[llvm] [LV] Peek through bitcasts when performing CSE (PR #146856)

[Pedro Lobo via llvm-commits]

pedroclobo wrote:

I'm not really sure about the consequences of the legacy CSE to VPlan-based CSE refactoring. 
I think @artagnon is more knowledgeable on that matter.

The byte type only introduced an additional `bitcast` from an `i8` to `b8`, which blocked the `bitcast` + `shufflevector` CSE. Before introducing the byte type, the store was over an integer type, so there was no need for a `bitcast`. As such, the CSE was able to eliminate the redundant `shufflevector`.

This pattern was found when diagnosing an assembly regression in 7zip.

https://github.com/llvm/llvm-project/pull/146856