paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-09-10 13:49:18 +02:00
Code Issues Packages Projects Releases Wiki Activity

x86: MOV constant optimisation

Browse Source
This commit is contained in:
J. Gareth "Kit" Moreton 2021-09-18 08:30:50 +00:00 committed by FPK
parent 33e25a3dfe
commit 962c5bd727
2 changed files with 49 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

62
compiler/x86/aoptx86.pas
View File

@ -2211,7 +2211,7 @@ unit aoptx86;
GetNextInstruction_p, TempRegUsed, CrossJump: Boolean; GetNextInstruction_p, TempRegUsed, CrossJump: Boolean;
PreMessage, RegName1, RegName2, InputVal, MaskNum: string; PreMessage, RegName1, RegName2, InputVal, MaskNum: string;
NewSize: topsize; NewSize: topsize;
CurrentReg: TRegister; CurrentReg, ActiveReg: TRegister;
begin begin
Result:=false; Result:=false;
@ -3019,8 +3019,8 @@ unit aoptx86;
{ mov x,reg1; mov y,reg1 -> mov y,reg1 is handled by the Mov2Nop 5 optimisation } { mov x,reg1; mov y,reg1 -> mov y,reg1 is handled by the Mov2Nop 5 optimisation }
end; end;
{ search further than the next instruction for a mov } { search further than the next instruction for a mov (as long as it's not a jump) }
if if not is_calljmpuncond(taicpu(hp1).opcode) and
{ check as much as possible before the expensive GetNextInstructionUsingRegCond call } { check as much as possible before the expensive GetNextInstructionUsingRegCond call }
(taicpu(p).oper[1]^.typ = top_reg) and (taicpu(p).oper[1]^.typ = top_reg) and
(taicpu(p).oper[0]^.typ in [top_reg,top_const]) and (taicpu(p).oper[0]^.typ in [top_reg,top_const]) and
@ -3031,15 +3031,18 @@ unit aoptx86;
hp3 := hp1; hp3 := hp1;
{ Initialise CrossJump (if it becomes True at any point, it will remain True) } { Initialise CrossJump (if it becomes True at any point, it will remain True) }
CrossJump := False; CrossJump := (taicpu(hp1).opcode = A_Jcc);
while GetNextInstructionUsingRegCond(hp3,hp2,taicpu(p).oper[1]^.reg,CrossJump) and { Saves on a large number of dereferences }
ActiveReg := taicpu(p).oper[1]^.reg;
while GetNextInstructionUsingRegCond(hp3,hp2,ActiveReg,CrossJump) and
{ GetNextInstructionUsingRegCond only searches one instruction ahead unless -O3 is specified } { GetNextInstructionUsingRegCond only searches one instruction ahead unless -O3 is specified }
(hp2.typ=ait_instruction) do (hp2.typ=ait_instruction) do
begin begin
case taicpu(hp2).opcode of case taicpu(hp2).opcode of
A_MOV: A_MOV:
if MatchOperand(taicpu(hp2).oper[0]^,taicpu(p).oper[1]^.reg) and if MatchOperand(taicpu(hp2).oper[0]^,ActiveReg) and
((taicpu(p).oper[0]^.typ=top_const) or ((taicpu(p).oper[0]^.typ=top_const) or
((taicpu(p).oper[0]^.typ=top_reg) and ((taicpu(p).oper[0]^.typ=top_reg) and
not(RegModifiedBetween(taicpu(p).oper[0]^.reg, p, hp2)) not(RegModifiedBetween(taicpu(p).oper[0]^.reg, p, hp2))
@ -3061,8 +3064,8 @@ unit aoptx86;
TempRegUsed := TempRegUsed :=
CrossJump { Assume the register is in use if it crossed a conditional jump } or CrossJump { Assume the register is in use if it crossed a conditional jump } or
RegUsedAfterInstruction(taicpu(p).oper[1]^.reg, hp2, TmpUsedRegs) or RegUsedAfterInstruction(ActiveReg, hp2, TmpUsedRegs) or
RegReadByInstruction(taicpu(p).oper[1]^.reg, hp1); RegReadByInstruction(ActiveReg, hp1);
case taicpu(p).oper[0]^.typ Of case taicpu(p).oper[0]^.typ Of
top_reg: top_reg:
@ -3165,11 +3168,40 @@ unit aoptx86;
else else
Internalerror(2019103001); Internalerror(2019103001);
end; end;
end; end
else
if MatchOperand(taicpu(hp2).oper[1]^, ActiveReg) then
begin
if not CrossJump and
not RegUsedBetween(ActiveReg, p, hp2) and
not RegReadByInstruction(ActiveReg, hp2) then
begin
{ Register is not used before it is overwritten }
DebugMsg(SPeepholeOptimization + 'Mov2Nop 3a done',p);
RemoveCurrentp(p, hp1);
Result := True;
Exit;
end;
if (taicpu(p).oper[0]^.typ = top_const) and
(taicpu(hp2).oper[0]^.typ = top_const) then
begin
if taicpu(p).oper[0]^.val = taicpu(hp2).oper[0]^.val then
begin
{ Same value - register hasn't changed }
DebugMsg(SPeepholeOptimization + 'Mov2Nop 2 done', hp2);
RemoveInstruction(hp2);
Result := True;
{ 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]^, taicpu(p).oper[1]^.reg) and MatchOperand(taicpu(hp2).oper[0]^, ActiveReg) and
SuperRegistersEqual(taicpu(hp2).oper[1]^.reg, taicpu(p).oper[1]^.reg) then SuperRegistersEqual(taicpu(hp2).oper[1]^.reg, ActiveReg) then
begin begin
{ {
Change from: Change from:
@ -3204,7 +3236,6 @@ unit aoptx86;
else else
if MatchOpType(taicpu(p), top_reg, top_reg) then if MatchOpType(taicpu(p), top_reg, top_reg) then
begin begin
CurrentReg := taicpu(p).oper[1]^.reg;
TransferUsedRegs(TmpUsedRegs); TransferUsedRegs(TmpUsedRegs);
TmpUsedRegs[R_INTREGISTER].Update(tai(p.Next)); TmpUsedRegs[R_INTREGISTER].Update(tai(p.Next));
if if
@ -3214,7 +3245,7 @@ unit aoptx86;
begin begin
{ Just in case something didn't get modified (e.g. an { Just in case something didn't get modified (e.g. an
implicit register) } implicit register) }
if not RegReadByInstruction(CurrentReg, hp2) and if not RegReadByInstruction(ActiveReg, hp2) and
{ If a conditional jump was crossed, do not delete { If a conditional jump was crossed, do not delete
the original MOV no matter what } the original MOV no matter what }
not CrossJump then not CrossJump then
@ -3226,9 +3257,10 @@ unit aoptx86;
if if
{ Make sure the original register isn't still present { Make sure the original register isn't still present
and has been written to (e.g. with SHRX) } and has been written to (e.g. with SHRX) }
RegLoadedWithNewValue(CurrentReg, hp2) or RegLoadedWithNewValue(ActiveReg, hp2) or
not RegUsedAfterInstruction(CurrentReg, hp2, TmpUsedRegs) then not RegUsedAfterInstruction(ActiveReg, hp2, TmpUsedRegs) then
begin begin
RegUsedAfterInstruction(ActiveReg, hp2, TmpUsedRegs);
{ We can remove the original MOV } { We can remove the original MOV }
DebugMsg(SPeepholeOptimization + 'Mov2Nop 3b done',p); DebugMsg(SPeepholeOptimization + 'Mov2Nop 3b done',p);
RemoveCurrentp(p, hp1); RemoveCurrentp(p, hp1);

2
compiler/x86/cpubase.pas
View File

@ -570,6 +570,7 @@ implementation
A_LOOPZ, A_LOOPZ,
A_LCALL, A_LCALL,
A_LJMP, A_LJMP,
A_RET,
A_Jcc : A_Jcc :
is_calljmp:=true; is_calljmp:=true;
else else
@ -583,6 +584,7 @@ implementation
case o of case o of
A_CALL, A_CALL,
A_JMP, A_JMP,
A_RET,
A_LCALL, A_LCALL,
A_LJMP: A_LJMP:
is_calljmpuncond:=true; is_calljmpuncond:=true;