This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
[Bug tree-optimization/79390] 10% performance drop in SciMark2 LU after r242550
- From: "jakub at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>
- To: gcc-bugs at gcc dot gnu dot org
- Date: Fri, 24 Feb 2017 14:49:32 +0000
- Subject: [Bug tree-optimization/79390] 10% performance drop in SciMark2 LU after r242550
- Auto-submitted: auto-generated
- References: <bug-79390-4@http.gcc.gnu.org/bugzilla/>
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79390
--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Perhaps we just want to limit number of cmov converted insns in a bb rather
than all noce* attempts, or we should also check how far is the user of the
conditional move; on this exact testcase we turn one of the conditional insns
into a maxsd (which supposedly has reasonable latency) and another one into
cmov, but that one is consumed only by itself from earlier iteration and then
whatever called the function.
Perhaps noce_convert_multiple_sets should call in addition to
noce_conversion_profitable_p some target hook that in the x86 case would count
real cmove insns and for the ONE_IF_CONV... targets reject it only if there are
more than one real cmove? I'm also surprised that the
noce_convert_multiple_sets doesn't try all those other cases like abs, minmax,
addcc, sign_mask etc. But apparently the backend manages to expand at least
the minmax as such.