[llvm] [VectorCombine] Fold binary op of reductions. (PR #121567)

[David Green via llvm-commits]

https://github.com/davemgreen commented:

Hi - Thanks for adding the cost model. It looks descent, and seems to work quite well for MVE. Unfortunately there are cases where it isn't working as well at least for one case for udot on aarch64, as we don't always generate the reduction pattern until during instruction selection.

I was going to suggest I added a phase ordering test, but it looks like it needs LTO to go wrong as it only sees the vectors after loopvect+unroll+vector-combine, which do not usually happen in that order. I'll see if I can add one, but we might need to end up undoing this in the backend (or add a bailout from the target, or just leave it to the backend to do).

https://github.com/llvm/llvm-project/pull/121567