https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98497
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed| |2021-01-05
Version|unknown |11.0
Known to fail| |11.0
Ever confirmed|0 |1
CC| |hubicka at gcc dot gnu.org,
| |rguenth at gcc dot gnu.org
Status|UNCONFIRMED |NEW
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue is a completely hosed profile:
<bb 2> [local count: 108459]:
_17 = x_15(D) + 131056;
ivtmp.25_16 = (unsigned long) _17;
ivtmp.26_41 = (unsigned long) x_15(D);
_53 = ivtmp.26_41 + 4294836212;
<bb 3> [local count: 108459]:
# _43 = PHI <0.0(2), sum_50(5)>
# ivtmp.25_45 = PHI <ivtmp.25_16(2), ivtmp.25_24(5)>
# ivtmp.26_46 = PHI <ivtmp.26_41(2), ivtmp.26_27(5)>
<bb 4> [local count: 650754]:
# vect__44.6_21 = PHI <{ 0.0, 0.0, 0.0, 0.0 }(3), vect_sum_39.10_25(4)>
# ivtmp.15_19 = PHI <ivtmp.26_46(3), ivtmp.15_44(4)>
_10 = (void *) ivtmp.15_19;
vect__34.9_28 = MEM <vector(4) float> [(float *)_10];
vect_sum_39.10_25 = vect__44.6_21 + vect__34.9_28;
ivtmp.15_44 = ivtmp.15_19 + 16;
if (ivtmp.15_44 != ivtmp.25_45)
goto <bb 4>; [100.00%]
else
goto <bb 5>; [0.00%]
<bb 5> [count: 0]:
_38 = .REDUC_PLUS (vect_sum_39.10_25);
_51 = (void *) ivtmp.25_45;
_34 = MEM[(float *)_51];
_5 = MEM[(float *)_51 + 4B];
_1 = _5 + _34;
_49 = MEM[(float *)_51 + 8B];
_52 = _43 + _49;
_35 = _1 + _52;
sum_50 = _35 + _38;
ivtmp.25_24 = ivtmp.25_45 + 131068;
ivtmp.26_27 = ivtmp.26_46 + 131068;
if (ivtmp.25_24 != _53)
goto <bb 3>; [100.00%]
else
goto <bb 6>; [0.00%]
<bb 6> [count: 0]:
return sum_50;
We recompute the local profile after graphite but that doesn't seem to result
in anything sensible:
Predictions for bb 21
first match heuristics: 99.00% exec 54230 (estimated locally) hit 54230
(estimated locally) (100.0%)
combined heuristics: 99.00% exec 54230 (estimated locally) hit 54230
(estimated locally) (100.0%)
loop iterations heuristics of edge 21->20: 1.00% exec 54230 (estimated
locally) hit 108460 (estimated locally) (200.0%)
Predictions for bb 22
1 edges in bb 22 predicted to even probabilities
Predictions for bb 20
first match heuristics: 99.00% exec 108459 (estimated locally) hit 108459
(estimated locally) (100.0%)
combined heuristics: 99.00% exec 108459 (estimated locally) hit 108459
(estimated locally) (100.0%)
loop iterations heuristics of edge 20->15: 1.00% exec 108459 (estimated
locally) hit 216918 (estimated locally) (200.0%)
GRAPHITE does (graphite.c:graphite_transform_loops)
if (changed)
{
cleanup_tree_cfg ();
profile_status_for_fn (cfun) = PROFILE_ABSENT;
release_recorded_exits (cfun);
tree_estimate_probability (false);
}
not releasing recorded exits doesn't help.
