[SLP] Vectorize non-power-of-2 ops with padding.

This patch introduces a new VectorizeWithPadding node type for root and
leave nodes to allow vectorizing loads/stores with non-power-of-2 number
of elements.

VectorizeWithPadding load nodes will pad the result to the next power of 2
with poison elements.

Non-leaf nodes will operate on normal power-of-2 vectors. For those
non-leaf nodes, we still track the number of padding elements needed to
go to the next power-of-2, to be used in various places, like cost
computation.

VectorizeWithPadding store nodes strip away the padding elements and
store the non-power-of-2 number of data elements.

Note that re-ordering and shuffling is not implemented for nodes
requiring padding yet to keep the initial implementation simpler.

The initial implementation also only tries to vectorize with padding if
original number of elements + 1 is a power-of-2, i.e. if only a single
padding element is needed.

The feature is guarded by a new flag, off by defaul for now.

Author: fhahn

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -179,6 +179,10 @@ static cl::opt<bool>
     ViewSLPTree("view-slp-tree", cl::Hidden,
                 cl::desc("Display the SLP trees with Graphviz"));
 
+static cl::opt<bool> VectorizeWithPadding(
+    "slp-vectorize-with-padding", cl::init(false), cl::Hidden,
+    cl::desc("Try to vectorize non-power-of-2 operations using padding."));
+
 // Limit the number of alias checks. The limit is chosen so that
 // it has no negative effect on the llvm benchmarks.
 static const unsigned AliasedCheckLimit = 10;
@@ -2733,6 +2737,9 @@ class BoUpSLP {
                           SmallVectorImpl<Value *> *OpScalars = nullptr,
                           SmallVectorImpl<Value *> *AltScalars = nullptr) const;
 
+    /// Return the number of padding lanes (containg poison) for this node.
+    unsigned isNonPowOf2Vec() const { return !isPowerOf2_32(Scalars.size()); }
+
 #ifndef NDEBUG
     /// Debug printer.
     LLVM_DUMP_METHOD void dump() const {
@@ -2891,9 +2898,13 @@ class BoUpSLP {
           ValueToGatherNodes.try_emplace(V).first->getSecond().insert(Last);
     }
 
-    if (UserTreeIdx.UserTE)
+    if (UserTreeIdx.UserTE) {
       Last->UserTreeIndices.push_back(UserTreeIdx);
-
+      if (!isPowerOf2_32(Last->Scalars.size())) {
+        assert((Last->ReorderIndices.empty()) &&
+               "Reodering isn't implemented for nodes with padding yet");
+      }
+    }
     return Last;
   }
 
@@ -2921,7 +2932,8 @@ class BoUpSLP {
   /// and fills required data before actual scheduling of the instructions.
   TreeEntry::EntryState getScalarsVectorizationState(
       InstructionsState &S, ArrayRef<Value *> VL, bool IsScatterVectorizeUserTE,
-      OrdersType &CurrentOrder, SmallVectorImpl<Value *> &PointerOps) const;
+      OrdersType &CurrentOrder, SmallVectorImpl<Value *> &PointerOps,
+      bool HasPadding) const;
 
   /// Maps a specific scalar to its tree entry.
   SmallDenseMap<Value *, TreeEntry *> ScalarToTreeEntry;
@@ -3881,6 +3893,9 @@ static LoadsState canVectorizeLoads(ArrayRef<Value *> VL, const Value *VL0,
   Order.clear();
   // Check the order of pointer operands or that all pointers are the same.
   bool IsSorted = sortPtrAccesses(PointerOps, ScalarTy, DL, SE, Order);
+  if (!Order.empty() && !isPowerOf2_32(VL.size()))
+    return LoadsState::Gather;
+
   if (IsSorted || all_of(PointerOps, [&](Value *P) {
         return arePointersCompatible(P, PointerOps.front(), TLI);
       })) {
@@ -4570,6 +4585,10 @@ bool BoUpSLP::canReorderOperands(
     TreeEntry *UserTE, SmallVectorImpl<std::pair<unsigned, TreeEntry *>> &Edges,
     ArrayRef<TreeEntry *> ReorderableGathers,
     SmallVectorImpl<TreeEntry *> &GatherOps) {
+  // Reordering isn't implemented for nodes with padding yet.
+  if (UserTE->isNonPowOf2Vec())
+    return false;
+
   for (unsigned I = 0, E = UserTE->getNumOperands(); I < E; ++I) {
     if (any_of(Edges, [I](const std::pair<unsigned, TreeEntry *> &OpData) {
           return OpData.first == I &&
@@ -4748,6 +4767,9 @@ void BoUpSLP::reorderBottomToTop(bool IgnoreReorder) {
         auto Res = OrdersUses.insert(std::make_pair(OrdersType(), 0));
         const auto &&AllowsReordering = [IgnoreReorder, &GathersToOrders](
                                             const TreeEntry *TE) {
+          // Reordering for nodes with padding not implemented yet.
+          if (TE->isNonPowOf2Vec())
+            return false;
           if (!TE->ReorderIndices.empty() || !TE->ReuseShuffleIndices.empty() ||
               (TE->State == TreeEntry::Vectorize && TE->isAltShuffle()) ||
               (IgnoreReorder && TE->Idx == 0))
@@ -5235,7 +5257,8 @@ static bool isAlternateInstruction(const Instruction *I,
 
 BoUpSLP::TreeEntry::EntryState BoUpSLP::getScalarsVectorizationState(
     InstructionsState &S, ArrayRef<Value *> VL, bool IsScatterVectorizeUserTE,
-    OrdersType &CurrentOrder, SmallVectorImpl<Value *> &PointerOps) const {
+    OrdersType &CurrentOrder, SmallVectorImpl<Value *> &PointerOps,
+    bool HasPadding) const {
   assert(S.MainOp && "Expected instructions with same/alternate opcodes only.");
 
   unsigned ShuffleOrOp =
@@ -5258,7 +5281,7 @@ BoUpSLP::TreeEntry::EntryState BoUpSLP::getScalarsVectorizationState(
   }
   case Instruction::ExtractValue:
   case Instruction::ExtractElement: {
-    bool Reuse = canReuseExtract(VL, VL0, CurrentOrder);
+    bool Reuse = !HasPadding && canReuseExtract(VL, VL0, CurrentOrder);
     if (Reuse || !CurrentOrder.empty())
       return TreeEntry::Vectorize;
     LLVM_DEBUG(dbgs() << "SLP: Gather extract sequence.\n");
@@ -5355,6 +5378,15 @@ BoUpSLP::TreeEntry::EntryState BoUpSLP::getScalarsVectorizationState(
     }
     return TreeEntry::Vectorize;
   }
+  case Instruction::UDiv:
+  case Instruction::SDiv:
+  case Instruction::URem:
+  case Instruction::SRem:
+    // The instruction may trigger immediate UB on the poison/undef padding
+    // elements, so force gather to avoid introducing new UB.
+    if (HasPadding)
+      return TreeEntry::NeedToGather;
+    [[fallthrough]];
   case Instruction::Select:
   case Instruction::FNeg:
   case Instruction::Add:
@@ -5363,11 +5395,7 @@ BoUpSLP::TreeEntry::EntryState BoUpSLP::getScalarsVectorizationState(
   case Instruction::FSub:
   case Instruction::Mul:
   case Instruction::FMul:
-  case Instruction::UDiv:
-  case Instruction::SDiv:
   case Instruction::FDiv:
-  case Instruction::URem:
-  case Instruction::SRem:
   case Instruction::FRem:
   case Instruction::Shl:
   case Instruction::LShr:
@@ -5550,6 +5578,7 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
                                  bool DoNotFail = false) {
     // Check that every instruction appears once in this bundle.
     DenseMap<Value *, unsigned> UniquePositions(VL.size());
+    auto OriginalVL = VL;
     for (Value *V : VL) {
       if (isConstant(V)) {
         ReuseShuffleIndicies.emplace_back(
@@ -5562,6 +5591,7 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
       if (Res.second)
         UniqueValues.emplace_back(V);
     }
+
     size_t NumUniqueScalarValues = UniqueValues.size();
     if (NumUniqueScalarValues == VL.size()) {
       ReuseShuffleIndicies.clear();
@@ -5589,6 +5619,14 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
             NonUniqueValueVL.append(PWSz - UniqueValues.size(),
                                     UniqueValues.back());
             VL = NonUniqueValueVL;
+
+            if (UserTreeIdx.UserTE && UserTreeIdx.UserTE->isNonPowOf2Vec()) {
+              LLVM_DEBUG(dbgs() << "SLP: Reshuffling scalars not yet supported "
+                                   "for nodes with padding.\n");
+              newTreeEntry(OriginalVL, std::nullopt /*not vectorized*/, S,
+                           UserTreeIdx);
+              return false;
+            }
           }
           return true;
         }
@@ -5597,6 +5635,13 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
         return false;
       }
       VL = UniqueValues;
+      if (UserTreeIdx.UserTE && UserTreeIdx.UserTE->isNonPowOf2Vec()) {
+        LLVM_DEBUG(dbgs() << "SLP: Reshuffling scalars not yet supported for "
+                             "nodes with padding.\n");
+        newTreeEntry(OriginalVL, std::nullopt /*not vectorized*/, S,
+                     UserTreeIdx);
+        return false;
+      }
     }
     return true;
   };
@@ -5861,7 +5906,8 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
   OrdersType CurrentOrder;
   SmallVector<Value *> PointerOps;
   TreeEntry::EntryState State = getScalarsVectorizationState(
-      S, VL, IsScatterVectorizeUserTE, CurrentOrder, PointerOps);
+      S, VL, IsScatterVectorizeUserTE, CurrentOrder, PointerOps,
+      UserTreeIdx.UserTE && UserTreeIdx.UserTE->isNonPowOf2Vec() > 0);
   if (State == TreeEntry::NeedToGather) {
     newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx,
                  ReuseShuffleIndicies);
@@ -6957,7 +7003,8 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
     return Constant::getAllOnesValue(Ty);
   }
 
-  InstructionCost getBuildVectorCost(ArrayRef<Value *> VL, Value *Root) {
+  InstructionCost getBuildVectorCost(ArrayRef<Value *> VL, Value *Root,
+                                     bool WithPadding = false) {
     if ((!Root && allConstant(VL)) || all_of(VL, UndefValue::classof))
       return TTI::TCC_Free;
     auto *VecTy = FixedVectorType::get(VL.front()->getType(), VL.size());
@@ -6968,7 +7015,7 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
     InstructionsState S = getSameOpcode(VL, *R.TLI);
     const unsigned Sz = R.DL->getTypeSizeInBits(VL.front()->getType());
     unsigned MinVF = R.getMinVF(2 * Sz);
-    if (VL.size() > 2 &&
+    if (!WithPadding && VL.size() > 2 &&
         ((S.getOpcode() == Instruction::Load && !S.isAltShuffle()) ||
          (InVectors.empty() &&
           any_of(seq<unsigned>(0, VL.size() / MinVF),
@@ -7079,7 +7126,7 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
         }
         GatherCost -= ScalarsCost;
       }
-    } else if (!Root && isSplat(VL)) {
+    } else if (!WithPadding && !Root && isSplat(VL)) {
       // Found the broadcasting of the single scalar, calculate the cost as
       // the broadcast.
       const auto *It =
@@ -7642,8 +7689,8 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
         CommonMask[Idx] = Mask[Idx] + VF;
   }
   Value *gather(ArrayRef<Value *> VL, unsigned MaskVF = 0,
-                Value *Root = nullptr) {
-    Cost += getBuildVectorCost(VL, Root);
+                Value *Root = nullptr, bool WithPadding = false) {
+    Cost += getBuildVectorCost(VL, Root, WithPadding);
     if (!Root) {
       // FIXME: Need to find a way to avoid use of getNullValue here.
       SmallVector<Constant *> Vals;
@@ -9712,6 +9759,9 @@ BoUpSLP::isGatherShuffledEntry(
   // No need to check for the topmost gather node.
   if (TE == VectorizableTree.front().get())
     return {};
+  // Gathering for nodes with padding is not implemented yet.
+  if (TE->isNonPowOf2Vec())
+    return {};
   Mask.assign(VL.size(), PoisonMaskElem);
   assert(TE->UserTreeIndices.size() == 1 &&
          "Expected only single user of the gather node.");
@@ -10433,7 +10483,7 @@ class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
     add(V1, NewMask);
   }
   Value *gather(ArrayRef<Value *> VL, unsigned MaskVF = 0,
-                Value *Root = nullptr) {
+                Value *Root = nullptr, bool WithPadding = false) {
     return R.gather(VL, Root);
   }
   Value *createFreeze(Value *V) { return Builder.CreateFreeze(V); }
@@ -10502,7 +10552,6 @@ Value *BoUpSLP::vectorizeOperand(TreeEntry *E, unsigned NodeIdx,
     SmallVector<int> Mask(E->ReorderIndices.begin(), E->ReorderIndices.end());
     reorderScalars(VL, Mask);
   }
-  const unsigned VF = VL.size();
   InstructionsState S = getSameOpcode(VL, *TLI);
   // Special processing for GEPs bundle, which may include non-gep values.
   if (!S.getOpcode() && VL.front()->getType()->isPointerTy()) {
@@ -10544,6 +10593,7 @@ Value *BoUpSLP::vectorizeOperand(TreeEntry *E, unsigned NodeIdx,
         ShuffleBuilder.add(V, Mask);
         return ShuffleBuilder.finalize(std::nullopt);
       };
+      const unsigned VF = VL.size();
       Value *V = vectorizeTree(VE, PostponedPHIs);
       if (VF != cast<FixedVectorType>(V->getType())->getNumElements()) {
         if (!VE->ReuseShuffleIndices.empty()) {
@@ -10670,6 +10720,14 @@ ResTy BoUpSLP::processBuildVector(const TreeEntry *E, Args &...Params) {
     return true;
   };
   BVTy ShuffleBuilder(Params...);
+  if (E->isNonPowOf2Vec()) {
+    Value *BV = ShuffleBuilder.gather(E->Scalars, 0, nullptr, true);
+    SmallVector<int> Mask(VF, PoisonMaskElem);
+    std::iota(Mask.begin(), Mask.begin() + E->Scalars.size(), 0);
+    ShuffleBuilder.add(BV, Mask);
+    return ShuffleBuilder.finalize(E->ReuseShuffleIndices);
+  }
+
   ResTy Res = ResTy();
   SmallVector<int> Mask;
   SmallVector<int> ExtractMask(GatheredScalars.size(), PoisonMaskElem);
@@ -13434,7 +13492,8 @@ bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R,
   unsigned VF = Chain.size();
 
   if (!isPowerOf2_32(Sz) || !isPowerOf2_32(VF) || VF < 2 || VF < MinVF)
-    return false;
+    if (!VectorizeWithPadding || (VF < MinVF && VF + 1 != MinVF))
+      return false;
 
   LLVM_DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idx
                     << "\n");
@@ -13530,9 +13589,36 @@ bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores,
                           << "MinVF (" << MinVF << ")\n");
       }
 
+      unsigned StartIdx = 0;
+      if (VectorizeWithPadding) {
+        unsigned CandVF = Operands.size() + 1;
+        if (isPowerOf2_32(CandVF) && CandVF <= MaxVF) {
+          assert(
+              all_of(
+                  Operands,
+                  [&](Value *V) {
+                    return cast<StoreInst>(V)->getValueOperand()->getType() ==
+                           cast<StoreInst>(Operands.front())
+                               ->getValueOperand()
+                               ->getType();
+                  }) &&
+              "Expected all operands of same type.");
+          if (!VectorizedStores.count(Operands.front()) &&
+              !VectorizedStores.count(Operands.back()) &&
+              TriedSequences
+                  .insert(std::make_pair(Operands.front(), Operands.back()))
+                  .second &&
+              vectorizeStoreChain(Operands, R, Operands.size(), MinVF)) {
+            // Mark the vectorized stores so that we don't vectorize them again.
+            VectorizedStores.insert(Operands.begin(), Operands.end());
+            Changed = true;
+            StartIdx += Operands.size();
+          }
+        }
+      }
+
       // FIXME: Is division-by-2 the correct step? Should we assert that the
       // register size is a power-of-2?
-      unsigned StartIdx = 0;
       for (unsigned Size = MaxVF; Size >= MinVF; Size /= 2) {
         for (unsigned Cnt = StartIdx, E = Operands.size(); Cnt + Size <= E;) {
           ArrayRef<Value *> Slice = ArrayRef(Operands).slice(Cnt, Size);

llvm/test/Transforms/SLPVectorizer/AArch64/vec15-base.ll

@@ -1,35 +1,45 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
-; RUN: opt -passes=slp-vectorizer -mtriple=arm64-apple-ios -S %s | FileCheck %s
+; RUN: opt -passes=slp-vectorizer -slp-vectorize-with-padding -mtriple=arm64-apple-ios -S %s | FileCheck --check-prefixes=PADDING %s
+; RUN: opt -passes=slp-vectorizer -slp-vectorize-with-padding=false -mtriple=arm64-apple-ios -S %s | FileCheck --check-prefixes=NO-PADDING %s
 
 define void @v15_load_i8_mul_by_constant_store(ptr %src, ptr noalias %dst) {
-; CHECK-LABEL: define void @v15_load_i8_mul_by_constant_store(
-; CHECK-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]]) {
-; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[GEP_SRC_0:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 0
-; CHECK-NEXT:    [[TMP0:%.*]] = load <8 x i8>, ptr [[GEP_SRC_0]], align 4
-; CHECK-NEXT:    [[TMP1:%.*]] = mul nsw <8 x i8> [[TMP0]], <i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10>
-; CHECK-NEXT:    store <8 x i8> [[TMP1]], ptr [[DST]], align 1
-; CHECK-NEXT:    [[GEP_SRC_8:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 8
-; CHECK-NEXT:    [[DST_8:%.*]] = getelementptr i8, ptr [[DST]], i8 8
-; CHECK-NEXT:    [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_SRC_8]], align 4
-; CHECK-NEXT:    [[TMP3:%.*]] = mul nsw <4 x i8> [[TMP2]], <i8 10, i8 10, i8 10, i8 10>
-; CHECK-NEXT:    store <4 x i8> [[TMP3]], ptr [[DST_8]], align 1
-; CHECK-NEXT:    [[GEP_SRC_12:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 12
-; CHECK-NEXT:    [[L_SRC_12:%.*]] = load i8, ptr [[GEP_SRC_12]], align 4
-; CHECK-NEXT:    [[MUL_12:%.*]] = mul nsw i8 [[L_SRC_12]], 10
-; CHECK-NEXT:    [[DST_12:%.*]] = getelementptr i8, ptr [[DST]], i8 12
-; CHECK-NEXT:    store i8 [[MUL_12]], ptr [[DST_12]], align 1
-; CHECK-NEXT:    [[GEP_SRC_13:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 13
-; CHECK-NEXT:    [[L_SRC_13:%.*]] = load i8, ptr [[GEP_SRC_13]], align 4
-; CHECK-NEXT:    [[MUL_13:%.*]] = mul nsw i8 [[L_SRC_13]], 10
-; CHECK-NEXT:    [[DST_13:%.*]] = getelementptr i8, ptr [[DST]], i8 13
-; CHECK-NEXT:    store i8 [[MUL_13]], ptr [[DST_13]], align 1
-; CHECK-NEXT:    [[GEP_SRC_14:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 14
-; CHECK-NEXT:    [[L_SRC_14:%.*]] = load i8, ptr [[GEP_SRC_14]], align 4
-; CHECK-NEXT:    [[MUL_14:%.*]] = mul nsw i8 [[L_SRC_14]], 10
-; CHECK-NEXT:    [[DST_14:%.*]] = getelementptr i8, ptr [[DST]], i8 14
-; CHECK-NEXT:    store i8 [[MUL_14]], ptr [[DST_14]], align 1
-; CHECK-NEXT:    ret void
+; PADDING-LABEL: define void @v15_load_i8_mul_by_constant_store(
+; PADDING-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]]) {
+; PADDING-NEXT:  entry:
+; PADDING-NEXT:    [[GEP_SRC_0:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 0
+; PADDING-NEXT:    [[TMP0:%.*]] = load <15 x i8>, ptr [[GEP_SRC_0]], align 4
+; PADDING-NEXT:    [[TMP1:%.*]] = mul nsw <15 x i8> [[TMP0]], <i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10>
+; PADDING-NEXT:    store <15 x i8> [[TMP1]], ptr [[DST]], align 1
+; PADDING-NEXT:    ret void
+;
+; NO-PADDING-LABEL: define void @v15_load_i8_mul_by_constant_store(
+; NO-PADDING-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]]) {
+; NO-PADDING-NEXT:  entry:
+; NO-PADDING-NEXT:    [[GEP_SRC_0:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 0
+; NO-PADDING-NEXT:    [[TMP0:%.*]] = load <8 x i8>, ptr [[GEP_SRC_0]], align 4
+; NO-PADDING-NEXT:    [[TMP1:%.*]] = mul nsw <8 x i8> [[TMP0]], <i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10, i8 10>
+; NO-PADDING-NEXT:    store <8 x i8> [[TMP1]], ptr [[DST]], align 1
+; NO-PADDING-NEXT:    [[GEP_SRC_8:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 8
+; NO-PADDING-NEXT:    [[DST_8:%.*]] = getelementptr i8, ptr [[DST]], i8 8
+; NO-PADDING-NEXT:    [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_SRC_8]], align 4
+; NO-PADDING-NEXT:    [[TMP3:%.*]] = mul nsw <4 x i8> [[TMP2]], <i8 10, i8 10, i8 10, i8 10>
+; NO-PADDING-NEXT:    store <4 x i8> [[TMP3]], ptr [[DST_8]], align 1
+; NO-PADDING-NEXT:    [[GEP_SRC_12:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 12
+; NO-PADDING-NEXT:    [[L_SRC_12:%.*]] = load i8, ptr [[GEP_SRC_12]], align 4
+; NO-PADDING-NEXT:    [[MUL_12:%.*]] = mul nsw i8 [[L_SRC_12]], 10
+; NO-PADDING-NEXT:    [[DST_12:%.*]] = getelementptr i8, ptr [[DST]], i8 12
+; NO-PADDING-NEXT:    store i8 [[MUL_12]], ptr [[DST_12]], align 1
+; NO-PADDING-NEXT:    [[GEP_SRC_13:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 13
+; NO-PADDING-NEXT:    [[L_SRC_13:%.*]] = load i8, ptr [[GEP_SRC_13]], align 4
+; NO-PADDING-NEXT:    [[MUL_13:%.*]] = mul nsw i8 [[L_SRC_13]], 10
+; NO-PADDING-NEXT:    [[DST_13:%.*]] = getelementptr i8, ptr [[DST]], i8 13
+; NO-PADDING-NEXT:    store i8 [[MUL_13]], ptr [[DST_13]], align 1
+; NO-PADDING-NEXT:    [[GEP_SRC_14:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i8 14
+; NO-PADDING-NEXT:    [[L_SRC_14:%.*]] = load i8, ptr [[GEP_SRC_14]], align 4
+; NO-PADDING-NEXT:    [[MUL_14:%.*]] = mul nsw i8 [[L_SRC_14]], 10
+; NO-PADDING-NEXT:    [[DST_14:%.*]] = getelementptr i8, ptr [[DST]], i8 14
+; NO-PADDING-NEXT:    store i8 [[MUL_14]], ptr [[DST_14]], align 1
+; NO-PADDING-NEXT:    ret void
 ;
 entry:
   %gep.src.0 = getelementptr inbounds i8, ptr %src, i8 0
@@ -123,5 +133,3 @@ entry:
 
   ret void
 }
-
-