https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118954
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Ever confirmed|0 |1
Last reconfirmed| |2025-02-20
Status|UNCONFIRMED |NEW
--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The store to g is gone.
I doubt r15-1763-g56670281c6db19 caused it because the store is gone before
isel.
Before pcom we had:
```
_23 = a.0_1 != -1; // 1 (a is 0)
_24 = (int) _23;
_27 = 3 - _24; // 2
g[0] = 5;
g[1] = 8;
...
_20 = (unsigned int) _27; // 2
_19 = _20 + 4294967295; // _20 - 1 == 1
d.4_46 = (int) _19; // 1
_8 = (long unsigned int) d.4_46; // 1
_9 = _8 * 4; // 4
_10 = &g + _9; // &g[1]
_11 = *_10; // g[1] == 8
``
After pcom:
```
_23 = a.0_1 != -1; // 1 (a is 0)
_24 = (int) _23;
_27 = 3 - _24; // 2
g[0] = 5;
g[1] = 8;
...
_38 = (sizetype) _23; // 1
_45 = _38 * 18446744073709551612; // -4
_29 = &g + _45; // &g[-1]
D__lsm0.14_28 = MEM[(int *)_29 + 8B]; // &g[1]
```
But aliasing gets confused by this as &g[-1] is bogus and undefined.
