[llvm] [LoopVectorize] Don't scalarize predicated instruction with optsize (PR #129265)

[Florian Hahn via llvm-commits]

https://github.com/fhahn commented:

This only deals with scalarizing memory instructions, but other instructions could also be scalarized, e.g. via `collectInstsToScalarize` which would be missed a the moment?

A more general alternative that would make it easier to catch all instance would be to check the generated VPlans for replicating recipes/regions. 

For `optsize`, we could also allow slightly increasing the size, in cases where scalarization only adds a few extra instructions in a large loop, so we could also compute the size of the vector loop and compare it against the size of the scalar loop.

https://github.com/llvm/llvm-project/pull/129265