paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-09-15 20:49:17 +02:00
Code Issues Packages Projects Releases Wiki Activity

* x86: New "shr %cl; shr x" swap optimisation

Browse Source
This commit is contained in:
J. Gareth "Curious Kit" Moreton 2022-08-16 03:40:29 +01:00 committed by FPK
parent 6757bf3832
commit b872322678
1 changed files with 26 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
compiler/x86/aoptx86.pas
View File

@ -13090,7 +13090,7 @@ unit aoptx86;
hp1 := p;
repeat
if not GetNextInstructionUsingReg(hp1, hp1, taicpu(p).oper[1]^.reg) or (hp1.typ <> ait_instruction) then
Exit;
Break;
{ Detect:
shr x, %reg
@ -13235,7 +13235,7 @@ unit aoptx86;
end;
if Shift < topsize2memsize[taicpu(p).opsize] - topsize2memsize[LimitSize] then
Exit;
Break;
{ Since we've established that the combined shift is within
limits, we can actually combine the adjacent SHR
@ -13259,6 +13259,30 @@ unit aoptx86;
Break;
until False;
{ Detect the following (looking backwards):
shr %cl,%reg
shr x, %reg
Swap the two SHR instructions to minimise a pipeline stall.
}
if GetLastInstruction(p, hp1) and
MatchInstruction(hp1, A_SHR, [taicpu(p).opsize]) and
MatchOpType(taicpu(hp1), top_reg, top_reg) and
{ First operand will be %cl }
(taicpu(hp1).oper[1]^.reg = taicpu(p).oper[1]^.reg) and
{ Just to be sure }
(getsupreg(taicpu(hp1).oper[1]^.reg) <> RS_ECX) then
begin
DebugMsg(SPeepholeOptimization + 'Swapped variable and constant SHR instructions to minimise pipeline stall (ShrShr2ShrShr)', hp1);
{ Moving the entries this way ensures the register tracking remains correct }
Asml.Remove(p);
Asml.InsertBefore(p, hp1);
p := hp1;
{ Don't set Result to True because the current instruction is now
"shr %cl,%reg" and there's nothing more we can do with it }
end;
end;