About

Tracking VL optimizer improvement.

The VL optimizer is MachineIR pass that runs before vsetvli insertion. The pass decreases the VL of vector instructions by looking at the minimum VL of all of the instructions that use its result. This especially important for arithmetic instructions as the loop vectorize does not generate VL predicated arithmetic operations. Naively this produces VLMAX operations. By propagating VL from stores, reductions, etc. we can avoid toggling VL to VLMAX.

Most instructions have been supported already.

Recent important patches in 2H205

https://github.com/llvm/llvm-project/pull/156618

https://github.com/llvm/llvm-project/pull/156719

https://github.com/llvm/llvm-project/pull/155467

https://github.com/llvm/llvm-project/pull/151285