paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-08-13 09:39:09 +02:00
Code Issues Packages Projects Releases Wiki Activity

* x86-64: Extended the movl/movq optimisation to cover more distance

Browse Source
This commit is contained in:
J. Gareth "Curious Kit" Moreton 2022-11-22 01:03:12 +00:00 committed by FPK
parent 42b052eb83
commit f3f9c68ddb
1 changed files with 93 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

122
compiler/x86/aoptx86.pas
View File

@ -4125,13 +4125,13 @@ unit aoptx86;
if (taicpu(p).oper[0]^.reg = taicpu(hp1).oper[1]^.reg) then if (taicpu(p).oper[0]^.reg = taicpu(hp1).oper[1]^.reg) then
begin begin
{ %reg1 = %reg3 } { %reg1 = %reg3 }
DebugMsg(SPeepholeOptimization + 'Made 32-to-64-bit zero extension more efficient (MovlMovq2MovlAndl)', hp1); DebugMsg(SPeepholeOptimization + 'Made 32-to-64-bit zero extension more efficient (MovlMovq2MovlAndl 1)', hp1);
taicpu(hp1).opcode := A_AND; taicpu(hp1).opcode := A_AND;
end end
else else
begin begin
{ %reg1 <> %reg3 } { %reg1 <> %reg3 }
DebugMsg(SPeepholeOptimization + 'Made 32-to-64-bit zero extension more efficient (MovlMovq2MovlMovl)', hp1); DebugMsg(SPeepholeOptimization + 'Made 32-to-64-bit zero extension more efficient (MovlMovq2MovlMovl 1)', hp1);
end; end;
if not RegUsedAfterInstruction(taicpu(p).oper[1]^.reg, hp1, TmpUsedRegs) then if not RegUsedAfterInstruction(taicpu(p).oper[1]^.reg, hp1, TmpUsedRegs) then
@ -4332,35 +4332,99 @@ unit aoptx86;
Internalerror(2019103001); Internalerror(2019103001);
end; end;
end end
else else if MatchOperand(taicpu(hp2).oper[1]^, p_TargetReg) then
if MatchOperand(taicpu(hp2).oper[1]^, p_TargetReg) then begin
begin if not CrossJump and
if not CrossJump and not RegUsedBetween(p_TargetReg, p, hp2) and
not RegUsedBetween(p_TargetReg, p, hp2) and not RegReadByInstruction(p_TargetReg, hp2) then
not RegReadByInstruction(p_TargetReg, hp2) then begin
begin { Register is not used before it is overwritten }
{ Register is not used before it is overwritten } DebugMsg(SPeepholeOptimization + 'Mov2Nop 3a done',p);
DebugMsg(SPeepholeOptimization + 'Mov2Nop 3a done',p); RemoveCurrentp(p, hp1);
RemoveCurrentp(p, hp1); Result := True;
Result := True; Exit;
Exit; end;
end;
if (taicpu(p).oper[0]^.typ = top_const) and if (taicpu(p).oper[0]^.typ = top_const) and
(taicpu(hp2).oper[0]^.typ = top_const) then (taicpu(hp2).oper[0]^.typ = top_const) then
begin begin
if taicpu(p).oper[0]^.val = taicpu(hp2).oper[0]^.val then if taicpu(p).oper[0]^.val = taicpu(hp2).oper[0]^.val then
begin begin
{ Same value - register hasn't changed } { Same value - register hasn't changed }
DebugMsg(SPeepholeOptimization + 'Mov2Nop 2 done', hp2); DebugMsg(SPeepholeOptimization + 'Mov2Nop 2 done', hp2);
RemoveInstruction(hp2); RemoveInstruction(hp2);
Result := True; Result := True;
{ See if there's more we can optimise }
Continue;
end;
end;
{$ifdef x86_64}
end
{ Change:
movl %reg1l,%reg2l
...
movq %reg2q,%reg3q (%reg1 <> %reg3)
To:
movl %reg1l,%reg2l
...
movl %reg1l,%reg3l (Upper 32 bits of %reg3q will be zero)
If %reg1 = %reg3, convert to:
movl %reg1l,%reg2l
...
andl %reg1l,%reg1l
}
else if (taicpu(p).opsize = S_L) and MatchInstruction(hp2,A_MOV,[S_Q]) and
(taicpu(p).oper[0]^.typ = top_reg) and
MatchOpType(taicpu(hp2), top_reg, top_reg) and
SuperRegistersEqual(p_TargetReg, taicpu(hp2).oper[0]^.reg) and
not RegModifiedBetween(p_TargetReg, p, hp2) then
begin
TempRegUsed :=
CrossJump { Assume the register is in use if it crossed a conditional jump } or
RegReadByInstruction(p_TargetReg, hp3) or
RegUsedAfterInstruction(p_TargetReg, hp2, TmpUsedRegs);
taicpu(hp2).opsize := S_L;
taicpu(hp2).loadreg(0, taicpu(p).oper[0]^.reg);
setsubreg(taicpu(hp2).oper[1]^.reg, R_SUBD);
AllocRegBetween(taicpu(p).oper[0]^.reg, p, hp2, UsedRegs);
if (taicpu(p).oper[0]^.reg = taicpu(hp2).oper[1]^.reg) then
begin
{ %reg1 = %reg3 }
DebugMsg(SPeepholeOptimization + 'Made 32-to-64-bit zero extension more efficient (MovlMovq2MovlAndl 2)', hp2);
taicpu(hp2).opcode := A_AND;
end
else
begin
{ %reg1 <> %reg3 }
DebugMsg(SPeepholeOptimization + 'Made 32-to-64-bit zero extension more efficient (MovlMovq2MovlMovl 2)', hp2);
end;
if not TempRegUsed then
begin
DebugMsg(SPeepholeOptimization + 'Mov2Nop 8a done', p);
RemoveCurrentP(p, hp1);
Result := True;
Exit;
end
else
begin
{ Initial instruction wasn't actually changed }
Include(OptsToCheck, aoc_ForceNewIteration);
{ if %reg1 = %reg3, don't do the long-distance lookahead that
appears below since %reg1 has technically changed }
if taicpu(hp2).opcode = A_AND then
Break;
end;
{$endif x86_64}
end;
{ See if there's more we can optimise }
Continue;
end;
end;
end;
A_MOVZX, A_MOVSX{$ifdef x86_64}, A_MOVSXD{$endif x86_64}: A_MOVZX, A_MOVSX{$ifdef x86_64}, A_MOVSXD{$endif x86_64}:
if MatchOpType(taicpu(hp2), top_reg, top_reg) and if MatchOpType(taicpu(hp2), top_reg, top_reg) and
MatchOperand(taicpu(hp2).oper[0]^, p_TargetReg) and MatchOperand(taicpu(hp2).oper[0]^, p_TargetReg) and