https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128022
Bug ID: 128022
Summary: [16/17 Regression] 1.6x slowdown of SPEC CPU2026
766.femflow_r on Zen5 with PGO
Product: gcc
Version: 16.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: vekumar at gcc dot gnu.org
Target Milestone: ---
Profile-guided optimization makes SPEC CPU2026 766.femflow_r substantially
slower than -O3. The slowdown is in IPA-SRA clones of
dealii::operator*(const double&, const dealii::VectorizedArray<double, 8>&)
At -O3 those clones are unrolled and vectorized (SSE2 mulpd). With
-fprofile-use, most of the clones become a 64-byte stack copy and a scalar
mulsd loop. Recompiling the profile-use pass with -fno-ipa-sra removes the
clones and turns the PGO slowdown into a speedup.
How to reproduce
----------------
Licensed SPEC CPU2026 tree required. No source edits.
default=base:
OPTIMIZE = -g -O3
default=peak:
OPTIMIZE = -g -O3
PASS1_FLAGS = -fprofile-generate
PASS2_FLAGS = -fprofile-use
runcpu -c <cfg> -n 1 -C 1 --size=train --tune=base,peak 766.femflow_r
