[llvm] [LV] Reduce register usage for scaled reductions (PR #133090)

Sam Tebbs via llvm-commits llvm-commits at lists.llvm.org
Fri Apr 11 01:46:41 PDT 2025


https://github.com/SamTebbs33 updated https://github.com/llvm/llvm-project/pull/133090

>From a994bf2e1f8caa6b884696b12dafd56791e7f227 Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Wed, 26 Mar 2025 14:01:59 +0000
Subject: [PATCH 01/19] [LV] Reduce register usage for scaled reductions

---
 .../Transforms/Vectorize/LoopVectorize.cpp    |  25 +++-
 .../Transforms/Vectorize/VPRecipeBuilder.h    |   3 +-
 llvm/lib/Transforms/Vectorize/VPlan.h         |  14 ++-
 .../partial-reduce-dot-product-neon.ll        | 116 ++++++++++++------
 .../AArch64/partial-reduce-dot-product.ll     | 114 ++++++++---------
 5 files changed, 166 insertions(+), 106 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 141a4fd83c833..d07fbe241ed10 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5014,7 +5014,6 @@ calculateRegisterUsage(VPlan &Plan, ArrayRef<ElementCount> VFs,
         if (isa<VPVectorPointerRecipe, VPVectorEndPointerRecipe,
                 VPBranchOnMaskRecipe>(R))
           continue;
-
         if (VFs[J].isScalar() ||
             isa<VPCanonicalIVPHIRecipe, VPReplicateRecipe, VPDerivedIVRecipe,
                 VPScalarIVStepsRecipe>(R) ||
@@ -5028,10 +5027,23 @@ calculateRegisterUsage(VPlan &Plan, ArrayRef<ElementCount> VFs,
           // even in the scalar case.
           RegUsage[ClassID] += 1;
         } else {
+          // The output from scaled phis and scaled reductions actually have
+          // fewer lanes than the VF.
+          auto VF = VFs[J];
+          if (auto *ReductionR = dyn_cast<VPReductionPHIRecipe>(R))
+            VF = VF.divideCoefficientBy(ReductionR->getVFScaleFactor());
+          else if (auto *PartialReductionR =
+                       dyn_cast<VPPartialReductionRecipe>(R))
+            VF = VF.divideCoefficientBy(PartialReductionR->getScaleFactor());
+          if (VF != VFs[J])
+            LLVM_DEBUG(dbgs() << "LV(REG): Scaled down VF from " << VFs[J]
+                              << " to " << VF << " for ";
+                       R->dump(););
+
           for (VPValue *DefV : R->definedValues()) {
             Type *ScalarTy = TypeInfo.inferScalarType(DefV);
             unsigned ClassID = TTI.getRegisterClassForType(true, ScalarTy);
-            RegUsage[ClassID] += GetRegUsage(ScalarTy, VFs[J]);
+            RegUsage[ClassID] += GetRegUsage(ScalarTy, VF);
           }
         }
       }
@@ -9137,8 +9149,8 @@ VPRecipeBase *VPRecipeBuilder::tryToCreateWidenRecipe(
   if (isa<LoadInst>(Instr) || isa<StoreInst>(Instr))
     return tryToWidenMemory(Instr, Operands, Range);
 
-  if (getScalingForReduction(Instr))
-    return tryToCreatePartialReduction(Instr, Operands);
+  if (auto ScaleFactor = getScalingForReduction(Instr))
+    return tryToCreatePartialReduction(Instr, Operands, ScaleFactor.value());
 
   if (!shouldWiden(Instr, Range))
     return nullptr;
@@ -9162,7 +9174,8 @@ VPRecipeBase *VPRecipeBuilder::tryToCreateWidenRecipe(
 
 VPRecipeBase *
 VPRecipeBuilder::tryToCreatePartialReduction(Instruction *Reduction,
-                                             ArrayRef<VPValue *> Operands) {
+                                             ArrayRef<VPValue *> Operands,
+                                             unsigned ScaleFactor) {
   assert(Operands.size() == 2 &&
          "Unexpected number of operands for partial reduction");
 
@@ -9195,7 +9208,7 @@ VPRecipeBuilder::tryToCreatePartialReduction(Instruction *Reduction,
     BinOp = Builder.createSelect(Mask, BinOp, Zero, Reduction->getDebugLoc());
   }
   return new VPPartialReductionRecipe(ReductionOpcode, BinOp, Accumulator,
-                                      Reduction);
+                                      ScaleFactor, Reduction);
 }
 
 void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF,
diff --git a/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h b/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h
index 334cfbad8bd7c..fd0064a34c4c9 100644
--- a/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h
+++ b/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h
@@ -178,7 +178,8 @@ class VPRecipeBuilder {
   /// Create and return a partial reduction recipe for a reduction instruction
   /// along with binary operation and reduction phi operands.
   VPRecipeBase *tryToCreatePartialReduction(Instruction *Reduction,
-                                            ArrayRef<VPValue *> Operands);
+                                            ArrayRef<VPValue *> Operands,
+                                            unsigned ScaleFactor);
 
   /// Set the recipe created for given ingredient.
   void setRecipe(Instruction *I, VPRecipeBase *R) {
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index da7aef73f9df3..35bb4ce85a6d8 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -2033,6 +2033,8 @@ class VPReductionPHIRecipe : public VPHeaderPHIRecipe,
   /// Generate the phi/select nodes.
   void execute(VPTransformState &State) override;
 
+  unsigned getVFScaleFactor() const { return VFScaleFactor; }
+
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
   /// Print the recipe.
   void print(raw_ostream &O, const Twine &Indent,
@@ -2063,17 +2065,19 @@ class VPReductionPHIRecipe : public VPHeaderPHIRecipe,
 /// scalar value.
 class VPPartialReductionRecipe : public VPSingleDefRecipe {
   unsigned Opcode;
+  unsigned ScaleFactor;
 
 public:
   VPPartialReductionRecipe(Instruction *ReductionInst, VPValue *Op0,
-                           VPValue *Op1)
+                           VPValue *Op1, unsigned ScaleFactor)
       : VPPartialReductionRecipe(ReductionInst->getOpcode(), Op0, Op1,
-                                 ReductionInst) {}
+                                 ScaleFactor, ReductionInst) {}
   VPPartialReductionRecipe(unsigned Opcode, VPValue *Op0, VPValue *Op1,
+                           unsigned ScaleFactor,
                            Instruction *ReductionInst = nullptr)
       : VPSingleDefRecipe(VPDef::VPPartialReductionSC,
                           ArrayRef<VPValue *>({Op0, Op1}), ReductionInst),
-        Opcode(Opcode) {
+        Opcode(Opcode), ScaleFactor(ScaleFactor) {
     [[maybe_unused]] auto *AccumulatorRecipe =
         getOperand(1)->getDefiningRecipe();
     assert((isa<VPReductionPHIRecipe>(AccumulatorRecipe) ||
@@ -2084,7 +2088,7 @@ class VPPartialReductionRecipe : public VPSingleDefRecipe {
 
   VPPartialReductionRecipe *clone() override {
     return new VPPartialReductionRecipe(Opcode, getOperand(0), getOperand(1),
-                                        getUnderlyingInstr());
+                                        ScaleFactor, getUnderlyingInstr());
   }
 
   VP_CLASSOF_IMPL(VPDef::VPPartialReductionSC)
@@ -2099,6 +2103,8 @@ class VPPartialReductionRecipe : public VPSingleDefRecipe {
   /// Get the binary op's opcode.
   unsigned getOpcode() const { return Opcode; }
 
+  unsigned getScaleFactor() const { return ScaleFactor; }
+
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
   /// Print the recipe.
   void print(raw_ostream &O, const Twine &Indent,
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
index a8476dbddb3c2..a55a5813044cb 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
@@ -770,10 +770,10 @@ define i32 @dotp_unrolled(i32 %num_out, i64 %num_in, ptr %a, ptr %b) {
 ; CHECK-INTERLEAVED-LABEL: define i32 @dotp_unrolled(
 ; CHECK-INTERLEAVED-SAME: i32 [[NUM_OUT:%.*]], i64 [[NUM_IN:%.*]], ptr [[A:%.*]], ptr [[B:%.*]]) #[[ATTR0]] {
 ; CHECK-INTERLEAVED-NEXT:  entry:
-; CHECK-INTERLEAVED-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[NUM_IN]], 16
+; CHECK-INTERLEAVED-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[NUM_IN]], 32
 ; CHECK-INTERLEAVED-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK-INTERLEAVED:       vector.ph:
-; CHECK-INTERLEAVED-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[NUM_IN]], 16
+; CHECK-INTERLEAVED-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[NUM_IN]], 32
 ; CHECK-INTERLEAVED-NEXT:    [[N_VEC:%.*]] = sub i64 [[NUM_IN]], [[N_MOD_VF]]
 ; CHECK-INTERLEAVED-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK-INTERLEAVED:       vector.body:
@@ -782,6 +782,10 @@ define i32 @dotp_unrolled(i32 %num_out, i64 %num_in, ptr %a, ptr %b) {
 ; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE10:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE7:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI4:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE16:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI5:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE17:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI6:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE1:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI7:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE11:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-INTERLEAVED-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[INDEX]]
 ; CHECK-INTERLEAVED-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[INDEX]]
 ; CHECK-INTERLEAVED-NEXT:    [[TMP3:%.*]] = or disjoint i64 [[INDEX]], 1
@@ -794,45 +798,81 @@ define i32 @dotp_unrolled(i32 %num_out, i64 %num_in, ptr %a, ptr %b) {
 ; CHECK-INTERLEAVED-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP9]]
 ; CHECK-INTERLEAVED-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP9]]
 ; CHECK-INTERLEAVED-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 0
+; CHECK-INTERLEAVED-NEXT:    [[TMP38:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 16
 ; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD:%.*]] = load <16 x i8>, ptr [[TMP12]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD8:%.*]] = load <16 x i8>, ptr [[TMP38]], align 1
 ; CHECK-INTERLEAVED-NEXT:    [[TMP13:%.*]] = sext <16 x i8> [[WIDE_LOAD]] to <16 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i8, ptr [[TMP2]], i32 0
-; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD4:%.*]] = load <16 x i8>, ptr [[TMP14]], align 1
-; CHECK-INTERLEAVED-NEXT:    [[TMP15:%.*]] = sext <16 x i8> [[WIDE_LOAD4]] to <16 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP16:%.*]] = mul nsw <16 x i32> [[TMP15]], [[TMP13]]
-; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI3]], <16 x i32> [[TMP16]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP17:%.*]] = getelementptr inbounds i8, ptr [[TMP4]], i32 0
-; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD5:%.*]] = load <16 x i8>, ptr [[TMP17]], align 1
-; CHECK-INTERLEAVED-NEXT:    [[TMP18:%.*]] = sext <16 x i8> [[WIDE_LOAD5]] to <16 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP19:%.*]] = getelementptr inbounds i8, ptr [[TMP5]], i32 0
-; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD6:%.*]] = load <16 x i8>, ptr [[TMP19]], align 1
-; CHECK-INTERLEAVED-NEXT:    [[TMP20:%.*]] = sext <16 x i8> [[WIDE_LOAD6]] to <16 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP21:%.*]] = mul nsw <16 x i32> [[TMP18]], [[TMP20]]
-; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE7]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI2]], <16 x i32> [[TMP21]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP22:%.*]] = getelementptr inbounds i8, ptr [[TMP7]], i32 0
-; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD8:%.*]] = load <16 x i8>, ptr [[TMP22]], align 1
-; CHECK-INTERLEAVED-NEXT:    [[TMP23:%.*]] = sext <16 x i8> [[WIDE_LOAD8]] to <16 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP24:%.*]] = getelementptr inbounds i8, ptr [[TMP8]], i32 0
-; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD9:%.*]] = load <16 x i8>, ptr [[TMP24]], align 1
-; CHECK-INTERLEAVED-NEXT:    [[TMP25:%.*]] = sext <16 x i8> [[WIDE_LOAD9]] to <16 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP26:%.*]] = mul nsw <16 x i32> [[TMP23]], [[TMP25]]
-; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE10]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI1]], <16 x i32> [[TMP26]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP27:%.*]] = getelementptr inbounds i8, ptr [[TMP10]], i32 0
-; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD11:%.*]] = load <16 x i8>, ptr [[TMP27]], align 1
-; CHECK-INTERLEAVED-NEXT:    [[TMP28:%.*]] = sext <16 x i8> [[WIDE_LOAD11]] to <16 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP29:%.*]] = getelementptr inbounds i8, ptr [[TMP11]], i32 0
-; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD12:%.*]] = load <16 x i8>, ptr [[TMP29]], align 1
-; CHECK-INTERLEAVED-NEXT:    [[TMP30:%.*]] = sext <16 x i8> [[WIDE_LOAD12]] to <16 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP31:%.*]] = mul nsw <16 x i32> [[TMP28]], [[TMP30]]
-; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE13]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI]], <16 x i32> [[TMP31]])
-; CHECK-INTERLEAVED-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 16
-; CHECK-INTERLEAVED-NEXT:    [[TMP32:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
-; CHECK-INTERLEAVED-NEXT:    br i1 [[TMP32]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP14:%.*]] = sext <16 x i8> [[WIDE_LOAD8]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP15:%.*]] = getelementptr inbounds i8, ptr [[TMP2]], i32 0
+; CHECK-INTERLEAVED-NEXT:    [[TMP16:%.*]] = getelementptr inbounds i8, ptr [[TMP2]], i32 16
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD9:%.*]] = load <16 x i8>, ptr [[TMP15]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD10:%.*]] = load <16 x i8>, ptr [[TMP16]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[TMP17:%.*]] = sext <16 x i8> [[WIDE_LOAD9]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP18:%.*]] = sext <16 x i8> [[WIDE_LOAD10]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP19:%.*]] = mul nsw <16 x i32> [[TMP17]], [[TMP13]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP20:%.*]] = mul nsw <16 x i32> [[TMP18]], [[TMP14]]
+; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE1]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI6]], <16 x i32> [[TMP19]])
+; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE11]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI7]], <16 x i32> [[TMP20]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[TMP4]], i32 0
+; CHECK-INTERLEAVED-NEXT:    [[TMP22:%.*]] = getelementptr inbounds i8, ptr [[TMP4]], i32 16
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD12:%.*]] = load <16 x i8>, ptr [[TMP21]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD13:%.*]] = load <16 x i8>, ptr [[TMP22]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[TMP23:%.*]] = sext <16 x i8> [[WIDE_LOAD12]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP24:%.*]] = sext <16 x i8> [[WIDE_LOAD13]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP25:%.*]] = getelementptr inbounds i8, ptr [[TMP5]], i32 0
+; CHECK-INTERLEAVED-NEXT:    [[TMP26:%.*]] = getelementptr inbounds i8, ptr [[TMP5]], i32 16
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD14:%.*]] = load <16 x i8>, ptr [[TMP25]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD15:%.*]] = load <16 x i8>, ptr [[TMP26]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[TMP27:%.*]] = sext <16 x i8> [[WIDE_LOAD14]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP28:%.*]] = sext <16 x i8> [[WIDE_LOAD15]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP29:%.*]] = mul nsw <16 x i32> [[TMP23]], [[TMP27]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP30:%.*]] = mul nsw <16 x i32> [[TMP24]], [[TMP28]]
+; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE16]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI4]], <16 x i32> [[TMP29]])
+; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE17]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI5]], <16 x i32> [[TMP30]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP31:%.*]] = getelementptr inbounds i8, ptr [[TMP7]], i32 0
+; CHECK-INTERLEAVED-NEXT:    [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[TMP7]], i32 16
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD18:%.*]] = load <16 x i8>, ptr [[TMP31]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD19:%.*]] = load <16 x i8>, ptr [[TMP32]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[TMP33:%.*]] = sext <16 x i8> [[WIDE_LOAD18]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP34:%.*]] = sext <16 x i8> [[WIDE_LOAD19]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP35:%.*]] = getelementptr inbounds i8, ptr [[TMP8]], i32 0
+; CHECK-INTERLEAVED-NEXT:    [[TMP36:%.*]] = getelementptr inbounds i8, ptr [[TMP8]], i32 16
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD20:%.*]] = load <16 x i8>, ptr [[TMP35]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD21:%.*]] = load <16 x i8>, ptr [[TMP36]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[TMP37:%.*]] = sext <16 x i8> [[WIDE_LOAD20]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP56:%.*]] = sext <16 x i8> [[WIDE_LOAD21]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP39:%.*]] = mul nsw <16 x i32> [[TMP33]], [[TMP37]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP40:%.*]] = mul nsw <16 x i32> [[TMP34]], [[TMP56]]
+; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE7]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI2]], <16 x i32> [[TMP39]])
+; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI3]], <16 x i32> [[TMP40]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP41:%.*]] = getelementptr inbounds i8, ptr [[TMP10]], i32 0
+; CHECK-INTERLEAVED-NEXT:    [[TMP42:%.*]] = getelementptr inbounds i8, ptr [[TMP10]], i32 16
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD24:%.*]] = load <16 x i8>, ptr [[TMP41]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD25:%.*]] = load <16 x i8>, ptr [[TMP42]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[TMP43:%.*]] = sext <16 x i8> [[WIDE_LOAD24]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP44:%.*]] = sext <16 x i8> [[WIDE_LOAD25]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP45:%.*]] = getelementptr inbounds i8, ptr [[TMP11]], i32 0
+; CHECK-INTERLEAVED-NEXT:    [[TMP46:%.*]] = getelementptr inbounds i8, ptr [[TMP11]], i32 16
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD26:%.*]] = load <16 x i8>, ptr [[TMP45]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD27:%.*]] = load <16 x i8>, ptr [[TMP46]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[TMP47:%.*]] = sext <16 x i8> [[WIDE_LOAD26]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP48:%.*]] = sext <16 x i8> [[WIDE_LOAD27]] to <16 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP49:%.*]] = mul nsw <16 x i32> [[TMP43]], [[TMP47]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP50:%.*]] = mul nsw <16 x i32> [[TMP44]], [[TMP48]]
+; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE13]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI]], <16 x i32> [[TMP49]])
+; CHECK-INTERLEAVED-NEXT:    [[PARTIAL_REDUCE10]] = call <4 x i32> @llvm.experimental.vector.partial.reduce.add.v4i32.v16i32(<4 x i32> [[VEC_PHI1]], <16 x i32> [[TMP50]])
+; CHECK-INTERLEAVED-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 32
+; CHECK-INTERLEAVED-NEXT:    [[TMP51:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-INTERLEAVED-NEXT:    br i1 [[TMP51]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
 ; CHECK-INTERLEAVED:       middle.block:
-; CHECK-INTERLEAVED-NEXT:    [[TMP33:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[PARTIAL_REDUCE13]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP34:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[PARTIAL_REDUCE10]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP35:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[PARTIAL_REDUCE7]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP36:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[PARTIAL_REDUCE]])
+; CHECK-INTERLEAVED-NEXT:    [[BIN_RDX:%.*]] = add <4 x i32> [[PARTIAL_REDUCE10]], [[PARTIAL_REDUCE13]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP52:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX]])
+; CHECK-INTERLEAVED-NEXT:    [[BIN_RDX30:%.*]] = add <4 x i32> [[PARTIAL_REDUCE]], [[PARTIAL_REDUCE7]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP53:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX30]])
+; CHECK-INTERLEAVED-NEXT:    [[BIN_RDX31:%.*]] = add <4 x i32> [[PARTIAL_REDUCE17]], [[PARTIAL_REDUCE16]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP54:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX31]])
+; CHECK-INTERLEAVED-NEXT:    [[BIN_RDX32:%.*]] = add <4 x i32> [[PARTIAL_REDUCE11]], [[PARTIAL_REDUCE1]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP55:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX32]])
 ; CHECK-INTERLEAVED-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[NUM_IN]], [[N_VEC]]
 ; CHECK-INTERLEAVED-NEXT:    br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
 ; CHECK-INTERLEAVED:       scalar.ph:
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
index bbcc6db020307..541cf034e0f3a 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
@@ -3256,65 +3256,65 @@ define dso_local void @not_dotp_high_register_pressure(ptr %a, ptr %b, ptr %sum,
 ; CHECK-INTERLEAVED-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK-INTERLEAVED:       vector.body:
 ; CHECK-INTERLEAVED-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ [[TMP0]], [[VECTOR_PH]] ], [ [[TMP64:%.*]], [[VECTOR_BODY]] ]
-; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x i32> [ [[TMP1]], [[VECTOR_PH]] ], [ [[TMP58:%.*]], [[VECTOR_BODY]] ]
-; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI4:%.*]] = phi <4 x i32> [ [[TMP2]], [[VECTOR_PH]] ], [ [[TMP52:%.*]], [[VECTOR_BODY]] ]
-; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI6:%.*]] = phi <4 x i32> [ [[TMP3]], [[VECTOR_PH]] ], [ [[TMP46:%.*]], [[VECTOR_BODY]] ]
-; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI8:%.*]] = phi <4 x i32> [ [[TMP4]], [[VECTOR_PH]] ], [ [[TMP40:%.*]], [[VECTOR_BODY]] ]
-; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI10:%.*]] = phi <4 x i32> [ [[TMP5]], [[VECTOR_PH]] ], [ [[TMP34:%.*]], [[VECTOR_BODY]] ]
-; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI12:%.*]] = phi <4 x i32> [ [[TMP6]], [[VECTOR_PH]] ], [ [[TMP28:%.*]], [[VECTOR_BODY]] ]
-; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI14:%.*]] = phi <4 x i32> [ [[TMP7]], [[VECTOR_PH]] ], [ [[TMP22:%.*]], [[VECTOR_BODY]] ]
-; CHECK-INTERLEAVED-NEXT:    [[TMP9:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 [[INDEX]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP10:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP9]], i32 0
-; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD16:%.*]] = load <4 x i8>, ptr [[TMP10]], align 1
-; CHECK-INTERLEAVED-NEXT:    [[TMP13:%.*]] = zext <4 x i8> [[WIDE_LOAD16]] to <4 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP15:%.*]] = shl nsw i64 [[INDEX]], 3
-; CHECK-INTERLEAVED-NEXT:    [[TMP17:%.*]] = getelementptr inbounds nuw i8, ptr [[B]], i64 [[TMP15]]
-; CHECK-INTERLEAVED-NEXT:    [[WIDE_VEC24:%.*]] = load <32 x i8>, ptr [[TMP17]], align 1
-; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC25:%.*]] = shufflevector <32 x i8> [[WIDE_VEC24]], <32 x i8> poison, <4 x i32> <i32 0, i32 8, i32 16, i32 24>
-; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC26:%.*]] = shufflevector <32 x i8> [[WIDE_VEC24]], <32 x i8> poison, <4 x i32> <i32 1, i32 9, i32 17, i32 25>
-; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC27:%.*]] = shufflevector <32 x i8> [[WIDE_VEC24]], <32 x i8> poison, <4 x i32> <i32 2, i32 10, i32 18, i32 26>
-; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC28:%.*]] = shufflevector <32 x i8> [[WIDE_VEC24]], <32 x i8> poison, <4 x i32> <i32 3, i32 11, i32 19, i32 27>
-; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC29:%.*]] = shufflevector <32 x i8> [[WIDE_VEC24]], <32 x i8> poison, <4 x i32> <i32 4, i32 12, i32 20, i32 28>
-; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC30:%.*]] = shufflevector <32 x i8> [[WIDE_VEC24]], <32 x i8> poison, <4 x i32> <i32 5, i32 13, i32 21, i32 29>
-; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC31:%.*]] = shufflevector <32 x i8> [[WIDE_VEC24]], <32 x i8> poison, <4 x i32> <i32 6, i32 14, i32 22, i32 30>
-; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC32:%.*]] = shufflevector <32 x i8> [[WIDE_VEC24]], <32 x i8> poison, <4 x i32> <i32 7, i32 15, i32 23, i32 31>
-; CHECK-INTERLEAVED-NEXT:    [[TMP19:%.*]] = sext <4 x i8> [[STRIDED_VEC25]] to <4 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP21:%.*]] = mul nsw <4 x i32> [[TMP19]], [[TMP13]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP22]] = add <4 x i32> [[TMP21]], [[VEC_PHI14]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP25:%.*]] = sext <4 x i8> [[STRIDED_VEC26]] to <4 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP27:%.*]] = mul nsw <4 x i32> [[TMP25]], [[TMP13]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP28]] = add <4 x i32> [[TMP27]], [[VEC_PHI12]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP31:%.*]] = sext <4 x i8> [[STRIDED_VEC27]] to <4 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP33:%.*]] = mul nsw <4 x i32> [[TMP31]], [[TMP13]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP34]] = add <4 x i32> [[TMP33]], [[VEC_PHI10]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP37:%.*]] = sext <4 x i8> [[STRIDED_VEC28]] to <4 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP39:%.*]] = mul nsw <4 x i32> [[TMP37]], [[TMP13]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP40]] = add <4 x i32> [[TMP39]], [[VEC_PHI8]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP43:%.*]] = sext <4 x i8> [[STRIDED_VEC29]] to <4 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP45:%.*]] = mul nsw <4 x i32> [[TMP43]], [[TMP13]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP46]] = add <4 x i32> [[TMP45]], [[VEC_PHI6]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP49:%.*]] = sext <4 x i8> [[STRIDED_VEC30]] to <4 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP51:%.*]] = mul nsw <4 x i32> [[TMP49]], [[TMP13]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP52]] = add <4 x i32> [[TMP51]], [[VEC_PHI4]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP55:%.*]] = sext <4 x i8> [[STRIDED_VEC31]] to <4 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP57:%.*]] = mul nsw <4 x i32> [[TMP55]], [[TMP13]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP58]] = add <4 x i32> [[TMP57]], [[VEC_PHI2]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP61:%.*]] = sext <4 x i8> [[STRIDED_VEC32]] to <4 x i32>
-; CHECK-INTERLEAVED-NEXT:    [[TMP63:%.*]] = mul nsw <4 x i32> [[TMP61]], [[TMP13]]
-; CHECK-INTERLEAVED-NEXT:    [[TMP64]] = add <4 x i32> [[TMP63]], [[VEC_PHI]]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ [[TMP0]], [[VECTOR_PH]] ], [ [[TMP36:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ [[TMP1]], [[VECTOR_PH]] ], [ [[TMP33:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x i32> [ [[TMP2]], [[VECTOR_PH]] ], [ [[TMP30:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI3:%.*]] = phi <4 x i32> [ [[TMP3]], [[VECTOR_PH]] ], [ [[TMP27:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI4:%.*]] = phi <4 x i32> [ [[TMP4]], [[VECTOR_PH]] ], [ [[TMP24:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI5:%.*]] = phi <4 x i32> [ [[TMP5]], [[VECTOR_PH]] ], [ [[TMP21:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI6:%.*]] = phi <4 x i32> [ [[TMP6]], [[VECTOR_PH]] ], [ [[TMP18:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[VEC_PHI7:%.*]] = phi <4 x i32> [ [[TMP7]], [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT:    [[TMP8:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 [[INDEX]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP9:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP8]], i32 0
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i8>, ptr [[TMP9]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[TMP10:%.*]] = zext <4 x i8> [[WIDE_LOAD]] to <4 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP11:%.*]] = shl nsw i64 [[INDEX]], 3
+; CHECK-INTERLEAVED-NEXT:    [[TMP12:%.*]] = getelementptr inbounds nuw i8, ptr [[B]], i64 [[TMP11]]
+; CHECK-INTERLEAVED-NEXT:    [[WIDE_VEC:%.*]] = load <32 x i8>, ptr [[TMP12]], align 1
+; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <32 x i8> [[WIDE_VEC]], <32 x i8> poison, <4 x i32> <i32 0, i32 8, i32 16, i32 24>
+; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC8:%.*]] = shufflevector <32 x i8> [[WIDE_VEC]], <32 x i8> poison, <4 x i32> <i32 1, i32 9, i32 17, i32 25>
+; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC9:%.*]] = shufflevector <32 x i8> [[WIDE_VEC]], <32 x i8> poison, <4 x i32> <i32 2, i32 10, i32 18, i32 26>
+; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC10:%.*]] = shufflevector <32 x i8> [[WIDE_VEC]], <32 x i8> poison, <4 x i32> <i32 3, i32 11, i32 19, i32 27>
+; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC11:%.*]] = shufflevector <32 x i8> [[WIDE_VEC]], <32 x i8> poison, <4 x i32> <i32 4, i32 12, i32 20, i32 28>
+; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC12:%.*]] = shufflevector <32 x i8> [[WIDE_VEC]], <32 x i8> poison, <4 x i32> <i32 5, i32 13, i32 21, i32 29>
+; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC13:%.*]] = shufflevector <32 x i8> [[WIDE_VEC]], <32 x i8> poison, <4 x i32> <i32 6, i32 14, i32 22, i32 30>
+; CHECK-INTERLEAVED-NEXT:    [[STRIDED_VEC14:%.*]] = shufflevector <32 x i8> [[WIDE_VEC]], <32 x i8> poison, <4 x i32> <i32 7, i32 15, i32 23, i32 31>
+; CHECK-INTERLEAVED-NEXT:    [[TMP13:%.*]] = sext <4 x i8> [[STRIDED_VEC]] to <4 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP14:%.*]] = mul nsw <4 x i32> [[TMP13]], [[TMP10]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP15]] = add <4 x i32> [[TMP14]], [[VEC_PHI7]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP16:%.*]] = sext <4 x i8> [[STRIDED_VEC8]] to <4 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP17:%.*]] = mul nsw <4 x i32> [[TMP16]], [[TMP10]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP18]] = add <4 x i32> [[TMP17]], [[VEC_PHI6]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP19:%.*]] = sext <4 x i8> [[STRIDED_VEC9]] to <4 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP20:%.*]] = mul nsw <4 x i32> [[TMP19]], [[TMP10]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP21]] = add <4 x i32> [[TMP20]], [[VEC_PHI5]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP22:%.*]] = sext <4 x i8> [[STRIDED_VEC10]] to <4 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP23:%.*]] = mul nsw <4 x i32> [[TMP22]], [[TMP10]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP24]] = add <4 x i32> [[TMP23]], [[VEC_PHI4]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP25:%.*]] = sext <4 x i8> [[STRIDED_VEC11]] to <4 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP26:%.*]] = mul nsw <4 x i32> [[TMP25]], [[TMP10]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP27]] = add <4 x i32> [[TMP26]], [[VEC_PHI3]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP28:%.*]] = sext <4 x i8> [[STRIDED_VEC12]] to <4 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP29:%.*]] = mul nsw <4 x i32> [[TMP28]], [[TMP10]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP30]] = add <4 x i32> [[TMP29]], [[VEC_PHI2]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP31:%.*]] = sext <4 x i8> [[STRIDED_VEC13]] to <4 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP32:%.*]] = mul nsw <4 x i32> [[TMP31]], [[TMP10]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP33]] = add <4 x i32> [[TMP32]], [[VEC_PHI1]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP34:%.*]] = sext <4 x i8> [[STRIDED_VEC14]] to <4 x i32>
+; CHECK-INTERLEAVED-NEXT:    [[TMP35:%.*]] = mul nsw <4 x i32> [[TMP34]], [[TMP10]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP36]] = add <4 x i32> [[TMP35]], [[VEC_PHI]]
 ; CHECK-INTERLEAVED-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
-; CHECK-INTERLEAVED-NEXT:    [[TMP66:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
-; CHECK-INTERLEAVED-NEXT:    br i1 [[TMP66]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP41:![0-9]+]]
+; CHECK-INTERLEAVED-NEXT:    [[TMP37:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-INTERLEAVED-NEXT:    br i1 [[TMP37]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP41:![0-9]+]]
 ; CHECK-INTERLEAVED:       middle.block:
-; CHECK-INTERLEAVED-NEXT:    [[TMP38:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP64]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP47:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP58]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP48:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP52]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP41:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP46]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP42:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP40]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP50:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP34]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP44:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP28]])
-; CHECK-INTERLEAVED-NEXT:    [[TMP53:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP22]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP38:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP36]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP39:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP33]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP40:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP30]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP41:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP27]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP42:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP24]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP43:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP21]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP44:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP18]])
+; CHECK-INTERLEAVED-NEXT:    [[TMP45:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP15]])
 ; CHECK-INTERLEAVED-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
 ; CHECK-INTERLEAVED-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_FOR_COND_CLEANUP_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK-INTERLEAVED:       scalar.ph:

>From 82151a8f22044e42302417a6fcbfade2f2a66cfd Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Mon, 31 Mar 2025 15:23:46 +0100
Subject: [PATCH 02/19] Rename to getVFScaleFactor

---
 llvm/lib/Transforms/Vectorize/LoopVectorize.cpp |  2 +-
 llvm/lib/Transforms/Vectorize/VPlan.h           | 14 +++++++-------
 2 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index d07fbe241ed10..ff96e98ef9cfa 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5034,7 +5034,7 @@ calculateRegisterUsage(VPlan &Plan, ArrayRef<ElementCount> VFs,
             VF = VF.divideCoefficientBy(ReductionR->getVFScaleFactor());
           else if (auto *PartialReductionR =
                        dyn_cast<VPPartialReductionRecipe>(R))
-            VF = VF.divideCoefficientBy(PartialReductionR->getScaleFactor());
+            VF = VF.divideCoefficientBy(PartialReductionR->getVFScaleFactor());
           if (VF != VFs[J])
             LLVM_DEBUG(dbgs() << "LV(REG): Scaled down VF from " << VFs[J]
                               << " to " << VF << " for ";
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 35bb4ce85a6d8..6928bea4bb092 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -2065,19 +2065,19 @@ class VPReductionPHIRecipe : public VPHeaderPHIRecipe,
 /// scalar value.
 class VPPartialReductionRecipe : public VPSingleDefRecipe {
   unsigned Opcode;
-  unsigned ScaleFactor;
+  unsigned VFScaleFactor;
 
 public:
   VPPartialReductionRecipe(Instruction *ReductionInst, VPValue *Op0,
-                           VPValue *Op1, unsigned ScaleFactor)
+                           VPValue *Op1, unsigned VFScaleFactor)
       : VPPartialReductionRecipe(ReductionInst->getOpcode(), Op0, Op1,
-                                 ScaleFactor, ReductionInst) {}
+                                 VFScaleFactor, ReductionInst) {}
   VPPartialReductionRecipe(unsigned Opcode, VPValue *Op0, VPValue *Op1,
-                           unsigned ScaleFactor,
+                           unsigned VFScaleFactor,
                            Instruction *ReductionInst = nullptr)
       : VPSingleDefRecipe(VPDef::VPPartialReductionSC,
                           ArrayRef<VPValue *>({Op0, Op1}), ReductionInst),
-        Opcode(Opcode), ScaleFactor(ScaleFactor) {
+        Opcode(Opcode), VFScaleFactor(VFScaleFactor) {
     [[maybe_unused]] auto *AccumulatorRecipe =
         getOperand(1)->getDefiningRecipe();
     assert((isa<VPReductionPHIRecipe>(AccumulatorRecipe) ||
@@ -2088,7 +2088,7 @@ class VPPartialReductionRecipe : public VPSingleDefRecipe {
 
   VPPartialReductionRecipe *clone() override {
     return new VPPartialReductionRecipe(Opcode, getOperand(0), getOperand(1),
-                                        ScaleFactor, getUnderlyingInstr());
+                                        VFScaleFactor, getUnderlyingInstr());
   }
 
   VP_CLASSOF_IMPL(VPDef::VPPartialReductionSC)
@@ -2103,7 +2103,7 @@ class VPPartialReductionRecipe : public VPSingleDefRecipe {
   /// Get the binary op's opcode.
   unsigned getOpcode() const { return Opcode; }
 
-  unsigned getScaleFactor() const { return ScaleFactor; }
+  unsigned getVFScaleFactor() const { return VFScaleFactor; }
 
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
   /// Print the recipe.

>From f5d5ed6f63cef1dec268ef19d5b553d6c91036eb Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Mon, 31 Mar 2025 15:06:54 +0100
Subject: [PATCH 03/19] Put if statement inside DEBUG

---
 llvm/lib/Transforms/Vectorize/LoopVectorize.cpp | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index ff96e98ef9cfa..05a8e999c6c2b 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5035,10 +5035,11 @@ calculateRegisterUsage(VPlan &Plan, ArrayRef<ElementCount> VFs,
           else if (auto *PartialReductionR =
                        dyn_cast<VPPartialReductionRecipe>(R))
             VF = VF.divideCoefficientBy(PartialReductionR->getVFScaleFactor());
-          if (VF != VFs[J])
-            LLVM_DEBUG(dbgs() << "LV(REG): Scaled down VF from " << VFs[J]
-                              << " to " << VF << " for ";
-                       R->dump(););
+          LLVM_DEBUG(if (VF != VFs[J]) {
+            dbgs() << "LV(REG): Scaled down VF from " << VFs[J] << " to " << VF
+                   << " for ";
+            R->dump();
+          });
 
           for (VPValue *DefV : R->definedValues()) {
             Type *ScalarTy = TypeInfo.inferScalarType(DefV);

>From 54af7d2f245d9f7323052d04526a4d3ce55b9da4 Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Wed, 2 Apr 2025 16:54:27 +0100
Subject: [PATCH 04/19] auto -> ElementCount

---
 llvm/lib/Transforms/Vectorize/LoopVectorize.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 05a8e999c6c2b..0c4cd232d91ef 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5029,7 +5029,7 @@ calculateRegisterUsage(VPlan &Plan, ArrayRef<ElementCount> VFs,
         } else {
           // The output from scaled phis and scaled reductions actually have
           // fewer lanes than the VF.
-          auto VF = VFs[J];
+          ElementCount VF = VFs[J];
           if (auto *ReductionR = dyn_cast<VPReductionPHIRecipe>(R))
             VF = VF.divideCoefficientBy(ReductionR->getVFScaleFactor());
           else if (auto *PartialReductionR =

>From b6b9063ba685ef270b3248637ac11a04cfbf9495 Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Wed, 2 Apr 2025 16:54:43 +0100
Subject: [PATCH 05/19] Add register pressure debug output checks

---
 .../AArch64/partial-reduce-dot-product-neon.ll             | 7 +++++++
 .../LoopVectorize/AArch64/partial-reduce-dot-product.ll    | 7 +++++++
 2 files changed, 14 insertions(+)

diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
index a55a5813044cb..c02e67f66a76a 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
@@ -2,6 +2,7 @@
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -mattr=+neon,+dotprod -force-vector-interleave=1 -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVE1
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -mattr=+neon,+dotprod -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVED
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -mattr=+neon,+dotprod -force-vector-interleave=1 -vectorizer-maximize-bandwidth -S < %s | FileCheck %s --check-prefixes=CHECK-MAXBW
+; RUN: opt -passes=loop-vectorize -debug-only=loop-vectorize --disable-output -S < %s 2>&1 | FileCheck %s --check-prefix=CHECK-REGS
 
 target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
 target triple = "aarch64-none-unknown-elf"
@@ -947,6 +948,12 @@ define i32 @dotp_unrolled(i32 %num_out, i64 %num_in, ptr %a, ptr %b) {
 ; CHECK-MAXBW-NEXT:    br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
 ; CHECK-MAXBW:       scalar.ph:
 ;
+; CHECK-REGS:      LV: Checking a loop in 'dotp_unrolled' from <stdin>
+; CHECK-REGS:      LV(REG): VF = 16
+; CHECK-REGS-NEXT: LV(REG): Found max usage: 2 item
+; CHECK-REGS-NEXT: LV(REG): RegisterClass: Generic::ScalarRC, 9 registers
+; CHECK-REGS-NEXT: LV(REG): RegisterClass: Generic::VectorRC, 24 registers
+; CHECK-REGS-NEXT: LV(REG): Found invariant usage: 1 item
 entry:
   br label %for.body
 
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
index 541cf034e0f3a..d17ac3dac8b9e 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
@@ -2,6 +2,7 @@
 ; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVE1
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVED
 ; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -vectorizer-maximize-bandwidth -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-MAXBW
+; RUN: opt -passes=loop-vectorize -debug-only=loop-vectorize --disable-output -S < %s 2>&1 | FileCheck %s --check-prefix=CHECK-REGS
 
 target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
 target triple = "aarch64-none-unknown-elf"
@@ -3420,6 +3421,12 @@ define dso_local void @not_dotp_high_register_pressure(ptr %a, ptr %b, ptr %sum,
 ; CHECK-MAXBW-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_FOR_COND_CLEANUP_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK-MAXBW:       scalar.ph:
 ;
+; CHECK-REGS-LABEL: LV: Checking a loop in 'dotp_high_register_pressure' from <stdin>
+; CHECK-REGS:       LV(REG): VF = 16
+; CHECK-REGS-NEXT:  LV(REG): Found max usage: 2 item
+; CHECK-REGS-NEXT:  LV(REG): RegisterClass: Generic::ScalarRC, 3 registers
+; CHECK-REGS-NEXT:  LV(REG): RegisterClass: Generic::VectorRC, 24 registers
+; CHECK-REGS-NEXT:  LV(REG): Found invariant usage: 1 item
 entry:
   %cmp100 = icmp sgt i32 %n, 0
   br i1 %cmp100, label %for.body.lr.ph, label %for.cond.cleanup

>From a6951248f86d30ac14ef0582e138ef69a73c51e3 Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Thu, 3 Apr 2025 17:35:00 +0100
Subject: [PATCH 06/19] Add REQUIRES: asserts

---
 .../LoopVectorize/AArch64/partial-reduce-dot-product.ll         | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
index d17ac3dac8b9e..2a9c488747244 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
@@ -1,4 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --filter-out-after "^scalar.ph:" --version 4
+; REQUIRES: asserts
+
 ; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVE1
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVED
 ; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -vectorizer-maximize-bandwidth -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-MAXBW

>From f05ebee46409a29beadb5543d4261db10844148f Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Mon, 7 Apr 2025 14:50:05 +0100
Subject: [PATCH 07/19] Update tests after rebasing on fhahn's changes

---
 .../AArch64/partial-reduce-dot-product-neon.ll |  2 +-
 .../AArch64/partial-reduce-dot-product.ll      |  6 +++---
 .../Transforms/LoopVectorize/X86/pr47437.ll    | 18 +++++++++---------
 3 files changed, 13 insertions(+), 13 deletions(-)

diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
index c02e67f66a76a..696d9eaf9a9cf 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
@@ -953,7 +953,7 @@ define i32 @dotp_unrolled(i32 %num_out, i64 %num_in, ptr %a, ptr %b) {
 ; CHECK-REGS-NEXT: LV(REG): Found max usage: 2 item
 ; CHECK-REGS-NEXT: LV(REG): RegisterClass: Generic::ScalarRC, 9 registers
 ; CHECK-REGS-NEXT: LV(REG): RegisterClass: Generic::VectorRC, 24 registers
-; CHECK-REGS-NEXT: LV(REG): Found invariant usage: 1 item
+; CHECK-REGS-NEXT: LV(REG): Found invariant usage: 0 item
 entry:
   br label %for.body
 
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
index 2a9c488747244..ccb539ad23456 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
@@ -3423,12 +3423,12 @@ define dso_local void @not_dotp_high_register_pressure(ptr %a, ptr %b, ptr %sum,
 ; CHECK-MAXBW-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_FOR_COND_CLEANUP_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK-MAXBW:       scalar.ph:
 ;
-; CHECK-REGS-LABEL: LV: Checking a loop in 'dotp_high_register_pressure' from <stdin>
+; CHECK-REGS-LABEL: LV: Checking a loop in 'not_dotp_high_register_pressure' from <stdin>
 ; CHECK-REGS:       LV(REG): VF = 16
 ; CHECK-REGS-NEXT:  LV(REG): Found max usage: 2 item
 ; CHECK-REGS-NEXT:  LV(REG): RegisterClass: Generic::ScalarRC, 3 registers
-; CHECK-REGS-NEXT:  LV(REG): RegisterClass: Generic::VectorRC, 24 registers
-; CHECK-REGS-NEXT:  LV(REG): Found invariant usage: 1 item
+; CHECK-REGS-NEXT:  LV(REG): RegisterClass: Generic::VectorRC, 40 registers
+; CHECK-REGS-NEXT:  LV(REG): Found invariant usage: 2 item
 entry:
   %cmp100 = icmp sgt i32 %n, 0
   br i1 %cmp100, label %for.body.lr.ph, label %for.cond.cleanup
diff --git a/llvm/test/Transforms/LoopVectorize/X86/pr47437.ll b/llvm/test/Transforms/LoopVectorize/X86/pr47437.ll
index c4291507e8d97..71e000a0272cc 100644
--- a/llvm/test/Transforms/LoopVectorize/X86/pr47437.ll
+++ b/llvm/test/Transforms/LoopVectorize/X86/pr47437.ll
@@ -212,16 +212,16 @@ define void @test_muladd(ptr noalias nocapture %d1, ptr noalias nocapture readon
 ; AVX1-NEXT:    [[TMP43:%.*]] = sext <4 x i16> [[STRIDED_VEC22]] to <4 x i32>
 ; AVX1-NEXT:    [[TMP46:%.*]] = mul nsw <4 x i32> [[TMP42]], [[TMP38]]
 ; AVX1-NEXT:    [[TMP47:%.*]] = mul nsw <4 x i32> [[TMP43]], [[TMP39]]
-; AVX1-NEXT:    [[TMP48:%.*]] = add nsw <4 x i32> [[TMP46]], [[TMP44]]
-; AVX1-NEXT:    [[TMP49:%.*]] = add nsw <4 x i32> [[TMP47]], [[TMP45]]
-; AVX1-NEXT:    [[TMP52:%.*]] = getelementptr inbounds i32, ptr [[D1:%.*]], i64 [[INDEX]]
-; AVX1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds i32, ptr [[TMP52]], i32 0
-; AVX1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds i32, ptr [[TMP52]], i32 4
-; AVX1-NEXT:    store <4 x i32> [[TMP48]], ptr [[TMP56]], align 4
-; AVX1-NEXT:    store <4 x i32> [[TMP49]], ptr [[TMP57]], align 4
+; AVX1-NEXT:    [[TMP19:%.*]] = add nsw <4 x i32> [[TMP46]], [[TMP44]]
+; AVX1-NEXT:    [[TMP20:%.*]] = add nsw <4 x i32> [[TMP47]], [[TMP45]]
+; AVX1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds i32, ptr [[D1:%.*]], i64 [[INDEX]]
+; AVX1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds i32, ptr [[TMP21]], i32 0
+; AVX1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds i32, ptr [[TMP21]], i32 4
+; AVX1-NEXT:    store <4 x i32> [[TMP19]], ptr [[TMP25]], align 4
+; AVX1-NEXT:    store <4 x i32> [[TMP20]], ptr [[TMP26]], align 4
 ; AVX1-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
-; AVX1-NEXT:    [[TMP60:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
-; AVX1-NEXT:    br i1 [[TMP60]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; AVX1-NEXT:    [[TMP24:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; AVX1-NEXT:    br i1 [[TMP24]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; AVX1:       middle.block:
 ; AVX1-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
 ; AVX1-NEXT:    br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[SCALAR_PH]]

>From ff432c92f832fe14d400049c20871fffac6df48c Mon Sep 17 00:00:00 2001
From: Sam Tebbs <samuel.tebbs at arm.com>
Date: Tue, 8 Apr 2025 10:16:11 +0100
Subject: [PATCH 08/19] Use one block for scaling factor consideration

---
 llvm/lib/Transforms/Vectorize/LoopVectorize.cpp | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 0c4cd232d91ef..e9f733eebe09d 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5030,11 +5030,15 @@ calculateRegisterUsage(VPlan &Plan, ArrayRef<ElementCount> VFs,
           // The output from scaled phis and scaled reductions actually have
           // fewer lanes than the VF.
           ElementCount VF = VFs[J];
-          if (auto *ReductionR = dyn_cast<VPReductionPHIRecipe>(R))
-            VF = VF.divideCoefficientBy(ReductionR->getVFScaleFactor());
-          else if (auto *PartialReductionR =
-                       dyn_cast<VPPartialReductionRecipe>(R))
-            VF = VF.divideCoefficientBy(PartialReductionR->getVFScaleFactor());
+          if (isa<VPPartialReductionRecipe, VPReductionPHIRecipe>(R)) {
+            auto *ReductionR = dyn_cast<VPReductionPHIRecipe>(R);
+            auto *PartialReductionR =
+                ReductionR ? nullptr : dyn_cast<VPPartialReductionRecipe>(R);
+            unsigned ScaleFactor = ReductionR
+                                       ? ReductionR->getVFScaleFactor()
+                                       : PartialReductionR->getVFScaleFactor();
+            VF = VF.divideCoefficientBy(ScaleFactor);
+          }
           LLVM_DEBUG(if (VF != VFs[J]) {
             dbgs() << "LV(REG): Scaled down VF from " << VFs[J] << " to " << VF
                    << " for ";

>From 0f68427099a951179a83c4b7f283fb32e9d8c3b8 Mon Sep 17 00:00:00 2001
From: Sam Tebbs <samuel.tebbs at arm.com>
Date: Tue, 8 Apr 2025 09:30:46 +0100
Subject: [PATCH 09/19] Add comments for getVFScaleFactor

---
 llvm/lib/Transforms/Vectorize/VPlan.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 6928bea4bb092..fd5a66d73e288 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -2033,6 +2033,7 @@ class VPReductionPHIRecipe : public VPHeaderPHIRecipe,
   /// Generate the phi/select nodes.
   void execute(VPTransformState &State) override;
 
+  /// Get the factor that the VF of this recipe's output should be scaled by
   unsigned getVFScaleFactor() const { return VFScaleFactor; }
 
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
@@ -2103,6 +2104,7 @@ class VPPartialReductionRecipe : public VPSingleDefRecipe {
   /// Get the binary op's opcode.
   unsigned getOpcode() const { return Opcode; }
 
+  /// Get the factor that the VF of this recipe's output should be scaled by
   unsigned getVFScaleFactor() const { return VFScaleFactor; }
 
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)

>From 84273625c2287355971ddbda0e48095a512fe81d Mon Sep 17 00:00:00 2001
From: Sam Tebbs <samuel.tebbs at arm.com>
Date: Tue, 8 Apr 2025 09:27:40 +0100
Subject: [PATCH 10/19] Add missing asserts requirement

---
 .../LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll    | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
index 696d9eaf9a9cf..70bc05922306f 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
@@ -1,4 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --filter-out-after "^scalar.ph:" --version 4
+; REQUIRES: asserts
+
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -mattr=+neon,+dotprod -force-vector-interleave=1 -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVE1
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -mattr=+neon,+dotprod -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVED
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -mattr=+neon,+dotprod -force-vector-interleave=1 -vectorizer-maximize-bandwidth -S < %s | FileCheck %s --check-prefixes=CHECK-MAXBW

>From b905806ab8616ca300564a8e05bb1ae27a5a7b1c Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Wed, 9 Apr 2025 15:13:23 +0100
Subject: [PATCH 11/19] Add << overload for VPDef and getVFScaleFactor utility
 function

---
 .../Transforms/Vectorize/LoopVectorize.cpp    | 31 +++++++++++--------
 llvm/lib/Transforms/Vectorize/VPlan.cpp       |  9 ++++++
 llvm/lib/Transforms/Vectorize/VPlanSLP.cpp    |  4 +--
 llvm/lib/Transforms/Vectorize/VPlanValue.h    |  1 +
 .../Transforms/Vectorize/VPlanTest.cpp        |  2 +-
 5 files changed, 31 insertions(+), 16 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index e9f733eebe09d..51d018e444161 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -4874,6 +4874,20 @@ void LoopVectorizationCostModel::collectElementTypesForWidening() {
   }
 }
 
+/// Get the VF scaling factor applied to the recipe's output, if the recipe has
+/// one.
+static unsigned getVFScaleFactor(VPRecipeBase *R) {
+  if (isa<VPPartialReductionRecipe, VPReductionPHIRecipe>(R)) {
+    auto *ReductionR = dyn_cast<VPReductionPHIRecipe>(R);
+    auto *PartialReductionR =
+        ReductionR ? nullptr : dyn_cast<VPPartialReductionRecipe>(R);
+    unsigned ScaleFactor = ReductionR ? ReductionR->getVFScaleFactor()
+                                      : PartialReductionR->getVFScaleFactor();
+    return ScaleFactor;
+  }
+  return 1;
+}
+
 /// Estimate the register usage for \p Plan and vectorization factors in \p VFs
 /// by calculating the highest number of values that are live at a single
 /// location as a rough estimate. Returns the register usage for each VF in \p
@@ -5027,22 +5041,13 @@ calculateRegisterUsage(VPlan &Plan, ArrayRef<ElementCount> VFs,
           // even in the scalar case.
           RegUsage[ClassID] += 1;
         } else {
-          // The output from scaled phis and scaled reductions actually have
+          // The output from scaled phis and scaled reductions actually has
           // fewer lanes than the VF.
-          ElementCount VF = VFs[J];
-          if (isa<VPPartialReductionRecipe, VPReductionPHIRecipe>(R)) {
-            auto *ReductionR = dyn_cast<VPReductionPHIRecipe>(R);
-            auto *PartialReductionR =
-                ReductionR ? nullptr : dyn_cast<VPPartialReductionRecipe>(R);
-            unsigned ScaleFactor = ReductionR
-                                       ? ReductionR->getVFScaleFactor()
-                                       : PartialReductionR->getVFScaleFactor();
-            VF = VF.divideCoefficientBy(ScaleFactor);
-          }
+          unsigned ScaleFactor = getVFScaleFactor(R);
+          ElementCount VF = VFs[J].divideCoefficientBy(ScaleFactor);
           LLVM_DEBUG(if (VF != VFs[J]) {
             dbgs() << "LV(REG): Scaled down VF from " << VFs[J] << " to " << VF
-                   << " for ";
-            R->dump();
+                   << " for " << *R << "\n";
           });
 
           for (VPValue *DefV : R->definedValues()) {
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.cpp b/llvm/lib/Transforms/Vectorize/VPlan.cpp
index 9474e7a171dff..26895b20ec708 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlan.cpp
@@ -71,6 +71,15 @@ raw_ostream &llvm::operator<<(raw_ostream &OS, const VPValue &V) {
   V.print(OS, SlotTracker);
   return OS;
 }
+raw_ostream &llvm::operator<<(raw_ostream &OS, const VPDef &D) {
+  /// If this def has a single value, we can cast it to an instruction and use
+  /// its plan for the slot tracker.
+  if (const VPValue *Val = D.getVPSingleValue())
+    return OS << *Val;
+  VPSlotTracker SlotTracker(nullptr);
+  D.print(OS, "", SlotTracker);
+  return OS;
+}
 #endif
 
 Value *VPLane::getAsRuntimeExpr(IRBuilderBase &Builder,
diff --git a/llvm/lib/Transforms/Vectorize/VPlanSLP.cpp b/llvm/lib/Transforms/Vectorize/VPlanSLP.cpp
index e943c7a29eb83..85e13a499471c 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanSLP.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanSLP.cpp
@@ -516,8 +516,8 @@ VPInstruction *VPlanSlp::buildGraph(ArrayRef<VPValue *> Values) {
   auto *Inst = cast<VPInstruction>(Values[0])->getUnderlyingInstr();
   auto *VPI = new VPInstruction(Opcode, CombinedOperands, Inst->getDebugLoc());
 
-  LLVM_DEBUG(dbgs() << "Create VPInstruction " << *VPI << " "
-                    << *cast<VPInstruction>(Values[0]) << "\n");
+  LLVM_DEBUG(dbgs() << "Create VPInstruction " << cast<VPValue>(*VPI) << " "
+                    << Values[0] << "\n");
   addCombined(Values, VPI);
   return VPI;
 }
diff --git a/llvm/lib/Transforms/Vectorize/VPlanValue.h b/llvm/lib/Transforms/Vectorize/VPlanValue.h
index d322fdfa727e4..6f0de80578913 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanValue.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanValue.h
@@ -195,6 +195,7 @@ typedef DenseMap<Value *, VPValue *> Value2VPValueTy;
 typedef DenseMap<VPValue *, Value *> VPValue2ValueTy;
 
 raw_ostream &operator<<(raw_ostream &OS, const VPValue &V);
+raw_ostream &operator<<(raw_ostream &OS, const VPDef &D);
 
 /// This class augments VPValue with operands which provide the inverse def-use
 /// edges from VPValue's users to their defs.
diff --git a/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp b/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp
index cb7545171744e..a2c66f939adb0 100644
--- a/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp
+++ b/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp
@@ -812,7 +812,7 @@ Successor(s): ir-bb<scalar.header>
   {
     std::string I4Dump;
     raw_string_ostream OS(I4Dump);
-    OS << *I4;
+    OS << *cast<VPDef>(I4);
     EXPECT_EQ("EMIT vp<%5> = mul vp<%3>, vp<%2>", I4Dump);
   }
 }

>From 5f14165168e2056ef40c99cd32a89f5a27aceda4 Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Thu, 10 Apr 2025 10:15:33 +0100
Subject: [PATCH 12/19] Simplify utility function

---
 llvm/lib/Transforms/Vectorize/LoopVectorize.cpp | 12 ++++--------
 1 file changed, 4 insertions(+), 8 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 51d018e444161..b595908abb325 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -4877,14 +4877,10 @@ void LoopVectorizationCostModel::collectElementTypesForWidening() {
 /// Get the VF scaling factor applied to the recipe's output, if the recipe has
 /// one.
 static unsigned getVFScaleFactor(VPRecipeBase *R) {
-  if (isa<VPPartialReductionRecipe, VPReductionPHIRecipe>(R)) {
-    auto *ReductionR = dyn_cast<VPReductionPHIRecipe>(R);
-    auto *PartialReductionR =
-        ReductionR ? nullptr : dyn_cast<VPPartialReductionRecipe>(R);
-    unsigned ScaleFactor = ReductionR ? ReductionR->getVFScaleFactor()
-                                      : PartialReductionR->getVFScaleFactor();
-    return ScaleFactor;
-  }
+  if (auto *RR = dyn_cast<VPReductionPHIRecipe>(R))
+    return RR->getVFScaleFactor();
+  if (auto *RR = dyn_cast<VPPartialReductionRecipe>(R))
+    return RR->getVFScaleFactor();
   return 1;
 }
 

>From 296e3ce09da48bd64b79746e76db0dfd11fbb44c Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Thu, 10 Apr 2025 10:58:28 +0100
Subject: [PATCH 13/19] Fix test by printing VPRecipeBase

---
 llvm/lib/Transforms/Vectorize/VPlan.cpp           | 10 +++-------
 llvm/lib/Transforms/Vectorize/VPlanValue.h        |  2 +-
 llvm/unittests/Transforms/Vectorize/VPlanTest.cpp |  2 +-
 3 files changed, 5 insertions(+), 9 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/VPlan.cpp b/llvm/lib/Transforms/Vectorize/VPlan.cpp
index 26895b20ec708..34aa3ff1a744a 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlan.cpp
@@ -71,13 +71,9 @@ raw_ostream &llvm::operator<<(raw_ostream &OS, const VPValue &V) {
   V.print(OS, SlotTracker);
   return OS;
 }
-raw_ostream &llvm::operator<<(raw_ostream &OS, const VPDef &D) {
-  /// If this def has a single value, we can cast it to an instruction and use
-  /// its plan for the slot tracker.
-  if (const VPValue *Val = D.getVPSingleValue())
-    return OS << *Val;
-  VPSlotTracker SlotTracker(nullptr);
-  D.print(OS, "", SlotTracker);
+raw_ostream &llvm::operator<<(raw_ostream &OS, const VPRecipeBase &R) {
+  VPSlotTracker SlotTracker(R.getParent()->getPlan());
+  R.print(OS, "", SlotTracker);
   return OS;
 }
 #endif
diff --git a/llvm/lib/Transforms/Vectorize/VPlanValue.h b/llvm/lib/Transforms/Vectorize/VPlanValue.h
index 6f0de80578913..33a049cc74aeb 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanValue.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanValue.h
@@ -195,7 +195,7 @@ typedef DenseMap<Value *, VPValue *> Value2VPValueTy;
 typedef DenseMap<VPValue *, Value *> VPValue2ValueTy;
 
 raw_ostream &operator<<(raw_ostream &OS, const VPValue &V);
-raw_ostream &operator<<(raw_ostream &OS, const VPDef &D);
+raw_ostream &operator<<(raw_ostream &OS, const VPRecipeBase &R);
 
 /// This class augments VPValue with operands which provide the inverse def-use
 /// edges from VPValue's users to their defs.
diff --git a/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp b/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp
index a2c66f939adb0..3c81376eb83e0 100644
--- a/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp
+++ b/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp
@@ -812,7 +812,7 @@ Successor(s): ir-bb<scalar.header>
   {
     std::string I4Dump;
     raw_string_ostream OS(I4Dump);
-    OS << *cast<VPDef>(I4);
+    OS << *cast<VPValue>(I4);
     EXPECT_EQ("EMIT vp<%5> = mul vp<%3>, vp<%2>", I4Dump);
   }
 }

>From c38fcabf65fb8c5d96fd9adc2000bdea99cb2662 Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Thu, 10 Apr 2025 15:19:41 +0100
Subject: [PATCH 14/19] Re-add new line

---
 llvm/lib/Transforms/Vectorize/LoopVectorize.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index b595908abb325..e07caeb0a2e1f 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5024,6 +5024,7 @@ calculateRegisterUsage(VPlan &Plan, ArrayRef<ElementCount> VFs,
         if (isa<VPVectorPointerRecipe, VPVectorEndPointerRecipe,
                 VPBranchOnMaskRecipe>(R))
           continue;
+
         if (VFs[J].isScalar() ||
             isa<VPCanonicalIVPHIRecipe, VPReplicateRecipe, VPDerivedIVRecipe,
                 VPScalarIVStepsRecipe>(R) ||

>From e6061f2c0b416f294f070f2d90d577d9ce1fb4fe Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Thu, 10 Apr 2025 15:19:56 +0100
Subject: [PATCH 15/19] auto -> std::optional<unsigned>

---
 llvm/lib/Transforms/Vectorize/LoopVectorize.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index e07caeb0a2e1f..8648401f4b23b 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -9156,7 +9156,7 @@ VPRecipeBase *VPRecipeBuilder::tryToCreateWidenRecipe(
   if (isa<LoadInst>(Instr) || isa<StoreInst>(Instr))
     return tryToWidenMemory(Instr, Operands, Range);
 
-  if (auto ScaleFactor = getScalingForReduction(Instr))
+  if (std::optional<unsigned> ScaleFactor = getScalingForReduction(Instr))
     return tryToCreatePartialReduction(Instr, Operands, ScaleFactor.value());
 
   if (!shouldWiden(Instr, Range))

>From f1e7e9be3e4b2d13dec7de512a4f669cd4eba534 Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Thu, 10 Apr 2025 15:20:15 +0100
Subject: [PATCH 16/19] Replace VPValue printer with VPRecipeBase

---
 llvm/lib/Transforms/Vectorize/VPlan.cpp           | 10 ++--------
 llvm/lib/Transforms/Vectorize/VPlanSLP.cpp        |  4 ++--
 llvm/lib/Transforms/Vectorize/VPlanValue.h        |  1 -
 llvm/unittests/Transforms/Vectorize/VPlanTest.cpp |  2 +-
 4 files changed, 5 insertions(+), 12 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/VPlan.cpp b/llvm/lib/Transforms/Vectorize/VPlan.cpp
index 34aa3ff1a744a..eb4394171d3be 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlan.cpp
@@ -64,15 +64,9 @@ static cl::opt<bool> PrintVPlansInDotFormat(
 #define DEBUG_TYPE "loop-vectorize"
 
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
-raw_ostream &llvm::operator<<(raw_ostream &OS, const VPValue &V) {
-  const VPInstruction *Instr = dyn_cast<VPInstruction>(&V);
-  VPSlotTracker SlotTracker(
-      (Instr && Instr->getParent()) ? Instr->getParent()->getPlan() : nullptr);
-  V.print(OS, SlotTracker);
-  return OS;
-}
 raw_ostream &llvm::operator<<(raw_ostream &OS, const VPRecipeBase &R) {
-  VPSlotTracker SlotTracker(R.getParent()->getPlan());
+  const VPBasicBlock *Parent = R.getParent();
+  VPSlotTracker SlotTracker(Parent ? Parent->getPlan() : nullptr);
   R.print(OS, "", SlotTracker);
   return OS;
 }
diff --git a/llvm/lib/Transforms/Vectorize/VPlanSLP.cpp b/llvm/lib/Transforms/Vectorize/VPlanSLP.cpp
index 85e13a499471c..c4d79526721ec 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanSLP.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanSLP.cpp
@@ -516,8 +516,8 @@ VPInstruction *VPlanSlp::buildGraph(ArrayRef<VPValue *> Values) {
   auto *Inst = cast<VPInstruction>(Values[0])->getUnderlyingInstr();
   auto *VPI = new VPInstruction(Opcode, CombinedOperands, Inst->getDebugLoc());
 
-  LLVM_DEBUG(dbgs() << "Create VPInstruction " << cast<VPValue>(*VPI) << " "
-                    << Values[0] << "\n");
+  LLVM_DEBUG(dbgs() << "Create VPInstruction " << *VPI << " " << Values[0]
+                    << "\n");
   addCombined(Values, VPI);
   return VPI;
 }
diff --git a/llvm/lib/Transforms/Vectorize/VPlanValue.h b/llvm/lib/Transforms/Vectorize/VPlanValue.h
index 33a049cc74aeb..70ce58900555c 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanValue.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanValue.h
@@ -194,7 +194,6 @@ class VPValue {
 typedef DenseMap<Value *, VPValue *> Value2VPValueTy;
 typedef DenseMap<VPValue *, Value *> VPValue2ValueTy;
 
-raw_ostream &operator<<(raw_ostream &OS, const VPValue &V);
 raw_ostream &operator<<(raw_ostream &OS, const VPRecipeBase &R);
 
 /// This class augments VPValue with operands which provide the inverse def-use
diff --git a/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp b/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp
index 3c81376eb83e0..cb7545171744e 100644
--- a/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp
+++ b/llvm/unittests/Transforms/Vectorize/VPlanTest.cpp
@@ -812,7 +812,7 @@ Successor(s): ir-bb<scalar.header>
   {
     std::string I4Dump;
     raw_string_ostream OS(I4Dump);
-    OS << *cast<VPValue>(I4);
+    OS << *I4;
     EXPECT_EQ("EMIT vp<%5> = mul vp<%3>, vp<%2>", I4Dump);
   }
 }

>From bde39b46fa02c362b2e01902952e355c5fd6baf3 Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Thu, 10 Apr 2025 15:20:48 +0100
Subject: [PATCH 17/19] Add VFScaleFactor comment

---
 llvm/lib/Transforms/Vectorize/VPlan.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index fd5a66d73e288..89a0c75b39c3d 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -2066,6 +2066,8 @@ class VPReductionPHIRecipe : public VPHeaderPHIRecipe,
 /// scalar value.
 class VPPartialReductionRecipe : public VPSingleDefRecipe {
   unsigned Opcode;
+  /// The divisor by which the VF of this recipe's output should be divided
+  /// during execution.
   unsigned VFScaleFactor;
 
 public:

>From f58b5e1a175b28208839e4e2fa0d3b7c3f58b4e9 Mon Sep 17 00:00:00 2001
From: Samuel Tebbs <samuel.tebbs at arm.com>
Date: Thu, 10 Apr 2025 15:22:04 +0100
Subject: [PATCH 18/19] Move reg usage tests to reg-usage.ll

---
 .../partial-reduce-dot-product-neon.ll        |  10 -
 .../AArch64/partial-reduce-dot-product.ll     |  10 -
 .../LoopVectorize/AArch64/reg-usage.ll        | 180 ++++++++++++++++++
 3 files changed, 180 insertions(+), 20 deletions(-)

diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
index 70bc05922306f..e6687fe767c0a 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product-neon.ll
@@ -1,10 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --filter-out-after "^scalar.ph:" --version 4
-; REQUIRES: asserts
-
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -mattr=+neon,+dotprod -force-vector-interleave=1 -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVE1
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -mattr=+neon,+dotprod -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVED
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -mattr=+neon,+dotprod -force-vector-interleave=1 -vectorizer-maximize-bandwidth -S < %s | FileCheck %s --check-prefixes=CHECK-MAXBW
-; RUN: opt -passes=loop-vectorize -debug-only=loop-vectorize --disable-output -S < %s 2>&1 | FileCheck %s --check-prefix=CHECK-REGS
 
 target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
 target triple = "aarch64-none-unknown-elf"
@@ -949,13 +946,6 @@ define i32 @dotp_unrolled(i32 %num_out, i64 %num_in, ptr %a, ptr %b) {
 ; CHECK-MAXBW-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[NUM_IN]], [[N_VEC]]
 ; CHECK-MAXBW-NEXT:    br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
 ; CHECK-MAXBW:       scalar.ph:
-;
-; CHECK-REGS:      LV: Checking a loop in 'dotp_unrolled' from <stdin>
-; CHECK-REGS:      LV(REG): VF = 16
-; CHECK-REGS-NEXT: LV(REG): Found max usage: 2 item
-; CHECK-REGS-NEXT: LV(REG): RegisterClass: Generic::ScalarRC, 9 registers
-; CHECK-REGS-NEXT: LV(REG): RegisterClass: Generic::VectorRC, 24 registers
-; CHECK-REGS-NEXT: LV(REG): Found invariant usage: 0 item
 entry:
   br label %for.body
 
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
index ccb539ad23456..1b22523e9f5bd 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-dot-product.ll
@@ -1,10 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --filter-out-after "^scalar.ph:" --version 4
-; REQUIRES: asserts
-
 ; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVE1
 ; RUN: opt -passes=loop-vectorize -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-INTERLEAVED
 ; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -vectorizer-maximize-bandwidth -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-MAXBW
-; RUN: opt -passes=loop-vectorize -debug-only=loop-vectorize --disable-output -S < %s 2>&1 | FileCheck %s --check-prefix=CHECK-REGS
 
 target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
 target triple = "aarch64-none-unknown-elf"
@@ -3422,13 +3419,6 @@ define dso_local void @not_dotp_high_register_pressure(ptr %a, ptr %b, ptr %sum,
 ; CHECK-MAXBW-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
 ; CHECK-MAXBW-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_FOR_COND_CLEANUP_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
 ; CHECK-MAXBW:       scalar.ph:
-;
-; CHECK-REGS-LABEL: LV: Checking a loop in 'not_dotp_high_register_pressure' from <stdin>
-; CHECK-REGS:       LV(REG): VF = 16
-; CHECK-REGS-NEXT:  LV(REG): Found max usage: 2 item
-; CHECK-REGS-NEXT:  LV(REG): RegisterClass: Generic::ScalarRC, 3 registers
-; CHECK-REGS-NEXT:  LV(REG): RegisterClass: Generic::VectorRC, 40 registers
-; CHECK-REGS-NEXT:  LV(REG): Found invariant usage: 2 item
 entry:
   %cmp100 = icmp sgt i32 %n, 0
   br i1 %cmp100, label %for.body.lr.ph, label %for.cond.cleanup
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/reg-usage.ll b/llvm/test/Transforms/LoopVectorize/AArch64/reg-usage.ll
index ab12ca4254a09..2e6efec297d61 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/reg-usage.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/reg-usage.ll
@@ -68,3 +68,183 @@ loop:
 exit:
   ret void
 }
+
+define dso_local void @dotp_high_register_pressure(ptr %a, ptr %b, ptr %sum, i32 %n) #1 {
+; CHECK-LABEL: LV: Checking a loop in 'dotp_high_register_pressure' from <stdin>
+; CHECK:       LV(REG): VF = 16
+; CHECK-NEXT:  LV(REG): Found max usage: 2 item
+; CHECK-NEXT:  LV(REG): RegisterClass: Generic::ScalarRC, 3 registers
+; CHECK-NEXT:  LV(REG): RegisterClass: Generic::VectorRC, 40 registers
+; CHECK-NEXT:  LV(REG): Found invariant usage: 2 item
+entry:
+  %cmp100 = icmp sgt i32 %n, 0
+  br i1 %cmp100, label %for.body.lr.ph, label %for.cond.cleanup
+
+for.body.lr.ph:                                   ; preds = %entry
+  %arrayidx13 = getelementptr inbounds nuw i8, ptr %sum, i64 4
+  %gep.b.12 = getelementptr inbounds nuw i8, ptr %sum, i64 8
+  %arrayidx31 = getelementptr inbounds nuw i8, ptr %sum, i64 12
+  %arrayidx40 = getelementptr inbounds nuw i8, ptr %sum, i64 16
+  %arrayidx49 = getelementptr inbounds nuw i8, ptr %sum, i64 20
+  %arrayidx58 = getelementptr inbounds nuw i8, ptr %sum, i64 24
+  %arrayidx67 = getelementptr inbounds nuw i8, ptr %sum, i64 28
+  %sum.promoted = load i32, ptr %sum, align 4
+  %arrayidx13.promoted = load i32, ptr %arrayidx13, align 4
+  %gep.b.12.promoted = load i32, ptr %gep.b.12, align 4
+  %arrayidx31.promoted = load i32, ptr %arrayidx31, align 4
+  %arrayidx40.promoted = load i32, ptr %arrayidx40, align 4
+  %arrayidx49.promoted = load i32, ptr %arrayidx49, align 4
+  %arrayidx58.promoted = load i32, ptr %arrayidx58, align 4
+  %arrayidx67.promoted = load i32, ptr %arrayidx67, align 4
+  %wide.trip.count = zext nneg i32 %n to i64
+  br label %for.body
+
+for.cond.for.cond.cleanup_crit_edge:              ; preds = %for.body
+  %add.lcssa = phi i32 [ %add.1, %for.body ]
+  %add.2.lcssa = phi i32 [ %add.2, %for.body ]
+  %add.3.lcssa = phi i32 [ %add.3, %for.body ]
+  %add.4.lcssa = phi i32 [ %add.4, %for.body ]
+  %add.5.lcssa = phi i32 [ %add.5, %for.body ]
+  %add.6.lcssa = phi i32 [ %add.6, %for.body ]
+  %add.7.lcssa = phi i32 [ %add.7, %for.body ]
+  %add.8.lcssa = phi i32 [ %add.8, %for.body ]
+  store i32 %add.lcssa, ptr %sum, align 4
+  store i32 %add.2.lcssa, ptr %arrayidx13, align 4
+  store i32 %add.3.lcssa, ptr %gep.b.12, align 4
+  store i32 %add.4.lcssa, ptr %arrayidx31, align 4
+  store i32 %add.5.lcssa, ptr %arrayidx40, align 4
+  store i32 %add.6.lcssa, ptr %arrayidx49, align 4
+  store i32 %add.7.lcssa, ptr %arrayidx58, align 4
+  store i32 %add.8.lcssa, ptr %arrayidx67, align 4
+  br label %for.cond.cleanup
+
+for.cond.cleanup:                                 ; preds = %for.cond.for.cond.cleanup_crit_edge, %entry
+  ret void
+
+for.body:                                         ; preds = %for.body.lr.ph, %for.body
+  %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
+  %0 = phi i32 [ %arrayidx67.promoted, %for.body.lr.ph ], [ %add.8, %for.body ]
+  %1 = phi i32 [ %arrayidx58.promoted, %for.body.lr.ph ], [ %add.7, %for.body ]
+  %2 = phi i32 [ %arrayidx49.promoted, %for.body.lr.ph ], [ %add.6, %for.body ]
+  %3 = phi i32 [ %arrayidx40.promoted, %for.body.lr.ph ], [ %add.5, %for.body ]
+  %4 = phi i32 [ %arrayidx31.promoted, %for.body.lr.ph ], [ %add.4, %for.body ]
+  %5 = phi i32 [ %gep.b.12.promoted, %for.body.lr.ph ], [ %add.3, %for.body ]
+  %6 = phi i32 [ %arrayidx13.promoted, %for.body.lr.ph ], [ %add.2, %for.body ]
+  %7 = phi i32 [ %sum.promoted, %for.body.lr.ph ], [ %add.1, %for.body ]
+  %arrayidx = getelementptr inbounds nuw i8, ptr %a, i64 %indvars.iv
+  %load.a = load i8, ptr %arrayidx, align 1
+  %ext.a = zext i8 %load.a to i32
+  %9 = shl nsw i64 %indvars.iv, 3
+  %gep.b.1 = getelementptr inbounds nuw i8, ptr %b, i64 %9
+  %load.b.1 = load i8, ptr %gep.b.1, align 1
+  %ext.b.1 = sext i8 %load.b.1 to i32
+  %mul.1 = mul nsw i32 %ext.b.1, %ext.a
+  %add.1 = add nsw i32 %mul.1, %7
+  %11 = or disjoint i64 %9, 1
+  %gep.b.2 = getelementptr inbounds nuw i8, ptr %b, i64 %11
+  %load.b.2 = load i8, ptr %gep.b.2, align 1
+  %ext.b.2 = sext i8 %load.b.2 to i32
+  %mul.2 = mul nsw i32 %ext.b.2, %ext.a
+  %add.2 = add nsw i32 %mul.2, %6
+  %13 = or disjoint i64 %9, 2
+  %gep.b.3 = getelementptr inbounds nuw i8, ptr %b, i64 %13
+  %load.b.3 = load i8, ptr %gep.b.3, align 1
+  %ext.b.3 = sext i8 %load.b.3 to i32
+  %mul.3 = mul nsw i32 %ext.b.3, %ext.a
+  %add.3 = add nsw i32 %mul.3, %5
+  %15 = or disjoint i64 %9, 3
+  %gep.b.4 = getelementptr inbounds nuw i8, ptr %b, i64 %15
+  %load.b.4 = load i8, ptr %gep.b.4, align 1
+  %ext.b.4 = sext i8 %load.b.4 to i32
+  %mul.4 = mul nsw i32 %ext.b.4, %ext.a
+  %add.4 = add nsw i32 %mul.4, %4
+  %17 = or disjoint i64 %9, 4
+  %gep.b.5 = getelementptr inbounds nuw i8, ptr %b, i64 %17
+  %load.b.5 = load i8, ptr %gep.b.5, align 1
+  %ext.b.5 = sext i8 %load.b.5 to i32
+  %mul.5 = mul nsw i32 %ext.b.5, %ext.a
+  %add.5 = add nsw i32 %mul.5, %3
+  %19 = or disjoint i64 %9, 5
+  %gep.b.6 = getelementptr inbounds nuw i8, ptr %b, i64 %19
+  %load.b.6 = load i8, ptr %gep.b.6, align 1
+  %ext.b.6 = sext i8 %load.b.6 to i32
+  %mul.6 = mul nsw i32 %ext.b.6, %ext.a
+  %add.6 = add nsw i32 %mul.6, %2
+  %21 = or disjoint i64 %9, 6
+  %gep.b.7 = getelementptr inbounds nuw i8, ptr %b, i64 %21
+  %load.b.7 = load i8, ptr %gep.b.7, align 1
+  %ext.b.7 = sext i8 %load.b.7 to i32
+  %mul.7 = mul nsw i32 %ext.b.7, %ext.a
+  %add.7 = add nsw i32 %mul.7, %1
+  %23 = or disjoint i64 %9, 7
+  %gep.b.8 = getelementptr inbounds nuw i8, ptr %b, i64 %23
+  %load.b.8 = load i8, ptr %gep.b.8, align 1
+  %ext.b.8 = sext i8 %load.b.8 to i32
+  %mul.8 = mul nsw i32 %ext.b.8, %ext.a
+  %add.8 = add nsw i32 %mul.8, %0
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond.not = icmp eq i64 %indvars.iv.next, %wide.trip.count
+  br i1 %exitcond.not, label %for.cond.for.cond.cleanup_crit_edge, label %for.body
+}
+
+define i32 @dotp_unrolled(i32 %num_out, i64 %num_in, ptr %a, ptr %b) {
+; CHECK-LABEL: LV: Checking a loop in 'dotp_unrolled' from <stdin>
+; CHECK:       LV(REG): VF = 16
+; CHECK-NEXT:  LV(REG): Found max usage: 2 item
+; CHECK-NEXT:  LV(REG): RegisterClass: Generic::ScalarRC, 9 registers
+; CHECK-NEXT:  LV(REG): RegisterClass: Generic::VectorRC, 24 registers
+; CHECK-NEXT:  LV(REG): Found invariant usage: 0 item
+entry:
+  br label %for.body
+
+for.body:                                    ; preds = %entry, %for.body
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
+  %accum3 = phi i32 [ 0, %entry ], [ %add.a3, %for.body ]
+  %accum2 = phi i32 [ 0, %entry ], [ %add.a2, %for.body ]
+  %accum1 = phi i32 [ 0, %entry ], [ %add.a1, %for.body ]
+  %accum0 = phi i32 [ 0, %entry ], [ %add.a0, %for.body ]
+  %gep.a0 = getelementptr inbounds i8, ptr %a, i64 %iv
+  %gep.b0 = getelementptr inbounds i8, ptr %b, i64 %iv
+  %offset.1 = or disjoint i64 %iv, 1
+  %gep.a1 = getelementptr inbounds i8, ptr %a, i64 %offset.1
+  %gep.b1 = getelementptr inbounds i8, ptr %b, i64 %offset.1
+  %offset.2 = or disjoint i64 %iv, 2
+  %gep.a2 = getelementptr inbounds i8, ptr %a, i64 %offset.2
+  %gep.b2 = getelementptr inbounds i8, ptr %b, i64 %offset.2
+  %offset.3 = or disjoint i64 %iv, 3
+  %gep.a3 = getelementptr inbounds i8, ptr %a, i64 %offset.3
+  %gep.b3 = getelementptr inbounds i8, ptr %b, i64 %offset.3
+  %load.a0 = load i8, ptr %gep.a0, align 1
+  %ext.a0 = sext i8 %load.a0 to i32
+  %load.b0 = load i8, ptr %gep.b0, align 1
+  %ext.b0 = sext i8 %load.b0 to i32
+  %mul.a0 = mul nsw i32 %ext.b0, %ext.a0
+  %add.a0 = add nsw i32 %mul.a0, %accum0
+  %load.a1 = load i8, ptr %gep.a1, align 1
+  %ext.a1 = sext i8 %load.a1 to i32
+  %load.b1 = load i8, ptr %gep.b1, align 1
+  %ext.b1 = sext i8 %load.b1 to i32
+  %mul.a1 = mul nsw i32 %ext.a1, %ext.b1
+  %add.a1 = add nsw i32 %mul.a1, %accum1
+  %load.a2 = load i8, ptr %gep.a2, align 1
+  %ext.a2 = sext i8 %load.a2 to i32
+  %load.b2 = load i8, ptr %gep.b2, align 1
+  %ext.b2 = sext i8 %load.b2 to i32
+  %mul.a2 = mul nsw i32 %ext.a2, %ext.b2
+  %add.a2 = add nsw i32 %mul.a2, %accum2
+  %load.a3 = load i8, ptr %gep.a3, align 1
+  %ext.a3 = sext i8 %load.a3 to i32
+  %load.b3 = load i8, ptr %gep.b3, align 1
+  %ext.b3 = sext i8 %load.b3 to i32
+  %mul.a3 = mul nsw i32 %ext.a3, %ext.b3
+  %add.a3 = add nsw i32 %mul.a3, %accum3
+  %iv.next = add nuw nsw i64 %iv, 1
+  %exitcond.not = icmp eq i64 %iv.next, %num_in
+  br i1 %exitcond.not, label %exit, label %for.body
+
+exit:                                        ; preds = %for.body
+  %result0 = add nsw i32 %add.a0, %add.a1
+  %result1 = add nsw i32 %add.a2, %add.a3
+  %result = add nsw i32 %result0, %result1
+  ret i32 %result
+}

>From aefea413d5dadeb3ea95034009d14ee484cb908b Mon Sep 17 00:00:00 2001
From: Sam Tebbs <samuel.tebbs at arm.com>
Date: Fri, 11 Apr 2025 09:45:18 +0100
Subject: [PATCH 19/19] Add full stops

---
 llvm/lib/Transforms/Vectorize/VPlan.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 89a0c75b39c3d..efb0789bc7041 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -2033,7 +2033,7 @@ class VPReductionPHIRecipe : public VPHeaderPHIRecipe,
   /// Generate the phi/select nodes.
   void execute(VPTransformState &State) override;
 
-  /// Get the factor that the VF of this recipe's output should be scaled by
+  /// Get the factor that the VF of this recipe's output should be scaled by.
   unsigned getVFScaleFactor() const { return VFScaleFactor; }
 
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
@@ -2106,7 +2106,7 @@ class VPPartialReductionRecipe : public VPSingleDefRecipe {
   /// Get the binary op's opcode.
   unsigned getOpcode() const { return Opcode; }
 
-  /// Get the factor that the VF of this recipe's output should be scaled by
+  /// Get the factor that the VF of this recipe's output should be scaled by.
   unsigned getVFScaleFactor() const { return VFScaleFactor; }
 
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)



More information about the llvm-commits mailing list