paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-04-14 12:39:31 +02:00
Code Issues Packages Projects Releases Wiki Activity

* unified TX86AsmOptimizer.OptPass1MOVAP and TX86AsmOptimizer.OptPass1VMOVAP

Browse Source 
git-svn-id: trunk@43370 -
This commit is contained in:
florian 2019-11-02 18:48:23 +00:00
parent f9530437c2
commit 3487c34ed9
3 changed files with 121 additions and 71 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
compiler/i386/aoptcpu.pas
View File

@ -547,9 +547,15 @@ begin
A_SUB:
if OptPass1Sub(p) then
continue;
A_MOVAPD,
A_MOVAPS,
A_MOVUPD,
A_MOVUPS,
A_VMOVAPS,
A_VMOVAPD:
if OptPass1VMOVAP(p) then
A_VMOVAPD,
A_VMOVUPS,
A_VMOVUPD:
if OptPass1_V_MOVAP(p) then
continue;
A_VDIVSD,
A_VDIVSS,
@ -573,10 +579,6 @@ begin
A_ADDSS:
if OptPass1OP(p) then
continue;
A_MOVAPD,
A_MOVAPS:
if OptPass1MOVAP(p) then
continue;
A_VMOVSD,
A_VMOVSS,
A_MOVSD,

167
compiler/x86/aoptx86.pas
View File

@ -62,11 +62,10 @@ unit aoptx86;
function PrePeepholeOptIMUL(var p : tai) : boolean;
function OptPass1AND(var p : tai) : boolean;
function OptPass1VMOVAP(var p : tai) : boolean;
function OptPass1_V_MOVAP(var p : tai) : boolean;
function OptPass1VOP(var p : tai) : boolean;
function OptPass1MOV(var p : tai) : boolean;
function OptPass1Movx(var p : tai) : boolean;
function OptPass1MOVAP(var p : tai) : boolean;
function OptPass1MOVXX(var p : tai) : boolean;
function OptPass1OP(var p : tai) : boolean;
function OptPass1LEA(var p : tai) : boolean;
@ -1130,61 +1129,7 @@ unit aoptx86;
end;
function TX86AsmOptimizer.OptPass1MOVAP(var p : tai) : boolean;
var
hp1,hp2 : tai;
begin
result:=false;
if MatchOpType(taicpu(p),top_reg,top_reg) and
GetNextInstruction(p, hp1) and
(hp1.typ = ait_instruction) and
GetNextInstruction(hp1, hp2) and
MatchInstruction(hp2,taicpu(p).opcode,[]) and
OpsEqual(taicpu(hp2).oper[1]^, taicpu(p).oper[0]^) and
MatchOpType(taicpu(hp2),top_reg,top_reg) and
MatchOperand(taicpu(hp2).oper[0]^,taicpu(p).oper[1]^) and
(((taicpu(p).opcode=A_MOVAPS) and
((taicpu(hp1).opcode=A_ADDSS) or (taicpu(hp1).opcode=A_SUBSS) or
(taicpu(hp1).opcode=A_MULSS) or (taicpu(hp1).opcode=A_DIVSS))) or
((taicpu(p).opcode=A_MOVAPD) and
((taicpu(hp1).opcode=A_ADDSD) or (taicpu(hp1).opcode=A_SUBSD) or
(taicpu(hp1).opcode=A_MULSD) or (taicpu(hp1).opcode=A_DIVSD)))
) then
{ change
movapX reg,reg2
addsX/subsX/... reg3, reg2
movapX reg2,reg
to
addsX/subsX/... reg3,reg
}
begin
TransferUsedRegs(TmpUsedRegs);
UpdateUsedRegs(TmpUsedRegs, tai(p.next));
UpdateUsedRegs(TmpUsedRegs, tai(hp1.next));
If not(RegUsedAfterInstruction(taicpu(p).oper[1]^.reg,hp2,TmpUsedRegs)) then
begin
DebugMsg(SPeepholeOptimization + 'MovapXOpMovapX2Op ('+
debug_op2str(taicpu(p).opcode)+' '+
debug_op2str(taicpu(hp1).opcode)+' '+
debug_op2str(taicpu(hp2).opcode)+') done',p);
{ we cannot eliminate the first move if
the operations uses the same register for source and dest }
if not(OpsEqual(taicpu(hp1).oper[1]^,taicpu(hp1).oper[0]^)) then
begin
asml.remove(p);
p.Free;
end;
taicpu(hp1).loadoper(1, taicpu(hp2).oper[1]^);
asml.remove(hp2);
hp2.Free;
p:=hp1;
result:=true;
end;
end
end;
function TX86AsmOptimizer.OptPass1VMOVAP(var p : tai) : boolean;
function TX86AsmOptimizer.OptPass1_V_MOVAP(var p : tai) : boolean;
var
hp1,hp2 : tai;
begin
@ -1234,12 +1179,72 @@ unit aoptx86;
result:=true;
end
end
else if MatchInstruction(hp1,[A_VFMADD132PD,A_VFNMADD231SD,A_VFMADD231SD],[S_NO]) and
else if MatchInstruction(hp1,[A_VFMADDPD,
A_VFMADD132PD,
A_VFMADD132PS,
A_VFMADD132SD,
A_VFMADD132SS,
A_VFMADD213PD,
A_VFMADD213PS,
A_VFMADD213SD,
A_VFMADD213SS,
A_VFMADD231PD,
A_VFMADD231PS,
A_VFMADD231SD,
A_VFMADD231SS,
A_VFMADDSUB132PD,
A_VFMADDSUB132PS,
A_VFMADDSUB213PD,
A_VFMADDSUB213PS,
A_VFMADDSUB231PD,
A_VFMADDSUB231PS,
A_VFMSUB132PD,
A_VFMSUB132PS,
A_VFMSUB132SD,
A_VFMSUB132SS,
A_VFMSUB213PD,
A_VFMSUB213PS,
A_VFMSUB213SD,
A_VFMSUB213SS,
A_VFMSUB231PD,
A_VFMSUB231PS,
A_VFMSUB231SD,
A_VFMSUB231SS,
A_VFMSUBADD132PD,
A_VFMSUBADD132PS,
A_VFMSUBADD213PD,
A_VFMSUBADD213PS,
A_VFMSUBADD231PD,
A_VFMSUBADD231PS,
A_VFNMADD132PD,
A_VFNMADD132PS,
A_VFNMADD132SD,
A_VFNMADD132SS,
A_VFNMADD213PD,
A_VFNMADD213PS,
A_VFNMADD213SD,
A_VFNMADD213SS,
A_VFNMADD231PD,
A_VFNMADD231PS,
A_VFNMADD231SD,
A_VFNMADD231SS,
A_VFNMSUB132PD,
A_VFNMSUB132PS,
A_VFNMSUB132SD,
A_VFNMSUB132SS,
A_VFNMSUB213PD,
A_VFNMSUB213PS,
A_VFNMSUB213SD,
A_VFNMSUB213SS,
A_VFNMSUB231PD,
A_VFNMSUB231PS,
A_VFNMSUB231SD,
A_VFNMSUB231SS],[S_NO]) and
{ we mix single and double opperations here because we assume that the compiler
generates vmovapd only after double operations and vmovaps only after single operations }
MatchOperand(taicpu(p).oper[1]^,taicpu(hp1).oper[2]^) and
GetNextInstruction(hp1,hp2) and
MatchInstruction(hp2,A_VMOVAPD,A_VMOVAPS,[S_NO]) and
MatchInstruction(hp2,[A_VMOVAPD,A_VMOVAPS,A_MOVAPD,A_MOVAPS],[S_NO]) and
MatchOperand(taicpu(p).oper[0]^,taicpu(hp2).oper[1]^) then
begin
TransferUsedRegs(TmpUsedRegs);
@ -1255,6 +1260,50 @@ unit aoptx86;
hp2.Free;
p:=hp1;
end;
end
else if (hp1.typ = ait_instruction) and
GetNextInstruction(hp1, hp2) and
MatchInstruction(hp2,taicpu(p).opcode,[]) and
OpsEqual(taicpu(hp2).oper[1]^, taicpu(p).oper[0]^) and
MatchOpType(taicpu(hp2),top_reg,top_reg) and
MatchOperand(taicpu(hp2).oper[0]^,taicpu(p).oper[1]^) and
(((taicpu(p).opcode=A_MOVAPS) and
((taicpu(hp1).opcode=A_ADDSS) or (taicpu(hp1).opcode=A_SUBSS) or
(taicpu(hp1).opcode=A_MULSS) or (taicpu(hp1).opcode=A_DIVSS))) or
((taicpu(p).opcode=A_MOVAPD) and
((taicpu(hp1).opcode=A_ADDSD) or (taicpu(hp1).opcode=A_SUBSD) or
(taicpu(hp1).opcode=A_MULSD) or (taicpu(hp1).opcode=A_DIVSD)))
) then
{ change
movapX reg,reg2
addsX/subsX/... reg3, reg2
movapX reg2,reg
to
addsX/subsX/... reg3,reg
}
begin
TransferUsedRegs(TmpUsedRegs);
UpdateUsedRegs(TmpUsedRegs, tai(p.next));
UpdateUsedRegs(TmpUsedRegs, tai(hp1.next));
If not(RegUsedAfterInstruction(taicpu(p).oper[1]^.reg,hp2,TmpUsedRegs)) then
begin
DebugMsg(SPeepholeOptimization + 'MovapXOpMovapX2Op ('+
debug_op2str(taicpu(p).opcode)+' '+
debug_op2str(taicpu(hp1).opcode)+' '+
debug_op2str(taicpu(hp2).opcode)+') done',p);
{ we cannot eliminate the first move if
the operations uses the same register for source and dest }
if not(OpsEqual(taicpu(hp1).oper[1]^,taicpu(hp1).oper[0]^)) then
begin
asml.remove(p);
p.Free;
end;
taicpu(hp1).loadoper(1, taicpu(hp2).oper[1]^);
asml.remove(hp2);
hp2.Free;
p:=hp1;
result:=true;
end;
end;
end;
end;

11
compiler/x86_64/aoptcpu.pas
View File

@ -79,16 +79,15 @@ uses
A_MOVSX,
A_MOVZX:
Result:=OptPass1Movx(p);
A_MOVAPD,
A_MOVAPS,
A_MOVUPD,
A_MOVUPS,
A_VMOVAPS,
A_VMOVAPD,
A_VMOVUPS,
A_VMOVUPD:
result:=OptPass1VMOVAP(p);
A_MOVAPD,
A_MOVAPS,
A_MOVUPD,
A_MOVUPS:
result:=OptPass1MOVAP(p);
result:=OptPass1_V_MOVAP(p);
A_VDIVSD,
A_VDIVSS,
A_VSUBSD,