[PATCH] "used" attribute saves decl from linker garbage collection

[Jozef Lawrynowicz]

On Tue, Nov 03, 2020 at 11:58:04AM -0800, H.J. Lu via Gcc-patches wrote:
> On Tue, Nov 3, 2020 at 10:22 AM Jozef Lawrynowicz
> <jozef.l@mittosystems.com> wrote:
> >
> > On Tue, Nov 03, 2020 at 09:57:58AM -0800, H.J. Lu via Gcc-patches wrote:
> > > On Tue, Nov 3, 2020 at 9:41 AM Jozef Lawrynowicz
> > > <jozef.l@mittosystems.com> wrote:
> > > >
> > > > The attached patch implements TARGET_ASM_MARK_DECL_PRESERVED for ELF GNU
> > > > OSABI targets, so that declarations that have the "used" attribute
> > > > applied will be saved from linker garbage collection.
> > > >
> > > > TARGET_ASM_MARK_DECL_PRESERVED will emit an assembler ".retain"
> > >
> > > Can you use the "R" flag instead?
> > >
> >
> > For the benefit of this mailing list, I have copied my response from the
> > Binutils mailing list regarding this.
> > The "comm_section" example I gave is actually innacurate, but you can
> > see the examples of the variety of sections that would need to be
> > handled by doing
> >
> > $ git grep -A2 "define.*SECTION_ASM_OP" gcc/ | grep "\".*\."
> >
> > > ... snip ...
> > > Secondly, for seamless integration with the "used" attribute, we must be
> > > able to to mark the symbol with the used attribute applied as "retained"
> > > without changing its section name. For GCC "named" sections, this is
> > > straightforward, but for "unnamed" sections it is a giant mess.
> > >
> > > The section name for a GCC "unnamed" section is not readily available,
> > > instead a string which contains the full assembly code to switch to one
> > > of these text/data/bss/rodata/comm etc. sections is encoded in the
> > > structure.
> > >
> > > Backends define the assembly code to switch to these sections (some
> > > "*ASM_OP*" macro) in a variety of ways. For example, the unnamed section
> > > "comm_section", might correspond to a .bss section, or emit a .comm
> > > directive. I even looked at trying to parse them to extract what the
> > > name of a section will be, but it would be very messy and not robust.
> > >
> > > Meanwhile, having a .retain <symbol_name> directive is a very simmple
> > > solution, and keeps the GCC implementation really concise (patch
> > > attached). The assembler will know for sure what the section containing
> > > the symbol will be, and can apply the SHF_GNU_RETAIN flag directly.
> > >
> 
> Please take a look at
> 
> https://gitlab.com/x86-gcc/gcc/-/commits/users/hjl/elf/shf_retain
> 
> which is built in top of
> 
> https://gcc.gnu.org/pipermail/gcc-patches/2020-February/539963.html
> 
> I think SECTION2_RETAIN matches SHF_GNU_RETAIN well.  If you
> want, you extract my flags2 change and use it for SHF_GNU_RETAIN.

In your patch you have to make the assumption that data_section, always
corresponds to a section named .data. For just this example, c6x (which
supports the GNU ELF OSABI) does not fit the rule:

> c6x/elf-common.h:#define DATA_SECTION_ASM_OP "\t.section\t\".fardata\",\"aw\""

data_section for c6x corresponds to .fardata, not .data. So the use of
"used" on a data declaration would place it in a different section, that
if the "used" attribute was not applied.

For c6x and mips, readonly_data_section does not correspond to .rodata,
so that assumption cannot be made either:
> c6x/elf-common.h:#define READONLY_DATA_SECTION_ASM_OP "\t.section\t\".const\",\"a\",@progbits"
> mips/mips.h:#define READONLY_DATA_SECTION_ASM_OP        "\t.rdata"      /* read-only data */

The same can be said for bss_section for c6x as well.

Furthermore, this is only considering the examples in
default_elf_select_section - the less standard unnamed section are used
in many backend's implementation of select_section, and we would need to
work out what section name they correspond to to properly support
SHF_GNU_RETAIN.

For every unnamed section, you either have to assume what the
corresponding section name is, or parse the associated assembly output
string for the section.

Given these edge cases which must be handled for GCC to robustly emit
the "R" flag for sections containing "used" symbols, surely it is
preferable to leverage the existing TARGET_ASM_MARK_DECL_PRESERVED and
emit a .retain <symname> directive, which is extremely simple and
doesn't require any handling of these edge cases and non-standard
backend implementations.

The point about multiple section directives, some with the "R" flag some
without, still applies as a downside to trying to emit the .section
directives for the "used" attribute.

Thanks,
Jozef

> 
> -- 
> H.J.