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

[Alexandros Lamprineas via llvm-commits]

labrinea wrote:

> I've removed the constant folding since it was indeed unnecessary computation. Also I found a testcase which exposes a potential issue: a function version (for example AES) having the same target features as the default. In TargetParser we have `ExtensionInfo->DependentFeatures` to model this information (I believe @ilinpv introduced this change) and it seems that for some entries we don't add the corresponding backend feature (perhaps intentionally?). I will try to address this in a separate ticket.

@andrewcarlotti already has a patch for ACLE https://github.com/ARM-software/acle/pull/315 which removes `pmull` since it is implied by `aes`, therefore we can add the missing `+aes` in ExtensionInfo->DependentFeatures for `aes`.

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