https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115259
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization
--- Comment #14 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
One thing I did note that cselim (which runs before sink1) on x86_64 (and
aarch64) was able to transform:
```
extent[n_11] = _9;
if (_9 < 0)
goto <bb 4>; [41.00%]
else
goto <bb 5>; [59.00%]
<bb 4> [local count: 391808389]:
extent[n_11] = 0;
<bb 5> [local count: 955630224]:
```
into:
```
extent[n_11] = _9;
if (_9 < 0)
goto <bb 5>; [41.00%]
else
goto <bb 4>; [59.00%]
<bb 4> [local count: 563821836]:
cstore_25 = MEM <index_type[15]> [(void *)&extent][n_11];
<bb 5> [local count: 955630224]:
# cstore_26 = PHI <cstore_25(4), 0(3)>
MEM <index_type[15]> [(void *)&extent][n_11] = cstore_26;
```
But I didn't look into why it didn't do that for gcn though.
Being able to do that, then sink would not do anything and you will not get the
MASKED_STORE mess that is currently being produced either and would be better
code generation in the first place.
