[llvm] [TTI][WebAssembly] Pairwise reduction expansion (PR #93948)

[Sam Parker via llvm-commits]

sparker-arm wrote:

I've changed the scope so that we're now only affecting FADD. This requires that code generators need to implement more pattern matchers, but it also means that, if they don't, the shuffles produced for i8x16 and i16x8 aren't awful.

For the FP workloads that I've been looking at, this change can be worth 3-7% performance uplift.

https://github.com/llvm/llvm-project/pull/93948