[llvm] [CodeGen] Expansion of scalable vector reductions (PR #129214)

[David Green via llvm-commits]

davemgreen wrote:

> So that something comparable to loop-fusion could happen during ISel for sequences of multiple such instructions? That sounds indeed interesting, is there already something similar implemented I could have a look at? I still think this reduction expansion would be useful now.

I think it would need to generate a single loop with multiple instructions inside it. Doing so inside isel (if it isn't global-isel) is more difficult. The problem here is "scalable vectors", not "reductions" and we shouldn't be special-casing reduction operations.

> My thought was that this concretely would still be beneficial for out-of-loop product reductions, or cases where code is compiled with -msve-vector-bits=.... If later the TTI is changed, the cost for reductions requiring such expansions should still be high so that NEON or so is preferred.

Why not just generate a fixed-width vector if we know the length? We should also be able to expand in ISel if we know it requires a certain number of steps.

I think we should be moving towards removing the ExpandReductionsPass, not expanding it. And like I said the problem here seems to be scalable-vectors with no backend support, not specific to reductions. The solution should be more general than just the ExpandReduction pass can handle.

https://github.com/llvm/llvm-project/pull/129214