Merge pull request #534 from bmx-ng/master

Sync with master

base.stringhelper.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2014-2019 Bruce A Henderson
+' Copyright (c) 2014-2020 Bruce A Henderson
 ' Copyright (c) 2014-2017 Ronny Otto
 '
 ' This software is provided 'as-is', without any express or implied

base64.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2008-2019 Bruce A Henderson
+' Copyright (c) 2008-2020 Bruce A Henderson
 ' 
 ' Permission is hereby granted, free of charge, to any person obtaining a copy
 ' of this software and associated documentation files (the "Software"), to deal

bcc.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -156,7 +156,13 @@ End Function
 Function SaveIncBinHeader(file:String, trans:TCTranslator, mung:String, app:TAppDecl)
 
 	If app.genIncBinHeader Then
-		Local path:String = OutputFilePath(file, mung, "incbin.c")
+		Local ibFile:String = "incbin"
+		If opt_legacy_incbin Then
+			ibFile :+ ".c"
+		Else
+			ibFile :+ "2.c"
+		End If
+		Local path:String = OutputFilePath(file, mung, ibFile)
 
 		SaveText(trans.JoinLines("incbin"), path)
 	End If

config.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -33,6 +33,7 @@ Import "options.bmx"
 Import "base.stringhelper.bmx"
 Import "base64.bmx"
 Import "enums.c"
+Import "hash.c"
 
 ' debugging help
 Const DEBUG:Int = False
@@ -54,6 +55,8 @@ Global POINTER_SIZE:Int = 4
 Global _symbols$[]=[ "..","[]",":*",":/",":+",":-",":|",":&",":~~",":shr",":shl",":sar",":mod"]
 Global _symbols_map$[]=[ "..","[]","*=","/=","+=","-=","|=","&=","^=",">>=", "<<=",">>=","%=" ]
 
+Global _fileHasher:TFileHash
+
 Function PushErr( errInfo$ )
 	_errStack.AddLast _errInfo
 	_errInfo=errInfo
@@ -207,13 +210,21 @@ Function BmxEnquote$( str$ )
 	Return str
 End Function
 
-Function BmxUnquote$( str$ )
-	If str.length = 1 Or str[str.length - 1] <> Asc("~q") Then
-		Err "Expecting expression but encountered malformed string literal"
+Function BmxUnquote$( str$, unquoted:Int = False )
+	Local length:Int
+	Local i:Int
+	If Not unquoted Then
+		If str.length < 2 Or str[str.length - 1] <> Asc("~q") Then
+			Err "Expecting expression but encountered malformed string literal"
+		End If
+		length = str.length - 1
+		i = 1
+	Else
+		length = str.length
 	End If
-	Local length:Int = str.length - 1
+
 	Local sb:TStringBuffer = New TStringBuffer
-	Local i:Int = 1
+
 	While i < length
 		Local c:Int = str[i]
 		i :+ 1
@@ -298,6 +309,90 @@ Function BmxUnquote$( str$ )
 	Return sb.ToString()
 End Function
 
+Function BmxProcessMultiString:String( str:String )
+	Local valid:Int
+	If str.length < 7 Then
+		Err "Expecting expression but encountered malformed multiline string literal"
+	End If
+	
+	For Local i:Int = 0 Until 3
+		If str[i] <> Asc("~q") Or str[str.length -1 -i] <> Asc("~q") Then
+			Err "Expecting expression but encountered malformed multiline string literal"
+		End If
+	Next
+	
+	str = str[3..str.length - 3]
+	' normalise line endings
+	str = str.Replace("~r~n", "~n").Replace("~r", "~n")
+
+	If str[0] <> Asc("~n") Then
+		Err "Expecting EOL but encountered malformed multiline string literal"
+	End If
+
+	str = str[1..]
+
+	Local LINES:String[] = str.Split("~n")
+
+	Local lineCount:Int = LINES.length - 1
+	Local last:String = LINES[lineCount]
+	
+	Local i:Int = last.length - 1
+	While i >= 0
+		If last[i] <> Asc(" ") And last[i] <> Asc("~t") Then
+			Err "Expecting trailing whitespace"
+		End If
+		i :- 1
+	Wend
+	
+	Local trailingIndent:String = last
+	
+	' strip indent
+	If trailingIndent Then
+		For i = 0 Until lineCount
+			Local line:String = LINES[i]
+			If line.StartsWith(trailingIndent) Then
+				line = line[trailingIndent.length..]
+				LINES[i] = line
+			End If
+		Next
+	End If
+
+	' right trim
+	For i = 0 Until lineCount
+		Local line:String = LINES[i]
+		Local index:Int = line.length
+		While index
+			index :- 1
+			If line[index] <> Asc(" ") And line[index] <> Asc("~t") Then
+				Exit
+			End If
+		Wend
+		If index < line.length - 1 Then
+			line = line[..index + 1]
+			LINES[i] = line
+		End If
+	Next
+
+	Local sb:TStringBuffer = New TStringBuffer
+	For i = 0 Until lineCount
+		Local line:String = LINES[i]
+		Local length:Int = line.length
+		Local softWrap:Int
+		If line And line[line.length-1] = Asc("\") Then
+			softWrap = True
+			length :- 1
+		End If
+		If line Then
+			sb.Append(line[..length])
+		End If
+		If Not softWrap And i < lineCount - 1 Then
+			sb.Append("~n")
+		End If
+	Next
+
+	Return BmxUnquote(sb.ToString(), True)
+End Function
+
 Type TStack Extends TList
 
 	Method Push(obj:Object)
@@ -492,6 +587,17 @@ Function HeaderComment:String()
 	' TODO
 End Function
 
+Global fileRegister:TMap = New TMap
+
+Function GenHash:String(file:String)
+	Local Hash:String = bmx_gen_hash(file)
+
+	If Not fileRegister.Contains(Hash) Then
+		fileRegister.Insert(Hash, file)
+	End If
+	
+	Return Hash
+End Function
 
 Type TTemplateRecord
 
@@ -561,7 +667,62 @@ Type TCallback
 	Method Callback(obj:Object) Abstract
 End Type
 
+Type TFileHash
+
+	Field statePtr:Byte Ptr
+	
+	Method Create:TFileHash()
+		statePtr = bmx_hash_createState()
+		Return Self
+	End Method
+	
+	Method CalculateHash:String(stream:TStream)
+		Const BUFFER_SIZE:Int = 8192
+	
+	
+		bmx_hash_reset(statePtr)
+		
+		Local data:Byte[BUFFER_SIZE]
+		
+		While True
+			Local read:Int = stream.Read(data, BUFFER_SIZE)
+
+			bmx_hash_update(statePtr, data, read)
+			
+			If read < BUFFER_SIZE Then
+				Exit
+			End If
+
+		Wend
+		
+		Return bmx_hash_digest(statePtr)
+		
+	End Method
+
+End Type
+
+Function CalculateFileHash:String(path:String)
+	If Not _fileHasher Then
+		_fileHasher = New TFileHash.Create()
+	End If
+
+	If FileType(path) = FILETYPE_FILE Then
+		Local stream:TStream = ReadStream(path)
+		Local fileHash:String = _fileHasher.CalculateHash(stream)
+		stream.Close()
+		
+		Return fileHash
+	End If
+	
+	Return Null
+End Function
+
 Extern
 	Function strlen_:Int(s:Byte Ptr)="strlen"
 	Function bmx_enum_next_power(char:Int, val:Long Var, ret:Long Var)
+	Function bmx_gen_hash:String(txt:String)
+	Function bmx_hash_createState:Byte Ptr()
+	Function bmx_hash_reset(state:Byte Ptr)
+	Function bmx_hash_update(state:Byte Ptr, data:Byte Ptr, length:Int)
+	Function bmx_hash_digest:String(state:Byte Ptr)
 End Extern

ctranslator.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -174,9 +174,13 @@ Type TCTranslator Extends TTranslator
 		If TStringType( ty ) Return "$"
 		If TArrayType( ty ) Then
 			Local s:String = "["
-			For Local i:Int = 0 Until TArrayType( ty ).dims - 1
-				s:+ ","
-			Next
+			If TArrayType( ty ).isStatic Then
+				s :+ TArrayType( ty ).length
+			Else
+				For Local i:Int = 0 Until TArrayType( ty ).dims - 1
+					s:+ ","
+				Next
+			End If
 			s:+ "]"
 			Return s + TransDebugScopeType(TArrayType( ty ).elemType)
 		End If
@@ -240,7 +244,13 @@ Type TCTranslator Extends TTranslator
 			End If
 			Return "BBSTRING" + p
 		End If
-		If TArrayType( ty ) Return "BBARRAY" + p
+		If TArrayType( ty ) Then
+			If TArrayType( ty ).isStatic Then
+				Return TransType(TArrayType( ty ).elemType, ident)
+			Else
+				Return "BBARRAY" + p
+			End If
+		End If
 		If TObjectType( ty ) Then
 			Return TransObject(TObjectType(ty).classdecl) + p
 		End If
@@ -334,9 +344,13 @@ Type TCTranslator Extends TTranslator
 		End If
 		If TArrayType( ty )  Then
 			Local s:String = TransIfcType(TArrayType( ty ).elemType) + "&["
-			For Local i:Int = 0 Until TArrayType( ty ).dims - 1
-				s:+ ","
-			Next
+			If TArrayType( ty ).isStatic Then
+				s :+ TArrayType( ty ).length
+			Else
+				For Local i:Int = 0 Until TArrayType( ty ).dims - 1
+					s:+ ","
+				Next
+			End If
 			Return s + "]" + p
 		End If
 		If TObjectType( ty ) Then
@@ -392,7 +406,7 @@ Type TCTranslator Extends TTranslator
 		Return TransType( ty, ident )
 	End Method
 
-	Method TransValue$( ty:TType,value$ )
+	Method TransValue$( ty:TType,value$, isStructInit:Int = False )
 		If value
 			If IsPointerType(ty, 0, TType.T_POINTER) Return value
 			If TBoolType( ty ) Return "1"
@@ -456,17 +470,69 @@ Type TCTranslator Extends TTranslator
 				End If
 			End If
 			If TNumericType( ty ) Return "0" ' numeric and pointers
-			If TStringType( ty ) Return Bra("&bbEmptyString")
-			If TArrayType( ty ) Return Bra("&bbEmptyArray")
+			If TStringType( ty ) Then
+				If isStructInit Then
+					Return "&bbEmptyString"
+				Else
+					Return Bra("&bbEmptyString")
+				End If
+			End If
+			If TArrayType( ty ) Then
+				If isStructInit Then 
+					If TArrayType( ty ).isStatic Then
+						Local t:String = "{"
+						For Local i:Int = 0 Until Int(TArrayType( ty ).length)
+							If i Then
+								t :+ ","
+							End If
+							t :+ TransValue(TArrayType( ty ).elemType, "", True)
+						Next
+						Return t + "}"
+					Else
+						Return "&bbEmptyArray"
+					End If
+				Else
+					Return Bra("&bbEmptyArray")
+				End If
+			End If
 			If TObjectType( ty ) Then
 				If TObjectType( ty ).classDecl.IsExtern() Or TObjectType( ty ).classDecl.IsStruct() Then
 					If TObjectType( ty ).classDecl.IsInterface() Or IsPointerType(ty,0,TType.T_POINTER) Or (Not TObjectType( ty ).classDecl.IsStruct()) Then
 						Return "0"
 					Else
-						Return "{}"
+						If TObjectType( ty ).classDecl.IsStruct() Then
+
+							Local t:String
+							If Not isStructInit Then
+								t = "((" + TransType(ty, "") + "){"
+							Else
+								t = "{"
+							End If
+							Local fields:Int
+							For Local f:TFieldDecl = EachIn TObjectType( ty ).classDecl.Decls()
+								If fields Then
+									t :+ ","
+								End If
+								fields = True
+								
+								t :+ TransValue(f.ty, "", True)
+							Next
+							If Not isStructInit Then
+								t :+ "})"
+							Else
+								t :+ "}"
+							End If
+							Return t 
+						Else
+							Return "{}"
+						End If
 					End If
 				Else
-					Return Bra("&bbNullObject")
+					If isStructInit Then
+						Return "&bbNullObject"
+					Else
+						Return Bra("&bbNullObject")
+					End If
 				End If
 			End If
 			If TFunctionPtrType( ty) Return "(&brl_blitz_NullFunctionError)" ' todo ??
@@ -581,7 +647,7 @@ Type TCTranslator Extends TTranslator
 
 					' Object -> Byte Ptr
 					If IsPointerType(ty, TType.T_BYTE) And TObjectType(arg.exprType) Then
-						t:+ varRef + Bra(Bra("(BBBYTE*)" + Bra(arg.Trans())) + "+" + Bra("sizeof(void*)"))
+						t:+ varRef + Bra("bbObjectToFieldOffset" + Bra(arg.Trans()))
 						Continue
 					End If
 
@@ -744,7 +810,11 @@ t:+"NULLNULLNULL"
 			If TInvokeExpr(init) And Not TInvokeExpr(init).invokedWithBraces Then
 				initTrans = "=" + cast + TInvokeExpr(init).decl.munged
 			Else
-				initTrans = "=" + cast + init.Trans()
+				If Not TArrayType(decl.ty) Or Not TArrayType(decl.ty).isStatic Then
+					initTrans = "=" + cast + init.Trans()
+				Else
+					initTrans = "[" + TArrayType(decl.ty).length + "]=" + TransValue(decl.ty, Null, True)
+				End If
 			End If
 		End If
 		
@@ -760,7 +830,7 @@ t:+"NULLNULLNULL"
 					If Not TConstExpr(init) Then
 						Return decl.munged + initTrans
 					End If
-				Else
+				Else If Not TArrayType(ty) Or Not TArrayType(ty).isStatic Then 
 					Return decl.munged + initTrans
 				End If
 			Else If TObjectType( ty ) And TObjectType( ty ).classDecl.IsStruct() Then
@@ -833,7 +903,11 @@ t:+"NULLNULLNULL"
 				If TLocalDecl(decl) And TLocalDecl(decl).volatile Then
 					Return TransType( decl.ty, decl.munged )+" volatile "+decl.munged + "=" + TransValue(decl.ty, "")
 				Else
-					Return TransType( decl.ty, decl.munged )+" "+decl.munged + "=" + TransValue(decl.ty, "")
+					If TArrayType(decl.ty) And TArrayType(decl.ty).isStatic Then
+						Return TransType( decl.ty, decl.munged )+" "+decl.munged + "[" + TArrayType(decl.ty).length + "]"
+					Else
+						Return TransType( decl.ty, decl.munged )+" "+decl.munged + "=" + TransValue(decl.ty, "")
+					End If
 				End If
 			End If
 		End If
@@ -1085,10 +1159,14 @@ t:+"NULLNULLNULL"
 '								Return ifc + "->" + TransFuncPrefix(cdecl, decl) + FuncDeclMangleIdent(decl)+TransArgs( args,decl, lvar )
 							Else
 								If cdecl And cdecl.IsStruct() Then
+									Local pref:String
+									If decl.IsMethod() Then
+										pref = "_"
+									End If
 									If Not isPointerType(lhs.exprType) Then
-										Return "_" + decl.munged+TransArgs( args,decl, "&" + TransSubExpr( lhs ) )
+										Return pref + decl.munged+TransArgs( args,decl, "&" + TransSubExpr( lhs ) )
 									Else
-										Return "_" + decl.munged+TransArgs( args,decl, TransSubExpr( lhs ) )
+										Return pref + decl.munged+TransArgs( args,decl, TransSubExpr( lhs ) )
 									End If
 								Else
 									If cdecl Then
@@ -1122,7 +1200,11 @@ t:+"NULLNULLNULL"
 					If cdecl.IsStruct() Then
 						' create a local variable of the inner invocation
 						Local lvar:String = CreateLocal(lhs)
-						Return "_" + decl.munged+TransArgs( args,decl, "&" + lvar )
+						Local t:String
+						If decl.IsMethod() Then
+							t = "_"
+						End If
+						Return t + decl.munged+TransArgs( args,decl, "&" + lvar )
 					Else
 						If decl.IsMethod() Then
 							Local class:String = cdecl.munged
@@ -1189,11 +1271,14 @@ t:+"NULLNULLNULL"
 						Else
 							If cdecl.IsStruct() Then
 
-								' baaaaaaaaaaaaaaaaa
+								Local pref:String
+								If decl.IsMethod() Then
+									pref = "_"
+								End If
 								If Not isPointerType(lhs.exprType) Then
-									Return "_" + decl.munged+TransArgs( args,decl, "&" + TransSubExpr( lhs ) )
+									Return pref + decl.munged+TransArgs( args,decl, "&" + TransSubExpr( lhs ) )
 								Else
-									Return "_" + decl.munged+TransArgs( args,decl, TransSubExpr( lhs ) )
+									Return pref + decl.munged+TransArgs( args,decl, TransSubExpr( lhs ) )
 								End If
 							
 							Else
@@ -1237,10 +1322,14 @@ t:+"NULLNULLNULL"
 						' create a local variable of the inner invocation
 						Local lvar:String = CreateLocal(lhs, True)
 
+						Local pref:String
+						If decl.IsMethod() Then
+							pref = "_"
+						End If
 						If Not isPointerType(lhs.exprType) Then
-							Return "_" + decl.munged+TransArgs( args,decl, "&" + lvar )
+							Return pref + decl.munged+TransArgs( args,decl, "&" + lvar )
 						Else
-							Return "_" + decl.munged+TransArgs( args,decl, lvar)
+							Return pref + decl.munged+TransArgs( args,decl, lvar)
 						End If
 					Else
 						' create a local variable of the inner invocation
@@ -1326,10 +1415,14 @@ t:+"NULLNULLNULL"
 					
 						Local lvar:String = CreateLocal(lhs, True, False)
 					
+						Local pref:String
+						If decl.IsMethod() Then
+							pref = "_"
+						End If
 						If Not isPointerType(lhs.exprType) Then
-							Return "_" + decl.munged+TransArgs( args,decl, "&" + lvar )
+							Return pref + decl.munged+TransArgs( args,decl, "&" + lvar )
 						Else
-							Return "_" + decl.munged+TransArgs( args,decl, lvar )
+							Return pref + decl.munged+TransArgs( args,decl, lvar )
 						End If
 					Else
 						Local lvar:String = CreateLocal(lhs, False, False)
@@ -1631,6 +1724,18 @@ t:+"NULLNULLNULL"
 	Method TransStackAllocExpr:String(expr:TStackAllocExpr)
 		Return "bbStackAlloc" + Bra(expr.expr.Trans())
 	End Method
+	
+	Method TransFieldOffsetExpr:String(expr:TFieldOffsetExpr)
+		Local t:String = "offsetof("
+		
+		Local cdecl:TClassDecl = TIdentTypeExpr(expr.typeExpr).cdecl
+		t :+ "struct " + cdecl.munged
+		If Not cdecl.IsStruct() Then
+			t :+ "_obj"
+		End If
+		
+		Return t + ", " + TVarExpr(expr.fieldExpr).decl.munged + ")"
+	End Method
 
 	'***** Expressions *****
 
@@ -1793,7 +1898,7 @@ t:+"NULLNULLNULL"
 						If TObjectType(dst) Then
 							Return Bra("&" + t)
 						Else
-							Return Bra(Bra("(BBBYTE*)" + Bra("&" + t)) + "+" + Bra("sizeof(void*)"))
+							Return Bra("bbObjectToFieldOffset" + Bra("&" + t))
 						End If
 					End If
 				End If
@@ -1806,7 +1911,11 @@ t:+"NULLNULLNULL"
 		Else If IsPointerType( dst, 0, TType.T_POINTER | TType.T_CHAR_PTR | TType.T_SHORT_PTR )
 
 			If TArrayType(src) Then
-				Return Bra(Bra(TransType(dst, "")) + "BBARRAYDATA(" + t + ",1)")
+				If TArrayType(src).isStatic Then
+					Return Bra("&" + Bra(t))
+				Else
+					Return Bra(Bra(TransType(dst, "")) + "BBARRAYDATA(" + t + ",1)")
+				End If
 			End If
 			'If TByteType(src) And Not IsPointerType(src, TType.T_BYTE, TType.T_POINTER) Return Bra("&"+t)
 
@@ -1831,7 +1940,7 @@ t:+"NULLNULLNULL"
 					Return Bra(t)
 				Else
 					If Not TObjectType(src).classDecl.IsStruct() Then
-						Return Bra(Bra("(BBBYTE*)" + t) + "+" + Bra("sizeof(void*)"))
+						Return Bra("bbObjectToFieldOffset" + Bra(t))
 					Else
 						Return Bra("(BBBYTE*)" + t)
 					End If
@@ -2195,6 +2304,9 @@ t:+"NULLNULLNULL"
 			'If TStringType( src ) Return Bra("(BBOBJECT)"+t)
 			'If TObjectType( src ) Return t
 			If Not TObjectType( dst ).classDecl.IsExtern() Then
+				If TObjectType( dst ).classDecl.IsStruct() Then
+					Return TransValue(dst, Null)
+				End If
 				If TNullType( src ) Return "&bbNullObject"
 				If TObjectType(dst).classDecl.IsInterface() Then
 					Return Bra(Bra(TransObject(TObjectType(dst).classDecl)) + "bbInterfaceDowncast" + Bra(t + ",&" + TObjectType(dst).classDecl.munged + "_ifc"))
@@ -2413,10 +2525,14 @@ t:+"NULLNULLNULL"
 					Return Bra(Bra(TransType(TArrayType( expr.expr.exprType).elemType, "*")) + Bra("BBARRAYDATA(" + t_expr + ",1)")) + "[" + t_index + "]"
 				End If
 			Else
-				If opt_debug Then
-					Return Bra("(" + TransType(expr.exprType, "") + "*)BBARRAYDATAINDEX(" + Bra(t_expr) + "," + Bra(t_expr) + "->dims," + t_index + ")") + "[" + t_index + "]"
+				If TArrayType( expr.expr.exprType ).isStatic Then
+					Return t_expr + "[" + t_index + "]"
 				Else
-					Return Bra("(" + TransType(expr.exprType, "") + "*)BBARRAYDATA(" + t_expr + ",1)") + "[" + t_index + "]"
+					If opt_debug Then
+						Return Bra("(" + TransType(expr.exprType, "") + "*)BBARRAYDATAINDEX(" + Bra(t_expr) + "," + Bra(t_expr) + "->dims," + t_index + ")") + "[" + t_index + "]"
+					Else
+						Return Bra("(" + TransType(expr.exprType, "") + "*)BBARRAYDATA(" + t_expr + ",1)") + "[" + t_index + "]"
+					End If
 				End If
 			End If
 		End If
@@ -2451,7 +2567,13 @@ t:+"NULLNULLNULL"
 		End If
 
 		If TArrayType(expr.exprType) Then
-			Return "bbArraySlice" + Bra(TransArrayType(TArrayType(expr.exprType).elemType) + "," + t_expr + "," + t_args)
+			Local ty:TType = TArrayType(expr.exprType).elemType
+			If TObjectType(ty) And TObjectType(ty).classDecl.IsStruct() Then
+				Return "bbArraySliceStruct" + Bra(TransArrayType(ty) + "," + t_expr + "," + t_args + ", sizeof" + ..
+						Bra(TransObject(TObjectType(ty).classdecl)) + ", _" + TObjectType(ty).classdecl.munged + "_New")
+			Else
+				Return "bbArraySlice" + Bra(TransArrayType(ty) + "," + t_expr + "," + t_args)
+			End If
 		'Else If TStringVarPtrType(expr.exprType) Then
 		'	Return "bbStringSlice" + Bra("*" + t_expr + "," + t_args)
 		Else
@@ -2718,9 +2840,6 @@ t:+"NULLNULLNULL"
 			If TLocalDecl(decl) Then
 				count :+ 1
 			End If
-			If TConstDecl(decl) Then
-				count :+ 1
-			End If
 			If TGlobalDecl(decl) Then
 				count :+ 1
 			End If
@@ -2731,17 +2850,6 @@ t:+"NULLNULLNULL"
 			count :+ 1
 		End If
 		
-		If _app.mainFunc = block Then
-			For Local decl:TDecl = EachIn _app.mainModule.Decls()
-				If TConstDecl(decl) Then
-					count :+ 1
-				End If
-				If TGlobalDecl(decl) Then
-					count :+ 1
-				End If
-			Next
-		End If
-		
 		If Not count Then
 			Emit "struct BBDebugScope __scope = {"
 		Else
@@ -2773,12 +2881,7 @@ t:+"NULLNULLNULL"
 				Emit "},"
 			End If
 			
-			' block consts and globals
-			' consts
-			For Local cdecl:TConstDecl = EachIn block.Decls()
-				EmitConstDebugScope(cdecl)
-			Next
-			' globals
+			' block globals
 			For Local gdecl:TGlobalDecl = EachIn block.Decls()
 				EmitGlobalDebugScope(gdecl)
 			Next
@@ -2802,18 +2905,6 @@ t:+"NULLNULLNULL"
 				End If
 			Next
 
-			' add module consts and globals
-			If _app.mainFunc = block Then
-				' consts
-				For Local cdecl:TConstDecl = EachIn _app.mainModule.Decls()
-					EmitConstDebugScope(cdecl)
-				Next
-				' globals
-				For Local gdecl:TGlobalDecl = EachIn _app.mainModule.Decls()
-					EmitGlobalDebugScope(gdecl)
-				Next
-			End If
-			
 			Emit "BBDEBUGDECL_END "
 			Emit "}"
 			
@@ -2841,7 +2932,7 @@ t:+"NULLNULLNULL"
 		Local s:String
 		Local cast:String
 		
-		If TObjectType(stmt.lhs.exprType) Then
+		If TObjectType(stmt.lhs.exprType) And (Not TObjectType(stmt.lhs.exprType).classdecl.IsStruct() Or IsPointerType(stmt.lhs.exprType)) Then
 			cast = Bra(TransType(stmt.lhs.exprType, ""))
 		End If
 
@@ -3144,6 +3235,9 @@ End Rem
 			If Not TFunctionPtrType(oarg.ty) Then
 				If Not odecl.castTo Then
 					args:+TransType( oarg.ty, arg.munged )
+					If TArrayType(oarg.ty) And TArrayType(oarg.ty).isStatic Then
+						args :+ "[" + TArrayType(oarg.ty).length + "]"
+					End If
 				Else
 					args:+ oarg.castTo + " " + arg.munged
 				End If
@@ -3303,6 +3397,9 @@ End Rem
 			If Not TFunctionPtrType(oarg.ty) Then
 				If Not odecl.castTo Then
 					args:+TransType( oarg.ty, arg.munged )+" "+arg.munged
+					If TArrayType(oarg.ty) And TArrayType(oarg.ty).isStatic Then
+						args :+ "[" + TArrayType(oarg.ty).length + "]"
+					End If
 				Else
 					args:+ oarg.castTo + " " + arg.munged
 				End If
@@ -3477,7 +3574,13 @@ End Rem
 				If classDecl.IsExtern() Then
 					Emit TransType(decl.ty, "") + " " + decl.ident + ";"
 				Else
-					Emit TransType(decl.ty, classDecl.actual.munged) + " _" + classDecl.actual.munged.ToLower() + "_" + decl.IdentLower() + ";"
+					Local t:String = TransType(decl.ty, classDecl.actual.munged) + " _" + classDecl.actual.munged.ToLower() + "_" + decl.IdentLower()
+					
+					If TArrayType(decl.ty) And TArrayType(decl.ty).isStatic Then
+						t :+ "[" + TArrayType(decl.ty).length + "]"
+					End If
+					
+					Emit t + ";"
 				End If
 			Else
 				If classDecl.IsExtern() Then
@@ -3496,7 +3599,7 @@ End Rem
 			decl.Semant()
 
 			If TFunctionPtrType(decl.ty) Then
-					Emit TransRefType( decl.ty, decl.munged ) + ";"
+					Emit "extern " + TransRefType( decl.ty, decl.munged ) + ";"
 			Else
 				Emit "extern "+TransRefType( decl.ty, "" )+" "+ decl.munged+";"
 			End If
@@ -3616,6 +3719,7 @@ End Rem
 		Emit "void*     extra;"
 		Emit "unsigned int obj_size;"
 		Emit "unsigned int instance_count;"
+		Emit "unsigned int fields_offset;"
 
 		EmitBBClassClassFuncProto(classDecl)
 
@@ -4448,6 +4552,8 @@ End Rem
 			Emit TransObjectSize(classDecl)
 			' instance_count
 			Emit ",0"
+			' fields_offset
+			Emit TransFirstFieldOffset(classDecl)
 	
 	
 			' methods/funcs
@@ -4618,21 +4724,47 @@ End Rem
 	Method TransObjectSize:String(classDecl:TClassDecl)
 		Local t:String
 		
-		Local fieldDecl:TFieldDecl
+		Local firstDecl:TFieldDecl
+		Local lastDecl:TFieldDecl
 
 		For Local decl:TFieldDecl = EachIn classDecl.Decls()
-			fieldDecl = decl
+			If Not firstDecl Then
+				firstDecl = decl
+			End If
+			lastDecl = decl
 		Next
 		
-		If fieldDecl Then
-			t = "offsetof" + Bra("struct " + classDecl.munged + "_obj," + fieldDecl.munged) + " - sizeof(void*) + sizeof" + Bra(TransType(fieldDecl.ty, ""))
+		If firstDecl Then
+			If firstDecl <> lastDecl Then
+				t = "offsetof" + Bra("struct " + classDecl.munged + "_obj," + lastDecl.munged) + " - offsetof" + Bra("struct " + classDecl.munged + "_obj," + firstDecl.munged) + " + sizeof" + Bra(TransType(lastDecl.ty, ""))
+			Else
+				t = "sizeof" + Bra(TransType(lastDecl.ty, ""))
+			End If
 		Else
 			t = "0"
 		End If
 
 		Return t
 	End Method
-	
+
+	Method TransFirstFieldOffset:String(classDecl:TClassDecl)
+		Local t:String
+		
+		Local fieldDecl:TFieldDecl
+
+		For Local decl:TFieldDecl = EachIn classDecl.Decls()
+			fieldDecl = decl
+			Exit
+		Next
+
+		If fieldDecl Then
+			t = ",offsetof" + Bra("struct " + classDecl.munged + "_obj," + fieldDecl.munged)
+		Else
+			t = ",sizeof(void*)"
+		End If
+		
+		Return t
+	End Method
 	
 	Method EmitClassDeclNew( classDecl:TClassDecl, fdecl:TFuncDecl )
 		Local classid$=classDecl.munged
@@ -4757,6 +4889,8 @@ End Rem
 							fld :+ "= " + TInvokeExpr(decl.init).decl.munged + ";"
 						Else If TObjectType(decl.ty) Then
 							fld :+ "= " + Bra(TransObject(TObjectType(decl.ty).classDecl)) + decl.init.Trans() + ";"
+						Else If TArrayType(decl.ty) And TArrayType(decl.ty).isStatic Then
+							fld = "for(int i=0;i<" + TArrayType(decl.ty).length + ";i++) " + TransFieldRef(decl, "o") + "[i]=" + TransValue(TArrayType(decl.ty).elemType,Null,False) + ";"
 						Else
 							fld :+ "= " + decl.init.Trans() + ";"
 						End If
@@ -4775,7 +4909,11 @@ End Rem
 					Else If TStringType(decl.ty) Then
 						fld :+ "= &bbEmptyString;"
 					Else If TArrayType(decl.ty) Then
-						fld :+ "= &bbEmptyArray;"
+						If TArrayType(decl.ty).isStatic Then
+							fld = "for(int i=0;i<" + TArrayType(decl.ty).length + ";i++) " + TransFieldRef(decl, "o") + "[i]=" + TransValue(TArrayType(decl.ty).elemType,Null,False) + ";"
+						Else
+							fld :+ "= &bbEmptyArray;"
+						End If
 					Else If TEnumType(decl.ty) Then
 						fld :+ "= " + TEnumType(decl.ty).decl.values[0].Value() + ";"
 					End If
@@ -4901,7 +5039,16 @@ End Rem
 		t = TransObject(classdecl) + " o"
 
 		If classDecl.IsStruct() Then
-			Emit t + " = {0};"
+			t :+ " = {"
+			Local fields:Int
+			For Local f:TFieldDecl = EachIn classDecl.Decls()
+				If fields Then
+					t :+ ","
+				End If
+				fields = True
+				t :+ TransValue(f.ty, "", True)
+			Next
+			Emit t + "};"
 		Else
 			t :+ " = "
 			If ClassHasObjectField(classDecl) Then
@@ -5117,7 +5264,11 @@ End Rem
 		' array.length
 		If decl.scope And decl.scope.ident = "___Array" Then
 			If decl.ident = "length" Then
-				Return Bra(variable + "->scales[0]")
+				If TArrayType(exprType) And TArrayType(exprType).isStatic Then
+					Return TArrayType(exprType).length
+				Else
+					Return Bra(variable + "->scales[0]")
+				End If
 			End If
 			If decl.ident = "numberOfDimensions" Then
 				Return Bra(variable + "->dims")
@@ -5363,7 +5514,11 @@ End Rem
 		EndIf
 
 		If TEnumType(expr.exprType) Then
-			Return Expr.Eval()
+			If TCastExpr(expr) And TNullExpr(TCastExpr(expr).expr) Then
+				Return TransValue(expr.exprType, Null)
+			Else
+				Return Expr.Eval()
+			End If
 		End If
 
 		'If TObjectType(expr.exprType) And TNullDecl(TObjectType(expr.exprType).classDecl) Then
@@ -5384,10 +5539,17 @@ End Rem
 	End Method
 
 	Method EmitIfcFieldDecl(fieldDecl:TFieldDecl)
+	
 		Local f:String
 		If fieldDecl.IsReadOnly() Then
 			f :+ "@"
-		Else
+		End If
+
+		If fieldDecl.IsStatic() Then
+			f :+ "~~"
+		End If
+		
+		If Not f Then
 			f :+ "."
 		End If
 		f :+ fieldDecl.ident + TransIfcType(fieldDecl.ty, fieldDecl.ModuleScope().IsSuperStrict())
@@ -5496,6 +5658,10 @@ End Rem
 				flags :+ "S"
 			End If
 			
+			If classDecl.IsPrivate() Then
+				flags :+ "P"
+			End If
+			
 			If classDecl.templateSource Then
 				flags :+ "G"
 			End If
@@ -5651,49 +5817,57 @@ End Rem
 
 		If FileType(ib.path) = FILETYPE_FILE Then
 
-			Local ident:String = _appInstance.munged + "_ib_" + ib.id
+			If Not opt_legacy_incbin Then
 
-			Local buf:Byte[] = LoadByteArray(ib.path)
-			ib.length = buf.length
+				Local ident:String = _appInstance.munged + "_" + ib.id
 
-			Emit "unsigned char " + ident + "[] = {"
-			Local sb:TStringBuffer = New TStringBuffer
+				Emit "INCBIN(" + ident + ", ~q" + ib.path + "~q);"
 
-			Local hx:Short[2]
-			Local lines:Int
-			Local count:Int
-			For Local i:Int = 0 Until buf.length
-				Local val:Int = buf[i]
+			Else
 
-				For Local k:Int=1 To 0 Step -1
-					Local n:Int=(val&15)+48
-					If n>57 n:+39
-					hx[k]=n
-					val:Shr 4
-				Next
-				sb.Append("0x").AppendShorts( hx,2 )
+				Local ident:String = _appInstance.munged + "_ib_" + ib.id
 
-				sb.Append(",")
-				
-				count :+ 5
+				Local buf:Byte[] = LoadByteArray(ib.path)
+				ib.length = buf.length
 
-				If count > 80 Then
-					sb.Append("~n")
-					count = 0
-					lines :+ 1
-				End If
-				
-				If lines > 100 Then
-					Emit sb.ToString()
-					sb.SetLength(0)
-					lines = 0
-				End If
-				
-			Next
+				Emit "unsigned char " + ident + "[] = {"
+				Local sb:TStringBuffer = New TStringBuffer
 
-			Emit sb.ToString()
-			Emit "};"
+				Local hx:Short[2]
+				Local LINES:Int
+				Local count:Int
+				For Local i:Int = 0 Until buf.length
+					Local val:Int = buf[i]
+
+					For Local k:Int=1 To 0 Step -1
+						Local n:Int=(val&15)+48
+						If n>57 n:+39
+						hx[k]=n
+						val:Shr 4
+					Next
+					sb.Append("0x").AppendShorts( hx,2 )
 
+					sb.Append(",")
+					
+					count :+ 5
+
+					If count > 80 Then
+						sb.Append("~n")
+						count = 0
+						LINES :+ 1
+					End If
+					
+					If LINES > 100 Then
+						Emit sb.ToString()
+						sb.SetLength(0)
+						LINES = 0
+					End If
+					
+				Next
+
+				Emit sb.ToString()
+				Emit "};"
+			End If
 		End If
 
 	End Method
@@ -5798,18 +5972,18 @@ End Rem
 			If gdecl
 				MungDecl gdecl
 				
-				If Not TFunctionPtrType(gdecl.ty) Then
 If Not gdecl.IsPrivate() Then
+				If Not TFunctionPtrType(gdecl.ty) Then
 					Emit "extern "+TransRefType( gdecl.ty, "" )+" "+gdecl.munged+";"	'forward reference...
-End If
 				Else
 					If Not TFunctionPtrType(gdecl.ty).func.noCastGen Then
 						' generate function pointer refs if we haven't been told not to
 '						If Not gdecl.IsExtern() Then
-							Emit TransRefType( gdecl.ty, gdecl.munged )+";"	'forward reference...
+							Emit "extern " + TransRefType( gdecl.ty, gdecl.munged )+";"	'forward reference...
 '						End If
 					End If
 				End If
+End If
 				Continue
 			EndIf
 
@@ -5861,6 +6035,7 @@ End If
 		' read lines until "// ----"
 		' TODO
 		Local files:TList = New TList
+		Local hashes:TMap = New TMap
 		Local stream:TStream = ReadFile(file)
 		While True
 			Local s:String = ReadLine(stream)
@@ -5870,7 +6045,17 @@ End If
 
 			Local ind:Int = s.Find("// FILE : ")
 			If ind = 0 Then
-				files.AddLast(s[10..].Replace("~q",""))
+				Local line:String = s[10..]
+				
+				Local parts:String[] = line.Split("~t")
+				If parts.length = 1 Then
+					Return True
+				End If
+				
+				Local filename:String = parts[0].Replace("~q","")
+				Local fileHash:String = parts[1]
+				files.AddLast(filename)
+				hashes.Insert(filename, fileHash)
 			End If
 		Wend
 		stream.Close()
@@ -5913,6 +6098,15 @@ End If
 			If timestamp < FileTime(ib.path) Then
 				Return True
 			End If
+			
+			Local fileHash:String = String(hashes.ValueForKey(ib.file))
+			If Not fileHash Then
+				Return True
+			End If
+			
+			If fileHash <> CalculateFileHash(ib.path) Then
+				Return True
+			End If
 
 			' set the length, as we will need this later if we aren't loading the files now.
 			ib.length = FileSize(ib.path)
@@ -5929,15 +6123,27 @@ End If
 			Local mung:String = FileMung(False)
 
 			Local name:String = StripAll(app.mainModule.filepath)
-			Local file:String = name + ".bmx" + mung + ".incbin.c"
-			Local filepath:String = OutputFilePath(opt_filepath, mung, "incbin.c")
+			Local file:String
+			If opt_legacy_incbin Then
+				file = "incbin.c"
+			Else
+				file = "incbin2.c"
+			End If
+			Local filepath:String = OutputFilePath(opt_filepath, mung, file)
 
 			If IncBinRequiresRebuild(filepath, app.incbins) Then
 
+				If Not opt_legacy_incbin Then
+					Emit "#define INCBIN_PREFIX _ib"
+					Emit "#define INCBIN_STYLE INCBIN_STYLE_SNAKE"
+					Emit "#include ~qbrl.mod/blitz.mod/incbin/incbin.h~q"
+				End If
+
 				app.genIncBinHeader = True
 
 				For Local ib:TIncbin = EachIn app.incbins
-					Emit "// FILE : " + Enquote(ib.file)
+					Local fileHash:String = CalculateFileHash(ib.path)
+					Emit "// FILE : " + Enquote(ib.file) + "~t" + fileHash
 				Next
 
 				Emit "// ----"
@@ -6022,7 +6228,9 @@ Else
 End If
 					End If
 				Else
-					'Emit TransRefType( gdecl.ty, gdecl.munged ) + ";"
+					If Not gdecl.IsExtern() Then
+						Emit TransRefType( gdecl.ty, gdecl.munged ) + ";"
+					EndIf
 				End If
 				Continue
 			EndIf
@@ -6050,6 +6258,8 @@ End If
 
 			Local gdecl:TGlobalDecl=TGlobalDecl( decl )
 			If gdecl
+				If gdecl.IsPrivate() Continue
+				
 				If Not TFunctionPtrType(gdecl.ty) And Not gdecl.IsPrivate() Then
 					If TConstExpr(gdecl.init) Then
 						Emit TransRefType( gdecl.ty, "WW" )+" "+TransGlobalDecl(gdecl)+";"
@@ -6064,7 +6274,9 @@ If Not gdecl.IsExtern() Then
 End If
 					End If
 				Else
-					'Emit TransRefType( gdecl.ty, gdecl.munged ) + ";"
+					If TFunctionPtrType(gdecl.ty) And Not gdecl.IsExtern() Then
+						Emit TransRefType( gdecl.ty, gdecl.munged ) + ";"
+					End If
 				End If
 				Continue
 			EndIf
@@ -6109,7 +6321,12 @@ End If
 
 		' incbin decls
 		For Local ib:TIncbin = EachIn app.incbins
-			Emit "extern unsigned char * " + app.munged + "_ib_" + ib.id + ";"
+			If opt_legacy_incbin Then
+				Emit "extern unsigned char * " + app.munged + "_ib_" + ib.id + ";"
+			Else
+				Emit "extern const unsigned char * " + ib.GeneratedDataName(app) + ";"
+				Emit "extern const unsigned int " + ib.GeneratedSizeName(app) + ";"
+			End If
 		Next
 
 		Emit "static int " + app.munged + "_inited" + " = 0;"
@@ -6143,7 +6360,11 @@ End If
 
 		' register incbins
 		For Local ib:TIncbin = EachIn app.incbins
-			Emit "bbIncbinAdd(&" + TStringConst(app.stringConsts.ValueForKey(ib.file)).id + ",&" + app.munged + "_ib_" + ib.id + "," + ib.length + ");"
+			If opt_legacy_incbin Then
+				Emit "bbIncbinAdd(&" + TStringConst(app.stringConsts.ValueForKey(ib.file)).id + ",&" + app.munged + "_ib_" + ib.id + "," + ib.length + ");"
+			Else
+				Emit "bbIncbinAdd(&" + TStringConst(app.stringConsts.ValueForKey(ib.file)).id + ",&" + ib.GeneratedDataName(app) + "," + ib.GeneratedSizeName(app) + ");"
+			End If
 		Next
 		
 		Local importOnce:TMap = New TMap
@@ -6192,6 +6413,14 @@ End If
 			End If
 		Next
 		'
+		
+		' register files
+		If opt_debug Then
+			For Local Hash:String = EachIn fileRegister.Keys()
+				Local file:String = String(fileRegister.ValueForKey(Hash))
+				Emit "bbRegisterSource(" + Hash + ", ~q" + file + "~q);"
+			Next
+		End If
 
 		' defdata init
 		If Not app.dataDefs.IsEmpty() Then
@@ -6268,7 +6497,7 @@ End If
 				If key.count > 0 Then
 					Emit "static BBString " + key.id + "={"
 					Emit "&bbStringClass,"
-					'Emit "2147483647,"
+					Emit bmx_gen_hash(s) + ","
 					Emit s.length + ","
 
 					Local t:String = "{"
@@ -6520,9 +6749,13 @@ End If
 
 		' classes
 		For Local decl:TDecl=EachIn app.Semanted()
-			If decl.IsPrivate() Continue
 
 			Local cdecl:TClassDecl=TClassDecl( decl )
+			
+			If cdecl And cdecl.IsPrivate() And (Not cdecl.IsStruct() Or (cdecl.IsStruct() And Not cdecl.exposed)) Then
+				Continue
+			End If
+			
 			If cdecl And Not cdecl.declImported
 				EmitIfcClassDecl(cdecl)
 			EndIf

decl.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -27,6 +27,7 @@ Const DECL_PRIVATE:Int=       $020000
 Const DECL_ABSTRACT:Int=      $040000
 Const DECL_FINAL:Int=         $080000
 Const DECL_READ_ONLY:Int=     $000100
+Const DECL_STATIC:Int=      $20000000
 Const DECL_OVERRIDE:Int=    $40000000
 Const DECL_INLINE:Int=      $80000000
 
@@ -107,6 +108,13 @@ Type TFuncDeclList Extends TList
 		End If
 		Return _identLower
 	End Method
+
+	Method AddLast:TLink( value:Object )
+		If Not Contains(value) Then
+			Return Super.AddLast(value)
+		End If
+	End Method
+
 End Type
 
 Type TMetadata
@@ -193,6 +201,10 @@ Type TDecl
 	Method IsAbstract:Int()
 		Return (attrs & DECL_ABSTRACT)<>0
 	End Method
+
+	Method IsStatic:Int()
+		Return (attrs & DECL_STATIC)<>0
+	End Method
 	
 	Method IsSemanted:Int()
 		Return (attrs & DECL_SEMANTED)<>0
@@ -515,19 +527,26 @@ Type TValDecl Extends TDecl
 					
 					
 				Else
-					If TArrayExpr(declInit) And TArrayType(ty) And TNumericType(TArrayType(ty).elemType) Then
-						TArrayExpr(declInit).toType = TArrayType(ty).elemType
-					End If
-				
-					init=declInit.Copy().SemantAndCast(ty)
-					
-					' check if struct has been initialised
-					If TObjectType(ty) And TObjectType(ty).classDecl.IsStruct() And Not TObjectType(ty).classDecl.IsExtern() Then
+					If TArrayType(ty) And TArrayType(ty).isStatic Then
+						init = declInit.Copy().Semant()
+						If Not TArrayType(ty).length Then
+							TArrayType(ty).length = init.Eval()
+						End If
+					Else
+						If TArrayExpr(declInit) And TArrayType(ty) And TNumericType(TArrayType(ty).elemType) Then
+							TArrayExpr(declInit).toType = TArrayType(ty).elemType
+						End If
 					
-						' new not used
-						If TConstExpr(init) And Not TConstExpr(init).value And Not IsPointerType(ty,0,TType.T_POINTER) Then
-							' always call the default constructor to init all the fields correctly
-							init = New TNewObjectExpr.Create(ty, Null).Semant()
+						init=declInit.Copy().SemantAndCast(ty)
+						
+						' check if struct has been initialised
+						If TObjectType(ty) And TObjectType(ty).classDecl.IsStruct() And Not TObjectType(ty).classDecl.IsExtern() Then
+						
+							' new not used
+							If TConstExpr(init) And Not TConstExpr(init).value And Not IsPointerType(ty,0,TType.T_POINTER) Then
+								' always call the default constructor to init all the fields correctly
+								init = New TNewObjectExpr.Create(ty, Null).Semant()
+							End If
 						End If
 					End If
 				End If
@@ -693,6 +712,17 @@ Type TArgDecl Extends TLocalDecl
 		If ty Then
 			ty = ty.Semant()
 		End If
+		
+		If attrs & DECL_STATIC Then
+			If Not TArrayType(ty) Then
+				Err "Expecting array"
+			End If
+			
+			If Not TNumericType(TArrayType(ty).elemType) Then
+				Err "Static array elements must be numeric"
+			End If
+		End If
+		
 		If init And Not TConstExpr(init) Then
 			If TCastExpr(init) Then
 				If TConstExpr(TCastExpr(init).expr) Or TNullExpr(TCastExpr(init).expr) Then
@@ -844,6 +874,14 @@ Type TFieldDecl Extends TVarDecl
 		End If
 		Return True
 	End Method
+	
+	Method OnSemant()
+		Super.OnSemant()
+		
+		If TObjectType(ty) And TObjectType(ty).classDecl.IsStruct() Then
+			TObjectType(ty).classDecl.exposed = True
+		End If
+	End Method
 
 End Type
 
@@ -1569,7 +1607,7 @@ End Rem
 	
 					Else ' for case of argdecls having default args
 						exact=False
-						If Not explicit Exit
+						Continue ' carry on to the next arg
 					EndIf
 				
 					possible=False
@@ -1745,6 +1783,12 @@ Type TBlockDecl Extends TScopeDecl
 		
 		For Local stmt:TStmt=EachIn stmts
 			stmt.Semant
+
+			If opt_debug And Not IsNoDebug() Then
+				If Not stmt.generated Then
+					GenHash(stmt.errInfo[1..].Split(";")[0])
+				End If
+			End If
 			
 			If TReturnStmt(stmt) Then
 				If SurroundingFinallyBlock(Self) Then PushErr stmt.errInfo; Err "Return cannot be used inside a Finally block."
@@ -2022,8 +2066,13 @@ Type TFuncDecl Extends TBlockDecl
 			If retTypeExpr And Not TVoidType(retTypeExpr) Then
 				Err ident + "() cannot specify a return type"
 			End If
-			If ClassScope() And ClassScope().IsInterface() Then
-				Err ident + "() cannot be declared in an Interface."
+			Local sc:TClassDecl = ClassScope()
+			If sc Then
+				If sc.IsInterface() Then
+					Err ident + "() cannot be declared in an Interface."
+				Else If sc.IsStruct() And isDtor() Then
+					Err ident + "() cannot be declared in a Struct."
+				End If
 			End If
 			If IsCtor() retTypeExpr=New TObjectType.Create( TNewDecl(Self).cDecl )
 		End If
@@ -2071,10 +2120,18 @@ Type TFuncDecl Extends TBlockDecl
 '			Err "Return type cannot be an array of generic objects."
 		EndIf
 
+		If ClassScope() And TObjectType(retType) And TObjectType(retType).classDecl.IsStruct() And TObjectType(retType).classDecl.IsPrivate() Then
+			TObjectType(retType).classDecl.exposed = True
+		End If
+
 		'semant args
 		For Local arg:TArgDecl=EachIn argDecls
 			InsertDecl arg
 			arg.Semant
+
+			If ClassScope() And TObjectType(arg.ty) And TObjectType(arg.ty).classDecl.IsStruct() And TObjectType(arg.ty).classDecl.IsPrivate() Then
+				TObjectType(arg.ty).classDecl.exposed = True
+			End If
 		Next
 
 		' if we are a function pointer declaration, we just want to semant the args here.
@@ -2403,6 +2460,8 @@ Type TClassDecl Extends TScopeDecl
 	Field objectType:TObjectType '"canned" objectType
 	Field globInit:Int
 	Field templateSource:TTemplateRecord
+	
+	Field exposed:Int
 
 	'Global nullObjectClass:TClassDecl=New TNullDecl.Create( "{NULL}",Null,Null,Null,DECL_ABSTRACT|DECL_EXTERN )
 	
@@ -3079,6 +3138,7 @@ End Rem
 			
 			Local arg:TArgDecl = New TArgDecl.Create("o1", TType.MapToVarType(New TObjectType.Create(Self)), Null)
 			func = New TFuncDecl.CreateF("Compare", New TIntType, [arg], FUNC_METHOD)
+			func.retType = New TIntType
 			BuildStructCompareStatements(func)
 			
 			Local found:Int
@@ -3095,7 +3155,7 @@ End Rem
 			End If
 			
 			' generate default comparator compare
-			BuildStructDefaultComparatorCompare()
+			BuildStructDefaultComparatorCompare(attrs & DECL_PRIVATE <> 0)
 			
 			attrs :~ DECL_CYCLIC
 		End If
@@ -3119,6 +3179,7 @@ End Rem
 		
 		' iterate fields
 		For Local fdecl:TFieldDecl = EachIn _decls
+			fdecl.Semant()
 			
 			'
 			' If cmp <> 0 Then
@@ -3141,6 +3202,11 @@ End Rem
 			Local expr2:TExpr = New TIdentExpr.Create( "o1")
 			expr2 = New TIdentExpr.Create( fdecl.ident, expr2)
 			
+			If TEnumType(fdecl.ty) Then
+				expr1 = New TFuncCallExpr.Create(New TIdentExpr.Create("Ordinal", expr1))
+				expr2 = New TFuncCallExpr.Create(New TIdentExpr.Create("Ordinal", expr2))
+			End If
+			
 			Local fcExpr:TExpr = New TIdentExpr.Create( "DefaultComparator_Compare")
 			fcExpr = New TFuncCallExpr.Create( fcExpr, [expr1, expr2])
 	
@@ -3157,10 +3223,13 @@ End Rem
 		func.stmts.AddLast returnStmt
 	End Method
 
-	Method BuildStructDefaultComparatorCompare()
+	Method BuildStructDefaultComparatorCompare(isPrivate:Int = False)
 		Local arg1:TArgDecl = New TArgDecl.Create("o1", TType.MapToVarType(New TObjectType.Create(Self)), Null)
 		Local arg2:TArgDecl = New TArgDecl.Create("o2", TType.MapToVarType(New TObjectType.Create(Self)), Null)
 		Local func:TFuncDecl = New TFuncDecl.CreateF("DefaultComparator_Compare", New TIntType, [arg1, arg2], 0)
+		If isPrivate Then
+			func.attrs :| DECL_PRIVATE
+		End If
 
 		Local expr:TExpr = New TIdentExpr.Create( "o1")
 		expr = New TIdentExpr.Create( "Compare" ,expr )
@@ -4157,7 +4226,9 @@ Type TAppDecl Extends TScopeDecl
 'DebugStop		
 		_env=Null
 		pushenv Self
-		
+	
+		SemantImports()
+
 		SemantDataDefs()	
 
 		mainModule.Semant
@@ -4183,6 +4254,17 @@ Type TAppDecl Extends TScopeDecl
 		Next
 	End Method
 	
+	Method SemantImports()
+		For Local decl:TModuleDecl = EachIn globalImports.Values()
+			For Local cdecl:TClassDecl = EachIn decl._decls
+				If Not cdecl.IsSemanted() Then
+					cdecl.Semant()
+					cdecl.SemantParts()
+				End If
+			Next
+		Next
+	End Method
+	
 	Method SemantDataDefs()
 		TDefDataDecl.count = 0
 		

expr.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -153,7 +153,7 @@ Type TExpr
 				If funcDecl.argDecls[i].ty._flags & TType.T_VAR Then
 
 					If TConstExpr(argExpr) Or TBinaryExpr(argExpr) Or (TIndexExpr(argExpr) And TStringType(TIndexExpr(argExpr).expr.exprType)) Or ..
-							TInvokeExpr(argExpr) Or TInvokeMemberExpr(argExpr) Then
+							TInvokeExpr(argExpr) Or TInvokeMemberExpr(argExpr) Or TSelfExpr(argExpr) Then
 						Err "Expression for 'Var' parameter must be a variable or an element of an array or pointer"
 					End If
 
@@ -1327,7 +1327,7 @@ Type TCastExpr Extends TExpr
 
 			EndIf
 			
-			If TNullType(src) And Not TVoidType(ty) Then
+			If TNullType(src) And Not TVoidType(ty) And Not (TArrayType(ty) And TArrayType(ty).IsStatic) Then
 				exprType = ty
 			End If
 			
@@ -1546,6 +1546,8 @@ Type TCastExpr Extends TExpr
 				End If
 			Else If TNumericType(src) And (src._flags & TType.T_VARPTR) Then
 				exprType = expr.exprType
+			Else If TObjectType(src) And (src._flags & TType.T_VARPTR) Then
+				exprType = expr.exprType
 			Else If TArrayType(src) Then
 			
 				' for functions and index access, use a new local variable
@@ -1907,11 +1909,7 @@ Type TBinaryMathExpr Extends TBinaryExpr
 				exprType=New TIntType
 			End If
 		Case "^"
-			If TIntegralType(lhs.exprType) And TIntegralType(rhs.exprType) Then
-				exprType=New TLongType
-			Else
-				exprType=New TDoubleType
-			End If
+			exprType=New TDoubleType
 		Default
 			exprType=BalanceTypes( lhs.exprType,rhs.exprType )
 			If TStringType( exprType )
@@ -1946,6 +1944,10 @@ Type TBinaryMathExpr Extends TBinaryExpr
 		End If
 
 		If TConstExpr( lhs ) And TConstExpr( rhs ) Return EvalConst()
+		
+		If TConstExpr( rhs ) And (op = "/" Or op = "mod") And TIntegralType(rhs.exprType) And Not Long(rhs.Eval()) Then
+			Err "Integer division by zero"
+		End If
 
 		Return Self
 	End Method
@@ -1956,7 +1958,7 @@ Type TBinaryMathExpr Extends TBinaryExpr
 		If TIntType( exprType ) Or TByteType( exprType ) Or TShortType( exprType )
 			Local x:Int=Int(lhs),y:Int=Int(rhs)
 			Select op
-			Case "^" Return x^y
+			Case "^" Return Double(lhs)^Double(rhs)
 			Case "*" Return x*y
 			Case "/" 
 				If Not y Then
@@ -1980,7 +1982,7 @@ Type TBinaryMathExpr Extends TBinaryExpr
 		Else If TLongType( exprType ) Or TSizeTType(exprType) Or TUIntType(exprType) Or TULongType(exprType) Or TInt128Type(exprType) Or TWParamType(exprType) Or TLParamType(exprType) 
 			Local x:Long=Long(lhs),y:Long=Long(rhs)
 			Select op
-			Case "^" Return x^y
+			Case "^" Return Double(lhs)^Double(rhs)
 			Case "*" Return x*y
 			Case "/"
 				If Not y Then
@@ -2002,14 +2004,14 @@ Type TBinaryMathExpr Extends TBinaryExpr
 			Case "|" Return x | y
 			End Select
 		Else If TFloatType( exprType )
-			Local x:Float=Float(lhs),y:Float=Float(rhs)
+			Local x:Double=Double(lhs),y:Double=Double(rhs)
 			Select op
-			Case "^" Return x^y
-			Case "*" Return x * y
-			Case "/" Return x / y
-			Case "mod" Return x Mod y
-			Case "+" Return x + y
-			Case "-" Return x - y
+			Case "^" Return Double(x^y)
+			Case "*" Return Float(x * y)
+			Case "/" Return Float(x / y)
+			Case "mod" Return Float(x Mod y)
+			Case "+" Return Float(x + y)
+			Case "-" Return Float(x - y)
 			End Select
 		Else If TDoubleType( exprType ) Or TFloat128Type(exprType) Or TDouble128Type(exprType) Or TFloat64Type(exprType)
 			Local x:Double=Double(lhs),y:Double=Double(rhs)
@@ -2100,6 +2102,8 @@ Type TBinaryCompareExpr Extends TBinaryExpr
 				End If
 				' otherwise, no overload, continue...
 			End Try
+		Else If (TArrayType(lhs.exprType) And TArrayType(lhs.exprType).isStatic) Or (TArrayType(rhs.exprType) And TArrayType(rhs.exprType).isStatic) Then
+			Err "Static arrays cannot be compared"
 		End If
 
 
@@ -2856,36 +2860,7 @@ Type TIdentExpr Extends TExpr
 	
 				Return New TVarExpr.Create( TVarDecl( vdecl ) ).Semant()
 			EndIf
-	
-			If op And op<>"="
-	
-				Local fdecl:TFuncDecl=scope.FindFuncDecl( IdentLower(),,,,,,SCOPE_ALL )
-				If Not fdecl IdentErr
-	
-				If _env.ModuleScope().IsStrict() And Not fdecl.IsProperty() Err "Identifier '"+ident+"' cannot be used in this way."
-	
-				Local lhs:TExpr
-	
-				If fdecl.IsStatic() Or (scope=_env And Not _env.FuncScope().IsStatic())
-					lhs=New TInvokeExpr.Create( fdecl )
-				Else If expr
-					Local tmp:TLocalDecl=New TLocalDecl.Create( "",Null,expr,, True )
-					lhs=New TInvokeMemberExpr.Create( New TVarExpr.Create( tmp ),fdecl )
-					lhs=New TStmtExpr.Create( New TDeclStmt.Create( tmp ),lhs )
-				Else
-					Return Null
-				EndIf
-	
-				Local bop$=op[..1]
-				Select bop
-				Case "*","/","shl","shr","+","-","&","|","~~"
-					rhs=New TBinaryMathExpr.Create( bop,lhs,rhs )
-				Default
-					InternalErr "TIdentExpr.SemantSet"
-				End Select
-				rhs=rhs.Semant()
-			EndIf
-	
+
 			Local args:TExpr[]
 			If rhs args=[rhs]
 
@@ -3069,10 +3044,12 @@ Type TIdentExpr Extends TExpr
 		'	If cdecl Return args[0].Cast( New TObjectType.Create(cdecl),CAST_EXPLICIT )
 		'EndIf
 
-		Local ty:TType=scope.FindType( IdentLower(),Null )
-		If ty Then
-			If args.Length=1 And args[0] Return args[0].Cast( ty,CAST_EXPLICIT )
-			Err "Illegal number of arguments for type conversion"
+		If Not expr Then
+			Local ty:TType=scope.FindType( IdentLower(),Null )
+			If ty Then
+				If args.Length=1 And args[0] Return args[0].Cast( ty,CAST_EXPLICIT )
+				Err "Illegal number of arguments for type conversion"
+			End If
 		End If
 
 		If throwError Then
@@ -3537,3 +3514,46 @@ Type TStackAllocExpr Extends TBuiltinExpr
 	End Method
 
 End Type
+
+Type TFieldOffsetExpr Extends TBuiltinExpr
+
+	Field typeExpr:TExpr
+	Field fieldExpr:TExpr
+
+	Method Create:TFieldOffsetExpr( typeExpr:TExpr, fieldExpr:TExpr )
+		Self.id="fieldoffset"
+		Self.typeExpr=typeExpr
+		Self.fieldExpr = fieldExpr
+		Return Self
+	End Method
+
+	Method Semant:TExpr(options:Int = 0)
+		If exprType Return Self
+		
+		' validate type and field
+		typeExpr = typeExpr.Semant()
+		
+		If Not TIdentTypeExpr(typeExpr) Then
+			Err "Expecting Type or Struct"
+		End If
+		
+		TIdentExpr(fieldExpr).scope = TIdentTypeExpr(typeExpr).cdecl
+
+		fieldExpr = fieldExpr.Semant()
+		
+		If Not TVarExpr(fieldExpr) Or Not TFieldDecl(TVarExpr(fieldExpr).decl) Then
+			Err "Expecting Field"
+		End If
+		
+		exprType = New TSizeTType
+		Return Self
+	End Method
+
+	Method Copy:TExpr()
+		Return New TFieldOffsetExpr.Create( typeExpr, fieldExpr )
+	End Method
+
+	Method ToString$()
+		Return "TFieldOffsetExpr("+typeExpr.ToString()+"," + fieldExpr.ToString() + ")"
+	End Method
+End Type

hash.c

@@ -0,0 +1,33 @@
+
+#include "brl.mod/blitz.mod/blitz.h"
+
+#ifndef XXHASH_H_5627135585666179
+#define XXH_STATIC_LINKING_ONLY
+#define XXH_IMPLEMENTATION
+
+#include "xxhash.h"
+#endif
+
+BBString * bmx_gen_hash(BBString * txt) {
+	char * buf[64];
+	snprintf(buf, 64, "0x%llx", XXH3_64bits(txt->buf, txt->length * sizeof(BBChar)));
+	return bbStringFromCString(buf);
+}
+
+XXH3_state_t * bmx_hash_createState() {
+	return XXH3_createState();
+}
+
+void bmx_hash_reset(XXH3_state_t * state) {
+	XXH3_64bits_reset(state);
+}
+
+void bmx_hash_update(XXH3_state_t * state, void * data, int length) {
+	XXH3_64bits_update(state, data, length);
+}
+
+BBString * bmx_hash_digest(XXH3_state_t * state) {
+	char * buf[64];
+	snprintf(buf, 64, "%llx", XXH3_64bits_digest(state));
+	return bbStringFromCString(buf);
+}

iparser.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 ' 
@@ -293,6 +293,9 @@ Type TIParser
 								Case Asc("G")
 									class.attrs :| CLASS_GENERIC
 									parsed = True
+								Case Asc("P")
+									class.attrs :| DECL_PRIVATE
+									parsed = True
 							End Select
 						Next
 
@@ -488,6 +491,10 @@ Type TIParser
 						If CParse( "&" ) Then
 						End If
 
+						If IstStaticArrayDef() Then
+							attrs :| DECL_STATIC
+						End If
+
 						While IsArrayDef()
 							ty = ParseArrayType(ty)
 				
@@ -578,6 +585,13 @@ Type TIParser
 			
 		Forever
 		
+		' semant imported classes
+		For Local cdecl:TClassDecl = EachIn _mod.decls
+			cdecl.Semant()
+			If Not cdecl.args Then
+				cdecl.FinalizeClass()
+			End If
+		Next
 		
 		Return True
 		
@@ -733,12 +747,21 @@ Type TIParser
 				'If decl.IsCtor() decl.retTypeExpr=New TObjectType.Create( classDecl )
 				classDecl.InsertDecl decl
 
-			Case ".", "@" ' field
+			Case ".", "@", "~~" ' field			
 				Local d_attrs:Int = decl_attrs | DECL_FIELD
-				If _toker._toke = "@" Then
-					d_attrs :| DECL_READ_ONLY
+				If _toker._toke = "." Then
+					NextToke
+				Else
+					While _toker._toke = "@" Or _toker._toke = "~~"
+						If _toker._toke = "@" Then
+							d_attrs :| DECL_READ_ONLY
+						End If
+						If _toker._toke = "~~" Then
+							d_attrs :| DECL_STATIC
+						End If
+						NextToke
+					Wend
 				End If
-				NextToke
 				Local decl:TDecl= ParseDecl( _toke,d_attrs )
 				classDecl.InsertDecl decl
 			Rem
@@ -1292,8 +1315,12 @@ Type TIParser
 					If CParse( "&" ) Then
 					End If
 
-					While IsArrayDef()
-						ty = ParseArrayType(ty)
+					If IstStaticArrayDef() Then
+						attrs :| DECL_STATIC
+					End If
+
+					While IsArrayDef(attrs & DECL_STATIC > 0)
+						ty = ParseArrayType(ty, attrs & DECL_STATIC > 0)
 			
 						If CParse( "&" ) Then
 						End If
@@ -1413,7 +1440,17 @@ End Rem
 	End Method
 
 	' replaces While CParse( "[]" ) sections, with support for multi-dimension arrays
-	Method ParseArrayType:TType(ty:TType)
+	Method ParseArrayType:TType(ty:TType, isStatic:Int = False)
+		If isStatic Then
+			Parse("[")
+			Local expr:TExpr = ParseUnaryExpr()
+			ty = New TArrayType.Create( ty )
+			TArrayType(ty).isStatic = True
+			TArrayType(ty).length = expr.Eval()
+			Parse("]")
+			Return ty
+		End If
+		
 		While True
 			Local dims:Int = 1
 			
@@ -1438,9 +1475,34 @@ End Rem
 		Return ty
 	End Method
 
-	Method IsArrayDef:Int()
+	Method IstStaticArrayDef:Int()
+		Local toker:TToker=New TToker.Copy(_toker)
+		If Not CParseToker(toker, "[") Then
+			Return False
+		End If
+		If toker.TokeType() <> TOKE_INTLIT Then
+			Return False
+		End If
+		NextTokeToker(toker)
+		If Not CParseToker(toker, "]") Then
+			Return False
+		End If
+		Return True
+	End Method
+
+	Method IsArrayDef:Int(isStatic:Int = False)
 		Local isDef:Int = True
 		Local toker:TToker=New TToker.Copy(_toker)
+		If isStatic Then
+			If Not CParseToker(toker, "[") Then
+				Return False
+			End If
+			NextTokeToker(toker)
+			If Not CParseToker(toker, "]") Then
+				Return False
+			End If
+			Return True
+		End If
 		While True
 			If CParseToker(toker, "[]") Then
 				Exit
@@ -1670,8 +1732,13 @@ End Rem
 		If CParse( "&" ) Then
 		End If
 
-		While IsArrayDef()
-			ty = ParseArrayType(ty)
+		If IstStaticArrayDef() Then
+			attrs :| DECL_STATIC
+		End If
+		
+		While IsArrayDef(attrs & DECL_STATIC > 0)
+
+			ty = ParseArrayType(ty, attrs & DECL_STATIC > 0)
 
 			If CParse( "&" ) Then
 			End If

options.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -25,7 +25,7 @@ SuperStrict
 
 Import "base.configmap.bmx"
 
-Const version:String = "0.118"
+Const version:String = "0.129"
 
 Const BUILDTYPE_APP:Int = 0
 Const BUILDTYPE_MODULE:Int = 1
@@ -62,6 +62,7 @@ Global opt_arch:String
 '    android
 '    raspberrypi
 '    nx
+'    haiku
 Global opt_platform:String
 ' framework
 Global opt_framework:String
@@ -127,6 +128,8 @@ Global opt_override_error:Int = False
 Global opt_userdefs:String
 ' 
 Global opt_need_strict:Int = False
+'
+Global opt_legacy_incbin:Int = False
 
 Global opt_filepath:String
 
@@ -240,6 +243,8 @@ Function ParseArgs:String[](args:String[])
 				opt_userdefs = args[count].ToLower()
 			Case "strict"
 				opt_need_strict=True
+			Case "ib"
+				opt_legacy_incbin=True
 		End Select
 	
 		count:+ 1
@@ -252,6 +257,11 @@ Function ParseArgs:String[](args:String[])
 	If opt_arch = "x64" Or opt_arch = "arm64v8a" Or opt_arch = "arm64" Then
 		WORD_SIZE = 8
 	End If
+	
+	' new incbin doesn't work on win32 x86
+	If opt_arch = "x86" And opt_platform = "win32" Then
+		opt_legacy_incbin = True
+	End If
 
 	If opt_makelib Then
 		If Not opt_nodef Then
@@ -297,6 +307,8 @@ Function DefaultOptions()
 	opt_platform = "android"
 ?raspberrypi
 	opt_platform = "raspberrypi"
+?haiku
+	opt_platform = "haiku"
 ?emscripten
 	opt_platform = "emscripten"
 ?
@@ -326,6 +338,8 @@ Function CheckConfig()
 		If tmp Then
 			osBmxPath = tmp
 		End If
+	?haiku
+		osBmxPath = config.GetString("BMXPATH_HAIKU")
 	?
 	'load default/generic path
 	If osBmxPath = "" Then osBmxPath = config.GetString("BMXPATH")

parser.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -67,19 +67,28 @@ Type TForEachinStmt Extends TLoopStmt
 	End Method
 
 	Method OnSemant()
-
+		Const NotIterableError:String = "EachIn requires a type that implements IIterable or has a suitable ObjectEnumerator method."
+		
 		expr=expr.Semant()
 
 		If TArrayType( expr.exprType ) Or TStringType( expr.exprType )
 
-			Local exprTmp:TLocalDecl=New TLocalDecl.Create( "",Null,expr,,True )
+			Local exprTmp:TLocalDecl
+			Local exprVar:TExpr
+			If TArrayType( expr.exprType ).isStatic And (TVarExpr(expr) Or TMemberVarExpr(expr)) Then ' TODO TSliceExpr
+				exprVar = expr
+			Else
+				exprTmp = New TLocalDecl.Create( "",Null,expr,,True )
+				exprVar = New TVarExpr.Create( exprTmp )
+			End If
+			
 			Local indexTmp:TLocalDecl=New TLocalDecl.Create( "",Null,New TConstExpr.Create( New TUIntType,"0" ),,True )
 
-			Local lenExpr:TExpr=New TIdentExpr.Create( "Length",New TVarExpr.Create( exprTmp ) )
+			Local lenExpr:TExpr=New TIdentExpr.Create( "Length",exprVar )
 
 			Local cmpExpr:TExpr=New TBinaryCompareExpr.Create( "<",New TVarExpr.Create( indexTmp ),lenExpr )
 
-			Local indexExpr:TExpr=New TIndexExpr.Create( New TVarExpr.Create( exprTmp ),[New TVarExpr.Create( indexTmp )] )
+			Local indexExpr:TExpr=New TIndexExpr.Create( exprVar,[New TVarExpr.Create( indexTmp )] )
 			Local addExpr:TExpr=New TBinaryMathExpr.Create( "+",New TVarExpr.Create( indexTmp ),New TConstExpr.Create( New TIntType,"1" ) )
 
 			Local cont:TContinueStmt
@@ -95,11 +104,39 @@ Type TForEachinStmt Extends TLoopStmt
 					Local isStruct:Int = TObjectType(TArrayType( expr.exprType ).elemType).classdecl.IsStruct()
 
 					Local cExpr:TExpr
+					Local varObjTmp:TLocalDecl
+					Local varObjStmt:TStmt
+					
+					If exprTmp Then
+						exprTmp.Semant()
+					End If
+					indexTmp.Semant()
 					
 					If TIdentType(varty) And TIdentType(varty).ident = "Object" Then
 						cExpr = indexExpr
 					Else
-						cExpr = New TCastExpr.Create( varty, indexExpr,CAST_EXPLICIT )
+						If TStringType(varty) Then
+							varObjTmp = New TLocalDecl.Create( "",TType.objectType,indexExpr,,True)
+							varObjTmp.Semant()
+							Local varObjExpr:TExpr=New TVarExpr.Create( varObjTmp )
+							
+							Local expr:TExpr = New TFuncCallExpr.Create( New TIdentExpr.Create( "ObjectIsString"), [varObjExpr])
+							expr=New TBinaryCompareExpr.Create( "=",expr, New TConstExpr.Create( New TIntType,"0" ))
+							
+							Local thenBlock:TBlockDecl=New TBlockDecl.Create( block.scope, True, BLOCK_IF )
+							Local elseBlock:TBlockDecl=New TBlockDecl.Create( block.scope, True, BLOCK_ELSE )
+							cont = New TContinueStmt.Create(Null, True)
+							thenBlock.AddStmt cont
+		
+							varObjStmt = New TIfStmt.Create( expr,thenBlock,elseBlock, True )
+							'block.stmts.AddFirst New TIfStmt.Create( expr,thenBlock,elseBlock, True )
+							
+							cExpr = New TCastExpr.Create( varty, varObjExpr,CAST_EXPLICIT )
+						Else
+							cExpr = New TCastExpr.Create( varty, indexExpr,CAST_EXPLICIT )
+						End If 
+
+						'cExpr = New TCastExpr.Create( varty, indexExpr,CAST_EXPLICIT )
 					End If
 
 					' local variable
@@ -108,7 +145,7 @@ Type TForEachinStmt Extends TLoopStmt
 					' local var as expression
 					Local expr:TExpr=New TVarExpr.Create( varTmp )
 
-					If Not isStruct Then
+					If Not isStruct And Not varObjTmp Then
 						' var = Null
 						expr=New TBinaryCompareExpr.Create( "=",expr, New TNullExpr.Create(TType.nullObjectType))
 	
@@ -122,6 +159,11 @@ Type TForEachinStmt Extends TLoopStmt
 					End If
 					block.stmts.AddFirst New TAssignStmt.Create( "=",New TVarExpr.Create( indexTmp ),addExpr )
 					block.stmts.AddFirst New TDeclStmt.Create( varTmp )
+					If varObjTmp Then
+						block.stmts.AddFirst varObjStmt
+						block.stmts.AddFirst New TDeclStmt.Create( varObjTmp, True )
+					End If
+
 
 				Else
 					Local varTmp:TLocalDecl=New TLocalDecl.Create( varid,varty,indexExpr )
@@ -175,7 +217,9 @@ Type TForEachinStmt Extends TLoopStmt
 			Local whileStmt:TWhileStmt=New TWhileStmt.Create( cmpExpr,block,loopLabel, True )
 
 			block=New TBlockDecl.Create( block.scope, True, BLOCK_LOOP )
-			block.AddStmt New TDeclStmt.Create( exprTmp, True )
+			If exprTmp Then
+				block.AddStmt New TDeclStmt.Create( exprTmp, True )
+			End If
 			block.AddStmt New TDeclStmt.Create( indexTmp, True )
 			block.AddStmt whileStmt
 			
@@ -195,7 +239,7 @@ Type TForEachinStmt Extends TLoopStmt
 			Else
 				Local declList:TFuncDeclList = TFuncDeclList(TObjectType(expr.exprType).classDecl.GetDecl("objectenumerator"))
 				If Not declList Then
-					Err "Use of EachIn requires enumerable Type with either ObjectEnumerator method or one which implements IIterable interface."
+					Err NotIterableError
 				End If
 			End If
 
@@ -213,6 +257,7 @@ Type TForEachinStmt Extends TLoopStmt
 				enumerInit = New TFuncCallExpr.Create( New TIdentExpr.Create( "ObjectEnumerator",expr ) )
 			End If
 			Local enumerTmp:TLocalDecl=New TLocalDecl.Create( "",Null,enumerInit,,True )
+			enumerTmp.Semant()
 
 			Local hasNextExpr:TExpr
 			If iterable Then
@@ -236,16 +281,38 @@ Type TForEachinStmt Extends TLoopStmt
 
 				Local cExpr:TExpr
 				
+				Local varObjTmp:TLocalDecl
+				Local varObjStmt:TStmt
+				
 				If iterable Or (TIdentType(varty) And TIdentType(varty).ident = "Object") Then
 					cExpr = nextObjExpr
 				Else
-					cExpr = New TCastExpr.Create( varty, nextObjExpr,CAST_EXPLICIT )
+					If TStringType(varty) Then
+						varObjTmp = New TLocalDecl.Create( "",TType.objectType,nextObjExpr,,True)
+						varObjTmp.Semant()
+						Local varObjExpr:TExpr=New TVarExpr.Create( varObjTmp )
+						
+						Local expr:TExpr = New TFuncCallExpr.Create( New TIdentExpr.Create( "ObjectIsString"), [varObjExpr])
+						expr=New TBinaryCompareExpr.Create( "=",expr, New TConstExpr.Create( New TIntType,"0" ))
+						
+						Local thenBlock:TBlockDecl=New TBlockDecl.Create( block.scope, True, BLOCK_IF )
+						Local elseBlock:TBlockDecl=New TBlockDecl.Create( block.scope, True, BLOCK_ELSE )
+						cont = New TContinueStmt.Create(Null, True)
+						thenBlock.AddStmt cont
+	
+						varObjStmt = New TIfStmt.Create( expr,thenBlock,elseBlock, True )
+						'block.stmts.AddFirst New TIfStmt.Create( expr,thenBlock,elseBlock, True )
+						
+						cExpr = New TCastExpr.Create( varty, varObjExpr,CAST_EXPLICIT )
+					Else
+						cExpr = New TCastExpr.Create( varty, nextObjExpr,CAST_EXPLICIT )
+					End If 
 				End If
 
 				' local variable
 				Local varTmp:TLocalDecl=New TLocalDecl.Create( varid,varty,cExpr)
 
-				If Not TNumericType(varty) Then
+				If Not TNumericType(varty) And Not varObjTmp Then
 					' local var as expression
 					Local expr:TExpr=New TVarExpr.Create( varTmp )
 	
@@ -261,6 +328,10 @@ Type TForEachinStmt Extends TLoopStmt
 					block.stmts.AddFirst New TIfStmt.Create( expr,thenBlock,elseBlock, True )
 				End If
 				block.stmts.AddFirst New TDeclStmt.Create( varTmp, True )
+				If varObjTmp Then
+					block.stmts.AddFirst varObjStmt
+					block.stmts.AddFirst New TDeclStmt.Create( varObjTmp, True )
+				End If
 			Else
 
 				If Not varty Then
@@ -268,6 +339,30 @@ Type TForEachinStmt Extends TLoopStmt
 					varty = varExpr.exprType
 				End If
 				
+				Local varObjTmp:TLocalDecl
+				Local varObjStmt:TStmt
+				Local cExpr:TExpr
+				
+				If TStringType(varty) Then
+					varObjTmp = New TLocalDecl.Create( "",TType.objectType,nextObjExpr,,True)
+					varObjTmp.Semant()
+					Local varObjExpr:TExpr=New TVarExpr.Create( varObjTmp )
+					
+					Local expr:TExpr = New TFuncCallExpr.Create( New TIdentExpr.Create( "ObjectIsString"), [varObjExpr])
+					expr=New TBinaryCompareExpr.Create( "=",expr, New TConstExpr.Create( New TIntType,"0" ))
+					
+					Local thenBlock:TBlockDecl=New TBlockDecl.Create( block.scope, True, BLOCK_IF )
+					Local elseBlock:TBlockDecl=New TBlockDecl.Create( block.scope, True, BLOCK_ELSE )
+					cont = New TContinueStmt.Create(Null, True)
+					thenBlock.AddStmt cont
+
+					varObjStmt = New TIfStmt.Create( expr,thenBlock,elseBlock, True )
+					'block.stmts.AddFirst New TIfStmt.Create( expr,thenBlock,elseBlock, True )
+					
+					cExpr = New TCastExpr.Create( varty, varObjExpr,CAST_EXPLICIT )
+				Else
+					cExpr = New TCastExpr.Create( varty, nextObjExpr,CAST_EXPLICIT )
+				End If 
 '				If Not varty Then
 '					Local decl:TValDecl = block.scope.FindValDecl(varid.ToLower())
 '					
@@ -280,7 +375,8 @@ Type TForEachinStmt Extends TLoopStmt
 				
 				' var = Null
 '				Local expr:TExpr=New TBinaryCompareExpr.Create( "=",New TIdentExpr.Create( varid ), New TNullExpr.Create(TType.nullObjectType))
-				If Not TNumericType(varty) Then
+				If Not TNumericType(varty) And Not varObjTmp Then
+
 					Local expr:TExpr=New TBinaryCompareExpr.Create( "=",varExpr, New TNullExpr.Create(TType.nullObjectType))
 	
 					' then continue
@@ -294,7 +390,11 @@ Type TForEachinStmt Extends TLoopStmt
 
 				End If
 '				block.stmts.AddFirst New TAssignStmt.Create( "=",New TIdentExpr.Create( varid ),New TCastExpr.Create( varty, nextObjExpr,CAST_EXPLICIT ))
-				block.stmts.AddFirst New TAssignStmt.Create( "=",varExpr,New TCastExpr.Create( varty, nextObjExpr,CAST_EXPLICIT ))
+				block.stmts.AddFirst New TAssignStmt.Create( "=",varExpr,cExpr)
+				If varObjTmp Then
+					block.stmts.AddFirst varObjStmt
+					block.stmts.AddFirst New TDeclStmt.Create( varObjTmp, True )
+				End If
 			EndIf
 
 			Local whileStmt:TWhileStmt=New TWhileStmt.Create( hasNextExpr,block, loopLabel, True )
@@ -311,7 +411,7 @@ Type TForEachinStmt Extends TLoopStmt
 			End If
 
 		Else
-			InternalErr "TForEachinStmt.OnSemant"
+			Err NotIterableError
 		EndIf
 
 		block.Semant
@@ -355,6 +455,14 @@ Type TIncbin
 		id = count
 		Return Self
 	End Method
+	
+	Method GeneratedDataName:String(app:TAppDecl)
+		Return "_ib" + app.munged + "_" + id + "_data"
+	End Method
+
+	Method GeneratedSizeName:String(app:TAppDecl)
+		Return "_ib" + app.munged + "_" + id + "_size"
+	End Method
 
 End Type
 
@@ -471,7 +579,11 @@ Type TParser Extends TGenProcessor
 			Case "~n"
 				Return "end-of-line"
 		End Select
-		Return "'" + toke + "'"
+		Local uni:String
+		If toke.length > 0 And toke[0] > 255 Then
+			uni = " (unicode : " + _toker._lastTCHR + ")"
+		End If
+		Return "'" + toke + "'" + uni
 	End Method
 
 	Method CParse:Int( toke$ )
@@ -492,7 +604,7 @@ Type TParser Extends TGenProcessor
 
 	Method Parse( toke$ )
 		If Not CParse( toke )
-			DoErr "Syntax error - expecting '"+toke+"'."
+			DoErr "Syntax error - expecting '"+toke+"' but found " + DescribeToke(toke)
 		EndIf
 	End Method
 
@@ -902,8 +1014,12 @@ Type TParser Extends TGenProcessor
 		
 		' array or function pointer?
 		Repeat
-			If (_toke = "[" Or _toke = "[]") And IsArrayDef()
-				ty = ParseArrayType(ty)
+			If (_toke = "[" Or _toke = "[]") And (IsArrayDef() Or IsArrayDef(attr & DECL_STATIC > 0)) Then
+				If Not IsArrayDef(attr & DECL_STATIC > 0) Then
+					Err "Invalid static array initialization."
+				Else
+					ty = ParseArrayType(ty, attr & DECL_STATIC > 0)
+				End If
 			Else If _toke = "(" Then
 				Local args:TArgDecl[] = ParseFuncParamDecl()
 				attr :| ParseCallConvention(attr & DECL_API_STDCALL)
@@ -931,7 +1047,16 @@ Type TParser Extends TGenProcessor
 	End Method
 
 	' replaces While CParse( "[]" ) sections, with support for multi-dimension arrays
-	Method ParseArrayType:TType(ty:TType)
+	Method ParseArrayType:TType(ty:TType, isStatic:Int = False)
+		If isStatic Then
+			Parse("[")
+			Local expr:TExpr = ParseUnaryExpr()
+			ty = New TArrayType.Create( ty )
+			TArrayType(ty).isStatic = True
+			TArrayType(ty).length = expr.Eval()
+			Parse("]")
+			Return ty
+		End If
 
 		While True
 			Local dims:Int = 1
@@ -977,9 +1102,19 @@ Type TParser Extends TGenProcessor
 		Return False
 	End Method
 	
-	Method IsArrayDef:Int()
+	Method IsArrayDef:Int(isStatic:Int = False)
 		Local isDef:Int = True
 		Local toker:TToker=New TToker.Copy(_toker)
+		If isStatic Then
+			If Not CParseToker(toker, "[") Then
+				Return False
+			End If
+			NextTokeToker(toker)
+			If Not CParseToker(toker, "]") Then
+				Return False
+			End If
+			Return True
+		End If
 		While True
 			'Local dims:Int = 1
 			
@@ -1351,6 +1486,24 @@ Type TParser Extends TGenProcessor
 				expr=ParseExpr()
 				expr=New TStackAllocExpr.Create( expr )
 			EndIf
+		Case "fieldoffset"
+			NextToke
+			
+			Local withBrackets:Int
+			
+			If CParse("(")
+				withBrackets = True
+			End If
+			
+			Local typeExpr:TExpr = ParseExpr()
+			Parse ","
+			Local fieldExpr:TExpr = ParseExpr()
+			
+			If withBrackets Then
+				Parse(")")
+			End If
+			
+			expr=New TFieldOffsetExpr.Create( typeExpr, fieldExpr )
 		Default
 			Select _tokeType
 			Case TOKE_IDENT
@@ -1406,6 +1559,11 @@ Type TParser Extends TGenProcessor
 				expr=New TConstExpr.Create( TType.stringType,s )
 				_app.mapStringConsts(s)
 				NextToke
+			Case TOKE_STRINGMULTI
+				Local s:String = BmxProcessMultiString( _toke )
+				expr=New TConstExpr.Create( TType.stringType,s )
+				_app.mapStringConsts(s)
+				NextToke
 			Default
 				Err "Expecting expression but encountered "+DescribeToke(_toke)
 			End Select
@@ -2050,16 +2208,16 @@ End Rem
 						ty=ParseArrayType(ty)
 					Wend
 				End If
+				PopBlock
 				Local init:TLocalDecl=New TLocalDecl.Create( id,ty,Null,0 )
 				Local block:TBlockDecl=New TBlockDecl.Create( _block, , BLOCK_CATCH )
 				catches.AddLast(New TCatchStmt.Create( init,block ))
-				PopBlock
 				PushBlock block
 			Else If CParse("finally") Then
 				If finallyStmt Then Err "Try statement cannot have more than one Finally block."
+				PopBlock
 				Local block:TBlockDecl = New TBlockDecl.Create(_block, , BLOCK_FINALLY)
 				finallyStmt = New TFinallyStmt.Create(block)
-				PopBlock
 				PushBlock block
 			Else
 				ParseStmt
@@ -2403,13 +2561,23 @@ End Rem
 
 		SetErr
 
+		If CParse("staticarray") Then
+			If toke = "const" Then
+				Err "Const cannot be used in this way"
+			End If
+			If attrs & DECL_STATIC Then
+				Err "Already declared as a static array"
+			End If
+			attrs :| DECL_STATIC
+		End If
+
 		Local id$=ParseIdent()
 		Local ty:TType
 		Local init:TExpr
 		
 		
 		If attrs & DECL_EXTERN
-			ty=ParseDeclType(attrs & DECL_API_STDCALL)
+			ty=ParseDeclType(attrs & (DECL_STATIC | DECL_API_STDCALL))
 			
 			If toke = "const" Then
 				If CParse("=") Then
@@ -2420,29 +2588,38 @@ End Rem
 '			init=ParseExpr()
 '			ty = init.exprType
 		Else
-			ty=ParseDeclType(attrs & DECL_API_STDCALL)
+			ty=ParseDeclType(attrs & (DECL_STATIC | DECL_API_STDCALL))
 
 			If CParse( "=" )
+				If (attrs & DECL_STATIC) Then
+					Err "Static arrays cannot be initialized in this way"
+				End If
 				init=ParseExpr()
 			Else If CParse( "[" ) ' an initialised array?
-				Local ln:TExpr[]
-				Repeat
-					If CParse(",") Then
-						ln = ln + [New TNullExpr]
-						Continue
-					End If
-					If CParse("]") Exit
-					ln = ln + [ParseExpr()]
-					If CParse("]") Exit
-					Parse(",")
-				Forever
-				'Parse "]"
-				ty = ParseArrayType(ty)
-				'While CParse( "[]" )
-				'	ty=New TArrayType.Create(ty)
-				'Wend
-				init=New TNewArrayExpr.Create( ty,ln)
-				ty=New TArrayType.Create( ty, ln.length )
+				If (attrs & DECL_STATIC) Then
+					init = ParseExpr()
+					Parse "]"
+					ty=New TArrayType.Create( ty,1,, attrs & DECL_STATIC > 0 )
+				Else
+					Local ln:TExpr[]
+					Repeat
+						If CParse(",") Then
+							ln = ln + [New TNullExpr]
+							Continue
+						End If
+						If CParse("]") Exit
+						ln = ln + [ParseExpr()]
+						If CParse("]") Exit
+						Parse(",")
+					Forever
+					'Parse "]"
+					ty = ParseArrayType(ty)
+					'While CParse( "[]" )
+					'	ty=New TArrayType.Create(ty)
+					'Wend
+					init=New TNewArrayExpr.Create( ty,ln)
+					ty=New TArrayType.Create( ty, ln.length,, attrs & DECL_STATIC > 0 )
+				End If
 			Else If toke <> "const"
 				If toke="global" Or toke="local" Then
 					init=New TConstExpr.Create( ty,"" )
@@ -2531,9 +2708,24 @@ End Rem
 		If toke Parse toke
 
 		If isField Then
-			If CParse("readonly") Then
-				attrs :| DECL_READ_ONLY
-			End If
+			Repeat
+				If CParse("readonly") Then
+					If attrs & DECL_READ_ONLY
+						Err "Duplicate modifier 'ReadOnly'."
+					End If
+
+					attrs :| DECL_READ_ONLY
+
+				Else If CParse("staticarray") Then
+					If attrs & DECL_STATIC
+						Err "Duplicate modifier 'Static'."
+					End If
+
+					attrs :| DECL_STATIC
+				Else
+					Exit
+				End If
+			Forever
 		End If
 
 		Local decls:TList=New TList'<Decl>
@@ -2998,7 +3190,7 @@ End Rem
 			Select opt_platform
 				Case "macos", "osx", "ios"
 					api = "macos"
-				Case "linux", "android", "raspberrypi"
+				Case "linux", "android", "raspberrypi", "haiku"
 					api = "linux"
 				Case "win32"
 					api = "win32"
@@ -3029,21 +3221,28 @@ End Rem
 		If _toke<>")"
 			Local nargs:Int
 			Repeat
+				Local attrs:Int
+				If CParse("staticarray") Then
+					attrs :| DECL_STATIC
+				End If
 				
 				Local argId$=ParseIdent()
 
-				Local ty:TType=ParseDeclType()
+				Local ty:TType=ParseDeclType(attrs)
 
 				Local init:TExpr
 				
 				' var argument?
 				If CParse("var") Then
+					If attrs & DECL_STATIC Then
+						Err "Unexpected 'Var' for static array argument"
+					End If
 					ty = TType.MapToVarType(ty)
 				Else If CParse( "=" )
 					init=ParseExpr()
 				End If
 				
-				Local arg:TArgDecl=New TArgDecl.Create( argId,ty,init )
+				Local arg:TArgDecl=New TArgDecl.Create( argId,ty,init,attrs )
 				If args.Length=nargs args=args + New TArgDecl[10]
 				args[nargs]=arg
 				nargs:+1
@@ -4386,23 +4585,8 @@ Function EvalS$( source$,ty:TType )
 
 	Local env:TScopeDecl=New TScopeDecl
 
-Rem
-Debug	True if program is being compiled in debug mode.
-Threaded	True if program is being compiled in threaded mode.
-Win32	True if program is being compiled for the Windows operating system.
-MacOS	True if program is being compiled for the MacOS operating system.
-Linux	True if program is being compiled for the Linux operating system.
-X86	True if program is being compiled for the Intel CPU.
-PPC	True if program is being compiled for the PowerPC CPU.
-MacOSX86	True if program is being compiled for an Intel Mac.
-MacOSPPC	True if program is being compiled for a PowerPC Mac.
-BigEndian	True if program is being compiled for a big endian CPU.
-LittleEndian
-End Rem
-
 	' debug/release
 	env.InsertDecl New TConstDecl.Create( "debug",New TIntType,New TConstExpr.Create( New TIntType,opt_debug ),0 )
-	'env.InsertDecl New TConstDecl.Create( "release",TType.intType,New TConstExpr.Create( TType.intType,opt_release ),0 )
 
 	' threaded
 	env.InsertDecl New TConstDecl.Create( "threaded",New TIntType,New TConstExpr.Create( New TIntType,opt_threaded ),0 )
@@ -4412,12 +4596,14 @@ End Rem
 	env.InsertDecl New TConstDecl.Create( "macosx86",New TIntType,New TConstExpr.Create( New TIntType,(opt_platform="macos" Or opt_platform="osx" Or opt_platform="ios") And opt_arch="x86"),0 )
 	env.InsertDecl New TConstDecl.Create( "macosppc",New TIntType,New TConstExpr.Create( New TIntType,(opt_platform="macos" Or opt_platform="osx") And opt_arch="ppc"),0 )
 	env.InsertDecl New TConstDecl.Create( "macosx64",New TIntType,New TConstExpr.Create( New TIntType,(opt_platform="macos" Or opt_platform="osx" Or opt_platform="ios") And opt_arch="x64"),0 )
+	env.InsertDecl New TConstDecl.Create( "macosarm64",New TIntType,New TConstExpr.Create( New TIntType,(opt_platform="macos" Or opt_platform="osx" Or opt_platform="ios") And opt_arch="arm64"),0 )
 
 	' osx
 	env.InsertDecl New TConstDecl.Create( "osx",New TIntType,New TConstExpr.Create( New TIntType,opt_platform="macos" Or opt_platform="osx" ),0 )
 	env.InsertDecl New TConstDecl.Create( "osxx86",New TIntType,New TConstExpr.Create( New TIntType,(opt_platform="macos" Or opt_platform="osx")  And opt_arch="x86"),0 )
 	env.InsertDecl New TConstDecl.Create( "osxppc",New TIntType,New TConstExpr.Create( New TIntType,(opt_platform="macos" Or opt_platform="osx") And opt_arch="ppc"),0 )
 	env.InsertDecl New TConstDecl.Create( "osxx64",New TIntType,New TConstExpr.Create( New TIntType,(opt_platform="macos" Or opt_platform="osx") And opt_arch="x64"),0 )
+	env.InsertDecl New TConstDecl.Create( "osxarm64",New TIntType,New TConstExpr.Create( New TIntType,(opt_platform="macos" Or opt_platform="osx") And opt_arch="arm64"),0 )
 
 	' ios
 	env.InsertDecl New TConstDecl.Create( "ios",New TIntType,New TConstExpr.Create( New TIntType,opt_platform="ios" ),0 )
@@ -4453,6 +4639,11 @@ End Rem
 	env.InsertDecl New TConstDecl.Create( "raspberrypiARM",New TIntType,New TConstExpr.Create( New TIntType,opt_platform="raspberrypi" And opt_arch="arm"),0 )
 	env.InsertDecl New TConstDecl.Create( "raspberrypiARM64",New TIntType,New TConstExpr.Create( New TIntType,opt_platform="raspberrypi" And opt_arch="arm64"),0 )
 
+	' haiku
+	env.InsertDecl New TConstDecl.Create( "haiku",New TIntType,New TConstExpr.Create( New TIntType,opt_platform="haiku" And (opt_arch="x86" Or opt_arch="x64")),0 )
+	env.InsertDecl New TConstDecl.Create( "haikux86",New TIntType,New TConstExpr.Create( New TIntType,opt_platform="haiku" And opt_arch="x86"),0 )
+	env.InsertDecl New TConstDecl.Create( "haikux64",New TIntType,New TConstExpr.Create( New TIntType,opt_platform="haiku" And opt_arch="x64"),0 )
+
 	' emscripten
 	env.InsertDecl New TConstDecl.Create( "emscripten",New TIntType,New TConstExpr.Create( New TIntType,opt_platform="emscripten" ),0 )
 	env.InsertDecl New TConstDecl.Create( "emscriptenjs",New TIntType,New TConstExpr.Create( New TIntType,opt_platform="emscripten" And opt_arch="js"),0 )
@@ -4510,10 +4701,6 @@ End Rem
 		Next
 	End If
 
-'	env.InsertDecl New TConstDecl.Create( "LANG",TType.stringType,New TConstExpr.Create( TType.stringType,ENV_LANG ),0 )
-'	env.InsertDecl New TConstDecl.Create( "TARGET",TType.stringType,New TConstExpr.Create( TType.stringType,ENV_TARGET ),0 )
-'	env.InsertDecl New TConstDecl.Create( "CONFIG",TType.stringType,New TConstExpr.Create( TType.stringType,ENV_CONFIG ),0 )
-
 	PushEnv env
 
 	Local toker:TToker=New TToker.Create( "",source )

stmt.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -135,19 +135,32 @@ Type TAssignStmt Extends TStmt
 				If TVarExpr(lhs) Then
 					decl = TVarExpr(lhs).decl
 				Else
-					decl = TMemberVarExpr(lhs).decl
+					Local mvExpr:TMemberVarExpr = TMemberVarExpr(lhs)
+					decl = mvExpr.decl
+					
+					If TFieldDecl(decl) And (TInvokeExpr(mvExpr.expr) Or TInvokeMemberExpr(mvExpr.expr)) Then
+						If TClassDecl(decl.scope) And TClassDecl(decl.scope).IsStruct() Then
+							rhs = Null
+							Warn "Discarding Field assignment of Struct returned via invocation"
+							Return
+						End If
+					End If
+					
 				End If
 				If decl And decl.IsReadOnly() Then
 					If TFieldDecl(decl) Then
 						' check scope for ctor
 						Local scope:TFuncDecl = _env.FuncScope()
-						If Not scope Or Not scope.IsCtor() Or (decl.ClassScope() <> scope.ClassScope()) Then
+						If Not scope Or Not scope.IsCtor() Or (Not scope.ClassScope().ExtendsClass(decl.ClassScope())) Then
 							Err "Cannot modify ReadOnly field " + decl.ident
 						End If
 					Else
 						Err "Cannot modify ReadOnly variable " + decl.ident
 					End If
 				End If
+				If TValDecl(decl) And TArrayType(TValDecl(decl).ty) And TArrayType(TValDecl(decl).ty).isStatic Then
+					Err "Static arrays cannot be assigned in this way."
+				End If
 			End If
 		
 			If IsPointerType(lhs.exprType, 0, TType.T_POINTER | TType.T_VARPTR) And TNumericType(rhs.exprType) Then
@@ -282,7 +295,7 @@ Type TReturnStmt Extends TStmt
 			expr=expr.SemantAndCast( fRetType  )
 			If TIdentTypeExpr(expr) Err "Function must return a value."
 		Else If fdecl.IsCtor()
-			expr=New TSelfExpr.Semant()
+			' ctors do not return anything
 		Else If Not TVoidType( fRetType  )
 			If _env.ModuleScope().IsSuperStrict() Err "Function must return a value"
 			expr=New TConstExpr.Create( fRetType ,"" ).Semant()

stringbuffer_common.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2016-2019 Bruce A Henderson
+' Copyright (c) 2016-2020 Bruce A Henderson
 ' 
 ' Permission is hereby granted, free of charge, to any person obtaining a copy
 ' of this software and associated documentation files (the "Software"), to deal

stringbuffer_core.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2016-2019 Bruce A Henderson
+' Copyright (c) 2016-2020 Bruce A Henderson
 ' 
 ' Permission is hereby granted, free of charge, to any person obtaining a copy
 ' of this software and associated documentation files (the "Software"), to deal

stringbuffer_glue.c

@@ -1,5 +1,5 @@
 /*
-  Copyright (c) 2016-2019 Bruce A Henderson
+  Copyright (c) 2016-2020 Bruce A Henderson
  
   Permission is hereby granted, free of charge, to any person obtaining a copy
   of this software and associated documentation files (the "Software"), to deal

toker.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -39,6 +39,7 @@ Const TOKE_SYMBOL:Int=8
 Const TOKE_LINECOMMENT:Int=9
 Const TOKE_LONGLIT:Int=10
 Const TOKE_NATIVE:Int=11
+Const TOKE_STRINGMULTI:Int=12
 
 '***** Tokenizer *****
 Type TToker
@@ -50,7 +51,7 @@ Type TToker
 		"next,return,alias,rem,endrem,throw,assert,try,catch,finally,nodebug,incbin,endselect,endmethod," + ..
 		"endfunction,endtype,endextern,endtry,endwhile,pi,release,defdata,readdata,restoredata,interface," + ..
 		"endinterface,implements,size_t,uint,ulong,struct,endstruct,operator,where,readonly,export,override," + ..
-		"enum,endenum,stackalloc,inline"
+		"enum,endenum,stackalloc,inline,fieldoffset"
 	Global _keywords:TMap
 
 	Field _path$
@@ -199,17 +200,51 @@ Type TToker
 				_tokePos:+1
 			Wend
 		Else If str="~q"
+			Local isMulti:Int
 			_tokeType=TOKE_STRINGLIT
 			Local _tstr:String = TSTR()
-			While _tstr And _tstr<>"~q"
-				' Strings can't cross line boundries
-				If _tstr="~n" Then
+			If _tstr = "~q" Then
+				_tokePos:+1
+				Local _tstr2:String = TSTR()
+				If _tstr2 = "~q" Then
+					isMulti = True
+				Else
 					_tokePos:-1
-					Exit
 				End If
-				_tokePos:+1
-				_tstr = TSTR()
-			Wend
+			End If
+			If Not isMulti Then
+				While _tstr And _tstr<>"~q"
+					' Strings can't cross line boundries
+					If _tstr="~n" Then
+						_tokePos:-1
+						Exit
+					End If
+					_tokePos:+1
+					_tstr = TSTR()
+				Wend
+			Else
+				Local lineCount:Int
+				Local count:Int
+				_tokeType = TOKE_STRINGMULTI
+				While _tstr
+					_tokePos:+1
+					_tstr = TSTR()
+					
+					If _tstr = "~n" Then
+						lineCount:+1
+					End If
+					
+					If _tstr = "~q" Then
+						count :+ 1
+						If count = 3 Then
+							_line :+ lineCount - 1
+							Exit
+						End If
+					Else
+						count = 0
+					End If
+				Wend
+			End If
 			If _tokePos<_source.Length _tokePos:+1 Else _tokeType=TOKE_STRINGLITEX
 		Else If str="'"
 			Local _tstr:String = TSTR()

transform.c

@@ -1,4 +1,4 @@
-/* Copyright (c) 2014-2019 Bruce A Henderson
+/* Copyright (c) 2014-2020 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

translator.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -1338,6 +1338,7 @@ End Rem
 		If TLenExpr(expr) Return TransLenExpr(TLenExpr(expr))
 		If TSizeOfExpr(expr) Return TransSizeOfExpr(TSizeOfExpr(expr))
 		If TStackAllocExpr(expr) Return TransStackAllocExpr(TStackAllocExpr(expr))
+		If TFieldOffsetExpr(expr) Return TransFieldOffsetExpr(TFieldOffsetExpr(expr))
 		Err "TODO : TransBuiltinExpr()"
 	End Method
 	
@@ -1355,6 +1356,9 @@ End Rem
 
 	Method TransStackAllocExpr:String(expr:TStackAllocExpr)
 	End Method
+
+	Method TransFieldOffsetExpr:String(expr:TFieldOffsetExpr)
+	End Method
 	
 	Method TransIdentTypeExpr:String(expr:TIdentTypeExpr) Abstract
 	
@@ -1961,7 +1965,7 @@ End Rem
 		Local infoArray:String[] = info.Split(";")
 
 		Local dbg:String = "struct BBDebugStm __stmt_" + count + " = {"
-		dbg :+ Enquote(infoArray[0]) + ", "
+		dbg :+ GenHash(infoArray[0]) + ", "
 		dbg :+ infoArray[1] + ", "
 		dbg :+ infoArray[2] + "};"
 		Emit dbg

type.bmx

@@ -1,4 +1,4 @@
-' Copyright (c) 2013-2019 Bruce A Henderson
+' Copyright (c) 2013-2020 Bruce A Henderson
 '
 ' Based on the public domain Monkey "trans" by Mark Sibly
 '
@@ -1406,11 +1406,15 @@ End Type
 Type TArrayType Extends TType
 	Field elemType:TType
 	Field dims:Int
+	Field isStatic:Int
+	Field length:String
 	
-	Method Create:TArrayType( elemType:TType, dims:Int = 1, flags:Int = 0 )
+	Method Create:TArrayType( elemType:TType, dims:Int = 1, flags:Int = 0, isStatic:Int = False, length:Int = 0 )
 		Self.elemType=elemType
 		Self.dims = dims
 		Self._flags = flags
+		Self.isStatic = isStatic
+		Self.length = length
 		Return Self
 	End Method
 	
@@ -1422,12 +1426,13 @@ Type TArrayType Extends TType
 		
 	Method EqualsType:Int( ty:TType )
 		Local arrayType:TArrayType=TArrayType( ty )
-		Return arrayType And elemType.EqualsType( arrayType.elemType ) And dims = arrayType.dims
+		Return arrayType And elemType.EqualsType( arrayType.elemType ) And dims = arrayType.dims And arrayType.isStatic = isStatic And arrayType.length = length
 	End Method
 	
 	Method ExtendsType:Int( ty:TType, noExtendString:Int = False, widensTest:Int = False )
 		Local arrayType:TArrayType=TArrayType( ty )
 		Return (arrayType And dims = arrayType.dims And ..
+			(arrayType.isStatic = isStatic And arrayType.length = length) And ..
 			( TVoidType( elemType ) ..
 				Or elemType.EqualsType( arrayType.elemType ) ..
 				Or ((TObjectType(elemType) Or TStringType(elemType) Or TArrayType(elemType)) And ..
@@ -1440,7 +1445,7 @@ Type TArrayType Extends TType
 	
 	Method Semant:TType(option:Int = False, callback:TCallback = Null)
 		Local ty:TType=elemType.Semant(option, callback)
-		If ty<>elemType Return New TArrayType.Create( ty, dims, _flags )
+		If ty<>elemType Return New TArrayType.Create( ty, dims, _flags, isStatic, Int(length) )
 		Return Self
 	End Method
 	
@@ -1453,11 +1458,19 @@ Type TArrayType Extends TType
 		Local ty:TArrayType = New TArrayType
 		ty.elemType = elemType
 		ty.dims = dims
+		ty.isStatic = isStatic
+		ty.length = length
 		Return ty
 	End Method
 
 	Method ToString$()
-		Return elemType.ToString()+" Array"
+		Local t:String = elemType.ToString()
+		If isStatic Then
+			t :+ " StaticArray[" + length + "]"
+		Else
+			t :+ " Array"
+		End If
+		Return t
 	End Method
 	
 	Method DistanceToType:Int(ty:TType)

xxh3.h

@@ -0,0 +1,2320 @@
+/*
+ * 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 */

xxhash.h

@@ -0,0 +1,1965 @@
+/*
+ * xxHash - Extremely Fast Hash algorithm
+ * Header File
+ * Copyright (C) 2012-2020 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
+ */
+
+/* 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
+*/
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* ****************************
+ *  INLINE mode
+ ******************************/
+/*!
+ * XXH_INLINE_ALL (and XXH_PRIVATE_API)
+ * 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
+ *
+ * It also keeps xxHash symbols private to the unit, so they are not exported.
+ *
+ * Usage:
+ *     #define XXH_INLINE_ALL
+ *     #include "xxhash.h"
+ *
+ * Do not compile and link xxhash.o as a separate object, as it is not useful.
+ */
+#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
+    && !defined(XXH_INLINE_ALL_31684351384)
+   /* this section should be traversed only once */
+#  define XXH_INLINE_ALL_31684351384
+   /* give access to the advanced API, required to compile implementations */
+#  undef XXH_STATIC_LINKING_ONLY   /* avoid macro redef */
+#  define XXH_STATIC_LINKING_ONLY
+   /* make all functions private */
+#  undef XXH_PUBLIC_API
+#  if defined(__GNUC__)
+#    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)
+#    define XXH_PUBLIC_API static __inline
+#  else
+     /* note: this version may generate warnings for unused static functions */
+#    define XXH_PUBLIC_API static
+#  endif
+
+   /*
+    * 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.
+    * 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
+#  define XXH_NAMESPACE XXH_INLINE_
+   /*
+    * Some identifiers (enums, type names) are not symbols, but they must
+    * still 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
+    */
+#  define XXH_IPREF(Id)   XXH_INLINE_ ## Id
+#  define XXH_OK XXH_IPREF(XXH_OK)
+#  define XXH_ERROR XXH_IPREF(XXH_ERROR)
+#  define XXH_errorcode XXH_IPREF(XXH_errorcode)
+#  define XXH32_canonical_t  XXH_IPREF(XXH32_canonical_t)
+#  define XXH64_canonical_t  XXH_IPREF(XXH64_canonical_t)
+#  define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t)
+#  define XXH32_state_s XXH_IPREF(XXH32_state_s)
+#  define XXH32_state_t XXH_IPREF(XXH32_state_t)
+#  define XXH64_state_s XXH_IPREF(XXH64_state_s)
+#  define XXH64_state_t XXH_IPREF(XXH64_state_t)
+#  define XXH3_state_s  XXH_IPREF(XXH3_state_s)
+#  define XXH3_state_t  XXH_IPREF(XXH3_state_t)
+#  define XXH128_hash_t XXH_IPREF(XXH128_hash_t)
+   /* Ensure the header is parsed again, even if it was previously included */
+#  undef XXHASH_H_5627135585666179
+#  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 */
+#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
+
+/*!
+ * 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)
+#  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)
+#  define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset)
+#  define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update)
+#  define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest)
+#  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)
+#  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)
+#  define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset)
+#  define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update)
+#  define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest)
+#  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)
+#endif
+
+
+/* *************************************
+*  Version
+***************************************/
+#define XXH_VERSION_MAJOR    0
+#define XXH_VERSION_MINOR    7
+#define XXH_VERSION_RELEASE  4
+#define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
+XXH_PUBLIC_API unsigned XXH_versionNumber (void);
+
+
+/* ****************************
+*  Definitions
+******************************/
+#include <stddef.h>   /* size_t */
+typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
+
+
+/*-**********************************************************************
+*  32-bit hash
+************************************************************************/
+#if !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   include <stdint.h>
+    typedef uint32_t XXH32_hash_t;
+#else
+#   include <limits.h>
+#   if UINT_MAX == 0xFFFFFFFFUL
+      typedef unsigned int XXH32_hash_t;
+#   else
+#     if ULONG_MAX == 0xFFFFFFFFUL
+        typedef unsigned long XXH32_hash_t;
+#     else
+#       error "unsupported platform: need a 32-bit type"
+#     endif
+#   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
+ *
+ * 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.
+ */
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
+
+/*******   Streaming   *******/
+
+/*
+ * 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.
+ *
+ * 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()`.
+ */
+
+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);
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);
+
+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   *******/
+
+/*
+ * The default return values from XXH functions are unsigned 32 and 64 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.
+ */
+
+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);
+
+
+#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
+
+/*!
+ * 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.
+ *
+ * This function usually runs faster on 64-bit systems, but slower on 32-bit
+ * systems (see benchmark).
+ *
+ * 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.
+ */
+XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t length, XXH64_hash_t seed);
+
+/*******   Streaming   *******/
+typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void);
+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);
+
+/*******   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);
+
+
+#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 an XXH
+ * state, for example, on the stack or in a struct.
+ * Never **ever** access members directly.
+ */
+
+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 */
+
+
+#ifndef XXH_NO_LONG_LONG  /* defined when there is no 64-bit support */
+
+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 */
+
+
+/*-**********************************************************************
+*  XXH3
+*  New experimental hash
+************************************************************************/
+
+/* ************************************************************************
+ * 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
+ *
+ * Speed analysis methodology is explained here:
+ *
+ *    https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ *
+ * 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.
+ *
+ * The algorithm is portable: Like XXH32 and XXH64, it generates the same hash
+ * on all platforms.
+ *
+ * It benefits greatly from SIMD and 64-bit arithmetic, but does not require it.
+ *
+ * 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.
+ *
+ * Optimized implementations are provided for AVX2, SSE2, NEON, POWER8, ZVector,
+ * and scalar targets. This can be controlled with the XXH_VECTOR macro.
+ *
+ * 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.
+ *
+ * It's also generally simpler to manipulate a scalar return type than a struct.
+ *
+ * The 128-bit version adds additional strength, but it is slightly slower.
+ *
+ * The XXH3 algorithm is still in development.
+ * The results it produces may still change in future versions.
+ *
+ * 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).
+ *
+ * Avoid storing values in long-term storage until the algorithm is finalized.
+ *
+ * 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.
+ *
+ * After which, return values of XXH3 and XXH128 will no longer change in
+ * future versions.
+ *
+ * XXH3's return values will be officially finalized upon reaching v0.8.0.
+ *
+ * The API supports one-shot hashing, streaming mode, and custom secrets.
+ */
+
+#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)
+
+#  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_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
+#  define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
+#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);
+
+/*
+ * 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.
+ */
+#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);
+
+/*
+ * 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().
+ */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+
+
+/* streaming 64-bit */
+
+#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))
+#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(__GNUC__)
+#   define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
+#else
+#   define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
+#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
+
+/*
+ * 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.
+ */
+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);
+
+
+/*
+ * XXH3_64bits_reset():
+ * Initialize with the default parameters.
+ * The result will be equivalent to `XXH3_64bits()`.
+ */
+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()`.
+ */
+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_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+
+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);
+
+
+/* 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)
+
+#  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 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
+
+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 */
+
+/*!
+ * XXH128_isEqual():
+ * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
+ */
+XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+
+/*!
+ * XXH128_cmp():
+ *
+ * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
+ *
+ * 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);
+
+
+/*******   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);
+
+
+#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 found in xxhash.c.
+ *
+ * However, code inlining requires the implementation to be visible to the
+ * compiler, usually within the header.
+ *
+ * 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.
+ *
+ * Therefore, the implementation is now directly integrated within xxhash.h.
+ * Another small advantage is that xxhash.c is no longer needed in /include.
+ ************************************************************************/
+
+#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
+   || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
+#  define XXH_IMPLEM_13a8737387
+
+/* *************************************
+*  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)
+ */
+#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
+
+/*!
+ *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.
+ */
+#ifndef XXH_ACCEPT_NULL_INPUT_POINTER   /* can be defined externally */
+#  define XXH_ACCEPT_NULL_INPUT_POINTER 0
+#endif
+
+/*!
+ * 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.
+ *
+ * This option does not affect XXH3.
+ */
+#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.
+ *
+ * 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), optimizing for size (-Os, -Oz), or using
+ * -fno-inline with GCC or Clang, this will automatically be defined.
+ */
+#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_REROLL:
+ * Whether to reroll XXH32_finalize, and XXH64_finalize,
+ * instead of using an unrolled jump table/if statement loop.
+ *
+ * This is automatically defined on -Os/-Oz on GCC and Clang.
+ */
+#ifndef XXH_REROLL
+#  if defined(__OPTIMIZE_SIZE__)
+#    define XXH_REROLL 1
+#  else
+#    define XXH_REROLL 0
+#  endif
+#endif
+
+
+/* *************************************
+*  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); }
+
+/*! and for memcpy() */
+#include <string.h>
+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 */
+#  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))
+#    else
+#      define XXH_FORCE_INLINE static inline
+#      define XXH_NO_INLINE static
+#    endif
+#  else
+#    define XXH_FORCE_INLINE static
+#    define XXH_NO_INLINE static
+#  endif /* __STDC_VERSION__ */
+#endif
+
+
+
+/* *************************************
+*  Debug
+***************************************/
+/*
+ * DEBUGLEVEL is expected to be defined externally, typically via the compiler's
+ * command line options. The value must be a number.
+ */
+#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
+
+/* note: use after variable declarations */
+#define XXH_STATIC_ASSERT(c)  do { enum { XXH_sa = 1/(int)(!!(c)) }; } while (0)
+
+
+/* *************************************
+*  Basic Types
+***************************************/
+#if !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# include <stdint.h>
+  typedef uint8_t  xxh_u8;
+#else
+  typedef unsigned char      xxh_u8;
+#endif
+typedef XXH32_hash_t xxh_u32;
+
+
+/* ***   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.
+ */
+#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))
+
+/*
+ * __pack instructions are safer but compiler specific, hence potentially
+ * problematic for some compilers.
+ *
+ * Currently only defined for GCC and ICC.
+ */
+typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
+static xxh_u32 XXH_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://stackoverflow.com/a/32095106/646947
+ */
+static xxh_u32 XXH_read32(const void* memPtr)
+{
+    xxh_u32 val;
+    memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+
+/* ***   Endianess   *** */
+typedef enum { XXH_bigEndian=0, XXH_littleEndian=1 } XXH_endianess;
+
+/*!
+ * 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)
+{
+    /*
+     * 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];
+}
+#   define XXH_CPU_LITTLE_ENDIAN   XXH_isLittleEndian()
+#  endif
+#endif
+
+
+
+
+/* ****************************************
+*  Compiler-specific Functions and Macros
+******************************************/
+#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+
+#ifndef __has_builtin
+#  define __has_builtin(x) 0
+#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
+
+#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
+*****************************/
+typedef enum { XXH_aligned, XXH_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
+***************************************/
+XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
+
+
+/* *******************************************************************
+*  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 */
+
+static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
+{
+    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;
+}
+
+/* mix all bits */
+static xxh_u32 XXH32_avalanche(xxh_u32 h32)
+{
+    h32 ^= h32 >> 15;
+    h32 *= PRIME32_2;
+    h32 ^= h32 >> 13;
+    h32 *= PRIME32_3;
+    h32 ^= h32 >> 16;
+    return(h32);
+}
+
+#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)
+{
+#define PROCESS1 do {                           \
+    h32 += (*ptr++) * PRIME32_5;                \
+    h32 = XXH_rotl32(h32, 11) * PRIME32_1;      \
+} while (0)
+
+#define PROCESS4 do {                           \
+    h32 += XXH_get32bits(ptr) * PRIME32_3;      \
+    ptr += 4;                                   \
+    h32  = XXH_rotl32(h32, 17) * PRIME32_4;     \
+} while (0)
+
+    /* 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 */
+    }
+}
+
+XXH_FORCE_INLINE xxh_u32
+XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
+{
+    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;
+    }
+#endif
+
+    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);
+
+        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;
+
+    return XXH32_finalize(h32, input, len&15, align);
+}
+
+
+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);
+
+#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   *******/
+
+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;
+}
+
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
+{
+    memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+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;
+}
+
+
+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;
+#endif
+
+    {   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->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;
+        }
+
+        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;
+
+            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);
+
+            state->v1 = v1;
+            state->v2 = v2;
+            state->v3 = v3;
+            state->v4 = v4;
+        }
+
+        if (p < bEnd) {
+            XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+XXH_PUBLIC_API XXH32_hash_t XXH32_digest (const XXH32_state_t* state)
+{
+    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;
+    }
+
+    h32 += state->total_len_32;
+
+    return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
+}
+
+
+/*******   Canonical representation   *******/
+
+/*
+ * 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).
+ *
+ * This way, hash values can be written into a file or buffer, remaining
+ * comparable across different systems.
+ *
+ * The following functions allow transformation of hash values to and from their
+ * canonical format.
+ */
+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);
+    memcpy(dst, &hash, sizeof(*dst));
+}
+
+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
+*********************************************************************/
+
+/*******   Memory access   *******/
+
+typedef XXH64_hash_t xxh_u64;
+
+
+/*!
+ * XXH_REROLL_XXH64:
+ * Whether to reroll the XXH64_finalize() loop.
+ *
+ * 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.
+ *
+ * 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.
+ *
+ * If XXH_REROLL is defined, this is ignored and the loop is always rerolled.
+ */
+#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) */
+
+#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))
+
+/*
+ * __pack instructions are safer, but compiler specific, hence potentially
+ * problematic for some compilers.
+ *
+ * Currently only defined for GCC and ICC.
+ */
+typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
+static xxh_u64 XXH_read64(const void* ptr) { return ((const unalign64*)ptr)->u64; }
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://stackoverflow.com/a/32095106/646947
+ */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    xxh_u64 val;
+    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   *******/
+
+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 */
+
+static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
+{
+    acc += input * PRIME64_2;
+    acc  = XXH_rotl64(acc, 31);
+    acc *= PRIME64_1;
+    return acc;
+}
+
+static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
+{
+    val  = XXH64_round(0, val);
+    acc ^= val;
+    acc  = acc * PRIME64_1 + PRIME64_4;
+    return acc;
+}
+
+static xxh_u64 XXH64_avalanche(xxh_u64 h64)
+{
+    h64 ^= h64 >> 33;
+    h64 *= PRIME64_2;
+    h64 ^= h64 >> 29;
+    h64 *= PRIME64_3;
+    h64 ^= h64 >> 32;
+    return h64;
+}
+
+
+#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)
+{
+#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)
+
+#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)
+
+    /* 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;
+        }
+        while (len > 0) {
+            PROCESS1_64;
+            --len;
+        }
+         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);
+        }
+    }
+    /* 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)
+{
+    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;
+    }
+#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;
+
+        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 + PRIME64_5;
+    }
+
+    h64 += (xxh_u64) len;
+
+    return XXH64_finalize(h64, input, len, align);
+}
+
+
+XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t len, XXH64_hash_t seed)
+{
+#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);
+
+#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   *******/
+
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
+{
+    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
+}
+XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dstState, const XXH64_state_t* srcState)
+{
+    memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t seed)
+{
+    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;
+}
+
+XXH_PUBLIC_API XXH_errorcode
+XXH64_update (XXH64_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;
+#endif
+
+    {   const xxh_u8* p = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = p + len;
+
+        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->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;
+        }
+
+        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;
+
+            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);
+
+            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);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+XXH_PUBLIC_API XXH64_hash_t XXH64_digest (const XXH64_state_t* state)
+{
+    xxh_u64 h64;
+
+    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;
+
+    return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
+}
+
+
+/******* Canonical representation   *******/
+
+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));
+}
+
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src)
+{
+    return XXH_readBE64(src);
+}
+
+
+
+/* *********************************************************************
+*  XXH3
+*  New generation hash designed for speed on small keys and vectorization
+************************************************************************ */
+
+#include "xxh3.h"
+
+
+#endif  /* XXH_NO_LONG_LONG */
+
+
+#endif  /* XXH_IMPLEMENTATION */
+
+
+#if defined (__cplusplus)
+}
+#endif