[llvm] [WebAssembly] Support partial-reduce accumulator (PR #158060)

[Sam Parker via llvm-commits]

sparker-arm wrote:

The TTI change only affects whether the partial.reduce intrinsic is used. MAX-BANDWIDTH is when we're forcing the vectorizer to vectorize based on the memory type, instead of the widest type and, AFAICT, these partial reductions will not happen unless we're optimising for bandwidth.

The test uses a vectorizer option, but it is also available as a TTI hook, which I'm exploring whether we could/should have it enabled.

https://github.com/llvm/llvm-project/pull/158060