paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-08-17 02:19:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

* x86: New XOR->BTC peephole optimisation

Browse Source
This commit is contained in:
J. Gareth "Curious Kit" Moreton 2022-11-12 04:00:50 +00:00 committed by FPK
parent f2bd099afb
commit 3cf8ae8433
2 changed files with 31 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
compiler/i386/aoptcpu.pas
View File

@ -386,6 +386,8 @@ unit aoptcpu;
A_ADD, A_ADD,
A_SUB: A_SUB:
Result:=PostPeepholeOptADDSUB(p); Result:=PostPeepholeOptADDSUB(p);
A_XOR:
Result:=PostPeepholeOptXor(p);
A_VPXOR: A_VPXOR:
Result:=PostPeepholeOptVPXOR(p); Result:=PostPeepholeOptVPXOR(p);
else else

34
compiler/x86/aoptx86.pas
View File

@ -196,9 +196,7 @@ unit aoptx86;
function PostPeepholeOptMov(var p : tai) : Boolean; function PostPeepholeOptMov(var p : tai) : Boolean;
function PostPeepholeOptMovzx(var p : tai) : Boolean; function PostPeepholeOptMovzx(var p : tai) : Boolean;
{$ifdef x86_64} { These post-peephole optimisations only affect 64-bit registers. [Kit] }
function PostPeepholeOptXor(var p : tai) : Boolean; function PostPeepholeOptXor(var p : tai) : Boolean;
{$endif x86_64}
function PostPeepholeOptAnd(var p : tai) : boolean; function PostPeepholeOptAnd(var p : tai) : boolean;
function PostPeepholeOptMOVSX(var p : tai) : boolean; function PostPeepholeOptMOVSX(var p : tai) : boolean;
function PostPeepholeOptCmp(var p : tai) : Boolean; function PostPeepholeOptCmp(var p : tai) : Boolean;
@ -14567,16 +14565,42 @@ unit aoptx86;
end; end;
{$ifdef x86_64}
function TX86AsmOptimizer.PostPeepholeOptXor(var p : tai) : Boolean; function TX86AsmOptimizer.PostPeepholeOptXor(var p : tai) : Boolean;
var var
hp1: tai;
{$ifdef x86_64}
PreMessage, RegName: string; PreMessage, RegName: string;
{$endif x86_64}
begin begin
Result := False;
{ If x is a power of 2 (popcnt = 1), change:
xor $x, %reg/ref
To:
btc lb(x), %reg/ref
}
if IsBTXAcceptable(p) and
{ IsBTXAcceptable checks to see if oper[0] is an immediate }
(PopCnt(QWord(taicpu(p).oper[0]^.val)) = 1) and
(
{ Don't optimise if a test instruction follows }
not GetNextInstruction(p, hp1) or
not MatchInstruction(hp1, A_TEST, [taicpu(p).opsize])
) then
begin
DebugMsg(SPeepholeOptimization + 'Changed XOR $0x' + hexstr(taicpu(p).oper[0]^.val, 2) + ' to BTC ' + debug_tostr(BsrQWord(taicpu(p).oper[0]^.val)) + ' to shrink instruction size (Xor2Btc)', p);
taicpu(p).opcode := A_BTC;
taicpu(p).oper[0]^.val := BsrQWord(taicpu(p).oper[0]^.val); { Essentially the base 2 logarithm }
Result := True;
Exit;
end;
{$ifdef x86_64}
{ Code size reduction by J. Gareth "Kit" Moreton } { Code size reduction by J. Gareth "Kit" Moreton }
{ change "xorq %reg,%reg" to "xorl %reg,%reg" for %rax, %rcx, %rdx, %rbx, %rsi, %rdi, %rbp and %rsp, { change "xorq %reg,%reg" to "xorl %reg,%reg" for %rax, %rcx, %rdx, %rbx, %rsi, %rdi, %rbp and %rsp,
as this removes the REX prefix } as this removes the REX prefix }
Result := False;
if not OpsEqual(taicpu(p).oper[0]^,taicpu(p).oper[1]^) then if not OpsEqual(taicpu(p).oper[0]^,taicpu(p).oper[1]^) then
Exit; Exit;
@ -14602,8 +14626,8 @@ unit aoptx86;
else else
; ;
end; end;
{$endif x86_64}
end; end;
{$endif}
function TX86AsmOptimizer.PostPeepholeOptVPXOR(var p : tai) : Boolean; function TX86AsmOptimizer.PostPeepholeOptVPXOR(var p : tai) : Boolean;
var var