[llvm] [X86][WinEH] Avoid stale RSP after Win64 dynalloca invokes (PR #191616)

[via llvm-commits]

taigacon wrote:

These cases were actually already covered in win64_alloca_dynalloca.ll, but I agree they were not very obvious.

I’ve now added more explicit coverage for them: one case with two dynamic allocas, and one case that makes the Win64 stack-argument shape and an over-aligned (>16) dynamic alloca more obvious in the test code.

I’m still a bit unsure about how tight the checks should be in this area, so if you have a chance, I’d really appreciate another look at whether the current test constraints seem appropriate.

https://github.com/llvm/llvm-project/pull/191616