[llvm] [FMV][GlobalOpt] Bypass the IFunc Resolver of MultiVersioned functions. (PR #87939)

[Alexandros Lamprineas via llvm-commits]

labrinea wrote:

> The testcase and description look valid, although the description seems a little more conservative than necessary.

"Subset" is perhaps not the right terminology. I meant to say the feature set is "less or equal to". In other words "implied".
 
> If you were to add a `bar._Mmops` version, would it's call be optimised to `test_multi_bb_resolver._Mmops`? 

Yes

> And would that also enable the call from `bar.default` to be optimised to point to `test_multi_bb_resolver.default`?

No, because there is no corresponding caller to pick `test_multi_bb_resolver._Msve2`. The feature set of `test_multi_bb_resolver._Msve2` is not implied by the feature set of `bar._sve`. 



https://github.com/llvm/llvm-project/pull/87939