paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-08-20 09:49:08 +02:00
Code Issues Packages Projects Releases Wiki Activity

* x86: Reversal of DoAddRefOpt optimisation if no pipeline saving is made

Browse Source
This commit is contained in:
J. Gareth "Curious Kit" Moreton 2022-06-16 19:44:11 +01:00
parent 45883c2cc0
commit 1d86db3100
1 changed files with 200 additions and 118 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

84
compiler/x86/aoptx86.pas
View File

@ -157,6 +157,7 @@ unit aoptx86;
function DoArithCombineOpt(var p : tai) : Boolean;
function DoAddRefOpt(var p, hp1: tai; Reg: TRegister; Value: TCGInt): Boolean;
function DoReverseAddRefOpt(var p: tai; Value: TCGInt): Boolean;
function DoMovCmpMemOpt(var p : tai; const hp1: tai) : Boolean;
function DoSETccLblRETOpt(var p: tai; const hp_label: tai_label) : Boolean;
@ -6449,6 +6450,73 @@ unit aoptx86;
end;
function TX86AsmOptimizer.DoReverseAddRefOpt(var p: tai; Value: TCGInt): Boolean;
var
CurrentRef: PReference;
OtherRegister: TRegister;
X: Integer;
hp1, hp2: tai;
begin
{ Assume that p has been checked and confirmed to be of the form
"ADD/SUB const,%reg" }
Result := False;
if GetLastInstruction(p, hp1) and (hp1.typ = ait_instruction) and
{ Make sure this instruction doesn't also modify the register used in
the ADD/SUB instruction }
not RegModifiedByInstruction(taicpu(p).oper[1]^.reg, hp1) then
begin
{ Find the reference }
for X := 0 to taicpu(Hp1).ops - 1 do
if (taicpu(hp1).oper[X]^.typ = top_ref) then
begin
{ Locally store the pointer to the reference }
CurrentRef := taicpu(hp1).oper[X]^.ref;
if
{ Only references of the form x(%reg1,%reg2,scale) can be
optimised here }
(CurrentRef^.refaddr <> addr_no) or
not RegInRef(taicpu(p).oper[1]^.reg, CurrentRef^) then
Exit;
{ Account for the scale factor on the value }
if SuperRegistersEqual(taicpu(p).oper[1]^.reg, CurrentRef^.index) then
begin
OtherRegister := CurrentRef^.base;
if SuperRegistersEqual(taicpu(p).oper[1]^.reg, CurrentRef^.base) then
Inc(Value, Value * max(CurrentRef^.scalefactor, 1))
else
Value := Value * max(CurrentRef^.scalefactor, 1);
end
else
OtherRegister := CurrentRef^.index;
if (OtherRegister <> NR_NO) and
{ If we can't set the offset to zero, this is wasted effort }
(CurrentRef^.offset = Value)
and GetLastInstruction(hp1, hp2) and
{ Make sure there is a pipeline stall between hp2 and hp1,
otherwise a saving won't be made }
RegModifiedByInstruction(OtherRegister, hp2) then
begin
taicpu(hp1).oper[X]^.ref^.offset := 0;
AsmL.Remove(hp1);
AsmL.InsertAfter(hp1, p);
{ In case OtherRegister got deallocated right after the reference }
AllocRegBetween(OtherRegister, hp2, hp1, UsedRegs);
DebugMsg(SPeepholeOptimization + 'Rearranged MOV; (ref); ADD/SUB to MOV; ADD/SUB; (ref) to remove offset and minimise cache pollution', hp2);
{ Keep p as the current instruction }
Result := True;
end;
end;
end;
end;
function TX86AsmOptimizer.DoArithCombineOpt(var p: tai): Boolean;
var
hp1 : tai;
@ -15495,6 +15563,19 @@ unit aoptx86;
begin
Result := False;
if taicpu(p).oper[0]^.typ = top_const then
begin
{ Sometimes, DoAddRefOpt makes an optimisation that doesn't
improve code speed and only increases cache pollution. If these
aren't cleared by other optimisations, rectify it here }
if taicpu(p).oper[1]^.typ = top_reg then
begin
if taicpu(p).opcode = A_ADD then
Result := DoReverseAddRefOpt(p, taicpu(p).oper[0]^.val)
else
Result := DoReverseAddRefOpt(p, -taicpu(p).oper[0]^.val);
end;
{ Change:
add/sub 128,(dest)
@ -15504,7 +15585,7 @@ unit aoptx86;
This generaally takes fewer bytes to encode because -128 can be stored
in a signed byte, whereas +128 cannot.
}
if (taicpu(p).opsize <> S_B) and MatchOperand(taicpu(p).oper[0]^, 128) then
if (taicpu(p).opsize <> S_B) and (taicpu(p).oper[0]^.val = 128) then
begin
if taicpu(p).opcode = A_ADD then
Opposite := A_SUB
@ -15621,6 +15702,7 @@ unit aoptx86;
Result := True;
Exit;
end;
end;
{ Detect:
add/sub %reg2,(dest)