[VPlan] Add ComputeFindLastIVResult opcode (NFC). (llvm#132689)

fhahn · pawosm-arm · commit 3675c5ffc490 · 2025-04-10T18:30:47.000Z
This moves the logic for computing the FindLastIV reduction result to its own opcode. A follow-up patch will update the new opcode to also take the start value, to fix llvm#126836. PR: llvm#132689
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7514,7 +7514,8 @@ static void fixReductionScalarResumeWhenVectorizingEpilog(
     BasicBlock *BypassBlock) {
   auto *EpiRedResult = dyn_cast<VPInstruction>(R);
   if (!EpiRedResult ||
-      EpiRedResult->getOpcode() != VPInstruction::ComputeReductionResult)
+      (EpiRedResult->getOpcode() != VPInstruction::ComputeReductionResult &&
+       EpiRedResult->getOpcode() != VPInstruction::ComputeFindLastIVResult))
     return;
 
   auto *EpiRedHeaderPhi =
@@ -9643,8 +9644,10 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
           Builder.createSelect(Cond, OrigExitingVPV, PhiR, {}, "", FMFs);
       OrigExitingVPV->replaceUsesWithIf(NewExitingVPV, [](VPUser &U, unsigned) {
         return isa<VPInstruction>(&U) &&
-               cast<VPInstruction>(&U)->getOpcode() ==
-                   VPInstruction::ComputeReductionResult;
+               (cast<VPInstruction>(&U)->getOpcode() ==
+                    VPInstruction::ComputeReductionResult ||
+                cast<VPInstruction>(&U)->getOpcode() ==
+                    VPInstruction::ComputeFindLastIVResult);
       });
       if (CM.usePredicatedReductionSelect(
               PhiR->getRecurrenceDescriptor().getOpcode(), PhiTy))
@@ -9689,8 +9692,12 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
     // also modeled in VPlan.
     VPBuilder::InsertPointGuard Guard(Builder);
     Builder.setInsertPoint(MiddleVPBB, IP);
-    auto *FinalReductionResult = Builder.createNaryOp(
-        VPInstruction::ComputeReductionResult, {PhiR, NewExitingVPV}, ExitDL);
+    auto *FinalReductionResult =
+        Builder.createNaryOp(RecurrenceDescriptor::isFindLastIVRecurrenceKind(
+                                 RdxDesc.getRecurrenceKind())
+                                 ? VPInstruction::ComputeFindLastIVResult
+                                 : VPInstruction::ComputeReductionResult,
+                             {PhiR, NewExitingVPV}, ExitDL);
     // Update all users outside the vector region.
     OrigExitingVPV->replaceUsesWithIf(
         FinalReductionResult, [FinalReductionResult](VPUser &User, unsigned) {
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -1209,6 +1209,7 @@ class VPInstruction : public VPRecipeWithIRFlags,
     CanonicalIVIncrementForPart,
     BranchOnCount,
     BranchOnCond,
+    ComputeFindLastIVResult,
     ComputeReductionResult,
     // Takes the VPValue to extract from as first operand and the lane or part
     // to extract as second operand, counting from the end starting with 1 for
diff --git a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
@@ -64,6 +64,7 @@ Type *VPTypeAnalysis::inferScalarTypeForRecipe(const VPInstruction *R) {
                inferScalarType(R->getOperand(1)) &&
            "different types inferred for different operands");
     return IntegerType::get(Ctx, 1);
+  case VPInstruction::ComputeFindLastIVResult:
   case VPInstruction::ComputeReductionResult: {
     auto *PhiR = cast<VPReductionPHIRecipe>(R->getOperand(0));
     auto *OrigPhi = cast<PHINode>(PhiR->getUnderlyingValue());
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -575,6 +575,29 @@ Value *VPInstruction::generate(VPTransformState &State) {
     Builder.GetInsertBlock()->getTerminator()->eraseFromParent();
     return CondBr;
   }
+  case VPInstruction::ComputeFindLastIVResult: {
+    // FIXME: The cross-recipe dependency on VPReductionPHIRecipe is temporary
+    // and will be removed by breaking up the recipe further.
+    auto *PhiR = cast<VPReductionPHIRecipe>(getOperand(0));
+    // Get its reduction variable descriptor.
+    const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor();
+    RecurKind RK = RdxDesc.getRecurrenceKind();
+    assert(RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK) &&
+           "Unexpected reduction kind");
+    assert(!PhiR->isInLoop() &&
+           "In-loop FindLastIV reduction is not supported yet");
+
+    // The recipe's operands are the reduction phi, followed by one operand for
+    // each part of the reduction.
+    unsigned UF = getNumOperands() - 1;
+    Value *ReducedPartRdx = State.get(getOperand(1));
+    for (unsigned Part = 1; Part < UF; ++Part) {
+      ReducedPartRdx = createMinMaxOp(Builder, RecurKind::SMax, ReducedPartRdx,
+                                      State.get(getOperand(1 + Part)));
+    }
+
+    return createFindLastIVReduction(Builder, ReducedPartRdx, RdxDesc);
+  }
   case VPInstruction::ComputeReductionResult: {
     // FIXME: The cross-recipe dependency on VPReductionPHIRecipe is temporary
     // and will be removed by breaking up the recipe further.
@@ -584,6 +607,8 @@ Value *VPInstruction::generate(VPTransformState &State) {
     const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor();
 
     RecurKind RK = RdxDesc.getRecurrenceKind();
+    assert(!RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK) &&
+           "should be handled by ComputeFindLastIVResult");
 
     Type *PhiTy = OrigPhi->getType();
     // The recipe's operands are the reduction phi, followed by one operand for
@@ -619,9 +644,6 @@ Value *VPInstruction::generate(VPTransformState &State) {
         if (Op != Instruction::ICmp && Op != Instruction::FCmp)
           ReducedPartRdx = Builder.CreateBinOp(
               (Instruction::BinaryOps)Op, RdxPart, ReducedPartRdx, "bin.rdx");
-        else if (RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK))
-          ReducedPartRdx =
-              createMinMaxOp(Builder, RecurKind::SMax, ReducedPartRdx, RdxPart);
         else
           ReducedPartRdx = createMinMaxOp(Builder, RK, ReducedPartRdx, RdxPart);
       }
@@ -630,8 +652,7 @@ Value *VPInstruction::generate(VPTransformState &State) {
     // Create the reduction after the loop. Note that inloop reductions create
     // the target reduction in the loop using a Reduction recipe.
     if ((State.VF.isVector() ||
-         RecurrenceDescriptor::isAnyOfRecurrenceKind(RK) ||
-         RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK)) &&
+         RecurrenceDescriptor::isAnyOfRecurrenceKind(RK)) &&
         !PhiR->isInLoop()) {
       // TODO: Support in-order reductions based on the recurrence descriptor.
       // All ops in the reduction inherit fast-math-flags from the recurrence
@@ -642,9 +663,6 @@ Value *VPInstruction::generate(VPTransformState &State) {
       if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RK))
         ReducedPartRdx =
             createAnyOfReduction(Builder, ReducedPartRdx, RdxDesc, OrigPhi);
-      else if (RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK))
-        ReducedPartRdx =
-            createFindLastIVReduction(Builder, ReducedPartRdx, RdxDesc);
       else
         ReducedPartRdx = createSimpleReduction(Builder, ReducedPartRdx, RK);
 
@@ -742,6 +760,7 @@ InstructionCost VPInstruction::computeCost(ElementCount VF,
 
 bool VPInstruction::isVectorToScalar() const {
   return getOpcode() == VPInstruction::ExtractFromEnd ||
+         getOpcode() == VPInstruction::ComputeFindLastIVResult ||
          getOpcode() == VPInstruction::ComputeReductionResult ||
          getOpcode() == VPInstruction::AnyOf;
 }
@@ -913,6 +932,9 @@ void VPInstruction::print(raw_ostream &O, const Twine &Indent,
   case VPInstruction::ExtractFromEnd:
     O << "extract-from-end";
     break;
+  case VPInstruction::ComputeFindLastIVResult:
+    O << "compute-find-last-iv-result";
+    break;
   case VPInstruction::ComputeReductionResult:
     O << "compute-reduction-result";
     break;
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
@@ -348,6 +348,8 @@ void UnrollState::unrollBlock(VPBlockBase *VPB) {
     // the parts to compute the final reduction value.
     VPValue *Op1;
     if (match(&R, m_VPInstruction<VPInstruction::ComputeReductionResult>(
+                      m_VPValue(), m_VPValue(Op1))) ||
+        match(&R, m_VPInstruction<VPInstruction::ComputeFindLastIVResult>(
                       m_VPValue(), m_VPValue(Op1)))) {
       addUniformForAllParts(cast<VPInstruction>(&R));
       for (unsigned Part = 1; Part != UF; ++Part)
diff --git a/llvm/test/Transforms/LoopVectorize/vplan-printing-reductions.ll b/llvm/test/Transforms/LoopVectorize/vplan-printing-reductions.ll
@@ -234,7 +234,7 @@ define i64 @find_last_iv(ptr %a, i64 %n, i64 %start) {
 ; CHECK-NEXT: Successor(s): middle.block
 ; CHECK-EMPTY:
 ; CHECK-NEXT: middle.block:
-; CHECK-NEXT:   EMIT vp<[[RDX_RES:%.+]]> = compute-reduction-result ir<%rdx>, ir<%cond>
+; CHECK-NEXT:   EMIT vp<[[RDX_RES:%.+]]> = compute-find-last-iv-result ir<%rdx>, ir<%cond>
 ; CHECK-NEXT:   EMIT vp<[[EXT:%.+]]> = extract-from-end vp<[[RDX_RES]]>, ir<1>
 ; CHECK-NEXT:   EMIT vp<%cmp.n> = icmp eq ir<%n>, vp<{{.+}}>
 ; CHECK-NEXT:   EMIT branch-on-cond vp<%cmp.n>
