https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100366
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue is that FRE sees
<bb 2> [local count: 1073741824]:
_6 = vec_2(D)->D.33436._M_impl.D.32749._M_start;
_7 = vec_2(D)->D.33436._M_impl.D.32749._M_finish;
if (_6 != _7)
goto <bb 3>; [70.00%]
else
goto <bb 4>; [30.00%]
<bb 3> [local count: 751619281]:
vec_2(D)->D.33436._M_impl.D.32749._M_finish = _6;
<bb 4> [local count: 1073741824]:
_5 = MEM[(char * const &)vec_2(D) + 8];
so clearly _5 is _not_ equal to _7 but on the path 2 -> 3 -> 4 it is equal to
_6.
Only PRE transforms the load in _5 to a select of both (a PHI node):
<bb 2> [local count: 1073741824]:
_6 = vec_2(D)->D.33436._M_impl.D.32749._M_start;
_7 = vec_2(D)->D.33436._M_impl.D.32749._M_finish;
if (_6 != _7)
goto <bb 4>; [70.00%]
else
goto <bb 3>; [30.00%]
<bb 3> [local count: 322122544]:
_86 = _7 - _6;
_125 = (sizetype) _86;
goto <bb 5>; [100.00%]
<bb 4> [local count: 751619281]:
vec_2(D)->D.33436._M_impl.D.32749._M_finish = _6;
<bb 5> [local count: 1073741824]:
# prephitmp_141 = PHI <_7(3), _6(4)>
# prephitmp_83 = PHI <_86(3), 0(4)>
# prephitmp_65 = PHI <_125(3), 0(4)>
_21 = vec_2(D)->D.33436._M_impl.D.32749._M_end_of_storage;
PRE also removes some partial redundancies here (the other two PHIs).