Set the retain attribute on _elf_set_element if CC supports [BZ #27492]

  So that text_set_element/data_set_element/bss_set_element defined
variables will be retained by the linker.

Note: 'used' and 'retain' are orthogonal: 'used' makes sure the variable
will not be optimized out; 'retain' prevents section garbage collection
if the linker support SHF_GNU_RETAIN.

GNU ld 2.36 and LLD 13 support -z start-stop-gc which allow C identifier
name sections to be GCed even if there are live __start_/__stop_
references.

Without the change, there are some static linking problems, e.g.
_IO_cleanup (libio/genops.c) may be discarded by ld --gc-sections, so
stdout is not flushed on exit.

Note: `#if defined (__has_attribute) && __has_attribute (retain)`
does not work in GCC assembly mode:

    error: missing binary operator before token "("
---
 include/libc-symbols.h | 24 ++++++++++++++++++++----
 1 file changed, 20 insertions(+), 4 deletions(-)
  

* Fangrui Song via Libc-alpha:

> So that text_set_element/data_set_element/bss_set_element defined
> variables will be retained by the linker.
>
> Note: 'used' and 'retain' are orthogonal: 'used' makes sure the variable
> will not be optimized out; 'retain' prevents section garbage collection
> if the linker support SHF_GNU_RETAIN.

This needs recent-ish GCC 11, see commit 6347f4a0904fce17e ("Add retain
attribute to place symbols in SHF_GNU_RETAIN section).

> +#ifdef __has_attribute
> +#if __has_attribute (retain)
> +# ifdef SHARED

Indentation is off there.

It's probably simpler to define a new macro, like
ELEMENT_SECTION_ATTRIBUTES, and use that unconditionally, perhaps like
this?

#ifndef __ASSEMBLER__
# if defined (__has_attribute) && __has_attribute (retain)
#  define ELEMENT_SECTION_ATTRIBUTES used, retain
# else
#  define ELEMENT_SECTION_ATTRIBUTES used
# endif
#endif

Have you checked that __has_attribute (retain) is only true on GCC if
the underlying linker has retain support and the attribute does not
produce a warning?

Thanks,
Florian
  

On Mon, Mar 15, 2021 at 12:28 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> * Fangrui Song via Libc-alpha:
>
> > So that text_set_element/data_set_element/bss_set_element defined
> > variables will be retained by the linker.
> >
> > Note: 'used' and 'retain' are orthogonal: 'used' makes sure the variable
> > will not be optimized out; 'retain' prevents section garbage collection
> > if the linker support SHF_GNU_RETAIN.
>
> This needs recent-ish GCC 11, see commit 6347f4a0904fce17e ("Add retain
> attribute to place symbols in SHF_GNU_RETAIN section).
>
> > +#ifdef __has_attribute
> > +#if __has_attribute (retain)
> > +# ifdef SHARED
>
> Indentation is off there.

OK. I can increase the level of indentation...

> It's probably simpler to define a new macro, like
> ELEMENT_SECTION_ATTRIBUTES, and use that unconditionally, perhaps like
> this?
>
> #ifndef __ASSEMBLER__
> # if defined (__has_attribute) && __has_attribute (retain)
> #  define ELEMENT_SECTION_ATTRIBUTES used, retain
> # else
> #  define ELEMENT_SECTION_ATTRIBUTES used
> # endif
> #endif

`#ifdef __has_attribute` should be fine without #ifndef __ASSEMBLER__ ?

> Have you checked that __has_attribute (retain) is only true on GCC if
> the underlying linker has retain support and the attribute does not
> produce a warning?

There is a GCC bug. I've reported
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99587

>
> Thanks,
> Florian
>
  

* Fāng-ruì Sòng:

>> It's probably simpler to define a new macro, like
>> ELEMENT_SECTION_ATTRIBUTES, and use that unconditionally, perhaps like
>> this?
>>
>> #ifndef __ASSEMBLER__
>> # if defined (__has_attribute) && __has_attribute (retain)
>> #  define ELEMENT_SECTION_ATTRIBUTES used, retain
>> # else
>> #  define ELEMENT_SECTION_ATTRIBUTES used
>> # endif
>> #endif
>
> `#ifdef __has_attribute` should be fine without #ifndef __ASSEMBLER__ ?

ELEMENT_SECTION_ATTRIBUTES is not useful to the assembler, and checking
for __ASSEMBLER__ on the outer level makes it simpler to define the
no-retain case.

>> Have you checked that __has_attribute (retain) is only true on GCC if
>> the underlying linker has retain support and the attribute does not
>> produce a warning?
>
> There is a GCC bug. I've reported
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99587

If this isn't fixed in GCC, we'll need a configure check because we want
to keep building with -Werror=attributes.

Thank,s
Florian