https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100263
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Though, looking at the r11-39 optimized dump, I actually don't see anything
wrong.
The path taken in the testcase should be:
<bb 2> [local count: 1073741824]:
b.0_1 = b;
if (b.0_1 != 2)
goto <bb 3>; [96.34%]
...
<bb 3> [local count: 1034442872]:
...
<bb 4> [local count: 10321168952]:
# b.12_44 = PHI <b.0_1(3), _32(42)>
# c__lsm.20_50 = PHI <c__lsm.20_75(D)(3), c__lsm.20_51(42)>
# c__lsm_flag.21_52 = PHI <0(3), c__lsm_flag.21_53(42)>
...
if (a.1_7 != 0)
goto <bb 5>; [50.00%]
else
goto <bb 19>; [50.00%]
...
<bb 5> [local count: 5160584476]:
if (h.4_5 != 0)
goto <bb 9>; [89.00%]
else
goto <bb 42>; [11.00%]
...
<bb 42> [local count: 9639533214]:
# c__lsm.20_51 = PHI <c__lsm.20_50(41), 1(5)>
# c__lsm_flag.21_53 = PHI <c__lsm_flag.21_52(41), 1(5)>
...
_32 = b.12_44 + 254;
if (_32 != 2)
goto <bb 4>; [96.34%]
else
goto <bb 43>; [3.66%]
<bb 43> [local count: 352806927]:
if (c__lsm_flag.21_53 != 0)
goto <bb 44>; [66.67%]
else
goto <bb 45>; [33.33%]
<bb 44> [local count: 235204617]:
MEM[(char *)&c + 107B] = c__lsm.20_51;
where bb will loop 126 times and then fall through to bb 43, but when bb 42 is
reached from bb 5 (should be always), then both SSA_NAMEs 51 and 53 are 1 and
so
the store of 1 should happen.
