paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-08-31 16:30:35 +02:00
Code Issues Packages Projects Releases Wiki Activity

* x86: TEST/Jcc/TEST optimisations can now look beyond the next instruction under -O3

Browse Source
This commit is contained in:
J. Gareth "Curious Kit" Moreton 2023-02-14 15:07:48 +00:00 committed by FPK
parent d372286159
commit 3f25438d35
1 changed files with 151 additions and 95 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

94
compiler/x86/aoptx86.pas
View File

@ -5223,7 +5223,7 @@ unit aoptx86;
function TX86AsmOptimizer.OptPass1Test(var p: tai) : boolean;
var
hp1, p_label, p_dist, hp1_dist: tai;
hp1, p_label, p_dist, hp1_dist, hp1_last: tai;
JumpLabel, JumpLabel_dist: TAsmLabel;
FirstValue, SecondValue: TCGInt;
TempBool: Boolean;
@ -5313,6 +5313,7 @@ unit aoptx86;
{ Search for:
test $x,(reg/ref)
jne @lbl1
...
test $y,(reg/ref) (same register or reference)
jne @lbl1
@ -5327,6 +5328,7 @@ unit aoptx86;
Also search for:
test $x,(reg/ref)
je @lbl1
...
test $y,(reg/ref)
je/jne @lbl2
@ -5355,11 +5357,18 @@ unit aoptx86;
after it might (e.g. test; jne @lbl1; test; jne @lbl2; test @lbl1),
so accommodate for this with a while loop.
}
hp1_dist := hp1;
hp1_last := hp1;
if GetNextInstruction(hp1, p_dist) and
(p_dist.typ = ait_instruction) and
while (
(
(taicpu(p).oper[1]^.typ = top_reg) and
GetNextInstructionUsingReg(hp1_last, p_dist, taicpu(p).oper[1]^.reg)
) or GetNextInstruction(hp1_last, p_dist)
) and (p_dist.typ = ait_instruction) do
begin
if (
(
(taicpu(p_dist).opcode = A_TEST) and
(
@ -5389,15 +5398,14 @@ unit aoptx86;
else
SecondValue := taicpu(p_dist).oper[0]^.val;
{ If both of the TEST constants are identical, delete the second
TEST that is unnecessary. }
{ If both of the TEST constants are identical, delete the
second TEST that is unnecessary (be careful though, just
in case the flags are modified in between) }
if (FirstValue = SecondValue) then
begin
DebugMsg(SPeepholeOptimization + 'TEST/Jcc/TEST; removed superfluous TEST', p_dist);
RemoveInstruction(p_dist);
{ Don't let the flags register become deallocated and reallocated between the jumps }
AllocRegBetween(NR_DEFAULTFLAGS, hp1, hp1_dist, UsedRegs);
Result := True;
{ We have to check the entire range }
TempBool := not RegModifiedBetween(NR_DEFAULTFLAGS, hp1, p_dist);
if condition_in(taicpu(hp1_dist).condition, taicpu(hp1).condition) then
begin
{ Since the second jump's condition is a subset of the first, we
@ -5409,18 +5417,60 @@ unit aoptx86;
DebugMsg(SPeepholeOptimization + 'Removed dominated jump (via TEST/Jcc/TEST)', hp1_dist);
RemoveInstruction(hp1_dist);
Result := True;
end
else if condition_in(inverse_cond(taicpu(hp1).condition), taicpu(hp1_dist).condition) then
begin
{ If the inverse of the first condition is a subset of the second,
the second one will definitely branch if the first one doesn't }
DebugMsg(SPeepholeOptimization + 'Conditional jump will always branch (via TEST/Jcc/TEST)', hp1_dist);
{ We can remove the TEST instruction too }
DebugMsg(SPeepholeOptimization + 'TEST/Jcc/TEST; removed superfluous TEST', p_dist);
RemoveInstruction(p_dist);
MakeUnconditional(taicpu(hp1_dist));
RemoveDeadCodeAfterJump(hp1_dist);
{ Since the jump is now unconditional, we can't
continue any further with this particular
optimisation. The original TEST is still intact
though, so there might be something else we can
do }
Include(OptsToCheck, aoc_ForceNewIteration);
Break;
end;
if Result or
{ If a jump wasn't removed or made unconditional, only
remove the identical TEST instruction if the flags
weren't modified }
TempBool then
begin
DebugMsg(SPeepholeOptimization + 'TEST/Jcc/TEST; removed superfluous TEST', p_dist);
RemoveInstruction(p_dist);
{ If the jump was removed or made unconditional, we
don't need to allocate NR_DEFAULTFLAGS over the
entire range }
if not Result then
begin
{ Mark the flags as 'in use' over the entire range }
AllocRegBetween(NR_DEFAULTFLAGS, hp1, hp1_dist, UsedRegs);
{ Speed gain - continue search from the Jcc instruction }
hp1_last := hp1_dist;
{ Only the TEST instruction was removed, and the
original was unchanged, so we can safely do
another iteration of the while loop }
Include(OptsToCheck, aoc_ForceNewIteration);
Continue;
end;
Exit;
end;
end;
if (taicpu(hp1).condition in [C_NE, C_NZ]) and
(taicpu(hp1_dist).condition in [C_NE, C_NZ]) and
@ -5456,6 +5506,19 @@ unit aoptx86;
Exit;
end;
end;
if { If -O2 and under, it may stop on any old instruction }
(cs_opt_level3 in current_settings.optimizerswitches) and
(taicpu(p).oper[1]^.typ = top_reg) and
not RegModifiedByInstruction(taicpu(p).oper[1]^.reg, p_dist) then
begin
hp1_last := p_dist;
Continue;
end;
Break;
end;
end;
{ Search for:
@ -6285,8 +6348,6 @@ unit aoptx86;
function TX86AsmOptimizer.DoMovCmpMemOpt(var p : tai; const hp1: tai) : Boolean;
var
hp2: tai;
begin
Result := False;
if MatchOpType(taicpu(p),top_ref,top_reg) and
@ -6334,13 +6395,8 @@ unit aoptx86;
RemoveCurrentP(p);
if (p <> hp1) then
begin
{ Correctly update TmpUsedRegs if p and hp1 aren't adjacent }
hp2 := p;
repeat
UpdateUsedRegs(TmpUsedRegs, tai(hp2.Next));
until not GetNextInstruction(hp2, hp2) or (hp2 = hp1);
end;
UpdateUsedRegsBetween(TmpUsedRegs, p, hp1);
{ Make sure the flags are allocated across the CMP instruction }
if not RegInUsedRegs(NR_DEFAULTFLAGS, TmpUsedRegs) then