building glibc with gold linker & -frecord-gcc-switches results in internal error

[Cary Coutant]

> Would the error be about a GOT32 relocation in a non-allocated section?
> Or would it be about an unrecognized relocation in a string merge
> section?  I can't quite decide if we should give an error for either of
> those cases or whether we should try to handle them as best we can.

I think both of those conditions are worthy of an error message. I
don't think we want to go allocating GOT and PLT entries based on
relocations in non-allocated sections. For string merge sections, the
only reasonable thing to do other than an error is to turn off the
merge property and process the section as a normal section -- but by
the time we are scanning relocations, I think gold is committed to the
merge processing. I suppose we could make an early check, but I doubt
it's worth it.

> And obviously it would be good to figure out what gcc is doing to create
> those relocations, and try to fix that.

Yes. I've just started looking at gcc, but don't see anything obvious.
Nothing in elf_record_gcc_switches() looks like it would be adding
those relocs to the end of the section.

-cary