Re: [PATCH 0/4] alpha: add IFUNC support

On Sun, Aug 9, 2026 at 5:41 PM Matt Turner <[email protected]> wrote:
>
> Testing this series against glibc turned up a case it got wrong, so I have
> sent v2 of two of the four patches, each in reply to the patch it replaces:
>
>   [PATCH v2 2/4] alpha: resolve IFUNCs this link resolves itself out of
>                  .rela.iplt
>   [PATCH v2 4/4] alpha: handle IFUNC symbols with local binding
>
> 1/4 and 3/4 are unchanged.  3/4 still applies on top of v2 2/4; only its
> hunk offsets move.
>
> An IFUNC referenced from a non-PIE executable killed the linker:
>
>   BFD assertion fail bfd/elf64-alpha.c:4216
>   collect2: fatal error: ld terminated with signal 11 [Segmentation fault]
>
> Such an IFUNC cannot be preempted, so it never becomes a dynamic symbol
> and the sizing code reserves nothing for it in .rela.got or .rela.<sec>.
> check_relocs knew that and reserved .rela.iplt space instead, but
> relocate_section decided where to emit by asking whether the link had
> dynamic sections at all, and so took the ordinary path and was handed the
> dynamic reloc section for the input section, which is a null pointer
> because nothing ever created one.
>
> v2 2/4 makes both sides ask the same question -- does this link resolve
> the IFUNC itself, which is so both when there is no dynamic linker and
> when the link is not position-independent.  It also sizes the .rela.iplt
> space for such a symbol's GOT entries in a dynamic link, which only the
> static path did before, and skips a symbol with a PLT entry in both the
> sizing and the emitting, since finish_dynamic_symbol handles those and
> .rela.iplt would otherwise overrun.
>
> Because the fix belongs where the mistake was, I folded it into 2/4 rather
> than adding a fifth patch: without it, 2/4 and 3/4 segfault the linker on
> an ordinary non-PIE executable, which is unpleasant to bisect through.
> 2/4 is retitled, since it no longer only concerns static links.
>
> Each patch builds on its own, and check-ld at the tip of the series has no
> unexpected failures.  New tests cover a dynamic executable that references
> and calls an IFUNC, with local and with global binding; both fail without
> this change.
>
> With the series applied, glibc builds on alpha with --enable-multi-arch
> and its IFUNC tests link, where before six of them did not.
>
> Thanks,
> Matt

Anything I can to do to help get this series (or any of my other
recent patches) reviewed?