[clang] [llvm] [AArch64][Driver] Fix behavior of +nofeat modifier on -mcpu (PR #203458)

[David Spickett via cfe-commits]

https://github.com/DavidSpickett commented:

I've been away from this area for too long to say whether the implementation is ok, that's for others to comment on.

I do see the use case. A feature might be "required" when designing the CPU, but with a compiler you're just selecting what you want to generate. Using a subset of features is fine.

Does this PR only apply to -mcpu, or -march too? If it applies to -march then that needs to be tested as well. If we are confident that both options end up using mostly the same code, a couple of -march is fine, no need to duplicate every test.

And if this doesn't apply to -march because -march works as expected, please:
* State that in the PR description.
* Check that there are tests showing that, and add them if they do not exist already.

https://github.com/llvm/llvm-project/pull/203458