[llvm] [InstCombine] Move extends across identity shuffles. (PR #146901)

[Florian Hahn via llvm-commits]

fhahn wrote:

> No objections, but I am concerned that we're going to eventually start finding cases where InstCombine will undo the cost driven decisions that VectorCombine has made.

I suppose that's a general concern for vector transforms in InstCombine? For this particular transform, I don't think we would have this problem at the moment, as I doubt that doing the extends on a wider type (with poison lanes) would be more profitable.

https://github.com/llvm/llvm-project/pull/146901