https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108523
--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
So the issue is that we oscillate between recognizing the PHI as degenerate
because of an equivalence recorded on the non-backedge(!) to the backedge value
which in turn changes because of this change:
Value numbering stmt = g_6_lsm.126_456 = PHI <g_6_lsm.126_304(19),
g_6_lsm.126_338(108)>
Setting value number of g_6_lsm.126_456 to _347
...
Value numbering stmt = _350 = g_6_lsm.126_456;
Setting value number of _350 to _347 (changed)
_347 is available for _347
Value numbering stmt = _352 = _348 ^ _350;
_347 is available for _347
Match-and-simplified _348 ^ _350 to _346
RHS _348 ^ _350 simplified to _346
Setting value number of _352 to _346 (changed)
_346 is available for _346
Value numbering stmt = g_6_lsm.126_304 = _352;
Setting value number of g_6_lsm.126_304 to _346 (changed)
_346 is available for _346
Looking for changed values of backedge 19->20 destination PHIs
Setting value number of g_481.64_448 to g_481.64_448
Setting value number of .MEM_464 to .MEM_409
Predication says _347 and _346 are equal on edge 108 -> 20
Setting value number of g_6_lsm.126_456 to _346 (changed)
Iterating to 25 BB20
...
Value numbering stmt = g_6_lsm.126_456 = PHI <g_6_lsm.126_304(19),
g_6_lsm.126_338(108)>
Predication says _347 and _346 are equal on edge 108 -> 20
Setting value number of g_6_lsm.126_456 to _346
...
Value numbering stmt = _350 = g_6_lsm.126_456;
Setting value number of _350 to _346 (changed)
_346 is available for _346
Value numbering stmt = _352 = _348 ^ _350;
_346 is available for _346
Match-and-simplified _348 ^ _350 to _347
RHS _348 ^ _350 simplified to _347
Setting value number of _352 to _347 (changed)
_347 is available for _347
Value numbering stmt = g_6_lsm.126_304 = _352;
Setting value number of g_6_lsm.126_304 to _347 (changed)
_347 is available for _347
Looking for changed values of backedge 19->20 destination PHIs
Setting value number of g_481.64_448 to g_481.64_448
Setting value number of .MEM_464 to .MEM_409
Setting value number of g_6_lsm.126_456 to _347 (changed)
Iterating to 25 BB20
and repeat.
And we have:
_348 = _346 ^ _347;
and
<bb 18> [local count: 42914375]:
# g_1198.66_449 = PHI <_358(141), 0(17)>
# g_6_lsm.126_131 = PHI <g_6_lsm.126_349(141), g_6_lsm.126_438(17)>
g_481_lsm.127_447 = 1;
_57 = g_6_lsm.126_131;
_58 = _57 ^ _348;
g_6_lsm.126_338 = _58;
if (_346 != _347)
goto <bb 150>; [5.50%]
else
goto <bb 108>; [94.50%]
<bb 108> [local count: 40554084]:
goto <bb 20>; [100.00%]
So when the value we compare to oscillates. Interestingly when I try
to feed a similar case into FRE1 that "works":
int foo (int a, int b, int n)
{
int val = a;
if (a == b)
{
do
{
val = (a ^ b) ^ val;
}
while (--n > 0);
}
return val;
}
but maybe this is by luck. So what we need to avoid is for an equivalence
a == b, to oscillate between 'a' and 'b' as result.
