[PATCH] D98512: [LoopVectorize] Simplify scalar cost calculation in getInstructionCost

[David Sherwood via Phabricator via llvm-commits]

david-arm added a comment.

Hi @dmgreen, thanks for taking a look at the patch. So you raise a good point - the loads/stores will be dealt with before reaching this variant getInstructionCost, however the GEPs/Bitcasts for those will still be costed here. However, for GEPs at least you can see this code in getInstructionCost:

  case Instruction::GetElementPtr:
    // We mark this instruction as zero-cost because the cost of GEPs in
    // vectorized code depends on whether the corresponding memory instruction
    // is scalarized or not. Therefore, we handle GEPs with the memory
    // instruction cost.
    return 0;

which only leaves us with Bitcasts with pointer type results. Normal ptr-ptr bitcasts will be 0 cost I think, so I imagine what's left are bitcasts of non-ptr types to ptr types. It's very unfortunate that the `Scalars` list includes so many different types of Scalars that need costing differently. What I could do here is have a single special case for Bitcasts where the cost (likely 0) is multiplied by VF, with an assertion that we're not dealing with scalable vectors.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D98512/new/

https://reviews.llvm.org/D98512