[clang] [FMV][AArch64] Don't optimize backward compatible features in resolver. (PR #90928)

[Eli Friedman via cfe-commits]

efriedma-quic wrote:

Burying this check in the middle of CodeGen seems error-prone: when we add new features in the future, someone is going to have to dig up this code.  Can we put this in on of the target feature tables, like in AArch64TargetParser.h?

If I'm understanding correctly, the issue here isn't really that the feature is "backwards-compatible"; the issue is that it's a feature the OS frequently disables, so even if you have compatible hardware, it might not actually be enabled. Other features have similar characteristics (like SVE), but we generally assume the OS will enable SVE on CPUs where it's available.  And the user can explicitly override this assumption if they want to.

It might make sense to allow the user to explicitly specify that we can assume the OS support is enabled.

https://github.com/llvm/llvm-project/pull/90928