[PATCH] D78651: [TTI] Devirtualize getInstructionLatency

[Sam Parker via Phabricator via llvm-commits]

samparker added a comment.

Thanks both.

> Is it a good idea to have getInstructionThroughput/getInstructionLatency/getUserCost being implemented through different mechanisms?

I don't think so. I like the idea of using the getInstructionCost interface which then calls into the respective functions, which then calls into the concrete implementation. This means the getUserCost would disappear though, and various cost methods would take the explicit TargetCostKind argument. I think AMDGPU is the only backend with a significant getUserCost implementation.

> Is there an llvm-dev thread that describes what we want the cost model to end up looking like?

I will try to get an RFC patch for what I described above and then post to the list.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D78651/new/

https://reviews.llvm.org/D78651