PING^n fwprop and pr/19653

Roger Sayle roger@eyesopen.com
Tue Apr 11 04:17:00 GMT 2006


Hi Paolo,

On Tue, 28 Mar 2006, Roger Sayle wrote:
> On Tue, 28 Mar 2006, Paolo Bonzini wrote:
> > [PATCH] PR/19653: x87 registers used for reloads when in SSE math mode
> > http://gcc.gnu.org/ml/gcc-patches/2006-01/msg01825.html
> > Pretty please?
>
> Changes to reload have traditionally been tricky, so be prepared to
> revert or fix these changes quickly should anyone report problems.

It looks like there's yet more fallout from your PR target/19653
patch.  This time I'm seeing three new testsuite failures on
x86_64-unknown-linux-gnu.

FAIL: gcc.c-torture/execute/20050121-1.c execution,  -O0
FAIL: gcc.c-torture/execute/complex-6.c execution,  -O0
FAIL: gcc.target/x86_64/abi/test_complex_returning.c execution,  -O0

Andrew Pinksi has managed to narrow down the regression to between
r112634: http://gcc.gnu.org/ml/gcc-testresults/2006-04/msg00176.html
r112655: http://gcc.gnu.org/ml/gcc-testresults/2006-04/msg00194.html

I investigated the problem with complex-6.c and the point of divergence
is in .136r.lreg where we choose different register preferences
from before your change.  I also tried the "force_reg" fix that
David Edelsohn and I looked into for the AIX fallout, but this
failure seems unrelated to his incorrect SUBREG issue.


I was wondering if you could a look?  It may be another latent bug
that's been exposed by your and Dale's changes.

Thanks in advance,

Roger
--



More information about the Gcc-patches mailing list