[llvm] [SLP]Fix fmul/fadd fusion costs and retry FMA seeds after all blocks (PR #226117)

[Tomas Matheson via llvm-commits]

tommat01 wrote:

@alexey-bataev we are seeing a cumulative ~5% performance regression in AArch64 SPEC LBM between this and https://github.com/llvm/llvm-project/pull/227083. I'll try to get reproducers for one or both of them.

There seem to be regular SLP patches which cause AArch64 regressions of this size in SPEC and other important benchmarks. You have consistently responded quickly and fixed issues once we have provided reproducers, which I appreciate. However, the current workflow of regression → reproducer → fix → further regression → further reproducer is not sustainable. Producing these reproducers takes substantial engineering effort, and we are falling behind the rate at which regressions are being introduced.

Would it be possible for you to add some aarch64 performance regression checking before merging changes, specifically SPEC17 and/or SPEC26 would be very helpful as an initial set?

In the meantime I think it would make sense if we could agree to revert any problematic patches and work together on fixes before landing them again, as per the [developer policy](https://llvm.org/docs/DeveloperPolicy.html#patch-reversion-policy).

I am open to other suggestions also.

https://github.com/llvm/llvm-project/pull/226117