paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-08-14 09:29:26 +02:00
Code Issues Packages Projects Releases Wiki Activity

* x86: Factored out the MovMov2MovMov 2 optimisation to

Browse Source 
catch an inefficiency in the "Deep MOV" optimisations
This commit is contained in:
J. Gareth "Curious Kit" Moreton 2023-10-29 09:26:10 +00:00 committed by FPK
parent af1eb33a75
commit f4da88726c
1 changed files with 45 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

60
compiler/x86/aoptx86.pas
View File

@ -92,9 +92,12 @@ unit aoptx86;
function RegModifiedByInstruction(Reg: TRegister; p1: tai): boolean; override; function RegModifiedByInstruction(Reg: TRegister; p1: tai): boolean; override;
private private
function SkipSimpleInstructions(var hp1: tai): Boolean; function SkipSimpleInstructions(var hp1: tai): Boolean;
protected protected
class function IsMOVZXAcceptable: Boolean; static; inline; class function IsMOVZXAcceptable: Boolean; static; inline;
function CheckMovMov2MovMov2(const p, hp1: tai): Boolean;
{ Attempts to allocate a volatile integer register for use between p and hp, { Attempts to allocate a volatile integer register for use between p and hp,
using AUsedRegs for the current register usage information. Returns NR_NO using AUsedRegs for the current register usage information. Returns NR_NO
if no free register could be found } if no free register could be found }
@ -2882,9 +2885,28 @@ unit aoptx86;
end; end;
function TX86AsmOptimizer.CheckMovMov2MovMov2(const p, hp1: tai) : boolean;
begin
Result := False;
if MatchOpType(taicpu(p),top_ref,top_reg) and
MatchOpType(taicpu(hp1),top_ref,top_reg) and
(taicpu(p).opsize = taicpu(hp1).opsize) and
RefsEqual(taicpu(p).oper[0]^.ref^,taicpu(hp1).oper[0]^.ref^) and
(taicpu(p).oper[0]^.ref^.volatility=[]) and
(taicpu(hp1).oper[0]^.ref^.volatility=[]) and
not(SuperRegistersEqual(taicpu(p).oper[1]^.reg,taicpu(hp1).oper[0]^.ref^.base)) and
not(SuperRegistersEqual(taicpu(p).oper[1]^.reg,taicpu(hp1).oper[0]^.ref^.index)) then
begin
DebugMsg(SPeepholeOptimization + 'MovMov2MovMov 2',p);
taicpu(hp1).loadReg(0,taicpu(p).oper[1]^.reg);
Result := True;
end;
end;
function TX86AsmOptimizer.OptPass1MOV(var p : tai) : boolean; function TX86AsmOptimizer.OptPass1MOV(var p : tai) : boolean;
var var
hp1, hp2, hp3: tai; hp1, hp2, hp3, hp4: tai;
DoOptimisation, TempBool: Boolean; DoOptimisation, TempBool: Boolean;
{$ifdef x86_64} {$ifdef x86_64}
NewConst: TCGInt; NewConst: TCGInt;
@ -4078,18 +4100,7 @@ unit aoptx86;
movl [mem1],reg1 movl [mem1],reg1
movl reg1,reg2 movl reg1,reg2
} }
else if MatchOpType(taicpu(p),top_ref,top_reg) and else if not CheckMovMov2MovMov2(p, hp1) and
MatchOpType(taicpu(hp1),top_ref,top_reg) and
(taicpu(p).opsize = taicpu(hp1).opsize) and
RefsEqual(taicpu(p).oper[0]^.ref^,taicpu(hp1).oper[0]^.ref^) and
(taicpu(p).oper[0]^.ref^.volatility=[]) and
(taicpu(hp1).oper[0]^.ref^.volatility=[]) and
not(SuperRegistersEqual(taicpu(p).oper[1]^.reg,taicpu(hp1).oper[0]^.ref^.base)) and
not(SuperRegistersEqual(taicpu(p).oper[1]^.reg,taicpu(hp1).oper[0]^.ref^.index)) then
begin
DebugMsg(SPeepholeOptimization + 'MovMov2MovMov 2',p);
taicpu(hp1).loadReg(0,taicpu(p).oper[1]^.reg);
end;
{ movl const1,[mem1] { movl const1,[mem1]
movl [mem1],reg1 movl [mem1],reg1
@ -4099,7 +4110,7 @@ unit aoptx86;
movl const1,reg1 movl const1,reg1
movl reg1,[mem1] movl reg1,[mem1]
} }
if MatchOpType(Taicpu(p),top_const,top_ref) and MatchOpType(Taicpu(p),top_const,top_ref) and
MatchOpType(Taicpu(hp1),top_ref,top_reg) and MatchOpType(Taicpu(hp1),top_ref,top_reg) and
(taicpu(p).opsize = taicpu(hp1).opsize) and (taicpu(p).opsize = taicpu(hp1).opsize) and
RefsEqual(taicpu(hp1).oper[0]^.ref^,taicpu(p).oper[1]^.ref^) and RefsEqual(taicpu(hp1).oper[0]^.ref^,taicpu(p).oper[1]^.ref^) and
@ -4521,7 +4532,26 @@ unit aoptx86;
Break; Break;
end; end;
{$endif x86_64} {$endif x86_64}
end; end
else if (taicpu(hp2).oper[0]^.typ = top_ref) and
GetNextInstruction(hp2, hp4) and
(taicpu(hp4).opcode = A_MOV) then
{ Optimise the following first:
movl [mem1],reg1
movl [mem1],reg2
to
movl [mem1],reg1
movl reg1,reg2
If [mem1] contains the target register and reg1 is the
the source register, this optimisation will get missed
and produce less efficient code later on.
}
if CheckMovMov2MovMov2(hp2, hp4) then
{ Initial instruction wasn't actually changed }
Include(OptsToCheck, aoc_ForceNewIteration);
A_MOVZX, A_MOVSX{$ifdef x86_64}, A_MOVSXD{$endif x86_64}: A_MOVZX, A_MOVSX{$ifdef x86_64}, A_MOVSXD{$endif x86_64}:
if MatchOpType(taicpu(hp2), top_reg, top_reg) and if MatchOpType(taicpu(hp2), top_reg, top_reg) and