https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105030
Drea Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Known to fail| |12.1.0
Summary|store motion if-change flag |[13/14/15/16/17 Regression]
|causes if-conversion |store motion if-change flag
|optimization can't be |causes if-conversion
|taken. |optimization can't be
| |taken.
Status|UNCONFIRMED |RESOLVED
Target Milestone|--- |17.0
Known to work| |11.4.0
Resolution|--- |FIXED
--- Comment #15 from Drea Pinski <pinskia at gcc dot gnu.org> ---
Phiopt1 now creates (at -Ofast):
atemp.0_5 = atemp;
_8 = MIN_EXPR <_4, atemp.0_5>;
MEM <double> [(void *)&atemp] = _8;
After
https://gcc.gnu.org/cgit/gcc/commit/?id=eb8011467d0ce70cae098a88569a0c8a5b7f5683
https://gcc.gnu.org/cgit/gcc/commit/?id=fd3e590360dc4f479e69142ed1bd6775c07213b6
lim2 gives:
```
if (n_12(D) > 0)
goto <bb 5>; [89.00%]
else
goto <bb 4>; [11.00%]
<bb 5> [local count: 105119324]:
atemp__lsm.4_7 = MEM <double> [(void *)&atemp];
<bb 3> [local count: 955630224]:
# i_19 = PHI <i_16(6), 0(5)>
# atemp__lsm.4_9 = PHI <atemp__lsm.4_18(6), atemp__lsm.4_7(5)>
_1 = (long unsigned int) i_19;
_2 = _1 * 8;
_3 = a_15(D) + _2;
_4 = *_3;
atemp.0_5 = atemp__lsm.4_9;
_6 = MIN_EXPR <_4, atemp.0_5>;
atemp__lsm.4_18 = _6;
i_16 = i_19 + 1;
if (n_12(D) > i_16)
goto <bb 6>; [89.00%]
else
goto <bb 7>; [11.00%]
<bb 6> [local count: 850510900]:
goto <bb 3>; [100.00%]
<bb 7> [local count: 105119324]:
# atemp__lsm.4_22 = PHI <atemp__lsm.4_18(3)>
MEM <double> [(void *)&atemp] = atemp__lsm.4_22;
<bb 4> [local count: 118111600]:
bar (&atemp, n_12(D));
atemp ={v} {CLOBBER(eos)};
return;
```
This is now vectorized too. Which it was not before except pre GCC 12 when
CSELIM didn't handle the decl.
So this was fixed by those 2 patches. Basically the load before the condition
allows to reuse the value and lim2 can do store motion (for -Ofast).
