Re: [Mips}Using DT tags for handling local ifuncs

[Richard Sandiford <rdsandiford at googlemail dot com>]

Jack Carter <Jack.Carter@imgtec.com> writes:
> This is what I think you want:
>
> *) A region in the GOT that only is fixed up by the relocation
>    mechanism, not by implicit fixup rules.

Right.

> *) This new region will come after the special GOT entries, but before
>    the traditional implicit local entries.

Right.

> *) This new region will have some type of DT pointers and count.

FWIW I was really thinking of just a single new tag to specify the start
of the implicitly-relocated local GOT entries.  That's all we need at
the moment and there doesn't seem much point adding anything fancier
since we can never rely on older ld.sos understanding it.

See the glibc change I sketched upthread.

> *) This new region will be visability neutral. It will contain both
>    global and local symbols.

Right.

> *) I am assuming that this new region will be sorted to put gp
>    relative entries towards the back to reduce multigot triggers.

There's no distinction really.  This new region will only be used
for entries that need to be accessed $gp-relative (i.e. it will only
be used for GOT relocations that are present in the input bfds).

FWIW that's also true of the ABI local GOT.  Only the global GOT is
special because of the "every symbol involved in a relocation must
have a GOT entry" rule.

ISTR the IRIX linker used the same trick as bfd and put the reloc-only
entries last.  I could be misremembering though, since it's many years
since I last had access to IRIX.

> *) I am assuming that this new region will be counted as part of
>    DT_MIPS_LOCAL_GOTNO since the special first entries are also a part of
>    it.

Right.

> If I this is so, I will start implementation.
>
> This discussion has been very useful to me. I am unlearning what I did
> at sgi and starting to understand how things are done for gcc. Many of
> my assumptions come from my past and don't apply here.

I think you're exaggerating the differences really.  Back in 2.16
I got very good test results using binutils on IRIX.  That included
multigot support IIRC.  There are certainly some genuine differences
between IRIX and GNU/Linux (handled by the various SGI_COMPAT and
IRIX_COMPAT tests you can see sprinkled around).  But the basics are
the very much the same.

Of course, GNU/Linux has sprouted various additions that weren't catered
for in the original ABI documents, such as TLS, MIPS16, microMIPS and
the PLT/copy-reloc scheme.  And each new feature has to interoperate
with the previous ones.  I don't think that's the problem here though;
this whole "new region" thing is working around the ABI GOT as originally
defined in the 32-bit MIPS psABI.  The psABI just didn't consider the
possibility that you'd want to apply dynamic relocations to the GOT,
since at the time the only dynamic relocation allowed was R_MIPS_REL32,
and the ABI GOT scheme basically acted like a series of fast R_MIPS_REL32s
(lazy-binding stubs aside).

Sounds like we're on the same page though. :-)  Thanks for working through it.

Thanks,
Richard