[llvm] [InstCombine] Fold patterns which uses <2N x iM> type for comparisons on <N x i2M> vector types (PR #184328)

[Yunbo Ni via llvm-commits]

================

----------------
cardigan1008 wrote:

> if the two are not equivalent, we should not do the transform right?

In the counterexample, lane 0: 

```
A: and 0, poison -> poison 
B: select false, poison, 0 -> 0
```

Not sure if I really get your question here, but I think it's not about equivalence but refinement. B is more defined/less poisonous than A. So A->B is accepted. 

C behaves more like the widened version of A: it speculates dependence on both half-lanes and can propagate poison that B would have masked through the unchosen select arm. 

https://github.com/llvm/llvm-project/pull/184328