Re: [PATCH v2] Handle private COMDAT function symbol reference in readonly data section

["H.J. Lu" <hjl.tools@gmail.com>]

On Wed, Jan 31, 2024 at 9:10 AM Jakub Jelinek <jakub@redhat.com> wrote:
>
> On Wed, Jan 31, 2024 at 08:48:33AM -0800, H.J. Lu wrote:
> > Which function (target hook) can I use to generate
> >
> >  .section        .data.rel.ro.local,"awG",@progbits,_ZN1AIxE3fooExx,comdat
>
> Just
>       if (decl)
>         return get_section (reloc == 1
>                             ? ".data.rel.ro.local" : ".data.rel.ro",
>                             SECTION_WRITE | SECTION_RELRO | SECTION_LINKONCE,
>                             decl);
> for the first hunk and
>   else if (decl)
>     return get_section (".rodata", SECTION_LINKONCE, decl);
> in the second case?
>
> Haven't tried it though, maybe the get_section section conflict stuff
> isn't able to handle it (and perhaps that is the reason why we simply don't
> emit functions into
>         .section        .text,"axG",@progbits,whatever,comdat
> sections instead of
>         .section        .text.function_name,"axG",@progbits,whatever,comdat

GNU binutils has no issues with it:

[hjl@gnu-cfl-3 tmp]$ cat x.s
.section .text,"axG",@progbits,whatever,comdat
nop
.text
nop
[hjl@gnu-cfl-3 tmp]$ gcc -c x.s
[hjl@gnu-cfl-3 tmp]$ readelf -SW x.o | grep text
  [ 2] .text             PROGBITS        0000000000000000 000048
000001 00  AX  0   0  1
  [ 5] .text             PROGBITS        0000000000000000 000049
000001 00 AXG  0   0  1
[hjl@gnu-cfl-3 tmp]$

If it doesn't work for some targets, we can use
targetm.asm_out.function_rodata_section.

> In such case, we could append something to those section names,
> like const.pool (or whatever else that couldn't clash with function names,
> so needs probably a dot somewhere).  Could be
> .data.rel.ro.local.const.pool
> .data.rel.ro.const.pool
> or
> .data.rel.ro.local..shared
> .data.rel.ro..shared
> or something similar, but .data.rel.ro.local.shared would not be ok,
> because it could clash with .data.rel.ro.local section for shared function.
>
> > > Another question is if we need to do anything about the
> > > DECL_COMDAT_GROUP (decl) && DECL_SECTION_NAME (decl)
> > > && startswith (DECL_SECTION_NAME (decl), ".gnu.linkonce.t.")
> > > case (older linkers) (i.e. when using years old GNU linkers).
> > >
> >
> > Should we support such targets? It is not easy for me to test it.
>
> Perhaps let's wait if somebody files an issue with such configuration.
>
>         Jakub
>


-- 
H.J.