Target feature implications for negative features are handled inconsistently between codegen and cfg(target_feature)

[RalfJung]

The logic that computes cfg(target_feature) takes into account target feature implications when handling something like -sse: it will also remove avx from the list of enabled target features in that case. However, the logic that computes which flags we set for codegen (which, unfortunately, is completely separate), does not do the same: it will add -sse to LLVM's target feature list, but does not do anything about Rust-level target feature implications.

This can't be correct -- either negative target features also imply that their "reverse dependencies" get disabled, or they do not. We shouldn't do one thing in codegen and a different thing for cfg. Or is there some good reason for this?

The logic for this was added in #128221.

Cc @calebzulawski @Amanieu @workingjubilee

[workingjubilee]

If I understand this issue description correctly (95% sure, will have to review the actual code to be ~99.9%) then no, there is no particularly good reason for this difference and we should do the exact same thing.