[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:

Updated the PR with the following changes:
1. Moved the code from AArch64InstrInfo to TargetInstrInfo
2. Added some LLVM IR test cases
3. Support uabal2 and sabal as well

I still need to:
1. Add tests for additional types
2. Support MLA

I don't think this case should kick in for loops that are not fully unrolled. If I'm not mistaken, when LLVM unrolls the loops it will create multiple accumulators, but I need to check that.

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