Rework the rewriting of loads and stores for vector and integer allocas

to properly handle the combinations of these with split integer loads
and stores. This essentially replaces Evan's r168227 by refactoring the
code in a different way, and trynig to mirror that refactoring in both
the load and store sides of the rewriting.

Generally speaking there was some really problematic duplicated code
here that led to poorly founded assumptions and then subtle bugs. Now
much of the code actually flows through and follows a more consistent
style and logical path. There is still a tiny bit of duplication on the
store side of things, but it is much less bad.

This also changes the logic to never re-use a load or store instruction
as that was simply too error prone in practice.

I've added a few tests (one a reduction of the one in Evan's original
patch, which happened to be the same as the report in PR14349). I'm
going to look at adding a few more tests for things I found and fixed in
passing (such as the volatile tests in the vectorizable predicate).

This patch has survived bootstrap, and modulo one bugfix survived
Duncan's test suite, but let me know if anything else explodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@168346 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 164 insertions, 194 deletions

diff --git a/lib/Transforms/Scalar/SROA.cpp b/lib/Transforms/Scalar/SROA.cpp
index b0b6188604..15103d784b 100644
--- a/lib/Transforms/Scalar/SROA.cpp
+++ b/lib/Transforms/Scalar/SROA.cpp
@@ -1382,11 +1382,7 @@ class SROA : public FunctionPass {
   /// \brief A collection of instructions to delete.
   /// We try to batch deletions to simplify code and make things a bit more
   /// efficient.
-  SmallVector<Instruction *, 8> DeadInsts;
-
-  /// \brief A set to prevent repeatedly marking an instruction split into many
-  /// uses as dead. Only used to guard insertion into DeadInsts.
-  SmallPtrSet<Instruction *, 4> DeadSplitInsts;
+  SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
 
   /// \brief Post-promotion worklist.
   ///
@@ -1573,7 +1569,7 @@ private:
     do {
       LoadInst *LI = Loads.pop_back_val();
       LI->replaceAllUsesWith(NewPN);
-      Pass.DeadInsts.push_back(LI);
+      Pass.DeadInsts.insert(LI);
     } while (!Loads.empty());
 
     // Inject loads into all of the pred blocks.
@@ -1717,7 +1713,7 @@ private:
 
       DEBUG(dbgs() << "          speculated to: " << *V << "\n");
       LI->replaceAllUsesWith(V);
-      Pass.DeadInsts.push_back(LI);
+      Pass.DeadInsts.insert(LI);
     }
   }
 };
@@ -2134,8 +2130,13 @@ static bool isVectorPromotionViable(const DataLayout &TD,
     } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
       // Disable vector promotion when there are loads or stores of an FCA.
       return false;
-    } else if (!isa<LoadInst>(I->U->getUser()) &&
-               !isa<StoreInst>(I->U->getUser())) {
+    } else if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
+      if (LI->isVolatile())
+        return false;
+    } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
+      if (SI->isVolatile())
+        return false;
+    } else {
       return false;
     }
   }
@@ -2241,18 +2242,23 @@ static bool isIntegerWideningViable(const DataLayout &TD,
 static Value *extractInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
                              IntegerType *Ty, uint64_t Offset,
                              const Twine &Name) {
+  DEBUG(dbgs() << "       start: " << *V << "\n");
   IntegerType *IntTy = cast<IntegerType>(V->getType());
   assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
          "Element extends past full value");
   uint64_t ShAmt = 8*Offset;
   if (DL.isBigEndian())
     ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
-  if (ShAmt)
+  if (ShAmt) {
     V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
+    DEBUG(dbgs() << "     shifted: " << *V << "\n");
+  }
   assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
          "Cannot extract to a larger integer!");
-  if (Ty != IntTy)
+  if (Ty != IntTy) {
     V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
+    DEBUG(dbgs() << "     trunced: " << *V << "\n");
+  }
   return V;
 }
 
@@ -2262,20 +2268,27 @@ static Value *insertInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *Old,
   IntegerType *Ty = cast<IntegerType>(V->getType());
   assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
          "Cannot insert a larger integer!");
-  if (Ty != IntTy)
+  DEBUG(dbgs() << "       start: " << *V << "\n");
+  if (Ty != IntTy) {
     V = IRB.CreateZExt(V, IntTy, Name + ".ext");
+    DEBUG(dbgs() << "    extended: " << *V << "\n");
+  }
   assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
          "Element store outside of alloca store");
   uint64_t ShAmt = 8*Offset;
   if (DL.isBigEndian())
     ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
-  if (ShAmt)
+  if (ShAmt) {
     V = IRB.CreateShl(V, ShAmt, Name + ".shift");
+    DEBUG(dbgs() << "     shifted: " << *V << "\n");
+  }
 
   if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
     APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
     Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
+    DEBUG(dbgs() << "      masked: " << *Old << "\n");
     V = IRB.CreateOr(Old, V, Name + ".insert");
+    DEBUG(dbgs() << "    inserted: " << *V << "\n");
   }
   return V;
 }
@@ -2442,30 +2455,21 @@ private:
   void deleteIfTriviallyDead(Value *V) {
     Instruction *I = cast<Instruction>(V);
     if (isInstructionTriviallyDead(I))
-      Pass.DeadInsts.push_back(I);
+      Pass.DeadInsts.insert(I);
   }
 
-  bool rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
-    Value *Result;
+  Value *rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
+    Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                     getName(".load"));
     if (LI.getType() == VecTy->getElementType() ||
         BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
-      Result = IRB.CreateExtractElement(
-        IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
-        getIndex(IRB, BeginOffset), getName(".extract"));
-    } else {
-      Result = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
-                                     getName(".load"));
+      V = IRB.CreateExtractElement(V, getIndex(IRB, BeginOffset),
+                                   getName(".extract"));
     }
-    if (Result->getType() != LI.getType())
-      Result = convertValue(TD, IRB, Result, LI.getType());
-    LI.replaceAllUsesWith(Result);
-    Pass.DeadInsts.push_back(&LI);
-
-    DEBUG(dbgs() << "          to: " << *Result << "\n");
-    return true;
+    return V;
   }
 
-  bool rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
+  Value *rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
     assert(IntTy && "We cannot insert an integer to the alloca");
     assert(!LI.isVolatile());
     Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
@@ -2473,12 +2477,10 @@ private:
     V = convertValue(TD, IRB, V, IntTy);
     assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
     uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
-    V = extractInteger(TD, IRB, V, cast<IntegerType>(LI.getType()), Offset,
-                       getName(".extract"));
-    LI.replaceAllUsesWith(V);
-    Pass.DeadInsts.push_back(&LI);
-    DEBUG(dbgs() << "          to: " << *V << "\n");
-    return true;
+    if (Offset > 0 || EndOffset < NewAllocaEndOffset)
+      V = extractInteger(TD, IRB, V, cast<IntegerType>(LI.getType()), Offset,
+                         getName(".extract"));
+    return V;
   }
 
   bool visitLoadInst(LoadInst &LI) {
@@ -2488,7 +2490,29 @@ private:
     IRBuilder<> IRB(&LI);
 
     uint64_t Size = EndOffset - BeginOffset;
-    if (Size < TD.getTypeStoreSize(LI.getType())) {
+    bool IsSplitIntLoad = Size < TD.getTypeStoreSize(LI.getType());
+    Type *TargetTy = IsSplitIntLoad ? Type::getIntNTy(LI.getContext(), Size * 8)
+                                    : LI.getType();
+    bool IsPtrAdjusted = false;
+    Value *V;
+    if (VecTy) {
+      V = rewriteVectorizedLoadInst(IRB, LI, OldOp);
+    } else if (IntTy && LI.getType()->isIntegerTy()) {
+      V = rewriteIntegerLoad(IRB, LI);
+    } else if (BeginOffset == NewAllocaBeginOffset &&
+               canConvertValue(TD, NewAllocaTy, LI.getType())) {
+      V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                LI.isVolatile(), getName(".load"));
+    } else {
+      Type *LTy = TargetTy->getPointerTo();
+      V = IRB.CreateAlignedLoad(getAdjustedAllocaPtr(IRB, LTy),
+                                getPartitionTypeAlign(TargetTy),
+                                LI.isVolatile(), getName(".load"));
+      IsPtrAdjusted = true;
+    }
+    V = convertValue(TD, IRB, V, TargetTy);
+
+    if (IsSplitIntLoad) {
       assert(!LI.isVolatile());
       assert(LI.getType()->isIntegerTy() &&
              "Only integer type loads and stores are split");
@@ -2498,21 +2522,8 @@ private:
       assert(LI.getType()->getIntegerBitWidth() ==
              TD.getTypeAllocSizeInBits(OldAI.getAllocatedType()) &&
              "Only alloca-wide loads can be split and recomposed");
-      IntegerType *NarrowTy = Type::getIntNTy(LI.getContext(), Size * 8);
-      bool IsConvertable = (BeginOffset - NewAllocaBeginOffset == 0) &&
-                           canConvertValue(TD, NewAllocaTy, NarrowTy);
-      Value *V;
       // Move the insertion point just past the load so that we can refer to it.
       IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
-      if (IsConvertable)
-        V = convertValue(TD, IRB,
-                         IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
-                                               getName(".load")),
-                         NarrowTy);
-      else
-        V = IRB.CreateAlignedLoad(
-          getAdjustedAllocaPtr(IRB, NarrowTy->getPointerTo()),
-          getPartitionTypeAlign(NarrowTy), getName(".load"));
       // Create a placeholder value with the same type as LI to use as the
       // basis for the new value. This allows us to replace the uses of LI with
       // the computed value, and then replace the placeholder with LI, leaving
@@ -2524,104 +2535,40 @@ private:
       LI.replaceAllUsesWith(V);
       Placeholder->replaceAllUsesWith(&LI);
       delete Placeholder;
-      if (Pass.DeadSplitInsts.insert(&LI))
-        Pass.DeadInsts.push_back(&LI);
-      DEBUG(dbgs() << "          to: " << *V << "\n");
-      return IsConvertable;
-    }
-
-    if (VecTy)
-      return rewriteVectorizedLoadInst(IRB, LI, OldOp);
-    if (IntTy && LI.getType()->isIntegerTy())
-      return rewriteIntegerLoad(IRB, LI);
-
-    if (BeginOffset == NewAllocaBeginOffset &&
-        canConvertValue(TD, NewAllocaTy, LI.getType())) {
-      Value *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
-                                           LI.isVolatile(), getName(".load"));
-      Value *NewV = convertValue(TD, IRB, NewLI, LI.getType());
-      LI.replaceAllUsesWith(NewV);
-      Pass.DeadInsts.push_back(&LI);
-
-      DEBUG(dbgs() << "          to: " << *NewLI << "\n");
-      return !LI.isVolatile();
+    } else {
+      LI.replaceAllUsesWith(V);
     }
 
-    assert(!IntTy && "Invalid load found with int-op widening enabled");
-
-    Value *NewPtr = getAdjustedAllocaPtr(IRB,
-                                         LI.getPointerOperand()->getType());
-    LI.setOperand(0, NewPtr);
-    LI.setAlignment(getPartitionTypeAlign(LI.getType()));
-    DEBUG(dbgs() << "          to: " << LI << "\n");
-
+    Pass.DeadInsts.insert(&LI);
     deleteIfTriviallyDead(OldOp);
-    return NewPtr == &NewAI && !LI.isVolatile();
-  }
-
-  bool rewriteWideStoreInst(IRBuilder<> &IRB, StoreInst &SI, Type *ValueTy,
-                            unsigned Size) {
-    assert(!SI.isVolatile());
-    assert(ValueTy->isIntegerTy() &&
-           "Only integer type loads and stores are split");
-    assert(ValueTy->getIntegerBitWidth() ==
-           TD.getTypeStoreSizeInBits(ValueTy) &&
-           "Non-byte-multiple bit width");
-    assert(ValueTy->getIntegerBitWidth() ==
-           TD.getTypeSizeInBits(OldAI.getAllocatedType()) &&
-           "Only alloca-wide stores can be split and recomposed");
-    IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
-    Value *V = extractInteger(TD, IRB, SI.getValueOperand(), NarrowTy,
-                              BeginOffset, getName(".extract"));
-    StoreInst *NewSI;
-    bool IsConvertable = (BeginOffset - NewAllocaBeginOffset == 0) &&
-      canConvertValue(TD, NarrowTy, NewAllocaTy);
-    if (IsConvertable)
-      NewSI = IRB.CreateAlignedStore(convertValue(TD, IRB, V, NewAllocaTy),
-                                     &NewAI, NewAI.getAlignment());
-    else
-      NewSI = IRB.CreateAlignedStore(
-        V, getAdjustedAllocaPtr(IRB, NarrowTy->getPointerTo()),
-        getPartitionTypeAlign(NarrowTy));
-    (void)NewSI;
-    if (Pass.DeadSplitInsts.insert(&SI))
-      Pass.DeadInsts.push_back(&SI);
-
-    DEBUG(dbgs() << "          to: " << *NewSI << "\n");
-    return IsConvertable;
+    DEBUG(dbgs() << "          to: " << *V << "\n");
+    return !LI.isVolatile() && !IsPtrAdjusted;
   }
 
-  bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, StoreInst &SI,
-                                  Value *OldOp) {
-    Value *V = SI.getValueOperand();
-    Type *ValueTy = V->getType();
-    if (ValueTy == ElementTy ||
+  bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, Value *V,
+                                  StoreInst &SI, Value *OldOp) {
+    if (V->getType() == ElementTy ||
         BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
-      if (ValueTy != ElementTy)
+      if (V->getType() != ElementTy)
         V = convertValue(TD, IRB, V, ElementTy);
       LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
                                            getName(".load"));
       V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset),
                                   getName(".insert"));
-    } else if (ValueTy != VecTy) {
-      uint64_t Size = EndOffset - BeginOffset;
-      if (Size < TD.getTypeStoreSize(ValueTy))
-        return rewriteWideStoreInst(IRB, SI, ValueTy, Size);
-
+    } else if (V->getType() != VecTy) {
       V = convertValue(TD, IRB, V, VecTy);
     }
     StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
-    Pass.DeadInsts.push_back(&SI);
+    Pass.DeadInsts.insert(&SI);
 
     (void)Store;
     DEBUG(dbgs() << "          to: " << *Store << "\n");
     return true;
   }
 
-  bool rewriteIntegerStore(IRBuilder<> &IRB, StoreInst &SI) {
+  bool rewriteIntegerStore(IRBuilder<> &IRB, Value *V, StoreInst &SI) {
     assert(IntTy && "We cannot extract an integer from the alloca");
     assert(!SI.isVolatile());
-    Value *V = SI.getValueOperand();
     if (TD.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
       Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
                                          getName(".oldload"));
@@ -2633,7 +2580,7 @@ private:
     }
     V = convertValue(TD, IRB, V, NewAllocaTy);
     StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
-    Pass.DeadInsts.push_back(&SI);
+    Pass.DeadInsts.insert(&SI);
     (void)Store;
     DEBUG(dbgs() << "          to: " << *Store << "\n");
     return true;
@@ -2645,46 +2592,53 @@ private:
     assert(OldOp == OldPtr);
     IRBuilder<> IRB(&SI);
 
-    if (VecTy)
-      return rewriteVectorizedStoreInst(IRB, SI, OldOp);
-    Type *ValueTy = SI.getValueOperand()->getType();
-
-    uint64_t Size = EndOffset - BeginOffset;
-    if (Size < TD.getTypeStoreSize(ValueTy))
-      return rewriteWideStoreInst(IRB, SI, ValueTy, Size);
-
-    if (IntTy && ValueTy->isIntegerTy())
-      return rewriteIntegerStore(IRB, SI);
+    Value *V = SI.getValueOperand();
 
     // Strip all inbounds GEPs and pointer casts to try to dig out any root
     // alloca that should be re-examined after promoting this alloca.
-    if (ValueTy->isPointerTy())
-      if (AllocaInst *AI = dyn_cast<AllocaInst>(SI.getValueOperand()
-                                                  ->stripInBoundsOffsets()))
+    if (V->getType()->isPointerTy())
+      if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
         Pass.PostPromotionWorklist.insert(AI);
 
-    if (BeginOffset == NewAllocaBeginOffset &&
-        canConvertValue(TD, ValueTy, NewAllocaTy)) {
-      Value *NewV = convertValue(TD, IRB, SI.getValueOperand(), NewAllocaTy);
-      StoreInst *NewSI = IRB.CreateAlignedStore(NewV, &NewAI, NewAI.getAlignment(),
-                                                SI.isVolatile());
-      (void)NewSI;
-      Pass.DeadInsts.push_back(&SI);
-
-      DEBUG(dbgs() << "          to: " << *NewSI << "\n");
-      return !SI.isVolatile();
+    uint64_t Size = EndOffset - BeginOffset;
+    if (Size < TD.getTypeStoreSize(V->getType())) {
+      assert(!SI.isVolatile());
+      assert(V->getType()->isIntegerTy() &&
+             "Only integer type loads and stores are split");
+      assert(V->getType()->getIntegerBitWidth() ==
+             TD.getTypeStoreSizeInBits(V->getType()) &&
+             "Non-byte-multiple bit width");
+      assert(V->getType()->getIntegerBitWidth() ==
+             TD.getTypeSizeInBits(OldAI.getAllocatedType()) &&
+             "Only alloca-wide stores can be split and recomposed");
+      IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
+      V = extractInteger(TD, IRB, V, NarrowTy, BeginOffset,
+                         getName(".extract"));
     }
 
-    assert(!IntTy && "Invalid store found with int-op widening enabled");
-
-    Value *NewPtr = getAdjustedAllocaPtr(IRB,
-                                         SI.getPointerOperand()->getType());
-    SI.setOperand(1, NewPtr);
-    SI.setAlignment(getPartitionTypeAlign(SI.getValueOperand()->getType()));
-    DEBUG(dbgs() << "          to: " << SI << "\n");
+    if (VecTy)
+      return rewriteVectorizedStoreInst(IRB, V, SI, OldOp);
+    if (IntTy && V->getType()->isIntegerTy())
+      return rewriteIntegerStore(IRB, V, SI);
 
+    StoreInst *NewSI;
+    if (BeginOffset == NewAllocaBeginOffset &&
+        canConvertValue(TD, V->getType(), NewAllocaTy)) {
+      V = convertValue(TD, IRB, V, NewAllocaTy);
+      NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
+                                     SI.isVolatile());
+    } else {
+      Value *NewPtr = getAdjustedAllocaPtr(IRB, V->getType()->getPointerTo());
+      NewSI = IRB.CreateAlignedStore(V, NewPtr,
+                                     getPartitionTypeAlign(V->getType()),
+                                     SI.isVolatile());
+    }
+    (void)NewSI;
+    Pass.DeadInsts.insert(&SI);
     deleteIfTriviallyDead(OldOp);
-    return NewPtr == &NewAI && !SI.isVolatile();
+
+    DEBUG(dbgs() << "          to: " << *NewSI << "\n");
+    return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
   }
 
   bool visitMemSetInst(MemSetInst &II) {
@@ -2704,8 +2658,7 @@ private:
     }
 
     // Record this instruction for deletion.
-    if (Pass.DeadSplitInsts.insert(&II))
-      Pass.DeadInsts.push_back(&II);
+    Pass.DeadInsts.insert(&II);
 
     Type *AllocaTy = NewAI.getAllocatedType();
     Type *ScalarTy = AllocaTy->getScalarType();
@@ -2861,8 +2814,7 @@ private:
       return false;
     }
     // Record this instruction for deletion.
-    if (Pass.DeadSplitInsts.insert(&II))
-      Pass.DeadInsts.push_back(&II);
+    Pass.DeadInsts.insert(&II);
 
     bool IsWholeAlloca = BeginOffset == NewAllocaBeginOffset &&
                          EndOffset == NewAllocaEndOffset;
@@ -2972,8 +2924,7 @@ private:
     assert(II.getArgOperand(1) == OldPtr);
 
     // Record this instruction for deletion.
-    if (Pass.DeadSplitInsts.insert(&II))
-      Pass.DeadInsts.push_back(&II);
+    Pass.DeadInsts.insert(&II);
 
     ConstantInt *Size
       = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
@@ -3542,7 +3493,7 @@ bool SROA::runOnAlloca(AllocaInst &AI) {
        DI != DE; ++DI) {
     Changed = true;
     (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
-    DeadInsts.push_back(*DI);
+    DeadInsts.insert(*DI);
   }
   for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
                                             DE = P.dead_op_end();
@@ -3553,7 +3504,7 @@ bool SROA::runOnAlloca(AllocaInst &AI) {
     if (Instruction *OldI = dyn_cast<Instruction>(OldV))
       if (isInstructionTriviallyDead(OldI)) {
         Changed = true;
-        DeadInsts.push_back(OldI);
+        DeadInsts.insert(OldI);
       }
   }
 
@@ -3574,7 +3525,6 @@ bool SROA::runOnAlloca(AllocaInst &AI) {
 /// We also record the alloca instructions deleted here so that they aren't
 /// subsequently handed to mem2reg to promote.
 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
-  DeadSplitInsts.clear();
   while (!DeadInsts.empty()) {
     Instruction *I = DeadInsts.pop_back_val();
     DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
@@ -3586,7 +3536,7 @@ void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
         // Zero out the operand and see if it becomes trivially dead.
         *OI = 0;
         if (isInstructionTriviallyDead(U))
-          DeadInsts.push_back(U);
+          DeadInsts.insert(U);
       }
 
     if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
diff --git a/test/Transforms/SROA/basictest.ll b/test/Transforms/SROA/basictest.ll
index 110950f76a..b363eefb3f 100644
--- a/test/Transforms/SROA/basictest.ll
+++ b/test/Transforms/SROA/basictest.ll
@@ -1100,12 +1100,12 @@ entry:
   %imag = getelementptr inbounds { float, float }* %retval, i32 0, i32 1
   store float %phi.real, float* %real
   store float %phi.imag, float* %imag
+  ; CHECK-NEXT: %[[real_convert:.*]] = bitcast float %[[real]] to i32
   ; CHECK-NEXT: %[[imag_convert:.*]] = bitcast float %[[imag]] to i32
   ; CHECK-NEXT: %[[imag_ext:.*]] = zext i32 %[[imag_convert]] to i64
   ; CHECK-NEXT: %[[imag_shift:.*]] = shl i64 %[[imag_ext]], 32
   ; CHECK-NEXT: %[[imag_mask:.*]] = and i64 undef, 4294967295
   ; CHECK-NEXT: %[[imag_insert:.*]] = or i64 %[[imag_mask]], %[[imag_shift]]
-  ; CHECK-NEXT: %[[real_convert:.*]] = bitcast float %[[real]] to i32
   ; CHECK-NEXT: %[[real_ext:.*]] = zext i32 %[[real_convert]] to i64
   ; CHECK-NEXT: %[[real_mask:.*]] = and i64 %[[imag_insert]], -4294967296
   ; CHECK-NEXT: %[[real_insert:.*]] = or i64 %[[real_mask]], %[[real_ext]]
diff --git a/test/Transforms/SROA/vector-extract.ll b/test/Transforms/SROA/vector-extract.ll
deleted file mode 100644
index 2ca6168b75..0000000000
--- a/test/Transforms/SROA/vector-extract.ll
+++ /dev/null
@@ -1,25 +0,0 @@
-; RUN: opt < %s -sroa -S | FileCheck %s
-; rdar://12713675
-
-target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
-
-define <2 x i16> @test1(i64 %x) nounwind ssp {
-; CHECK: @test1
-entry:
-  %tmp = alloca i64, align 8
-  br i1 undef, label %bb1, label %bb2
-; CHECK-NOT: alloca
-
-bb1:
-  store i64 %x, i64* %tmp, align 8
-; CHECK-NOT: store
-  %0 = bitcast i64* %tmp to <2 x i16>*
-  %1 = load <2 x i16>* %0, align 8
-; CHECK-NOT: load
-; CHECK: trunc i64 %x to i32
-; CHECK: bitcast i32
-  ret <2 x i16> %1
-
-bb2:
-  ret <2 x i16> < i16 0, i16 0 >
-}
diff --git a/test/Transforms/SROA/vector-promotion.ll b/test/Transforms/SROA/vector-promotion.ll
index 02e084bf11..ea28f5d1a6 100644
--- a/test/Transforms/SROA/vector-promotion.ll
+++ b/test/Transforms/SROA/vector-promotion.ll
@@ -220,3 +220,48 @@ entry:
   ret i32 %load
 ; CHECK: ret i32
 }
+
+define <2 x i8> @PR14349.1(i32 %x) {
+; CEHCK: @PR14349.1
+; The first testcase for broken SROA rewriting of split integer loads and
+; stores due to smaller vector loads and stores. This particular test ensures
+; that we can rewrite a split store of an integer to a store of a vector.
+entry:
+  %a = alloca i32
+; CHECK-NOT: alloca
+
+  store i32 %x, i32* %a
+; CHECK-NOT: store
+
+  %cast = bitcast i32* %a to <2 x i8>*
+  %vec = load <2 x i8>* %cast
+; CHECK-NOT: load
+
+  ret <2 x i8> %vec
+; CHECK: %[[trunc:.*]] = trunc i32 %x to i16
+; CHECK: %[[cast:.*]] = bitcast i16 %[[trunc]] to <2 x i8>
+; CHECK: ret <2 x i8> %[[cast]]
+}
+
+define i32 @PR14349.2(<2 x i8> %x) {
+; CEHCK: @PR14349.2
+; The first testcase for broken SROA rewriting of split integer loads and
+; stores due to smaller vector loads and stores. This particular test ensures
+; that we can rewrite a split load of an integer to a load of a vector.
+entry:
+  %a = alloca i32
+; CHECK-NOT: alloca
+
+  %cast = bitcast i32* %a to <2 x i8>*
+  store <2 x i8> %x, <2 x i8>* %cast
+; CHECK-NOT: store
+
+  %int = load i32* %a
+; CHECK-NOT: load
+
+  ret i32 %int
+; CHECK: %[[cast:.*]] = bitcast <2 x i8> %x to i16
+; CHECK: %[[trunc:.*]] = zext i16 %[[cast]] to i32
+; CHECK: %[[insert:.*]] = or i32 %{{.*}}, %[[trunc]]
+; CHECK: ret i32 %[[insert]]
+}