Re: Assertion failed when building r-base on Aranym
On Sun, 4 Dec 2016, John Paul Adrian Glaubitz wrote:
> On 12/04/2016 12:50 AM, Finn Thain wrote:
> > When I googled that assertion failure, I got the impression that
> > whilst the assertion is in glibc code, the bug usually isn't. The
> > assertion gets triggered when glibc data structures get clobbered by a
> > program that overflows an allocated block etc.
> >
> >> /bin/bash: line 1: 17541 Done echo
> >> "tools:::data2LazyLoadDB(\"datasets\", compress=3)"
> >> 17542 Aborted | R_DEFAULT_PACKAGES=NULL LC_ALL=C ../../../bin/R --vanilla --slave > /dev/null
> >> Makefile:21: recipe for target 'all' failed
> >>
> >
> > ... which leads me to think the bug is probably in this R binary.
>
> Sounds reasonable, however, there is one problem with that theory.
I can see how these facts support alternative theories, but they don't
actually refute mine...
> As I have already mentioned, the problem also affects a version of
> r-base which is known to be good as it previously built fine:
>
> good: https://buildd.debian.org/status/fetch.php?pkg=r-base&arch=m68k&ver=3.2.2-1&stamp=1445983541
> bad: https://buildd.debian.org/status/fetch.php?pkg=r-base&arch=m68k&ver=3.2.2-1&stamp=1480420610
>
> Same version. Different default gcc (gcc-5 vs. gcc-6), different glibc
> (2.19 vs. 2.24), different default compiler options (dpkg is injecting
> -fPIE. Kernel is identical.
>
Since the assertion is in glibc, it isn't surprising when the bug becomes
visible when glibc is upgraded.
> I'm currently testing whether this is a result of dpkg defaulting the
> build to -fPIE.
>
Well, that should be interesting.
If gcc-6 with -fPIE failed and gcc-5 without -fPIE succeeded, I'd be
curious about both gcc-5 with -fPIE and gcc-6 without -fPIE.
--
Reply to: