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

[Alexandros Lamprineas via llvm-commits]

labrinea wrote:

Hi, @jroelofs. This patch has been quiet in a while and the reason is I have been collaborating with @tmatheson-arm in refactoring the TargetParser as a preliminary step. However the landscape doesn't seem to be changing much, it's mostly NFC. What I mean is that for this patch to work we need to be able to express every FMV feature in the LLVM IR attribute "target-features". This isn't the case for a few FMV features and I believe it won't be easy to change. For example the FMV feature AES in ExtensionInfo has no DependentFeatures because the SubtargetFeature for AES is fused with PMULL, meaning we can't link them (the fmv feature and the subtarget counterpart) together as they mean different things. I do not want to rely on the cleanup in https://github.com/ARM-software/acle/pull/315 as features may diverge again between FMV and the AAch64 compiler backend. Spliting up backend features will break backwards compatibility in LLVM IR (because they will also have to be renamed), so maybe we need some other way to propagate FMV information from clang to LLVM (not via the target-features attribute). Is this worth an RFC, like introducing a new attribute for FMV in IR? Is this optimization still valuable to you? Shall we pursue this further or just give up?

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