X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FSROA.cpp;h=1ab7715168213f245a5e5bf77c9172f736914deb;hp=cb6da3b95615415a169b9c122fd99295afe76c71;hb=56e1394c8861ecdc551815ae875d2c3db2fa9cdb;hpb=ac104272d9fc42af8dc6853853b96d489685e5a7

diff --git a/lib/Transforms/Scalar/SROA.cpp b/lib/Transforms/Scalar/SROA.cpp
index cb6da3b9561..1ab77151682 100644
--- a/lib/Transforms/Scalar/SROA.cpp
+++ b/lib/Transforms/Scalar/SROA.cpp
@@ -25,44 +25,48 @@
 
 #define DEBUG_TYPE "sroa"
 #include "llvm/Transforms/Scalar.h"
-#include "llvm/Constants.h"
-#include "llvm/DIBuilder.h"
-#include "llvm/DebugInfo.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/Function.h"
-#include "llvm/IRBuilder.h"
-#include "llvm/Instructions.h"
-#include "llvm/IntrinsicInst.h"
-#include "llvm/LLVMContext.h"
-#include "llvm/Module.h"
-#include "llvm/Operator.h"
-#include "llvm/Pass.h"
+#include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SetVector.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/Statistic.h"
-#include "llvm/ADT/STLExtras.h"
-#include "llvm/Analysis/Dominators.h"
 #include "llvm/Analysis/Loads.h"
+#include "llvm/Analysis/PtrUseVisitor.h"
 #include "llvm/Analysis/ValueTracking.h"
+#include "llvm/DIBuilder.h"
+#include "llvm/DebugInfo.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Dominators.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/Operator.h"
+#include "llvm/InstVisitor.h"
+#include "llvm/Pass.h"
 #include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Compiler.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
-#include "llvm/Support/GetElementPtrTypeIterator.h"
-#include "llvm/Support/InstVisitor.h"
 #include "llvm/Support/MathExtras.h"
 #include "llvm/Support/raw_ostream.h"
-#include "llvm/DataLayout.h"
 #include "llvm/Transforms/Utils/Local.h"
 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
 #include "llvm/Transforms/Utils/SSAUpdater.h"
 using namespace llvm;
 
 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
-STATISTIC(NumNewAllocas,      "Number of new, smaller allocas introduced");
-STATISTIC(NumPromoted,        "Number of allocas promoted to SSA values");
+STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
+STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
+STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
+STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
+STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
+STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
-STATISTIC(NumDeleted,         "Number of instructions deleted");
-STATISTIC(NumVectorized,      "Number of vectorized aggregates");
+STATISTIC(NumDeleted, "Number of instructions deleted");
+STATISTIC(NumVectorized, "Number of vectorized aggregates");
 
 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
 /// forming SSA values through the SSAUpdater infrastructure.
@@ -70,176 +74,144 @@ static cl::opt<bool>
 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
 
 namespace {
-/// \brief Alloca partitioning representation.
-///
-/// This class represents a partitioning of an alloca into slices, and
-/// information about the nature of uses of each slice of the alloca. The goal
-/// is that this information is sufficient to decide if and how to split the
-/// alloca apart and replace slices with scalars. It is also intended that this
-/// structure can capture the relevant information needed both to decide about
-/// and to enact these transformations.
-class AllocaPartitioning {
+/// \brief A custom IRBuilder inserter which prefixes all names if they are
+/// preserved.
+template <bool preserveNames = true>
+class IRBuilderPrefixedInserter :
+    public IRBuilderDefaultInserter<preserveNames> {
+  std::string Prefix;
+
 public:
-  /// \brief A common base class for representing a half-open byte range.
-  struct ByteRange {
-    /// \brief The beginning offset of the range.
-    uint64_t BeginOffset;
+  void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
 
-    /// \brief The ending offset, not included in the range.
-    uint64_t EndOffset;
+protected:
+  void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
+                    BasicBlock::iterator InsertPt) const {
+    IRBuilderDefaultInserter<preserveNames>::InsertHelper(
+        I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
+  }
+};
 
-    ByteRange() : BeginOffset(), EndOffset() {}
-    ByteRange(uint64_t BeginOffset, uint64_t EndOffset)
-        : BeginOffset(BeginOffset), EndOffset(EndOffset) {}
+// Specialization for not preserving the name is trivial.
+template <>
+class IRBuilderPrefixedInserter<false> :
+    public IRBuilderDefaultInserter<false> {
+public:
+  void SetNamePrefix(const Twine &P) {}
+};
 
-    /// \brief Support for ordering ranges.
-    ///
-    /// This provides an ordering over ranges such that start offsets are
-    /// always increasing, and within equal start offsets, the end offsets are
-    /// decreasing. Thus the spanning range comes first in a cluster with the
-    /// same start position.
-    bool operator<(const ByteRange &RHS) const {
-      if (BeginOffset < RHS.BeginOffset) return true;
-      if (BeginOffset > RHS.BeginOffset) return false;
-      if (EndOffset > RHS.EndOffset) return true;
-      return false;
-    }
+/// \brief Provide a typedef for IRBuilder that drops names in release builds.
+#ifndef NDEBUG
+typedef llvm::IRBuilder<true, ConstantFolder,
+                        IRBuilderPrefixedInserter<true> > IRBuilderTy;
+#else
+typedef llvm::IRBuilder<false, ConstantFolder,
+                        IRBuilderPrefixedInserter<false> > IRBuilderTy;
+#endif
+}
 
-    /// \brief Support comparison with a single offset to allow binary searches.
-    friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) {
-      return LHS.BeginOffset < RHSOffset;
-    }
+namespace {
+/// \brief A used slice of an alloca.
+///
+/// This structure represents a slice of an alloca used by some instruction. It
+/// stores both the begin and end offsets of this use, a pointer to the use
+/// itself, and a flag indicating whether we can classify the use as splittable
+/// or not when forming partitions of the alloca.
+class Slice {
+  /// \brief The beginning offset of the range.
+  uint64_t BeginOffset;
 
-    friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
-                                                const ByteRange &RHS) {
-      return LHSOffset < RHS.BeginOffset;
-    }
+  /// \brief The ending offset, not included in the range.
+  uint64_t EndOffset;
 
-    bool operator==(const ByteRange &RHS) const {
-      return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset;
-    }
-    bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); }
-  };
+  /// \brief Storage for both the use of this slice and whether it can be
+  /// split.
+  PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
 
-  /// \brief A partition of an alloca.
-  ///
-  /// This structure represents a contiguous partition of the alloca. These are
-  /// formed by examining the uses of the alloca. During formation, they may
-  /// overlap but once an AllocaPartitioning is built, the Partitions within it
-  /// are all disjoint.
-  struct Partition : public ByteRange {
-    /// \brief Whether this partition is splittable into smaller partitions.
-    ///
-    /// We flag partitions as splittable when they are formed entirely due to
-    /// accesses by trivially splittable operations such as memset and memcpy.
-    ///
-    /// FIXME: At some point we should consider loads and stores of FCAs to be
-    /// splittable and eagerly split them into scalar values.
-    bool IsSplittable;
-
-    /// \brief Test whether a partition has been marked as dead.
-    bool isDead() const {
-      if (BeginOffset == UINT64_MAX) {
-        assert(EndOffset == UINT64_MAX);
-        return true;
-      }
-      return false;
-    }
+public:
+  Slice() : BeginOffset(), EndOffset() {}
+  Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
+      : BeginOffset(BeginOffset), EndOffset(EndOffset),
+        UseAndIsSplittable(U, IsSplittable) {}
 
-    /// \brief Kill a partition.
-    /// This is accomplished by setting both its beginning and end offset to
-    /// the maximum possible value.
-    void kill() {
-      assert(!isDead() && "He's Dead, Jim!");
-      BeginOffset = EndOffset = UINT64_MAX;
-    }
+  uint64_t beginOffset() const { return BeginOffset; }
+  uint64_t endOffset() const { return EndOffset; }
 
-    Partition() : ByteRange(), IsSplittable() {}
-    Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
-        : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
-  };
+  bool isSplittable() const { return UseAndIsSplittable.getInt(); }
+  void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
 
-  /// \brief A particular use of a partition of the alloca.
-  ///
-  /// This structure is used to associate uses of a partition with it. They
-  /// mark the range of bytes which are referenced by a particular instruction,
-  /// and includes a handle to the user itself and the pointer value in use.
-  /// The bounds of these uses are determined by intersecting the bounds of the
-  /// memory use itself with a particular partition. As a consequence there is
-  /// intentionally overlap between various uses of the same partition.
-  struct PartitionUse : public ByteRange {
-    /// \brief The use in question. Provides access to both user and used value.
-    ///
-    /// Note that this may be null if the partition use is *dead*, that is, it
-    /// should be ignored.
-    Use *U;
+  Use *getUse() const { return UseAndIsSplittable.getPointer(); }
 
-    PartitionUse() : ByteRange(), U() {}
-    PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U)
-        : ByteRange(BeginOffset, EndOffset), U(U) {}
-  };
+  bool isDead() const { return getUse() == 0; }
+  void kill() { UseAndIsSplittable.setPointer(0); }
 
-  /// \brief Construct a partitioning of a particular alloca.
+  /// \brief Support for ordering ranges.
   ///
-  /// Construction does most of the work for partitioning the alloca. This
-  /// performs the necessary walks of users and builds a partitioning from it.
-  AllocaPartitioning(const DataLayout &TD, AllocaInst &AI);
+  /// This provides an ordering over ranges such that start offsets are
+  /// always increasing, and within equal start offsets, the end offsets are
+  /// decreasing. Thus the spanning range comes first in a cluster with the
+  /// same start position.
+  bool operator<(const Slice &RHS) const {
+    if (beginOffset() < RHS.beginOffset()) return true;
+    if (beginOffset() > RHS.beginOffset()) return false;
+    if (isSplittable() != RHS.isSplittable()) return !isSplittable();
+    if (endOffset() > RHS.endOffset()) return true;
+    return false;
+  }
+
+  /// \brief Support comparison with a single offset to allow binary searches.
+  friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
+                                              uint64_t RHSOffset) {
+    return LHS.beginOffset() < RHSOffset;
+  }
+  friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
+                                              const Slice &RHS) {
+    return LHSOffset < RHS.beginOffset();
+  }
+
+  bool operator==(const Slice &RHS) const {
+    return isSplittable() == RHS.isSplittable() &&
+           beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
+  }
+  bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
+};
+} // end anonymous namespace
+
+namespace llvm {
+template <typename T> struct isPodLike;
+template <> struct isPodLike<Slice> {
+   static const bool value = true;
+};
+}
+
+namespace {
+/// \brief Representation of the alloca slices.
+///
+/// This class represents the slices of an alloca which are formed by its
+/// various uses. If a pointer escapes, we can't fully build a representation
+/// for the slices used and we reflect that in this structure. The uses are
+/// stored, sorted by increasing beginning offset and with unsplittable slices
+/// starting at a particular offset before splittable slices.
+class AllocaSlices {
+public:
+  /// \brief Construct the slices of a particular alloca.
+  AllocaSlices(const DataLayout &DL, AllocaInst &AI);
 
   /// \brief Test whether a pointer to the allocation escapes our analysis.
   ///
-  /// If this is true, the partitioning is never fully built and should be
+  /// If this is true, the slices are never fully built and should be
   /// ignored.
   bool isEscaped() const { return PointerEscapingInstr; }
 
-  /// \brief Support for iterating over the partitions.
-  /// @{
-  typedef SmallVectorImpl<Partition>::iterator iterator;
-  iterator begin() { return Partitions.begin(); }
-  iterator end() { return Partitions.end(); }
-
-  typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
-  const_iterator begin() const { return Partitions.begin(); }
-  const_iterator end() const { return Partitions.end(); }
-  /// @}
-
-  /// \brief Support for iterating over and manipulating a particular
-  /// partition's uses.
-  ///
-  /// The iteration support provided for uses is more limited, but also
-  /// includes some manipulation routines to support rewriting the uses of
-  /// partitions during SROA.
+  /// \brief Support for iterating over the slices.
   /// @{
-  typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
-  use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
-  use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
-  use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
-  use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
+  typedef SmallVectorImpl<Slice>::iterator iterator;
+  iterator begin() { return Slices.begin(); }
+  iterator end() { return Slices.end(); }
 
-  typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
-  const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
-  const_use_iterator use_begin(const_iterator I) const {
-    return Uses[I - begin()].begin();
-  }
-  const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
-  const_use_iterator use_end(const_iterator I) const {
-    return Uses[I - begin()].end();
-  }
-
-  unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); }
-  unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); }
-  const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const {
-    return Uses[PIdx][UIdx];
-  }
-  const PartitionUse &getUse(const_iterator I, unsigned UIdx) const {
-    return Uses[I - begin()][UIdx];
-  }
-
-  void use_push_back(unsigned Idx, const PartitionUse &PU) {
-    Uses[Idx].push_back(PU);
-  }
-  void use_push_back(const_iterator I, const PartitionUse &PU) {
-    Uses[I - begin()].push_back(PU);
-  }
+  typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
+  const_iterator begin() const { return Slices.begin(); }
+  const_iterator end() const { return Slices.end(); }
   /// @}
 
   /// \brief Allow iterating the dead users for this alloca.
@@ -265,114 +237,49 @@ public:
   dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
   /// @}
 
-  /// \brief MemTransferInst auxiliary data.
-  /// This struct provides some auxiliary data about memory transfer
-  /// intrinsics such as memcpy and memmove. These intrinsics can use two
-  /// different ranges within the same alloca, and provide other challenges to
-  /// correctly represent. We stash extra data to help us untangle this
-  /// after the partitioning is complete.
-  struct MemTransferOffsets {
-    /// The destination begin and end offsets when the destination is within
-    /// this alloca. If the end offset is zero the destination is not within
-    /// this alloca.
-    uint64_t DestBegin, DestEnd;
-
-    /// The source begin and end offsets when the source is within this alloca.
-    /// If the end offset is zero, the source is not within this alloca.
-    uint64_t SourceBegin, SourceEnd;
-
-    /// Flag for whether an alloca is splittable.
-    bool IsSplittable;
-  };
-  MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
-    return MemTransferInstData.lookup(&II);
-  }
-
-  /// \brief Map from a PHI or select operand back to a partition.
-  ///
-  /// When manipulating PHI nodes or selects, they can use more than one
-  /// partition of an alloca. We store a special mapping to allow finding the
-  /// partition referenced by each of these operands, if any.
-  iterator findPartitionForPHIOrSelectOperand(Use *U) {
-    SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
-      = PHIOrSelectOpMap.find(U);
-    if (MapIt == PHIOrSelectOpMap.end())
-      return end();
-
-    return begin() + MapIt->second.first;
-  }
-
-  /// \brief Map from a PHI or select operand back to the specific use of
-  /// a partition.
-  ///
-  /// Similar to mapping these operands back to the partitions, this maps
-  /// directly to the use structure of that partition.
-  use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
-    SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
-      = PHIOrSelectOpMap.find(U);
-    assert(MapIt != PHIOrSelectOpMap.end());
-    return Uses[MapIt->second.first].begin() + MapIt->second.second;
-  }
-
-  /// \brief Compute a common type among the uses of a particular partition.
-  ///
-  /// This routines walks all of the uses of a particular partition and tries
-  /// to find a common type between them. Untyped operations such as memset and
-  /// memcpy are ignored.
-  Type *getCommonType(iterator I) const;
-
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
   void print(raw_ostream &OS, const_iterator I, StringRef Indent = "  ") const;
-  void printUsers(raw_ostream &OS, const_iterator I,
+  void printSlice(raw_ostream &OS, const_iterator I,
                   StringRef Indent = "  ") const;
+  void printUse(raw_ostream &OS, const_iterator I,
+                StringRef Indent = "  ") const;
   void print(raw_ostream &OS) const;
-  void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
-  void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
+  void dump(const_iterator I) const;
+  void dump() const;
 #endif
 
 private:
   template <typename DerivedT, typename RetT = void> class BuilderBase;
-  class PartitionBuilder;
-  friend class AllocaPartitioning::PartitionBuilder;
-  class UseBuilder;
-  friend class AllocaPartitioning::UseBuilder;
+  class SliceBuilder;
+  friend class AllocaSlices::SliceBuilder;
 
-#ifndef NDEBUG
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
   /// \brief Handle to alloca instruction to simplify method interfaces.
   AllocaInst &AI;
 #endif
 
-  /// \brief The instruction responsible for this alloca having no partitioning.
+  /// \brief The instruction responsible for this alloca not having a known set
+  /// of slices.
   ///
   /// When an instruction (potentially) escapes the pointer to the alloca, we
-  /// store a pointer to that here and abort trying to partition the alloca.
-  /// This will be null if the alloca is partitioned successfully.
+  /// store a pointer to that here and abort trying to form slices of the
+  /// alloca. This will be null if the alloca slices are analyzed successfully.
   Instruction *PointerEscapingInstr;
 
-  /// \brief The partitions of the alloca.
-  ///
-  /// We store a vector of the partitions over the alloca here. This vector is
-  /// sorted by increasing begin offset, and then by decreasing end offset. See
-  /// the Partition inner class for more details. Initially (during
-  /// construction) there are overlaps, but we form a disjoint sequence of
-  /// partitions while finishing construction and a fully constructed object is
-  /// expected to always have this as a disjoint space.
-  SmallVector<Partition, 8> Partitions;
-
-  /// \brief The uses of the partitions.
+  /// \brief The slices of the alloca.
   ///
-  /// This is essentially a mapping from each partition to a list of uses of
-  /// that partition. The mapping is done with a Uses vector that has the exact
-  /// same number of entries as the partition vector. Each entry is itself
-  /// a vector of the uses.
-  SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
+  /// We store a vector of the slices formed by uses of the alloca here. This
+  /// vector is sorted by increasing begin offset, and then the unsplittable
+  /// slices before the splittable ones. See the Slice inner class for more
+  /// details.
+  SmallVector<Slice, 8> Slices;
 
   /// \brief Instructions which will become dead if we rewrite the alloca.
   ///
-  /// Note that these are not separated by partition. This is because we expect
-  /// a partitioned alloca to be completely rewritten or not rewritten at all.
-  /// If rewritten, all these instructions can simply be removed and replaced
-  /// with undef as they come from outside of the allocated space.
+  /// Note that these are not separated by slice. This is because we expect an
+  /// alloca to be completely rewritten or not rewritten at all. If rewritten,
+  /// all these instructions can simply be removed and replaced with undef as
+  /// they come from outside of the allocated space.
   SmallVector<Instruction *, 8> DeadUsers;
 
   /// \brief Operands which will become dead if we rewrite the alloca.
@@ -384,356 +291,250 @@ private:
   /// want to swap this particular input for undef to simplify the use lists of
   /// the alloca.
   SmallVector<Use *, 8> DeadOperands;
-
-  /// \brief The underlying storage for auxiliary memcpy and memset info.
-  SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
-
-  /// \brief A side datastructure used when building up the partitions and uses.
-  ///
-  /// This mapping is only really used during the initial building of the
-  /// partitioning so that we can retain information about PHI and select nodes
-  /// processed.
-  SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
-
-  /// \brief Auxiliary information for particular PHI or select operands.
-  SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
-
-  /// \brief A utility routine called from the constructor.
-  ///
-  /// This does what it says on the tin. It is the key of the alloca partition
-  /// splitting and merging. After it is called we have the desired disjoint
-  /// collection of partitions.
-  void splitAndMergePartitions();
 };
 }
 
-template <typename DerivedT, typename RetT>
-class AllocaPartitioning::BuilderBase
-    : public InstVisitor<DerivedT, RetT> {
-public:
-  BuilderBase(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
-      : TD(TD),
-        AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
-        P(P) {
-    enqueueUsers(AI, 0);
-  }
-
-protected:
-  const DataLayout &TD;
-  const uint64_t AllocSize;
-  AllocaPartitioning &P;
-
-  SmallPtrSet<Use *, 8> VisitedUses;
-
-  struct OffsetUse {
-    Use *U;
-    int64_t Offset;
-  };
-  SmallVector<OffsetUse, 8> Queue;
-
-  // The active offset and use while visiting.
-  Use *U;
-  int64_t Offset;
-
-  void enqueueUsers(Instruction &I, int64_t UserOffset) {
-    for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
-         UI != UE; ++UI) {
-      if (VisitedUses.insert(&UI.getUse())) {
-        OffsetUse OU = { &UI.getUse(), UserOffset };
-        Queue.push_back(OU);
-      }
-    }
-  }
-
-  bool computeConstantGEPOffset(GetElementPtrInst &GEPI, int64_t &GEPOffset) {
-    GEPOffset = Offset;
-    for (gep_type_iterator GTI = gep_type_begin(GEPI), GTE = gep_type_end(GEPI);
-         GTI != GTE; ++GTI) {
-      ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
-      if (!OpC)
-        return false;
-      if (OpC->isZero())
-        continue;
-
-      // Handle a struct index, which adds its field offset to the pointer.
-      if (StructType *STy = dyn_cast<StructType>(*GTI)) {
-        unsigned ElementIdx = OpC->getZExtValue();
-        const StructLayout *SL = TD.getStructLayout(STy);
-        uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
-        // Check that we can continue to model this GEP in a signed 64-bit offset.
-        if (ElementOffset > INT64_MAX ||
-            (GEPOffset >= 0 &&
-             ((uint64_t)GEPOffset + ElementOffset) > INT64_MAX)) {
-          DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
-                       << "what can be represented in an int64_t!\n"
-                       << "  alloca: " << P.AI << "\n");
-          return false;
-        }
-        if (GEPOffset < 0)
-          GEPOffset = ElementOffset + (uint64_t)-GEPOffset;
-        else
-          GEPOffset += ElementOffset;
-        continue;
-      }
+static Value *foldSelectInst(SelectInst &SI) {
+  // If the condition being selected on is a constant or the same value is
+  // being selected between, fold the select. Yes this does (rarely) happen
+  // early on.
+  if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
+    return SI.getOperand(1+CI->isZero());
+  if (SI.getOperand(1) == SI.getOperand(2))
+    return SI.getOperand(1);
 
-      APInt Index = OpC->getValue().sextOrTrunc(TD.getPointerSizeInBits());
-      Index *= APInt(Index.getBitWidth(),
-                     TD.getTypeAllocSize(GTI.getIndexedType()));
-      Index += APInt(Index.getBitWidth(), (uint64_t)GEPOffset,
-                     /*isSigned*/true);
-      // Check if the result can be stored in our int64_t offset.
-      if (!Index.isSignedIntN(sizeof(GEPOffset) * 8)) {
-        DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
-                     << "what can be represented in an int64_t!\n"
-                     << "  alloca: " << P.AI << "\n");
-        return false;
-      }
+  return 0;
+}
 
-      GEPOffset = Index.getSExtValue();
-    }
-    return true;
-  }
+/// \brief Builder for the alloca slices.
+///
+/// This class builds a set of alloca slices by recursively visiting the uses
+/// of an alloca and making a slice for each load and store at each offset.
+class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
+  friend class PtrUseVisitor<SliceBuilder>;
+  friend class InstVisitor<SliceBuilder>;
+  typedef PtrUseVisitor<SliceBuilder> Base;
 
-  Value *foldSelectInst(SelectInst &SI) {
-    // If the condition being selected on is a constant or the same value is
-    // being selected between, fold the select. Yes this does (rarely) happen
-    // early on.
-    if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
-      return SI.getOperand(1+CI->isZero());
-    if (SI.getOperand(1) == SI.getOperand(2)) {
-      assert(*U == SI.getOperand(1));
-      return SI.getOperand(1);
-    }
-    return 0;
-  }
-};
+  const uint64_t AllocSize;
+  AllocaSlices &S;
 
-/// \brief Builder for the alloca partitioning.
-///
-/// This class builds an alloca partitioning by recursively visiting the uses
-/// of an alloca and splitting the partitions for each load and store at each
-/// offset.
-class AllocaPartitioning::PartitionBuilder
-    : public BuilderBase<PartitionBuilder, bool> {
-  friend class InstVisitor<PartitionBuilder, bool>;
+  SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
+  SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
 
-  SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
+  /// \brief Set to de-duplicate dead instructions found in the use walk.
+  SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
 
 public:
-  PartitionBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
-      : BuilderBase<PartitionBuilder, bool>(TD, AI, P) {}
-
-  /// \brief Run the builder over the allocation.
-  bool operator()() {
-    // Note that we have to re-evaluate size on each trip through the loop as
-    // the queue grows at the tail.
-    for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
-      U = Queue[Idx].U;
-      Offset = Queue[Idx].Offset;
-      if (!visit(cast<Instruction>(U->getUser())))
-        return false;
-    }
-    return true;
-  }
+  SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &S)
+      : PtrUseVisitor<SliceBuilder>(DL),
+        AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), S(S) {}
 
 private:
-  bool markAsEscaping(Instruction &I) {
-    P.PointerEscapingInstr = &I;
-    return false;
+  void markAsDead(Instruction &I) {
+    if (VisitedDeadInsts.insert(&I))
+      S.DeadUsers.push_back(&I);
   }
 
-  void insertUse(Instruction &I, int64_t Offset, uint64_t Size,
+  void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
                  bool IsSplittable = false) {
-    // Completely skip uses which have a zero size or don't overlap the
-    // allocation.
-    if (Size == 0 ||
-        (Offset >= 0 && (uint64_t)Offset >= AllocSize) ||
-        (Offset < 0 && (uint64_t)-Offset >= Size)) {
+    // Completely skip uses which have a zero size or start either before or
+    // past the end of the allocation.
+    if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) {
       DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
-                   << " which starts past the end of the " << AllocSize
-                   << " byte alloca:\n"
-                   << "    alloca: " << P.AI << "\n"
-                   << "       use: " << I << "\n");
-      return;
-    }
-
-    // Clamp the start to the beginning of the allocation.
-    if (Offset < 0) {
-      DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
-                   << " to start at the beginning of the alloca:\n"
-                   << "    alloca: " << P.AI << "\n"
+                   << " which has zero size or starts outside of the "
+                   << AllocSize << " byte alloca:\n"
+                   << "    alloca: " << S.AI << "\n"
                    << "       use: " << I << "\n");
-      Size -= (uint64_t)-Offset;
-      Offset = 0;
+      return markAsDead(I);
     }
 
-    uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
+    uint64_t BeginOffset = Offset.getZExtValue();
+    uint64_t EndOffset = BeginOffset + Size;
 
     // Clamp the end offset to the end of the allocation. Note that this is
     // formulated to handle even the case where "BeginOffset + Size" overflows.
+    // This may appear superficially to be something we could ignore entirely,
+    // but that is not so! There may be widened loads or PHI-node uses where
+    // some instructions are dead but not others. We can't completely ignore
+    // them, and so have to record at least the information here.
     assert(AllocSize >= BeginOffset); // Established above.
     if (Size > AllocSize - BeginOffset) {
       DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
                    << " to remain within the " << AllocSize << " byte alloca:\n"
-                   << "    alloca: " << P.AI << "\n"
+                   << "    alloca: " << S.AI << "\n"
                    << "       use: " << I << "\n");
       EndOffset = AllocSize;
     }
 
-    Partition New(BeginOffset, EndOffset, IsSplittable);
-    P.Partitions.push_back(New);
+    S.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
   }
 
-  bool handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
-    uint64_t Size = TD.getTypeStoreSize(Ty);
-
-    // If this memory access can be shown to *statically* extend outside the
-    // bounds of of the allocation, it's behavior is undefined, so simply
-    // ignore it. Note that this is more strict than the generic clamping
-    // behavior of insertUse. We also try to handle cases which might run the
-    // risk of overflow.
-    // FIXME: We should instead consider the pointer to have escaped if this
-    // function is being instrumented for addressing bugs or race conditions.
-    if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
-        Size > (AllocSize - (uint64_t)Offset)) {
-      DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
-                   << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
-                   << " which extends past the end of the " << AllocSize
-                   << " byte alloca:\n"
-                   << "    alloca: " << P.AI << "\n"
-                   << "       use: " << I << "\n");
-      return true;
-    }
+  void visitBitCastInst(BitCastInst &BC) {
+    if (BC.use_empty())
+      return markAsDead(BC);
 
-    insertUse(I, Offset, Size);
-    return true;
+    return Base::visitBitCastInst(BC);
   }
 
-  bool visitBitCastInst(BitCastInst &BC) {
-    enqueueUsers(BC, Offset);
-    return true;
+  void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
+    if (GEPI.use_empty())
+      return markAsDead(GEPI);
+
+    return Base::visitGetElementPtrInst(GEPI);
   }
 
-  bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
-    int64_t GEPOffset;
-    if (!computeConstantGEPOffset(GEPI, GEPOffset))
-      return markAsEscaping(GEPI);
+  void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
+                         uint64_t Size, bool IsVolatile) {
+    // We allow splitting of loads and stores where the type is an integer type
+    // and cover the entire alloca. This prevents us from splitting over
+    // eagerly.
+    // FIXME: In the great blue eventually, we should eagerly split all integer
+    // loads and stores, and then have a separate step that merges adjacent
+    // alloca partitions into a single partition suitable for integer widening.
+    // Or we should skip the merge step and rely on GVN and other passes to
+    // merge adjacent loads and stores that survive mem2reg.
+    bool IsSplittable =
+        Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
 
-    enqueueUsers(GEPI, GEPOffset);
-    return true;
+    insertUse(I, Offset, Size, IsSplittable);
   }
 
-  bool visitLoadInst(LoadInst &LI) {
+  void visitLoadInst(LoadInst &LI) {
     assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
            "All simple FCA loads should have been pre-split");
-    return handleLoadOrStore(LI.getType(), LI, Offset);
+
+    if (!IsOffsetKnown)
+      return PI.setAborted(&LI);
+
+    uint64_t Size = DL.getTypeStoreSize(LI.getType());
+    return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
   }
 
-  bool visitStoreInst(StoreInst &SI) {
+  void visitStoreInst(StoreInst &SI) {
     Value *ValOp = SI.getValueOperand();
     if (ValOp == *U)
-      return markAsEscaping(SI);
+      return PI.setEscapedAndAborted(&SI);
+    if (!IsOffsetKnown)
+      return PI.setAborted(&SI);
+
+    uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
+
+    // If this memory access can be shown to *statically* extend outside the
+    // bounds of of the allocation, it's behavior is undefined, so simply
+    // ignore it. Note that this is more strict than the generic clamping
+    // behavior of insertUse. We also try to handle cases which might run the
+    // risk of overflow.
+    // FIXME: We should instead consider the pointer to have escaped if this
+    // function is being instrumented for addressing bugs or race conditions.
+    if (Offset.isNegative() || Size > AllocSize ||
+        Offset.ugt(AllocSize - Size)) {
+      DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
+                   << " which extends past the end of the " << AllocSize
+                   << " byte alloca:\n"
+                   << "    alloca: " << S.AI << "\n"
+                   << "       use: " << SI << "\n");
+      return markAsDead(SI);
+    }
 
     assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
            "All simple FCA stores should have been pre-split");
-    return handleLoadOrStore(ValOp->getType(), SI, Offset);
+    handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
   }
 
 
-  bool visitMemSetInst(MemSetInst &II) {
+  void visitMemSetInst(MemSetInst &II) {
     assert(II.getRawDest() == *U && "Pointer use is not the destination?");
     ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
-    uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
-    insertUse(II, Offset, Size, Length);
-    return true;
+    if ((Length && Length->getValue() == 0) ||
+        (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
+      // Zero-length mem transfer intrinsics can be ignored entirely.
+      return markAsDead(II);
+
+    if (!IsOffsetKnown)
+      return PI.setAborted(&II);
+
+    insertUse(II, Offset,
+              Length ? Length->getLimitedValue()
+                     : AllocSize - Offset.getLimitedValue(),
+              (bool)Length);
   }
 
-  bool visitMemTransferInst(MemTransferInst &II) {
+  void visitMemTransferInst(MemTransferInst &II) {
     ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
-    uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
-    if (!Size)
+    if ((Length && Length->getValue() == 0) ||
+        (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
       // Zero-length mem transfer intrinsics can be ignored entirely.
-      return true;
+      return markAsDead(II);
 
-    MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
+    if (!IsOffsetKnown)
+      return PI.setAborted(&II);
 
-    // Only intrinsics with a constant length can be split.
-    Offsets.IsSplittable = Length;
+    uint64_t RawOffset = Offset.getLimitedValue();
+    uint64_t Size = Length ? Length->getLimitedValue()
+                           : AllocSize - RawOffset;
 
-    if (*U == II.getRawDest()) {
-      Offsets.DestBegin = Offset;
-      Offsets.DestEnd = Offset + Size;
-    }
-    if (*U == II.getRawSource()) {
-      Offsets.SourceBegin = Offset;
-      Offsets.SourceEnd = Offset + Size;
+    // Check for the special case where the same exact value is used for both
+    // source and dest.
+    if (*U == II.getRawDest() && *U == II.getRawSource()) {
+      // For non-volatile transfers this is a no-op.
+      if (!II.isVolatile())
+        return markAsDead(II);
+
+      return insertUse(II, Offset, Size, /*IsSplittable=*/false);
     }
 
-    // If we have set up end offsets for both the source and the destination,
-    // we have found both sides of this transfer pointing at the same alloca.
-    bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd;
-    if (SeenBothEnds && II.getRawDest() != II.getRawSource()) {
-      unsigned PrevIdx = MemTransferPartitionMap[&II];
+    // If we have seen both source and destination for a mem transfer, then
+    // they both point to the same alloca.
+    bool Inserted;
+    SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
+    llvm::tie(MTPI, Inserted) =
+        MemTransferSliceMap.insert(std::make_pair(&II, S.Slices.size()));
+    unsigned PrevIdx = MTPI->second;
+    if (!Inserted) {
+      Slice &PrevP = S.Slices[PrevIdx];
 
       // Check if the begin offsets match and this is a non-volatile transfer.
       // In that case, we can completely elide the transfer.
-      if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) {
-        P.Partitions[PrevIdx].kill();
-        return true;
+      if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
+        PrevP.kill();
+        return markAsDead(II);
       }
 
       // Otherwise we have an offset transfer within the same alloca. We can't
       // split those.
-      P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false;
-    } else if (SeenBothEnds) {
-      // Handle the case where this exact use provides both ends of the
-      // operation.
-      assert(II.getRawDest() == II.getRawSource());
-
-      // For non-volatile transfers this is a no-op.
-      if (!II.isVolatile())
-        return true;
-
-      // Otherwise just suppress splitting.
-      Offsets.IsSplittable = false;
+      PrevP.makeUnsplittable();
     }
 
-
     // Insert the use now that we've fixed up the splittable nature.
-    insertUse(II, Offset, Size, Offsets.IsSplittable);
-
-    // Setup the mapping from intrinsic to partition of we've not seen both
-    // ends of this transfer.
-    if (!SeenBothEnds) {
-      unsigned NewIdx = P.Partitions.size() - 1;
-      bool Inserted
-        = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second;
-      assert(Inserted &&
-             "Already have intrinsic in map but haven't seen both ends");
-      (void)Inserted;
-    }
+    insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
 
-    return true;
+    // Check that we ended up with a valid index in the map.
+    assert(S.Slices[PrevIdx].getUse()->getUser() == &II &&
+           "Map index doesn't point back to a slice with this user.");
   }
 
   // Disable SRoA for any intrinsics except for lifetime invariants.
-  // FIXME: What about debug instrinsics? This matches old behavior, but
+  // FIXME: What about debug intrinsics? This matches old behavior, but
   // doesn't make sense.
-  bool visitIntrinsicInst(IntrinsicInst &II) {
+  void visitIntrinsicInst(IntrinsicInst &II) {
+    if (!IsOffsetKnown)
+      return PI.setAborted(&II);
+
     if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
         II.getIntrinsicID() == Intrinsic::lifetime_end) {
       ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
-      uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue());
+      uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
+                               Length->getLimitedValue());
       insertUse(II, Offset, Size, true);
-      return true;
+      return;
     }
 
-    return markAsEscaping(II);
+    Base::visitIntrinsicInst(II);
   }
 
   Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
     // We consider any PHI or select that results in a direct load or store of
-    // the same offset to be a viable use for partitioning purposes. These uses
+    // the same offset to be a viable use for slicing purposes. These uses
     // are considered unsplittable and the size is the maximum loaded or stored
     // size.
     SmallPtrSet<Instruction *, 4> Visited;
@@ -748,498 +549,176 @@ private:
       llvm::tie(UsedI, I) = Uses.pop_back_val();
 
       if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
-        Size = std::max(Size, TD.getTypeStoreSize(LI->getType()));
+        Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
         continue;
       }
       if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
         Value *Op = SI->getOperand(0);
         if (Op == UsedI)
           return SI;
-        Size = std::max(Size, TD.getTypeStoreSize(Op->getType()));
+        Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
         continue;
       }
 
       if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
         if (!GEP->hasAllZeroIndices())
-          return GEP;
-      } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
-                 !isa<SelectInst>(I)) {
-        return I;
-      }
-
-      for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
-           ++UI)
-        if (Visited.insert(cast<Instruction>(*UI)))
-          Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
-    } while (!Uses.empty());
-
-    return 0;
-  }
-
-  bool visitPHINode(PHINode &PN) {
-    // See if we already have computed info on this node.
-    std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
-    if (PHIInfo.first) {
-      PHIInfo.second = true;
-      insertUse(PN, Offset, PHIInfo.first);
-      return true;
-    }
-
-    // Check for an unsafe use of the PHI node.
-    if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
-      return markAsEscaping(*EscapingI);
-
-    insertUse(PN, Offset, PHIInfo.first);
-    return true;
-  }
-
-  bool visitSelectInst(SelectInst &SI) {
-    if (Value *Result = foldSelectInst(SI)) {
-      if (Result == *U)
-        // If the result of the constant fold will be the pointer, recurse
-        // through the select as if we had RAUW'ed it.
-        enqueueUsers(SI, Offset);
-
-      return true;
-    }
-
-    // See if we already have computed info on this node.
-    std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
-    if (SelectInfo.first) {
-      SelectInfo.second = true;
-      insertUse(SI, Offset, SelectInfo.first);
-      return true;
-    }
-
-    // Check for an unsafe use of the PHI node.
-    if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
-      return markAsEscaping(*EscapingI);
-
-    insertUse(SI, Offset, SelectInfo.first);
-    return true;
-  }
-
-  /// \brief Disable SROA entirely if there are unhandled users of the alloca.
-  bool visitInstruction(Instruction &I) { return markAsEscaping(I); }
-};
-
-
-/// \brief Use adder for the alloca partitioning.
-///
-/// This class adds the uses of an alloca to all of the partitions which they
-/// use. For splittable partitions, this can end up doing essentially a linear
-/// walk of the partitions, but the number of steps remains bounded by the
-/// total result instruction size:
-/// - The number of partitions is a result of the number unsplittable
-///   instructions using the alloca.
-/// - The number of users of each partition is at worst the total number of
-///   splittable instructions using the alloca.
-/// Thus we will produce N * M instructions in the end, where N are the number
-/// of unsplittable uses and M are the number of splittable. This visitor does
-/// the exact same number of updates to the partitioning.
-///
-/// In the more common case, this visitor will leverage the fact that the
-/// partition space is pre-sorted, and do a logarithmic search for the
-/// partition needed, making the total visit a classical ((N + M) * log(N))
-/// complexity operation.
-class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> {
-  friend class InstVisitor<UseBuilder>;
-
-  /// \brief Set to de-duplicate dead instructions found in the use walk.
-  SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
-
-public:
-  UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
-      : BuilderBase<UseBuilder>(TD, AI, P) {}
-
-  /// \brief Run the builder over the allocation.
-  void operator()() {
-    // Note that we have to re-evaluate size on each trip through the loop as
-    // the queue grows at the tail.
-    for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
-      U = Queue[Idx].U;
-      Offset = Queue[Idx].Offset;
-      this->visit(cast<Instruction>(U->getUser()));
-    }
-  }
-
-private:
-  void markAsDead(Instruction &I) {
-    if (VisitedDeadInsts.insert(&I))
-      P.DeadUsers.push_back(&I);
-  }
-
-  void insertUse(Instruction &User, int64_t Offset, uint64_t Size) {
-    // If the use has a zero size or extends outside of the allocation, record
-    // it as a dead use for elimination later.
-    if (Size == 0 || (uint64_t)Offset >= AllocSize ||
-        (Offset < 0 && (uint64_t)-Offset >= Size))
-      return markAsDead(User);
-
-    // Clamp the start to the beginning of the allocation.
-    if (Offset < 0) {
-      Size -= (uint64_t)-Offset;
-      Offset = 0;
-    }
-
-    uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
-
-    // Clamp the end offset to the end of the allocation. Note that this is
-    // formulated to handle even the case where "BeginOffset + Size" overflows.
-    assert(AllocSize >= BeginOffset); // Established above.
-    if (Size > AllocSize - BeginOffset)
-      EndOffset = AllocSize;
-
-    // NB: This only works if we have zero overlapping partitions.
-    iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
-    if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
-      B = llvm::prior(B);
-    for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
-         ++I) {
-      PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
-                         std::min(I->EndOffset, EndOffset), U);
-      P.use_push_back(I, NewPU);
-      if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
-        P.PHIOrSelectOpMap[U]
-          = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
-    }
-  }
-
-  void handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
-    uint64_t Size = TD.getTypeStoreSize(Ty);
-
-    // If this memory access can be shown to *statically* extend outside the
-    // bounds of of the allocation, it's behavior is undefined, so simply
-    // ignore it. Note that this is more strict than the generic clamping
-    // behavior of insertUse.
-    if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
-        Size > (AllocSize - (uint64_t)Offset))
-      return markAsDead(I);
-
-    insertUse(I, Offset, Size);
-  }
-
-  void visitBitCastInst(BitCastInst &BC) {
-    if (BC.use_empty())
-      return markAsDead(BC);
-
-    enqueueUsers(BC, Offset);
-  }
-
-  void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
-    if (GEPI.use_empty())
-      return markAsDead(GEPI);
-
-    int64_t GEPOffset;
-    if (!computeConstantGEPOffset(GEPI, GEPOffset))
-      llvm_unreachable("Unable to compute constant offset for use");
-
-    enqueueUsers(GEPI, GEPOffset);
-  }
-
-  void visitLoadInst(LoadInst &LI) {
-    handleLoadOrStore(LI.getType(), LI, Offset);
-  }
-
-  void visitStoreInst(StoreInst &SI) {
-    handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
-  }
-
-  void visitMemSetInst(MemSetInst &II) {
-    ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
-    uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
-    insertUse(II, Offset, Size);
-  }
-
-  void visitMemTransferInst(MemTransferInst &II) {
-    ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
-    uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
-    if (!Size)
-      return markAsDead(II);
+          return GEP;
+      } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
+                 !isa<SelectInst>(I)) {
+        return I;
+      }
 
-    MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
-    if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd &&
-        Offsets.DestBegin == Offsets.SourceBegin)
-      return markAsDead(II); // Skip identity transfers without side-effects.
+      for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
+           ++UI)
+        if (Visited.insert(cast<Instruction>(*UI)))
+          Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
+    } while (!Uses.empty());
 
-    insertUse(II, Offset, Size);
+    return 0;
   }
 
-  void visitIntrinsicInst(IntrinsicInst &II) {
-    assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
-           II.getIntrinsicID() == Intrinsic::lifetime_end);
-
-    ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
-    insertUse(II, Offset,
-              std::min(AllocSize - Offset, Length->getLimitedValue()));
-  }
+  void visitPHINode(PHINode &PN) {
+    if (PN.use_empty())
+      return markAsDead(PN);
+    if (!IsOffsetKnown)
+      return PI.setAborted(&PN);
 
-  void insertPHIOrSelect(Instruction &User, uint64_t Offset) {
-    uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
+    // See if we already have computed info on this node.
+    uint64_t &PHISize = PHIOrSelectSizes[&PN];
+    if (!PHISize) {
+      // This is a new PHI node, check for an unsafe use of the PHI node.
+      if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHISize))
+        return PI.setAborted(UnsafeI);
+    }
 
     // For PHI and select operands outside the alloca, we can't nuke the entire
     // phi or select -- the other side might still be relevant, so we special
     // case them here and use a separate structure to track the operands
     // themselves which should be replaced with undef.
-    if (Offset >= AllocSize) {
-      P.DeadOperands.push_back(U);
+    // FIXME: This should instead be escaped in the event we're instrumenting
+    // for address sanitization.
+    if ((Offset.isNegative() && (-Offset).uge(PHISize)) ||
+        (!Offset.isNegative() && Offset.uge(AllocSize))) {
+      S.DeadOperands.push_back(U);
       return;
     }
 
-    insertUse(User, Offset, Size);
+    insertUse(PN, Offset, PHISize);
   }
-  void visitPHINode(PHINode &PN) {
-    if (PN.use_empty())
-      return markAsDead(PN);
 
-    insertPHIOrSelect(PN, Offset);
-  }
   void visitSelectInst(SelectInst &SI) {
     if (SI.use_empty())
       return markAsDead(SI);
-
     if (Value *Result = foldSelectInst(SI)) {
       if (Result == *U)
         // If the result of the constant fold will be the pointer, recurse
         // through the select as if we had RAUW'ed it.
-        enqueueUsers(SI, Offset);
+        enqueueUsers(SI);
       else
         // Otherwise the operand to the select is dead, and we can replace it
         // with undef.
-        P.DeadOperands.push_back(U);
+        S.DeadOperands.push_back(U);
 
       return;
     }
+    if (!IsOffsetKnown)
+      return PI.setAborted(&SI);
 
-    insertPHIOrSelect(SI, Offset);
-  }
-
-  /// \brief Unreachable, we've already visited the alloca once.
-  void visitInstruction(Instruction &I) {
-    llvm_unreachable("Unhandled instruction in use builder.");
-  }
-};
-
-void AllocaPartitioning::splitAndMergePartitions() {
-  size_t NumDeadPartitions = 0;
-
-  // Track the range of splittable partitions that we pass when accumulating
-  // overlapping unsplittable partitions.
-  uint64_t SplitEndOffset = 0ull;
-
-  Partition New(0ull, 0ull, false);
-
-  for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
-    ++j;
-
-    if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
-      assert(New.BeginOffset == New.EndOffset);
-      New = Partitions[i];
-    } else {
-      assert(New.IsSplittable);
-      New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
-    }
-    assert(New.BeginOffset != New.EndOffset);
-
-    // Scan the overlapping partitions.
-    while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
-      // If the new partition we are forming is splittable, stop at the first
-      // unsplittable partition.
-      if (New.IsSplittable && !Partitions[j].IsSplittable)
-        break;
-
-      // Grow the new partition to include any equally splittable range. 'j' is
-      // always equally splittable when New is splittable, but when New is not
-      // splittable, we may subsume some (or part of some) splitable partition
-      // without growing the new one.
-      if (New.IsSplittable == Partitions[j].IsSplittable) {
-        New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
-      } else {
-        assert(!New.IsSplittable);
-        assert(Partitions[j].IsSplittable);
-        SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
-      }
-
-      Partitions[j].kill();
-      ++NumDeadPartitions;
-      ++j;
-    }
-
-    // If the new partition is splittable, chop off the end as soon as the
-    // unsplittable subsequent partition starts and ensure we eventually cover
-    // the splittable area.
-    if (j != e && New.IsSplittable) {
-      SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
-      New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
+    // See if we already have computed info on this node.
+    uint64_t &SelectSize = PHIOrSelectSizes[&SI];
+    if (!SelectSize) {
+      // This is a new Select, check for an unsafe use of it.
+      if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectSize))
+        return PI.setAborted(UnsafeI);
     }
 
-    // Add the new partition if it differs from the original one and is
-    // non-empty. We can end up with an empty partition here if it was
-    // splittable but there is an unsplittable one that starts at the same
-    // offset.
-    if (New != Partitions[i]) {
-      if (New.BeginOffset != New.EndOffset)
-        Partitions.push_back(New);
-      // Mark the old one for removal.
-      Partitions[i].kill();
-      ++NumDeadPartitions;
+    // For PHI and select operands outside the alloca, we can't nuke the entire
+    // phi or select -- the other side might still be relevant, so we special
+    // case them here and use a separate structure to track the operands
+    // themselves which should be replaced with undef.
+    // FIXME: This should instead be escaped in the event we're instrumenting
+    // for address sanitization.
+    if ((Offset.isNegative() && Offset.uge(SelectSize)) ||
+        (!Offset.isNegative() && Offset.uge(AllocSize))) {
+      S.DeadOperands.push_back(U);
+      return;
     }
 
-    New.BeginOffset = New.EndOffset;
-    if (!New.IsSplittable) {
-      New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
-      if (j != e && !Partitions[j].IsSplittable)
-        New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
-      New.IsSplittable = true;
-      // If there is a trailing splittable partition which won't be fused into
-      // the next splittable partition go ahead and add it onto the partitions
-      // list.
-      if (New.BeginOffset < New.EndOffset &&
-          (j == e || !Partitions[j].IsSplittable ||
-           New.EndOffset < Partitions[j].BeginOffset)) {
-        Partitions.push_back(New);
-        New.BeginOffset = New.EndOffset = 0ull;
-      }
-    }
+    insertUse(SI, Offset, SelectSize);
   }
 
-  // Re-sort the partitions now that they have been split and merged into
-  // disjoint set of partitions. Also remove any of the dead partitions we've
-  // replaced in the process.
-  std::sort(Partitions.begin(), Partitions.end());
-  if (NumDeadPartitions) {
-    assert(Partitions.back().isDead());
-    assert((ptrdiff_t)NumDeadPartitions ==
-           std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
+  /// \brief Disable SROA entirely if there are unhandled users of the alloca.
+  void visitInstruction(Instruction &I) {
+    PI.setAborted(&I);
   }
-  Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
-}
+};
 
-AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI)
+AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
     :
-#ifndef NDEBUG
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
       AI(AI),
 #endif
       PointerEscapingInstr(0) {
-  PartitionBuilder PB(TD, AI, *this);
-  if (!PB())
+  SliceBuilder PB(DL, AI, *this);
+  SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
+  if (PtrI.isEscaped() || PtrI.isAborted()) {
+    // FIXME: We should sink the escape vs. abort info into the caller nicely,
+    // possibly by just storing the PtrInfo in the AllocaSlices.
+    PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
+                                                  : PtrI.getAbortingInst();
+    assert(PointerEscapingInstr && "Did not track a bad instruction");
     return;
+  }
+
+  Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
+                              std::mem_fun_ref(&Slice::isDead)),
+               Slices.end());
 
   // Sort the uses. This arranges for the offsets to be in ascending order,
   // and the sizes to be in descending order.
-  std::sort(Partitions.begin(), Partitions.end());
-
-  // Remove any partitions from the back which are marked as dead.
-  while (!Partitions.empty() && Partitions.back().isDead())
-    Partitions.pop_back();
-
-  if (Partitions.size() > 1) {
-    // Intersect splittability for all partitions with equal offsets and sizes.
-    // Then remove all but the first so that we have a sequence of non-equal but
-    // potentially overlapping partitions.
-    for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
-         I = J) {
-      ++J;
-      while (J != E && *I == *J) {
-        I->IsSplittable &= J->IsSplittable;
-        ++J;
-      }
-    }
-    Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
-                     Partitions.end());
-
-    // Split splittable and merge unsplittable partitions into a disjoint set
-    // of partitions over the used space of the allocation.
-    splitAndMergePartitions();
-  }
-
-  // Now build up the user lists for each of these disjoint partitions by
-  // re-walking the recursive users of the alloca.
-  Uses.resize(Partitions.size());
-  UseBuilder UB(TD, AI, *this);
-  UB();
+  std::sort(Slices.begin(), Slices.end());
 }
 
-Type *AllocaPartitioning::getCommonType(iterator I) const {
-  Type *Ty = 0;
-  for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
-    if (!UI->U)
-      continue; // Skip dead uses.
-    if (isa<IntrinsicInst>(*UI->U->getUser()))
-      continue;
-    if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
-      continue;
-
-    Type *UserTy = 0;
-    if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) {
-      UserTy = LI->getType();
-    } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) {
-      UserTy = SI->getValueOperand()->getType();
-    }
-
-    if (Ty && Ty != UserTy)
-      return 0;
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
 
-    Ty = UserTy;
-  }
-  return Ty;
+void AllocaSlices::print(raw_ostream &OS, const_iterator I,
+                         StringRef Indent) const {
+  printSlice(OS, I, Indent);
+  printUse(OS, I, Indent);
 }
 
-#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
-
-void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
-                               StringRef Indent) const {
-  OS << Indent << "partition #" << (I - begin())
-     << " [" << I->BeginOffset << "," << I->EndOffset << ")"
-     << (I->IsSplittable ? " (splittable)" : "")
-     << (Uses[I - begin()].empty() ? " (zero uses)" : "")
-     << "\n";
+void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
+                              StringRef Indent) const {
+  OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
+     << " slice #" << (I - begin())
+     << (I->isSplittable() ? " (splittable)" : "") << "\n";
 }
 
-void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
-                                    StringRef Indent) const {
-  for (const_use_iterator UI = use_begin(I), UE = use_end(I);
-       UI != UE; ++UI) {
-    if (!UI->U)
-      continue; // Skip dead uses.
-    OS << Indent << "  [" << UI->BeginOffset << "," << UI->EndOffset << ") "
-       << "used by: " << *UI->U->getUser() << "\n";
-    if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
-      const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
-      bool IsDest;
-      if (!MTO.IsSplittable)
-        IsDest = UI->BeginOffset == MTO.DestBegin;
-      else
-        IsDest = MTO.DestBegin != 0u;
-      OS << Indent << "    (original " << (IsDest ? "dest" : "source") << ": "
-         << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
-         << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
-    }
-  }
+void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
+                            StringRef Indent) const {
+  OS << Indent << "  used by: " << *I->getUse()->getUser() << "\n";
 }
 
-void AllocaPartitioning::print(raw_ostream &OS) const {
+void AllocaSlices::print(raw_ostream &OS) const {
   if (PointerEscapingInstr) {
-    OS << "No partitioning for alloca: " << AI << "\n"
+    OS << "Can't analyze slices for alloca: " << AI << "\n"
        << "  A pointer to this alloca escaped by:\n"
        << "  " << *PointerEscapingInstr << "\n";
     return;
   }
 
-  OS << "Partitioning of alloca: " << AI << "\n";
-  unsigned Num = 0;
-  for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
+  OS << "Slices of alloca: " << AI << "\n";
+  for (const_iterator I = begin(), E = end(); I != E; ++I)
     print(OS, I);
-    printUsers(OS, I);
-  }
 }
 
-void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
-void AllocaPartitioning::dump() const { print(dbgs()); }
+LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
+  print(dbgs(), I);
+}
+LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
 
 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
 
-
 namespace {
 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
 ///
@@ -1256,12 +735,13 @@ class AllocaPromoter : public LoadAndStorePromoter {
   SmallVector<DbgValueInst *, 4> DVIs;
 
 public:
-  AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
+  AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
                  AllocaInst &AI, DIBuilder &DIB)
-    : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
+      : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
 
   void run(const SmallVectorImpl<Instruction*> &Insts) {
-    // Remember which alloca we're promoting (for isInstInList).
+    // Retain the debug information attached to the alloca for use when
+    // rewriting loads and stores.
     if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
       for (Value::use_iterator UI = DebugNode->use_begin(),
                                UE = DebugNode->use_end();
@@ -1273,7 +753,9 @@ public:
     }
 
     LoadAndStorePromoter::run(Insts);
-    AI.eraseFromParent();
+
+    // While we have the debug information, clear it off of the alloca. The
+    // caller takes care of deleting the alloca.
     while (!DDIs.empty())
       DDIs.pop_back_val()->eraseFromParent();
     while (!DVIs.empty())
@@ -1282,13 +764,34 @@ public:
 
   virtual bool isInstInList(Instruction *I,
                             const SmallVectorImpl<Instruction*> &Insts) const {
+    Value *Ptr;
     if (LoadInst *LI = dyn_cast<LoadInst>(I))
-      return LI->getOperand(0) == &AI;
-    return cast<StoreInst>(I)->getPointerOperand() == &AI;
+      Ptr = LI->getOperand(0);
+    else
+      Ptr = cast<StoreInst>(I)->getPointerOperand();
+
+    // Only used to detect cycles, which will be rare and quickly found as
+    // we're walking up a chain of defs rather than down through uses.
+    SmallPtrSet<Value *, 4> Visited;
+
+    do {
+      if (Ptr == &AI)
+        return true;
+
+      if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
+        Ptr = BCI->getOperand(0);
+      else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
+        Ptr = GEPI->getPointerOperand();
+      else
+        return false;
+
+    } while (Visited.insert(Ptr));
+
+    return false;
   }
 
   virtual void updateDebugInfo(Instruction *Inst) const {
-    for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
+    for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
            E = DDIs.end(); I != E; ++I) {
       DbgDeclareInst *DDI = *I;
       if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
@@ -1296,21 +799,21 @@ public:
       else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
         ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
     }
-    for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
+    for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
            E = DVIs.end(); I != E; ++I) {
       DbgValueInst *DVI = *I;
-      Value *Arg = NULL;
+      Value *Arg = 0;
       if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
         // If an argument is zero extended then use argument directly. The ZExt
         // may be zapped by an optimization pass in future.
         if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
           Arg = dyn_cast<Argument>(ZExt->getOperand(0));
-        if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
+        else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
           Arg = dyn_cast<Argument>(SExt->getOperand(0));
         if (!Arg)
-          Arg = SI->getOperand(0);
+          Arg = SI->getValueOperand();
       } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
-        Arg = LI->getOperand(0);
+        Arg = LI->getPointerOperand();
       } else {
         continue;
       }
@@ -1334,7 +837,7 @@ namespace {
 /// 1) It takes allocations of aggregates and analyzes the ways in which they
 ///    are used to try to split them into smaller allocations, ideally of
 ///    a single scalar data type. It will split up memcpy and memset accesses
-///    as necessary and try to isolate invidual scalar accesses.
+///    as necessary and try to isolate individual scalar accesses.
 /// 2) It will transform accesses into forms which are suitable for SSA value
 ///    promotion. This can be replacing a memset with a scalar store of an
 ///    integer value, or it can involve speculating operations on a PHI or
@@ -1347,7 +850,7 @@ class SROA : public FunctionPass {
   const bool RequiresDomTree;
 
   LLVMContext *C;
-  const DataLayout *TD;
+  const DataLayout *DL;
   DominatorTree *DT;
 
   /// \brief Worklist of alloca instructions to simplify.
@@ -1362,11 +865,7 @@ 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.
   ///
@@ -1381,10 +880,25 @@ class SROA : public FunctionPass {
   /// \brief A collection of alloca instructions we can directly promote.
   std::vector<AllocaInst *> PromotableAllocas;
 
+  /// \brief A worklist of PHIs to speculate prior to promoting allocas.
+  ///
+  /// All of these PHIs have been checked for the safety of speculation and by
+  /// being speculated will allow promoting allocas currently in the promotable
+  /// queue.
+  SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
+
+  /// \brief A worklist of select instructions to speculate prior to promoting
+  /// allocas.
+  ///
+  /// All of these select instructions have been checked for the safety of
+  /// speculation and by being speculated will allow promoting allocas
+  /// currently in the promotable queue.
+  SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
+
 public:
   SROA(bool RequiresDomTree = true)
       : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
-        C(0), TD(0), DT(0) {
+        C(0), DL(0), DT(0) {
     initializeSROAPass(*PassRegistry::getPassRegistry());
   }
   bool runOnFunction(Function &F);
@@ -1395,13 +909,13 @@ public:
 
 private:
   friend class PHIOrSelectSpeculator;
-  friend class AllocaPartitionRewriter;
-  friend class AllocaPartitionVectorRewriter;
+  friend class AllocaSliceRewriter;
 
-  bool rewriteAllocaPartition(AllocaInst &AI,
-                              AllocaPartitioning &P,
-                              AllocaPartitioning::iterator PI);
-  bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
+  bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
+                        AllocaSlices::iterator B, AllocaSlices::iterator E,
+                        int64_t BeginOffset, int64_t EndOffset,
+                        ArrayRef<AllocaSlices::iterator> SplitUses);
+  bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
   bool runOnAlloca(AllocaInst &AI);
   void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
   bool promoteAllocas(Function &F);
@@ -1420,334 +934,263 @@ INITIALIZE_PASS_DEPENDENCY(DominatorTree)
 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
                     false, false)
 
-namespace {
-/// \brief Visitor to speculate PHIs and Selects where possible.
-class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
-  // Befriend the base class so it can delegate to private visit methods.
-  friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
-
-  const DataLayout &TD;
-  AllocaPartitioning &P;
-  SROA &Pass;
+/// Walk the range of a partitioning looking for a common type to cover this
+/// sequence of slices.
+static Type *findCommonType(AllocaSlices::const_iterator B,
+                            AllocaSlices::const_iterator E,
+                            uint64_t EndOffset) {
+  Type *Ty = 0;
+  bool IgnoreNonIntegralTypes = false;
+  for (AllocaSlices::const_iterator I = B; I != E; ++I) {
+    Use *U = I->getUse();
+    if (isa<IntrinsicInst>(*U->getUser()))
+      continue;
+    if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
+      continue;
 
-public:
-  PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass)
-    : TD(TD), P(P), Pass(Pass) {}
-
-  /// \brief Visit the users of an alloca partition and rewrite them.
-  void visitUsers(AllocaPartitioning::const_iterator PI) {
-    // Note that we need to use an index here as the underlying vector of uses
-    // may be grown during speculation. However, we never need to re-visit the
-    // new uses, and so we can use the initial size bound.
-    for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
-      const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
-      if (!PU.U)
-        continue; // Skip dead use.
-
-      visit(cast<Instruction>(PU.U->getUser()));
+    Type *UserTy = 0;
+    if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
+      UserTy = LI->getType();
+    } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
+      UserTy = SI->getValueOperand()->getType();
+    } else {
+      IgnoreNonIntegralTypes = true; // Give up on anything but an iN type.
+      continue;
     }
-  }
 
-private:
-  // By default, skip this instruction.
-  void visitInstruction(Instruction &I) {}
-
-  /// PHI instructions that use an alloca and are subsequently loaded can be
-  /// rewritten to load both input pointers in the pred blocks and then PHI the
-  /// results, allowing the load of the alloca to be promoted.
-  /// From this:
-  ///   %P2 = phi [i32* %Alloca, i32* %Other]
-  ///   %V = load i32* %P2
-  /// to:
-  ///   %V1 = load i32* %Alloca      -> will be mem2reg'd
-  ///   ...
-  ///   %V2 = load i32* %Other
-  ///   ...
-  ///   %V = phi [i32 %V1, i32 %V2]
-  ///
-  /// We can do this to a select if its only uses are loads and if the operands
-  /// to the select can be loaded unconditionally.
-  ///
-  /// FIXME: This should be hoisted into a generic utility, likely in
-  /// Transforms/Util/Local.h
-  bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
-    // For now, we can only do this promotion if the load is in the same block
-    // as the PHI, and if there are no stores between the phi and load.
-    // TODO: Allow recursive phi users.
-    // TODO: Allow stores.
-    BasicBlock *BB = PN.getParent();
-    unsigned MaxAlign = 0;
-    for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
-         UI != UE; ++UI) {
-      LoadInst *LI = dyn_cast<LoadInst>(*UI);
-      if (LI == 0 || !LI->isSimple()) return false;
-
-      // For now we only allow loads in the same block as the PHI.  This is
-      // a common case that happens when instcombine merges two loads through
-      // a PHI.
-      if (LI->getParent() != BB) return false;
-
-      // Ensure that there are no instructions between the PHI and the load that
-      // could store.
-      for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
-        if (BBI->mayWriteToMemory())
-          return false;
-
-      MaxAlign = std::max(MaxAlign, LI->getAlignment());
-      Loads.push_back(LI);
+    if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
+      // If the type is larger than the partition, skip it. We only encounter
+      // this for split integer operations where we want to use the type of the
+      // entity causing the split. Also skip if the type is not a byte width
+      // multiple.
+      if (ITy->getBitWidth() % 8 != 0 ||
+          ITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
+        continue;
+
+      // If we have found an integer type use covering the alloca, use that
+      // regardless of the other types, as integers are often used for
+      // a "bucket of bits" type.
+      //
+      // NB: This *must* be the only return from inside the loop so that the
+      // order of slices doesn't impact the computed type.
+      return ITy;
+    } else if (IgnoreNonIntegralTypes) {
+      continue;
     }
 
-    // We can only transform this if it is safe to push the loads into the
-    // predecessor blocks. The only thing to watch out for is that we can't put
-    // a possibly trapping load in the predecessor if it is a critical edge.
-    for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
-         ++Idx) {
-      TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
-      Value *InVal = PN.getIncomingValue(Idx);
-
-      // If the value is produced by the terminator of the predecessor (an
-      // invoke) or it has side-effects, there is no valid place to put a load
-      // in the predecessor.
-      if (TI == InVal || TI->mayHaveSideEffects())
-        return false;
+    if (Ty && Ty != UserTy)
+      IgnoreNonIntegralTypes = true; // Give up on anything but an iN type.
 
-      // If the predecessor has a single successor, then the edge isn't
-      // critical.
-      if (TI->getNumSuccessors() == 1)
-        continue;
+    Ty = UserTy;
+  }
+  return Ty;
+}
 
-      // If this pointer is always safe to load, or if we can prove that there
-      // is already a load in the block, then we can move the load to the pred
-      // block.
-      if (InVal->isDereferenceablePointer() ||
-          isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
-        continue;
+/// PHI instructions that use an alloca and are subsequently loaded can be
+/// rewritten to load both input pointers in the pred blocks and then PHI the
+/// results, allowing the load of the alloca to be promoted.
+/// From this:
+///   %P2 = phi [i32* %Alloca, i32* %Other]
+///   %V = load i32* %P2
+/// to:
+///   %V1 = load i32* %Alloca      -> will be mem2reg'd
+///   ...
+///   %V2 = load i32* %Other
+///   ...
+///   %V = phi [i32 %V1, i32 %V2]
+///
+/// We can do this to a select if its only uses are loads and if the operands
+/// to the select can be loaded unconditionally.
+///
+/// FIXME: This should be hoisted into a generic utility, likely in
+/// Transforms/Util/Local.h
+static bool isSafePHIToSpeculate(PHINode &PN,
+                                 const DataLayout *DL = 0) {
+  // For now, we can only do this promotion if the load is in the same block
+  // as the PHI, and if there are no stores between the phi and load.
+  // TODO: Allow recursive phi users.
+  // TODO: Allow stores.
+  BasicBlock *BB = PN.getParent();
+  unsigned MaxAlign = 0;
+  bool HaveLoad = false;
+  for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end(); UI != UE;
+       ++UI) {
+    LoadInst *LI = dyn_cast<LoadInst>(*UI);
+    if (LI == 0 || !LI->isSimple())
+      return false;
 
+    // For now we only allow loads in the same block as the PHI.  This is
+    // a common case that happens when instcombine merges two loads through
+    // a PHI.
+    if (LI->getParent() != BB)
       return false;
-    }
 
-    return true;
+    // Ensure that there are no instructions between the PHI and the load that
+    // could store.
+    for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
+      if (BBI->mayWriteToMemory())
+        return false;
+
+    MaxAlign = std::max(MaxAlign, LI->getAlignment());
+    HaveLoad = true;
   }
 
-  void visitPHINode(PHINode &PN) {
-    DEBUG(dbgs() << "    original: " << PN << "\n");
+  if (!HaveLoad)
+    return false;
 
-    SmallVector<LoadInst *, 4> Loads;
-    if (!isSafePHIToSpeculate(PN, Loads))
-      return;
+  // We can only transform this if it is safe to push the loads into the
+  // predecessor blocks. The only thing to watch out for is that we can't put
+  // a possibly trapping load in the predecessor if it is a critical edge.
+  for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
+    TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
+    Value *InVal = PN.getIncomingValue(Idx);
+
+    // If the value is produced by the terminator of the predecessor (an
+    // invoke) or it has side-effects, there is no valid place to put a load
+    // in the predecessor.
+    if (TI == InVal || TI->mayHaveSideEffects())
+      return false;
 
-    assert(!Loads.empty());
+    // If the predecessor has a single successor, then the edge isn't
+    // critical.
+    if (TI->getNumSuccessors() == 1)
+      continue;
 
-    Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
-    IRBuilder<> PHIBuilder(&PN);
-    PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
-                                          PN.getName() + ".sroa.speculated");
+    // If this pointer is always safe to load, or if we can prove that there
+    // is already a load in the block, then we can move the load to the pred
+    // block.
+    if (InVal->isDereferenceablePointer() ||
+        isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
+      continue;
 
-    // Get the TBAA tag and alignment to use from one of the loads.  It doesn't
-    // matter which one we get and if any differ, it doesn't matter.
-    LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
-    MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
-    unsigned Align = SomeLoad->getAlignment();
+    return false;
+  }
 
-    // Rewrite all loads of the PN to use the new PHI.
-    do {
-      LoadInst *LI = Loads.pop_back_val();
-      LI->replaceAllUsesWith(NewPN);
-      Pass.DeadInsts.push_back(LI);
-    } while (!Loads.empty());
-
-    // Inject loads into all of the pred blocks.
-    for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
-      BasicBlock *Pred = PN.getIncomingBlock(Idx);
-      TerminatorInst *TI = Pred->getTerminator();
-      Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
-      Value *InVal = PN.getIncomingValue(Idx);
-      IRBuilder<> PredBuilder(TI);
-
-      LoadInst *Load
-        = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
-                                         Pred->getName()));
-      ++NumLoadsSpeculated;
-      Load->setAlignment(Align);
-      if (TBAATag)
-        Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
-      NewPN->addIncoming(Load, Pred);
-
-      Instruction *Ptr = dyn_cast<Instruction>(InVal);
-      if (!Ptr)
-        // No uses to rewrite.
-        continue;
+  return true;
+}
 
-      // Try to lookup and rewrite any partition uses corresponding to this phi
-      // input.
-      AllocaPartitioning::iterator PI
-        = P.findPartitionForPHIOrSelectOperand(InUse);
-      if (PI == P.end())
-        continue;
+static void speculatePHINodeLoads(PHINode &PN) {
+  DEBUG(dbgs() << "    original: " << PN << "\n");
+
+  Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
+  IRBuilderTy PHIBuilder(&PN);
+  PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
+                                        PN.getName() + ".sroa.speculated");
+
+  // Get the TBAA tag and alignment to use from one of the loads.  It doesn't
+  // matter which one we get and if any differ.
+  LoadInst *SomeLoad = cast<LoadInst>(*PN.use_begin());
+  MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
+  unsigned Align = SomeLoad->getAlignment();
+
+  // Rewrite all loads of the PN to use the new PHI.
+  while (!PN.use_empty()) {
+    LoadInst *LI = cast<LoadInst>(*PN.use_begin());
+    LI->replaceAllUsesWith(NewPN);
+    LI->eraseFromParent();
+  }
+
+  // Inject loads into all of the pred blocks.
+  for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
+    BasicBlock *Pred = PN.getIncomingBlock(Idx);
+    TerminatorInst *TI = Pred->getTerminator();
+    Value *InVal = PN.getIncomingValue(Idx);
+    IRBuilderTy PredBuilder(TI);
+
+    LoadInst *Load = PredBuilder.CreateLoad(
+        InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
+    ++NumLoadsSpeculated;
+    Load->setAlignment(Align);
+    if (TBAATag)
+      Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
+    NewPN->addIncoming(Load, Pred);
+  }
+
+  DEBUG(dbgs() << "          speculated to: " << *NewPN << "\n");
+  PN.eraseFromParent();
+}
 
-      // Replace the Use in the PartitionUse for this operand with the Use
-      // inside the load.
-      AllocaPartitioning::use_iterator UI
-        = P.findPartitionUseForPHIOrSelectOperand(InUse);
-      assert(isa<PHINode>(*UI->U->getUser()));
-      UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
-    }
-    DEBUG(dbgs() << "          speculated to: " << *NewPN << "\n");
-  }
-
-  /// Select instructions that use an alloca and are subsequently loaded can be
-  /// rewritten to load both input pointers and then select between the result,
-  /// allowing the load of the alloca to be promoted.
-  /// From this:
-  ///   %P2 = select i1 %cond, i32* %Alloca, i32* %Other
-  ///   %V = load i32* %P2
-  /// to:
-  ///   %V1 = load i32* %Alloca      -> will be mem2reg'd
-  ///   %V2 = load i32* %Other
-  ///   %V = select i1 %cond, i32 %V1, i32 %V2
-  ///
-  /// We can do this to a select if its only uses are loads and if the operand
-  /// to the select can be loaded unconditionally.
-  bool isSafeSelectToSpeculate(SelectInst &SI,
-                               SmallVectorImpl<LoadInst *> &Loads) {
-    Value *TValue = SI.getTrueValue();
-    Value *FValue = SI.getFalseValue();
-    bool TDerefable = TValue->isDereferenceablePointer();
-    bool FDerefable = FValue->isDereferenceablePointer();
-
-    for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
-         UI != UE; ++UI) {
-      LoadInst *LI = dyn_cast<LoadInst>(*UI);
-      if (LI == 0 || !LI->isSimple()) return false;
-
-      // Both operands to the select need to be dereferencable, either
-      // absolutely (e.g. allocas) or at this point because we can see other
-      // accesses to it.
-      if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
-                                                      LI->getAlignment(), &TD))
-        return false;
-      if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
-                                                      LI->getAlignment(), &TD))
-        return false;
-      Loads.push_back(LI);
-    }
+/// Select instructions that use an alloca and are subsequently loaded can be
+/// rewritten to load both input pointers and then select between the result,
+/// allowing the load of the alloca to be promoted.
+/// From this:
+///   %P2 = select i1 %cond, i32* %Alloca, i32* %Other
+///   %V = load i32* %P2
+/// to:
+///   %V1 = load i32* %Alloca      -> will be mem2reg'd
+///   %V2 = load i32* %Other
+///   %V = select i1 %cond, i32 %V1, i32 %V2
+///
+/// We can do this to a select if its only uses are loads and if the operand
+/// to the select can be loaded unconditionally.
+static bool isSafeSelectToSpeculate(SelectInst &SI, const DataLayout *DL = 0) {
+  Value *TValue = SI.getTrueValue();
+  Value *FValue = SI.getFalseValue();
+  bool TDerefable = TValue->isDereferenceablePointer();
+  bool FDerefable = FValue->isDereferenceablePointer();
+
+  for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end(); UI != UE;
+       ++UI) {
+    LoadInst *LI = dyn_cast<LoadInst>(*UI);
+    if (LI == 0 || !LI->isSimple())
+      return false;
 
-    return true;
+    // Both operands to the select need to be dereferencable, either
+    // absolutely (e.g. allocas) or at this point because we can see other
+    // accesses to it.
+    if (!TDerefable &&
+        !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
+      return false;
+    if (!FDerefable &&
+        !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
+      return false;
   }
 
-  void visitSelectInst(SelectInst &SI) {
-    DEBUG(dbgs() << "    original: " << SI << "\n");
-    IRBuilder<> IRB(&SI);
-
-    // If the select isn't safe to speculate, just use simple logic to emit it.
-    SmallVector<LoadInst *, 4> Loads;
-    if (!isSafeSelectToSpeculate(SI, Loads))
-      return;
+  return true;
+}
 
-    Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
-    AllocaPartitioning::iterator PIs[2];
-    AllocaPartitioning::PartitionUse PUs[2];
-    for (unsigned i = 0, e = 2; i != e; ++i) {
-      PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
-      if (PIs[i] != P.end()) {
-        // If the pointer is within the partitioning, remove the select from
-        // its uses. We'll add in the new loads below.
-        AllocaPartitioning::use_iterator UI
-          = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
-        PUs[i] = *UI;
-        // Clear out the use here so that the offsets into the use list remain
-        // stable but this use is ignored when rewriting.
-        UI->U = 0;
-      }
-    }
+static void speculateSelectInstLoads(SelectInst &SI) {
+  DEBUG(dbgs() << "    original: " << SI << "\n");
 
-    Value *TV = SI.getTrueValue();
-    Value *FV = SI.getFalseValue();
-    // Replace the loads of the select with a select of two loads.
-    while (!Loads.empty()) {
-      LoadInst *LI = Loads.pop_back_val();
+  IRBuilderTy IRB(&SI);
+  Value *TV = SI.getTrueValue();
+  Value *FV = SI.getFalseValue();
+  // Replace the loads of the select with a select of two loads.
+  while (!SI.use_empty()) {
+    LoadInst *LI = cast<LoadInst>(*SI.use_begin());
+    assert(LI->isSimple() && "We only speculate simple loads");
 
-      IRB.SetInsertPoint(LI);
-      LoadInst *TL =
+    IRB.SetInsertPoint(LI);
+    LoadInst *TL =
         IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
-      LoadInst *FL =
+    LoadInst *FL =
         IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
-      NumLoadsSpeculated += 2;
-
-      // Transfer alignment and TBAA info if present.
-      TL->setAlignment(LI->getAlignment());
-      FL->setAlignment(LI->getAlignment());
-      if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
-        TL->setMetadata(LLVMContext::MD_tbaa, Tag);
-        FL->setMetadata(LLVMContext::MD_tbaa, Tag);
-      }
-
-      Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
-                                  LI->getName() + ".sroa.speculated");
-
-      LoadInst *Loads[2] = { TL, FL };
-      for (unsigned i = 0, e = 2; i != e; ++i) {
-        if (PIs[i] != P.end()) {
-          Use *LoadUse = &Loads[i]->getOperandUse(0);
-          assert(PUs[i].U->get() == LoadUse->get());
-          PUs[i].U = LoadUse;
-          P.use_push_back(PIs[i], PUs[i]);
-        }
-      }
-
-      DEBUG(dbgs() << "          speculated to: " << *V << "\n");
-      LI->replaceAllUsesWith(V);
-      Pass.DeadInsts.push_back(LI);
-    }
-  }
-};
-}
+    NumLoadsSpeculated += 2;
 
-/// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
-///
-/// If the provided GEP is all-constant, the total byte offset formed by the
-/// GEP is computed and Offset is set to it. If the GEP has any non-constant
-/// operands, the function returns false and the value of Offset is unmodified.
-static bool accumulateGEPOffsets(const DataLayout &TD, GEPOperator &GEP,
-                                 APInt &Offset) {
-  APInt GEPOffset(Offset.getBitWidth(), 0);
-  for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
-       GTI != GTE; ++GTI) {
-    ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
-    if (!OpC)
-      return false;
-    if (OpC->isZero()) continue;
-
-    // Handle a struct index, which adds its field offset to the pointer.
-    if (StructType *STy = dyn_cast<StructType>(*GTI)) {
-      unsigned ElementIdx = OpC->getZExtValue();
-      const StructLayout *SL = TD.getStructLayout(STy);
-      GEPOffset += APInt(Offset.getBitWidth(),
-                         SL->getElementOffset(ElementIdx));
-      continue;
+    // Transfer alignment and TBAA info if present.
+    TL->setAlignment(LI->getAlignment());
+    FL->setAlignment(LI->getAlignment());
+    if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
+      TL->setMetadata(LLVMContext::MD_tbaa, Tag);
+      FL->setMetadata(LLVMContext::MD_tbaa, Tag);
     }
 
-    APInt TypeSize(Offset.getBitWidth(),
-                   TD.getTypeAllocSize(GTI.getIndexedType()));
-    if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
-      assert((VTy->getScalarSizeInBits() % 8) == 0 &&
-             "vector element size is not a multiple of 8, cannot GEP over it");
-      TypeSize = VTy->getScalarSizeInBits() / 8;
-    }
+    Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
+                                LI->getName() + ".sroa.speculated");
 
-    GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
+    DEBUG(dbgs() << "          speculated to: " << *V << "\n");
+    LI->replaceAllUsesWith(V);
+    LI->eraseFromParent();
   }
-  Offset = GEPOffset;
-  return true;
+  SI.eraseFromParent();
 }
 
 /// \brief Build a GEP out of a base pointer and indices.
 ///
 /// This will return the BasePtr if that is valid, or build a new GEP
 /// instruction using the IRBuilder if GEP-ing is needed.
-static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
-                       SmallVectorImpl<Value *> &Indices,
-                       const Twine &Prefix) {
+static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
+                       SmallVectorImpl<Value *> &Indices) {
   if (Indices.empty())
     return BasePtr;
 
@@ -1756,7 +1199,7 @@ static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
   if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
     return BasePtr;
 
-  return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
+  return IRB.CreateInBoundsGEP(BasePtr, Indices, "idx");
 }
 
 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
@@ -1768,12 +1211,11 @@ static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
 /// TargetTy. If we can't find one with the same type, we at least try to use
 /// one with the same size. If none of that works, we just produce the GEP as
 /// indicated by Indices to have the correct offset.
-static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
+static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
                                     Value *BasePtr, Type *Ty, Type *TargetTy,
-                                    SmallVectorImpl<Value *> &Indices,
-                                    const Twine &Prefix) {
+                                    SmallVectorImpl<Value *> &Indices) {
   if (Ty == TargetTy)
-    return buildGEP(IRB, BasePtr, Indices, Prefix);
+    return buildGEP(IRB, BasePtr, Indices);
 
   // See if we can descend into a struct and locate a field with the correct
   // type.
@@ -1784,7 +1226,9 @@ static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
       break;
     if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
       ElementTy = SeqTy->getElementType();
-      Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(), 0)));
+      // Note that we use the default address space as this index is over an
+      // array or a vector, not a pointer.
+      Indices.push_back(IRB.getInt(APInt(DL.getPointerSizeInBits(0), 0)));
     } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
       if (STy->element_begin() == STy->element_end())
         break; // Nothing left to descend into.
@@ -1798,20 +1242,19 @@ static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
   if (ElementTy != TargetTy)
     Indices.erase(Indices.end() - NumLayers, Indices.end());
 
-  return buildGEP(IRB, BasePtr, Indices, Prefix);
+  return buildGEP(IRB, BasePtr, Indices);
 }
 
 /// \brief Recursively compute indices for a natural GEP.
 ///
 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
 /// element types adding appropriate indices for the GEP.
-static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
+static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
                                        Value *Ptr, Type *Ty, APInt &Offset,
                                        Type *TargetTy,
-                                       SmallVectorImpl<Value *> &Indices,
-                                       const Twine &Prefix) {
+                                       SmallVectorImpl<Value *> &Indices) {
   if (Offset == 0)
-    return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
+    return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices);
 
   // We can't recurse through pointer types.
   if (Ty->isPointerTy())
@@ -1821,49 +1264,49 @@ static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
   // extremely poorly defined currently. The long-term goal is to remove GEPing
   // over a vector from the IR completely.
   if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
-    unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
+    unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
     if (ElementSizeInBits % 8)
       return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
     APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
-    APInt NumSkippedElements = Offset.udiv(ElementSize);
+    APInt NumSkippedElements = Offset.sdiv(ElementSize);
     if (NumSkippedElements.ugt(VecTy->getNumElements()))
       return 0;
     Offset -= NumSkippedElements * ElementSize;
     Indices.push_back(IRB.getInt(NumSkippedElements));
-    return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
-                                    Offset, TargetTy, Indices, Prefix);
+    return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
+                                    Offset, TargetTy, Indices);
   }
 
   if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
     Type *ElementTy = ArrTy->getElementType();
-    APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
-    APInt NumSkippedElements = Offset.udiv(ElementSize);
+    APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
+    APInt NumSkippedElements = Offset.sdiv(ElementSize);
     if (NumSkippedElements.ugt(ArrTy->getNumElements()))
       return 0;
 
     Offset -= NumSkippedElements * ElementSize;
     Indices.push_back(IRB.getInt(NumSkippedElements));
-    return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
-                                    Indices, Prefix);
+    return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
+                                    Indices);
   }
 
   StructType *STy = dyn_cast<StructType>(Ty);
   if (!STy)
     return 0;
 
-  const StructLayout *SL = TD.getStructLayout(STy);
+  const StructLayout *SL = DL.getStructLayout(STy);
   uint64_t StructOffset = Offset.getZExtValue();
   if (StructOffset >= SL->getSizeInBytes())
     return 0;
   unsigned Index = SL->getElementContainingOffset(StructOffset);
   Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
   Type *ElementTy = STy->getElementType(Index);
-  if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
+  if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
     return 0; // The offset points into alignment padding.
 
   Indices.push_back(IRB.getInt32(Index));
-  return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
-                                  Indices, Prefix);
+  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
+                                  Indices);
 }
 
 /// \brief Get a natural GEP from a base pointer to a particular offset and
@@ -1876,10 +1319,9 @@ static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
 /// Indices, and setting Ty to the result subtype.
 ///
 /// If no natural GEP can be constructed, this function returns null.
-static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
+static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
                                       Value *Ptr, APInt Offset, Type *TargetTy,
-                                      SmallVectorImpl<Value *> &Indices,
-                                      const Twine &Prefix) {
+                                      SmallVectorImpl<Value *> &Indices) {
   PointerType *Ty = cast<PointerType>(Ptr->getType());
 
   // Don't consider any GEPs through an i8* as natural unless the TargetTy is
@@ -1890,15 +1332,15 @@ static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
   Type *ElementTy = Ty->getElementType();
   if (!ElementTy->isSized())
     return 0; // We can't GEP through an unsized element.
-  APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
+  APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
   if (ElementSize == 0)
     return 0; // Zero-length arrays can't help us build a natural GEP.
-  APInt NumSkippedElements = Offset.udiv(ElementSize);
+  APInt NumSkippedElements = Offset.sdiv(ElementSize);
 
   Offset -= NumSkippedElements * ElementSize;
   Indices.push_back(IRB.getInt(NumSkippedElements));
-  return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
-                                  Indices, Prefix);
+  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
+                                  Indices);
 }
 
 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
@@ -1913,12 +1355,11 @@ static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
 /// The strategy for finding the more natural GEPs is to peel off layers of the
 /// pointer, walking back through bit casts and GEPs, searching for a base
 /// pointer from which we can compute a natural GEP with the desired
-/// properities. The algorithm tries to fold as many constant indices into
+/// properties. The algorithm tries to fold as many constant indices into
 /// a single GEP as possible, thus making each GEP more independent of the
 /// surrounding code.
-static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
-                             Value *Ptr, APInt Offset, Type *PointerTy,
-                             const Twine &Prefix) {
+static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL,
+                             Value *Ptr, APInt Offset, Type *PointerTy) {
   // Even though we don't look through PHI nodes, we could be called on an
   // instruction in an unreachable block, which may be on a cycle.
   SmallPtrSet<Value *, 4> Visited;
@@ -1941,7 +1382,7 @@ static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
     // First fold any existing GEPs into the offset.
     while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
       APInt GEPOffset(Offset.getBitWidth(), 0);
-      if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
+      if (!GEP->accumulateConstantOffset(DL, GEPOffset))
         break;
       Offset += GEPOffset;
       Ptr = GEP->getPointerOperand();
@@ -1951,8 +1392,8 @@ static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
 
     // See if we can perform a natural GEP here.
     Indices.clear();
-    if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
-                                           Indices, Prefix)) {
+    if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
+                                           Indices)) {
       if (P->getType() == PointerTy) {
         // Zap any offset pointer that we ended up computing in previous rounds.
         if (OffsetPtr && OffsetPtr->use_empty())
@@ -1987,24 +1428,183 @@ static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
   if (!OffsetPtr) {
     if (!Int8Ptr) {
       Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
-                                  Prefix + ".raw_cast");
+                                  "raw_cast");
       Int8PtrOffset = Offset;
     }
 
     OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
       IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
-                            Prefix + ".raw_idx");
+                            "raw_idx");
   }
   Ptr = OffsetPtr;
 
   // On the off chance we were targeting i8*, guard the bitcast here.
   if (Ptr->getType() != PointerTy)
-    Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
+    Ptr = IRB.CreateBitCast(Ptr, PointerTy, "cast");
 
   return Ptr;
 }
 
-/// \brief Test whether the given alloca partition can be promoted to a vector.
+/// \brief Test whether we can convert a value from the old to the new type.
+///
+/// This predicate should be used to guard calls to convertValue in order to
+/// ensure that we only try to convert viable values. The strategy is that we
+/// will peel off single element struct and array wrappings to get to an
+/// underlying value, and convert that value.
+static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
+  if (OldTy == NewTy)
+    return true;
+  if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
+    if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
+      if (NewITy->getBitWidth() >= OldITy->getBitWidth())
+        return true;
+  if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
+    return false;
+  if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
+    return false;
+
+  // We can convert pointers to integers and vice-versa. Same for vectors
+  // of pointers and integers.
+  OldTy = OldTy->getScalarType();
+  NewTy = NewTy->getScalarType();
+  if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
+    if (NewTy->isPointerTy() && OldTy->isPointerTy())
+      return true;
+    if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
+      return true;
+    return false;
+  }
+
+  return true;
+}
+
+/// \brief Generic routine to convert an SSA value to a value of a different
+/// type.
+///
+/// This will try various different casting techniques, such as bitcasts,
+/// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
+/// two types for viability with this routine.
+static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
+                           Type *NewTy) {
+  Type *OldTy = V->getType();
+  assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
+
+  if (OldTy == NewTy)
+    return V;
+
+  if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
+    if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
+      if (NewITy->getBitWidth() > OldITy->getBitWidth())
+        return IRB.CreateZExt(V, NewITy);
+
+  // See if we need inttoptr for this type pair. A cast involving both scalars
+  // and vectors requires and additional bitcast.
+  if (OldTy->getScalarType()->isIntegerTy() &&
+      NewTy->getScalarType()->isPointerTy()) {
+    // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
+    if (OldTy->isVectorTy() && !NewTy->isVectorTy())
+      return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
+                                NewTy);
+
+    // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
+    if (!OldTy->isVectorTy() && NewTy->isVectorTy())
+      return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
+                                NewTy);
+
+    return IRB.CreateIntToPtr(V, NewTy);
+  }
+
+  // See if we need ptrtoint for this type pair. A cast involving both scalars
+  // and vectors requires and additional bitcast.
+  if (OldTy->getScalarType()->isPointerTy() &&
+      NewTy->getScalarType()->isIntegerTy()) {
+    // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
+    if (OldTy->isVectorTy() && !NewTy->isVectorTy())
+      return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
+                               NewTy);
+
+    // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
+    if (!OldTy->isVectorTy() && NewTy->isVectorTy())
+      return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
+                               NewTy);
+
+    return IRB.CreatePtrToInt(V, NewTy);
+  }
+
+  return IRB.CreateBitCast(V, NewTy);
+}
+
+/// \brief Test whether the given slice use can be promoted to a vector.
+///
+/// This function is called to test each entry in a partioning which is slated
+/// for a single slice.
+static bool isVectorPromotionViableForSlice(
+    const DataLayout &DL, AllocaSlices &S, uint64_t SliceBeginOffset,
+    uint64_t SliceEndOffset, VectorType *Ty, uint64_t ElementSize,
+    AllocaSlices::const_iterator I) {
+  // First validate the slice offsets.
+  uint64_t BeginOffset =
+      std::max(I->beginOffset(), SliceBeginOffset) - SliceBeginOffset;
+  uint64_t BeginIndex = BeginOffset / ElementSize;
+  if (BeginIndex * ElementSize != BeginOffset ||
+      BeginIndex >= Ty->getNumElements())
+    return false;
+  uint64_t EndOffset =
+      std::min(I->endOffset(), SliceEndOffset) - SliceBeginOffset;
+  uint64_t EndIndex = EndOffset / ElementSize;
+  if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
+    return false;
+
+  assert(EndIndex > BeginIndex && "Empty vector!");
+  uint64_t NumElements = EndIndex - BeginIndex;
+  Type *SliceTy =
+      (NumElements == 1) ? Ty->getElementType()
+                         : VectorType::get(Ty->getElementType(), NumElements);
+
+  Type *SplitIntTy =
+      Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
+
+  Use *U = I->getUse();
+
+  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
+    if (MI->isVolatile())
+      return false;
+    if (!I->isSplittable())
+      return false; // Skip any unsplittable intrinsics.
+  } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
+    // Disable vector promotion when there are loads or stores of an FCA.
+    return false;
+  } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
+    if (LI->isVolatile())
+      return false;
+    Type *LTy = LI->getType();
+    if (SliceBeginOffset > I->beginOffset() ||
+        SliceEndOffset < I->endOffset()) {
+      assert(LTy->isIntegerTy());
+      LTy = SplitIntTy;
+    }
+    if (!canConvertValue(DL, SliceTy, LTy))
+      return false;
+  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
+    if (SI->isVolatile())
+      return false;
+    Type *STy = SI->getValueOperand()->getType();
+    if (SliceBeginOffset > I->beginOffset() ||
+        SliceEndOffset < I->endOffset()) {
+      assert(STy->isIntegerTy());
+      STy = SplitIntTy;
+    }
+    if (!canConvertValue(DL, STy, SliceTy))
+      return false;
+  } else {
+    return false;
+  }
+
+  return true;
+}
+
+/// \brief Test whether the given alloca partitioning and range of slices can be
+/// promoted to a vector.
 ///
 /// This is a quick test to check whether we can rewrite a particular alloca
 /// partition (and its newly formed alloca) into a vector alloca with only
@@ -2012,140 +1612,303 @@ static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
 /// SSA value. We only can ensure this for a limited set of operations, and we
 /// don't want to do the rewrites unless we are confident that the result will
 /// be promotable, so we have an early test here.
-static bool isVectorPromotionViable(const DataLayout &TD,
-                                    Type *AllocaTy,
-                                    AllocaPartitioning &P,
-                                    uint64_t PartitionBeginOffset,
-                                    uint64_t PartitionEndOffset,
-                                    AllocaPartitioning::const_use_iterator I,
-                                    AllocaPartitioning::const_use_iterator E) {
+static bool
+isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy, AllocaSlices &S,
+                        uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
+                        AllocaSlices::const_iterator I,
+                        AllocaSlices::const_iterator E,
+                        ArrayRef<AllocaSlices::iterator> SplitUses) {
   VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
   if (!Ty)
     return false;
 
-  uint64_t VecSize = TD.getTypeSizeInBits(Ty);
-  uint64_t ElementSize = Ty->getScalarSizeInBits();
+  uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
 
   // While the definition of LLVM vectors is bitpacked, we don't support sizes
   // that aren't byte sized.
   if (ElementSize % 8)
     return false;
-  assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
-  VecSize /= 8;
+  assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
+         "vector size not a multiple of element size?");
   ElementSize /= 8;
 
-  for (; I != E; ++I) {
-    if (!I->U)
-      continue; // Skip dead use.
-
-    uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
-    uint64_t BeginIndex = BeginOffset / ElementSize;
-    if (BeginIndex * ElementSize != BeginOffset ||
-        BeginIndex >= Ty->getNumElements())
+  for (; I != E; ++I)
+    if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
+                                         SliceEndOffset, Ty, ElementSize, I))
       return false;
-    uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
-    uint64_t EndIndex = EndOffset / ElementSize;
-    if (EndIndex * ElementSize != EndOffset ||
-        EndIndex > Ty->getNumElements())
+
+  for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
+                                                        SUE = SplitUses.end();
+       SUI != SUE; ++SUI)
+    if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
+                                         SliceEndOffset, Ty, ElementSize, *SUI))
       return false;
 
-    // FIXME: We should build shuffle vector instructions to handle
-    // non-element-sized accesses.
-    if ((EndOffset - BeginOffset) != ElementSize &&
-        (EndOffset - BeginOffset) != VecSize)
+  return true;
+}
+
+/// \brief Test whether a slice of an alloca is valid for integer widening.
+///
+/// This implements the necessary checking for the \c isIntegerWideningViable
+/// test below on a single slice of the alloca.
+static bool isIntegerWideningViableForSlice(const DataLayout &DL,
+                                            Type *AllocaTy,
+                                            uint64_t AllocBeginOffset,
+                                            uint64_t Size, AllocaSlices &S,
+                                            AllocaSlices::const_iterator I,
+                                            bool &WholeAllocaOp) {
+  uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
+  uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
+
+  // We can't reasonably handle cases where the load or store extends past
+  // the end of the aloca's type and into its padding.
+  if (RelEnd > Size)
+    return false;
+
+  Use *U = I->getUse();
+
+  if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
+    if (LI->isVolatile())
+      return false;
+    if (RelBegin == 0 && RelEnd == Size)
+      WholeAllocaOp = true;
+    if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
+      if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
+        return false;
+    } else if (RelBegin != 0 || RelEnd != Size ||
+               !canConvertValue(DL, AllocaTy, LI->getType())) {
+      // Non-integer loads need to be convertible from the alloca type so that
+      // they are promotable.
+      return false;
+    }
+  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
+    Type *ValueTy = SI->getValueOperand()->getType();
+    if (SI->isVolatile())
       return false;
-
-    if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
-      if (MI->isVolatile())
+    if (RelBegin == 0 && RelEnd == Size)
+      WholeAllocaOp = true;
+    if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
+      if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
         return false;
-      if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
-        const AllocaPartitioning::MemTransferOffsets &MTO
-          = P.getMemTransferOffsets(*MTI);
-        if (!MTO.IsSplittable)
-          return false;
-      }
-    } 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 (RelBegin != 0 || RelEnd != Size ||
+               !canConvertValue(DL, ValueTy, AllocaTy)) {
+      // Non-integer stores need to be convertible to the alloca type so that
+      // they are promotable.
       return false;
     }
+  } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
+    if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
+      return false;
+    if (!I->isSplittable())
+      return false; // Skip any unsplittable intrinsics.
+  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
+    if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
+        II->getIntrinsicID() != Intrinsic::lifetime_end)
+      return false;
+  } else {
+    return false;
   }
+
   return true;
 }
 
-/// \brief Test whether the given alloca partition can be promoted to an int.
+/// \brief Test whether the given alloca partition's integer operations can be
+/// widened to promotable ones.
 ///
-/// This is a quick test to check whether we can rewrite a particular alloca
-/// partition (and its newly formed alloca) into an integer alloca suitable for
-/// promotion to an SSA value. We only can ensure this for a limited set of
-/// operations, and we don't want to do the rewrites unless we are confident
-/// that the result will be promotable, so we have an early test here.
-static bool isIntegerPromotionViable(const DataLayout &TD,
-                                     Type *AllocaTy,
-                                     uint64_t AllocBeginOffset,
-                                     AllocaPartitioning &P,
-                                     AllocaPartitioning::const_use_iterator I,
-                                     AllocaPartitioning::const_use_iterator E) {
-  IntegerType *Ty = dyn_cast<IntegerType>(AllocaTy);
-  if (!Ty || 8*TD.getTypeStoreSize(Ty) != Ty->getBitWidth())
+/// This is a quick test to check whether we can rewrite the integer loads and
+/// stores to a particular alloca into wider loads and stores and be able to
+/// promote the resulting alloca.
+static bool
+isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
+                        uint64_t AllocBeginOffset, AllocaSlices &S,
+                        AllocaSlices::const_iterator I,
+                        AllocaSlices::const_iterator E,
+                        ArrayRef<AllocaSlices::iterator> SplitUses) {
+  uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
+  // Don't create integer types larger than the maximum bitwidth.
+  if (SizeInBits > IntegerType::MAX_INT_BITS)
+    return false;
+
+  // Don't try to handle allocas with bit-padding.
+  if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
+    return false;
+
+  // We need to ensure that an integer type with the appropriate bitwidth can
+  // be converted to the alloca type, whatever that is. We don't want to force
+  // the alloca itself to have an integer type if there is a more suitable one.
+  Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
+  if (!canConvertValue(DL, AllocaTy, IntTy) ||
+      !canConvertValue(DL, IntTy, AllocaTy))
     return false;
 
-  // Check the uses to ensure the uses are (likely) promoteable integer uses.
-  // Also ensure that the alloca has a covering load or store. We don't want
-  // promote because of some other unsplittable entry (which we may make
-  // splittable later) and lose the ability to promote each element access.
-  bool WholeAllocaOp = false;
-  for (; I != E; ++I) {
-    if (!I->U)
-      continue; // Skip dead use.
-
-    // We can't reasonably handle cases where the load or store extends past
-    // the end of the aloca's type and into its padding.
-    if ((I->EndOffset - AllocBeginOffset) > TD.getTypeStoreSize(Ty))
+  uint64_t Size = DL.getTypeStoreSize(AllocaTy);
+
+  // While examining uses, we ensure that the alloca has a covering load or
+  // store. We don't want to widen the integer operations only to fail to
+  // promote due to some other unsplittable entry (which we may make splittable
+  // later). However, if there are only splittable uses, go ahead and assume
+  // that we cover the alloca.
+  bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
+
+  for (; I != E; ++I)
+    if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
+                                         S, I, WholeAllocaOp))
       return false;
 
-    if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
-      if (LI->isVolatile() || !LI->getType()->isIntegerTy())
-        return false;
-      if (LI->getType() == Ty)
-        WholeAllocaOp = true;
-    } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
-      if (SI->isVolatile() || !SI->getValueOperand()->getType()->isIntegerTy())
-        return false;
-      if (SI->getValueOperand()->getType() == Ty)
-        WholeAllocaOp = true;
-    } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
-      if (MI->isVolatile())
-        return false;
-      if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
-        const AllocaPartitioning::MemTransferOffsets &MTO
-          = P.getMemTransferOffsets(*MTI);
-        if (!MTO.IsSplittable)
-          return false;
-      }
-    } else {
+  for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
+                                                        SUE = SplitUses.end();
+       SUI != SUE; ++SUI)
+    if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
+                                         S, *SUI, WholeAllocaOp))
       return false;
-    }
-  }
+
   return WholeAllocaOp;
 }
 
+static Value *extractInteger(const DataLayout &DL, IRBuilderTy &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) {
+    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) {
+    V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
+    DEBUG(dbgs() << "     trunced: " << *V << "\n");
+  }
+  return V;
+}
+
+static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
+                            Value *V, uint64_t Offset, const Twine &Name) {
+  IntegerType *IntTy = cast<IntegerType>(Old->getType());
+  IntegerType *Ty = cast<IntegerType>(V->getType());
+  assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
+         "Cannot insert a larger integer!");
+  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) {
+    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;
+}
+
+static Value *extractVector(IRBuilderTy &IRB, Value *V,
+                            unsigned BeginIndex, unsigned EndIndex,
+                            const Twine &Name) {
+  VectorType *VecTy = cast<VectorType>(V->getType());
+  unsigned NumElements = EndIndex - BeginIndex;
+  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
+
+  if (NumElements == VecTy->getNumElements())
+    return V;
+
+  if (NumElements == 1) {
+    V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
+                                 Name + ".extract");
+    DEBUG(dbgs() << "     extract: " << *V << "\n");
+    return V;
+  }
+
+  SmallVector<Constant*, 8> Mask;
+  Mask.reserve(NumElements);
+  for (unsigned i = BeginIndex; i != EndIndex; ++i)
+    Mask.push_back(IRB.getInt32(i));
+  V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
+                              ConstantVector::get(Mask),
+                              Name + ".extract");
+  DEBUG(dbgs() << "     shuffle: " << *V << "\n");
+  return V;
+}
+
+static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
+                           unsigned BeginIndex, const Twine &Name) {
+  VectorType *VecTy = cast<VectorType>(Old->getType());
+  assert(VecTy && "Can only insert a vector into a vector");
+
+  VectorType *Ty = dyn_cast<VectorType>(V->getType());
+  if (!Ty) {
+    // Single element to insert.
+    V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
+                                Name + ".insert");
+    DEBUG(dbgs() <<  "     insert: " << *V << "\n");
+    return V;
+  }
+
+  assert(Ty->getNumElements() <= VecTy->getNumElements() &&
+         "Too many elements!");
+  if (Ty->getNumElements() == VecTy->getNumElements()) {
+    assert(V->getType() == VecTy && "Vector type mismatch");
+    return V;
+  }
+  unsigned EndIndex = BeginIndex + Ty->getNumElements();
+
+  // When inserting a smaller vector into the larger to store, we first
+  // use a shuffle vector to widen it with undef elements, and then
+  // a second shuffle vector to select between the loaded vector and the
+  // incoming vector.
+  SmallVector<Constant*, 8> Mask;
+  Mask.reserve(VecTy->getNumElements());
+  for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
+    if (i >= BeginIndex && i < EndIndex)
+      Mask.push_back(IRB.getInt32(i - BeginIndex));
+    else
+      Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
+  V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
+                              ConstantVector::get(Mask),
+                              Name + ".expand");
+  DEBUG(dbgs() << "    shuffle: " << *V << "\n");
+
+  Mask.clear();
+  for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
+    Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
+
+  V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
+
+  DEBUG(dbgs() << "    blend: " << *V << "\n");
+  return V;
+}
+
 namespace {
-/// \brief Visitor to rewrite instructions using a partition of an alloca to
-/// use a new alloca.
+/// \brief Visitor to rewrite instructions using p particular slice of an alloca
+/// to use a new alloca.
 ///
 /// Also implements the rewriting to vector-based accesses when the partition
 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
 /// lives here.
-class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
-                                                   bool> {
+class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
   // Befriend the base class so it can delegate to private visit methods.
-  friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
+  friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
+  typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
 
-  const DataLayout &TD;
-  AllocaPartitioning &P;
+  const DataLayout &DL;
+  AllocaSlices &S;
   SROA &Pass;
   AllocaInst &OldAI, &NewAI;
   const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
@@ -2153,7 +1916,7 @@ class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
 
   // If we are rewriting an alloca partition which can be written as pure
   // vector operations, we stash extra information here. When VecTy is
-  // non-null, we have some strict guarantees about the rewriten alloca:
+  // non-null, we have some strict guarantees about the rewritten alloca:
   //   - The new alloca is exactly the size of the vector type here.
   //   - The accesses all either map to the entire vector or to a single
   //     element.
@@ -2165,100 +1928,117 @@ class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
   uint64_t ElementSize;
 
   // This is a convenience and flag variable that will be null unless the new
-  // alloca has a promotion-targeted integer type due to passing
-  // isIntegerPromotionViable above. If it is non-null does, the desired
+  // alloca's integer operations should be widened to this integer type due to
+  // passing isIntegerWideningViable above. If it is non-null, the desired
   // integer type will be stored here for easy access during rewriting.
-  IntegerType *IntPromotionTy;
+  IntegerType *IntTy;
 
-  // The offset of the partition user currently being rewritten.
+  // The offset of the slice currently being rewritten.
   uint64_t BeginOffset, EndOffset;
+  bool IsSplittable;
+  bool IsSplit;
   Use *OldUse;
   Instruction *OldPtr;
 
-  // The name prefix to use when rewriting instructions for this alloca.
-  std::string NamePrefix;
+  // Output members carrying state about the result of visiting and rewriting
+  // the slice of the alloca.
+  bool IsUsedByRewrittenSpeculatableInstructions;
+
+  // Utility IR builder, whose name prefix is setup for each visited use, and
+  // the insertion point is set to point to the user.
+  IRBuilderTy IRB;
 
 public:
-  AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P,
-                          AllocaPartitioning::iterator PI,
-                          SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
-                          uint64_t NewBeginOffset, uint64_t NewEndOffset)
-    : TD(TD), P(P), Pass(Pass),
-      OldAI(OldAI), NewAI(NewAI),
-      NewAllocaBeginOffset(NewBeginOffset),
-      NewAllocaEndOffset(NewEndOffset),
-      NewAllocaTy(NewAI.getAllocatedType()),
-      VecTy(), ElementTy(), ElementSize(), IntPromotionTy(),
-      BeginOffset(), EndOffset() {
-  }
-
-  /// \brief Visit the users of the alloca partition and rewrite them.
-  bool visitUsers(AllocaPartitioning::const_use_iterator I,
-                  AllocaPartitioning::const_use_iterator E) {
-    if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
-                                NewAllocaBeginOffset, NewAllocaEndOffset,
-                                I, E)) {
-      ++NumVectorized;
-      VecTy = cast<VectorType>(NewAI.getAllocatedType());
-      ElementTy = VecTy->getElementType();
-      assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
+  AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
+                      AllocaInst &OldAI, AllocaInst &NewAI,
+                      uint64_t NewBeginOffset, uint64_t NewEndOffset,
+                      bool IsVectorPromotable = false,
+                      bool IsIntegerPromotable = false)
+      : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
+        NewAllocaBeginOffset(NewBeginOffset), NewAllocaEndOffset(NewEndOffset),
+        NewAllocaTy(NewAI.getAllocatedType()),
+        VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : 0),
+        ElementTy(VecTy ? VecTy->getElementType() : 0),
+        ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
+        IntTy(IsIntegerPromotable
+                  ? Type::getIntNTy(
+                        NewAI.getContext(),
+                        DL.getTypeSizeInBits(NewAI.getAllocatedType()))
+                  : 0),
+        BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
+        OldPtr(), IsUsedByRewrittenSpeculatableInstructions(false),
+        IRB(NewAI.getContext(), ConstantFolder()) {
+    if (VecTy) {
+      assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
              "Only multiple-of-8 sized vector elements are viable");
-      ElementSize = VecTy->getScalarSizeInBits() / 8;
-    } else if (isIntegerPromotionViable(TD, NewAI.getAllocatedType(),
-                                        NewAllocaBeginOffset, P, I, E)) {
-      IntPromotionTy = cast<IntegerType>(NewAI.getAllocatedType());
+      ++NumVectorized;
     }
+    assert((!IsVectorPromotable && !IsIntegerPromotable) ||
+           IsVectorPromotable != IsIntegerPromotable);
+  }
+
+  bool visit(AllocaSlices::const_iterator I) {
     bool CanSROA = true;
-    for (; I != E; ++I) {
-      if (!I->U)
-        continue; // Skip dead uses.
-      BeginOffset = I->BeginOffset;
-      EndOffset = I->EndOffset;
-      OldUse = I->U;
-      OldPtr = cast<Instruction>(I->U->get());
-      NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
-      CanSROA &= visit(cast<Instruction>(I->U->getUser()));
-    }
-    if (VecTy) {
+    BeginOffset = I->beginOffset();
+    EndOffset = I->endOffset();
+    IsSplittable = I->isSplittable();
+    IsSplit =
+        BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
+
+    OldUse = I->getUse();
+    OldPtr = cast<Instruction>(OldUse->get());
+
+    Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
+    IRB.SetInsertPoint(OldUserI);
+    IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
+    IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
+
+    CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
+    if (VecTy || IntTy)
       assert(CanSROA);
-      VecTy = 0;
-      ElementTy = 0;
-      ElementSize = 0;
-    }
     return CanSROA;
   }
 
+  /// \brief Query whether this slice is used by speculatable instructions after
+  /// rewriting.
+  ///
+  /// These instructions (PHIs and Selects currently) require the alloca slice
+  /// to run back through the rewriter. Thus, they are promotable, but not on
+  /// this iteration. This is distinct from a slice which is unpromotable for
+  /// some other reason, in which case we don't even want to perform the
+  /// speculation. This can be querried at any time and reflects whether (at
+  /// that point) a visit call has rewritten a speculatable instruction on the
+  /// current slice.
+  bool isUsedByRewrittenSpeculatableInstructions() const {
+    return IsUsedByRewrittenSpeculatableInstructions;
+  }
+
 private:
+  // Make sure the other visit overloads are visible.
+  using Base::visit;
+
   // Every instruction which can end up as a user must have a rewrite rule.
   bool visitInstruction(Instruction &I) {
     DEBUG(dbgs() << "    !!!! Cannot rewrite: " << I << "\n");
     llvm_unreachable("No rewrite rule for this instruction!");
   }
 
-  Twine getName(const Twine &Suffix) {
-    return NamePrefix + Suffix;
-  }
-
-  Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
-    assert(BeginOffset >= NewAllocaBeginOffset);
-    APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
-    return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
+  Value *getAdjustedAllocaPtr(IRBuilderTy &IRB, uint64_t Offset,
+                              Type *PointerTy) {
+    assert(Offset >= NewAllocaBeginOffset);
+    return getAdjustedPtr(IRB, DL, &NewAI, APInt(DL.getPointerSizeInBits(),
+                                                 Offset - NewAllocaBeginOffset),
+                          PointerTy);
   }
 
   /// \brief Compute suitable alignment to access an offset into the new alloca.
   unsigned getOffsetAlign(uint64_t Offset) {
     unsigned NewAIAlign = NewAI.getAlignment();
     if (!NewAIAlign)
-      NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
+      NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
     return MinAlign(NewAIAlign, Offset);
   }
 
-  /// \brief Compute suitable alignment to access this partition of the new
-  /// alloca.
-  unsigned getPartitionAlign() {
-    return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
-  }
-
   /// \brief Compute suitable alignment to access a type at an offset of the
   /// new alloca.
   ///
@@ -2266,211 +2046,161 @@ private:
   /// otherwise returns the maximal suitable alignment.
   unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
     unsigned Align = getOffsetAlign(Offset);
-    return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
-  }
-
-  /// \brief Compute suitable alignment to access a type at the beginning of
-  /// this partition of the new alloca.
-  ///
-  /// See \c getOffsetTypeAlign for details; this routine delegates to it.
-  unsigned getPartitionTypeAlign(Type *Ty) {
-    return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
+    return Align == DL.getABITypeAlignment(Ty) ? 0 : Align;
   }
 
-  ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
+  unsigned getIndex(uint64_t Offset) {
     assert(VecTy && "Can only call getIndex when rewriting a vector");
     uint64_t RelOffset = Offset - NewAllocaBeginOffset;
     assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
     uint32_t Index = RelOffset / ElementSize;
     assert(Index * ElementSize == RelOffset);
-    return IRB.getInt32(Index);
-  }
-
-  Value *extractInteger(IRBuilder<> &IRB, IntegerType *TargetTy,
-                        uint64_t Offset) {
-    assert(IntPromotionTy && "Alloca is not an integer we can extract from");
-    Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
-                                     getName(".load"));
-    assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
-    uint64_t RelOffset = Offset - NewAllocaBeginOffset;
-    assert(TD.getTypeStoreSize(TargetTy) + RelOffset <=
-           TD.getTypeStoreSize(IntPromotionTy) &&
-           "Element load outside of alloca store");
-    uint64_t ShAmt = 8*RelOffset;
-    if (TD.isBigEndian())
-      ShAmt = 8*(TD.getTypeStoreSize(IntPromotionTy) -
-                 TD.getTypeStoreSize(TargetTy) - RelOffset);
-    if (ShAmt)
-      V = IRB.CreateLShr(V, ShAmt, getName(".shift"));
-    if (TargetTy != IntPromotionTy) {
-      assert(TargetTy->getBitWidth() < IntPromotionTy->getBitWidth() &&
-             "Cannot extract to a larger integer!");
-      V = IRB.CreateTrunc(V, TargetTy, getName(".trunc"));
-    }
-    return V;
-  }
-
-  StoreInst *insertInteger(IRBuilder<> &IRB, Value *V, uint64_t Offset) {
-    IntegerType *Ty = cast<IntegerType>(V->getType());
-    if (Ty == IntPromotionTy)
-      return IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
-
-    assert(Ty->getBitWidth() < IntPromotionTy->getBitWidth() &&
-           "Cannot insert a larger integer!");
-    V = IRB.CreateZExt(V, IntPromotionTy, getName(".ext"));
-    assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
-    uint64_t RelOffset = Offset - NewAllocaBeginOffset;
-    assert(TD.getTypeStoreSize(Ty) + RelOffset <=
-           TD.getTypeStoreSize(IntPromotionTy) &&
-           "Element store outside of alloca store");
-    uint64_t ShAmt = 8*RelOffset;
-    if (TD.isBigEndian())
-      ShAmt = 8*(TD.getTypeStoreSize(IntPromotionTy) - TD.getTypeStoreSize(Ty)
-                 - RelOffset);
-    if (ShAmt)
-      V = IRB.CreateShl(V, ShAmt, getName(".shift"));
-
-    APInt Mask = ~Ty->getMask().zext(IntPromotionTy->getBitWidth()).shl(ShAmt);
-    Value *Old = IRB.CreateAnd(IRB.CreateAlignedLoad(&NewAI,
-                                                     NewAI.getAlignment(),
-                                                     getName(".oldload")),
-                               Mask, getName(".mask"));
-    return IRB.CreateAlignedStore(IRB.CreateOr(Old, V, getName(".insert")),
-                                  &NewAI, NewAI.getAlignment());
+    return Index;
   }
 
   void deleteIfTriviallyDead(Value *V) {
     Instruction *I = cast<Instruction>(V);
     if (isInstructionTriviallyDead(I))
-      Pass.DeadInsts.push_back(I);
-  }
-
-  /// \brief Test whether we can convert a value from the old to the new type.
-  ///
-  /// This predicate should be used to guard calls to convertValue in order to
-  /// ensure that we only try to convert viable values. The strategy is that we
-  /// will peel off single element struct and array wrappings to get to an
-  /// underlying value, and convert that value.
-  bool canConvertValue(Type *OldTy, Type *NewTy) {
-    if (OldTy == NewTy)
-      return true;
-    if (TD.getTypeSizeInBits(NewTy) != TD.getTypeSizeInBits(OldTy))
-      return false;
-    if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
-      return false;
-
-    if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
-      if (NewTy->isPointerTy() && OldTy->isPointerTy())
-        return true;
-      if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
-        return true;
-      return false;
-    }
-
-    return true;
-  }
-
-  Value *convertValue(IRBuilder<> &IRB, Value *V, Type *Ty) {
-    assert(canConvertValue(V->getType(), Ty) && "Value not convertable to type");
-    if (V->getType() == Ty)
-      return V;
-    if (V->getType()->isIntegerTy() && Ty->isPointerTy())
-      return IRB.CreateIntToPtr(V, Ty);
-    if (V->getType()->isPointerTy() && Ty->isIntegerTy())
-      return IRB.CreatePtrToInt(V, Ty);
-
-    return IRB.CreateBitCast(V, Ty);
+      Pass.DeadInsts.insert(I);
   }
 
-  bool rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
-    Value *Result;
-    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"));
-    }
-    if (Result->getType() != LI.getType())
-      Result = convertValue(IRB, Result, LI.getType());
-    LI.replaceAllUsesWith(Result);
-    Pass.DeadInsts.push_back(&LI);
+  Value *rewriteVectorizedLoadInst(uint64_t NewBeginOffset,
+                                   uint64_t NewEndOffset) {
+    unsigned BeginIndex = getIndex(NewBeginOffset);
+    unsigned EndIndex = getIndex(NewEndOffset);
+    assert(EndIndex > BeginIndex && "Empty vector!");
 
-    DEBUG(dbgs() << "          to: " << *Result << "\n");
-    return true;
+    Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                     "load");
+    return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
   }
 
-  bool rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
+  Value *rewriteIntegerLoad(LoadInst &LI, uint64_t NewBeginOffset,
+                            uint64_t NewEndOffset) {
+    assert(IntTy && "We cannot insert an integer to the alloca");
     assert(!LI.isVolatile());
-    Value *Result = extractInteger(IRB, cast<IntegerType>(LI.getType()),
-                                   BeginOffset);
-    LI.replaceAllUsesWith(Result);
-    Pass.DeadInsts.push_back(&LI);
-    DEBUG(dbgs() << "          to: " << *Result << "\n");
-    return true;
+    Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                     "load");
+    V = convertValue(DL, IRB, V, IntTy);
+    assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
+    uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
+    if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
+      V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
+                         "extract");
+    return V;
   }
 
   bool visitLoadInst(LoadInst &LI) {
     DEBUG(dbgs() << "    original: " << LI << "\n");
     Value *OldOp = LI.getOperand(0);
     assert(OldOp == OldPtr);
-    IRBuilder<> IRB(&LI);
 
-    if (VecTy)
-      return rewriteVectorizedLoadInst(IRB, LI, OldOp);
-    if (IntPromotionTy)
-      return rewriteIntegerLoad(IRB, LI);
-
-    if (BeginOffset == NewAllocaBeginOffset &&
-        canConvertValue(NewAllocaTy, LI.getType())) {
-      Value *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
-                                           LI.isVolatile(), getName(".load"));
-      Value *NewV = convertValue(IRB, NewLI, LI.getType());
-      LI.replaceAllUsesWith(NewV);
-      Pass.DeadInsts.push_back(&LI);
-
-      DEBUG(dbgs() << "          to: " << *NewLI << "\n");
-      return !LI.isVolatile();
-    }
+    // Compute the intersecting offset range.
+    assert(BeginOffset < NewAllocaEndOffset);
+    assert(EndOffset > NewAllocaBeginOffset);
+    uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
+    uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
 
-    Value *NewPtr = getAdjustedAllocaPtr(IRB,
-                                         LI.getPointerOperand()->getType());
-    LI.setOperand(0, NewPtr);
-    LI.setAlignment(getPartitionTypeAlign(LI.getType()));
-    DEBUG(dbgs() << "          to: " << LI << "\n");
+    uint64_t Size = NewEndOffset - NewBeginOffset;
 
-    deleteIfTriviallyDead(OldOp);
-    return NewPtr == &NewAI && !LI.isVolatile();
-  }
+    Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), Size * 8)
+                             : LI.getType();
+    bool IsPtrAdjusted = false;
+    Value *V;
+    if (VecTy) {
+      V = rewriteVectorizedLoadInst(NewBeginOffset, NewEndOffset);
+    } else if (IntTy && LI.getType()->isIntegerTy()) {
+      V = rewriteIntegerLoad(LI, NewBeginOffset, NewEndOffset);
+    } else if (NewBeginOffset == NewAllocaBeginOffset &&
+               canConvertValue(DL, NewAllocaTy, LI.getType())) {
+      V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                LI.isVolatile(), "load");
+    } else {
+      Type *LTy = TargetTy->getPointerTo();
+      V = IRB.CreateAlignedLoad(
+          getAdjustedAllocaPtr(IRB, NewBeginOffset, LTy),
+          getOffsetTypeAlign(TargetTy, NewBeginOffset - NewAllocaBeginOffset),
+          LI.isVolatile(), "load");
+      IsPtrAdjusted = true;
+    }
+    V = convertValue(DL, IRB, V, TargetTy);
+
+    if (IsSplit) {
+      assert(!LI.isVolatile());
+      assert(LI.getType()->isIntegerTy() &&
+             "Only integer type loads and stores are split");
+      assert(Size < DL.getTypeStoreSize(LI.getType()) &&
+             "Split load isn't smaller than original load");
+      assert(LI.getType()->getIntegerBitWidth() ==
+             DL.getTypeStoreSizeInBits(LI.getType()) &&
+             "Non-byte-multiple bit width");
+      // Move the insertion point just past the load so that we can refer to it.
+      IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
+      // 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
+      // LI only used for this computation.
+      Value *Placeholder
+        = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
+      V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
+                        "insert");
+      LI.replaceAllUsesWith(V);
+      Placeholder->replaceAllUsesWith(&LI);
+      delete Placeholder;
+    } else {
+      LI.replaceAllUsesWith(V);
+    }
 
-  bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, StoreInst &SI,
-                                  Value *OldOp) {
-    Value *V = SI.getValueOperand();
-    if (V->getType() == ElementTy ||
-        BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
-      if (V->getType() != ElementTy)
-        V = convertValue(IRB, V, ElementTy);
-      LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
-                                           getName(".load"));
-      V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset),
-                                  getName(".insert"));
-    } else if (V->getType() != VecTy) {
-      V = convertValue(IRB, V, VecTy);
+    Pass.DeadInsts.insert(&LI);
+    deleteIfTriviallyDead(OldOp);
+    DEBUG(dbgs() << "          to: " << *V << "\n");
+    return !LI.isVolatile() && !IsPtrAdjusted;
+  }
+
+  bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
+                                  uint64_t NewBeginOffset,
+                                  uint64_t NewEndOffset) {
+    if (V->getType() != VecTy) {
+      unsigned BeginIndex = getIndex(NewBeginOffset);
+      unsigned EndIndex = getIndex(NewEndOffset);
+      assert(EndIndex > BeginIndex && "Empty vector!");
+      unsigned NumElements = EndIndex - BeginIndex;
+      assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
+      Type *SliceTy =
+          (NumElements == 1) ? ElementTy
+                             : VectorType::get(ElementTy, NumElements);
+      if (V->getType() != SliceTy)
+        V = convertValue(DL, IRB, V, SliceTy);
+
+      // Mix in the existing elements.
+      Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                         "load");
+      V = insertVector(IRB, Old, V, BeginIndex, "vec");
     }
     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(Value *V, StoreInst &SI,
+                           uint64_t NewBeginOffset, uint64_t NewEndOffset) {
+    assert(IntTy && "We cannot extract an integer from the alloca");
     assert(!SI.isVolatile());
-    StoreInst *Store = insertInteger(IRB, SI.getValueOperand(), BeginOffset);
-    Pass.DeadInsts.push_back(&SI);
+    if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
+      Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                         "oldload");
+      Old = convertValue(DL, IRB, Old, IntTy);
+      assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
+      uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
+      V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
+                        "insert");
+    }
+    V = convertValue(DL, IRB, V, NewAllocaTy);
+    StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
+    Pass.DeadInsts.insert(&SI);
     (void)Store;
     DEBUG(dbgs() << "          to: " << *Store << "\n");
     return true;
@@ -2480,80 +2210,141 @@ private:
     DEBUG(dbgs() << "    original: " << SI << "\n");
     Value *OldOp = SI.getOperand(1);
     assert(OldOp == OldPtr);
-    IRBuilder<> IRB(&SI);
-
-    if (VecTy)
-      return rewriteVectorizedStoreInst(IRB, SI, OldOp);
-    if (IntPromotionTy)
-      return rewriteIntegerStore(IRB, SI);
 
-    Type *ValueTy = SI.getValueOperand()->getType();
+    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(ValueTy, NewAllocaTy)) {
-      Value *NewV = convertValue(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();
+    // Compute the intersecting offset range.
+    assert(BeginOffset < NewAllocaEndOffset);
+    assert(EndOffset > NewAllocaBeginOffset);
+    uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
+    uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
+
+    uint64_t Size = NewEndOffset - NewBeginOffset;
+    if (Size < DL.getTypeStoreSize(V->getType())) {
+      assert(!SI.isVolatile());
+      assert(V->getType()->isIntegerTy() &&
+             "Only integer type loads and stores are split");
+      assert(V->getType()->getIntegerBitWidth() ==
+             DL.getTypeStoreSizeInBits(V->getType()) &&
+             "Non-byte-multiple bit width");
+      IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
+      V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
+                         "extract");
     }
 
-    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(V, SI, OldOp, NewBeginOffset,
+                                        NewEndOffset);
+    if (IntTy && V->getType()->isIntegerTy())
+      return rewriteIntegerStore(V, SI, NewBeginOffset, NewEndOffset);
+
+    StoreInst *NewSI;
+    if (NewBeginOffset == NewAllocaBeginOffset &&
+        NewEndOffset == NewAllocaEndOffset &&
+        canConvertValue(DL, V->getType(), NewAllocaTy)) {
+      V = convertValue(DL, IRB, V, NewAllocaTy);
+      NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
+                                     SI.isVolatile());
+    } else {
+      Value *NewPtr = getAdjustedAllocaPtr(IRB, NewBeginOffset,
+                                           V->getType()->getPointerTo());
+      NewSI = IRB.CreateAlignedStore(
+          V, NewPtr, getOffsetTypeAlign(
+                         V->getType(), NewBeginOffset - NewAllocaBeginOffset),
+          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();
+  }
+
+  /// \brief Compute an integer value from splatting an i8 across the given
+  /// number of bytes.
+  ///
+  /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
+  /// call this routine.
+  /// FIXME: Heed the advice above.
+  ///
+  /// \param V The i8 value to splat.
+  /// \param Size The number of bytes in the output (assuming i8 is one byte)
+  Value *getIntegerSplat(Value *V, unsigned Size) {
+    assert(Size > 0 && "Expected a positive number of bytes.");
+    IntegerType *VTy = cast<IntegerType>(V->getType());
+    assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
+    if (Size == 1)
+      return V;
+
+    Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
+    V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
+                      ConstantExpr::getUDiv(
+                        Constant::getAllOnesValue(SplatIntTy),
+                        ConstantExpr::getZExt(
+                          Constant::getAllOnesValue(V->getType()),
+                          SplatIntTy)),
+                      "isplat");
+    return V;
+  }
+
+  /// \brief Compute a vector splat for a given element value.
+  Value *getVectorSplat(Value *V, unsigned NumElements) {
+    V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
+    DEBUG(dbgs() << "       splat: " << *V << "\n");
+    return V;
   }
 
   bool visitMemSetInst(MemSetInst &II) {
     DEBUG(dbgs() << "    original: " << II << "\n");
-    IRBuilder<> IRB(&II);
     assert(II.getRawDest() == OldPtr);
 
     // If the memset has a variable size, it cannot be split, just adjust the
     // pointer to the new alloca.
     if (!isa<Constant>(II.getLength())) {
-      II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
+      assert(!IsSplit);
+      assert(BeginOffset >= NewAllocaBeginOffset);
+      II.setDest(
+          getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
       Type *CstTy = II.getAlignmentCst()->getType();
-      II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
+      II.setAlignment(ConstantInt::get(CstTy, getOffsetAlign(BeginOffset)));
 
       deleteIfTriviallyDead(OldPtr);
       return false;
     }
 
     // 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();
 
+    // Compute the intersecting offset range.
+    assert(BeginOffset < NewAllocaEndOffset);
+    assert(EndOffset > NewAllocaBeginOffset);
+    uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
+    uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
+    uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
+
     // If this doesn't map cleanly onto the alloca type, and that type isn't
     // a single value type, just emit a memset.
-    if (!VecTy && (BeginOffset != NewAllocaBeginOffset ||
-                   EndOffset != NewAllocaEndOffset ||
-                   !AllocaTy->isSingleValueType() ||
-                   !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
+    if (!VecTy && !IntTy &&
+        (BeginOffset > NewAllocaBeginOffset ||
+         EndOffset < NewAllocaEndOffset ||
+         !AllocaTy->isSingleValueType() ||
+         !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
+         DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
       Type *SizeTy = II.getLength()->getType();
-      Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
-      CallInst *New
-        = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
-                                                II.getRawDest()->getType()),
-                           II.getValue(), Size, getPartitionAlign(),
-                           II.isVolatile());
+      Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
+      CallInst *New = IRB.CreateMemSet(
+          getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getRawDest()->getType()),
+          II.getValue(), Size, getOffsetAlign(SliceOffset), II.isVolatile());
       (void)New;
       DEBUG(dbgs() << "          to: " << *New << "\n");
       return false;
@@ -2562,56 +2353,59 @@ private:
     // If we can represent this as a simple value, we have to build the actual
     // value to store, which requires expanding the byte present in memset to
     // a sensible representation for the alloca type. This is essentially
-    // splatting the byte to a sufficiently wide integer, bitcasting to the
-    // desired scalar type, and splatting it across any desired vector type.
-    Value *V = II.getValue();
-    IntegerType *VTy = cast<IntegerType>(V->getType());
-    Type *IntTy = Type::getIntNTy(VTy->getContext(),
-                                  TD.getTypeSizeInBits(ScalarTy));
-    if (TD.getTypeSizeInBits(ScalarTy) > VTy->getBitWidth())
-      V = IRB.CreateMul(IRB.CreateZExt(V, IntTy, getName(".zext")),
-                        ConstantExpr::getUDiv(
-                          Constant::getAllOnesValue(IntTy),
-                          ConstantExpr::getZExt(
-                            Constant::getAllOnesValue(V->getType()),
-                            IntTy)),
-                        getName(".isplat"));
-    if (V->getType() != ScalarTy) {
-      if (ScalarTy->isPointerTy())
-        V = IRB.CreateIntToPtr(V, ScalarTy);
-      else if (ScalarTy->isPrimitiveType() || ScalarTy->isVectorTy())
-        V = IRB.CreateBitCast(V, ScalarTy);
-      else if (ScalarTy->isIntegerTy())
-        llvm_unreachable("Computed different integer types with equal widths");
-      else
-        llvm_unreachable("Invalid scalar type");
-    }
+    // splatting the byte to a sufficiently wide integer, splatting it across
+    // any desired vector width, and bitcasting to the final type.
+    Value *V;
 
-    // If this is an element-wide memset of a vectorizable alloca, insert it.
-    if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
-                  EndOffset < NewAllocaEndOffset)) {
-      StoreInst *Store = IRB.CreateAlignedStore(
-        IRB.CreateInsertElement(IRB.CreateAlignedLoad(&NewAI,
-                                                      NewAI.getAlignment(),
-                                                      getName(".load")),
-                                V, getIndex(IRB, BeginOffset),
-                                getName(".insert")),
-        &NewAI, NewAI.getAlignment());
-      (void)Store;
-      DEBUG(dbgs() << "          to: " << *Store << "\n");
-      return true;
-    }
+    if (VecTy) {
+      // If this is a memset of a vectorized alloca, insert it.
+      assert(ElementTy == ScalarTy);
+
+      unsigned BeginIndex = getIndex(NewBeginOffset);
+      unsigned EndIndex = getIndex(NewEndOffset);
+      assert(EndIndex > BeginIndex && "Empty vector!");
+      unsigned NumElements = EndIndex - BeginIndex;
+      assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
+
+      Value *Splat =
+          getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
+      Splat = convertValue(DL, IRB, Splat, ElementTy);
+      if (NumElements > 1)
+        Splat = getVectorSplat(Splat, NumElements);
+
+      Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                         "oldload");
+      V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
+    } else if (IntTy) {
+      // If this is a memset on an alloca where we can widen stores, insert the
+      // set integer.
+      assert(!II.isVolatile());
+
+      uint64_t Size = NewEndOffset - NewBeginOffset;
+      V = getIntegerSplat(II.getValue(), Size);
+
+      if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
+                    EndOffset != NewAllocaBeginOffset)) {
+        Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                           "oldload");
+        Old = convertValue(DL, IRB, Old, IntTy);
+        uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
+        V = insertInteger(DL, IRB, Old, V, Offset, "insert");
+      } else {
+        assert(V->getType() == IntTy &&
+               "Wrong type for an alloca wide integer!");
+      }
+      V = convertValue(DL, IRB, V, AllocaTy);
+    } else {
+      // Established these invariants above.
+      assert(NewBeginOffset == NewAllocaBeginOffset);
+      assert(NewEndOffset == NewAllocaEndOffset);
+
+      V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
+      if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
+        V = getVectorSplat(V, AllocaVecTy->getNumElements());
 
-    // Splat to a vector if needed.
-    if (VectorType *VecTy = dyn_cast<VectorType>(AllocaTy)) {
-      VectorType *SplatSourceTy = VectorType::get(V->getType(), 1);
-      V = IRB.CreateShuffleVector(
-        IRB.CreateInsertElement(UndefValue::get(SplatSourceTy), V,
-                                IRB.getInt32(0), getName(".vsplat.insert")),
-        UndefValue::get(SplatSourceTy),
-        ConstantVector::getSplat(VecTy->getNumElements(), IRB.getInt32(0)),
-        getName(".vsplat.shuffle"));
-      assert(V->getType() == VecTy);
+      V = convertValue(DL, IRB, V, AllocaTy);
     }
 
     Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
@@ -2626,23 +2420,26 @@ private:
     // them into two categories: split intrinsics and unsplit intrinsics.
 
     DEBUG(dbgs() << "    original: " << II << "\n");
-    IRBuilder<> IRB(&II);
+
+    // Compute the intersecting offset range.
+    assert(BeginOffset < NewAllocaEndOffset);
+    assert(EndOffset > NewAllocaBeginOffset);
+    uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
+    uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
 
     assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
     bool IsDest = II.getRawDest() == OldPtr;
 
-    const AllocaPartitioning::MemTransferOffsets &MTO
-      = P.getMemTransferOffsets(II);
-
     // Compute the relative offset within the transfer.
-    unsigned IntPtrWidth = TD.getPointerSizeInBits();
-    APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
-                                                       : MTO.SourceBegin));
+    unsigned IntPtrWidth = DL.getPointerSizeInBits();
+    APInt RelOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
 
     unsigned Align = II.getAlignment();
+    uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
     if (Align > 1)
-      Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
-                       MinAlign(II.getAlignment(), getPartitionAlign()));
+      Align =
+          MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
+                   MinAlign(II.getAlignment(), getOffsetAlign(SliceOffset)));
 
     // For unsplit intrinsics, we simply modify the source and destination
     // pointers in place. This isn't just an optimization, it is a matter of
@@ -2651,12 +2448,14 @@ private:
     // a variable length. We may also be dealing with memmove instead of
     // memcpy, and so simply updating the pointers is the necessary for us to
     // update both source and dest of a single call.
-    if (!MTO.IsSplittable) {
+    if (!IsSplittable) {
       Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
       if (IsDest)
-        II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
+        II.setDest(
+            getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
       else
-        II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
+        II.setSource(getAdjustedAllocaPtr(IRB, BeginOffset,
+                                          II.getRawSource()->getType()));
 
       Type *CstTy = II.getAlignmentCst()->getType();
       II.setAlignment(ConstantInt::get(CstTy, Align));
@@ -2674,57 +2473,46 @@ private:
     // If this doesn't map cleanly onto the alloca type, and that type isn't
     // a single value type, just emit a memcpy.
     bool EmitMemCpy
-      = !VecTy && (BeginOffset != NewAllocaBeginOffset ||
-                   EndOffset != NewAllocaEndOffset ||
-                   !NewAI.getAllocatedType()->isSingleValueType());
+      = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
+                             EndOffset < NewAllocaEndOffset ||
+                             !NewAI.getAllocatedType()->isSingleValueType());
 
     // If we're just going to emit a memcpy, the alloca hasn't changed, and the
     // size hasn't been shrunk based on analysis of the viable range, this is
     // a no-op.
     if (EmitMemCpy && &OldAI == &NewAI) {
-      uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
-      uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
       // Ensure the start lines up.
-      assert(BeginOffset == OrigBegin);
-      (void)OrigBegin;
+      assert(NewBeginOffset == BeginOffset);
 
       // Rewrite the size as needed.
-      if (EndOffset != OrigEnd)
+      if (NewEndOffset != EndOffset)
         II.setLength(ConstantInt::get(II.getLength()->getType(),
-                                      EndOffset - BeginOffset));
+                                      NewEndOffset - NewBeginOffset));
       return false;
     }
     // Record this instruction for deletion.
-    if (Pass.DeadSplitInsts.insert(&II))
-      Pass.DeadInsts.push_back(&II);
-
-    bool IsVectorElement = VecTy && (BeginOffset > NewAllocaBeginOffset ||
-                                     EndOffset < NewAllocaEndOffset);
-
-    Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
-                              : II.getRawDest()->getType();
-    if (!EmitMemCpy)
-      OtherPtrTy = IsVectorElement ? VecTy->getElementType()->getPointerTo()
-                                   : NewAI.getType();
-
-    // Compute the other pointer, folding as much as possible to produce
-    // a single, simple GEP in most cases.
-    Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
-    OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
-                              getName("." + OtherPtr->getName()));
+    Pass.DeadInsts.insert(&II);
 
     // Strip all inbounds GEPs and pointer casts to try to dig out any root
     // alloca that should be re-examined after rewriting this instruction.
+    Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
     if (AllocaInst *AI
           = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
       Pass.Worklist.insert(AI);
 
     if (EmitMemCpy) {
-      Value *OurPtr
-        = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
-                                           : II.getRawSource()->getType());
+      Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
+                                : II.getRawDest()->getType();
+
+      // Compute the other pointer, folding as much as possible to produce
+      // a single, simple GEP in most cases.
+      OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
+
+      Value *OurPtr = getAdjustedAllocaPtr(
+          IRB, NewBeginOffset,
+          IsDest ? II.getRawDest()->getType() : II.getRawSource()->getType());
       Type *SizeTy = II.getLength()->getType();
-      Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
+      Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
 
       CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
                                        IsDest ? OtherPtr : OurPtr,
@@ -2740,29 +2528,59 @@ private:
     if (!Align)
       Align = 1;
 
-    Value *SrcPtr = OtherPtr;
+    bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
+                         NewEndOffset == NewAllocaEndOffset;
+    uint64_t Size = NewEndOffset - NewBeginOffset;
+    unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
+    unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
+    unsigned NumElements = EndIndex - BeginIndex;
+    IntegerType *SubIntTy
+      = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
+
+    Type *OtherPtrTy = NewAI.getType();
+    if (VecTy && !IsWholeAlloca) {
+      if (NumElements == 1)
+        OtherPtrTy = VecTy->getElementType();
+      else
+        OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
+
+      OtherPtrTy = OtherPtrTy->getPointerTo();
+    } else if (IntTy && !IsWholeAlloca) {
+      OtherPtrTy = SubIntTy->getPointerTo();
+    }
+
+    Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
     Value *DstPtr = &NewAI;
     if (!IsDest)
       std::swap(SrcPtr, DstPtr);
 
     Value *Src;
-    if (IsVectorElement && !IsDest) {
-      // We have to extract rather than load.
-      Src = IRB.CreateExtractElement(
-        IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")),
-        getIndex(IRB, BeginOffset),
-        getName(".copyextract"));
+    if (VecTy && !IsWholeAlloca && !IsDest) {
+      Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                  "load");
+      Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
+    } else if (IntTy && !IsWholeAlloca && !IsDest) {
+      Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                  "load");
+      Src = convertValue(DL, IRB, Src, IntTy);
+      uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
+      Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
     } else {
       Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
-                                  getName(".copyload"));
+                                  "copyload");
     }
 
-    if (IsVectorElement && IsDest) {
-      // We have to insert into a loaded copy before storing.
-      Src = IRB.CreateInsertElement(
-        IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
-        Src, getIndex(IRB, BeginOffset),
-        getName(".insert"));
+    if (VecTy && !IsWholeAlloca && IsDest) {
+      Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                         "oldload");
+      Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
+    } else if (IntTy && !IsWholeAlloca && IsDest) {
+      Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                         "oldload");
+      Old = convertValue(DL, IRB, Old, IntTy);
+      uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
+      Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
+      Src = convertValue(DL, IRB, Src, NewAllocaTy);
     }
 
     StoreInst *Store = cast<StoreInst>(
@@ -2776,64 +2594,93 @@ private:
     assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
            II.getIntrinsicID() == Intrinsic::lifetime_end);
     DEBUG(dbgs() << "    original: " << II << "\n");
-    IRBuilder<> IRB(&II);
     assert(II.getArgOperand(1) == OldPtr);
 
+    // Compute the intersecting offset range.
+    assert(BeginOffset < NewAllocaEndOffset);
+    assert(EndOffset > NewAllocaBeginOffset);
+    uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
+    uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
+
     // 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()),
-                         EndOffset - BeginOffset);
-    Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
+                         NewEndOffset - NewBeginOffset);
+    Value *Ptr =
+        getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getArgOperand(1)->getType());
     Value *New;
     if (II.getIntrinsicID() == Intrinsic::lifetime_start)
       New = IRB.CreateLifetimeStart(Ptr, Size);
     else
       New = IRB.CreateLifetimeEnd(Ptr, Size);
 
+    (void)New;
     DEBUG(dbgs() << "          to: " << *New << "\n");
     return true;
   }
 
   bool visitPHINode(PHINode &PN) {
     DEBUG(dbgs() << "    original: " << PN << "\n");
+    assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
+    assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
 
     // We would like to compute a new pointer in only one place, but have it be
     // as local as possible to the PHI. To do that, we re-use the location of
     // the old pointer, which necessarily must be in the right position to
     // dominate the PHI.
-    IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
+    IRBuilderTy PtrBuilder(OldPtr);
+    PtrBuilder.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) +
+                             ".");
 
-    Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
+    Value *NewPtr =
+        getAdjustedAllocaPtr(PtrBuilder, BeginOffset, OldPtr->getType());
     // Replace the operands which were using the old pointer.
-    User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
-    for (; OI != OE; ++OI)
-      if (*OI == OldPtr)
-        *OI = NewPtr;
+    std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
 
     DEBUG(dbgs() << "          to: " << PN << "\n");
     deleteIfTriviallyDead(OldPtr);
-    return false;
+
+    // Check whether we can speculate this PHI node, and if so remember that
+    // fact and queue it up for another iteration after the speculation
+    // occurs.
+    if (isSafePHIToSpeculate(PN, &DL)) {
+      Pass.SpeculatablePHIs.insert(&PN);
+      IsUsedByRewrittenSpeculatableInstructions = true;
+      return true;
+    }
+
+    return false; // PHIs can't be promoted on their own.
   }
 
   bool visitSelectInst(SelectInst &SI) {
     DEBUG(dbgs() << "    original: " << SI << "\n");
-    IRBuilder<> IRB(&SI);
+    assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
+           "Pointer isn't an operand!");
+    assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
+    assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
 
-    // Find the operand we need to rewrite here.
-    bool IsTrueVal = SI.getTrueValue() == OldPtr;
-    if (IsTrueVal)
-      assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
-    else
-      assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
+    Value *NewPtr = getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType());
+    // Replace the operands which were using the old pointer.
+    if (SI.getOperand(1) == OldPtr)
+      SI.setOperand(1, NewPtr);
+    if (SI.getOperand(2) == OldPtr)
+      SI.setOperand(2, NewPtr);
 
-    Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
-    SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
     DEBUG(dbgs() << "          to: " << SI << "\n");
     deleteIfTriviallyDead(OldPtr);
-    return false;
+
+    // Check whether we can speculate this select instruction, and if so
+    // remember that fact and queue it up for another iteration after the
+    // speculation occurs.
+    if (isSafeSelectToSpeculate(SI, &DL)) {
+      Pass.SpeculatableSelects.insert(&SI);
+      IsUsedByRewrittenSpeculatableInstructions = true;
+      return true;
+    }
+
+    return false; // Selects can't be promoted on their own.
   }
 
 };
@@ -2849,7 +2696,7 @@ class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
   // Befriend the base class so it can delegate to private visit methods.
   friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
 
-  const DataLayout &TD;
+  const DataLayout &DL;
 
   /// Queue of pointer uses to analyze and potentially rewrite.
   SmallVector<Use *, 8> Queue;
@@ -2862,7 +2709,7 @@ class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
   Use *U;
 
 public:
-  AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {}
+  AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
 
   /// Rewrite loads and stores through a pointer and all pointers derived from
   /// it.
@@ -2895,7 +2742,7 @@ private:
   class OpSplitter {
   protected:
     /// The builder used to form new instructions.
-    IRBuilder<> IRB;
+    IRBuilderTy IRB;
     /// The indices which to be used with insert- or extractvalue to select the
     /// appropriate value within the aggregate.
     SmallVector<unsigned, 4> Indices;
@@ -2972,9 +2819,8 @@ private:
     void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
       assert(Ty->isSingleValueType());
       // Load the single value and insert it using the indices.
-      Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
-                                                         Name + ".gep"),
-                                   Name + ".load");
+      Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
+      Value *Load = IRB.CreateLoad(GEP, Name + ".load");
       Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
       DEBUG(dbgs() << "          to: " << *Load << "\n");
     }
@@ -3049,6 +2895,36 @@ private:
 };
 }
 
+/// \brief Strip aggregate type wrapping.
+///
+/// This removes no-op aggregate types wrapping an underlying type. It will
+/// strip as many layers of types as it can without changing either the type
+/// size or the allocated size.
+static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
+  if (Ty->isSingleValueType())
+    return Ty;
+
+  uint64_t AllocSize = DL.getTypeAllocSize(Ty);
+  uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
+
+  Type *InnerTy;
+  if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
+    InnerTy = ArrTy->getElementType();
+  } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
+    const StructLayout *SL = DL.getStructLayout(STy);
+    unsigned Index = SL->getElementContainingOffset(0);
+    InnerTy = STy->getElementType(Index);
+  } else {
+    return Ty;
+  }
+
+  if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
+      TypeSize > DL.getTypeSizeInBits(InnerTy))
+    return Ty;
+
+  return stripAggregateTypeWrapping(DL, InnerTy);
+}
+
 /// \brief Try to find a partition of the aggregate type passed in for a given
 /// offset and size.
 ///
@@ -3062,10 +2938,13 @@ private:
 /// when the size or offset cause either end of type-based partition to be off.
 /// Also, this is a best-effort routine. It is reasonable to give up and not
 /// return a type if necessary.
-static Type *getTypePartition(const DataLayout &TD, Type *Ty,
+static Type *getTypePartition(const DataLayout &DL, Type *Ty,
                               uint64_t Offset, uint64_t Size) {
-  if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
-    return Ty;
+  if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
+    return stripAggregateTypeWrapping(DL, Ty);
+  if (Offset > DL.getTypeAllocSize(Ty) ||
+      (DL.getTypeAllocSize(Ty) - Offset) < Size)
+    return 0;
 
   if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
     // We can't partition pointers...
@@ -3073,14 +2952,15 @@ static Type *getTypePartition(const DataLayout &TD, Type *Ty,
       return 0;
 
     Type *ElementTy = SeqTy->getElementType();
-    uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
+    uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
     uint64_t NumSkippedElements = Offset / ElementSize;
-    if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
+    if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
       if (NumSkippedElements >= ArrTy->getNumElements())
         return 0;
-    if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
+    } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
       if (NumSkippedElements >= VecTy->getNumElements())
         return 0;
+    }
     Offset -= NumSkippedElements * ElementSize;
 
     // First check if we need to recurse.
@@ -3089,12 +2969,12 @@ static Type *getTypePartition(const DataLayout &TD, Type *Ty,
       if ((Offset + Size) > ElementSize)
         return 0;
       // Recurse through the element type trying to peel off offset bytes.
-      return getTypePartition(TD, ElementTy, Offset, Size);
+      return getTypePartition(DL, ElementTy, Offset, Size);
     }
     assert(Offset == 0);
 
     if (Size == ElementSize)
-      return ElementTy;
+      return stripAggregateTypeWrapping(DL, ElementTy);
     assert(Size > ElementSize);
     uint64_t NumElements = Size / ElementSize;
     if (NumElements * ElementSize != Size)
@@ -3106,7 +2986,7 @@ static Type *getTypePartition(const DataLayout &TD, Type *Ty,
   if (!STy)
     return 0;
 
-  const StructLayout *SL = TD.getStructLayout(STy);
+  const StructLayout *SL = DL.getStructLayout(STy);
   if (Offset >= SL->getSizeInBytes())
     return 0;
   uint64_t EndOffset = Offset + Size;
@@ -3117,7 +2997,7 @@ static Type *getTypePartition(const DataLayout &TD, Type *Ty,
   Offset -= SL->getElementOffset(Index);
 
   Type *ElementTy = STy->getElementType(Index);
-  uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
+  uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
   if (Offset >= ElementSize)
     return 0; // The offset points into alignment padding.
 
@@ -3125,12 +3005,12 @@ static Type *getTypePartition(const DataLayout &TD, Type *Ty,
   if (Offset > 0 || Size < ElementSize) {
     if ((Offset + Size) > ElementSize)
       return 0;
-    return getTypePartition(TD, ElementTy, Offset, Size);
+    return getTypePartition(DL, ElementTy, Offset, Size);
   }
   assert(Offset == 0);
 
   if (Size == ElementSize)
-    return ElementTy;
+    return stripAggregateTypeWrapping(DL, ElementTy);
 
   StructType::element_iterator EI = STy->element_begin() + Index,
                                EE = STy->element_end();
@@ -3151,13 +3031,9 @@ static Type *getTypePartition(const DataLayout &TD, Type *Ty,
   }
 
   // Try to build up a sub-structure.
-  SmallVector<Type *, 4> ElementTys;
-  do {
-    ElementTys.push_back(*EI++);
-  } while (EI != EE);
-  StructType *SubTy = StructType::get(STy->getContext(), ElementTys,
+  StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
                                       STy->isPacked());
-  const StructLayout *SubSL = TD.getStructLayout(SubTy);
+  const StructLayout *SubSL = DL.getStructLayout(SubTy);
   if (Size != SubSL->getSizeInBytes())
     return 0; // The sub-struct doesn't have quite the size needed.
 
@@ -3174,113 +3050,280 @@ static Type *getTypePartition(const DataLayout &TD, Type *Ty,
 /// appropriate new offsets. It also evaluates how successful the rewrite was
 /// at enabling promotion and if it was successful queues the alloca to be
 /// promoted.
-bool SROA::rewriteAllocaPartition(AllocaInst &AI,
-                                  AllocaPartitioning &P,
-                                  AllocaPartitioning::iterator PI) {
-  uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
-  bool IsLive = false;
-  for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
-                                        UE = P.use_end(PI);
-       UI != UE && !IsLive; ++UI)
-    if (UI->U)
-      IsLive = true;
-  if (!IsLive)
-    return false; // No live uses left of this partition.
-
-  DEBUG(dbgs() << "Speculating PHIs and selects in partition "
-               << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
-
-  PHIOrSelectSpeculator Speculator(*TD, P, *this);
-  DEBUG(dbgs() << "  speculating ");
-  DEBUG(P.print(dbgs(), PI, ""));
-  Speculator.visitUsers(PI);
+bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
+                            AllocaSlices::iterator B, AllocaSlices::iterator E,
+                            int64_t BeginOffset, int64_t EndOffset,
+                            ArrayRef<AllocaSlices::iterator> SplitUses) {
+  assert(BeginOffset < EndOffset);
+  uint64_t SliceSize = EndOffset - BeginOffset;
 
   // Try to compute a friendly type for this partition of the alloca. This
   // won't always succeed, in which case we fall back to a legal integer type
   // or an i8 array of an appropriate size.
-  Type *AllocaTy = 0;
-  if (Type *PartitionTy = P.getCommonType(PI))
-    if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
-      AllocaTy = PartitionTy;
-  if (!AllocaTy)
-    if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
-                                             PI->BeginOffset, AllocaSize))
-      AllocaTy = PartitionTy;
-  if ((!AllocaTy ||
-       (AllocaTy->isArrayTy() &&
-        AllocaTy->getArrayElementType()->isIntegerTy())) &&
-      TD->isLegalInteger(AllocaSize * 8))
-    AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
-  if (!AllocaTy)
-    AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
-  assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
+  Type *SliceTy = 0;
+  if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
+    if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
+      SliceTy = CommonUseTy;
+  if (!SliceTy)
+    if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
+                                                 BeginOffset, SliceSize))
+      SliceTy = TypePartitionTy;
+  if ((!SliceTy || (SliceTy->isArrayTy() &&
+                    SliceTy->getArrayElementType()->isIntegerTy())) &&
+      DL->isLegalInteger(SliceSize * 8))
+    SliceTy = Type::getIntNTy(*C, SliceSize * 8);
+  if (!SliceTy)
+    SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
+  assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
+
+  bool IsVectorPromotable = isVectorPromotionViable(
+      *DL, SliceTy, S, BeginOffset, EndOffset, B, E, SplitUses);
+
+  bool IsIntegerPromotable =
+      !IsVectorPromotable &&
+      isIntegerWideningViable(*DL, SliceTy, BeginOffset, S, B, E, SplitUses);
 
   // Check for the case where we're going to rewrite to a new alloca of the
   // exact same type as the original, and with the same access offsets. In that
   // case, re-use the existing alloca, but still run through the rewriter to
-  // performe phi and select speculation.
+  // perform phi and select speculation.
   AllocaInst *NewAI;
-  if (AllocaTy == AI.getAllocatedType()) {
-    assert(PI->BeginOffset == 0 &&
+  if (SliceTy == AI.getAllocatedType()) {
+    assert(BeginOffset == 0 &&
            "Non-zero begin offset but same alloca type");
-    assert(PI == P.begin() && "Begin offset is zero on later partition");
     NewAI = &AI;
+    // FIXME: We should be able to bail at this point with "nothing changed".
+    // FIXME: We might want to defer PHI speculation until after here.
   } else {
     unsigned Alignment = AI.getAlignment();
     if (!Alignment) {
       // The minimum alignment which users can rely on when the explicit
       // alignment is omitted or zero is that required by the ABI for this
       // type.
-      Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
+      Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
     }
-    Alignment = MinAlign(Alignment, PI->BeginOffset);
+    Alignment = MinAlign(Alignment, BeginOffset);
     // If we will get at least this much alignment from the type alone, leave
     // the alloca's alignment unconstrained.
-    if (Alignment <= TD->getABITypeAlignment(AllocaTy))
+    if (Alignment <= DL->getABITypeAlignment(SliceTy))
       Alignment = 0;
-    NewAI = new AllocaInst(AllocaTy, 0, Alignment,
-                           AI.getName() + ".sroa." + Twine(PI - P.begin()),
-                           &AI);
+    NewAI = new AllocaInst(SliceTy, 0, Alignment,
+                           AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
     ++NumNewAllocas;
   }
 
   DEBUG(dbgs() << "Rewriting alloca partition "
-               << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
-               << *NewAI << "\n");
+               << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
+               << "\n");
 
-  // Track the high watermark of the post-promotion worklist. We will reset it
-  // to this point if the alloca is not in fact scheduled for promotion.
+  // Track the high watermark on several worklists that are only relevant for
+  // promoted allocas. We will reset it to this point if the alloca is not in
+  // fact scheduled for promotion.
   unsigned PPWOldSize = PostPromotionWorklist.size();
-
-  AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
-                                   PI->BeginOffset, PI->EndOffset);
-  DEBUG(dbgs() << "  rewriting ");
-  DEBUG(P.print(dbgs(), PI, ""));
-  bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
-  if (Promotable) {
+  unsigned SPOldSize = SpeculatablePHIs.size();
+  unsigned SSOldSize = SpeculatableSelects.size();
+  unsigned NumUses = 0;
+
+  AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
+                               EndOffset, IsVectorPromotable,
+                               IsIntegerPromotable);
+  bool Promotable = true;
+  for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
+                                                        SUE = SplitUses.end();
+       SUI != SUE; ++SUI) {
+    DEBUG(dbgs() << "  rewriting split ");
+    DEBUG(S.printSlice(dbgs(), *SUI, ""));
+    Promotable &= Rewriter.visit(*SUI);
+    ++NumUses;
+  }
+  for (AllocaSlices::iterator I = B; I != E; ++I) {
+    DEBUG(dbgs() << "  rewriting ");
+    DEBUG(S.printSlice(dbgs(), I, ""));
+    Promotable &= Rewriter.visit(I);
+    ++NumUses;
+  }
+
+  NumAllocaPartitionUses += NumUses;
+  MaxUsesPerAllocaPartition =
+      std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
+
+  if (Promotable && !Rewriter.isUsedByRewrittenSpeculatableInstructions()) {
     DEBUG(dbgs() << "  and queuing for promotion\n");
     PromotableAllocas.push_back(NewAI);
-  } else if (NewAI != &AI) {
+  } else if (NewAI != &AI ||
+             (Promotable &&
+              Rewriter.isUsedByRewrittenSpeculatableInstructions())) {
     // If we can't promote the alloca, iterate on it to check for new
     // refinements exposed by splitting the current alloca. Don't iterate on an
     // alloca which didn't actually change and didn't get promoted.
+    //
+    // Alternatively, if we could promote the alloca but have speculatable
+    // instructions then we will speculate them after finishing our processing
+    // of the original alloca. Mark the new one for re-visiting in the next
+    // iteration so the speculated operations can be rewritten.
+    //
+    // FIXME: We should actually track whether the rewriter changed anything.
     Worklist.insert(NewAI);
   }
 
   // Drop any post-promotion work items if promotion didn't happen.
-  if (!Promotable)
+  if (!Promotable) {
     while (PostPromotionWorklist.size() > PPWOldSize)
       PostPromotionWorklist.pop_back();
+    while (SpeculatablePHIs.size() > SPOldSize)
+      SpeculatablePHIs.pop_back();
+    while (SpeculatableSelects.size() > SSOldSize)
+      SpeculatableSelects.pop_back();
+  }
 
   return true;
 }
 
-/// \brief Walks the partitioning of an alloca rewriting uses of each partition.
-bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
+namespace {
+struct IsSliceEndLessOrEqualTo {
+  uint64_t UpperBound;
+
+  IsSliceEndLessOrEqualTo(uint64_t UpperBound) : UpperBound(UpperBound) {}
+
+  bool operator()(const AllocaSlices::iterator &I) {
+    return I->endOffset() <= UpperBound;
+  }
+};
+}
+
+static void
+removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
+                        uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
+  if (Offset >= MaxSplitUseEndOffset) {
+    SplitUses.clear();
+    MaxSplitUseEndOffset = 0;
+    return;
+  }
+
+  size_t SplitUsesOldSize = SplitUses.size();
+  SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
+                                 IsSliceEndLessOrEqualTo(Offset)),
+                  SplitUses.end());
+  if (SplitUsesOldSize == SplitUses.size())
+    return;
+
+  // Recompute the max. While this is linear, so is remove_if.
+  MaxSplitUseEndOffset = 0;
+  for (SmallVectorImpl<AllocaSlices::iterator>::iterator
+           SUI = SplitUses.begin(),
+           SUE = SplitUses.end();
+       SUI != SUE; ++SUI)
+    MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
+}
+
+/// \brief Walks the slices of an alloca and form partitions based on them,
+/// rewriting each of their uses.
+bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
+  if (S.begin() == S.end())
+    return false;
+
+  unsigned NumPartitions = 0;
   bool Changed = false;
-  for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
-       ++PI)
-    Changed |= rewriteAllocaPartition(AI, P, PI);
+  SmallVector<AllocaSlices::iterator, 4> SplitUses;
+  uint64_t MaxSplitUseEndOffset = 0;
+
+  uint64_t BeginOffset = S.begin()->beginOffset();
+
+  for (AllocaSlices::iterator SI = S.begin(), SJ = llvm::next(SI), SE = S.end();
+       SI != SE; SI = SJ) {
+    uint64_t MaxEndOffset = SI->endOffset();
+
+    if (!SI->isSplittable()) {
+      // When we're forming an unsplittable region, it must always start at the
+      // first slice and will extend through its end.
+      assert(BeginOffset == SI->beginOffset());
+
+      // Form a partition including all of the overlapping slices with this
+      // unsplittable slice.
+      while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
+        if (!SJ->isSplittable())
+          MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
+        ++SJ;
+      }
+    } else {
+      assert(SI->isSplittable()); // Established above.
+
+      // Collect all of the overlapping splittable slices.
+      while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
+             SJ->isSplittable()) {
+        MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
+        ++SJ;
+      }
+
+      // Back up MaxEndOffset and SJ if we ended the span early when
+      // encountering an unsplittable slice.
+      if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
+        assert(!SJ->isSplittable());
+        MaxEndOffset = SJ->beginOffset();
+      }
+    }
+
+    // Check if we have managed to move the end offset forward yet. If so,
+    // we'll have to rewrite uses and erase old split uses.
+    if (BeginOffset < MaxEndOffset) {
+      // Rewrite a sequence of overlapping slices.
+      Changed |=
+          rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
+      ++NumPartitions;
+
+      removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
+    }
+
+    // Accumulate all the splittable slices from the [SI,SJ) region which
+    // overlap going forward.
+    for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
+      if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
+        SplitUses.push_back(SK);
+        MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
+      }
+
+    // If we're already at the end and we have no split uses, we're done.
+    if (SJ == SE && SplitUses.empty())
+      break;
+
+    // If we have no split uses or no gap in offsets, we're ready to move to
+    // the next slice.
+    if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
+      BeginOffset = SJ->beginOffset();
+      continue;
+    }
+
+    // Even if we have split slices, if the next slice is splittable and the
+    // split slices reach it, we can simply set up the beginning offset of the
+    // next iteration to bridge between them.
+    if (SJ != SE && SJ->isSplittable() &&
+        MaxSplitUseEndOffset > SJ->beginOffset()) {
+      BeginOffset = MaxEndOffset;
+      continue;
+    }
+
+    // Otherwise, we have a tail of split slices. Rewrite them with an empty
+    // range of slices.
+    uint64_t PostSplitEndOffset =
+        SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
+
+    Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
+                                SplitUses);
+    ++NumPartitions;
+
+    if (SJ == SE)
+      break; // Skip the rest, we don't need to do any cleanup.
+
+    removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
+                            PostSplitEndOffset);
+
+    // Now just reset the begin offset for the next iteration.
+    BeginOffset = SJ->beginOffset();
+  }
+
+  NumAllocaPartitions += NumPartitions;
+  MaxPartitionsPerAlloca =
+      std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
 
   return Changed;
 }
@@ -3288,7 +3331,7 @@ bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
 /// \brief Analyze an alloca for SROA.
 ///
 /// This analyzes the alloca to ensure we can reason about it, builds
-/// a partitioning of the alloca, and then hands it off to be split and
+/// the slices of the alloca, and then hands it off to be split and
 /// rewritten as needed.
 bool SROA::runOnAlloca(AllocaInst &AI) {
   DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
@@ -3302,32 +3345,32 @@ bool SROA::runOnAlloca(AllocaInst &AI) {
 
   // Skip alloca forms that this analysis can't handle.
   if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
-      TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
+      DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
     return false;
 
   bool Changed = false;
 
   // First, split any FCA loads and stores touching this alloca to promote
   // better splitting and promotion opportunities.
-  AggLoadStoreRewriter AggRewriter(*TD);
+  AggLoadStoreRewriter AggRewriter(*DL);
   Changed |= AggRewriter.rewrite(AI);
 
-  // Build the partition set using a recursive instruction-visiting builder.
-  AllocaPartitioning P(*TD, AI);
-  DEBUG(P.print(dbgs()));
-  if (P.isEscaped())
+  // Build the slices using a recursive instruction-visiting builder.
+  AllocaSlices S(*DL, AI);
+  DEBUG(S.print(dbgs()));
+  if (S.isEscaped())
     return Changed;
 
   // Delete all the dead users of this alloca before splitting and rewriting it.
-  for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
-                                              DE = P.dead_user_end();
+  for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
+                                        DE = S.dead_user_end();
        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();
+  for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
+                                      DE = S.dead_op_end();
        DO != DE; ++DO) {
     Value *OldV = **DO;
     // Clobber the use with an undef value.
@@ -3335,15 +3378,25 @@ bool SROA::runOnAlloca(AllocaInst &AI) {
     if (Instruction *OldI = dyn_cast<Instruction>(OldV))
       if (isInstructionTriviallyDead(OldI)) {
         Changed = true;
-        DeadInsts.push_back(OldI);
+        DeadInsts.insert(OldI);
       }
   }
 
-  // No partitions to split. Leave the dead alloca for a later pass to clean up.
-  if (P.begin() == P.end())
+  // No slices to split. Leave the dead alloca for a later pass to clean up.
+  if (S.begin() == S.end())
     return Changed;
 
-  return splitAlloca(AI, P) || Changed;
+  Changed |= splitAlloca(AI, S);
+
+  DEBUG(dbgs() << "  Speculating PHIs\n");
+  while (!SpeculatablePHIs.empty())
+    speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
+
+  DEBUG(dbgs() << "  Speculating Selects\n");
+  while (!SpeculatableSelects.empty())
+    speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
+
+  return Changed;
 }
 
 /// \brief Delete the dead instructions accumulated in this run.
@@ -3356,17 +3409,18 @@ 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");
 
+    I->replaceAllUsesWith(UndefValue::get(I->getType()));
+
     for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
       if (Instruction *U = dyn_cast<Instruction>(*OI)) {
         // 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))
@@ -3377,6 +3431,15 @@ void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
   }
 }
 
+static void enqueueUsersInWorklist(Instruction &I,
+                                   SmallVectorImpl<Instruction *> &Worklist,
+                                   SmallPtrSet<Instruction *, 8> &Visited) {
+  for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
+       ++UI)
+    if (Visited.insert(cast<Instruction>(*UI)))
+      Worklist.push_back(cast<Instruction>(*UI));
+}
+
 /// \brief Promote the allocas, using the best available technique.
 ///
 /// This attempts to promote whatever allocas have been identified as viable in
@@ -3384,7 +3447,7 @@ void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
 /// If there is a domtree available, we attempt to promote using the full power
 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
 /// based on the SSAUpdater utilities. This function returns whether any
-/// promotion occured.
+/// promotion occurred.
 bool SROA::promoteAllocas(Function &F) {
   if (PromotableAllocas.empty())
     return false;
@@ -3401,25 +3464,28 @@ bool SROA::promoteAllocas(Function &F) {
   DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
   SSAUpdater SSA;
   DIBuilder DIB(*F.getParent());
-  SmallVector<Instruction*, 64> Insts;
+  SmallVector<Instruction *, 64> Insts;
+
+  // We need a worklist to walk the uses of each alloca.
+  SmallVector<Instruction *, 8> Worklist;
+  SmallPtrSet<Instruction *, 8> Visited;
+  SmallVector<Instruction *, 32> DeadInsts;
 
   for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
     AllocaInst *AI = PromotableAllocas[Idx];
-    for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
-         UI != UE;) {
-      Instruction *I = cast<Instruction>(*UI++);
+    Insts.clear();
+    Worklist.clear();
+    Visited.clear();
+
+    enqueueUsersInWorklist(*AI, Worklist, Visited);
+
+    while (!Worklist.empty()) {
+      Instruction *I = Worklist.pop_back_val();
+
       // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
       // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
       // leading to them) here. Eventually it should use them to optimize the
       // scalar values produced.
-      if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
-        assert(onlyUsedByLifetimeMarkers(I) &&
-               "Found a bitcast used outside of a lifetime marker.");
-        while (!I->use_empty())
-          cast<Instruction>(*I->use_begin())->eraseFromParent();
-        I->eraseFromParent();
-        continue;
-      }
       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
         assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
                II->getIntrinsicID() == Intrinsic::lifetime_end);
@@ -3427,10 +3493,30 @@ bool SROA::promoteAllocas(Function &F) {
         continue;
       }
 
-      Insts.push_back(I);
+      // Push the loads and stores we find onto the list. SROA will already
+      // have validated that all loads and stores are viable candidates for
+      // promotion.
+      if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
+        assert(LI->getType() == AI->getAllocatedType());
+        Insts.push_back(LI);
+        continue;
+      }
+      if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
+        assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
+        Insts.push_back(SI);
+        continue;
+      }
+
+      // For everything else, we know that only no-op bitcasts and GEPs will
+      // make it this far, just recurse through them and recall them for later
+      // removal.
+      DeadInsts.push_back(I);
+      enqueueUsersInWorklist(*I, Worklist, Visited);
     }
     AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
-    Insts.clear();
+    while (!DeadInsts.empty())
+      DeadInsts.pop_back_val()->eraseFromParent();
+    AI->eraseFromParent();
   }
 
   PromotableAllocas.clear();
@@ -3454,8 +3540,8 @@ namespace {
 bool SROA::runOnFunction(Function &F) {
   DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
   C = &F.getContext();
-  TD = getAnalysisIfAvailable<DataLayout>();
-  if (!TD) {
+  DL = getAnalysisIfAvailable<DataLayout>();
+  if (!DL) {
     DEBUG(dbgs() << "  Skipping SROA -- no target data!\n");
     return false;
   }