[llvm] [ARM] Fix CMSE S->NS calls when CONTROL_S.SFPA==0 (CVE-2024-7883) (PR #114433)

[Eli Friedman via llvm-commits]

https://github.com/efriedma-quic approved this pull request.

The logic here seems like it's very dependent on the assumption that nothing will try to optimize the expanded instructions after pseudo-expansion.  Which is... probably okay, I guess.  Probably we'd notice if that sort of optimization started happening.

It might be worth adding a testcase where an function argument is undef/poison, to make sure we don't optimize out the "no-op" there?

Otherwise LGTM.

https://github.com/llvm/llvm-project/pull/114433