[llvm] [LLVM][AARCH64]Replace +sme2p1+smef16f16 by +smef16f16 (PR #88860)

[Momchil Velikov via llvm-commits]

momchil-velikov wrote:

> Apparently `FEAT_INIT` indicates that the extension is a non-FMV extension, and the string list of dependencies that
> follows it is a list of dependencies for FMV. Therefore, historically, extensions with `FEAT_INIT` would always have an
> empty list of FMV dependencies.

Maybe. But I'd rather think the issue here is that we have `FEAT_INIT` instead of `FEAT_SME_F16`, analogous to the following two lines (cannot think of a logical reason to treat it differently). Which can be thought of as a temporary state, FMV is very much still in development.
 
> The effect of this is allowing a `-target-feature` to be specified without specifying any of its dependencies on the `-cc1`
> command line.

I cannot see such an effect here: https://gcc.godbolt.org/z/fPc1d5fqa

>  Every necessary `SubtargetFeature` should be specified explicitly with `-target-feature` and not rely on
> any dependency resolution at this stage.

If that's what we want to achieve, the dependency resolution must not **exist** at this stage. There's no point of in saying "ok, this functionality exists, but you must not use it".

> This means that tests like `clang/test/CodeGen/aarch64-sve2p1-intrinsics/acle_sve2p1_bfclamp.c` are wrong, because
> they should specify all the `-target-feature` flags that they need. 

Perhaps you meant some other test, this one is for S**V**E2.1.

https://github.com/llvm/llvm-project/pull/88860