[llvm] [RISCV] Insert a freeze before converting select to AND/OR. (PR #84232)

[Philip Reames via llvm-commits]

preames wrote:

> This looks fine to me, as long as RISCV folks are fine with having these regressions in the meantime. The alternative would be to land mitigations first.

LGTM specifically on the taking the regressions for correctness point.  I'm assuming Nikita's correct on the poison propagation rules.  I glanced at the code, didn't see any obvious errors, but this stuff makes my head hurt.  

https://github.com/llvm/llvm-project/pull/84232