* support for the individual alignment flags for the memcpy intrinsic of

    LLVM 7.0

git-svn-id: branches/debug_eh@41728 -

compiler/llvm/hlcgllvm.pas

@@ -164,7 +164,7 @@ implementation
     verbose,cutils,globals,fmodule,constexp,systems,
     defutil,llvmdef,llvmsym,
     aasmtai,aasmcpu,
-    aasmllvm,llvmbase,tgllvm,
+    aasmllvm,llvmbase,llvminfo,tgllvm,
     symtable,symllvm,
     paramgr,
     pass_2,procinfo,llvmpi,cpuinfo,cgobj,cgllvm,cghlcpu,
@@ -1169,6 +1169,7 @@ implementation
       pd: tprocdef;
       sourcepara, destpara, sizepara, alignpara, volatilepara: tcgpara;
       maxalign: longint;
+      indivalign: boolean;
     begin
       { perform small copies directly; not larger ones, because then llvm
         will try to load the entire large datastructure into registers and
@@ -1180,7 +1181,11 @@ implementation
           a_load_ref_ref(list,size,size,source,dest);
           exit;
         end;
-      pd:=search_system_proc('llvm_memcpy64');
+      indivalign:=llvmflag_memcpy_indiv_align in llvmversion_properties[current_settings.llvmversion];
+      if indivalign then
+        pd:=search_system_proc('llvm_memcpy64_indivalign')
+      else
+        pd:=search_system_proc('llvm_memcpy64');
       sourcepara.init;
       destpara.init;
       sizepara.init;
@@ -1189,15 +1194,27 @@ implementation
       paramanager.getintparaloc(list,pd,1,destpara);
       paramanager.getintparaloc(list,pd,2,sourcepara);
       paramanager.getintparaloc(list,pd,3,sizepara);
-      paramanager.getintparaloc(list,pd,4,alignpara);
-      paramanager.getintparaloc(list,pd,5,volatilepara);
+      if indivalign then
+        begin
+          paramanager.getintparaloc(list,pd,4,volatilepara);
+          destpara.Alignment:=dest.alignment;
+          sourcepara.Alignment:=source.alignment;
+        end
+      else
+        begin
+          paramanager.getintparaloc(list,pd,4,alignpara);
+          paramanager.getintparaloc(list,pd,5,volatilepara);
+          maxalign:=newalignment(max(source.alignment,dest.alignment),min(source.alignment,dest.alignment));
+          a_load_const_cgpara(list,u32inttype,maxalign,alignpara);
+        end;
       a_loadaddr_ref_cgpara(list,size,dest,destpara);
       a_loadaddr_ref_cgpara(list,size,source,sourcepara);
       a_load_const_cgpara(list,u64inttype,size.size,sizepara);
-      maxalign:=newalignment(max(source.alignment,dest.alignment),min(source.alignment,dest.alignment));
-      a_load_const_cgpara(list,u32inttype,maxalign,alignpara);
       a_load_const_cgpara(list,llvmbool1type,ord((vol_read in source.volatility) or (vol_write in dest.volatility)),volatilepara);
-      g_call_system_proc(list,pd,[@destpara,@sourcepara,@sizepara,@alignpara,@volatilepara],nil).resetiftemp;
+      if indivalign then
+        g_call_system_proc(list,pd,[@destpara,@sourcepara,@sizepara,@volatilepara],nil).resetiftemp
+      else
+        g_call_system_proc(list,pd,[@destpara,@sourcepara,@sizepara,@alignpara,@volatilepara],nil).resetiftemp;
       sourcepara.done;
       destpara.done;
       sizepara.done;

compiler/llvm/llvminfo.pas

@@ -70,7 +70,8 @@ type
      llvmflag_load_getelptr_type,  { the return type of loads and the base type of getelementptr must be specified }
      llvmflag_call_no_ptr,         { with direct calls, the function type is not a function pointer }
      llvmflag_alias_double_type,   { with "alias" declarations, have to print both aliasee and aliasee* types }
-     llvmflag_fembed_bitcode       { support embedding bitcode in object files }
+     llvmflag_fembed_bitcode,      { support embedding bitcode in object files }
+     llvmflag_memcpy_indiv_align   { memcpy intrinsic supports separate alignment for source and dest }
    );
    tllvmversionflags = set of tllvmversionflag;
 
@@ -129,7 +130,7 @@ Const
        { llvmver_xc_9_2  } [llvmflag_load_getelptr_type,llvmflag_call_no_ptr,llvmflag_alias_double_type,llvmflag_fembed_bitcode],
        { llvmver_6_0     } [llvmflag_load_getelptr_type,llvmflag_call_no_ptr,llvmflag_alias_double_type,llvmflag_fembed_bitcode],
        { llvmver_xc_10_0 } [llvmflag_load_getelptr_type,llvmflag_call_no_ptr,llvmflag_alias_double_type,llvmflag_fembed_bitcode],
-       { llvmver_7_0     } [llvmflag_load_getelptr_type,llvmflag_call_no_ptr,llvmflag_alias_double_type,llvmflag_fembed_bitcode]
+       { llvmver_7_0     } [llvmflag_load_getelptr_type,llvmflag_call_no_ptr,llvmflag_alias_double_type,llvmflag_fembed_bitcode,llvmflag_memcpy_indiv_align]
      );
 
    { Supported optimizations, only used for information }

rtl/inc/llvmintr.inc

@@ -15,6 +15,8 @@
 
 
 procedure llvm_memcpy64(dest, source: pointer; len: qword; align: cardinal; isvolatile: LLVMBool1); compilerproc; external name 'llvm.memcpy.p0i8.p0i8.i64';
+{ LLVM 7.0+, alignment specified as part of the dest and source parameters }
+procedure llvm_memcpy64_indivalign(dest, source: pointer; len: qword; isvolatile: LLVMBool1); compilerproc; external name 'llvm.memcpy.p0i8.p0i8.i64';
 
 function llvm_frameaddress(level: longint): pointer; compilerproc; external name 'llvm.frameaddress';