[llvm] [RISCV] Run another VLOptimizer before machine SSA optimizations (PR #179377)

[Craig Topper via llvm-commits]

topperc wrote:

> Can you add the case you saw? From the diffs, I don't see it beneficial to performance. Or we can collect some statistics on llvm-testsuite.

I spoke to Min earlier. When the FMA formmation is moved to machineCombiner in https://github.com/llvm/llvm-project/pull/177238, we miss cases where the VectorPeephole pass turned an vfadd into a tail undisturbed operation. The vfadd will have a non-VLMAX VL, but the vfmul will have VLMAX until VLOptimizer uns.

https://github.com/llvm/llvm-project/pull/179377