[Bug tree-optimization/118852] [15 regression] Train run of 502.gcc_r compiled with -Ofast -fprofile-generate -march=x86_64-v3 fails

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118852

--- Comment #15 from Richard Biener <rguenth at gcc dot gnu.org> ---
So we're failing to vectorize

hi$slot_78 = PHI <_75(24), hi$slot_71(15)>

which we lack SLP discovery for.  That's because the stmt isn't live but
only forced-live by early break and that's because we start SLP discovery
from the latch def, but here we have

<bb 16> [local count: 1014686024]:
# _75 = PHI <_18(24), _62(15)>
# hi$slot_78 = PHI <_75(24), hi$slot_71(15)>
...
<bb 18> [local count: 958878294]:
PROF_edge_counter_35 = __gcov0.set_hashtable_value_ids_1_I_lsm.24_5 + 1;
_18 = _75 + 8;
...
<bb 24> [local count: 906139988]:
goto <bb 16>; [100.00%]

so discovery for _18 will get us the _75 SSA cycle, not including the
hi$slot_78 PHI.  I'm testing a fix.