[llvm] [LV] Reland "Add costs for VPInstructionWithType::computeCost" (PR #202952)

[Florian Hahn via llvm-commits]

================
@@ -1790,6 +1790,38 @@ void VPInstruction::printRecipe(raw_ostream &O, const Twine &Indent,
 }
 #endif
 
+InstructionCost VPInstructionWithType::computeCost(ElementCount VF,
+                                                   VPCostContext &Ctx) const {
+  // NOTE: At the moment it seems only possible to expose this path for
+  // the trunc, zext and sext opcodes. However, isScalarCast also covers
+  // int<>fp conversions, bitcasts, ptr<>int conversions, etc.
+  if (Instruction::isCast(getOpcode()))
+    return getCostForRecipeWithOpcode(getOpcode(), ElementCount::getFixed(1),
+                                      Ctx);
+
+  switch (getOpcode()) {
+  case VPInstruction::VScale: {
+    Type *Ty = this->getScalarType();
+    ArrayRef<Type *> Tys;
+    IntrinsicCostAttributes Attrs(Intrinsic::vscale, Ty, Tys);
+    return Ctx.TTI.getIntrinsicInstrCost(Attrs, Ctx.CostKind);
+  }
+  case VPInstruction::StepVector:
+    // TODO: This isn't quite right since even if the step-vector is hoisted
+    // out of the loop it has a non-zero cost in the middle block, etc.
+    // Once the stepvector is correctly hoisted out of the vector loop by the
+    // licm transform we can add the cost here so that it doesn't incorrectly
+    // affect the choice of VF.
+    return 0;
+  default:
+    // Although VPInstructionWithType is also used for
+    // VPInstruction::WideIVStep it isn't currently possible to expose cases
+    // where the cost is queried.
+    break;
+  }
+  return 0;
----------------
fhahn wrote:

I think it would still be good to keep the urneachable; this would help us make sure we do not miss other opcodes by accident in the future.

https://github.com/llvm/llvm-project/pull/202952