Re: [PATCH] x86: Keep __patchable_function_entries sections with --gc-sections

[Fangrui Song <i at maskray dot me>]

On 2020-02-01, H.J. Lu wrote:
> On Sat, Feb 1, 2020 at 9:34 AM Fangrui Song <i@maskray.me> wrote:
> > On 2020-02-01, H.J. Lu wrote:
> > >After all text sections have been garbage collected, if a
> > >__patchable_function_entries section references a section which
> > >wasn't marked, mark it with SEC_EXCLUDE and return NULL.  Otherwise,
> > >keep it.
> > >
> > >Should it be handled in _bfd_elf_gc_mark_extra_sections?
> >
> > Thanks for paying attention to these feature requests.
> >
> > I referenced GNU as and ld requests at
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93492#c2
> > If we
> >
> > * implement SHF_LINK_ORDER
>
> I am not sure if overloading (abusing?) SHF_LINK_ORDER is a good idea.

It is an extension, but it is agreed by multiple parties on
https://groups.google.com/d/msg/generic-abi/_CbBM6T6WeM/eGF9A0AnAAAJ ,
including HP-UX and Solaris developers.

> > * allow multiple sections with the same name ("unique")
>
> This is orthogonal to this.  I have a question on assembly syntax:
>
> https://sourceware.org/bugzilla/show_bug.cgi?id=25380#c1
>
> > * teach GCC to use SHF_LINK_ORDER and "unique" (see https://gcc.gnu.org/ml/gcc/2020-01/msg00067.html)
> >
> > An ad-hoc gc marking will be unnecessary.
>
> We need to scan relocations in _patchable_function_entries section for
> references to garbage collected sections.   We can either check section
> name or a SHF_XXX.  But I don't know if SHF_LINK_ORDER is a good
> approach.

We don't need to if we use multiple __patchable_function_entries
sections. Multiple sections are a must due to
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93195#c1 (COMDAT)

> A symbol table entry with STB_LOCAL binding that is defined relative
> to one of a group's sections, and that is contained in a symbol table
> section that is not part of the group, must be discarded if the group
> members are discarded. References to this symbol table entry from
> outside the group are not allowed.

The __patchable_function_entries creation logic I implemented in clang:

foreach function foo
  find the first function label defined in foo's section, name it $associated
    ($associated can have 2 reasonable values, w/ or w/o -ffunction-sections)
  get or create an id for $associated, say, $unique

  if foo is in a COMDAT named $comdat
    .section __patchable_function_entries,"awo",@progbits,$associated,comdat,$comdat,unique,$unique
  else
    .section __patchable_function_entries,"awo",@progbits,$associated,unique,$unique

This approach uses feature requests I referenced in *direct* links of
https://gcc.gnu.org/ml/gcc/2020-01/msg00067.html plus
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93492#c2 ,
and addresses all bugs I filed.

> > SHF_LINK_ORDER has been used in a few sanitizers. Now we know
> > __patchable_function_entries can benefit from it.  In the future, they
> > may be instances. We really need a general solution.
>
>
> --
> H.J.