of file for license information.
+
+#pragma once
+
+#include <vector>
+
+namespace rect_pack {
+
+enum class Method {
+  Best,
+  Best_Skyline,
+  Best_MaxRects,
+  Skyline_BottomLeft,
+  Skyline_BestFit,
+  MaxRects_BestShortSideFit,
+  MaxRects_BestLongSideFit,
+  MaxRects_BestAreaFit,
+  MaxRects_BottomLeftRule,
+  MaxRects_ContactPointRule
+};
+
+struct Size {
+  int id;
+  int width;
+  int height;
+};
+
+struct Rect {
+  int id;
+  int x;
+  int y;
+  int width;
+  int height;
+  bool rotated;
+};
+
+struct Sheet {
+  int width;
+  int height;
+  std::vector<Rect> rects;
+};
+
+struct Settings {
+  Method method;
+  int max_sheets;
+  bool power_of_two;
+  bool square;
+  bool allow_rotate;
+  int align_width;
+  int border_padding;
+  int over_allocate;
+  int min_width;
+  int min_height;
+  int max_width;
+  int max_height;
+};
+
+std::vector<Sheet> pack(Settings settings, std::vector<Size> sizes);
+
+} // namespace
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses -- choose whichever you prefer.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2021 Albert Kalchmair
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+------------------------------------------------------------------------------
+ALTERNATIVE B - Public Domain (www.unlicense.org)
+This is free and unencumbered software released into the public domain.
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+------------------------------------------------------------------------------
+*/

rectpacker.mod/rect_pack/rect_pack_11.cpp

@@ -0,0 +1,658 @@
+// c++11 friendly version of rect_pack.cpp
+// Bruce A Henderson 2024
+
+#define STBRP_LARGE_RECTS
+#include "rect_pack.h"
+#include "MaxRectsBinPack.h"
+#include "stb_rect_pack.h"
+#include <algorithm>
+#include <cmath>
+#include <cassert>
+#include <memory> // for std::unique_ptr
+
+namespace rect_pack
+{
+    namespace
+    {
+        const auto first_Skyline_method = Method::Skyline_BottomLeft;
+        const auto last_Skyline_method = Method::Skyline_BestFit;
+        const auto first_MaxRects_method = Method::MaxRects_BestShortSideFit;
+        const auto last_MaxRects_method = Method::MaxRects_ContactPointRule;
+
+        int floor(int v, int q) { return (v / q) * q; };
+        int ceil(int v, int q) { return ((v + q - 1) / q) * q; };
+        int sqrt(int a) { return static_cast<int>(std::sqrt(a)); }
+        int div_ceil(int a, int b) { return (b > 0 ? (a + b - 1) / b : -1); }
+
+        template <typename T>
+        T clamp(T v, T lo, T hi)
+        {
+            return std::min(hi, std::max(lo, v));
+        }
+
+        int ceil_to_pot(int value)
+        {
+            for (int pot = 1;; pot <<= 1)
+                if (pot >= value)
+                    return pot;
+        }
+
+        int floor_to_pot(int value)
+        {
+            for (int pot = 1;; pot <<= 1)
+                if (pot > value)
+                    return (pot >> 1);
+        }
+
+        bool is_stb_method(Method method)
+        {
+            const auto first = static_cast<int>(first_Skyline_method);
+            const auto last = static_cast<int>(last_Skyline_method);
+            const auto index = static_cast<int>(method);
+            return (index >= first && index <= last);
+        }
+
+        bool is_rbp_method(Method method)
+        {
+            const auto first = static_cast<int>(first_MaxRects_method);
+            const auto last = static_cast<int>(last_MaxRects_method);
+            const auto index = static_cast<int>(method);
+            return (index >= first && index <= last);
+        }
+
+        int to_stb_method(Method method)
+        {
+            assert(is_stb_method(method));
+            return static_cast<int>(method) - static_cast<int>(first_Skyline_method);
+        }
+
+        rbp::MaxRectsBinPack::FreeRectChoiceHeuristic to_rbp_method(Method method)
+        {
+            assert(is_rbp_method(method));
+            return static_cast<rbp::MaxRectsBinPack::FreeRectChoiceHeuristic>(
+                static_cast<int>(method) - static_cast<int>(first_MaxRects_method));
+        }
+
+        std::vector<Method> get_concrete_methods(Method settings_method)
+        {
+            std::vector<Method> methods;
+            const auto add_skyline_methods = [&methods]()
+            {
+                methods.insert(end(methods), {Method::Skyline_BottomLeft,
+                                              Method::Skyline_BestFit});
+            };
+            const auto add_maxrect_methods = [&methods]()
+            {
+                methods.insert(end(methods), {
+                                                 Method::MaxRects_BestShortSideFit,
+                                                 Method::MaxRects_BestLongSideFit,
+                                                 Method::MaxRects_BestAreaFit,
+                                                 Method::MaxRects_BottomLeftRule,
+                                                 // do not automatically try costy contact point rule
+                                                 // Method::MaxRects_ContactPointRule,
+                                             });
+            };
+            switch (settings_method)
+            {
+            case Method::Best:
+                add_skyline_methods();
+                add_maxrect_methods();
+                break;
+
+            case Method::Best_Skyline:
+                add_skyline_methods();
+                break;
+
+            case Method::Best_MaxRects:
+                add_maxrect_methods();
+                break;
+
+            default:
+                methods.push_back(settings_method);
+                break;
+            }
+            return methods;
+        }
+
+        bool can_fit(const Settings &settings, int width, int height)
+        {
+            return ((width <= settings.max_width &&
+                     height <= settings.max_height) ||
+                    (settings.allow_rotate &&
+                     width <= settings.max_height &&
+                     height <= settings.max_width));
+        }
+
+        void apply_padding(const Settings &settings, int &width, int &height, bool indent)
+        {
+            const auto dir = (indent ? 1 : -1);
+            width -= dir * settings.border_padding * 2;
+            height -= dir * settings.border_padding * 2;
+            width += dir * settings.over_allocate;
+            height += dir * settings.over_allocate;
+        }
+
+        bool correct_settings(Settings &settings, std::vector<Size> &sizes)
+        {
+            // clamp max to far less than numeric_limits<int>::max() to prevent overflow
+            const auto size_limit = 1000000000;
+            if (settings.max_width <= 0 || settings.max_width > size_limit)
+                settings.max_width = size_limit;
+            if (settings.max_height <= 0 || settings.max_height > size_limit)
+                settings.max_height = size_limit;
+
+            if (settings.min_width < 0 ||
+                settings.min_height < 0 ||
+                settings.min_width > settings.max_width ||
+                settings.min_height > settings.max_height)
+                return false;
+
+            // immediately apply padding and over allocation, only relevant for power-of-two and alignment constraint
+            apply_padding(settings, settings.min_width, settings.min_height, true);
+            apply_padding(settings, settings.max_width, settings.max_height, true);
+
+            int max_rect_width = 0;
+            int max_rect_height = 0;
+            for (auto it = begin(sizes); it != end(sizes);)
+                if (it->width <= 0 ||
+                    it->height <= 0 ||
+                    !can_fit(settings, it->width, it->height))
+                {
+                    it = sizes.erase(it);
+                }
+                else
+                {
+                    if (settings.allow_rotate &&
+                        it->height > it->width &&
+                        it->height <= settings.max_width &&
+                        it->width <= settings.max_height)
+                    {
+                        max_rect_width = std::max(max_rect_width, it->height);
+                        max_rect_height = std::max(max_rect_height, it->width);
+                    }
+                    else
+                    {
+                        max_rect_width = std::max(max_rect_width, it->width);
+                        max_rect_height = std::max(max_rect_height, it->height);
+                    }
+                    ++it;
+                }
+
+            settings.min_width = std::max(settings.min_width, max_rect_width);
+            settings.min_height = std::max(settings.min_height, max_rect_height);
+
+            // clamp min to max and still pack the sprites which fit
+            settings.min_width = std::min(settings.min_width, settings.max_width);
+            settings.min_height = std::min(settings.min_height, settings.max_height);
+            return true;
+        }
+
+        struct Run
+        {
+            Method method;
+            int width;
+            int height;
+            std::vector<Sheet> sheets;
+            int total_area;
+        };
+
+        void correct_size(const Settings &settings, int &width, int &height)
+        {
+            width = std::max(width, settings.min_width);
+            height = std::max(height, settings.min_height);
+            apply_padding(settings, width, height, false);
+
+            if (settings.power_of_two)
+            {
+                width = ceil_to_pot(width);
+                height = ceil_to_pot(height);
+            }
+
+            if (settings.align_width)
+                width = ceil(width, settings.align_width);
+
+            if (settings.square)
+                width = height = std::max(width, height);
+
+            apply_padding(settings, width, height, true);
+            width = std::min(width, settings.max_width);
+            height = std::min(height, settings.max_height);
+            apply_padding(settings, width, height, false);
+
+            if (settings.power_of_two)
+            {
+                width = floor_to_pot(width);
+                height = floor_to_pot(height);
+            }
+
+            if (settings.align_width)
+                width = floor(width, settings.align_width);
+
+            if (settings.square)
+                width = height = std::min(width, height);
+
+            apply_padding(settings, width, height, true);
+        }
+
+        bool is_better_than(const Run &a, const Run &b, bool a_incomplete = false)
+        {
+            if (a_incomplete)
+            {
+                if (b.sheets.size() <= a.sheets.size())
+                    return false;
+            }
+            else
+            {
+                if (a.sheets.size() < b.sheets.size())
+                    return true;
+                if (b.sheets.size() < a.sheets.size())
+                    return false;
+            }
+            return (a.total_area < b.total_area);
+        }
+
+        int get_perfect_area(const std::vector<Size> &sizes)
+        {
+            int area = 0;
+            for (const auto &size : sizes)
+                area += size.width * size.height;
+            return area;
+        }
+
+        std::pair<int, int> get_run_size(const Settings &settings, int area)
+        {
+            int width = sqrt(area);
+            int height = div_ceil(area, width);
+            if (width < settings.min_width || width > settings.max_width)
+            {
+                width = clamp(width, settings.min_width, settings.max_width);
+                height = div_ceil(area, width);
+            }
+            else if (height < settings.min_height || height > settings.max_height)
+            {
+                height = clamp(height, settings.min_height, settings.max_height);
+                width = div_ceil(area, height);
+            }
+            correct_size(settings, width, height);
+            return std::make_pair(width, height);
+        }
+
+        std::pair<int, int> get_initial_run_size(const Settings &settings, int perfect_area)
+        {
+            return get_run_size(settings, perfect_area * 5 / 4);
+        }
+
+        enum class OptimizationStage
+        {
+            first_run,
+            minimize_sheet_count,
+            shrink_square,
+            shrink_width_fast,
+            shrink_height_fast,
+            shrink_width_slow,
+            shrink_height_slow,
+            end
+        };
+
+        struct OptimizationState
+        {
+            const int perfect_area;
+            int width;
+            int height;
+            OptimizationStage stage;
+            int iteration;
+        };
+
+        bool advance(OptimizationStage &stage)
+        {
+            if (stage == OptimizationStage::end)
+                return false;
+            stage = static_cast<OptimizationStage>(static_cast<int>(stage) + 1);
+            return true;
+        }
+
+        // returns true when stage should be kept, false to advance
+        bool optimize_stage(OptimizationState &state,
+                            const Settings &pack_settings, const Run &best_run)
+        {
+            switch (state.stage)
+            {
+            case OptimizationStage::first_run:
+            case OptimizationStage::end:
+                return false;
+
+            case OptimizationStage::minimize_sheet_count:
+            {
+                if (best_run.sheets.size() <= 1 ||
+                    state.iteration > 5)
+                    return false;
+
+                const auto &last_sheet = best_run.sheets.back();
+                int area = last_sheet.width * last_sheet.height;
+                for (int i = 0; area > 0; ++i)
+                {
+                    if (state.width == pack_settings.max_width &&
+                        state.height == pack_settings.max_height)
+                        break;
+                    if (state.height == pack_settings.max_height ||
+                        (state.width < pack_settings.max_width && i % 2))
+                    {
+                        ++state.width;
+                        area -= state.height;
+                    }
+                    else
+                    {
+                        ++state.height;
+                        area -= state.width;
+                    }
+                }
+                return true;
+            }
+
+            case OptimizationStage::shrink_square:
+            {
+                if (state.width != best_run.width ||
+                    state.height != best_run.height ||
+                    state.iteration > 5)
+                    return false;
+
+                std::pair<int, int> size = get_run_size(pack_settings, state.perfect_area);
+                state.width = (state.width + size.first) / 2;
+                state.height = (state.height + size.second) / 2;
+                return true;
+            }
+
+            case OptimizationStage::shrink_width_fast:
+            case OptimizationStage::shrink_height_fast:
+            case OptimizationStage::shrink_width_slow:
+            case OptimizationStage::shrink_height_slow:
+            {
+                if (state.iteration > 5)
+                    return false;
+
+                std::pair<int, int> size = get_run_size(pack_settings, state.perfect_area);
+                switch (state.stage)
+                {
+                default:
+                case OptimizationStage::shrink_width_fast:
+                    if (state.width > size.first + 4)
+                        state.width = (state.width + size.first) / 2;
+                    break;
+                case OptimizationStage::shrink_height_fast:
+                    if (state.height > size.second + 4)
+                        state.height = (state.height + size.second) / 2;
+                    break;
+                case OptimizationStage::shrink_width_slow:
+                    if (state.width > size.first)
+                        --state.width;
+                    break;
+                case OptimizationStage::shrink_height_slow:
+                    if (state.height > size.second)
+                        --state.height;
+                    break;
+                }
+                return true;
+            }
+            }
+            return false;
+        }
+
+        bool optimize_run_settings(OptimizationState &state,
+                                   const Settings &pack_settings, const Run &best_run)
+        {
+            const auto previous_state = state;
+            for (;;)
+            {
+                if (!optimize_stage(state, pack_settings, best_run))
+                    if (advance(state.stage))
+                    {
+                        state.width = best_run.width;
+                        state.height = best_run.height;
+                        state.iteration = 0;
+                        continue;
+                    }
+
+                if (state.stage == OptimizationStage::end)
+                    return false;
+
+                ++state.iteration;
+
+                int width = state.width;
+                int height = state.height;
+                correct_size(pack_settings, width, height);
+
+                if (width != previous_state.width || height != previous_state.height)
+                {
+                    state.width = width;
+                    state.height = height;
+                    return true;
+                }
+            }
+        }
+
+        template <typename T>
+        void copy_vector(const T &source, T &dest)
+        {
+            dest.resize(source.size());
+            std::copy(begin(source), end(source), begin(dest));
+        }
+
+        struct RbpState
+        {
+            rbp::MaxRectsBinPack max_rects;
+            std::vector<rbp::Rect> rects;
+            std::vector<rbp::RectSize> rect_sizes;
+            std::vector<rbp::RectSize> run_rect_sizes;
+        };
+
+        RbpState init_rbp_state(const std::vector<Size> &sizes)
+        {
+            RbpState rbp;
+            rbp.rects.reserve(sizes.size());
+            rbp.rect_sizes.reserve(sizes.size());
+            for (const auto &size : sizes)
+                rbp.rect_sizes.push_back({size.width, size.height,
+                                          static_cast<int>(rbp.rect_sizes.size())});
+
+            // to preserve order of identical rects (RBP_REVERSE_ORDER is also defined)
+            std::reverse(begin(rbp.rect_sizes), end(rbp.rect_sizes));
+            return rbp;
+        }
+
+        bool run_rbp_method(RbpState &rbp, const Settings &settings, Run &run,
+                            const std::unique_ptr<Run> &best_run, const std::vector<Size> &sizes)
+        {
+            copy_vector(rbp.rect_sizes, rbp.run_rect_sizes);
+            bool cancelled = false;
+            while (!rbp.run_rect_sizes.empty())
+            {
+                rbp.rects.clear();
+                rbp.max_rects.Init(run.width, run.height, settings.allow_rotate);
+                rbp.max_rects.Insert(rbp.run_rect_sizes, rbp.rects, to_rbp_method(run.method));
+                int width, height;
+                std::tie(width, height) = rbp.max_rects.BottomRight();
+
+                correct_size(settings, width, height);
+                run.total_area += width * height;
+
+                apply_padding(settings, width, height, false);
+                run.sheets.emplace_back(Sheet{width, height, {}});
+                auto &sheet = run.sheets.back(); // This is the updated line
+
+                // cancel when not making any progress
+                if (rbp.rects.empty())
+                    return false;
+
+                // cancel when already worse than best run
+                const bool done = rbp.run_rect_sizes.empty();
+                if (best_run && !is_better_than(run, *best_run, !done))
+                {
+                    cancelled = true;
+                    break;
+                }
+
+                sheet.rects.reserve(rbp.rects.size());
+                for (auto &rbp_rect : rbp.rects)
+                {
+                    const auto &size = sizes[static_cast<size_t>(rbp_rect.id)];
+                    sheet.rects.push_back({size.id,
+                                           rbp_rect.x + settings.border_padding,
+                                           rbp_rect.y + settings.border_padding,
+                                           rbp_rect.width,
+                                           rbp_rect.height,
+                                           (rbp_rect.width != size.width)});
+                }
+            }
+            return !cancelled;
+        }
+
+        struct StbState
+        {
+            stbrp_context context;
+            std::vector<stbrp_node> nodes;
+            std::vector<stbrp_rect> rects;
+            std::vector<stbrp_rect> run_rects;
+        };
+
+        StbState init_stb_state(const Settings &settings, const std::vector<Size> &sizes)
+        {
+            StbState stb;
+            stb.rects.reserve(sizes.size());
+            stb.run_rects.reserve(sizes.size());
+            for (const auto &size : sizes)
+                stb.rects.push_back({static_cast<int>(stb.rects.size()),
+                                     size.width, size.height, 0, 0, false});
+
+            if (settings.allow_rotate)
+                for (auto &rect : stb.rects)
+                    if (rect.w > settings.max_width || rect.h > settings.max_height)
+                        std::swap(rect.w, rect.h);
+
+            return stb;
+        }
+
+        bool run_stb_method(StbState &stb, const Settings &settings, Run &run,
+                            const std::unique_ptr<Run> &best_run, const std::vector<Size> &sizes)
+        {
+            copy_vector(stb.rects, stb.run_rects);
+            stb.nodes.resize(std::max(stb.nodes.size(), static_cast<size_t>(run.width)));
+
+            bool cancelled = false;
+            while (!stb.run_rects.empty())
+            {
+                stbrp_init_target(&stb.context, run.width, run.height,
+                                  stb.nodes.data(), static_cast<int>(stb.nodes.size()));
+                stbrp_setup_heuristic(&stb.context, to_stb_method(run.method));
+
+                const bool all_packed =
+                    (stbrp_pack_rects(&stb.context, stb.run_rects.data(),
+                                      static_cast<int>(stb.run_rects.size())) == 1);
+
+                int width = 0;
+                int height = 0;
+                std::vector<Rect> rects;
+                rects.reserve(stb.run_rects.size());
+                stb.run_rects.erase(std::remove_if(begin(stb.run_rects), end(stb.run_rects),
+                                                   [&](const stbrp_rect &stb_rect)
+                                                   {
+                                                       if (!stb_rect.was_packed)
+                                                           return false;
+
+                                                       width = std::max(width, stb_rect.x + stb_rect.w);
+                                                       height = std::max(height, stb_rect.y + stb_rect.h);
+
+                                                       const auto &size = sizes[static_cast<size_t>(stb_rect.id)];
+                                                       rects.push_back({size.id,
+                                                                        stb_rect.x + settings.border_padding,
+                                                                        stb_rect.y + settings.border_padding,
+                                                                        stb_rect.w, stb_rect.h,
+                                                                        (stb_rect.w != size.width)});
+                                                       return true;
+                                                   }),
+                                    end(stb.run_rects));
+
+                correct_size(settings, width, height);
+                run.total_area += width * height;
+
+                apply_padding(settings, width, height, false);
+                run.sheets.emplace_back(Sheet{width, height, std::move(rects)});
+                const bool done = stb.run_rects.empty();
+                if (run.sheets.back().rects.empty() ||
+                    (best_run && !is_better_than(run, *best_run, !done)))
+                {
+                    cancelled = true;
+                    break;
+                }
+            }
+            return !cancelled;
+        }
+    } // namespace
+
+    std::vector<Sheet> pack(Settings settings, std::vector<Size> sizes)
+    {
+        if (!correct_settings(settings, sizes))
+            return {};
+
+        if (sizes.empty())
+            return {};
+
+        std::unique_ptr<StbState> stb_state;
+        if (settings.method == Method::Best ||
+            settings.method == Method::Best_Skyline ||
+            is_stb_method(settings.method))
+            stb_state.reset(new StbState(init_stb_state(settings, sizes)));
+
+        std::unique_ptr<RbpState> rbp_state;
+        if (settings.method == Method::Best ||
+            settings.method == Method::Best_MaxRects ||
+            is_rbp_method(settings.method))
+            rbp_state.reset(new RbpState(init_rbp_state(sizes)));
+
+        const int perfect_area = get_perfect_area(sizes);
+        const int target_area = perfect_area + perfect_area / 100;
+        int initial_width, initial_height;
+        std::tie(initial_width, initial_height) = get_initial_run_size(settings, perfect_area);
+
+        std::unique_ptr<Run> total_best_run;
+        const std::vector<Method> methods = get_concrete_methods(settings.method);
+        for (const auto &method : methods)
+        {
+            std::unique_ptr<Run> best_run;
+            OptimizationState state{
+                perfect_area,
+                initial_width,
+                initial_height,
+                OptimizationStage::first_run,
+                0,
+            };
+            for (;;)
+            {
+                if (best_run && best_run->sheets.size() == 1 && best_run->total_area <= target_area)
+                    break;
+
+                Run run{method, state.width, state.height, {}, 0};
+                const bool succeeded = is_rbp_method(run.method) ? run_rbp_method(*rbp_state, settings, run, best_run, sizes) : run_stb_method(*stb_state, settings, run, best_run, sizes);
+
+                if (succeeded && (!best_run || is_better_than(run, *best_run)))
+                    best_run.reset(new Run(std::move(run)));
+
+                if (!best_run || !optimize_run_settings(state, settings, *best_run))
+                    break;
+            }
+            if (best_run && (!total_best_run || is_better_than(*best_run, *total_best_run)))
+            {
+                total_best_run.reset(new Run(std::move(*best_run)));
+            }
+        }
+
+        if (!total_best_run)
+            return {};
+
+        if (settings.max_sheets &&
+            settings.max_sheets < static_cast<int>(total_best_run->sheets.size()))
+            total_best_run->sheets.resize(static_cast<size_t>(settings.max_sheets));
+
+        return std::move(total_best_run->sheets);
+    }
+
+} // namespace rect_pack

rectpacker.mod/rect_pack/stb_rect_pack.cpp

@@ -0,0 +1,16 @@
+
+#include <array>
+#include <algorithm>
+
+template<size_t size, typename T, typename C>
+void my_stbrp_sort(T* ptr, std::size_t count, const C& comp) {
+  static_assert(sizeof(T) == size);
+  const auto begin = static_cast<T*>(ptr);
+  const auto end = begin + count;
+  std::sort(begin, end, [&](const T& a, const T& b) { return (comp(&a, &b) < 0); });
+}
+#define STBRP_SORT(PTR, COUNT, SIZE, COMP) my_stbrp_sort<(SIZE)>((PTR), (COUNT), (COMP))
+
+#define STB_RECT_PACK_IMPLEMENTATION
+#define STBRP_LARGE_RECTS
+#include "stb_rect_pack.h"

rectpacker.mod/rect_pack/stb_rect_pack.h

@@ -0,0 +1,623 @@
+// stb_rect_pack.h - v1.01 - public domain - rectangle packing
+// Sean Barrett 2014
+//
+// Useful for e.g. packing rectangular textures into an atlas.
+// Does not do rotation.
+//
+// Before #including,
+//
+//    #define STB_RECT_PACK_IMPLEMENTATION
+//
+// in the file that you want to have the implementation.
+//
+// Not necessarily the awesomest packing method, but better than
+// the totally naive one in stb_truetype (which is primarily what
+// this is meant to replace).
+//
+// Has only had a few tests run, may have issues.
+//
+// More docs to come.
+//
+// No memory allocations; uses qsort() and assert() from stdlib.
+// Can override those by defining STBRP_SORT and STBRP_ASSERT.
+//
+// This library currently uses the Skyline Bottom-Left algorithm.
+//
+// Please note: better rectangle packers are welcome! Please
+// implement them to the same API, but with a different init
+// function.
+//
+// Credits
+//
+//  Library
+//    Sean Barrett
+//  Minor features
+//    Martins Mozeiko
+//    github:IntellectualKitty
+//
+//  Bugfixes / warning fixes
+//    Jeremy Jaussaud
+//    Fabian Giesen
+//
+// Version history:
+//
+//     1.01  (2021-07-11)  always use large rect mode, expose STBRP__MAXVAL in public section
+//     1.00  (2019-02-25)  avoid small space waste; gracefully fail too-wide rectangles
+//     0.99  (2019-02-07)  warning fixes
+//     0.11  (2017-03-03)  return packing success/fail result
+//     0.10  (2016-10-25)  remove cast-away-const to avoid warnings
+//     0.09  (2016-08-27)  fix compiler warnings
+//     0.08  (2015-09-13)  really fix bug with empty rects (w=0 or h=0)
+//     0.07  (2015-09-13)  fix bug with empty rects (w=0 or h=0)
+//     0.06  (2015-04-15)  added STBRP_SORT to allow replacing qsort
+//     0.05:  added STBRP_ASSERT to allow replacing assert
+//     0.04:  fixed minor bug in STBRP_LARGE_RECTS support
+//     0.01:  initial release
+//
+// LICENSE
+//
+//   See end of file for license information.
+
+//////////////////////////////////////////////////////////////////////////////
+//
+//       INCLUDE SECTION
+//
+
+#ifndef STB_INCLUDE_STB_RECT_PACK_H
+#define STB_INCLUDE_STB_RECT_PACK_H
+
+#define STB_RECT_PACK_VERSION  1
+
+#ifdef STBRP_STATIC
+#define STBRP_DEF static
+#else
+#define STBRP_DEF extern
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct stbrp_context stbrp_context;
+typedef struct stbrp_node    stbrp_node;
+typedef struct stbrp_rect    stbrp_rect;
+
+typedef int            stbrp_coord;
+
+#define STBRP__MAXVAL  0x7fffffff
+// Mostly for internal use, but this is the maximum supported coordinate value.
+
+STBRP_DEF int stbrp_pack_rects (stbrp_context *context, stbrp_rect *rects, int num_rects);
+// Assign packed locations to rectangles. The rectangles are of type
+// 'stbrp_rect' defined below, stored in the array 'rects', and there
+// are 'num_rects' many of them.
+//
+// Rectangles which are successfully packed have the 'was_packed' flag
+// set to a non-zero value and 'x' and 'y' store the minimum location
+// on each axis (i.e. bottom-left in cartesian coordinates, top-left
+// if you imagine y increasing downwards). Rectangles which do not fit
+// have the 'was_packed' flag set to 0.
+//
+// You should not try to access the 'rects' array from another thread
+// while this function is running, as the function temporarily reorders
+// the array while it executes.
+//
+// To pack into another rectangle, you need to call stbrp_init_target
+// again. To continue packing into the same rectangle, you can call
+// this function again. Calling this multiple times with multiple rect
+// arrays will probably produce worse packing results than calling it
+// a single time with the full rectangle array, but the option is
+// available.
+//
+// The function returns 1 if all of the rectangles were successfully
+// packed and 0 otherwise.
+
+struct stbrp_rect
+{
+   // reserved for your use:
+   int            id;
+
+   // input:
+   stbrp_coord    w, h;
+
+   // output:
+   stbrp_coord    x, y;
+   int            was_packed;  // non-zero if valid packing
+
+}; // 16 bytes, nominally
+
+
+STBRP_DEF void stbrp_init_target (stbrp_context *context, int width, int height, stbrp_node *nodes, int num_nodes);
+// Initialize a rectangle packer to:
+//    pack a rectangle that is 'width' by 'height' in dimensions
+//    using temporary storage provided by the array 'nodes', which is 'num_nodes' long
+//
+// You must call this function every time you start packing into a new target.
+//
+// There is no "shutdown" function. The 'nodes' memory must stay valid for
+// the following stbrp_pack_rects() call (or calls), but can be freed after
+// the call (or calls) finish.
+//
+// Note: to guarantee best results, either:
+//       1. make sure 'num_nodes' >= 'width'
+//   or  2. call stbrp_allow_out_of_mem() defined below with 'allow_out_of_mem = 1'
+//
+// If you don't do either of the above things, widths will be quantized to multiples
+// of small integers to guarantee the algorithm doesn't run out of temporary storage.
+//
+// If you do #2, then the non-quantized algorithm will be used, but the algorithm
+// may run out of temporary storage and be unable to pack some rectangles.
+
+STBRP_DEF void stbrp_setup_allow_out_of_mem (stbrp_context *context, int allow_out_of_mem);
+// Optionally call this function after init but before doing any packing to
+// change the handling of the out-of-temp-memory scenario, described above.
+// If you call init again, this will be reset to the default (false).
+
+
+STBRP_DEF void stbrp_setup_heuristic (stbrp_context *context, int heuristic);
+// Optionally select which packing heuristic the library should use. Different
+// heuristics will produce better/worse results for different data sets.
+// If you call init again, this will be reset to the default.
+
+enum
+{
+   STBRP_HEURISTIC_Skyline_default=0,
+   STBRP_HEURISTIC_Skyline_BL_sortHeight = STBRP_HEURISTIC_Skyline_default,
+   STBRP_HEURISTIC_Skyline_BF_sortHeight
+};
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// the details of the following structures don't matter to you, but they must
+// be visible so you can handle the memory allocations for them
+
+struct stbrp_node
+{
+   stbrp_coord  x,y;
+   stbrp_node  *next;
+};
+
+struct stbrp_context
+{
+   int width;
+   int height;
+   int align;
+   int init_mode;
+   int heuristic;
+   int num_nodes;
+   stbrp_node *active_head;
+   stbrp_node *free_head;
+   stbrp_node extra[2]; // we allocate two extra nodes so optimal user-node-count is 'width' not 'width+2'
+};
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+//
+//     IMPLEMENTATION SECTION
+//
+
+#ifdef STB_RECT_PACK_IMPLEMENTATION
+#ifndef STBRP_SORT
+#include <stdlib.h>
+#define STBRP_SORT qsort
+#endif
+
+#ifndef STBRP_ASSERT
+#include <assert.h>
+#define STBRP_ASSERT assert
+#endif
+
+#ifdef _MSC_VER
+#define STBRP__NOTUSED(v)  (void)(v)
+#define STBRP__CDECL       __cdecl
+#else
+#define STBRP__NOTUSED(v)  (void)sizeof(v)
+#define STBRP__CDECL
+#endif
+
+enum
+{
+   STBRP__INIT_skyline = 1
+};
+
+STBRP_DEF void stbrp_setup_heuristic(stbrp_context *context, int heuristic)
+{
+   switch (context->init_mode) {
+      case STBRP__INIT_skyline:
+         STBRP_ASSERT(heuristic == STBRP_HEURISTIC_Skyline_BL_sortHeight || heuristic == STBRP_HEURISTIC_Skyline_BF_sortHeight);
+         context->heuristic = heuristic;
+         break;
+      default:
+         STBRP_ASSERT(0);
+   }
+}
+
+STBRP_DEF void stbrp_setup_allow_out_of_mem(stbrp_context *context, int allow_out_of_mem)
+{
+   if (allow_out_of_mem)
+      // if it's ok to run out of memory, then don't bother aligning them;
+      // this gives better packing, but may fail due to OOM (even though
+      // the rectangles easily fit). @TODO a smarter approach would be to only
+      // quantize once we've hit OOM, then we could get rid of this parameter.
+      context->align = 1;
+   else {
+      // if it's not ok to run out of memory, then quantize the widths
+      // so that num_nodes is always enough nodes.
+      //
+      // I.e. num_nodes * align >= width
+      //                  align >= width / num_nodes
+      //                  align = ceil(width/num_nodes)
+
+      context->align = (context->width + context->num_nodes-1) / context->num_nodes;
+   }
+}
+
+STBRP_DEF void stbrp_init_target(stbrp_context *context, int width, int height, stbrp_node *nodes, int num_nodes)
+{
+   int i;
+
+   for (i=0; i < num_nodes-1; ++i)
+      nodes[i].next = &nodes[i+1];
+   nodes[i].next = NULL;
+   context->init_mode = STBRP__INIT_skyline;
+   context->heuristic = STBRP_HEURISTIC_Skyline_default;
+   context->free_head = &nodes[0];
+   context->active_head = &context->extra[0];
+   context->width = width;
+   context->height = height;
+   context->num_nodes = num_nodes;
+   stbrp_setup_allow_out_of_mem(context, 0);
+
+   // node 0 is the full width, node 1 is the sentinel (lets us not store width explicitly)
+   context->extra[0].x = 0;
+   context->extra[0].y = 0;
+   context->extra[0].next = &context->extra[1];
+   context->extra[1].x = (stbrp_coord) width;
+   context->extra[1].y = (1<<30);
+   context->extra[1].next = NULL;
+}
+
+// find minimum y position if it starts at x1
+static int stbrp__skyline_find_min_y(stbrp_context *c, stbrp_node *first, int x0, int width, int *pwaste)
+{
+   stbrp_node *node = first;
+   int x1 = x0 + width;
+   int min_y, visited_width, waste_area;
+
+   STBRP__NOTUSED(c);
+
+   STBRP_ASSERT(first->x <= x0);
+
+   #if 0
+   // skip in case we're past the node
+   while (node->next->x <= x0)
+      ++node;
+   #else
+   STBRP_ASSERT(node->next->x > x0); // we ended up handling this in the caller for efficiency
+   #endif
+
+   STBRP_ASSERT(node->x <= x0);
+
+   min_y = 0;
+   waste_area = 0;
+   visited_width = 0;
+   while (node->x < x1) {
+      if (node->y > min_y) {
+         // raise min_y higher.
+         // we've accounted for all waste up to min_y,
+         // but we'll now add more waste for everything we've visted
+         waste_area += visited_width * (node->y - min_y);
+         min_y = node->y;
+         // the first time through, visited_width might be reduced
+         if (node->x < x0)
+            visited_width += node->next->x - x0;
+         else
+            visited_width += node->next->x - node->x;
+      } else {
+         // add waste area
+         int under_width = node->next->x - node->x;
+         if (under_width + visited_width > width)
+            under_width = width - visited_width;
+         waste_area += under_width * (min_y - node->y);
+         visited_width += under_width;
+      }
+      node = node->next;
+   }
+
+   *pwaste = waste_area;
+   return min_y;
+}
+
+typedef struct
+{
+   int x,y;
+   stbrp_node **prev_link;
+} stbrp__findresult;
+
+static stbrp__findresult stbrp__skyline_find_best_pos(stbrp_context *c, int width, int height)
+{
+   int best_waste = (1<<30), best_x, best_y = (1 << 30);
+   stbrp__findresult fr;
+   stbrp_node **prev, *node, *tail, **best = NULL;
+
+   // align to multiple of c->align
+   width = (width + c->align - 1);
+   width -= width % c->align;
+   STBRP_ASSERT(width % c->align == 0);
+
+   // if it can't possibly fit, bail immediately
+   if (width > c->width || height > c->height) {
+      fr.prev_link = NULL;
+      fr.x = fr.y = 0;
+      return fr;
+   }
+
+   node = c->active_head;
+   prev = &c->active_head;
+   while (node->x + width <= c->width) {
+      int y,waste;
+      y = stbrp__skyline_find_min_y(c, node, node->x, width, &waste);
+      if (c->heuristic == STBRP_HEURISTIC_Skyline_BL_sortHeight) { // actually just want to test BL
+         // bottom left
+         if (y < best_y) {
+            best_y = y;
+            best = prev;
+         }
+      } else {
+         // best-fit
+         if (y + height <= c->height) {
+            // can only use it if it first vertically
+            if (y < best_y || (y == best_y && waste < best_waste)) {
+               best_y = y;
+               best_waste = waste;
+               best = prev;
+            }
+         }
+      }
+      prev = &node->next;
+      node = node->next;
+   }
+
+   best_x = (best == NULL) ? 0 : (*best)->x;
+
+   // if doing best-fit (BF), we also have to try aligning right edge to each node position
+   //
+   // e.g, if fitting
+   //
+   //     ____________________
+   //    |____________________|
+   //
+   //            into
+   //
+   //   |                         |
+   //   |             ____________|
+   //   |____________|
+   //
+   // then right-aligned reduces waste, but bottom-left BL is always chooses left-aligned
+   //
+   // This makes BF take about 2x the time
+
+   if (c->heuristic == STBRP_HEURISTIC_Skyline_BF_sortHeight) {
+      tail = c->active_head;
+      node = c->active_head;
+      prev = &c->active_head;
+      // find first node that's admissible
+      while (tail->x < width)
+         tail = tail->next;
+      while (tail) {
+         int xpos = tail->x - width;
+         int y,waste;
+         STBRP_ASSERT(xpos >= 0);
+         // find the left position that matches this
+         while (node->next->x <= xpos) {
+            prev = &node->next;
+            node = node->next;
+         }
+         STBRP_ASSERT(node->next->x > xpos && node->x <= xpos);
+         y = stbrp__skyline_find_min_y(c, node, xpos, width, &waste);
+         if (y + height <= c->height) {
+            if (y <= best_y) {
+               if (y < best_y || waste < best_waste || (waste==best_waste && xpos < best_x)) {
+                  best_x = xpos;
+                  STBRP_ASSERT(y <= best_y);
+                  best_y = y;
+                  best_waste = waste;
+                  best = prev;
+               }
+            }
+         }
+         tail = tail->next;
+      }
+   }
+
+   fr.prev_link = best;
+   fr.x = best_x;
+   fr.y = best_y;
+   return fr;
+}
+
+static stbrp__findresult stbrp__skyline_pack_rectangle(stbrp_context *context, int width, int height)
+{
+   // find best position according to heuristic
+   stbrp__findresult res = stbrp__skyline_find_best_pos(context, width, height);
+   stbrp_node *node, *cur;
+
+   // bail if:
+   //    1. it failed
+   //    2. the best node doesn't fit (we don't always check this)
+   //    3. we're out of memory
+   if (res.prev_link == NULL || res.y + height > context->height || context->free_head == NULL) {
+      res.prev_link = NULL;
+      return res;
+   }
+
+   // on success, create new node
+   node = context->free_head;
+   node->x = (stbrp_coord) res.x;
+   node->y = (stbrp_coord) (res.y + height);
+
+   context->free_head = node->next;
+
+   // insert the new node into the right starting point, and
+   // let 'cur' point to the remaining nodes needing to be
+   // stiched back in
+
+   cur = *res.prev_link;
+   if (cur->x < res.x) {
+      // preserve the existing one, so start testing with the next one
+      stbrp_node *next = cur->next;
+      cur->next = node;
+      cur = next;
+   } else {
+      *res.prev_link = node;
+   }
+
+   // from here, traverse cur and free the nodes, until we get to one
+   // that shouldn't be freed
+   while (cur->next && cur->next->x <= res.x + width) {
+      stbrp_node *next = cur->next;
+      // move the current node to the free list
+      cur->next = context->free_head;
+      context->free_head = cur;
+      cur = next;
+   }
+
+   // stitch the list back in
+   node->next = cur;
+
+   if (cur->x < res.x + width)
+      cur->x = (stbrp_coord) (res.x + width);
+
+#ifdef _DEBUG
+   cur = context->active_head;
+   while (cur->x < context->width) {
+      STBRP_ASSERT(cur->x < cur->next->x);
+      cur = cur->next;
+   }
+   STBRP_ASSERT(cur->next == NULL);
+
+   {
+      int count=0;
+      cur = context->active_head;
+      while (cur) {
+         cur = cur->next;
+         ++count;
+      }
+      cur = context->free_head;
+      while (cur) {
+         cur = cur->next;
+         ++count;
+      }
+      STBRP_ASSERT(count == context->num_nodes+2);
+   }
+#endif
+
+   return res;
+}
+
+static int STBRP__CDECL rect_height_compare(const void *a, const void *b)
+{
+   const stbrp_rect *p = (const stbrp_rect *) a;
+   const stbrp_rect *q = (const stbrp_rect *) b;
+   if (p->h > q->h)
+      return -1;
+   if (p->h < q->h)
+      return  1;
+   return (p->w > q->w) ? -1 : (p->w < q->w);
+}
+
+static int STBRP__CDECL rect_original_order(const void *a, const void *b)
+{
+   const stbrp_rect *p = (const stbrp_rect *) a;
+   const stbrp_rect *q = (const stbrp_rect *) b;
+   return (p->was_packed < q->was_packed) ? -1 : (p->was_packed > q->was_packed);
+}
+
+STBRP_DEF int stbrp_pack_rects(stbrp_context *context, stbrp_rect *rects, int num_rects)
+{
+   int i, all_rects_packed = 1;
+
+   // we use the 'was_packed' field internally to allow sorting/unsorting
+   for (i=0; i < num_rects; ++i) {
+      rects[i].was_packed = i;
+   }
+
+   // sort according to heuristic
+   STBRP_SORT(rects, num_rects, sizeof(rects[0]), rect_height_compare);
+
+   for (i=0; i < num_rects; ++i) {
+      if (rects[i].w == 0 || rects[i].h == 0) {
+         rects[i].x = rects[i].y = 0;  // empty rect needs no space
+      } else {
+         stbrp__findresult fr = stbrp__skyline_pack_rectangle(context, rects[i].w, rects[i].h);
+         if (fr.prev_link) {
+            rects[i].x = (stbrp_coord) fr.x;
+            rects[i].y = (stbrp_coord) fr.y;
+         } else {
+            rects[i].x = rects[i].y = STBRP__MAXVAL;
+         }
+      }
+   }
+
+   // unsort
+   STBRP_SORT(rects, num_rects, sizeof(rects[0]), rect_original_order);
+
+   // set was_packed flags and all_rects_packed status
+   for (i=0; i < num_rects; ++i) {
+      rects[i].was_packed = !(rects[i].x == STBRP__MAXVAL && rects[i].y == STBRP__MAXVAL);
+      if (!rects[i].was_packed)
+         all_rects_packed = 0;
+   }
+
+   // return the all_rects_packed status
+   return all_rects_packed;
+}
+#endif
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses -- choose whichever you prefer.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2017 Sean Barrett
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+------------------------------------------------------------------------------
+ALTERNATIVE B - Public Domain (www.unlicense.org)
+This is free and unencumbered software released into the public domain.
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+------------------------------------------------------------------------------
+*/

rectpacker.mod/rectpacker.bmx

@@ -0,0 +1,295 @@
+' Copyright (c) 2024-2025 Bruce A Henderson
+' 
+' This software is provided 'as-is', without any express or implied
+' warranty. In no event will the authors be held liable for any damages
+' arising from the use of this software.
+' 
+' Permission is granted to anyone to use this software for any purpose,
+' including commercial applications, and to alter it and redistribute it
+' freely, subject to the following restrictions:
+' 
+' 1. The origin of this software must not be misrepresented; you must not
+'    claim that you wrote the original software. If you use this software
+'    in a product, an acknowledgment in the product documentation would be
+'    appreciated but is not required.
+' 2. Altered source versions must be plainly marked as such, and must not be
+'    misrepresented as being the original software.
+' 3. This notice may not be removed or altered from any source distribution.
+' 
+SuperStrict
+
+Rem
+bbdoc: A module for packing rectangles into sheets.
+about: Useful for creating texture atlases, sprite sheets, and other similar things.
+End Rem
+Module BRL.RectPacker
+
+ModuleInfo "Version: 1.01"
+ModuleInfo "License: zlib/libpng"
+ModuleInfo "Copyright: 2024-2025 Bruce A Henderson"
+ModuleInfo "rect_pack: Albert Kalchmair 2021, Sean Barrett 2014, Jukka Jylänki"
+
+ModuleInfo "History: 1.01"
+ModuleInfo "History: borderPadding now applies to individual rects."
+ModuleInfo "History: Added sheetPadding to add padding around the edge of the sheet."
+ModuleInfo "History: 1.00 Initial Release"
+
+ModuleInfo "CPP_OPTS: -std=c++11"
+
+Import BRL.Collections
+
+Import "source.bmx"
+
+Rem
+bbdoc: Packs rectangles into sheets.
+about: The packer provides a number of settings that can be used to control how the rectangles are packed.
+The rectangles are added to the packer using the #Add method, and then the #Pack method is called to pack them into sheets.
+The packer will return an array of #TPackedSheet objects, each of which contains the rectangles that have been packed into it.
+An @id can be assigned to each rectangle, which can be used to identify the rectangle in the packed sheets.
+End Rem
+Type TRectPacker
+
+	Rem
+	bbdoc: The packing method to use.
+	End Rem
+	Field packingMethod:EPackingMethod = EPackingMethod.Best
+
+	Rem
+	bbdoc: The maximum number of sheets to produce.
+	about: If the packer is unable to fit all the rectangles into the specified number of sheets, those that don't fit will be discarded.
+	End Rem
+	Field maxSheets:Int = 1
+
+	Rem
+	bbdoc: Whether to pack into power-of-two sized sheets.
+	about: If this is set to #True, the width and height of the sheets will be rounded up to the nearest power of two.
+	This is useful for creating sheets that are intended to be used for creating textures.
+	End Rem
+	Field powerOfTwo:Int = True
+
+	Rem
+	bbdoc: Whether to pack into square sheets.
+	about: If this is set to #True, the width and height of the sheets will be the same.
+	End Rem
+	Field square:Int = False
+
+	Rem
+	bbdoc: Whether to allow rectangles to be rotated.
+	about: If this is set to #True, the packer may attempt to rotate rectangles to help fit them into the sheets.
+	End Rem
+	Field allowRotate:Int = False
+
+	Rem
+	bbdoc: Whether to align the width of the rectangles.
+	about: If this is set to #True, the packer will attempt to align the width of the rectangles to the width of the sheet.
+	This can help to reduce the amount of wasted space in the sheet.
+	End Rem
+	Field alignWidth:Int = False
+
+	Rem
+	bbdoc: The amount of padding to add around individual rects.
+	End Rem
+	Field borderPadding:Int
+
+	Rem
+	bbdoc: The amount of padding to add around the edge of the sheet.
+	End Rem
+	Field sheetPadding:Int
+
+	Rem
+	bbdoc: The amount to over-allocate the sheet by.
+	about: This is useful if you want to add a border around the sheet, or if you want to add some padding around the rectangles.
+	End Rem
+	Field overAllocate:Int
+
+	Rem
+	bbdoc: The minimum width of the sheets.
+	End Rem
+	Field minWidth:Int
+
+	Rem
+	bbdoc: The minimum height of the sheets.
+	End Rem
+	Field minHeight:Int
+
+	Rem
+	bbdoc: The maximum width of the sheets.
+	End Rem
+	Field maxWidth:Int
+
+	Rem
+	bbdoc: The maximum height of the sheets.
+	End Rem
+	Field maxHeight:Int
+
+	Field sizes:TArrayList<SRectSize> = New TArrayList<SRectSize>
+
+	Rem
+	bbdoc: Adds a rectangle with the given @id to the packer.
+	End Rem
+	Method Add(width:Int, height:Int, id:Int)
+
+		Local size:SRectSize = New SRectSize(width, height, id)
+		sizes.Add(size)
+
+	End Method
+
+	Rem
+	bbdoc: Packs the rectangles into sheets, based on the settings of the packer.
+	about: This method will return an array of #TPackedSheet objects, each of which contains the rectangles that have been packed into it.
+	Any rectangles that don't fit into the sheets will be discarded, and not be included in the returned array.
+	End Rem
+	Method Pack:TPackedSheet[]()
+		Return bmx_rectpacker_pack(Self, packingMethod, maxSheets, powerOfTwo, square, allowRotate, alignWidth, borderPadding, sheetPadding, overAllocate, minWidth, minHeight, maxWidth, maxHeight, sizes.Count())
+	End Method
+
+Private
+	Function _GetSize(packer:TRectPacker, index:Int, width:Int Var, height:Int Var, id:Int Var) { nomangle }
+		Local size:SRectSize = packer.sizes[index]
+		width = size.width
+		height = size.height
+		id = size.id
+	End Function
+
+	Function _NewSheetArray:TPackedSheet[](size:Int) { nomangle }
+		Return New TPackedSheet[size]
+	End Function
+
+	Function _SetSheet(sheets:TPackedSheet[], index:Int, sheet:TPackedSheet) { nomangle }
+		sheets[index] = sheet
+	End Function
+
+End Type
+
+Struct SRectSize
+	Field width:Int
+	Field height:Int
+	Field id:Int
+
+	Method New(width:Int, height:Int, id:Int)
+		Self.width = width
+		Self.height = height
+		Self.id = id
+	End Method
+
+	Method Operator=:Int(other:SRectSize)
+		Return width = other.width And height = other.height And id = other.id
+	End Method
+End Struct
+
+Rem
+bbdoc: The packing method to use.
+about: The packing method determines how the rectangles are packed into the sheets.
+
+| Value                         | Description                                  |
+|-------------------------------|----------------------------------------------|
+| #Best                         | The best fitting from all of the available methods. |
+| #BestSkyline                  | The best available skyline method.           |
+| #BestMaxRects                 | The best available max rects method.         |
+| #SkylineBottomLeft            | The skyline bottom-left method.              |
+| #SkylineBestFit               | The skyline best-fit method.                 |
+| #MaxRectsBestShortSideFit     | The max rects best short-side fit method.    |
+| #MaxRectsBestLongSideFit      | The max rects best long-side fit method.     |
+| #MaxRectsBestAreaFit          | The max rects best area fit method.          |
+| #MaxRectsBottomLeftRule       | The max rects bottom-left rule method.       |
+| #MaxRectsContactPointRule     | The max rects contact-point rule method.     |
+End Rem
+Enum EPackingMethod
+	Best
+	BestSkyline
+	BestMaxRects
+	SkylineBottomLeft
+	SkylineBestFit
+	MaxRectsBestShortSideFit
+	MaxRectsBestLongSideFit
+	MaxRectsBestAreaFit
+	MaxRectsBottomLeftRule
+	MaxRectsContactPointRule
+End Enum
+
+Rem
+bbdoc: Represents a rectangle that has been packed into a sheet.
+End Rem
+Struct SPackedRect
+
+	Rem
+	bbdoc: The ID of the rectangle.
+	End Rem
+	Field id:Int
+
+	Rem
+	bbdoc: The X position of the rectangle.
+	End Rem
+	Field x:Int
+
+	Rem
+	bbdoc: The Y position of the rectangle.
+	End Rem
+	Field y:Int
+
+	Rem
+	bbdoc: The width of the rectangle.
+	End Rem
+	Field width:Int
+
+	Rem
+	bbdoc: The height of the rectangle.
+	End Rem
+	Field height:Int
+
+	Rem
+	bbdoc: Whether the rectangle has been rotated.
+	End Rem
+	Field rotated:Int
+
+	Method New(id:Int, x:Int, y:Int, width:Int, height:Int, rotated:Int)
+		Self.id = id
+		Self.x = x
+		Self.y = y
+		Self.width = width
+		Self.height = height
+		Self.rotated = rotated
+	End Method
+End Struct
+
+Rem
+bbdoc: Represents a sheet that has been packed with rectangles.
+End Rem
+Type TPackedSheet
+
+	Rem
+	bbdoc: The width of the sheet.
+	End Rem
+	Field width:Int
+
+	Rem
+	bbdoc: The height of the sheet.
+	End Rem
+	Field height:Int
+
+	Rem
+	bbdoc: The rectangles that have been packed into the sheet.
+	End Rem
+	Field rects:SPackedRect[]
+
+Private
+	Function _Create:TPackedSheet(width:Int, height:Int, size:Int) { nomangle }
+		Local sheet:TPackedSheet = New TPackedSheet
+		sheet.width = width
+		sheet.height = height
+		sheet.rects = New SPackedRect[size]
+		Return sheet
+	End Function
+
+	Function _SetRect(sheet:TPackedSheet, index:Int, id:Int, x:Int, y:Int, width:Int, height:Int, rotated:Int) { nomangle }
+		Local rect:SPackedRect = New SPackedRect(id, x, y, width, height, rotated)
+		sheet.rects[index] = rect
+	End Function
+
+End Type
+
+Extern
+
+	Function bmx_rectpacker_pack:TPackedSheet[](packer:TRectPacker, packingMethod:EPackingMethod, maxSheets:Int, powerOfTwo:Int, square:Int, allowRotate:Int, alignWidth:Int, borderPadding:Int, sheetPadding:Int, overAllocate:Int, minWidth:Int, minHeight:Int, maxWidth:Int, maxHeight:Int, count:Int)
+
+End Extern

rectpacker.mod/source.bmx

@@ -0,0 +1,27 @@
+' Copyright (c) 2024-2025 Bruce A Henderson
+' 
+' This software is provided 'as-is', without any express or implied
+' warranty. In no event will the authors be held liable for any damages
+' arising from the use of this software.
+' 
+' Permission is granted to anyone to use this software for any purpose,
+' including commercial applications, and to alter it and redistribute it
+' freely, subject to the following restrictions:
+' 
+' 1. The origin of this software must not be misrepresented; you must not
+'    claim that you wrote the original software. If you use this software
+'    in a product, an acknowledgment in the product documentation would be
+'    appreciated but is not required.
+' 2. Altered source versions must be plainly marked as such, and must not be
+'    misrepresented as being the original software.
+' 3. This notice may not be removed or altered from any source distribution.
+' 
+SuperStrict
+
+Import "rect_pack/*.h"
+
+Import "glue.cpp"
+
+Import "rect_pack/rect_pack_11.cpp"
+Import "rect_pack/MaxRectsBinPack.cpp"
+Import "rect_pack/stb_rect_pack.cpp"

stringbuilder.mod/common.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2018-2023 Bruce A Henderson
+' Copyright (c) 2018-2024 Bruce A Henderson
 ' 
 ' This software is provided 'as-is', without any express or implied
 ' warranty. In no event will the authors be held liable for any damages
@@ -89,6 +89,16 @@ Extern
 	Function bmx_stringbuilder_splitbuffer_text:String(splitPtr:Byte Ptr, index:Int)
 	Function bmx_stringbuilder_splitbuffer_free(splitPtr:Byte Ptr)
 	Function bmx_stringbuilder_splitbuffer_toarray:String[](splitPtr:Byte Ptr)
+	Function bmx_stringbuilder_splitbuffer_split:Byte Ptr(splitPtr:Byte Ptr, separator:String, index:Int)
+	Function bmx_stringbuilder_splitbuffer_toint:Int(splitPtr:Byte Ptr, index:Int)
+	Function bmx_stringbuilder_splitbuffer_tofloat:Float(splitPtr:Byte Ptr, index:Int)
+	Function bmx_stringbuilder_splitbuffer_todouble:Double(splitPtr:Byte Ptr, index:Int)
+	Function bmx_stringbuilder_splitbuffer_tolong:Long(splitPtr:Byte Ptr, index:Int)
+	Function bmx_stringbuilder_splitbuffer_toulong:ULong(splitPtr:Byte Ptr, index:Int)
+	Function bmx_stringbuilder_splitbuffer_toshort:Short(splitPtr:Byte Ptr, index:Int)
+	Function bmx_stringbuilder_splitbuffer_touint:UInt(splitPtr:Byte Ptr, index:Int)
+	Function bmx_stringbuilder_splitbuffer_tosizet:Size_T(splitPtr:Byte Ptr, index:Int)
+	Function bmx_stringbuilder_splitbuffer_tobyte:Byte(splitPtr:Byte Ptr, index:Int)
 
 End Extern
 

stringbuilder.mod/glue.c

@@ -1,5 +1,5 @@
 /*
-  Copyright (c) 2018-2023 Bruce A Henderson
+  Copyright (c) 2018-2024 Bruce A Henderson
   
   This software is provided 'as-is', without any express or implied
   warranty. In no event will the authors be held liable for any damages
@@ -19,6 +19,9 @@
 */ 
 
 #include "glue.h"
+#include <errno.h>
+#include <limits.h>
+#include <stdlib.h>
 
 static int utf32strlen( const BBUINT *p ){
 	const BBUINT *t=p;
@@ -810,6 +813,58 @@ void bmx_stringbuilder_toutf8_buffer(BBString *str, char * buf, size_t length) {
 	*q=0;
 }
 
+void bmx_stringbuilder_toutf8_sbuffer(BBChar * p, int len, char * buf, size_t length) {
+	int i=0;
+	int out=0;
+	char *q=buf;
+
+	while (i < len && out < length) {
+		unsigned int c=*p++;
+		if(0xd800 <= c && c <= 0xdbff && i < len - 1) {
+			/* surrogate pair */
+			unsigned int c2 = *p;
+			if(0xdc00 <= c2 && c2 <= 0xdfff) {
+				/* valid second surrogate */
+				c = ((c - 0xd800) << 10) + (c2 - 0xdc00) + 0x10000;
+				++p;
+				++i;
+			}
+		}
+		if( c<0x80 ){
+			*q++=c;
+			out++;
+		}else if( c<0x800 ){
+			if (out > length - 2) {
+				break;
+			}
+			*q++=0xc0|(c>>6);
+			*q++=0x80|(c&0x3f);
+			out += 2;
+		}else if(c < 0x10000) { 
+			if (out > length - 3) {
+				break;
+			}
+			*q++=0xe0|(c>>12);
+			*q++=0x80|((c>>6)&0x3f);
+			*q++=0x80|(c&0x3f);
+			out += 3;
+		}else if(c <= 0x10ffff) {
+			if (out > length - 4) {
+				break;
+			}
+			*q++ = 0xf0|(c>>18);
+			*q++ = 0x80|((c>>12)&0x3f);
+			*q++ = 0x80|((c>>6)&0x3f);
+			*q++ = 0x80|((c&0x3f));
+			out += 4;
+		}else{
+			bbExThrowCString( "Unicode character out of UTF-8 range" );
+		}
+		++i;
+	}
+	*q=0;
+}
+
 void bmx_stringbuilder_format_string(struct MaxStringBuilder * buf, BBString * formatText, BBString * value) {
 	char formatBuf[256];
 	bmx_stringbuilder_toutf8_buffer(formatText, formatBuf, sizeof(formatBuf));
@@ -974,3 +1029,325 @@ BBArray * bmx_stringbuilder_splitbuffer_toarray(struct MaxSplitBuffer * buf) {
 	}
 	return bits;
 }
+
+struct MaxSplitBuffer * bmx_stringbuilder_splitbuffer_split(struct MaxSplitBuffer * splitBuffer, BBString * separator, int index) {
+    if (index < 0 || index >= splitBuffer->count) {
+        return NULL;
+    }
+
+    // Extract the segment we want to split further
+    int start = splitBuffer->startIndex[index];
+    int end = splitBuffer->endIndex[index];
+    
+    // Create a temporary buffer for just this segment
+    struct MaxStringBuilder tempBuf;
+    tempBuf.buffer = splitBuffer->buffer->buffer + start;
+    tempBuf.count = end - start;
+    tempBuf.capacity = end - start;
+    tempBuf.hash = 0;
+
+    // First, count how many new segments we will create
+    int count = 1;
+    int offset = 0;
+    int i = 0;
+
+    while ((offset = bmx_stringbuilder_find(&tempBuf, separator, i)) != -1) {
+        ++count;
+        i = offset + separator->length;
+    }
+
+    // Allocate memory for new split buffer
+    struct MaxSplitBuffer * newSplitBuffer = malloc(sizeof(struct MaxSplitBuffer));
+    newSplitBuffer->buffer = splitBuffer->buffer; // Reference the original buffer
+    newSplitBuffer->count = count;
+    newSplitBuffer->startIndex = malloc(count * sizeof(int));
+    newSplitBuffer->endIndex = malloc(count * sizeof(int));
+
+    int * bufferStartIndex = newSplitBuffer->startIndex;
+    int * bufferEndIndex = newSplitBuffer->endIndex;
+
+    // Perform the actual split
+    i = 0;
+    int subSegmentStart = 0;
+    while ((offset = bmx_stringbuilder_find(&tempBuf, separator, i)) != -1) {
+        *bufferStartIndex++ = start + subSegmentStart;
+        *bufferEndIndex++ = start + offset;
+        subSegmentStart = offset + separator->length;
+        i = subSegmentStart;
+    }
+
+    // Handle the last segment (or the whole segment if no separator was found)
+    *bufferStartIndex++ = start + subSegmentStart;
+    *bufferEndIndex++ = end;
+
+    return newSplitBuffer;
+}
+
+int bmx_stringbuilder_splitbuffer_toint(struct MaxSplitBuffer * splitBuffer, int index) {
+    if (index < 0 || index >= splitBuffer->count) {
+        return 0;
+    }
+
+    // Get the start and end positions of the segment in the original buffer
+    int start = splitBuffer->startIndex[index];
+    int end = splitBuffer->endIndex[index];
+
+    int length = end - start;
+
+    // If the segment is empty, return 0
+    if (length == 0) {
+        return 0;
+    }
+
+    BBChar *segment = (BBChar *)&(splitBuffer->buffer->buffer[start]);
+
+	char numbuf[256];
+	bmx_stringbuilder_toutf8_sbuffer(segment, length, numbuf, sizeof(numbuf));
+
+    char *endPtr;
+    errno = 0;
+    long result = strtol(numbuf, &endPtr, 10);
+
+    // Make sure that endPtr does not exceed the bounds of the segment
+    if (endPtr > numbuf + length) {
+        return 0;
+    }
+
+    if (errno == ERANGE || result > INT_MAX || result < INT_MIN || endPtr == numbuf) {
+        return 0;
+    }
+
+    return (int)result;
+}
+
+unsigned int bmx_stringbuilder_splitbuffer_touint(struct MaxSplitBuffer * splitBuffer, int index) {
+    if (index < 0 || index >= splitBuffer->count) {
+        return 0;
+    }
+
+    int start = splitBuffer->startIndex[index];
+    int end = splitBuffer->endIndex[index];
+    int length = end - start;
+
+    if (length == 0) {
+        return 0;
+    }
+
+    BBChar *segment = (BBChar *) &(splitBuffer->buffer->buffer[start]);
+
+	char numbuf[256];
+	bmx_stringbuilder_toutf8_sbuffer(segment, length, numbuf, sizeof(numbuf));
+
+    char *endPtr;
+    errno = 0;
+    unsigned long result = strtoul(numbuf, &endPtr, 10);
+
+    if (endPtr > numbuf + length || errno == ERANGE || endPtr == numbuf || result > UINT_MAX) {
+        return 0;
+    }
+
+    return (unsigned int)result;
+}
+
+float bmx_stringbuilder_splitbuffer_tofloat(struct MaxSplitBuffer * splitBuffer, int index) {
+    if (index < 0 || index >= splitBuffer->count) {
+        return 0.0f;
+    }
+
+    int start = splitBuffer->startIndex[index];
+    int end = splitBuffer->endIndex[index];
+    int length = end - start;
+
+    if (length == 0) {
+        return 0.0f;
+    }
+
+    BBChar *segment = (BBChar *) &(splitBuffer->buffer->buffer[start]);
+
+	char numbuf[256];
+	bmx_stringbuilder_toutf8_sbuffer(segment, length, numbuf, sizeof(numbuf));
+
+    char *endPtr;
+    errno = 0;
+    float result = strtof(numbuf, &endPtr);
+
+    if (endPtr > numbuf + length || errno == ERANGE || endPtr == numbuf) {
+        return 0.0f;
+    }
+
+    return result;
+}
+
+double bmx_stringbuilder_splitbuffer_todouble(struct MaxSplitBuffer * splitBuffer, int index) {
+    if (index < 0 || index >= splitBuffer->count) {
+        return 0.0;
+    }
+
+    int start = splitBuffer->startIndex[index];
+    int end = splitBuffer->endIndex[index];
+    int length = end - start;
+
+    if (length == 0) {
+        return 0.0;
+    }
+
+    BBChar *segment = (BBChar *) &(splitBuffer->buffer->buffer[start]);
+	
+	char numbuf[256];
+	bmx_stringbuilder_toutf8_sbuffer(segment, length, numbuf, sizeof(numbuf));
+
+    char *endPtr;
+    errno = 0;
+    double result = strtod(numbuf, &endPtr);
+
+    if (endPtr > numbuf + length || errno == ERANGE || endPtr == numbuf) {
+        return 0.0;
+    }
+
+    return result;
+}
+
+BBInt64 bmx_stringbuilder_splitbuffer_tolong(struct MaxSplitBuffer * splitBuffer, int index) {
+    if (index < 0 || index >= splitBuffer->count) {
+        return 0;
+    }
+
+    int start = splitBuffer->startIndex[index];
+    int end = splitBuffer->endIndex[index];
+    int length = end - start;
+
+    if (length == 0) {
+        return 0;
+    }
+
+    BBChar *segment = (BBChar *) &(splitBuffer->buffer->buffer[start]);
+
+	char numbuf[256];
+	bmx_stringbuilder_toutf8_sbuffer(segment, length, numbuf, sizeof(numbuf));
+
+    char *endPtr;
+    errno = 0;
+    BBInt64 result = strtoll(numbuf, &endPtr, 10);
+
+    if (endPtr > numbuf + length || errno == ERANGE || endPtr == numbuf) {
+        return 0;
+    }
+
+    return result;
+}
+
+BBUInt64 bmx_stringbuilder_splitbuffer_toulong(struct MaxSplitBuffer * splitBuffer, int index) {
+    if (index < 0 || index >= splitBuffer->count) {
+        return 0;
+    }
+
+    int start = splitBuffer->startIndex[index];
+    int end = splitBuffer->endIndex[index];
+    int length = end - start;
+
+    if (length == 0) {
+        return 0;
+    }
+
+    BBChar *segment = (BBChar *) &(splitBuffer->buffer->buffer[start]);
+
+	char numbuf[256];
+	bmx_stringbuilder_toutf8_sbuffer(segment, length, numbuf, sizeof(numbuf));
+
+    char *endPtr;
+    errno = 0;
+    BBUInt64 result = strtoull(numbuf, &endPtr, 10);
+
+    if (endPtr > numbuf + length || errno == ERANGE || endPtr == numbuf) {
+        return 0;
+    }
+
+    return result;
+}
+
+size_t bmx_stringbuilder_splitbuffer_tosizet(struct MaxSplitBuffer * splitBuffer, int index) {
+    if (index < 0 || index >= splitBuffer->count) {
+        return 0;
+    }
+
+    int start = splitBuffer->startIndex[index];
+    int end = splitBuffer->endIndex[index];
+    int length = end - start;
+
+    if (length == 0) {
+        return 0;
+    }
+
+    BBChar *segment = (BBChar *) &(splitBuffer->buffer->buffer[start]);
+
+	char numbuf[256];
+	bmx_stringbuilder_toutf8_sbuffer(segment, length, numbuf, sizeof(numbuf));
+
+    char *endPtr;
+    errno = 0;
+    unsigned long long result = strtoull(numbuf, &endPtr, 10);
+
+    if (endPtr > numbuf + length || errno == ERANGE || endPtr == numbuf) {
+        return 0;
+    }
+
+    return (size_t)result;
+}
+
+BBSHORT bmx_stringbuilder_splitbuffer_toshort(struct MaxSplitBuffer * splitBuffer, int index) {
+    if (index < 0 || index >= splitBuffer->count) {
+        return 0;
+    }
+
+    int start = splitBuffer->startIndex[index];
+    int end = splitBuffer->endIndex[index];
+    int length = end - start;
+
+    if (length == 0) {
+        return 0;
+    }
+
+    BBChar *segment = (BBChar *) &(splitBuffer->buffer->buffer[start]);
+
+	char numbuf[256];
+	bmx_stringbuilder_toutf8_sbuffer(segment, length, numbuf, sizeof(numbuf));
+
+    char *endPtr;
+    errno = 0;
+    long result = strtol(numbuf, &endPtr, 10);
+
+    if (endPtr > numbuf + length || errno == ERANGE || endPtr == numbuf || result < 0 || result > USHRT_MAX) {
+        return 0;
+    }
+
+    return (BBSHORT)result;
+}
+
+BBBYTE bmx_stringbuilder_splitbuffer_tobyte(struct MaxSplitBuffer * splitBuffer, int index) {
+    if (index < 0 || index >= splitBuffer->count) {
+        return 0;
+    }
+
+    int start = splitBuffer->startIndex[index];
+    int end = splitBuffer->endIndex[index];
+    int length = end - start;
+
+    if (length == 0) {
+        return 0;
+    }
+
+    BBChar *segment = (BBChar *) &(splitBuffer->buffer->buffer[start]);
+
+	char numbuf[256];
+	bmx_stringbuilder_toutf8_sbuffer(segment, length, numbuf, sizeof(numbuf));
+
+    char *endPtr;
+    errno = 0;
+    long result = strtol(numbuf, &endPtr, 10);
+
+    if (endPtr > numbuf + length || errno == ERANGE || endPtr == numbuf || result < 0 || result > UCHAR_MAX) {
+        return 0;
+    }
+
+    return (BBBYTE)result;
+}

stringbuilder.mod/glue.h

@@ -1,5 +1,5 @@
 /*
-  Copyright (c) 2018-2023 Bruce A Henderson
+  Copyright (c) 2018-2024 Bruce A Henderson
   
   This software is provided 'as-is', without any express or implied
   warranty. In no event will the authors be held liable for any damages
@@ -89,6 +89,7 @@ void bmx_stringbuilder_rightalign(struct MaxStringBuilder * buf, int length);
 char * bmx_stringbuilder_toutf8string(struct MaxStringBuilder * buf);
 BBChar * bmx_stringbuilder_towstring(struct MaxStringBuilder * buf);
 void bmx_stringbuilder_toutf8_buffer(BBString *str, char * buf, size_t length);
+void bmx_stringbuilder_toutf8_sbuffer(BBChar * p, int len, char * buf, size_t length);
 void bmx_stringbuilder_format_string(struct MaxStringBuilder * buf, BBString * formatText, BBString * value);
 void bmx_stringbuilder_format_byte(struct MaxStringBuilder * buf, BBString * formatText, BBBYTE value);
 void bmx_stringbuilder_format_short(struct MaxStringBuilder * buf, BBString * formatText, BBSHORT value);
@@ -109,5 +110,15 @@ int bmx_stringbuilder_splitbuffer_length(struct MaxSplitBuffer * buf);
 BBString * bmx_stringbuilder_splitbuffer_text(struct MaxSplitBuffer * buf, int index);
 void bmx_stringbuilder_splitbuffer_free(struct MaxSplitBuffer * buf);
 BBArray * bmx_stringbuilder_splitbuffer_toarray(struct MaxSplitBuffer * buf);
+struct MaxSplitBuffer * bmx_stringbuilder_splitbuffer_split(struct MaxSplitBuffer * splitBuffer, BBString * separator, int index);
+int bmx_stringbuilder_splitbuffer_toint(struct MaxSplitBuffer * splitBuffer, int index);
+unsigned int bmx_stringbuilder_splitbuffer_touint(struct MaxSplitBuffer * splitBuffer, int index);
+float bmx_stringbuilder_splitbuffer_tofloat(struct MaxSplitBuffer * splitBuffer, int index);
+double bmx_stringbuilder_splitbuffer_todouble(struct MaxSplitBuffer * splitBuffer, int index);
+BBInt64 bmx_stringbuilder_splitbuffer_tolong(struct MaxSplitBuffer * splitBuffer, int index);
+BBUInt64 bmx_stringbuilder_splitbuffer_toulong(struct MaxSplitBuffer * splitBuffer, int index);
+size_t bmx_stringbuilder_splitbuffer_tosizet(struct MaxSplitBuffer * splitBuffer, int index);
+BBSHORT bmx_stringbuilder_splitbuffer_toshort(struct MaxSplitBuffer * splitBuffer, int index);
+BBBYTE bmx_stringbuilder_splitbuffer_tobyte(struct MaxSplitBuffer * splitBuffer, int index);
 
 #endif

stringbuilder.mod/stringbuilder.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2018-2023 Bruce A Henderson
+' Copyright (c) 2018-2024 Bruce A Henderson
 ' 
 ' This software is provided 'as-is', without any express or implied
 ' warranty. In no event will the authors be held liable for any damages
@@ -23,10 +23,13 @@ bbdoc: A string builder.
 End Rem	
 Module BRL.StringBuilder
 
-ModuleInfo "Version: 1.18"
+ModuleInfo "Version: 1.19"
 ModuleInfo "License: zlib/libpng"
-ModuleInfo "Copyright: 2018-2023 Bruce A Henderson"
+ModuleInfo "Copyright: 2018-2024 Bruce A Henderson"
 
+ModuleInfo "History: 1.19"
+ModuleInfo "History: Added TSplitBuffer Split() method."
+ModuleInfo "History: Added TSplitBuffer to number methods."
 ModuleInfo "History: 1.18"
 ModuleInfo "History: Added optional startIndex to StartsWith()."
 ModuleInfo "History: 1.17"
@@ -972,6 +975,88 @@ Public
 		Return bmx_stringbuilder_splitbuffer_toarray(splitPtr)
 	End Method
 
+	Rem
+	bbdoc: Returns the split element at the specified index as an #Int.
+	about: If the element is not a valid #Int, 0 is returned.
+	End Rem
+	Method ToInt:Int(index:Int)
+		Return bmx_stringbuilder_splitbuffer_toint(splitPtr, index)
+	End Method
+
+	Rem
+	bbdoc: Returns the split element at the specified index as a #Float.
+	about: If the element is not a valid #Float, 0.0 is returned.
+	End Rem
+	Method ToFloat:Float(index:Int)
+		Return bmx_stringbuilder_splitbuffer_tofloat(splitPtr, index)
+	End Method
+
+	Rem
+	bbdoc: Returns the split element at the specified index as a #Double.
+	about: If the element is not a valid #Double, 0.0 is returned.
+	End Rem
+	Method ToDouble:Double(index:Int)
+		Return bmx_stringbuilder_splitbuffer_todouble(splitPtr, index)
+	End Method
+
+	Rem
+	bbdoc: Returns the split element at the specified index as a #Long.
+	about: If the element is not a valid #Long, 0 is returned.
+	End Rem
+	Method ToLong:Long(index:Int)
+		Return bmx_stringbuilder_splitbuffer_tolong(splitPtr, index)
+	End Method
+
+	Rem
+	bbdoc: Returns the split element at the specified index as a #ULong.
+	about: If the element is not a valid #ULong, 0 is returned.
+	End Rem
+	Method ToULong:ULong(index:Int)
+		Return bmx_stringbuilder_splitbuffer_toulong(splitPtr, index)
+	End Method
+
+	Rem
+	bbdoc: Returns the split element at the specified index as a #UInt.
+	about: If the element is not a valid #UInt, 0 is returned.
+	End Rem
+	Method ToUInt:UInt(index:Int)
+		Return bmx_stringbuilder_splitbuffer_touint(splitPtr, index)
+	End Method
+
+	Rem
+	bbdoc: Returns the split element at the specified index as a #Short.
+	about: If the element is not a valid #Short, 0 is returned.
+	End Rem
+	Method ToShort:Short(index:Int)
+		Return bmx_stringbuilder_splitbuffer_toshort(splitPtr, index)
+	End Method
+
+	Rem
+	bbdoc: Returns the split element at the specified index as a #Byte.
+	about: If the element is not a valid #Byte, 0 is returned.
+	End Rem
+	Method ToByte:Byte(index:Int)
+		Return bmx_stringbuilder_splitbuffer_tobyte(splitPtr, index)
+	End Method
+
+	Rem
+	bbdoc: Returns the split element at the specified index as a #Size_T.
+	about: If the element is not a valid #Size_T, 0 is returned.
+	End Rem
+	Method ToSizeT:Size_T(index:Int)
+		Return bmx_stringbuilder_splitbuffer_tosizet(splitPtr, index)
+	End Method
+
+	Rem
+	bbdoc: Creates a new split buffer of the split element at the specified index.
+	End Rem
+	Method Split:TSplitBuffer(index:Int, separator:String)
+		Local buf:TSplitBuffer = New TSplitBuffer
+		buf.buffer = buffer
+		buf.splitPtr = bmx_stringbuilder_splitbuffer_split(splitPtr, separator, index)
+		Return buf
+	End Method
+
 	Method ObjectEnumerator:TSplitBufferEnum()
 		Local enumeration:TSplitBufferEnum = New TSplitBufferEnum
 		enumeration.buffer = Self

stringbuilder.mod/tests/test.bmx

@@ -188,3 +188,213 @@ Type TStringBuilderTest Extends TTest
 	End Method
 
 End Type
+
+Type TSplitBufferTest Extends TTest
+
+	Field sb:TStringBuilder
+	
+	Method setup() { before }
+		sb = New TStringBuilder
+	End Method
+
+	Method testSplitBuffer() { test }
+		sb.Append("a,b,c,d,e,f,g")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		assertEquals(7, split.Length())
+		assertEquals("a", split.Text(0))
+		assertEquals("d", split.Text(3))
+		assertEquals("g", split.Text(6))
+		
+	End Method
+
+	Method testSplitBufferEmpty() { test }
+		sb.Append("a,b,c,d,e,f,g")
+
+		Local split:TSplitBuffer = sb.Split(" ")
+
+		assertEquals(1, split.Length())
+		assertEquals("a,b,c,d,e,f,g", split.Text(0))
+		
+	End Method
+
+	Method testSplitBufferEmptyString() { test }
+		sb.Append("")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		assertEquals(1, split.Length())
+		assertEquals("", split.Text(0))
+		
+	End Method
+
+	Method testSplitBufferEmptySeparator() { test }
+		sb.Append("a,b,c,d,e,f,g")
+
+		Local split:TSplitBuffer = sb.Split("")
+
+		assertEquals(1, split.Length())
+		assertEquals("a,b,c,d,e,f,g", split.Text(0))
+		
+	End Method
+
+	Method testSplitBufferEmptyFields() { test }
+		sb.Append("1,,,3,4,5")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		assertEquals(6, split.Length())
+		assertEquals("1", split.Text(0))
+		assertEquals("", split.Text(1))
+		assertEquals("3", split.Text(3))
+		assertEquals("5", split.Text(5))
+
+	End Method
+
+	Method testSplitBufferSplit() { test }
+		sb.Append("1,2,3|4,5,6|7,8,9")
+
+		Local split:TSplitBuffer = sb.Split("|")
+
+		assertEquals(3, split.Length())
+		assertEquals("1,2,3", split.Text(0))
+
+		Local split2:TSplitBuffer = split.Split(0, ",")
+
+		assertEquals(3, split2.Length())
+
+		assertEquals("1", split2.Text(0))
+		assertEquals("2", split2.Text(1))
+		assertEquals("3", split2.Text(2))
+	End Method
+
+	Method testSplitBufferSplitEmptyFields() { test }
+		sb.Append("1,2,3|4,,6|7,8,9")
+
+		Local split:TSplitBuffer = sb.Split("|")
+
+		assertEquals(3, split.Length())
+		assertEquals("4,,6", split.Text(1))
+
+		Local split2:TSplitBuffer = split.Split(1, ",")
+
+		assertEquals(3, split2.Length())
+
+		assertEquals("4", split2.Text(0))
+		assertEquals("", split2.Text(1))
+		assertEquals("6", split2.Text(2))
+	End Method
+
+	Method testSplitBufferSplitEmptyFields2() { test }
+		sb.Append("1,2,3||7,8,9")
+
+		Local split:TSplitBuffer = sb.Split("|")
+
+		assertEquals(3, split.Length())
+		assertEquals("", split.Text(1))
+
+		Local split2:TSplitBuffer = split.Split(1, ",")
+
+		assertEquals(1, split2.Length())
+		assertEquals("", split2.Text(0))
+	End Method
+
+	Method testSplitBufferEnumeration() { test }
+		sb.Append("a,b,c,d,e,f,g")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		Local txt:String
+		For Local s:String = EachIn split
+			txt :+ s
+		Next
+
+		assertEquals("abcdefg", txt)
+	End Method
+
+	Method testSplitBufferSplitEnumeration() { test }
+		sb.Append("1,2,3|4,5,6|7,8,9")
+
+		Local split:TSplitBuffer = sb.Split("|")
+		Local split2:TSplitBuffer = split.Split(1, ",")
+
+		Local txt:String
+		For Local s:String = EachIn split2
+			txt :+ s
+		Next
+
+		assertEquals("456", txt)
+	End Method
+
+	Method testSplitBufferToInt() { test }
+		sb.Append("1,22,333,4444,-55555,666666,777777,8888888,99999999")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		assertEquals(1, split.ToInt(0))
+		assertEquals(-55555, split.ToInt(4))
+		assertEquals(99999999, split.ToInt(8))
+	End Method
+
+	Method testSplitBufferToFloat() { test }
+		sb.Append("1.1,2.2,3.3,4.4,5.5,6.6,7.7,8.8,-9.9")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		assertEquals(1.1, split.ToFloat(0), 0.0001)
+		assertEquals(5.5, split.ToFloat(4), 0.0001)
+		assertEquals(-9.9, split.ToFloat(8), 0.0001)
+	End Method
+
+	Method testSplitBufferToDouble() { test }
+		sb.Append("1.1,2.2,3.3,4.4,-5.5,6.6,7.7,8.8,9.9")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		assertEquals(1.1, split.ToDouble(0), 0.0001)
+		assertEquals(-5.5, split.ToDouble(4), 0.0001)
+		assertEquals(9.9, split.ToDouble(8), 0.0001)
+	End Method
+
+	Method testSplitBufferToShort() { test }
+		sb.Append("1,2,3,4,5,6,7,8,9")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		assertEquals(1, split.ToShort(0))
+		assertEquals(5, split.ToShort(4))
+		assertEquals(9, split.ToShort(8))
+	End Method
+
+	Method testSplitBufferToByte() { test }
+		sb.Append("1,2,3,4,5,6,7,8,9")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		assertEquals(1, split.ToByte(0))
+		assertEquals(5, split.ToByte(4))
+		assertEquals(9, split.ToByte(8))
+	End Method
+
+	Method testSplitBufferToLong() { test }
+		sb.Append("-1,2,3,4,5,6,7,8,9")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		assertEquals(-1, split.ToLong(0))
+		assertEquals(5, split.ToLong(4))
+		assertEquals(9, split.ToLong(8))
+	End Method
+
+	Method testSplitBufferToULong() { test }
+		sb.Append("1111,22222,333333,4444444,55555555,666666666,777777777,8888888,99999999999")
+
+		Local split:TSplitBuffer = sb.Split(",")
+
+		assertEquals(1111, split.ToULong(0))
+		assertEquals(55555555, split.ToULong(4))
+		assertEquals(99999999999:ULong, split.ToULong(8))
+	End Method
+
+End Type

textstream.mod/textstream.bmx

@@ -512,7 +512,10 @@ Function IsProbablyUTF8:Int(data:Byte Ptr, size:Int)
 	Next
 
 	If count Then
-		Return False
+		' If there was no new-line or non-multi-byte-character at the 
+		' end of the buffer then count will be > 0. So we also have to
+		' check if we can decode the remaining buffer content.
+		If Decode(buf, count) = -1 Then Return False
 	End If
 	
 	Return True

threadpool.mod/examples/scheduled_01.bmx

@@ -0,0 +1,34 @@
+'
+' Demostrates use of a scheduled thread pool with single-shot tasks.
+'
+SuperStrict
+
+Framework BRL.Standardio
+Import BRL.ThreadPool
+
+Local pool:TScheduledThreadPoolExecutor = TScheduledThreadPoolExecutor.newFixedThreadPool(11)
+
+For Local i:Int = 10 Until 0 Step -1
+	pool.schedule(New TTask(i), 10 - i, ETimeUnit.Seconds)
+Next
+
+Print "Shutting down the pool..."
+
+pool.shutdown()
+
+Print "Done"
+
+Type TTask Extends TRunnable
+
+	Field value:Int
+
+	Method New(value:Int)
+		Self.value = value
+	End Method
+
+	Method run()
+		Print "Number " + value
+	End Method
+	
+End Type
+

threadpool.mod/examples/scheduled_02.bmx

@@ -0,0 +1,35 @@
+'
+' Demostrates use of a scheduled thread pool with recurring tasks.
+'
+SuperStrict
+
+Framework BRL.Standardio
+Import BRL.ThreadPool
+
+Local pool:TScheduledThreadPoolExecutor = TScheduledThreadPoolExecutor.newFixedThreadPool(11)
+
+pool.schedule(New TTask("One-shot Task"), 5, ETimeUnit.Seconds) ' after 5 seconds
+pool.schedule(New TTask("Recurring Task"), 3, 5, ETimeUnit.Seconds) ' after 3 seconds and then every 5 seconds
+
+Delay(10 * 1000) ' wait for 10 seconds and then shutdown the pool
+
+Print "Shutting down the pool..."
+
+pool.shutdown()
+
+Print "Done"
+
+Type TTask Extends TRunnable
+
+	Field message:String
+
+	Method New(message:String)
+		Self.message = message
+	End Method
+
+	Method run()
+		Print message
+	End Method
+	
+End Type
+

threadpool.mod/threadpool.bmx

@@ -26,11 +26,13 @@ bbdoc: System/ThreadPool
 End Rem
 Module BRL.ThreadPool
 
-ModuleInfo "Version: 1.01"
+ModuleInfo "Version: 1.02"
 ModuleInfo "Author: Bruce A Henderson"
 ModuleInfo "License: zlib/libpng"
 ModuleInfo "Copyright: Bruce A Henderson"
 
+ModuleInfo "History: 1.02"
+ModuleInfo "History: Added scheduled pool executor"
 ModuleInfo "History: 1.01"
 ModuleInfo "History: Added cached pool executor"
 ModuleInfo "History: 1.00"
@@ -40,7 +42,8 @@ ModuleInfo "History: Initial Release"
 
 Import BRL.Threads
 Import BRL.LinkedList
-
+Import BRL.Time
+Import pub.stdc
 
 Rem
 bbdoc: An object that is intended to be executed by a thread pool.
@@ -181,24 +184,30 @@ Type TThreadPoolExecutor Extends TExecutor
 	End Rem
 	Method execute(command:TRunnable) Override
 		If Not isShutdown Then
-			If maxThreads < 0 Then
-				Local newThread:Int
-				countLock.Lock()
-				If threadsWorking = threadsAlive Then
-					newThread = True
-				End If
-				countLock.Unlock()
-				If newThread Then
-					threadsLock.Lock()
-					threads.AddLast(New TPooledThread(Self, _processThread))
-					threadsLock.Unlock()
-				End If
+			doExecute(command)
+		End If
+	End Method
+
+Private
+	Method doExecute(command:TRunnable)
+		If maxThreads < 0 Then
+			Local newThread:Int
+			countLock.Lock()
+			If threadsWorking = threadsAlive Then
+				newThread = True
+			End If
+			countLock.Unlock()
+			If newThread Then
+				threadsLock.Lock()
+				threads.AddLast(New TPooledThread(Self, _processThread))
+				threadsLock.Unlock()
 			End If
-			jobQueue.Lock()
-			jobQueue.Add(command)
-			jobQueue.Unlock()
 		End If
+		jobQueue.Lock()
+		jobQueue.Add(command)
+		jobQueue.Unlock()
 	End Method
+Public
 
 	Rem
 	bbdoc: Creates an executor that uses a single worker thread operating off an unbounded queue.
@@ -279,6 +288,177 @@ Type TThreadPoolExecutor Extends TExecutor
 	end method
 End Type
 
+Rem
+bbdoc: An executor that can be used to schedule commands to run after a given delay, or to execute commands periodically.
+End Rem
+Type TScheduledThreadPoolExecutor Extends TThreadPoolExecutor
+
+	Field tasks:TScheduledTask
+
+	Field taskMutex:TMutex
+	Field taskCond:TCondVar
+
+	Field schedulerThread:TThread
+
+	Method New(initial:Int, idleWait:Int = 0)
+		Super.New(initial, idleWait)
+		taskMutex = TMutex.Create()
+		taskCond = TCondVar.Create()
+
+		schedulerThread = CreateThread(taskScheduler, Self)
+	End Method
+
+	Method schedule(command:TRunnable, delay_:Int, unit:ETimeUnit = ETimeUnit.Milliseconds)
+		schedule(command, ULong(delay_), 0, unit)
+	End Method
+
+	Method schedule(command:TRunnable, initialDelay:Int, period:Int, unit:ETimeUnit = ETimeUnit.Milliseconds)
+		schedule(command, ULong(initialDelay), ULong(period), unit)
+	End Method
+	
+	Rem
+	bbdoc: Schedules a one-shot command to run after a given delay.
+	End Rem
+	Method schedule(command:TRunnable, delay_:ULong, unit:ETimeUnit = ETimeUnit.Milliseconds)
+		schedule(command, delay_, 0, unit)
+	End Method
+
+	Rem
+	bbdoc: Schedules a recurring command to run after a given initial delay, and subsequently with the given period.
+	End Rem
+	Method schedule(command:TRunnable, initialDelay:ULong, period:ULong, unit:ETimeUnit = ETimeUnit.Milliseconds)
+		Local now:ULong = CurrentUnixTime()
+
+		Local newTask:TScheduledTask = New TScheduledTask
+
+		Local delayMs:ULong = TimeUnitToMillis(initialDelay, unit)
+		Local periodMs:ULong = TimeUnitToMillis(period, unit)
+
+		newTask.executeAt = now + delayMs
+		newTask.intervalMs = periodMs
+		newTask.command = command
+		
+		taskMutex.Lock()
+		
+		insertTask(newTask)
+
+		taskMutex.Unlock()
+		
+	End Method
+
+	Rem
+	bbdoc: Initiates an orderly shutdown in which previously submitted tasks are executed, but no new tasks will be accepted.
+	End Rem
+	Method shutdown() Override
+		isShutdown = True
+		schedulerThread.Wait()
+		Super.shutdown()
+	End Method
+
+Private
+	Method insertTask(newTask:TScheduledTask)
+		Local headChanged:Int = False
+		If Not tasks Or newTask.executeAt < tasks.executeAt Then
+			newTask.nextTask = tasks
+			tasks = newTask
+			headChanged = True
+		Else
+			Local current:TScheduledTask = tasks
+			While current.nextTask And current.nextTask.executeAt < newTask.executeAt
+				current = current.nextTask
+			Wend
+			newTask.nextTask = current.nextTask
+			current.nextTask = newTask
+		End If
+
+		If headChanged Then
+			taskCond.Signal()
+		End If
+	End Method
+
+	Function taskScheduler:Object( data:Object )
+		Local exec:TScheduledThreadPoolExecutor = TScheduledThreadPoolExecutor(data)
+
+		While True
+
+			exec.taskMutex.Lock()
+
+			While Not exec.tasks
+				
+				If exec.isShutdown Then
+					exec.taskMutex.Unlock()
+					Return Null
+				End If
+
+				exec.taskCond.Wait(exec.taskMutex)
+			Wend
+
+			Local now:ULong = CurrentUnixTime()
+
+			If now < exec.tasks.executeAt Then
+				' Wait until the next task is due or a new task is scheduled
+				exec.taskCond.TimedWait(exec.taskMutex, Int(exec.tasks.executeAt - now))
+			End If
+
+			now = CurrentUnixTime()
+
+			While exec.tasks And exec.tasks.executeAt <= now
+				Local task:TScheduledTask = exec.tasks
+
+				exec.doExecute(task.command)
+
+				If task.intervalMs And Not exec.isShutdown Then
+					' If the task is recurring, reschedule it, unless the executor is shutting down
+					task.executeAt = now + task.intervalMs
+					exec.tasks = task.nextTask
+					exec.insertTask(task)
+				Else
+					' Otherwise, remove it from the list
+					exec.tasks = task.nextTask
+				End If
+			Wend
+
+			exec.taskMutex.Unlock()
+		Wend
+	End Function
+Public
+
+Rem
+	bbdoc: Creates an executor that uses a single worker thread operating off an unbounded queue.
+	End Rem
+	Function newSingleThreadExecutor:TScheduledThreadPoolExecutor()
+		Return New TScheduledThreadPoolExecutor(1)
+	End Function
+	
+	Rem
+	bbdoc: Creates a thread pool that reuses a fixed number of threads operating off a shared unbounded queue.
+	about: At any point, at most @threads threads will be active processing tasks. If additional tasks are
+	submitted when all threads are active, they will wait in the queue until a thread is available.
+	End Rem
+	Function newFixedThreadPool:TScheduledThreadPoolExecutor(threads:Int)
+		Assert threads > 0
+		Return New TScheduledThreadPoolExecutor(threads)
+	End Function
+	
+	Rem
+	bbdoc: Creates a thread pool that creates new threads as needed, but will reuse previously constructed threads when they are available.
+	about: These pools will typically improve the performance of programs that execute many short-lived asynchronous tasks.
+	Threads that remain idle for more than the specified @idleWait time will be terminated and removed from the pool.
+	End Rem
+	Function newCachedThreadPool:TScheduledThreadPoolExecutor(idleWait:Int = 60000)
+		Return New TScheduledThreadPoolExecutor(-1, idleWait)
+	End Function
+End Type
+
+Type TScheduledTask
+	Field executeAt:Long ' the time to execute the task, in ms since the epoch
+	Field command:TRunnable
+	
+	Field intervalMs:ULong ' zero for one-shot tasks
+
+	Field nextTask:TScheduledTask
+End Type
+
 Private
 
 Type TBinarySemaphore

threads.mod/examples/future_01.bmx

@@ -0,0 +1,40 @@
+SuperStrict
+
+Framework brl.standardio
+Import brl.threads
+
+' Define a function to perform some computation in a background thread
+Function ComputeSomethingAsync:Object( data:Object )
+	Local future:TFuture<Float> = TFuture<Float>(data)
+
+	' Simulate a time-consuming computation with Delay
+	Local result:Float = 0.0
+	For Local i:Int = 0 Until 10
+		result :+ 0.1
+		Print "Computing... " + result
+		Delay(500)
+	Next
+	' Set the result in the future object
+	future.SetResult( result )
+End Function
+
+
+' Create a TFuture instance for Float
+Local future:TFuture<Float> = New TFuture<Float>
+
+' Start the computation in a background thread
+Local thread:TThread = CreateThread( ComputeSomethingAsync, future )
+
+' Simulate doing some other work in the main thread
+Print "Main thread is doing other work..."
+For Local j:Int = 0 Until 5
+	Print "Main work step " + j
+	Delay(300)
+Next
+
+' Wait for and retrieve the result from the future object
+Local result:Float = future.GetResult()
+Print "The result of the computation is: " + result
+
+' Wait for the background thread to finish
+thread.Wait()

threads.mod/threads.bmx

@@ -6,10 +6,12 @@ bbdoc: System/Threads
 End Rem
 Module BRL.Threads
 
-ModuleInfo "Version: 1.02"
+ModuleInfo "Version: 1.03"
 ModuleInfo "License: zlib/libpng"
 ModuleInfo "Copyright: Blitz Research Ltd"
 
+ModuleInfo "History: 1.03"
+ModuleInfo "History: Added TFuture type."
 ModuleInfo "History: 1.02"
 ModuleInfo "History: Changed to use macOS dispatch semphores."
 ModuleInfo "History: 1.01"
@@ -23,6 +25,8 @@ Import "threads_mac.m"
 
 ?Threaded
 
+Import Pub.Stdc
+Import BRL.Time
 Import "threads.c"
 
 Private
@@ -363,6 +367,130 @@ Type TCondVar
 
 End Type
 
+Rem
+bbdoc: A generic type for asynchronous result handling, allowing threads to wait for and retrieve results safely.
+about: It provides a mechanism for one thread to produce a result that another thread can wait for and retrieve
+at a later time. This is particularly useful for tasks that are executed in parallel, where the completion
+time may vary, and the consumer needs to wait for the result before proceeding.
+End Rem
+Type TFuture<V>
+Private
+	Field value:V
+
+	Field ready:Int
+
+	Field condvar:TCondVar=CreateCondVar()
+	Field mutex:TMutex=CreateMutex()
+Public
+	Rem
+	bbdoc: Waits for the result to become available and then returns it.
+	about: This method blocks the calling thread until result is available.
+	End Rem
+	Method GetResult:V()
+		mutex.Lock()
+		While Not ready
+			condvar.Wait( mutex )
+		Wend
+		Local result:V = value
+		mutex.Unlock()
+		Return result
+	End Method
+
+	Rem
+	bbdoc: Sets the result of the asynchronous operation and signals any waiting threads.
+	End Rem
+	Method SetResult( value:V )
+		mutex.Lock()
+		Self.value=value
+		Self.ready=True
+		condvar.Signal()
+		mutex.Unlock()
+	End Method
+
+End Type
+
+Rem
+bbdoc: A thread event object.
+about: A basic synchronization object that allows one thread to signal an event to other threads.
+It manages an internal flag that can be set or cleared, and provides methods to wait for the event to be set.
+End rem
+Type TThreadEvent
+    Private
+        Field lock:TMutex
+        Field condition:TCondVar
+        Field _isSet:Int
+	Public
+    Method New()
+        lock = TMutex.Create()
+        condition = TCondVar.Create()
+        _isSet = False
+    End Method
+
+    Rem
+	bboc: Sets the internal flag to #True and signals any waiting threads.
+	about: All threads waiting for it to become #True are awakened. Threads that call #Wait once the flag is true will not block at all.
+	End Rem
+    Method Set()
+        lock.Lock()
+        _isSet = True
+        condition.Broadcast()
+        lock.Unlock()
+    End Method
+
+    Rem
+	bbdoc: Resets the internal flag to false.
+	about: After clearing, threads calling #Wait will block until #Set is called to set the internal flag to #True again.
+	End Rem
+    Method Clear()
+        lock.Lock()
+        _isSet = False
+        lock.Unlock()
+    End Method
+
+    Rem
+	bbdoc: Waits for the event to be set.
+	about: This method could block indefinitely if the event is never set.
+	If the event is already set, the method returns immediately.
+	End Rem
+    Method Wait()
+        lock.Lock()
+        While Not _isSet
+            condition.Wait(lock)
+        Wend
+        lock.Unlock()
+    End Method
+
+    Rem
+	bbdoc: Waits for the event to be set, with a timeout.
+	about: If the timeout is reached before the event is set, the method returns #False.
+	End Rem
+    Method Wait:Int(timeout:ULong, unit:ETimeUnit = ETimeUnit.Milliseconds)
+        lock.Lock()
+		Local timeoutMs:ULong = TimeUnitToMillis(timeout, unit)
+        Local endTime:ULong = CurrentUnixTime() + timeoutMs
+        While Not _isSet
+            Local now:ULong = CurrentUnixTime()
+            If now >= timeoutMs Then
+                lock.Unlock()
+                Return False
+            End If
+            condition.TimedWait(lock, Int(timeoutMs - now))
+        Wend
+        lock.Unlock()
+        Return True
+    End Method
+
+    Rem
+	bbdoc: Returns whether the event is set or not.
+	End Rem
+    Method IsSet:Int()
+        lock.Lock()
+        Local result:Int = _isSet
+        lock.Unlock()
+        Return result
+    End Method
+End Type
+
 Rem
 bbdoc: Create a thread
 returns: A new thread object.

time.mod/time.bmx

@@ -0,0 +1,70 @@
+' Copyright (c)2024 Bruce A Henderson
+'
+' This software is provided 'as-is', without any express or implied
+' warranty. In no event will the authors be held liable for any damages
+' arising from the use of this software.
+'
+' Permission is granted to anyone to use this software for any purpose,
+' including commercial applications, and to alter it and redistribute it
+' freely, subject to the following restrictions:
+'
+'    1. The origin of this software must not be misrepresented; you must not
+'    claim that you wrote the original software. If you use this software
+'    in a product, an acknowledgment in the product documentation would be
+'    appreciated but is not required.
+'
+'    2. Altered source versions must be plainly marked as such, and must not be
+'    misrepresented as being the original software.
+'
+'    3. This notice may not be removed or altered from any source
+'    distribution.
+'
+SuperStrict
+
+Module BRL.Time
+
+ModuleInfo "Version: 1.0"
+ModuleInfo "Author: Bruce A Henderson"
+ModuleInfo "License: zlib/libpng"
+ModuleInfo "Copyright: Bruce A Henderson"
+
+ModuleInfo "History: 1.00"
+ModuleInfo "History: Initial Release"
+
+
+Rem
+bbdoc: A unit of date-time, such as Days or Hours.
+End Rem
+Enum ETimeUnit
+	Milliseconds
+	Seconds
+	Minutes
+	Hours
+	Days
+End Enum
+
+Rem
+bbdoc: Converts a time value to milliseconds.
+End Rem
+Function TimeUnitToMillis:ULong( value:ULong, unit:ETimeUnit )
+	Select unit
+		Case ETimeUnit.Milliseconds
+			Return value
+		Case ETimeUnit.Seconds
+			Return value * 1000
+		Case ETimeUnit.Minutes
+			Return value * 60000
+		Case ETimeUnit.Hours
+			Return value * 3600000
+		Case ETimeUnit.Days
+			Return value * 86400000
+	End Select
+End Function
+
+Type TTimeoutException Extends TRuntimeException
+
+	Method New(message:String)
+		Super.New(message)
+	End Method
+	
+End Type