Re: [RFC 0/2] Transparent multi-version symbol support

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 18/03/2021 17:06, Florian Weimer via Libc-alpha wrote:
> This turned out to be way harder than it should be.  Older binutils does
> not support multiple .symver directives for the same first symbol, so we
> have various kludges like (from time/clock_settime.c):
> 
> | versioned_symbol (libc, __clock_settime, clock_settime, GLIBC_2_17);
> | 
> | /* clock_settime moved to libc in version 2.17;
> |    old binaries may expect the symbol version it had in librt.  */
> | #if SHLIB_COMPAT (libc, GLIBC_2_2, GLIBC_2_17)
> | strong_alias (__clock_settime, __clock_settime_2);
> | compat_symbol (libc, __clock_settime_2, clock_settime, GLIBC_2_2);
> | #endif
> 
> I think I have found a way to do this with assembler hacks, so that it
> applies to function and data symbols alike.  With this, the above
> snipper becomes:
> 
> | versioned_symbol (libc, __clock_settime, clock_settime, GLIBC_2_17);
> | 
> | /* clock_settime moved to libc in version 2.17;
> |    old binaries may expect the symbol version it had in librt.  */
> | #if SHLIB_COMPAT (libc, GLIBC_2_2, GLIBC_2_17)
> | compat_symbol (libc, __clock_settime, clock_settime, GLIBC_2_2);
> | #endif
> 
> So the explicit aliases are gone.
> 
> Built with build-many-glibcs.py with binutils 2.36 and 2.34.  Tested on
> i686-linux-gnu and x86_64-linux-gnu with binutils 2.34 and 2.35.
> 
> The object files aren't unchanged due to the __malloc_initialize_hook
> change (and presumably others, I haven't looked closely).

Thanks for working on this, I have checked on the all architectures that
generated abilist and I saw not regression on a make check (with 
run-built-tests=no).

The only detail I am not sure is if SYMVER_NEEDS_ALIAS does yields any
substantial gain (and it require some refactoring and code to handle it).
What kind of improvement do you expect?