https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55334
vp at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |vp at gcc dot gnu.org
--- Comment #37 from vp at gcc dot gnu.org ---
I see this issue in AArch64 as well on 4.9 branch and on the trunk causing
performance regression.
Since MG3P is inlined at MAIN__, all the four calls to RESID are in MAIN__. Two
of which are to the clone resid_.constprop.1 (not vectorized) and one to resid_
(vectorized) and the last one inlined (not vectorized and it looks very similar
to the resid_.constprop.1 version).
I could however, improve by passing -fno-inline-functions and -fno-ipa-cp at
-Ofast (so that all the calls are to resid_ which gets vectorized).
Anyone working on this issue?