https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98121
jozefl at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jozefl at gcc dot gnu.org
--- Comment #1 from jozefl at gcc dot gnu.org ---
The feedback after I originally submitted a patch that implements a "retain"
attribute to apply SHF_GNU_RETIAN, was that "used" should instead apply
SHF_GNU_RETAIN:
https://gcc.gnu.org/pipermail/gcc-patches/2020-October/557097.html
The discussions went on about other aspects of the implementation in other
threads, but there were no objections to the idea of "used" applying
SHF_GNU_RETAIN.
I think the logic is sound that a declaration marked with the "used" attribute
should not be removed by the compiler, or garbage collected by the linker.
Basically, if you mark a declaration as "used", it should be retained in the
final program, even if it was not referenced.
The issue lies in the fact that the "used" attribute has been around for so
long, that the change in behavior has the potential to affect existing
applications that make use of it.
However, the new behavior is only enabled when using the latest GCC and latest
Binutils. Some change in behavior can be expected when using the latest
versions of the tools. This new behavior should really be documented in the
release notes though.
I didn't want to have "used" change section attributes in assembler directives,
the original plan was to use a separate assembler ".retain" directive to apply
SHF_GNU_RETAIN. This would mean we could leave .section assembler directives
for a particular "used" declaration untouched, and just emit a ".retain
<symbol_name>" directive. The assembler therefore handles the application of
the SHF_GNU_RETAIN flag to a particular section, instead of forcing the
compiler to have to make these additional considerations about section
attributes.
I even did a patch for that, but there was staunch objection from H.J. on this,
on multiple occasions, I realized we weren't going to get anywhere:
https://gcc.gnu.org/pipermail/gcc-patches/2020-November/557931.html
https://gcc.gnu.org/pipermail/gcc-patches/2020-November/558400.html
https://sourceware.org/pipermail/binutils/2020-November/114077.html
I would be more than happy to add the .retain attribute to Binutils, but we
would need a Binutils global maintainer to approve that, and no one gave the OK
for my original ".retain <symbol_name>" patch the first time round. I could
then change the GCC implementation to use the ".retain" directive, which would
ensure that section attributes in assembler code output by GCC are not
contaminated with the "R" flag when the "used" attribute is applied.
Thanks,
Jozef
