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

[David Green via llvm-commits]

https://github.com/davemgreen commented:

> @davemgreen I added the cost model for the reduce.add(ext(mul(ext(A), ext(B))) pattern. There are other dot product patterns. I would be happy to work on them if only you told me what is your preferred solution to https://github.com/llvm/llvm-project/pull/121567#issuecomment-2650197122 Can you please let me know? Or maybe you can address this later yourself?

Thanks. I think this might just always be a bit of a bad thing for Arm/AArch64 (for add's at least, the others are fine), and is more likely to make things worse than better.  The backend already handles it in the cases we need it, and controlling it with a cost model is a little unreliable. But I ran some tests on this version and didn't see any problems in the tests I tried. The code you have looks decent and all the cases that were a problem before no longer are.

(I'm not sure of the most ideal way to structure it, if I was writing the whole compiler myself I would be tempted to add a smuladd/umuladd reduction intrinsic so that the cost modelling can be more accurate, but that sounds like quite a lot of work and might not even perfectly solve the problem in every case without effectively doing type/operation legalization earlier. Undoing arbitrarily reassociated/combined add trees can be difficult but is maybe better because it might capture more cases. "We can add that in the future if we need it" is probably the best way forward).

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