--gc-sections causes assertion failure

[slide]

When running with the Linaro GCC 4.8.1 2013.08 release, I get the following assertion when using -ffunction-sections and -fdata-sections to GCC along with --gc-sections to LD. From searching the web, it looks like there are various possible reasons for the assertion in upstream binutils, but I couldn't narrow it down directly as they were slightly different.

c:/gnu/arm-none-eabi-4.8.1/bin/../lib/gcc/arm-none-eabi/4.8.1/../../../../arm-none-eabi/bin/ld.exe: BFD (GNU Binutils) 2.23.2 assertion fail ../../../src/binutils-2.23.2/bfd/elf32-arm.c:12334

This corresponds to the first assert in the code below:

if (may_need_local_target_p
   && elf32_arm_get_plt_info (abfd, eh, r_symndx, &root_plt, &arm_plt))
 {
   /* If PLT refcount book-keeping is wrong and too low, we'll
      see a zero value (going to -1) for the root PLT reference
      count. */
   if (root_plt->refcount >= 0)
     {
       BFD_ASSERT (root_plt->refcount != 0);
       root_plt->refcount -= 1;
     }
   else
     /* A value of -1 means the symbol has become local, forced
        or seeing a hidden definition. Any other negative value
        is an error. */
     BFD_ASSERT (root_plt->refcount == -1);

[Charles Baylis (cbaylis)]

Thanks for the bug report. Unfortunately, I can't reproduce this problem, could you provide an example or a link to code which causes this?

[slide (slide-o-mix)]

I will try and create a minimal test case to reproduce the issue and attach it. I forgot to mention that I am using the mingw32 version of the compiler.

[Charles Baylis (cbaylis)]

Did you manage to reproduce this?

[slide (slide-o-mix)]

I haven't been able to reproduce it outside of my environment yet. Is there something I can look for in my code that might cause this issue to appear so I can zero in a test case a little better?

[slide (slide-o-mix)]

I'm more than willing to add in additional debug to the binutils source as well if that will help, I just don't know what add and to where.

[Will Newton (will-newton)]

Can you still reproduce this issue?

I haven't been able to come up with a way that the code could get the refcounts out of sync in this way (which doesn't mean its not possible, just that its not obvious).

Even a moderately complex reproducer is fine as long as it comes with clear instructions.

Thanks,

[slide (slide-o-mix)]

I can't seem to replicate this with the 4.9 release from 2014-07, so I think it can be closed out.