bug#38055: patchelf: Assertion failed when setting interpreter

["Ludovic Courtès" <ludo@gnu.org>]

Bona nit!

Ivan Vilata i Balaguer <ivan@selidor.net> skribis:

> Yeah, as I mentioned in the original mail that particular problem does indeed
> seem to be fixed in 0.10.  However when I try to build that source with `guix
> build patchelf --with-source=…`, tests fail.
>
> If I run `guix environment -C --pure patchelf` then unpack and build the
> source, the only test that actually fails is `no-rpath.sh`.  If I run `sh -x
> no-rpath.sh` I get this:
>
> ```
> ++ basename no-rpath.sh .sh
> + SCRATCH=scratch/no-rpath
> + rm -rf scratch/no-rpath
> + mkdir -p scratch/no-rpath
> + cp no-rpath scratch/no-rpath/
> ++ ../src/patchelf --print-rpath scratch/no-rpath/no-rpath
> + oldRPath=/gnu/store/ahqgl4h89xqj695lgqvsaf6zh2nhy4pj-glibc-2.29/lib:/gnu/store/2plcy91lypnbbysb18ymnhaw3zwk8pg1-gcc-7.4.0-lib/lib:/gnu/store/2plcy91lypnbbysb18ymnhaw3zwk8pg1-gcc-7.4.0-lib/lib/gcc/x86_64-unknown-linux-gnu/7.4.0/../../..:/gnu/store/dcrwf5irwh39knld1wim1qkny659af9g-profile/lib
> + test -n /gnu/store/ahqgl4h89xqj695lgqvsaf6zh2nhy4pj-glibc-2.29/lib:/gnu/store/2plcy91lypnbbysb18ymnhaw3zwk8pg1-gcc-7.4.0-lib/lib:/gnu/store/2plcy91lypnbbysb18ymnhaw3zwk8pg1-gcc-7.4.0-lib/lib/gcc/x86_64-unknown-linux-gnu/7.4.0/../../..:/gnu/store/dcrwf5irwh39knld1wim1qkny659af9g-profile/lib
> + exit 1
> ```
>
> To succeed, the output of `…/patchelf --print-rpath …/no-rpath`
> (i.e. `oldRPath`) should be empty.  I'm not that familiar with Guix's GNU
> build system, but is that at all possible under Guix?  I mean, maybe the test
> is pointless or must be altered in some Guix-specific way for the `no-rpath`
> binary not to have an rpath.

Guix’ ‘gcc’ automatically adds RUNPATH entries to glibc/lib and
gcc:lib/lib:

--8<---------------cut here---------------start------------->8---
$ gcc -dumpspecs |grep -e -rpath
%{!mandroid|tno-android-ld:-L/gnu/store/ahqgl4h89xqj695lgqvsaf6zh2nhy4pj-glibc-2.29/lib %{!static:-rpath=/gnu/store/ahqgl4h89xqj695lgqvsaf6zh2nhy4pj-glibc-2.29/lib %{!static-libgcc:-rpath=/gnu/store/347y0zr1a9s2f5pkcncgi3gd0r33qq81-gcc-9.2.0-lib/lib -lgcc_s}} %{pthread:-lpthread} %{shared:-lc}    %{!shared:%{profile:-lc_p}%{!profile:-lc}};:%{shared:-lc}    %{!shared:%{profile:-lc_p}%{!profile:-lc}} %{!static: -ldl}}
--8<---------------cut here---------------end--------------->8---

The code that modifies GCC to do that is in (gnu packages gcc).

Thus, RUNPATH is never empty, and the test above is bound to fail.

Two possibilities: change the test to ensure that ‘--print-rpath’
returns precisely libc/lib:gcc/lib, or, if that turns out to be tricky,
skip the test (in both cases, add a comment linking to this discussion
for future reference.)

How does that sound?

Thank you!

Ludo’.