[llvm] [AArch64][MachineCombiner] Reassociate long chains of accumulation instructions into a tree to increase ILP (PR #126060)

[Jonathan Cohen via llvm-commits]

jcohen-apple wrote:

MLA is indeed more common, but less likely to get linearized by the compiler. See this godbolt example where the compiler linearizes multiple accumulators due to reassociate pass running after inlining a function - https://godbolt.org/z/1aPbb3YY1. This same bug should appear for other intrinsics whose ISEL pattern also contains `add` instructions. If you find it confusing that the first commit handles UABAL and not something more obvious, I can start with MLA and then add UABAL.
Adding the additional instruction does indeed make it slightly more tricky than the other patterns handled by MachineCombiner. I thought that the checks on profitability of the pattern (that it reduces critical path length and does not increase resource length) still apply here. 
I'll move the main logic into MachineCombiner pass and add the tests you suggested, and MLA as well.

https://github.com/llvm/llvm-project/pull/126060