Add an optimization that does CSE in a group of similar GEPs.

This optimization merges the common part of a group of GEPs, so we can compute
each pointer address by adding a simple offset to the common part.

The optimization is currently only enabled for the NVPTX backend, where it has
a large payoff on some benchmarks.

Review: http://reviews.llvm.org/D3462

Patch by Jingyue Wu.




git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@207783 91177308-0d34-0410-b5e6-96231b3b80d8

9 files changed, 774 insertions, 4 deletions

diff --git a/include/llvm/InitializePasses.h b/include/llvm/InitializePasses.h
index 232e422de1..8e536159db 100644
--- a/include/llvm/InitializePasses.h
+++ b/include/llvm/InitializePasses.h
@@ -238,6 +238,7 @@ void initializeSimpleInlinerPass(PassRegistry&);
 void initializeRegisterCoalescerPass(PassRegistry&);
 void initializeSingleLoopExtractorPass(PassRegistry&);
 void initializeSinkingPass(PassRegistry&);
+void initializeSeparateConstOffsetFromGEPPass(PassRegistry &);
 void initializeSlotIndexesPass(PassRegistry&);
 void initializeSpillPlacementPass(PassRegistry&);
 void initializeStackProtectorPass(PassRegistry&);
diff --git a/include/llvm/LinkAllPasses.h b/include/llvm/LinkAllPasses.h
index 9cb1c5c8a8..2616ebd1fa 100644
--- a/include/llvm/LinkAllPasses.h
+++ b/include/llvm/LinkAllPasses.h
@@ -156,6 +156,7 @@ namespace {
       (void) llvm::createBBVectorizePass();
       (void) llvm::createPartiallyInlineLibCallsPass();
       (void) llvm::createScalarizerPass();
+      (void) llvm::createSeparateConstOffsetFromGEPPass();
 
       (void)new llvm::IntervalPartition();
       (void)new llvm::FindUsedTypes();
diff --git a/include/llvm/Transforms/Scalar.h b/include/llvm/Transforms/Scalar.h
index 453de03972..6aea643c42 100644
--- a/include/llvm/Transforms/Scalar.h
+++ b/include/llvm/Transforms/Scalar.h
@@ -377,6 +377,12 @@ FunctionPass *createScalarizerPass();
 // AddDiscriminators - Add DWARF path discriminators to the IR.
 FunctionPass *createAddDiscriminatorsPass();
 
+//===----------------------------------------------------------------------===//
+//
+// SeparateConstOffsetFromGEP - Split GEPs for better CSE
+//
+FunctionPass *createSeparateConstOffsetFromGEPPass();
+
 } // End llvm namespace
 
 #endif
diff --git a/lib/Target/NVPTX/NVPTXTargetMachine.cpp b/lib/Target/NVPTX/NVPTXTargetMachine.cpp
index 0cc5c51629..26a4f84052 100644
--- a/lib/Target/NVPTX/NVPTXTargetMachine.cpp
+++ b/lib/Target/NVPTX/NVPTXTargetMachine.cpp
@@ -147,10 +147,23 @@ void NVPTXPassConfig::addIRPasses() {
   addPass(createNVPTXAssignValidGlobalNamesPass());
   addPass(createGenericToNVVMPass());
   addPass(createNVPTXFavorNonGenericAddrSpacesPass());
-  // The FavorNonGenericAddrSpaces pass may remove instructions and leave some
-  // values unused. Therefore, we run a DCE pass right afterwards. We could
-  // remove unused values in an ad-hoc manner, but it requires manual work and
-  // might be error-prone.
+  addPass(createSeparateConstOffsetFromGEPPass());
+  // The SeparateConstOffsetFromGEP pass creates variadic bases that can be used
+  // by multiple GEPs. Run GVN or EarlyCSE to really reuse them. GVN generates
+  // significantly better code than EarlyCSE for some of our benchmarks.
+  if (getOptLevel() == CodeGenOpt::Aggressive)
+    addPass(createGVNPass());
+  else
+    addPass(createEarlyCSEPass());
+  // Both FavorNonGenericAddrSpaces and SeparateConstOffsetFromGEP may leave
+  // some dead code.  We could remove dead code in an ad-hoc manner, but that
+  // requires manual work and might be error-prone.
+  //
+  // The FavorNonGenericAddrSpaces pass shortcuts unnecessary addrspacecasts,
+  // and leave them unused.
+  //
+  // SeparateConstOffsetFromGEP rebuilds a new index from the old index, and the
+  // old index and some of its intermediate results may become unused.
   addPass(createDeadCodeEliminationPass());
 }
 
diff --git a/lib/Transforms/Scalar/Scalar.cpp b/lib/Transforms/Scalar/Scalar.cpp
index 09167b9e82..f8f828c840 100644
--- a/lib/Transforms/Scalar/Scalar.cpp
+++ b/lib/Transforms/Scalar/Scalar.cpp
@@ -64,6 +64,7 @@ void llvm::initializeScalarOpts(PassRegistry &Registry) {
   initializeStructurizeCFGPass(Registry);
   initializeSinkingPass(Registry);
   initializeTailCallElimPass(Registry);
+  initializeSeparateConstOffsetFromGEPPass(Registry);
 }
 
 void LLVMInitializeScalarOpts(LLVMPassRegistryRef R) {
diff --git a/lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp b/lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp
new file mode 100644
index 0000000000..0465f237ec
--- /dev/null
+++ b/lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp
@@ -0,0 +1,583 @@
+//===-- SeparateConstOffsetFromGEP.cpp - ------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// Loop unrolling may create many similar GEPs for array accesses.
+// e.g., a 2-level loop
+//
+// float a[32][32]; // global variable
+//
+// for (int i = 0; i < 2; ++i) {
+//   for (int j = 0; j < 2; ++j) {
+//     ...
+//     ... = a[x + i][y + j];
+//     ...
+//   }
+// }
+//
+// will probably be unrolled to:
+//
+// gep %a, 0, %x, %y; load
+// gep %a, 0, %x, %y + 1; load
+// gep %a, 0, %x + 1, %y; load
+// gep %a, 0, %x + 1, %y + 1; load
+//
+// LLVM's GVN does not use partial redundancy elimination yet, and is thus
+// unable to reuse (gep %a, 0, %x, %y). As a result, this misoptimization incurs
+// significant slowdown in targets with limited addressing modes. For instance,
+// because the PTX target does not support the reg+reg addressing mode, the
+// NVPTX backend emits PTX code that literally computes the pointer address of
+// each GEP, wasting tons of registers. It emits the following PTX for the
+// first load and similar PTX for other loads.
+//
+// mov.u32         %r1, %x;
+// mov.u32         %r2, %y;
+// mul.wide.u32    %rl2, %r1, 128;
+// mov.u64         %rl3, a;
+// add.s64         %rl4, %rl3, %rl2;
+// mul.wide.u32    %rl5, %r2, 4;
+// add.s64         %rl6, %rl4, %rl5;
+// ld.global.f32   %f1, [%rl6];
+//
+// To reduce the register pressure, the optimization implemented in this file
+// merges the common part of a group of GEPs, so we can compute each pointer
+// address by adding a simple offset to the common part, saving many registers.
+//
+// It works by splitting each GEP into a variadic base and a constant offset.
+// The variadic base can be computed once and reused by multiple GEPs, and the
+// constant offsets can be nicely folded into the reg+immediate addressing mode
+// (supported by most targets) without using any extra register.
+//
+// For instance, we transform the four GEPs and four loads in the above example
+// into:
+//
+// base = gep a, 0, x, y
+// load base
+// laod base + 1  * sizeof(float)
+// load base + 32 * sizeof(float)
+// load base + 33 * sizeof(float)
+//
+// Given the transformed IR, a backend that supports the reg+immediate
+// addressing mode can easily fold the pointer arithmetics into the loads. For
+// example, the NVPTX backend can easily fold the pointer arithmetics into the
+// ld.global.f32 instructions, and the resultant PTX uses much fewer registers.
+//
+// mov.u32         %r1, %tid.x;
+// mov.u32         %r2, %tid.y;
+// mul.wide.u32    %rl2, %r1, 128;
+// mov.u64         %rl3, a;
+// add.s64         %rl4, %rl3, %rl2;
+// mul.wide.u32    %rl5, %r2, 4;
+// add.s64         %rl6, %rl4, %rl5;
+// ld.global.f32   %f1, [%rl6]; // so far the same as unoptimized PTX
+// ld.global.f32   %f2, [%rl6+4]; // much better
+// ld.global.f32   %f3, [%rl6+128]; // much better
+// ld.global.f32   %f4, [%rl6+132]; // much better
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Operator.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Transforms/Scalar.h"
+
+using namespace llvm;
+
+static cl::opt<bool> DisableSeparateConstOffsetFromGEP(
+    "disable-separate-const-offset-from-gep", cl::init(false),
+    cl::desc("Do not separate the constant offset from a GEP instruction"),
+    cl::Hidden);
+
+namespace {
+
+/// \brief A helper class for separating a constant offset from a GEP index.
+///
+/// In real programs, a GEP index may be more complicated than a simple addition
+/// of something and a constant integer which can be trivially splitted. For
+/// example, to split ((a << 3) | 5) + b, we need to search deeper for the
+/// constant offset, so that we can seperate the index to (a << 3) + b and 5.
+///
+/// Therefore, this class looks into the expression that computes a given GEP
+/// index, and tries to find a constant integer that can be hoisted to the
+/// outermost level of the expression as an addition. Not every constant in an
+/// expression can jump out. e.g., we cannot transform (b * (a + 5)) to (b * a +
+/// 5); nor can we transform (3 * (a + 5)) to (3 * a + 5), however in this case,
+/// -instcombine probably already optimized (3 * (a + 5)) to (3 * a + 15).
+class ConstantOffsetExtractor {
+ public:
+  /// Extracts a constant offset from the given GEP index. It outputs the
+  /// numeric value of the extracted constant offset (0 if failed), and a
+  /// new index representing the remainder (equal to the original index minus
+  /// the constant offset).
+  /// \p Idx The given GEP index
+  /// \p NewIdx The new index to replace
+  /// \p DL The datalayout of the module
+  /// \p IP Calculating the new index requires new instructions. IP indicates
+  /// where to insert them (typically right before the GEP).
+  static int64_t Extract(Value *Idx, Value *&NewIdx, const DataLayout *DL,
+                         Instruction *IP);
+  /// Looks for a constant offset without extracting it. The meaning of the
+  /// arguments and the return value are the same as Extract.
+  static int64_t Find(Value *Idx, const DataLayout *DL);
+
+ private:
+  ConstantOffsetExtractor(const DataLayout *Layout, Instruction *InsertionPt)
+      : DL(Layout), IP(InsertionPt) {}
+  /// Searches the expression that computes V for a constant offset. If the
+  /// searching is successful, update UserChain as a path from V to the constant
+  /// offset.
+  int64_t find(Value *V);
+  /// A helper function to look into both operands of a binary operator U.
+  /// \p IsSub Whether U is a sub operator. If so, we need to negate the
+  /// constant offset at some point.
+  int64_t findInEitherOperand(User *U, bool IsSub);
+  /// After finding the constant offset and how it is reached from the GEP
+  /// index, we build a new index which is a clone of the old one except the
+  /// constant offset is removed. For example, given (a + (b + 5)) and knowning
+  /// the constant offset is 5, this function returns (a + b).
+  ///
+  /// We cannot simply change the constant to zero because the expression that
+  /// computes the index or its intermediate result may be used by others.
+  Value *rebuildWithoutConstantOffset();
+  // A helper function for rebuildWithoutConstantOffset that rebuilds the direct
+  // user (U) of the constant offset (C).
+  Value *rebuildLeafWithoutConstantOffset(User *U, Value *C);
+  /// Returns a clone of U except the first occurrence of From with To.
+  Value *cloneAndReplace(User *U, Value *From, Value *To);
+
+  /// Returns true if LHS and RHS have no bits in common, i.e., LHS | RHS == 0.
+  bool NoCommonBits(Value *LHS, Value *RHS) const;
+  /// Computes which bits are known to be one or zero.
+  /// \p KnownOne Mask of all bits that are known to be one.
+  /// \p KnownZero Mask of all bits that are known to be zero.
+  void ComputeKnownBits(Value *V, APInt &KnownOne, APInt &KnownZero) const;
+  /// Finds the first use of Used in U. Returns -1 if not found.
+  static unsigned FindFirstUse(User *U, Value *Used);
+
+  /// The path from the constant offset to the old GEP index. e.g., if the GEP
+  /// index is "a * b + (c + 5)". After running function find, UserChain[0] will
+  /// be the constant 5, UserChain[1] will be the subexpression "c + 5", and
+  /// UserChain[2] will be the entire expression "a * b + (c + 5)".
+  ///
+  /// This path helps rebuildWithoutConstantOffset rebuild the new GEP index.
+  SmallVector<User *, 8> UserChain;
+  /// The data layout of the module. Used in ComputeKnownBits.
+  const DataLayout *DL;
+  Instruction *IP;  /// Insertion position of cloned instructions.
+};
+
+/// \brief A pass that tries to split every GEP in the function into a variadic
+/// base and a constant offset. It is a FuntionPass because searching for the
+/// constant offset may inspect other basic blocks.
+class SeparateConstOffsetFromGEP : public FunctionPass {
+ public:
+  static char ID;
+  SeparateConstOffsetFromGEP() : FunctionPass(ID) {
+    initializeSeparateConstOffsetFromGEPPass(*PassRegistry::getPassRegistry());
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<DataLayoutPass>();
+    AU.addRequired<TargetTransformInfo>();
+  }
+  bool runOnFunction(Function &F) override;
+
+ private:
+  /// Tries to split the given GEP into a variadic base and a constant offset,
+  /// and returns true if the splitting succeeds.
+  bool splitGEP(GetElementPtrInst *GEP);
+  /// Finds the constant offset within each index, and accumulates them. This
+  /// function only inspects the GEP without changing it. The output
+  /// NeedsExtraction indicates whether we can extract a non-zero constant
+  /// offset from any index.
+  int64_t accumulateByteOffset(GetElementPtrInst *GEP, const DataLayout *DL,
+                               bool &NeedsExtraction);
+};
+}  // anonymous namespace
+
+char SeparateConstOffsetFromGEP::ID = 0;
+INITIALIZE_PASS_BEGIN(
+    SeparateConstOffsetFromGEP, "separate-const-offset-from-gep",
+    "Split GEPs to a variadic base and a constant offset for better CSE", false,
+    false)
+INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
+INITIALIZE_PASS_DEPENDENCY(DataLayoutPass)
+INITIALIZE_PASS_END(
+    SeparateConstOffsetFromGEP, "separate-const-offset-from-gep",
+    "Split GEPs to a variadic base and a constant offset for better CSE", false,
+    false)
+
+FunctionPass *llvm::createSeparateConstOffsetFromGEPPass() {
+  return new SeparateConstOffsetFromGEP();
+}
+
+int64_t ConstantOffsetExtractor::findInEitherOperand(User *U, bool IsSub) {
+  assert(U->getNumOperands() == 2);
+  int64_t ConstantOffset = find(U->getOperand(0));
+  // If we found a constant offset in the left operand, stop and return that.
+  // This shortcut might cause us to miss opportunities of combining the
+  // constant offsets in both operands, e.g., (a + 4) + (b + 5) => (a + b) + 9.
+  // However, such cases are probably already handled by -instcombine,
+  // given this pass runs after the standard optimizations.
+  if (ConstantOffset != 0) return ConstantOffset;
+  ConstantOffset = find(U->getOperand(1));
+  // If U is a sub operator, negate the constant offset found in the right
+  // operand.
+  return IsSub ? -ConstantOffset : ConstantOffset;
+}
+
+int64_t ConstantOffsetExtractor::find(Value *V) {
+  // TODO(jingyue): We can even trace into integer/pointer casts, such as
+  // inttoptr, ptrtoint, bitcast, and addrspacecast. We choose to handle only
+  // integers because it gives good enough results for our benchmarks.
+  assert(V->getType()->isIntegerTy());
+
+  User *U = dyn_cast<User>(V);
+  // We cannot do much with Values that are not a User, such as BasicBlock and
+  // MDNode.
+  if (U == nullptr) return 0;
+
+  int64_t ConstantOffset = 0;
+  if (ConstantInt *CI = dyn_cast<ConstantInt>(U)) {
+    // Hooray, we found it!
+    ConstantOffset = CI->getSExtValue();
+  } else if (Operator *O = dyn_cast<Operator>(U)) {
+    // The GEP index may be more complicated than a simple addition of a
+    // varaible and a constant. Therefore, we trace into subexpressions for more
+    // hoisting opportunities.
+    switch (O->getOpcode()) {
+      case Instruction::Add: {
+        ConstantOffset = findInEitherOperand(U, false);
+        break;
+      }
+      case Instruction::Sub: {
+        ConstantOffset = findInEitherOperand(U, true);
+        break;
+      }
+      case Instruction::Or: {
+        // If LHS and RHS don't have common bits, (LHS | RHS) is equivalent to
+        // (LHS + RHS).
+        if (NoCommonBits(U->getOperand(0), U->getOperand(1)))
+          ConstantOffset = findInEitherOperand(U, false);
+        break;
+      }
+      case Instruction::SExt: {
+        // For safety, we trace into sext only when its operand is marked
+        // "nsw" because xxx.nsw guarantees no signed wrap. e.g., we can safely
+        // transform "sext (add nsw a, 5)" into "add nsw (sext a), 5".
+        if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0))) {
+          if (BO->hasNoSignedWrap())
+            ConstantOffset = find(U->getOperand(0));
+        }
+        break;
+      }
+      case Instruction::ZExt: {
+        // Similarly, we trace into zext only when its operand is marked with
+        // "nuw" because zext (add nuw a, b) == add nuw (zext a), (zext b).
+        if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0))) {
+          if (BO->hasNoUnsignedWrap())
+            ConstantOffset = find(U->getOperand(0));
+        }
+        break;
+      }
+    }
+  }
+  // If we found a non-zero constant offset, adds it to the path for future
+  // transformation (rebuildWithoutConstantOffset). Zero is a valid constant
+  // offset, but doesn't help this optimization.
+  if (ConstantOffset != 0)
+    UserChain.push_back(U);
+  return ConstantOffset;
+}
+
+unsigned ConstantOffsetExtractor::FindFirstUse(User *U, Value *Used) {
+  for (unsigned I = 0, E = U->getNumOperands(); I < E; ++I) {
+    if (U->getOperand(I) == Used)
+      return I;
+  }
+  return -1;
+}
+
+Value *ConstantOffsetExtractor::cloneAndReplace(User *U, Value *From,
+                                                Value *To) {
+  // Finds in U the first use of From. It is safe to ignore future occurrences
+  // of From, because findInEitherOperand similarly stops searching the right
+  // operand when the first operand has a non-zero constant offset.
+  unsigned OpNo = FindFirstUse(U, From);
+  assert(OpNo != (unsigned)-1 && "UserChain wasn't built correctly");
+
+  // ConstantOffsetExtractor::find only follows Operators (i.e., Instructions
+  // and ConstantExprs). Therefore, U is either an Instruction or a
+  // ConstantExpr.
+  if (Instruction *I = dyn_cast<Instruction>(U)) {
+    Instruction *Clone = I->clone();
+    Clone->setOperand(OpNo, To);
+    Clone->insertBefore(IP);
+    return Clone;
+  }
+  // cast<Constant>(To) is safe because a ConstantExpr only uses Constants.
+  return cast<ConstantExpr>(U)
+      ->getWithOperandReplaced(OpNo, cast<Constant>(To));
+}
+
+Value *ConstantOffsetExtractor::rebuildLeafWithoutConstantOffset(User *U,
+                                                                 Value *C) {
+  assert(U->getNumOperands() <= 2 &&
+         "We didn't trace into any operator with more than 2 operands");
+  // If U has only one operand which is the constant offset, removing the
+  // constant offset leaves U as a null value.
+  if (U->getNumOperands() == 1)
+    return Constant::getNullValue(U->getType());
+
+  // U->getNumOperands() == 2
+  unsigned OpNo = FindFirstUse(U, C); // U->getOperand(OpNo) == C
+  assert(OpNo < 2 && "UserChain wasn't built correctly");
+  Value *TheOther = U->getOperand(1 - OpNo); // The other operand of U
+  // If U = C - X, removing C makes U = -X; otherwise U will simply be X.
+  if (!isa<SubOperator>(U) || OpNo == 1)
+    return TheOther;
+  if (isa<ConstantExpr>(U))
+    return ConstantExpr::getNeg(cast<Constant>(TheOther));
+  return BinaryOperator::CreateNeg(TheOther, "", IP);
+}
+
+Value *ConstantOffsetExtractor::rebuildWithoutConstantOffset() {
+  assert(UserChain.size() > 0 && "you at least found a constant, right?");
+  // Start with the constant and go up through UserChain, each time building a
+  // clone of the subexpression but with the constant removed.
+  // e.g., to build a clone of (a + (b + (c + 5)) but with the 5 removed, we
+  // first c, then (b + c), and finally (a + (b + c)).
+  //
+  // Fast path: if the GEP index is a constant, simply returns 0.
+  if (UserChain.size() == 1)
+    return ConstantInt::get(UserChain[0]->getType(), 0);
+
+  Value *Remainder =
+      rebuildLeafWithoutConstantOffset(UserChain[1], UserChain[0]);
+  for (size_t I = 2; I < UserChain.size(); ++I)
+    Remainder = cloneAndReplace(UserChain[I], UserChain[I - 1], Remainder);
+  return Remainder;
+}
+
+int64_t ConstantOffsetExtractor::Extract(Value *Idx, Value *&NewIdx,
+                                         const DataLayout *DL,
+                                         Instruction *IP) {
+  ConstantOffsetExtractor Extractor(DL, IP);
+  // Find a non-zero constant offset first.
+  int64_t ConstantOffset = Extractor.find(Idx);
+  if (ConstantOffset == 0)
+    return 0;
+  // Then rebuild a new index with the constant removed.
+  NewIdx = Extractor.rebuildWithoutConstantOffset();
+  return ConstantOffset;
+}
+
+int64_t ConstantOffsetExtractor::Find(Value *Idx, const DataLayout *DL) {
+  return ConstantOffsetExtractor(DL, nullptr).find(Idx);
+}
+
+void ConstantOffsetExtractor::ComputeKnownBits(Value *V, APInt &KnownOne,
+                                               APInt &KnownZero) const {
+  IntegerType *IT = cast<IntegerType>(V->getType());
+  KnownOne = APInt(IT->getBitWidth(), 0);
+  KnownZero = APInt(IT->getBitWidth(), 0);
+  llvm::ComputeMaskedBits(V, KnownZero, KnownOne, DL, 0);
+}
+
+bool ConstantOffsetExtractor::NoCommonBits(Value *LHS, Value *RHS) const {
+  assert(LHS->getType() == RHS->getType() &&
+         "LHS and RHS should have the same type");
+  APInt LHSKnownOne, LHSKnownZero, RHSKnownOne, RHSKnownZero;
+  ComputeKnownBits(LHS, LHSKnownOne, LHSKnownZero);
+  ComputeKnownBits(RHS, RHSKnownOne, RHSKnownZero);
+  return (LHSKnownZero | RHSKnownZero).isAllOnesValue();
+}
+
+int64_t SeparateConstOffsetFromGEP::accumulateByteOffset(
+    GetElementPtrInst *GEP, const DataLayout *DL, bool &NeedsExtraction) {
+  NeedsExtraction = false;
+  int64_t AccumulativeByteOffset = 0;
+  gep_type_iterator GTI = gep_type_begin(*GEP);
+  for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
+    if (isa<SequentialType>(*GTI)) {
+      // Tries to extract a constant offset from this GEP index.
+      int64_t ConstantOffset =
+          ConstantOffsetExtractor::Find(GEP->getOperand(I), DL);
+      if (ConstantOffset != 0) {
+        NeedsExtraction = true;
+        // A GEP may have multiple indices.  We accumulate the extracted
+        // constant offset to a byte offset, and later offset the remainder of
+        // the original GEP with this byte offset.
+        AccumulativeByteOffset +=
+            ConstantOffset * DL->getTypeAllocSize(GTI.getIndexedType());
+      }
+    }
+  }
+  return AccumulativeByteOffset;
+}
+
+bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
+  // Skip vector GEPs.
+  if (GEP->getType()->isVectorTy())
+    return false;
+
+  // The backend can already nicely handle the case where all indices are
+  // constant.
+  if (GEP->hasAllConstantIndices())
+    return false;
+
+  bool Changed = false;
+
+  // Shortcuts integer casts. Eliminating these explicit casts can make
+  // subsequent optimizations more obvious: ConstantOffsetExtractor needn't
+  // trace into these casts.
+  if (GEP->isInBounds()) {
+    // Doing this to inbounds GEPs is safe because their indices are guaranteed
+    // to be non-negative and in bounds.
+    gep_type_iterator GTI = gep_type_begin(*GEP);
+    for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
+      if (isa<SequentialType>(*GTI)) {
+        if (Operator *O = dyn_cast<Operator>(GEP->getOperand(I))) {
+          if (O->getOpcode() == Instruction::SExt ||
+              O->getOpcode() == Instruction::ZExt) {
+            GEP->setOperand(I, O->getOperand(0));
+            Changed = true;
+          }
+        }
+      }
+    }
+  }
+
+  const DataLayout *DL = &getAnalysis<DataLayoutPass>().getDataLayout();
+  bool NeedsExtraction;
+  int64_t AccumulativeByteOffset =
+      accumulateByteOffset(GEP, DL, NeedsExtraction);
+
+  if (!NeedsExtraction)
+    return Changed;
+  // Before really splitting the GEP, check whether the backend supports the
+  // addressing mode we are about to produce. If no, this splitting probably
+  // won't be beneficial.
+  TargetTransformInfo &TTI = getAnalysis<TargetTransformInfo>();
+  if (!TTI.isLegalAddressingMode(GEP->getType()->getElementType(),
+                                 /*BaseGV=*/nullptr, AccumulativeByteOffset,
+                                 /*HasBaseReg=*/true, /*Scale=*/0)) {
+    return Changed;
+  }
+
+  // Remove the constant offset in each GEP index. The resultant GEP computes
+  // the variadic base.
+  gep_type_iterator GTI = gep_type_begin(*GEP);
+  for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
+    if (isa<SequentialType>(*GTI)) {
+      Value *NewIdx = nullptr;
+      // Tries to extract a constant offset from this GEP index.
+      int64_t ConstantOffset =
+          ConstantOffsetExtractor::Extract(GEP->getOperand(I), NewIdx, DL, GEP);
+      if (ConstantOffset != 0) {
+        assert(NewIdx && "ConstantOffset != 0 implies NewIdx is set");
+        GEP->setOperand(I, NewIdx);
+        // Clear the inbounds attribute because the new index may be off-bound.
+        // e.g.,
+        //
+        // b = add i64 a, 5
+        // addr = gep inbounds float* p, i64 b
+        //
+        // is transformed to:
+        //
+        // addr2 = gep float* p, i64 a
+        // addr = gep float* addr2, i64 5
+        //
+        // If a is -4, although the old index b is in bounds, the new index a is
+        // off-bound. http://llvm.org/docs/LangRef.html#id181 says "if the
+        // inbounds keyword is not present, the offsets are added to the base
+        // address with silently-wrapping two's complement arithmetic".
+        // Therefore, the final code will be a semantically equivalent.
+        //
+        // TODO(jingyue): do some range analysis to keep as many inbounds as
+        // possible. GEPs with inbounds are more friendly to alias analysis.
+        GEP->setIsInBounds(false);
+        Changed = true;
+      }
+    }
+  }
+
+  // Offsets the base with the accumulative byte offset.
+  //
+  //   %gep                        ; the base
+  //   ... %gep ...
+  //
+  // => add the offset
+  //
+  //   %gep2                       ; clone of %gep
+  //   %0       = ptrtoint %gep2
+  //   %1       = add %0, <offset>
+  //   %new.gep = inttoptr %1
+  //   %gep                        ; will be removed
+  //   ... %gep ...
+  //
+  // => replace all uses of %gep with %new.gep and remove %gep
+  //
+  //   %gep2                       ; clone of %gep
+  //   %0       = ptrtoint %gep2
+  //   %1       = add %0, <offset>
+  //   %new.gep = inttoptr %1
+  //   ... %new.gep ...
+  //
+  // TODO(jingyue): Emit a GEP instead of an "uglygep"
+  // (http://llvm.org/docs/GetElementPtr.html#what-s-an-uglygep) to make the IR
+  // prettier and more alias analysis friendly. One caveat: if the original GEP
+  // ends with a StructType, we need to split the GEP at the last
+  // SequentialType. For instance, consider the following IR:
+  //
+  //   %struct.S = type { float, double }
+  //   @array = global [1024 x %struct.S]
+  //   %p = getelementptr %array, 0, %i + 5, 1
+  //
+  // To separate the constant 5 from %p, we would need to split %p at the last
+  // array index so that we have:
+  //
+  //   %addr = gep %array, 0, %i
+  //   %p = gep %addr, 5, 1
+  Instruction *NewGEP = GEP->clone();
+  NewGEP->insertBefore(GEP);
+  Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
+  Value *Addr = new PtrToIntInst(NewGEP, IntPtrTy, "", GEP);
+  Addr = BinaryOperator::CreateAdd(
+      Addr, ConstantInt::get(IntPtrTy, AccumulativeByteOffset, true), "", GEP);
+  Addr = new IntToPtrInst(Addr, GEP->getType(), "", GEP);
+
+  GEP->replaceAllUsesWith(Addr);
+  GEP->eraseFromParent();
+
+  return true;
+}
+
+bool SeparateConstOffsetFromGEP::runOnFunction(Function &F) {
+  if (DisableSeparateConstOffsetFromGEP)
+    return false;
+
+  bool Changed = false;
+  for (Function::iterator B = F.begin(), BE = F.end(); B != BE; ++B) {
+    for (BasicBlock::iterator I = B->begin(), IE = B->end(); I != IE; ) {
+      if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I++)) {
+        Changed |= splitGEP(GEP);
+      }
+      // No need to split GEP ConstantExprs because all its indices are constant
+      // already.
+    }
+  }
+  return Changed;
+}
diff --git a/test/Transforms/SeparateConstOffsetFromGEP/NVPTX/lit.local.cfg b/test/Transforms/SeparateConstOffsetFromGEP/NVPTX/lit.local.cfg
new file mode 100644
index 0000000000..40532cdaa2
--- /dev/null
+++ b/test/Transforms/SeparateConstOffsetFromGEP/NVPTX/lit.local.cfg
@@ -0,0 +1,4 @@
+targets = set(config.root.targets_to_build.split())
+if not 'NVPTX' in targets:
+    config.unsupported = True
+
diff --git a/test/Transforms/SeparateConstOffsetFromGEP/NVPTX/split-gep-and-gvn.ll b/test/Transforms/SeparateConstOffsetFromGEP/NVPTX/split-gep-and-gvn.ll
new file mode 100644
index 0000000000..66f4096fa9
--- /dev/null
+++ b/test/Transforms/SeparateConstOffsetFromGEP/NVPTX/split-gep-and-gvn.ll
@@ -0,0 +1,60 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s --check-prefix=PTX
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 | FileCheck %s --check-prefix=PTX
+; RUN: opt < %s -S -separate-const-offset-from-gep -gvn -dce | FileCheck %s --check-prefix=IR
+
+; Verifies the SeparateConstOffsetFromGEP pass.
+; The following code computes
+; *output = array[x][y] + array[x][y+1] + array[x+1][y] + array[x+1][y+1]
+;
+; We expect SeparateConstOffsetFromGEP to transform it to
+;
+; float *base = &a[x][y];
+; *output = base[0] + base[1] + base[32] + base[33];
+;
+; so the backend can emit PTX that uses fewer virtual registers.
+
+target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64-unknown-unknown"
+
+@array = internal addrspace(3) constant [32 x [32 x float]] zeroinitializer, align 4
+
+define void @sum_of_array(i32 %x, i32 %y, float* nocapture %output) {
+.preheader:
+  %0 = zext i32 %y to i64
+  %1 = zext i32 %x to i64
+  %2 = getelementptr inbounds [32 x [32 x float]] addrspace(3)* @array, i64 0, i64 %1, i64 %0
+  %3 = addrspacecast float addrspace(3)* %2 to float*
+  %4 = load float* %3, align 4
+  %5 = fadd float %4, 0.000000e+00
+  %6 = add i32 %y, 1
+  %7 = zext i32 %6 to i64
+  %8 = getelementptr inbounds [32 x [32 x float]] addrspace(3)* @array, i64 0, i64 %1, i64 %7
+  %9 = addrspacecast float addrspace(3)* %8 to float*
+  %10 = load float* %9, align 4
+  %11 = fadd float %5, %10
+  %12 = add i32 %x, 1
+  %13 = zext i32 %12 to i64
+  %14 = getelementptr inbounds [32 x [32 x float]] addrspace(3)* @array, i64 0, i64 %13, i64 %0
+  %15 = addrspacecast float addrspace(3)* %14 to float*
+  %16 = load float* %15, align 4
+  %17 = fadd float %11, %16
+  %18 = getelementptr inbounds [32 x [32 x float]] addrspace(3)* @array, i64 0, i64 %13, i64 %7
+  %19 = addrspacecast float addrspace(3)* %18 to float*
+  %20 = load float* %19, align 4
+  %21 = fadd float %17, %20
+  store float %21, float* %output, align 4
+  ret void
+}
+
+; PTX-LABEL: sum_of_array(
+; PTX: ld.shared.f32 {{%f[0-9]+}}, {{\[}}[[BASE_REG:%(rl|r)[0-9]+]]{{\]}}
+; PTX: ld.shared.f32 {{%f[0-9]+}}, {{\[}}[[BASE_REG]]+4{{\]}}
+; PTX: ld.shared.f32 {{%f[0-9]+}}, {{\[}}[[BASE_REG]]+128{{\]}}
+; PTX: ld.shared.f32 {{%f[0-9]+}}, {{\[}}[[BASE_REG]]+132{{\]}}
+
+; IR-LABEL: @sum_of_array(
+; IR: [[BASE_PTR:%[0-9]+]] = getelementptr inbounds [32 x [32 x float]] addrspace(3)* @array, i64 0, i32 %x, i32 %y
+; IR: [[BASE_INT:%[0-9]+]] = ptrtoint float addrspace(3)* [[BASE_PTR]] to i64
+; IR: %5 = add i64 [[BASE_INT]], 4
+; IR: %10 = add i64 [[BASE_INT]], 128
+; IR: %15 = add i64 [[BASE_INT]], 132
diff --git a/test/Transforms/SeparateConstOffsetFromGEP/NVPTX/split-gep.ll b/test/Transforms/SeparateConstOffsetFromGEP/NVPTX/split-gep.ll
new file mode 100644
index 0000000000..f4020019c9
--- /dev/null
+++ b/test/Transforms/SeparateConstOffsetFromGEP/NVPTX/split-gep.ll
@@ -0,0 +1,101 @@
+; RUN: opt < %s -separate-const-offset-from-gep -dce -S | FileCheck %s
+
+; Several unit tests for -separate-const-offset-from-gep. The transformation
+; heavily relies on TargetTransformInfo, so we put these tests under
+; target-specific folders.
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+; target triple is necessary; otherwise TargetTransformInfo rejects any
+; addressing mode.
+target triple = "nvptx64-unknown-unknown"
+
+%struct.S = type { float, double }
+
+@struct_array = global [1024 x %struct.S] zeroinitializer, align 16
+@float_2d_array = global [32 x [32 x float]] zeroinitializer, align 4
+
+; We should not extract any struct field indices, because fields in a struct
+; may have different types.
+define double* @struct(i32 %i) {
+entry:
+  %add = add nsw i32 %i, 5
+  %idxprom = sext i32 %add to i64
+  %p = getelementptr inbounds [1024 x %struct.S]* @struct_array, i64 0, i64 %idxprom, i32 1
+  ret double* %p
+}
+; CHECK-LABEL: @struct
+; CHECK: getelementptr [1024 x %struct.S]* @struct_array, i64 0, i32 %i, i32 1
+
+; We should be able to trace into sext/zext if it's directly used as a GEP
+; index.
+define float* @sext_zext(i32 %i, i32 %j) {
+entry:
+  %i1 = add i32 %i, 1
+  %j2 = add i32 %j, 2
+  %i1.ext = sext i32 %i1 to i64
+  %j2.ext = zext i32 %j2 to i64
+  %p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i1.ext, i64 %j2.ext
+  ret float* %p
+}
+; CHECK-LABEL: @sext_zext
+; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i32 %i, i32 %j
+; CHECK: add i64 %{{[0-9]+}}, 136
+
+; We should be able to trace into sext/zext if it can be distributed to both
+; operands, e.g., sext (add nsw a, b) == add nsw (sext a), (sext b)
+define float* @ext_add_no_overflow(i64 %a, i32 %b, i64 %c, i32 %d) {
+  %b1 = add nsw i32 %b, 1
+  %b2 = sext i32 %b1 to i64
+  %i = add i64 %a, %b2
+  %d1 = add nuw i32 %d, 1
+  %d2 = zext i32 %d1 to i64
+  %j = add i64 %c, %d2
+  %p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i, i64 %j
+  ret float* %p
+}
+; CHECK-LABEL: @ext_add_no_overflow
+; CHECK: [[BASE_PTR:%[0-9]+]] = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %{{[0-9]+}}, i64 %{{[0-9]+}}
+; CHECK: [[BASE_INT:%[0-9]+]] = ptrtoint float* [[BASE_PTR]] to i64
+; CHECK: add i64 [[BASE_INT]], 132
+
+; We should treat "or" with no common bits (%k) as "add", and leave "or" with
+; potentially common bits (%l) as is.
+define float* @or(i64 %i) {
+entry:
+  %j = shl i64 %i, 2
+  %k = or i64 %j, 3 ; no common bits
+  %l = or i64 %j, 4 ; potentially common bits
+  %p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %k, i64 %l
+  ret float* %p
+}
+; CHECK-LABEL: @or
+; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %j, i64 %l
+; CHECK: add i64 %{{[0-9]+}}, 384
+
+; The subexpression (b + 5) is used in both "i = a + (b + 5)" and "*out = b +
+; 5". When extracting the constant offset 5, make sure "*out = b + 5" isn't
+; affected.
+define float* @expr(i64 %a, i64 %b, i64* %out) {
+entry:
+  %b5 = add i64 %b, 5
+  %i = add i64 %b5, %a
+  %p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i, i64 0
+  store i64 %b5, i64* %out
+  ret float* %p
+}
+; CHECK-LABEL: @expr
+; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %0, i64 0
+; CHECK: add i64 %{{[0-9]+}}, 640
+; CHECK: store i64 %b5, i64* %out
+
+; Verifies we handle "sub" correctly.
+define float* @sub(i64 %i, i64 %j) {
+  %i2 = sub i64 %i, 5 ; i - 5
+  %j2 = sub i64 5, %j ; 5 - i
+  %p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i2, i64 %j2
+  ret float* %p
+}
+; CHECK-LABEL: @sub
+; CHECK: %[[j2:[0-9]+]] = sub i64 0, %j
+; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i, i64 %[[j2]]
+; CHECK: add i64 %{{[0-9]+}}, -620