[llvm] [VPlan] Fix issues around scalar-cast VPInstructions (PR #195629)

[Ramkumar Ramachandra via llvm-commits]

https://github.com/artagnon updated https://github.com/llvm/llvm-project/pull/195629

>From e32f9ac243057f3f293503e3cf252bcf090fcc3e Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <artagnon at tenstorrent.com>
Date: Mon, 4 May 2026 11:03:20 +0100
Subject: [PATCH] [VPlan] Fix issues around scalar-cast VPInstructions

Compute a cost for them, mark them alwaysUnmasked, and apply flags when
executing them. While at it, strip VPRecipeBase::isScalarCast.
---
 .../Transforms/Vectorize/LoopVectorize.cpp    |  3 ++-
 llvm/lib/Transforms/Vectorize/VPlan.h         | 15 ++++----------
 .../lib/Transforms/Vectorize/VPlanRecipes.cpp | 20 ++++++++++++-------
 3 files changed, 19 insertions(+), 19 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 78163b5fe35d5..afbdb88d7640e 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7845,7 +7845,8 @@ static SmallVector<Instruction *> preparePlanForEpilogueVectorLoop(
                 [](const VPUser *U) {
                   return isa<VPScalarIVStepsRecipe>(U) ||
                          isa<VPDerivedIVRecipe>(U) ||
-                         cast<VPRecipeBase>(U)->isScalarCast() ||
+                         Instruction::isCast(
+                             cast<VPInstruction>(U)->getOpcode()) ||
                          cast<VPInstruction>(U)->getOpcode() ==
                              Instruction::Add;
                 }) &&
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 4a5420185224b..8f04357da099c 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -552,9 +552,6 @@ class LLVM_ABI_FOR_TEST VPRecipeBase
   /// Returns the debug location of the recipe.
   DebugLoc getDebugLoc() const { return DL; }
 
-  /// Return true if the recipe is a scalar cast.
-  bool isScalarCast() const;
-
   /// Set the recipe's debug location to \p NewDL.
   void setDebugLoc(DebugLoc NewDL) { DL = NewDL; }
 
@@ -1376,7 +1373,8 @@ class LLVM_ABI_FOR_TEST VPInstruction : public VPRecipeWithIRFlags,
     if (!getUnderlyingValue())
       return true;
 
-    return Opcode == Instruction::PHI || Opcode == Instruction::GetElementPtr;
+    return Instruction::isCast(Opcode) || Opcode == Instruction::PHI ||
+           Opcode == Instruction::GetElementPtr;
   }
 
 public:
@@ -1524,8 +1522,6 @@ class VPInstructionWithType : public VPInstruction {
   static inline bool classof(const VPRecipeBase *R) {
     // VPInstructionWithType are VPInstructions with specific opcodes requiring
     // type information.
-    if (R->isScalarCast())
-      return true;
     auto *VPI = dyn_cast<VPInstruction>(R);
     if (!VPI)
       return false;
@@ -1536,7 +1532,7 @@ class VPInstructionWithType : public VPInstruction {
     case Instruction::Load:
       return true;
     default:
-      return false;
+      return Instruction::isCast(VPI->getOpcode());
     }
   }
 
@@ -1556,10 +1552,7 @@ class VPInstructionWithType : public VPInstruction {
 
   /// Return the cost of this VPInstruction.
   InstructionCost computeCost(ElementCount VF,
-                              VPCostContext &Ctx) const override {
-    // TODO: Compute accurate cost after retiring the legacy cost model.
-    return 0;
-  }
+                              VPCostContext &Ctx) const override;
 
   Type *getResultType() const { return ResultTy; }
 
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 2225dfa310c6c..41102932fd930 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -312,11 +312,6 @@ bool VPRecipeBase::isPhi() const {
          isa<VPPhi, VPIRPhi>(this);
 }
 
-bool VPRecipeBase::isScalarCast() const {
-  auto *VPI = dyn_cast<VPInstruction>(this);
-  return VPI && Instruction::isCast(VPI->getOpcode());
-}
-
 void VPIRFlags::intersectFlags(const VPIRFlags &Other) {
   assert(OpType == Other.OpType && "OpType must match");
   switch (OpType) {
@@ -1306,7 +1301,7 @@ bool VPInstruction::isSingleScalar() const {
   case VPInstruction::VScale:
     return true;
   default:
-    return isScalarCast();
+    return Instruction::isCast(getOpcode());
   }
 }
 
@@ -1589,10 +1584,14 @@ void VPInstruction::printRecipe(raw_ostream &O, const Twine &Indent,
 
 void VPInstructionWithType::execute(VPTransformState &State) {
   State.setDebugLocFrom(getDebugLoc());
-  if (isScalarCast()) {
+  if (Instruction::isCast(getOpcode())) {
     Value *Op = State.get(getOperand(0), VPLane(0));
     Value *Cast = State.Builder.CreateCast(Instruction::CastOps(getOpcode()),
                                            Op, ResultTy);
+    if (auto *CastOp = dyn_cast<Instruction>(Cast)) {
+      applyFlags(*CastOp);
+      applyMetadata(*CastOp);
+    }
     State.set(this, Cast, VPLane(0));
     return;
   }
@@ -1614,6 +1613,13 @@ void VPInstructionWithType::execute(VPTransformState &State) {
   }
 }
 
+InstructionCost VPInstructionWithType::computeCost(ElementCount VF,
+                                                   VPCostContext &Ctx) const {
+  if (!getUnderlyingValue())
+    return 0;
+  return getCostForRecipeWithOpcode(getOpcode(), VF, Ctx);
+}
+
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
 void VPInstructionWithType::printRecipe(raw_ostream &O, const Twine &Indent,
                                         VPSlotTracker &SlotTracker) const {