https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123070
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|rtl-optimization |tree-optimization
--- Comment #20 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So I am not sure if the best place to resolve this is cse. The problem is
really ivopts which causes the overlapping live ranges to show up.
````
<bb 3> [local count: 1014686025]:
# ivtmp.7_17 = PHI <ivtmp.7_6(7), ivtmp.7_20(6)>
_3 = (int *) ivtmp.7_17;
_21 = (void *) ivtmp.7_17;
_4 = MEM[(int *)_21];
if (_4 == val_13)
goto <bb 9>; [5.50%] // exit
else
goto <bb 4>; [94.50%]
<bb 9> [local count: 55807731]:
# _19 = PHI <_3(3)>
goto <bb 5>; [100.00%]
<bb 4> [local count: 958878295]:
ivtmp.7_6 = ivtmp.7_17 + 4;
if (ivtmp.7_6 != _25)
goto <bb 7>; [94.50%]
else
goto <bb 12>; [5.50%] // exit
<bb 12> [local count: 52738306]:
goto <bb 8>; [100.00%]
<bb 7> [local count: 906139989]:
goto <bb 3>; [100.00%]
```
Notice how _3 (which is also _21 later on) and ivtmp.7_17 are alive at the same
time [though they are the same value].
(this is what got "fixed" up by cse2 before and is not currently).
The big question is how to fix this in IV-OPTs. Using the pointer type for the
IV might cause an overflow but maybe that would be undefined in the beginning.
