On Tue, Jul 08, 2014 at 10:54:36PM +0300, Niko Tyni wrote:
> On Mon, Jul 07, 2014 at 08:00:37AM +0300, Niko Tyni wrote:
> > Package: perl
> > Version: 5.18.2-5
> > Severity: serious
> > X-Debbugs-Cc: debian-m...@lists.debian.org
> >
> > The perl package failed to build on mips:
> >
> > LD_LIBRARY_PATH=/build/perl\-sdOt09/perl\-5.18.2${LD_LIBRARY_PATH:+:}$LD_LIBRARY_PATH
> > ./miniperl -Ilib make_patchnum.pl
> > Bad mojo! at make_patchnum.pl line 111.
>
> > This is reproducible on gabrielli.
>
> A test case would be
>
> gabrielli% cat ~/t.pl
> print ( (q{foo'} =~ /[^']+'/) ? "ok\n" : "not ok\n" );
> gabrielli% perl ~/t.pl
> ok
> gabrielli% ./miniperl ~/t.pl
> not ok
>
> This is a shared build; the static build doesn't exhibit the bug and
> passes the full test suite. The bug only shows up with -fPIC.
>
> An educated guess narrowed it down to regcomp.c. The above test fails
> when regcomp.c is built with gcc-4.9 at -O2 but works with gcc-4.9 at
> -O0 and gcc-4.8 at -O2.
>
> I'll patch the package to lower the optimization level of regcomp.c
> for now. Assuming a test build succeeds on gabrielli, I hope to upload
> tomorrow.
>
> Distilling a proper gcc bug report about this seems rather arduous, and
> I don't really expect to get around to it. Help from the MIPS porters
> on that would be appreciated.
The problem is due to the -ftree-vrp optimization, enabled at -O2. I
will continue investigating, but in the meantime using -fno-tree-vrp
should also workaround the problem without lowering too much the
optimization level.
--
Aurelien Jarno GPG: 4096R/1DDD8C9B
aurel...@aurel32.net http://www.aurel32.net
--
To UNSUBSCRIBE, email to debian-bugs-dist-requ...@lists.debian.org
with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org
