On Mon, Jan 11, 2016 at 10:39:58AM -0700, Jeff Law wrote:
> On 01/11/2016 10:10 AM, Jakub Jelinek wrote:
> >Hi!
> >
> >As mentioned in the PR, this is a test new in GCC 6, that relies on rtx
> >costs to be right to succeed, and on i?86 (32-bit) they are not right.
> >While the costs look weird on i?86, it is cost code in generic code, and I'm
> >afraid changing those is too risky now so late in stage3, so I'm just
> >proposing to restrict this test to the architectures where it is known to
> >pass right now.
> >
> >Tested on x86_64-linux and i686-linux, ok for trunk?
> >
> >2016-01-11 Jakub Jelinek <ja...@redhat.com>
> >
> > PR target/67462
> > * gcc.dg/ifcvt-3.c: Only compile on lp64 targets, include also i?86
> > if lp64.
> >
> >--- gcc/testsuite/gcc.dg/ifcvt-3.c.jj 2015-09-03 17:18:33.000000000
> >+0200
> >+++ gcc/testsuite/gcc.dg/ifcvt-3.c 2016-01-09 09:05:42.234836729 +0100
> >@@ -1,4 +1,4 @@
> >-/* { dg-do compile { target aarch64*-*-* x86_64-*-* } } */
> >+/* { dg-do compile { target { { aarch64*-*-* i?86-*-* x86_64-*-* } && lp64
> >} } } */
> > /* { dg-options "-fdump-rtl-ce1 -O2" } */
> Would it make more sense to xfail the other targets rather than not running
> it at all on those targets?
The testcase already is limited to just 2 targets, and anything relating to
conditional moves is extremely controversial on i?86/x86_64 anyway, because
often it hurts performance instead of improving it. Plus for pre-i686 there
are no conditional moves and it might be likely even more expensive in real
world. So, I view it as primarily an aarch64 testcase that happens to pass
these days on x86_64 64-bit too, but is really very beneficial only on
aarch64.
Jakub
