[llvm] [AMDGPU] Cache results in a bitvector for VOPDPairingMutation (PR #201837)

[Igor Wodiany via llvm-commits]

IgWod wrote:

> Do you get measurable compile time saving from this? I assume that is the motivation, right?

Yes, Antonio was primarily working on this particular idea and I think the improvement was something like 4-5% in the initial test. I think it'll be less now since my patches improving VOPD eligibility checking went in, but I would expect it to be a similar ballpark as the other caching PR (#199714).

https://github.com/llvm/llvm-project/pull/201837