http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45988
Rainer Orth <ro at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|RESOLVED |REOPENED
Last reconfirmed| |2012-03-02
CC| |ubizjak at gmail dot com
Version|4.6.0 |4.7.0
Resolution|FIXED |
Ever Confirmed|0 |1
--- Comment #4 from Rainer Orth <ro at gcc dot gnu.org> 2012-03-02 16:24:24 UTC
---
I'm still seeing this bug on a particular machine, sometimes also for the
64-bit
case.
E.g. on zingana (4 x Quad-Core 2.93 GHz Intel Xeon X7350 running Solaris 10),
fast-math-pr38968.exe takes
real 1:38.46
user 1:37.61
sys 0.53
The exact same binary and libraries on arenal (8 x Dual-Core 2.593 GHz AMD
Opteron 8218 running Solaris 11) takes
real 38.09
user 37.60
sys 0.40
instead. I've used the collect tool from the Studio 12.3 compilers to
investigate
and find that a single insn seems to dominate the runtime:
On zingana:
## 64.925 64.925 [?] 400f0c: movaps 0x10(%rcx,%rax),%xmm1
On arenal:
## 13.810 13.810 [?] 400f0c: movaps 0x10(%rcx,%rax),%xmm1
I'll attach the complete profiles.
Rainer
