https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118551
Bug ID: 118551
Summary: Autofdo regressed 538.imagick_r by ~10% with
-march=x86-64-v3 -O2
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: gcov-profile
Assignee: unassigned at gcc dot gnu.org
Reporter: liuhongt at gcc dot gnu.org
Target Milestone: ---
Similar like PR116743, it's related to ipa scaling, but in different
place(estimate_bb_frequencies).
/* If we have profile feedback in which this function was never
executed, then preserve this info. */
if (!(bb->count == profile_count::zero ()))
bb->count = count.guessed_local ().combine_with_ipa_count
(ipa_count);
ipa_count is the count of the entry block and it's zero, but the whole function
is hot. estimate_bb_frequencies scale every non-zero bb count to zero since
ipa_count is zero, and GCC optimized those BB for size since it thought the bb
is cold.
part to gcov_dump from the hot function:
MeanShiftImage total:17008941 head:0
2: 0
25: 0
26: 0
29: 0
30: 0
31: 0
32: 0
32.1: 0
34: 0
35: 0
39: 0
40: 0
41: 0
42: 0
47.1: 0
47.2: 0
61: 0
63: 0
64: 1
66: 1
71: 1 GetCacheViewVirtualIndexQueue:1
72.1: 2189
72.2: 2189
85: 2197 GetMagickPixelPacket:2268
87: 2171
88: 2171
89.1: 2171
105: 3788
106: 3788
107: 3811 GetMagickPixelPacket:3931
109: 3788
110: 3788
111: 0
111.1: 72754
111.2: 72754
116: 72731
116.1: 816545
116.2: 816545
118: 1381111
123: 967682 GetOneCacheViewVirtualPixel:998671
