[llvm] [X86][ISel] Fix logic for optimizing `movmsk(bitcast(shuffle(x)))` (PR #68369)

[via llvm-commits]

goldsteinn wrote:

> Have you any cases where all of this is necessary vs my smaller patch? This code was to handle a uncommon edge case in the first place (a movmsk variant of reduce(permute(x)) -> reduce(x)), and we're adding a lot of code to handle it. Ideally the SimplifyDemanded* calls will have simplified the shuffle anyhow if we're only demanding the sign bits .

Okay, changed to your version + updated commit/PR message

https://github.com/llvm/llvm-project/pull/68369