[clang] [llvm] [clang-tools-extra] [VPlan] Add new VPUniformPerUFRecipe, use for step truncation. (PR #78113)

[via cfe-commits]

================
@@ -1469,6 +1465,47 @@ void VPReplicateRecipe::print(raw_ostream &O, const Twine &Indent,
 }
 #endif
 
+Value *VPScalarCastRecipe ::generate(VPTransformState &State, unsigned Part) {
+  assert(vputils::onlyFirstLaneUsed(this) &&
+         "Codegen only implemented for first lane only.");
+  switch (Opcode) {
+  case Instruction::SExt:
+  case Instruction::ZExt:
+  case Instruction::Trunc: {
+    Value *Op = State.get(getOperand(0), VPIteration(Part, 0));
+    return State.Builder.CreateCast(Instruction::CastOps(Opcode), Op, ResultTy);
+  }
+  default:
+    llvm_unreachable("opcode not implemented yet");
+  }
+}
+
+void VPScalarCastRecipe ::execute(VPTransformState &State) {
+  bool UniformAcrossUFs = all_of(operands(), [](VPValue *Op) {
+    return Op->isDefinedOutsideVectorRegions();
+  });
+  for (unsigned Part = 0; Part != State.UF; ++Part) {
+    Value *Res;
+    // Only generate a single instance, if the recipe is uniform across all UFs.
----------------
ayalz wrote:

This is currently untested / dead code.

The value returned by an EVL recipe should arguably be of the same type as that of its first operand, even if the underlying intrinsic it currently wraps insists on always returning an i32? I.e., that case should arguably be fixed by performing a trunc/sext/zext as needed when providing the adequate EVL value given a current Phi/remaining-trip-count values, to facilitate adding/subtracting them together.

https://github.com/llvm/llvm-project/pull/78113