Re: [RFC] Function Multi Versioning on Arm

["Martin Liška" <mliska@suse.cz>]

On 7/18/22 12:36, Daniel Kiss via Gcc wrote:
> Hello,
> 
> We are going to add Function Multiversioning [1] support to Arm architectures.
> The specification is made public as beta[2] to ensure toolchain that follows Arm
> C Language Extension will implement it in the same way.
> 
> A few tweaks considered to make the developers' life easier.
> Since the `target` attribute is used widely on Arm, we would like to introduce a
> new attribute `target_version` to avoid confusion and possible deployment
> problems. The `target_clones` attribute will be supported too. Also the “default”
> version to be made optional.
> 
> We are looking for feedback on the specification (reply, github works too).
> 
> Thanks so much,
> Daniel
> 
> [1] https://gcc.gnu.org/onlinedocs/gcc/Function-Multiversioning.html 
> [2] https://github.com/ARM-software/acle/blob/main/main/acle.md#function-multi-versioning
> 

Hello.

Thanks for working on the feature, it will be nice to cover the gap in between x86_64 and powerpc,
which implement the FMV feature.

As the person who's been involved with the current MVC code in the GCC, I have a few comments/questions
about it:

1) both i386 and Powerpc also allow specifying an equivalent to -march (like `arch=bdver2`),
   in Arm case it seems to me only individual features are considered

2) about 'target_version' attribute - I like the idea as one can have a completely independent
   function implementation optimized for an ISA;
   note it's very close to 'target' attribute (supported in C++), where one needs to provide
   a resolver and have the pretty same functionality (see e.g. gcc/testsuite/g++.target/i386/mv1.C).
   However, the feature does not work in C and you will have the very same problem with target_version
   attribute (multiple functions with the same declaration):
  
  mv1.c:76:1: error: redefinition of ‘foo’
     76 | foo ()
        | ^~~

3) If you will implement 'target_version' attribute, I would like to see it available also for the
   existing targets supporting MVC

4) A small note about the mangling, the existing i386 naming scheme looks like:

...
_Z3foov.avx2_ssse3
...

5) Can you please define how will you evaluate priorities for a situation where multiple features
   are used (e.g. dotprod+sm)?

   Note we face the very same problem on i386, where it's very hard to specify a priority
   for the family of avx512 features. That said, an optional priority specifier might be possible.

6) Note that as opposed to i385 and Powerpc, we don't allow a combination of ISA flags for target_clone
   attribute (like sse2+avx512f).

7) FMV may be disabled in compile time by a compiler flag. In this case the default version shall be used.

   I would like to see the functionality also target agnostic.

Anyway, looking forward to the Arm implementation.
Hope the comments are constructive.

Cheers,
Martin