https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117935
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Summary|[[likely]] attribute not |[[likely]] attribute is
|propagated from function |lost by phiopt1 in some
|body to callers |cases
CC| |pinskia at gcc dot gnu.org
Component|middle-end |tree-optimization
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |pinskia at gcc dot
gnu.org
--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
clang/LLVM optimize all 3 cases the same.
For aarch64 even uses csel:
ldrb w8, [x0, #4]
ldr w9, [x0]
cmp w8, #0
csel w0, w9, wzr, ne
phiopt1 makes the likely predicate to be lost.
For this:
```
if (b_2(D) != 0)
goto <bb 3>; [INV]
else
goto <bb 4>; [INV]
<bb 3> :
// predicted likely by hot label predictor.
// predicted unlikely by early return (on trees) predictor.
goto <bb 5>; [INV]
<bb 4> :
// predicted unlikely by early return (on trees) predictor.
<bb 5> :
# _1 = PHI <1(3), 0(4)>
```
phiopt1 turns that into just:
return b_2(D);
I can think of a way to fix this is by rejecting it in (early) phiopt1 if the
predicates are not return types.
This will allow jump threading to handle it to get the correct predicate on the
branch.
And yes -fdisable-tree-phiopt1 fixes the issue.
