paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-09-02 18:30:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

Fix is_calljmp to include call instructions. This was breaking some optimizations previously.

Browse Source 
Add debug messages to peephole optimizations and add some extra optimizations.

git-svn-id: trunk@31244 -
This commit is contained in:
Jeppe Johansen 2015-07-30 06:29:21 +00:00
parent 6f3da41769
commit ffe67a93b4
2 changed files with 241 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

242
compiler/avr/aoptcpu.pas
View File

@ -26,12 +26,17 @@ Unit aoptcpu;
{$i fpcdefs.inc} {$i fpcdefs.inc}
{$define DEBUG_AOPTCPU}
Interface Interface
uses cpubase, cgbase, aasmtai, aopt, aoptcpub; uses cpubase, cgbase, aasmtai, aopt, aoptcpub;
Type Type
TCpuAsmOptimizer = class(TAsmOptimizer) TCpuAsmOptimizer = class(TAsmOptimizer)
{ outputs a debug message into the assembler file }
procedure DebugMsg(const s: string; p: tai);
Function GetNextInstructionUsingReg(Current: tai; Var Next: tai;reg : TRegister): Boolean; Function GetNextInstructionUsingReg(Current: tai; Var Next: tai;reg : TRegister): Boolean;
function RegInInstruction(Reg: TRegister; p1: tai): Boolean; override; function RegInInstruction(Reg: TRegister; p1: tai): Boolean; override;
@ -109,6 +114,18 @@ Implementation
end; end;
{$ifdef DEBUG_AOPTCPU}
procedure TCpuAsmOptimizer.DebugMsg(const s: string;p : tai);
begin
asml.insertbefore(tai_comment.Create(strpnew(s)), p);
end;
{$else DEBUG_AOPTCPU}
procedure TCpuAsmOptimizer.DebugMsg(const s: string;p : tai);inline;
begin
end;
{$endif DEBUG_AOPTCPU}
function TCpuAsmOptimizer.RegInInstruction(Reg: TRegister; p1: tai): Boolean; function TCpuAsmOptimizer.RegInInstruction(Reg: TRegister; p1: tai): Boolean;
begin begin
If (p1.typ = ait_instruction) and (taicpu(p1).opcode in [A_MUL,A_MULS,A_FMUL,A_FMULS,A_FMULSU]) and If (p1.typ = ait_instruction) and (taicpu(p1).opcode in [A_MUL,A_MULS,A_FMUL,A_FMULS,A_FMULSU]) and
@ -155,8 +172,13 @@ Implementation
A_OR,A_ORI,A_ROL,A_ROR,A_SBC,A_SBCI,A_SUB,A_SUBI]) and A_OR,A_ORI,A_ROL,A_ROR,A_SBC,A_SBCI,A_SUB,A_SUBI]) and
GetNextInstruction(p, hp1) and GetNextInstruction(p, hp1) and
MatchInstruction(hp1, A_CP) and MatchInstruction(hp1, A_CP) and
(taicpu(p).oper[0]^.reg = taicpu(hp1).oper[0]^.reg) and (((taicpu(p).oper[0]^.reg = taicpu(hp1).oper[0]^.reg) and
(taicpu(hp1).oper[1]^.reg = NR_R1) and (taicpu(hp1).oper[1]^.reg = NR_R1)) or
((taicpu(p).oper[0]^.reg = taicpu(hp1).oper[1]^.reg) and
(taicpu(hp1).oper[0]^.reg = NR_R1) and
(taicpu(p).opcode in [A_ADC,A_ADD,A_AND,A_ANDI,A_ASR,A_COM,A_EOR,
A_LSL,A_LSR,
A_OR,A_ORI,A_ROL,A_ROR]))) and
GetNextInstruction(hp1, hp2) and GetNextInstruction(hp1, hp2) and
{ be careful here, following instructions could use other flags { be careful here, following instructions could use other flags
however after a jump fpc never depends on the value of flags } however after a jump fpc never depends on the value of flags }
@ -181,6 +203,8 @@ Implementation
end; end;
} }
DebugMsg('Peephole OpCp2Op performed', p);
asml.remove(hp1); asml.remove(hp1);
hp1.free; hp1.free;
Result:=true; Result:=true;
@ -223,6 +247,8 @@ Implementation
dealloc.Free; dealloc.Free;
end; end;
DebugMsg('Peephole LdiCp2Cpi performed', p);
GetNextInstruction(p,hp1); GetNextInstruction(p,hp1);
asml.Remove(p); asml.Remove(p);
p.Free; p.Free;
@ -240,6 +266,8 @@ Implementation
(taicpu(p).oper[0]^.ref^.offset>=32) and (taicpu(p).oper[0]^.ref^.offset>=32) and
(taicpu(p).oper[0]^.ref^.offset<=95) then (taicpu(p).oper[0]^.ref^.offset<=95) then
begin begin
DebugMsg('Peephole Sts2Out performed', p);
taicpu(p).opcode:=A_OUT; taicpu(p).opcode:=A_OUT;
taicpu(p).loadconst(0,taicpu(p).oper[0]^.ref^.offset-32); taicpu(p).loadconst(0,taicpu(p).oper[0]^.ref^.offset-32);
end; end;
@ -252,6 +280,8 @@ Implementation
(taicpu(p).oper[1]^.ref^.offset>=32) and (taicpu(p).oper[1]^.ref^.offset>=32) and
(taicpu(p).oper[1]^.ref^.offset<=95) then (taicpu(p).oper[1]^.ref^.offset<=95) then
begin begin
DebugMsg('Peephole Lds2In performed', p);
taicpu(p).opcode:=A_IN; taicpu(p).opcode:=A_IN;
taicpu(p).loadconst(1,taicpu(p).oper[1]^.ref^.offset-32); taicpu(p).loadconst(1,taicpu(p).oper[1]^.ref^.offset-32);
end; end;
@ -275,6 +305,8 @@ Implementation
MatchOperand(taicpu(hp2).oper[1]^,taicpu(p).oper[0]^) and MatchOperand(taicpu(hp2).oper[1]^,taicpu(p).oper[0]^) and
MatchOperand(taicpu(hp2).oper[0]^,taicpu(p).oper[1]^) then MatchOperand(taicpu(hp2).oper[0]^,taicpu(p).oper[1]^) then
begin begin
DebugMsg('Peephole InOriOut2Sbi performed', p);
taicpu(p).opcode:=A_SBI; taicpu(p).opcode:=A_SBI;
taicpu(p).loadconst(0,taicpu(p).oper[1]^.val); taicpu(p).loadconst(0,taicpu(p).oper[1]^.val);
taicpu(p).loadconst(1,BsrByte(taicpu(hp1).oper[1]^.val)); taicpu(p).loadconst(1,BsrByte(taicpu(hp1).oper[1]^.val));
@ -301,6 +333,8 @@ Implementation
MatchOperand(taicpu(hp2).oper[1]^,taicpu(p).oper[0]^) and MatchOperand(taicpu(hp2).oper[1]^,taicpu(p).oper[0]^) and
MatchOperand(taicpu(hp2).oper[0]^,taicpu(p).oper[1]^) then MatchOperand(taicpu(hp2).oper[0]^,taicpu(p).oper[1]^) then
begin begin
DebugMsg('Peephole InAndiOut2Cbi performed', p);
taicpu(p).opcode:=A_CBI; taicpu(p).opcode:=A_CBI;
taicpu(p).loadconst(0,taicpu(p).oper[1]^.val); taicpu(p).loadconst(0,taicpu(p).oper[1]^.val);
taicpu(p).loadconst(1,BsrByte(not(taicpu(hp1).oper[1]^.val))); taicpu(p).loadconst(1,BsrByte(not(taicpu(hp1).oper[1]^.val)));
@ -333,6 +367,8 @@ Implementation
else else
taicpu(p).opcode:=A_SBIC; taicpu(p).opcode:=A_SBIC;
DebugMsg('Peephole InAndiBrx2SbixJmp performed', p);
taicpu(p).loadconst(0,taicpu(p).oper[1]^.val); taicpu(p).loadconst(0,taicpu(p).oper[1]^.val);
taicpu(p).loadconst(1,BsrByte(taicpu(hp1).oper[1]^.val)); taicpu(p).loadconst(1,BsrByte(taicpu(hp1).oper[1]^.val));
asml.Remove(hp1); asml.Remove(hp1);
@ -351,6 +387,73 @@ Implementation
result:=true; result:=true;
end; end;
end; end;
A_ANDI:
begin
{
Turn
andi rx, #pow2
brne l
<op>
l:
Into
sbrs rx, #(1 shl imm)
<op>
l:
}
if (taicpu(p).ops=2) and
(taicpu(p).oper[1]^.typ=top_const) and
ispowerof2(taicpu(p).oper[1]^.val,i) and
assigned(FindRegDeAlloc(taicpu(p).oper[0]^.reg,tai(p.next))) and
GetNextInstruction(p,hp1) and
(hp1.typ=ait_instruction) and
(taicpu(hp1).opcode=A_BRxx) and
(taicpu(hp1).condition in [C_EQ,C_NE]) and
(taicpu(hp1).ops>0) and
(taicpu(hp1).oper[0]^.typ = top_ref) and
(taicpu(hp1).oper[0]^.ref^.symbol is TAsmLabel) and
GetNextInstruction(hp1,hp2) and
(hp2.typ=ait_instruction) and
GetNextInstruction(hp2,hp3) and
(hp3.typ=ait_label) and
(taicpu(hp1).oper[0]^.ref^.symbol=tai_label(hp3).labsym) then
begin
DebugMsg('Peephole AndiBr2Sbr performed', p);
taicpu(p).oper[1]^.val:=i;
if taicpu(hp1).condition=C_NE then
taicpu(p).opcode:=A_SBRS
else
taicpu(p).opcode:=A_SBRC;
asml.Remove(hp1);
hp1.free;
result:=true;
end
{
Remove
andi rx, #y
dealloc rx
}
else if (taicpu(p).ops=2) and
(taicpu(p).oper[0]^.typ=top_reg) and
assigned(FindRegDeAlloc(taicpu(p).oper[0]^.reg,tai(p.next))) and
(assigned(FindRegDeAlloc(NR_DEFAULTFLAGS,tai(p.Next))) or
(not RegInUsedRegs(NR_DEFAULTFLAGS,UsedRegs))) then
begin
DebugMsg('Redundant Andi removed', p);
GetNextInstruction(p,hp1);
AsmL.Remove(p);
p.free;
p:=hp1;
result:=true;
end;
end;
A_CLR: A_CLR:
begin begin
{ turn the common { turn the common
@ -369,6 +472,8 @@ Implementation
(taicpu(hp1).oper[0]^.typ=top_reg) and (taicpu(hp1).oper[0]^.typ=top_reg) and
(taicpu(hp1).oper[0]^.reg=taicpu(p).oper[0]^.reg) then (taicpu(hp1).oper[0]^.reg=taicpu(p).oper[0]^.reg) then
begin begin
DebugMsg('Peephole ClrMov2Mov performed', p);
asml.Remove(p); asml.Remove(p);
p.Free; p.Free;
p:=hp1; p:=hp1;
@ -394,6 +499,8 @@ Implementation
(taicpu(hp1).oper[0]^.reg<>taicpu(p).oper[0]^.reg) and (taicpu(hp1).oper[0]^.reg<>taicpu(p).oper[0]^.reg) and
assigned(FindRegDeAlloc(taicpu(p).oper[0]^.reg,tai(hp1.Next))) then assigned(FindRegDeAlloc(taicpu(p).oper[0]^.reg,tai(hp1.Next))) then
begin begin
DebugMsg('Peephole ClrAdc2Adc performed', p);
taicpu(hp1).oper[1]^.reg:=NR_R1; taicpu(hp1).oper[1]^.reg:=NR_R1;
alloc:=FindRegAllocBackward(taicpu(p).oper[0]^.reg,tai(p.Previous)); alloc:=FindRegAllocBackward(taicpu(p).oper[0]^.reg,tai(p.Previous));
@ -443,6 +550,8 @@ Implementation
(getsupreg(taicpu(hp2).oper[0]^.reg)=getsupreg(taicpu(hp3).oper[0]^.reg)+1) and (getsupreg(taicpu(hp2).oper[0]^.reg)=getsupreg(taicpu(hp3).oper[0]^.reg)+1) and
((getsupreg(taicpu(hp3).oper[0]^.reg) mod 2)=0) then ((getsupreg(taicpu(hp3).oper[0]^.reg) mod 2)=0) then
begin begin
DebugMsg('Peephole PushPushPopPop2Movw performed', p);
taicpu(p).ops:=2; taicpu(p).ops:=2;
taicpu(p).opcode:=A_MOVW; taicpu(p).opcode:=A_MOVW;
@ -474,7 +583,8 @@ Implementation
GetNextInstructionUsingReg(p,hp1,taicpu(p).oper[0]^.reg) and GetNextInstructionUsingReg(p,hp1,taicpu(p).oper[0]^.reg) and
(not RegModifiedBetween(taicpu(p).oper[1]^.reg, p, hp1)) and (not RegModifiedBetween(taicpu(p).oper[1]^.reg, p, hp1)) and
(hp1.typ = ait_instruction) and (hp1.typ = ait_instruction) and
(taicpu(hp1).opcode in [A_PUSH,A_MOV,A_CP,A_CPC,A_ADD,A_SUB,A_EOR,A_AND,A_OR]) and (taicpu(hp1).opcode in [A_PUSH,A_MOV,A_CP,A_CPC,A_ADD,A_SUB,A_ADC,A_SBC,A_EOR,A_AND,A_OR,
A_OUT,A_IN]) and
RegInInstruction(taicpu(p).oper[0]^.reg, hp1) and RegInInstruction(taicpu(p).oper[0]^.reg, hp1) and
(not RegModifiedByInstruction(taicpu(p).oper[0]^.reg, hp1)) and (not RegModifiedByInstruction(taicpu(p).oper[0]^.reg, hp1)) and
{(taicpu(hp1).ops=1) and {(taicpu(hp1).ops=1) and
@ -482,6 +592,8 @@ Implementation
(taicpu(hp1).oper[0]^.reg=taicpu(p).oper[0]^.reg) and } (taicpu(hp1).oper[0]^.reg=taicpu(p).oper[0]^.reg) and }
assigned(FindRegDeAlloc(taicpu(p).oper[0]^.reg,tai(hp1.Next))) then assigned(FindRegDeAlloc(taicpu(p).oper[0]^.reg,tai(hp1.Next))) then
begin begin
DebugMsg('Peephole MovPush2Push performed', p);
for i := 0 to taicpu(hp1).ops-1 do for i := 0 to taicpu(hp1).ops-1 do
if taicpu(hp1).oper[i]^.typ=top_reg then if taicpu(hp1).oper[i]^.typ=top_reg then
if taicpu(hp1).oper[i]^.reg=taicpu(p).oper[0]^.reg then if taicpu(hp1).oper[i]^.reg=taicpu(p).oper[0]^.reg then
@ -512,10 +624,126 @@ Implementation
(taicpu(p).oper[1]^.typ = top_reg) and (taicpu(p).oper[1]^.typ = top_reg) and
(taicpu(p).oper[0]^.reg = taicpu(p).oper[1]^.reg) then (taicpu(p).oper[0]^.reg = taicpu(p).oper[1]^.reg) then
begin begin
DebugMsg('Peephole RedundantMov performed', p);
GetNextInstruction(p,hp1); GetNextInstruction(p,hp1);
asml.remove(p); asml.remove(p);
p.free; p.free;
p:=hp1; p:=hp1;
result:=true;
end
{
Turn
mov rx,ry
op rx,rz
mov ry, rx
Into
op ry,rz
}
else if (taicpu(p).ops=2) and
(taicpu(p).oper[0]^.typ = top_reg) and
(taicpu(p).oper[1]^.typ = top_reg) and
GetNextInstructionUsingReg(p,hp1,taicpu(p).oper[0]^.reg) and
(hp1.typ=ait_instruction) and
(taicpu(hp1).ops >= 1) and
(taicpu(hp1).oper[0]^.typ = top_reg) and
GetNextInstructionUsingReg(hp1,hp2,taicpu(hp1).oper[0]^.reg) and
(hp2.typ=ait_instruction) and
(taicpu(hp2).opcode=A_MOV) and
(taicpu(hp2).oper[0]^.typ = top_reg) and
(taicpu(hp2).oper[1]^.typ = top_reg) and
(taicpu(hp2).oper[0]^.reg = taicpu(p).oper[1]^.reg) and
(taicpu(hp2).oper[1]^.reg = taicpu(hp1).oper[0]^.reg) and
(taicpu(hp2).oper[1]^.reg = taicpu(p).oper[0]^.reg) and
(not RegModifiedBetween(taicpu(p).oper[1]^.reg,p,hp2)) and
(taicpu(hp1).opcode in [A_ADD,A_ADC,A_SUB,A_SBC,A_AND,A_OR,A_EOR,
A_LSL,A_LSR,A_ASR,A_ROR,A_ROL]) and
assigned(FindRegDeAlloc(taicpu(p).oper[0]^.reg, tai(hp2.Next))) then
begin
DebugMsg('Peephole MovOpMov2Op performed', p);
if (taicpu(hp1).ops=2) and
(taicpu(hp1).oper[1]^.typ=top_reg) and
(taicpu(hp1).oper[1]^.reg = taicpu(p).oper[1]^.reg) then
taicpu(hp1).oper[1]^.reg:=taicpu(p).oper[1]^.reg;
taicpu(hp1).oper[0]^.reg:=taicpu(p).oper[1]^.reg;
alloc:=FindRegAllocBackward(taicpu(p).oper[0]^.reg,tai(p.Previous));
dealloc:=FindRegDeAlloc(taicpu(p).oper[0]^.reg,tai(hp2.Next));
if assigned(alloc) and assigned(dealloc) then
begin
asml.Remove(alloc);
alloc.Free;
asml.Remove(dealloc);
dealloc.Free;
end;
GetNextInstruction(p,hp1);
asml.remove(p);
p.free;
asml.remove(hp2);
hp2.free;
p:=hp1;
result:=true;
end
{
Turn
mov rx,ry
op rx,rw
mov rw,rx
Into
op rw,ry
}
else if (taicpu(p).ops=2) and
(taicpu(p).oper[0]^.typ = top_reg) and
(taicpu(p).oper[1]^.typ = top_reg) and
GetNextInstructionUsingReg(p,hp1,taicpu(p).oper[0]^.reg) and
(hp1.typ=ait_instruction) and
(taicpu(hp1).ops = 2) and
(taicpu(hp1).oper[0]^.typ = top_reg) and
(taicpu(hp1).oper[1]^.typ = top_reg) and
GetNextInstructionUsingReg(hp1,hp2,taicpu(hp1).oper[0]^.reg) and
(hp2.typ=ait_instruction) and
(taicpu(hp2).opcode=A_MOV) and
(taicpu(hp2).oper[0]^.typ = top_reg) and
(taicpu(hp2).oper[1]^.typ = top_reg) and
(taicpu(hp2).oper[0]^.reg = taicpu(hp1).oper[1]^.reg) and
(taicpu(hp2).oper[1]^.reg = taicpu(hp1).oper[0]^.reg) and
(taicpu(hp2).oper[1]^.reg = taicpu(p).oper[0]^.reg) and
(not RegModifiedBetween(taicpu(p).oper[1]^.reg,p,hp1)) and
(taicpu(hp1).opcode in [A_ADD,A_ADC,A_AND,A_OR,A_EOR]) and
assigned(FindRegDeAlloc(taicpu(p).oper[0]^.reg, tai(hp2.Next))) then
begin
DebugMsg('Peephole MovOpMov2Op2 performed', p);
taicpu(hp1).oper[0]^.reg:=taicpu(hp2).oper[0]^.reg;
taicpu(hp1).oper[1]^.reg:=taicpu(p).oper[1]^.reg;
alloc:=FindRegAllocBackward(taicpu(p).oper[0]^.reg,tai(p.Previous));
dealloc:=FindRegDeAlloc(taicpu(p).oper[0]^.reg,tai(hp2.Next));
if assigned(alloc) and assigned(dealloc) then
begin
asml.Remove(alloc);
alloc.Free;
asml.Remove(dealloc);
dealloc.Free;
end;
GetNextInstruction(p,hp1);
asml.remove(p);
p.free;
asml.remove(hp2);
hp2.free;
p:=hp1;
result:=true; result:=true;
end end
{ fold { fold
@ -539,6 +767,8 @@ Implementation
((getsupreg(taicpu(p).oper[1]^.reg) mod 2)=0) and ((getsupreg(taicpu(p).oper[1]^.reg) mod 2)=0) and
(getsupreg(taicpu(hp1).oper[1]^.reg)=getsupreg(taicpu(p).oper[1]^.reg)+1) then (getsupreg(taicpu(hp1).oper[1]^.reg)=getsupreg(taicpu(p).oper[1]^.reg)+1) then
begin begin
DebugMsg('Peephole MovMov2Movw performed', p);
alloc:=FindRegAllocBackward(taicpu(hp1).oper[0]^.reg,tai(hp1.Previous)); alloc:=FindRegAllocBackward(taicpu(hp1).oper[0]^.reg,tai(hp1.Previous));
if assigned(alloc) then if assigned(alloc) then
begin begin
@ -562,6 +792,8 @@ Implementation
{ don't remove the first mov if the second is a mov rX,rX } { don't remove the first mov if the second is a mov rX,rX }
not(MatchOperand(taicpu(hp1).oper[0]^,taicpu(hp1).oper[1]^)) do not(MatchOperand(taicpu(hp1).oper[0]^,taicpu(hp1).oper[1]^)) do
begin begin
DebugMsg('Peephole MovMov2Mov performed', p);
asml.remove(p); asml.remove(p);
p.free; p.free;
p:=hp1; p:=hp1;
@ -599,6 +831,8 @@ Implementation
(hp3.typ=ait_label) and (hp3.typ=ait_label) and
(taicpu(hp1).oper[0]^.ref^.symbol=tai_label(hp3).labsym) then (taicpu(hp1).oper[0]^.ref^.symbol=tai_label(hp3).labsym) then
begin begin
DebugMsg('Peephole SbiJmp2Sbi performed',p);
if taicpu(p).opcode=A_SBIC then if taicpu(p).opcode=A_SBIC then
taicpu(p).opcode:=A_SBIS taicpu(p).opcode:=A_SBIS
else else
@ -651,6 +885,8 @@ Implementation
(hp3.typ=ait_label) and (hp3.typ=ait_label) and
(taicpu(hp2).oper[0]^.ref^.symbol=tai_label(hp5).labsym) then (taicpu(hp2).oper[0]^.ref^.symbol=tai_label(hp5).labsym) then
begin begin
DebugMsg('Peephole SbiJmpJmp2Sbi performed',p);
tai_label(hp3).labsym.decrefs; tai_label(hp3).labsym.decrefs;
tai_label(hp5).labsym.decrefs; tai_label(hp5).labsym.decrefs;

3
compiler/avr/cpubase.pas
View File

@ -68,6 +68,7 @@ unit cpubase;
{ call/reg instructions (A_RCALL,A_ICALL,A_CALL,A_RET,A_RETI) are not considered as jmp instructions for the usage cases of { call/reg instructions (A_RCALL,A_ICALL,A_CALL,A_RET,A_RETI) are not considered as jmp instructions for the usage cases of
this set } this set }
jmp_instructions = [A_BRxx,A_SBIC,A_SBIS,A_JMP,A_EIJMP,A_RJMP,A_CPSE,A_IJMP]; jmp_instructions = [A_BRxx,A_SBIC,A_SBIS,A_JMP,A_EIJMP,A_RJMP,A_CPSE,A_IJMP];
call_jmp_instructions = [A_ICALL,A_RCALL,A_CALL,A_RET,A_RETI]+jmp_instructions;
{***************************************************************************** {*****************************************************************************
Registers Registers
@ -487,7 +488,7 @@ unit cpubase;
function is_calljmp(o:tasmop):boolean;{$ifdef USEINLINE}inline;{$endif USEINLINE} function is_calljmp(o:tasmop):boolean;{$ifdef USEINLINE}inline;{$endif USEINLINE}
begin begin
is_calljmp:= o in jmp_instructions; is_calljmp:= o in call_jmp_instructions;
end; end;