This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
Re: optimization/8599: [3.2 regression] loop unroll bug with -march=k6-3
On Wednesday 04 December 2002 16:34, Eric Botcazou wrote:
> > Well, Dale's patch1 is already checked in, this was the regression
> > causing one. If that's what you mean, yes, that's what I meant too :-).
>
> Not exactly, I was considering the patches in chronological order. Here's
> the lexicon:
> - rth's patch: the backported patch from rth you posted,
> - dale's patch1: the backported patch from Dale I posted,
> - dale's patch2: the backported patch already checked in,
OK, but the net result is the same :-).
> > My bootstraps on powerpc-linux-gnu and i686-linux-gnu with this setup
> > completed without regressions.
>
> gcc 3.2.1: 5 failures
> FAIL: g77.f-torture/execute/intrinsic-unix-bessel.f execution, -O0
> FAIL: g77.f-torture/execute/intrinsic-unix-bessel.f execution, -O1
> FAIL: g77.f-torture/execute/intrinsic-unix-bessel.f execution, -O2
> FAIL: g77.f-torture/execute/intrinsic-unix-bessel.f execution, -O3 -g
> FAIL gcc.dg/i386-unroll-1.c
>
> gcc 3.2.1 + rth's patch: 1 failure
> FAIL: g77.f-torture/execute/intrinsic-unix-bessel.f execution, -O0
>
> gcc 3.2.1 + rth's patch + Dale's patch1: 1 failure
> FAIL: g77.f-torture/execute/intrinsic-unix-bessel.f execution, -O0
>
> gcc 3.2.1 + rth's patch + Dale's patch1 + Dale's patch2: 2 failures
> FAIL: g77.f-torture/execute/intrinsic-unix-bessel.f execution, -O0
> FAIL: g77.f-torture/execute/intrinsic-unix-bessel.f execution, -O2
>
>
> However the failures of g77.f-torture/execute/intrinsic-unix-bessel.f are
> NOT reproducible: sometimes the execution is ok, sometimes it is not. And
> this happens with mainline too on i586 (except at -Os).
Well, on rh73/i686 I see a similar thing with:
FAIL: Thread_Interrupt output from source compiled test
It moved from a -funsigned-char failure to a -O2 failure for me. But this test
seems to be very fragile in general, as a lot more variations of this test
seem to fail if the machine is loaded.
But I agree that all these effects are unrelated to the patches, so now we
just need somebody for the branch approval as the patches fix the testsuite
regression on PPC without introducing a code quality regression.
Franz.