paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-04-21 11:29:24 +02:00
Code Issues Packages Projects Releases Wiki Activity

* improve copy_avx gencopy mode

Browse Source 
git-svn-id: trunk@45486 -
This commit is contained in:
florian 2020-05-24 17:52:07 +00:00
parent 93eb40b05c
commit 88c98a5741
1 changed files with 32 additions and 48 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

80
compiler/x86/cgx86.pas
View File

@ -1554,7 +1554,7 @@ unit cgx86;
if GetRefAlignment(tmpref) = 16 then
op := A_VMOVDQA
else
op := A_VMOVDQU
op := A_VMOVDQU;
end
else
begin
@ -1570,7 +1570,7 @@ unit cgx86;
if GetRefAlignment(tmpref) = 32 then
op := A_VMOVDQA
else
op := A_VMOVDQU
op := A_VMOVDQU;
end
else
{ SSE doesn't support 256-bit vectors }
@ -1582,7 +1582,7 @@ unit cgx86;
if GetRefAlignment(tmpref) = 64 then
op := A_VMOVDQA
else
op := A_VMOVDQU
op := A_VMOVDQU;
end
else
{ SSE doesn't support 512-bit vectors }
@ -1648,13 +1648,13 @@ unit cgx86;
if GetRefAlignment(tmpref) = 16 then
op := A_VMOVDQA
else
op := A_VMOVDQU
op := A_VMOVDQU;
end else
begin
if GetRefAlignment(tmpref) = 16 then
op := A_MOVDQA
else
op := A_MOVDQU
op := A_MOVDQU;
end;
OS_M256:
{ Use XMM integer transfer }
@ -1663,7 +1663,7 @@ unit cgx86;
if GetRefAlignment(tmpref) = 32 then
op := A_VMOVDQA
else
op := A_VMOVDQU
op := A_VMOVDQU;
end else
{ SSE doesn't support 256-bit vectors }
InternalError(2018012942);
@ -1674,7 +1674,7 @@ unit cgx86;
if GetRefAlignment(tmpref) = 64 then
op := A_VMOVDQA
else
op := A_VMOVDQU
op := A_VMOVDQU;
end else
{ SSE doesn't support 512-bit vectors }
InternalError(2018012945);
@ -2703,6 +2703,7 @@ unit cgx86;
cgsize:Tcgsize;
cm:copymode;
saved_ds,saved_es: Boolean;
hlist: TAsmList;
begin
srcref:=source;
@ -2926,54 +2927,37 @@ unit cgx86;
copy_avx:
begin
r0:=NR_NO;
r1:=NR_NO;
r2:=NR_NO;
r3:=NR_NO;
if len>=16 then
hlist:=TAsmList.create;
while (len>=32) and (srcref.alignment>=32) and (dstref.alignment>=32) do
begin
r0:=getmmregister(list,OS_M256);
a_loadmm_ref_reg(list,OS_M256,OS_M256,srcref,r0,nil);
a_loadmm_reg_ref(hlist,OS_M256,OS_M256,r0,dstref,nil);
inc(srcref.offset,32);
inc(dstref.offset,32);
dec(len,32);
Include(current_procinfo.flags,pi_uses_ymm);
end;
while len>=16 do
begin
r0:=getmmregister(list,OS_M128);
{ we want to force the use of vmovups, so do not use a_loadmm_ref_reg }
list.concat(taicpu.op_ref_reg(A_VMOVUPS,S_NO,srcref,r0));
a_loadmm_ref_reg(list,OS_M128,OS_M128,srcref,r0,nil);
a_loadmm_reg_ref(hlist,OS_M128,OS_M128,r0,dstref,nil);
inc(srcref.offset,16);
end;
if len>=32 then
begin
r1:=getmmregister(list,OS_M128);
list.concat(taicpu.op_ref_reg(A_VMOVUPS,S_NO,srcref,r1));
inc(srcref.offset,16);
end;
if len>=48 then
begin
r2:=getmmregister(list,OS_M128);
list.concat(taicpu.op_ref_reg(A_VMOVUPS,S_NO,srcref,r2));
inc(srcref.offset,16);
end;
if (len=8) or (len=24) or (len=40) then
begin
r3:=getmmregister(list,OS_M64);
list.concat(taicpu.op_ref_reg(A_VMOVSD,S_NO,srcref,r3));
end;
if len>=16 then
begin
list.concat(taicpu.op_reg_ref(A_VMOVUPS,S_NO,r0,dstref));
inc(dstref.offset,16);
dec(len,16);
end;
if len>=32 then
if len>=8 then
begin
list.concat(taicpu.op_reg_ref(A_VMOVUPS,S_NO,r1,dstref));
inc(dstref.offset,16);
end;
if len>=48 then
begin
list.concat(taicpu.op_reg_ref(A_VMOVUPS,S_NO,r2,dstref));
inc(dstref.offset,16);
end;
if (len=8) or (len=24) or (len=40) then
begin
list.concat(taicpu.op_reg_ref(A_VMOVSD,S_NO,r3,dstref));
r0:=getmmregister(list,OS_M64);
a_loadmm_ref_reg(list,OS_M64,OS_M64,srcref,r0,nil);
a_loadmm_reg_ref(hlist,OS_M64,OS_M64,r0,dstref,nil);
inc(srcref.offset,8);
inc(dstref.offset,8);
dec(len,8);
end;
list.concatList(hlist);
hlist.free;
end
else {copy_string, should be a good fallback in case of unhandled}
begin