* i386: more clean up of TCPUAsmOPtimizer.PeepHoleOptPass1Cpu

git-svn-id: trunk@43465 -
This commit is contained in:
florian 2019-11-14 21:26:48 +00:00
parent 632f13c47a
commit 87b3b089d6

View File

@ -245,324 +245,244 @@ unit aoptcpu;
Result:=true;
exit;
end;
{ Handle Jmp Optimizations }
if taicpu(p).is_jmp then
begin
{ the following if-block removes all code between a jmp and the next label,
because it can never be executed }
if (taicpu(p).opcode = A_JMP) then
begin
hp2:=p;
while GetNextInstruction(hp2, hp1) and
(hp1.typ <> ait_label) do
if not(hp1.typ in ([ait_label]+skipinstr)) then
begin
{ don't kill start/end of assembler block,
no-line-info-start/end, cfi end, etc }
if not(hp1.typ in [ait_align,ait_marker]) and
((hp1.typ<>ait_cfi) or
(tai_cfi_base(hp1).cfityp<>cfi_endproc)) then
begin
asml.remove(hp1);
hp1.free;
end
else
hp2:=hp1;
end
else break;
end;
{ remove jumps to a label coming right after them }
if GetNextInstruction(p, hp1) then
begin
if FindLabel(tasmlabel(taicpu(p).oper[0]^.ref^.symbol), hp1) and
{ TODO: FIXME removing the first instruction fails}
(p<>blockstart) then
begin
hp2:=tai(hp1.next);
asml.remove(p);
p.free;
p:=hp2;
Result:=true;
exit;
end
case taicpu(p).opcode Of
A_AND:
Result:=OptPass1And(p);
A_CMP:
begin
{ cmp register,$8000 neg register
je target --> jo target
.... only if register is deallocated before jump.}
case Taicpu(p).opsize of
S_B: v:=$80;
S_W: v:=$8000;
S_L: v:=aint($80000000);
else
begin
if hp1.typ = ait_label then
SkipLabels(hp1,hp1);
if (tai(hp1).typ=ait_instruction) and
(taicpu(hp1).opcode=A_JMP) and
GetNextInstruction(hp1, hp2) and
FindLabel(tasmlabel(taicpu(p).oper[0]^.ref^.symbol), hp2) then
begin
if taicpu(p).opcode=A_Jcc then
begin
taicpu(p).condition:=inverse_cond(taicpu(p).condition);
tai_label(hp2).labsym.decrefs;
taicpu(p).oper[0]^.ref^.symbol:=taicpu(hp1).oper[0]^.ref^.symbol;
{ when free'ing hp1, the ref. isn't decresed, so we don't
increase it (FK)
taicpu(p).oper[0]^.ref^.symbol.increfs;
}
asml.remove(hp1);
hp1.free;
GetFinalDestination(asml, taicpu(p),0);
end
else
begin
GetFinalDestination(asml, taicpu(p),0);
p:=tai(p.next);
Result:=true;
exit;
end;
end
else
GetFinalDestination(asml, taicpu(p),0);
end;
internalerror(2013112905);
end;
end
else
{ All other optimizes }
begin
case taicpu(p).opcode Of
A_AND:
Result:=OptPass1And(p);
A_CMP:
if (taicpu(p).oper[0]^.typ=Top_const) and
(taicpu(p).oper[0]^.val=v) and
(Taicpu(p).oper[1]^.typ=top_reg) and
GetNextInstruction(p, hp1) and
(hp1.typ=ait_instruction) and
(taicpu(hp1).opcode=A_Jcc) and
(Taicpu(hp1).condition in [C_E,C_NE]) and
not(RegInUsedRegs(Taicpu(p).oper[1]^.reg, UsedRegs)) then
begin
{ cmp register,$8000 neg register
je target --> jo target
.... only if register is deallocated before jump.}
case Taicpu(p).opsize of
S_B: v:=$80;
S_W: v:=$8000;
S_L: v:=aint($80000000);
else
internalerror(2013112905);
end;
if (taicpu(p).oper[0]^.typ=Top_const) and
(taicpu(p).oper[0]^.val=v) and
(Taicpu(p).oper[1]^.typ=top_reg) and
GetNextInstruction(p, hp1) and
(hp1.typ=ait_instruction) and
(taicpu(hp1).opcode=A_Jcc) and
(Taicpu(hp1).condition in [C_E,C_NE]) and
not(RegInUsedRegs(Taicpu(p).oper[1]^.reg, UsedRegs)) then
begin
Taicpu(p).opcode:=A_NEG;
Taicpu(p).loadoper(0,Taicpu(p).oper[1]^);
Taicpu(p).clearop(1);
Taicpu(p).ops:=1;
if Taicpu(hp1).condition=C_E then
Taicpu(hp1).condition:=C_O
else
Taicpu(hp1).condition:=C_NO;
Result:=true;
end;
{
@@2: @@2:
.... ....
cmp operand1,0
jle/jbe @@1
dec operand1 --> sub operand1,1
jmp @@2 jge/jae @@2
@@1: @@1:
... ....}
if (taicpu(p).oper[0]^.typ = top_const) and
(taicpu(p).oper[1]^.typ in [top_reg,top_ref]) and
(taicpu(p).oper[0]^.val = 0) and
GetNextInstruction(p, hp1) and
(hp1.typ = ait_instruction) and
(taicpu(hp1).is_jmp) and
(taicpu(hp1).opcode=A_Jcc) and
(taicpu(hp1).condition in [C_LE,C_BE]) and
GetNextInstruction(hp1,hp2) and
(hp2.typ = ait_instruction) and
(taicpu(hp2).opcode = A_DEC) and
OpsEqual(taicpu(hp2).oper[0]^,taicpu(p).oper[1]^) and
GetNextInstruction(hp2, hp3) and
(hp3.typ = ait_instruction) and
(taicpu(hp3).is_jmp) and
(taicpu(hp3).opcode = A_JMP) and
GetNextInstruction(hp3, hp4) and
FindLabel(tasmlabel(taicpu(hp1).oper[0]^.ref^.symbol),hp4) then
begin
taicpu(hp2).Opcode := A_SUB;
taicpu(hp2).loadoper(1,taicpu(hp2).oper[0]^);
taicpu(hp2).loadConst(0,1);
taicpu(hp2).ops:=2;
taicpu(hp3).Opcode := A_Jcc;
case taicpu(hp1).condition of
C_LE: taicpu(hp3).condition := C_GE;
C_BE: taicpu(hp3).condition := C_AE;
else
internalerror(2019050903);
end;
asml.remove(p);
asml.remove(hp1);
p.free;
hp1.free;
p := hp2;
Result:=true;
end
Taicpu(p).opcode:=A_NEG;
Taicpu(p).loadoper(0,Taicpu(p).oper[1]^);
Taicpu(p).clearop(1);
Taicpu(p).ops:=1;
if Taicpu(hp1).condition=C_E then
Taicpu(hp1).condition:=C_O
else
Taicpu(hp1).condition:=C_NO;
Result:=true;
end;
A_FLD:
Result:=OptPass1FLD(p);
A_FSTP,A_FISTP:
Result:=OptPass1FSTP(p);
A_LEA:
Result:=OptPass1LEA(p);
A_MOV:
Result:=OptPass1MOV(p);
A_MOVSX,
A_MOVZX :
Result:=OptPass1Movx(p);
(* should not be generated anymore by the current code generator
A_POP:
{
@@2: @@2:
.... ....
cmp operand1,0
jle/jbe @@1
dec operand1 --> sub operand1,1
jmp @@2 jge/jae @@2
@@1: @@1:
... ....}
if (taicpu(p).oper[0]^.typ = top_const) and
(taicpu(p).oper[1]^.typ in [top_reg,top_ref]) and
(taicpu(p).oper[0]^.val = 0) and
GetNextInstruction(p, hp1) and
(hp1.typ = ait_instruction) and
(taicpu(hp1).is_jmp) and
(taicpu(hp1).opcode=A_Jcc) and
(taicpu(hp1).condition in [C_LE,C_BE]) and
GetNextInstruction(hp1,hp2) and
(hp2.typ = ait_instruction) and
(taicpu(hp2).opcode = A_DEC) and
OpsEqual(taicpu(hp2).oper[0]^,taicpu(p).oper[1]^) and
GetNextInstruction(hp2, hp3) and
(hp3.typ = ait_instruction) and
(taicpu(hp3).is_jmp) and
(taicpu(hp3).opcode = A_JMP) and
GetNextInstruction(hp3, hp4) and
FindLabel(tasmlabel(taicpu(hp1).oper[0]^.ref^.symbol),hp4) then
begin
if target_info.system=system_i386_go32v2 then
taicpu(hp2).Opcode := A_SUB;
taicpu(hp2).loadoper(1,taicpu(hp2).oper[0]^);
taicpu(hp2).loadConst(0,1);
taicpu(hp2).ops:=2;
taicpu(hp3).Opcode := A_Jcc;
case taicpu(hp1).condition of
C_LE: taicpu(hp3).condition := C_GE;
C_BE: taicpu(hp3).condition := C_AE;
else
internalerror(2019050903);
end;
asml.remove(p);
asml.remove(hp1);
p.free;
hp1.free;
p := hp2;
Result:=true;
end
end;
A_FLD:
Result:=OptPass1FLD(p);
A_FSTP,A_FISTP:
Result:=OptPass1FSTP(p);
A_LEA:
Result:=OptPass1LEA(p);
A_MOV:
Result:=OptPass1MOV(p);
A_MOVSX,
A_MOVZX :
Result:=OptPass1Movx(p);
(* should not be generated anymore by the current code generator
A_POP:
begin
if target_info.system=system_i386_go32v2 then
begin
{ Transform a series of pop/pop/pop/push/push/push to }
{ 'movl x(%esp),%reg' for go32v2 (not for the rest, }
{ because I'm not sure whether they can cope with }
{ 'movl x(%esp),%reg' with x > 0, I believe we had }
{ such a problem when using esp as frame pointer (JM) }
if (taicpu(p).oper[0]^.typ = top_reg) then
begin
{ Transform a series of pop/pop/pop/push/push/push to }
{ 'movl x(%esp),%reg' for go32v2 (not for the rest, }
{ because I'm not sure whether they can cope with }
{ 'movl x(%esp),%reg' with x > 0, I believe we had }
{ such a problem when using esp as frame pointer (JM) }
if (taicpu(p).oper[0]^.typ = top_reg) then
hp1 := p;
hp2 := p;
l := 0;
while getNextInstruction(hp1,hp1) and
(hp1.typ = ait_instruction) and
(taicpu(hp1).opcode = A_POP) and
(taicpu(hp1).oper[0]^.typ = top_reg) do
begin
hp1 := p;
hp2 := p;
l := 0;
while getNextInstruction(hp1,hp1) and
(hp1.typ = ait_instruction) and
(taicpu(hp1).opcode = A_POP) and
(taicpu(hp1).oper[0]^.typ = top_reg) do
hp2 := hp1;
inc(l,4);
end;
getLastInstruction(p,hp3);
l1 := 0;
while (hp2 <> hp3) and
assigned(hp1) and
(hp1.typ = ait_instruction) and
(taicpu(hp1).opcode = A_PUSH) and
(taicpu(hp1).oper[0]^.typ = top_reg) and
(taicpu(hp1).oper[0]^.reg.enum = taicpu(hp2).oper[0]^.reg.enum) do
begin
{ change it to a two op operation }
taicpu(hp2).oper[1]^.typ:=top_none;
taicpu(hp2).ops:=2;
taicpu(hp2).opcode := A_MOV;
taicpu(hp2).loadoper(1,taicpu(hp1).oper[0]^);
reference_reset(tmpref);
tmpRef.base.enum:=R_INTREGISTER;
tmpRef.base.number:=NR_STACK_POINTER_REG;
convert_register_to_enum(tmpref.base);
tmpRef.offset := l;
taicpu(hp2).loadRef(0,tmpRef);
hp4 := hp1;
getNextInstruction(hp1,hp1);
asml.remove(hp4);
hp4.free;
getLastInstruction(hp2,hp2);
dec(l,4);
inc(l1);
end;
if l <> -4 then
begin
inc(l,4);
for l1 := l1 downto 1 do
begin
hp2 := hp1;
inc(l,4);
end;
getLastInstruction(p,hp3);
l1 := 0;
while (hp2 <> hp3) and
assigned(hp1) and
(hp1.typ = ait_instruction) and
(taicpu(hp1).opcode = A_PUSH) and
(taicpu(hp1).oper[0]^.typ = top_reg) and
(taicpu(hp1).oper[0]^.reg.enum = taicpu(hp2).oper[0]^.reg.enum) do
begin
{ change it to a two op operation }
taicpu(hp2).oper[1]^.typ:=top_none;
taicpu(hp2).ops:=2;
taicpu(hp2).opcode := A_MOV;
taicpu(hp2).loadoper(1,taicpu(hp1).oper[0]^);
reference_reset(tmpref);
tmpRef.base.enum:=R_INTREGISTER;
tmpRef.base.number:=NR_STACK_POINTER_REG;
convert_register_to_enum(tmpref.base);
tmpRef.offset := l;
taicpu(hp2).loadRef(0,tmpRef);
hp4 := hp1;
getNextInstruction(hp1,hp1);
asml.remove(hp4);
hp4.free;
getLastInstruction(hp2,hp2);
dec(l,4);
inc(l1);
end;
if l <> -4 then
begin
inc(l,4);
for l1 := l1 downto 1 do
begin
getNextInstruction(hp2,hp2);
dec(taicpu(hp2).oper[0]^.ref^.offset,l);
end
getNextInstruction(hp2,hp2);
dec(taicpu(hp2).oper[0]^.ref^.offset,l);
end
end
end
else
begin
if (taicpu(p).oper[0]^.typ = top_reg) and
GetNextInstruction(p, hp1) and
(tai(hp1).typ=ait_instruction) and
(taicpu(hp1).opcode=A_PUSH) and
(taicpu(hp1).oper[0]^.typ = top_reg) and
(taicpu(hp1).oper[0]^.reg.enum=taicpu(p).oper[0]^.reg.enum) then
begin
{ change it to a two op operation }
taicpu(p).oper[1]^.typ:=top_none;
taicpu(p).ops:=2;
taicpu(p).opcode := A_MOV;
taicpu(p).loadoper(1,taicpu(p).oper[0]^);
reference_reset(tmpref);
TmpRef.base.enum := R_ESP;
taicpu(p).loadRef(0,TmpRef);
asml.remove(hp1);
hp1.free;
end;
end;
end;
*)
A_PUSH:
end
end
else
begin
if (taicpu(p).opsize = S_W) and
(taicpu(p).oper[0]^.typ = Top_Const) and
GetNextInstruction(p, hp1) and
(tai(hp1).typ = ait_instruction) and
(taicpu(hp1).opcode = A_PUSH) and
(taicpu(hp1).oper[0]^.typ = Top_Const) and
(taicpu(hp1).opsize = S_W) then
if (taicpu(p).oper[0]^.typ = top_reg) and
GetNextInstruction(p, hp1) and
(tai(hp1).typ=ait_instruction) and
(taicpu(hp1).opcode=A_PUSH) and
(taicpu(hp1).oper[0]^.typ = top_reg) and
(taicpu(hp1).oper[0]^.reg.enum=taicpu(p).oper[0]^.reg.enum) then
begin
taicpu(p).changeopsize(S_L);
taicpu(p).loadConst(0,taicpu(p).oper[0]^.val shl 16 + word(taicpu(hp1).oper[0]^.val));
{ change it to a two op operation }
taicpu(p).oper[1]^.typ:=top_none;
taicpu(p).ops:=2;
taicpu(p).opcode := A_MOV;
taicpu(p).loadoper(1,taicpu(p).oper[0]^);
reference_reset(tmpref);
TmpRef.base.enum := R_ESP;
taicpu(p).loadRef(0,TmpRef);
asml.remove(hp1);
hp1.free;
Result:=true;
end;
end;
A_SHL, A_SAL:
Result:=OptPass1SHLSAL(p);
A_SUB:
Result:=OptPass1Sub(p);
A_MOVAPD,
A_MOVAPS,
A_MOVUPD,
A_MOVUPS,
A_VMOVAPS,
A_VMOVAPD,
A_VMOVUPS,
A_VMOVUPD:
Result:=OptPass1_V_MOVAP(p);
A_VDIVSD,
A_VDIVSS,
A_VSUBSD,
A_VSUBSS,
A_VMULSD,
A_VMULSS,
A_VADDSD,
A_VADDSS,
A_VANDPD,
A_VANDPS,
A_VORPD,
A_VORPS,
A_VXORPD,
A_VXORPS:
Result:=OptPass1VOP(p);
A_MULSD,
A_MULSS,
A_ADDSD,
A_ADDSS:
Result:=OptPass1OP(p);
A_VMOVSD,
A_VMOVSS,
A_MOVSD,
A_MOVSS:
Result:=OptPass1MOVXX(p);
A_SETcc:
Result:=OptPass1SETcc(p);
else
;
end;
end; { if is_jmp }
*)
A_PUSH:
begin
if (taicpu(p).opsize = S_W) and
(taicpu(p).oper[0]^.typ = Top_Const) and
GetNextInstruction(p, hp1) and
(tai(hp1).typ = ait_instruction) and
(taicpu(hp1).opcode = A_PUSH) and
(taicpu(hp1).oper[0]^.typ = Top_Const) and
(taicpu(hp1).opsize = S_W) then
begin
taicpu(p).changeopsize(S_L);
taicpu(p).loadConst(0,taicpu(p).oper[0]^.val shl 16 + word(taicpu(hp1).oper[0]^.val));
asml.remove(hp1);
hp1.free;
Result:=true;
end;
end;
A_SHL, A_SAL:
Result:=OptPass1SHLSAL(p);
A_SUB:
Result:=OptPass1Sub(p);
A_MOVAPD,
A_MOVAPS,
A_MOVUPD,
A_MOVUPS,
A_VMOVAPS,
A_VMOVAPD,
A_VMOVUPS,
A_VMOVUPD:
Result:=OptPass1_V_MOVAP(p);
A_VDIVSD,
A_VDIVSS,
A_VSUBSD,
A_VSUBSS,
A_VMULSD,
A_VMULSS,
A_VADDSD,
A_VADDSS,
A_VANDPD,
A_VANDPS,
A_VORPD,
A_VORPS,
A_VXORPD,
A_VXORPS:
Result:=OptPass1VOP(p);
A_MULSD,
A_MULSS,
A_ADDSD,
A_ADDSS:
Result:=OptPass1OP(p);
A_VMOVSD,
A_VMOVSS,
A_MOVSD,
A_MOVSS:
Result:=OptPass1MOVXX(p);
A_SETcc:
Result:=OptPass1SETcc(p);
else
;
end;
end;
else
;