https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97160
Bug ID: 97160
Summary: Regression from GCC 8 optimizing to sincos on ppc64le
Product: gcc
Version: 10.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: fx at gnu dot org
Target Milestone: ---
Target: ppc64le
Created attachment 49254
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49254&action=edit
Fortran benchmark
I looked at the other case in a Fortran benchmark set where GCC does badly on
ppc64le with GCC 10. The problem seems to be in the middle end, though, not
Fortran.
This case is heavy on trig functions, and I found that gfortran 10 and 9 miss
the optimization of converting sin+cos to sincos which is done by 8 on ppc64le
and also by 10 on x86_64. I tried a simple example rather than the whole
thing, but couldn't quickly reproduce the effect out of context, so it may be
rather specific to this case. I'll attach the original 1400-line single
routine (!) but I could take quick suggestions to narrow down the issue.
With gfortran 10.2 (and similarly with 9) on RHEL7 POWER9 the top of the perf
profile resulting from "gfortran -Ofast" is
43.60% mp_prop_design libm-2.17.so [.] __cos
28.85% mp_prop_design mp_prop_design [.] MAIN__
24.40% mp_prop_design libm-2.17.so [.] __sin
1.04% mp_prop_design libm-2.17.so [.] __sincos
With gfortran 8 (from "advanced toolkit 12") it's
78.49% a.out mp_prop_design [.] MAIN__
20.29% a.out libm-2.28.so [.] __sincos
1.14% a.out libm-2.28.so [.] __tan
(Those look as if they're from different systems because AT 12 picks up its own
glibc when you set up the environment.)
