[llvm] [InstCombine] Scalarize `(vec_ops (insert ?, X, Idx))` when only one element is demanded (PR #84645)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> Are you saying that SLP is creating these vectorizations but shouldn't? Isnt that a cost issue with the alt opcode cases?

Does not look like cost issue, looks like a corner case that should be handled separately in SLP. We have some kind of it already, but not all. There is something similar with sext/zext/trunc, I'm going to fix later, since it requires some extra stuff in minbitwidth analysis. This one should not require extra work, just need to automatically create gather node for 2-element altshuffle, if we see that the operand is going to be gather/buildvector node. Can do it tomorrow.

https://github.com/llvm/llvm-project/pull/84645