paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-04-14 07:59:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'mov-const' into 'main'

Browse Source 
[x86] MOV constant manipulation

See merge request freepascal.org/fpc/source!181
This commit is contained in:
J. Gareth "Kit" Moreton 2025-04-02 10:27:47 +00:00
commit 1d3083a41c
1 changed files with 112 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

116
compiler/x86/aoptx86.pas
View File

@ -3223,6 +3223,43 @@ unit aoptx86;
MovAligned, MovUnaligned: TAsmOp;
ThisRef: TReference;
JumpTracking: TLinkedList;
OldVal, NewVal: TCGInt;
X, Range: Integer;
{ Nested function depends on NewVal and hp2. Use local parameters to
minimise reads from the stack }
function CheckConstantShifts(LocalOld: TCGInt; Shift: Integer): Boolean;
begin
Result := False;
if (LocalOld shl Shift) = NewVal then
begin
DebugMsg(SPeepholeOptimization + tostr(NewVal) + ' = ' + tostr(LocalOld) + ' shl ' + tostr(Shift) + '; changed to reduce code size without speed loss (Mov2Shl)', hp2);
taicpu(hp2).opcode := A_SHL;
taicpu(hp2).oper[0]^.val := Shift;
Result := True;
end
else if (LocalOld shr Shift) = NewVal then
begin
DebugMsg(SPeepholeOptimization + tostr(NewVal) + ' = ' + tostr(LocalOld) + ' shr ' + tostr(Shift) + '; changed to reduce code size without speed loss (Mov2Shr)', hp2);
taicpu(hp2).opcode := A_SHR;
taicpu(hp2).oper[0]^.val := Shift;
Result := True;
end
else if SarInt64(LocalOld, Shift) = NewVal then
begin
DebugMsg(SPeepholeOptimization + tostr(NewVal) + ' = ' + tostr(LocalOld) + ' sar ' + tostr(Shift) + '; changed to reduce code size without speed loss (Mov2Sar)', hp2);
taicpu(hp2).opcode := A_SAR;
taicpu(hp2).oper[0]^.val := Shift;
Result := True;
end;
end;
begin
Result:=false;
@ -4880,20 +4917,91 @@ unit aoptx86;
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
TransferUsedRegs(TmpUsedRegs);
hp3 := p;
{ Make sure the flags register is properly tracked up to hp2 }
repeat
TmpUsedRegs[R_SPECIALREGISTER].Update(tai(hp3.Next));
if not GetNextInstruction(hp3, hp3) then
InternalError(2021091301);
until hp3 = hp2;
OldVal := taicpu(p).oper[0]^.val;
NewVal := taicpu(hp2).oper[0]^.val;
if OldVal = NewVal then
begin
{ Same value - register hasn't changed }
DebugMsg(SPeepholeOptimization + 'Mov2Nop 2 done', hp2);
RemoveInstruction(hp2);
Include(OptsToCheck, aoc_ForceNewIteration);
{ See if there's more we can optimise }
Continue;
end;
if (
{ Only do the MOV-to-arithmetic optimisation under
-Os or on very old processors }
(cs_opt_size in current_settings.optimizerswitches)
{$ifndef x86_64}
or (current_settings.cputype <= cpu_486)
{$endif not x86_64}
) and
{ mov 0,%reg can be efficiently encoded as xor %reg,%reg later }
(NewVal <> 0) and not RegInUsedRegs(NR_DEFAULTFLAGS, TmpUsedRegs) then
begin
Range := tcgsize2size[reg_cgsize(p_TargetReg)] * 4; { Only half of the range }
if Range > 8 then { Only a saving for 32-bit and 64-bit really }
begin
{ For the first half of the bit range, check for SHL, SHR, SAR, ROL and ROR permutations }
for X := 1 to Range - 1 do
begin
{ Check SHL, SHR and SAR }
if CheckConstantShifts(OldVal, X) then
begin
Result := True;
Exit;
end;
if RolQWord(QWord(OldVal), X) = NewVal then
begin
DebugMsg(SPeepholeOptimization + tostr(NewVal) + ' = ' + tostr(OldVal) + ' rol ' + tostr(X) + '; changed to reduce code size without speed loss (Mov2Rol)', hp2);
taicpu(hp2).opcode := A_ROL;
taicpu(hp2).oper[0]^.val := X;
Result := True;
Exit;
end;
if RorQWord(QWord(OldVal), X) = NewVal then
begin
DebugMsg(SPeepholeOptimization + tostr(NewVal) + ' = ' + tostr(OldVal) + ' ror ' + tostr(X) + '; changed to reduce code size (Mov2Ror)', hp2);
taicpu(hp2).opcode := A_ROR;
taicpu(hp2).oper[0]^.val := X;
Result := True;
Exit;
end;
end;
{ Don't check for ROR and ROL for the remainder of the range, since the
instructions overlap in the first half (e.g. rorq $33,%reg = rolq $31,%reg }
Range := Range shl 1;
for X := (Range shr 2) to Range - 1 do
{ Check SHL, SHR and SAR }
if CheckConstantShifts(OldVal, X) then
begin
Result := True;
Exit;
end;
end;
end;
end;
{$ifdef x86_64}
end
{$ifdef x86_64}
{ Change:
movl %reg1l,%reg2l
...
@ -4955,8 +5063,8 @@ unit aoptx86;
if taicpu(hp2).opcode = A_AND then
Break;
end;
{$endif x86_64}
end
{$endif x86_64}
else if (taicpu(hp2).oper[0]^.typ = top_ref) and
GetNextInstruction(hp2, hp4) and
(hp4.typ = ait_instruction) and (taicpu(hp4).opcode = A_MOV) then