[llvm] [ISel] Commute FMUL and inserting zero into vector lane (PR #146096)

[Julian Nagele via llvm-commits]

juliannagele wrote:

Thanks! I've added more tests, including one that is better when doing the the insert early (distilled down from where we found the pattern). I've also added some logic to possibly insert into the other operand and deal with both operands being the same. We could probably add better heuristics for which operand to choose, but if you don't mind I'd leave that as a follow up for now (since it should never be worse as is).

https://github.com/llvm/llvm-project/pull/146096