[llvm] [SelectionDAG] Avoid illegal wide vectors expanding get.active.lane.mask (PR #213670)

[Philip Reames via llvm-commits]

preames wrote:

> The case was found by @preames, and it can't reproduced/vectorized by upstream any more, but I think the point is it's valid IR but crash.

We had a bit of a game of telephone here.  This was a crazy observed on a previous version (19.x if I remember correctly), and a change to vectorizer costing fixed it in the meantime.  The EVL work also side steps the issue, so we have two independent fixes for the original case.  So the vectorizer _isn't_ emitting this today.

However, the reduced IR crash (after original vectorization strategy) is arguably valid.  That's the case Kito is working on here.  I say "arguably" because I'm not sure on whether it is actually valid or not.  What exactly our contract is around illegally typed vector intrinsics is a bit opaque to me.  

https://github.com/llvm/llvm-project/pull/213670