llvm-project/llvm/lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp

//===- SeparateConstOffsetFromGEP.cpp -------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// 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
// load 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
//
// Another improvement enabled by the LowerGEP flag is to lower a GEP with
// multiple indices to multiple GEPs with a single index.
// Such transformation can have following benefits:
// (1) It can always extract constants in the indices of structure type.
// (2) After such Lowering, there are more optimization opportunities such as
//     CSE, LICM and CGP.
//
// E.g. The following GEPs have multiple indices:
//  BB1:
//    %p = getelementptr [10 x %struct], ptr %ptr, i64 %i, i64 %j1, i32 3
//    load %p
//    ...
//  BB2:
//    %p2 = getelementptr [10 x %struct], ptr %ptr, i64 %i, i64 %j1, i32 2
//    load %p2
//    ...
//
// We can not do CSE to the common part related to index "i64 %i". Lowering
// GEPs can achieve such goals.
//
// This pass will lower a GEP with multiple indices into multiple GEPs with a
// single index:
//  BB1:
//    %2 = mul i64 %i, length_of_10xstruct          ; CSE opportunity
//    %3 = getelementptr i8, ptr %ptr, i64 %2       ; CSE opportunity
//    %4 = mul i64 %j1, length_of_struct
//    %5 = getelementptr i8, ptr %3, i64 %4
//    %p = getelementptr i8, ptr %5, struct_field_3 ; Constant offset
//    load %p
//    ...
//  BB2:
//    %8 = mul i64 %i, length_of_10xstruct            ; CSE opportunity
//    %9 = getelementptr i8, ptr %ptr, i64 %8         ; CSE opportunity
//    %10 = mul i64 %j2, length_of_struct
//    %11 = getelementptr i8, ptr %9, i64 %10
//    %p2 = getelementptr i8, ptr %11, struct_field_2 ; Constant offset
//    load %p2
//    ...
//
// Lowering GEPs can also benefit other passes such as LICM and CGP.
// LICM (Loop Invariant Code Motion) can not hoist/sink a GEP of multiple
// indices if one of the index is variant. If we lower such GEP into invariant
// parts and variant parts, LICM can hoist/sink those invariant parts.
// CGP (CodeGen Prepare) tries to sink address calculations that match the
// target's addressing modes. A GEP with multiple indices may not match and will
// not be sunk. If we lower such GEP into smaller parts, CGP may sink some of
// them. So we end up with a better addressing mode.
//
//===----------------------------------------------------------------------===//

#include "llvm/Transforms/Scalar/SeparateConstOffsetFromGEP.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/MemoryBuiltins.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/Constant.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/GetElementPtrTypeIterator.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstrTypes.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/PassManager.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/User.h"
#include "llvm/IR/Value.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Utils/Local.h"
#include <cassert>
#include <cstdint>
#include <string>

using namespace llvm;
using namespace llvm::PatternMatch;

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);

// Setting this flag may emit false positives when the input module already
// contains dead instructions. Therefore, we set it only in unit tests that are
// free of dead code.
static cl::opt<bool>
    VerifyNoDeadCode("reassociate-geps-verify-no-dead-code", cl::init(false),
                     cl::desc("Verify this pass produces no dead code"),
                     cl::Hidden);

namespace {

/// 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 separate 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 returns the
  /// new index representing the remainder (equal to the original index minus
  /// the constant offset), or nullptr if we cannot extract a constant offset.
  /// \p Idx The given GEP index
  /// \p GEP The given GEP
  /// \p UserChainTail Outputs the tail of UserChain so that we can
  ///                  garbage-collect unused instructions in UserChain.
  /// \p PreservesNUW  Outputs whether the extraction allows preserving the
  ///                  GEP's nuw flag, if it has one.
  static Value *Extract(Value *Idx, GetElementPtrInst *GEP,
                        User *&UserChainTail, bool &PreservesNUW);

  /// Looks for a constant offset from the given GEP index without extracting
  /// it. It returns the numeric value of the extracted constant offset (0 if
  /// failed). The meaning of the arguments are the same as Extract.
  static int64_t Find(Value *Idx, GetElementPtrInst *GEP);

private:
  ConstantOffsetExtractor(BasicBlock::iterator InsertionPt)
      : IP(InsertionPt), DL(InsertionPt->getDataLayout()) {}

  /// Searches the expression that computes V for a non-zero constant C s.t.
  /// V can be reassociated into the form V' + C. If the searching is
  /// successful, returns C and update UserChain as a def-use chain from C to V;
  /// otherwise, UserChain is empty.
  ///
  /// \p V            The given expression
  /// \p SignExtended Whether V will be sign-extended in the computation of the
  ///                 GEP index
  /// \p ZeroExtended Whether V will be zero-extended in the computation of the
  ///                 GEP index
  /// \p NonNegative  Whether V is guaranteed to be non-negative. For example,
  ///                 an index of an inbounds GEP is guaranteed to be
  ///                 non-negative. Levaraging this, we can better split
  ///                 inbounds GEPs.
  APInt find(Value *V, bool SignExtended, bool ZeroExtended, bool NonNegative);

  /// A helper function to look into both operands of a binary operator.
  APInt findInEitherOperand(BinaryOperator *BO, bool SignExtended,
                            bool ZeroExtended);

  /// After finding the constant offset C from the GEP index I, we build a new
  /// index I' s.t. I' + C = I. This function builds and returns the new
  /// index I' according to UserChain produced by function "find".
  ///
  /// The building conceptually takes two steps:
  /// 1) iteratively distribute s/zext towards the leaves of the expression tree
  /// that computes I
  /// 2) reassociate the expression tree to the form I' + C.
  ///
  /// For example, to extract the 5 from sext(a + (b + 5)), we first distribute
  /// sext to a, b and 5 so that we have
  ///   sext(a) + (sext(b) + 5).
  /// Then, we reassociate it to
  ///   (sext(a) + sext(b)) + 5.
  /// Given this form, we know I' is sext(a) + sext(b).
  Value *rebuildWithoutConstOffset();

  /// After the first step of rebuilding the GEP index without the constant
  /// offset, distribute s/zext to the operands of all operators in UserChain.
  /// e.g., zext(sext(a + (b + 5)) (assuming no overflow) =>
  /// zext(sext(a)) + (zext(sext(b)) + zext(sext(5))).
  ///
  /// The function also updates UserChain to point to new subexpressions after
  /// distributing s/zext. e.g., the old UserChain of the above example is
  /// 5 -> b + 5 -> a + (b + 5) -> sext(...) -> zext(sext(...)),
  /// and the new UserChain is
  /// zext(sext(5)) -> zext(sext(b)) + zext(sext(5)) ->
  ///   zext(sext(a)) + (zext(sext(b)) + zext(sext(5))
  ///
  /// \p ChainIndex The index to UserChain. ChainIndex is initially
  ///               UserChain.size() - 1, and is decremented during
  ///               the recursion.
  Value *distributeExtsAndCloneChain(unsigned ChainIndex);

  /// Reassociates the GEP index to the form I' + C and returns I'.
  Value *removeConstOffset(unsigned ChainIndex);

  /// A helper function to apply ExtInsts, a list of s/zext, to value V.
  /// e.g., if ExtInsts = [sext i32 to i64, zext i16 to i32], this function
  /// returns "sext i32 (zext i16 V to i32) to i64".
  Value *applyExts(Value *V);

  /// A helper function that returns whether we can trace into the operands
  /// of binary operator BO for a constant offset.
  ///
  /// \p SignExtended Whether BO is surrounded by sext
  /// \p ZeroExtended Whether BO is surrounded by zext
  /// \p NonNegative Whether BO is known to be non-negative, e.g., an in-bound
  ///                array index.
  bool CanTraceInto(bool SignExtended, bool ZeroExtended, BinaryOperator *BO,
                    bool NonNegative);

  /// Analyze XOR instruction to extract disjoint constant bits that behave
  /// like addition operations for improved address mode folding.
  APInt extractDisjointBitsFromXor(BinaryOperator *XorInst);

  /// 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 to rebuild the new GEP index.
  SmallVector<User *, 8> UserChain;

  /// A data structure used in rebuildWithoutConstOffset. Contains all
  /// sext/zext instructions along UserChain.
  SmallVector<CastInst *, 16> ExtInsts;

  /// Insertion position of cloned instructions.
  BasicBlock::iterator IP;

  const DataLayout &DL;
};

/// A pass that tries to split every GEP in the function into a variadic
/// base and a constant offset. It is a FunctionPass because searching for the
/// constant offset may inspect other basic blocks.
class SeparateConstOffsetFromGEPLegacyPass : public FunctionPass {
public:
  static char ID;

  SeparateConstOffsetFromGEPLegacyPass(bool LowerGEP = false)
      : FunctionPass(ID), LowerGEP(LowerGEP) {
    initializeSeparateConstOffsetFromGEPLegacyPassPass(
        *PassRegistry::getPassRegistry());
  }

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addRequired<DominatorTreeWrapperPass>();
    AU.addRequired<TargetTransformInfoWrapperPass>();
    AU.addRequired<LoopInfoWrapperPass>();
    AU.setPreservesCFG();
    AU.addRequired<TargetLibraryInfoWrapperPass>();
  }

  bool runOnFunction(Function &F) override;

private:
  bool LowerGEP;
};

/// A pass that tries to split every GEP in the function into a variadic
/// base and a constant offset. It is a FunctionPass because searching for the
/// constant offset may inspect other basic blocks.
class SeparateConstOffsetFromGEP {
public:
  SeparateConstOffsetFromGEP(
      DominatorTree *DT, LoopInfo *LI, TargetLibraryInfo *TLI,
      function_ref<TargetTransformInfo &(Function &)> GetTTI, bool LowerGEP)
      : DT(DT), LI(LI), TLI(TLI), GetTTI(GetTTI), LowerGEP(LowerGEP) {}

  bool run(Function &F);

private:
  /// Track the operands of an add or sub.
  using ExprKey = std::pair<Value *, Value *>;

  /// Create a pair for use as a map key for a commutable operation.
  static ExprKey createNormalizedCommutablePair(Value *A, Value *B) {
    if (A < B)
      return {A, B};
    return {B, A};
  }

  /// 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);

  /// Tries to reorder the given GEP with the GEP that produces the base if
  /// doing so results in producing a constant offset as the outermost
  /// index.
  bool reorderGEP(GetElementPtrInst *GEP, TargetTransformInfo &TTI);

  /// Lower a GEP with multiple indices into multiple GEPs with a single index.
  /// Function splitGEP already split the original GEP into a variadic part and
  /// a constant offset (i.e., AccumulativeByteOffset). This function lowers the
  /// variadic part into a set of GEPs with a single index and applies
  /// AccumulativeByteOffset to it.
  /// \p Variadic                  The variadic part of the original GEP.
  /// \p AccumulativeByteOffset    The constant offset.
  void lowerToSingleIndexGEPs(GetElementPtrInst *Variadic,
                              int64_t AccumulativeByteOffset);

  /// Finds the constant offset within each index and accumulates them. If
  /// LowerGEP is true, it finds in indices of both sequential and structure
  /// types, otherwise it only finds in sequential indices. The output
  /// NeedsExtraction indicates whether we successfully find a non-zero constant
  /// offset.
  int64_t accumulateByteOffset(GetElementPtrInst *GEP, bool &NeedsExtraction);

  /// Canonicalize array indices to pointer-size integers. This helps to
  /// simplify the logic of splitting a GEP. For example, if a + b is a
  /// pointer-size integer, we have
  ///   gep base, a + b = gep (gep base, a), b
  /// However, this equality may not hold if the size of a + b is smaller than
  /// the pointer size, because LLVM conceptually sign-extends GEP indices to
  /// pointer size before computing the address
  /// (http://llvm.org/docs/LangRef.html#id181).
  ///
  /// This canonicalization is very likely already done in clang and
  /// instcombine. Therefore, the program will probably remain the same.
  ///
  /// Returns true if the module changes.
  ///
  /// Verified in @i32_add in split-gep.ll
  bool canonicalizeArrayIndicesToIndexSize(GetElementPtrInst *GEP);

  /// Optimize sext(a)+sext(b) to sext(a+b) when a+b can't sign overflow.
  /// SeparateConstOffsetFromGEP distributes a sext to leaves before extracting
  /// the constant offset. After extraction, it becomes desirable to reunion the
  /// distributed sexts. For example,
  ///
  ///                              &a[sext(i +nsw (j +nsw 5)]
  ///   => distribute              &a[sext(i) +nsw (sext(j) +nsw 5)]
  ///   => constant extraction     &a[sext(i) + sext(j)] + 5
  ///   => reunion                 &a[sext(i +nsw j)] + 5
  bool reuniteExts(Function &F);

  /// A helper that reunites sexts in an instruction.
  bool reuniteExts(Instruction *I);

  /// Find the closest dominator of <Dominatee> that is equivalent to <Key>.
  Instruction *findClosestMatchingDominator(
      ExprKey Key, Instruction *Dominatee,
      DenseMap<ExprKey, SmallVector<Instruction *, 2>> &DominatingExprs);

  /// Verify F is free of dead code.
  void verifyNoDeadCode(Function &F);

  bool hasMoreThanOneUseInLoop(Value *v, Loop *L);

  // Swap the index operand of two GEP.
  void swapGEPOperand(GetElementPtrInst *First, GetElementPtrInst *Second);

  // Check if it is safe to swap operand of two GEP.
  bool isLegalToSwapOperand(GetElementPtrInst *First, GetElementPtrInst *Second,
                            Loop *CurLoop);

  const DataLayout *DL = nullptr;
  DominatorTree *DT = nullptr;
  LoopInfo *LI;
  TargetLibraryInfo *TLI;
  // Retrieved lazily since not always used.
  function_ref<TargetTransformInfo &(Function &)> GetTTI;

  /// Whether to lower a GEP with multiple indices into arithmetic operations or
  /// multiple GEPs with a single index.
  bool LowerGEP;

  DenseMap<ExprKey, SmallVector<Instruction *, 2>> DominatingAdds;
  DenseMap<ExprKey, SmallVector<Instruction *, 2>> DominatingSubs;
};

} // end anonymous namespace

char SeparateConstOffsetFromGEPLegacyPass::ID = 0;

INITIALIZE_PASS_BEGIN(
    SeparateConstOffsetFromGEPLegacyPass, "separate-const-offset-from-gep",
    "Split GEPs to a variadic base and a constant offset for better CSE", false,
    false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
INITIALIZE_PASS_END(
    SeparateConstOffsetFromGEPLegacyPass, "separate-const-offset-from-gep",
    "Split GEPs to a variadic base and a constant offset for better CSE", false,
    false)

FunctionPass *llvm::createSeparateConstOffsetFromGEPPass(bool LowerGEP) {
  return new SeparateConstOffsetFromGEPLegacyPass(LowerGEP);
}

bool ConstantOffsetExtractor::CanTraceInto(bool SignExtended,
                                            bool ZeroExtended,
                                            BinaryOperator *BO,
                                            bool NonNegative) {
  // We only consider ADD, SUB and OR, because a non-zero constant found in
  // expressions composed of these operations can be easily hoisted as a
  // constant offset by reassociation.
  if (BO->getOpcode() != Instruction::Add &&
      BO->getOpcode() != Instruction::Sub &&
      BO->getOpcode() != Instruction::Or) {
    return false;
  }

  Value *LHS = BO->getOperand(0), *RHS = BO->getOperand(1);
  // Do not trace into "or" unless it is equivalent to "add".
  // This is the case if the or's disjoint flag is set.
  if (BO->getOpcode() == Instruction::Or &&
      !cast<PossiblyDisjointInst>(BO)->isDisjoint())
    return false;

  // FIXME: We don't currently support constants from the RHS of subs,
  // when we are zero-extended, because we need a way to zero-extended
  // them before they are negated.
  if (ZeroExtended && !SignExtended && BO->getOpcode() == Instruction::Sub)
    return false;

  // In addition, tracing into BO requires that its surrounding s/zext (if
  // any) is distributable to both operands.
  //
  // Suppose BO = A op B.
  //  SignExtended | ZeroExtended | Distributable?
  // --------------+--------------+----------------------------------
  //       0       |      0       | true because no s/zext exists
  //       0       |      1       | zext(BO) == zext(A) op zext(B)
  //       1       |      0       | sext(BO) == sext(A) op sext(B)
  //       1       |      1       | zext(sext(BO)) ==
  //               |              |     zext(sext(A)) op zext(sext(B))
  if (BO->getOpcode() == Instruction::Add && !ZeroExtended && NonNegative) {
    // If a + b >= 0 and (a >= 0 or b >= 0), then
    //   sext(a + b) = sext(a) + sext(b)
    // even if the addition is not marked nsw.
    //
    // Leveraging this invariant, we can trace into an sext'ed inbound GEP
    // index if the constant offset is non-negative.
    //
    // Verified in @sext_add in split-gep.ll.
    if (ConstantInt *ConstLHS = dyn_cast<ConstantInt>(LHS)) {
      if (!ConstLHS->isNegative())
        return true;
    }
    if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(RHS)) {
      if (!ConstRHS->isNegative())
        return true;
    }
  }

  // sext (add/sub nsw A, B) == add/sub nsw (sext A), (sext B)
  // zext (add/sub nuw A, B) == add/sub nuw (zext A), (zext B)
  if (BO->getOpcode() == Instruction::Add ||
      BO->getOpcode() == Instruction::Sub) {
    if (SignExtended && !BO->hasNoSignedWrap())
      return false;
    if (ZeroExtended && !BO->hasNoUnsignedWrap())
      return false;
  }

  return true;
}

APInt ConstantOffsetExtractor::findInEitherOperand(BinaryOperator *BO,
                                                   bool SignExtended,
                                                   bool ZeroExtended) {
  // Save off the current height of the chain, in case we need to restore it.
  size_t ChainLength = UserChain.size();

  // BO being non-negative does not shed light on whether its operands are
  // non-negative. Clear the NonNegative flag here.
  APInt ConstantOffset = find(BO->getOperand(0), SignExtended, ZeroExtended,
                              /* NonNegative */ false);
  // 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;

  // Reset the chain back to where it was when we started exploring this node,
  // since visiting the LHS didn't pan out.
  UserChain.resize(ChainLength);

  ConstantOffset = find(BO->getOperand(1), SignExtended, ZeroExtended,
                        /* NonNegative */ false);
  // If U is a sub operator, negate the constant offset found in the right
  // operand.
  if (BO->getOpcode() == Instruction::Sub)
    ConstantOffset = -ConstantOffset;

  // If RHS wasn't a suitable candidate either, reset the chain again.
  if (ConstantOffset == 0)
    UserChain.resize(ChainLength);

  return ConstantOffset;
}

APInt ConstantOffsetExtractor::find(Value *V, bool SignExtended,
                                    bool ZeroExtended, bool NonNegative) {
  // TODO(jingyue): We could 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.
  unsigned BitWidth = cast<IntegerType>(V->getType())->getBitWidth();

  // We cannot do much with Values that are not a User, such as an Argument.
  User *U = dyn_cast<User>(V);
  if (U == nullptr) return APInt(BitWidth, 0);

  APInt ConstantOffset(BitWidth, 0);
  if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
    // Hooray, we found it!
    ConstantOffset = CI->getValue();
  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V)) {
    // Trace into subexpressions for more hoisting opportunities.
    if (CanTraceInto(SignExtended, ZeroExtended, BO, NonNegative))
      ConstantOffset = findInEitherOperand(BO, SignExtended, ZeroExtended);
    // Handle XOR with disjoint bits that can be treated as addition.
    else if (BO->getOpcode() == Instruction::Xor)
      ConstantOffset = extractDisjointBitsFromXor(BO);
  } else if (isa<TruncInst>(V)) {
    ConstantOffset =
        find(U->getOperand(0), SignExtended, ZeroExtended, NonNegative)
            .trunc(BitWidth);
  } else if (isa<SExtInst>(V)) {
    ConstantOffset = find(U->getOperand(0), /* SignExtended */ true,
                          ZeroExtended, NonNegative).sext(BitWidth);
  } else if (isa<ZExtInst>(V)) {
    // As an optimization, we can clear the SignExtended flag because
    // sext(zext(a)) = zext(a). Verified in @sext_zext in split-gep.ll.
    //
    // Clear the NonNegative flag, because zext(a) >= 0 does not imply a >= 0.
    ConstantOffset =
        find(U->getOperand(0), /* SignExtended */ false,
             /* ZeroExtended */ true, /* NonNegative */ false).zext(BitWidth);
  }

  // If we found a non-zero constant offset, add it to the path for
  // rebuildWithoutConstOffset. Zero is a valid constant offset, but doesn't
  // help this optimization.
  if (ConstantOffset != 0)
    UserChain.push_back(U);
  return ConstantOffset;
}

Value *ConstantOffsetExtractor::applyExts(Value *V) {
  Value *Current = V;
  // ExtInsts is built in the use-def order. Therefore, we apply them to V
  // in the reversed order.
  for (CastInst *I : llvm::reverse(ExtInsts)) {
    if (Constant *C = dyn_cast<Constant>(Current)) {
      // Try to constant fold the cast.
      Current = ConstantFoldCastOperand(I->getOpcode(), C, I->getType(), DL);
      if (Current)
        continue;
    }

    Instruction *Ext = I->clone();
    Ext->setOperand(0, Current);
    // In ConstantOffsetExtractor::find we do not analyze nuw/nsw for trunc, so
    // we assume that it is ok to redistribute trunc over add/sub/or. But for
    // example (add (trunc nuw A), (trunc nuw B)) is more poisonous than (trunc
    // nuw (add A, B))). To make such redistributions legal we drop all the
    // poison generating flags from cloned trunc instructions here.
    if (isa<TruncInst>(Ext))
      Ext->dropPoisonGeneratingFlags();
    Ext->insertBefore(*IP->getParent(), IP);
    Current = Ext;
  }
  return Current;
}

Value *ConstantOffsetExtractor::rebuildWithoutConstOffset() {
  distributeExtsAndCloneChain(UserChain.size() - 1);
  // Remove all nullptrs (used to be s/zext) from UserChain.
  unsigned NewSize = 0;
  for (User *I : UserChain) {
    if (I != nullptr) {
      UserChain[NewSize] = I;
      NewSize++;
    }
  }
  UserChain.resize(NewSize);
  return removeConstOffset(UserChain.size() - 1);
}

Value *
ConstantOffsetExtractor::distributeExtsAndCloneChain(unsigned ChainIndex) {
  User *U = UserChain[ChainIndex];
  if (ChainIndex == 0) {
    assert(isa<ConstantInt>(U));
    // If U is a ConstantInt, applyExts will return a ConstantInt as well.
    return UserChain[ChainIndex] = cast<ConstantInt>(applyExts(U));
  }

  if (CastInst *Cast = dyn_cast<CastInst>(U)) {
    assert(
        (isa<SExtInst>(Cast) || isa<ZExtInst>(Cast) || isa<TruncInst>(Cast)) &&
        "Only following instructions can be traced: sext, zext & trunc");
    ExtInsts.push_back(Cast);
    UserChain[ChainIndex] = nullptr;
    return distributeExtsAndCloneChain(ChainIndex - 1);
  }

  // Function find only trace into BinaryOperator and CastInst.
  BinaryOperator *BO = cast<BinaryOperator>(U);
  // OpNo = which operand of BO is UserChain[ChainIndex - 1]
  unsigned OpNo = (BO->getOperand(0) == UserChain[ChainIndex - 1] ? 0 : 1);
  Value *TheOther = applyExts(BO->getOperand(1 - OpNo));
  Value *NextInChain = distributeExtsAndCloneChain(ChainIndex - 1);

  BinaryOperator *NewBO = nullptr;
  if (OpNo == 0) {
    NewBO = BinaryOperator::Create(BO->getOpcode(), NextInChain, TheOther,
                                   BO->getName(), IP);
  } else {
    NewBO = BinaryOperator::Create(BO->getOpcode(), TheOther, NextInChain,
                                   BO->getName(), IP);
  }
  return UserChain[ChainIndex] = NewBO;
}

Value *ConstantOffsetExtractor::removeConstOffset(unsigned ChainIndex) {
  if (ChainIndex == 0) {
    assert(isa<ConstantInt>(UserChain[ChainIndex]));
    return ConstantInt::getNullValue(UserChain[ChainIndex]->getType());
  }

  BinaryOperator *BO = cast<BinaryOperator>(UserChain[ChainIndex]);
  assert((BO->use_empty() || BO->hasOneUse()) &&
         "distributeExtsAndCloneChain clones each BinaryOperator in "
         "UserChain, so no one should be used more than "
         "once");

  unsigned OpNo = (BO->getOperand(0) == UserChain[ChainIndex - 1] ? 0 : 1);
  assert(BO->getOperand(OpNo) == UserChain[ChainIndex - 1]);
  Value *NextInChain = removeConstOffset(ChainIndex - 1);
  Value *TheOther = BO->getOperand(1 - OpNo);

  if (ConstantInt *CI = dyn_cast<ConstantInt>(NextInChain)) {
    if (CI->isZero()) {
      // Custom XOR handling for disjoint bits - preserves original XOR
      // with non-disjoint constant bits.
      // TODO: The design should be updated to support partial constant
      // extraction.
      if (BO->getOpcode() == Instruction::Xor)
        return BO;

      // If NextInChain is 0 and not the LHS of a sub, we can simplify the
      // sub-expression to be just TheOther.
      if (!(BO->getOpcode() == Instruction::Sub && OpNo == 0))
        return TheOther;
    }
  }

  BinaryOperator::BinaryOps NewOp = BO->getOpcode();
  if (BO->getOpcode() == Instruction::Or) {
    // Rebuild "or" as "add", because "or" may be invalid for the new
    // expression.
    //
    // For instance, given
    //   a | (b + 5) where a and b + 5 have no common bits,
    // we can extract 5 as the constant offset.
    //
    // However, reusing the "or" in the new index would give us
    //   (a | b) + 5
    // which does not equal a | (b + 5).
    //
    // Replacing the "or" with "add" is fine, because
    //   a | (b + 5) = a + (b + 5) = (a + b) + 5
    NewOp = Instruction::Add;
  }

  BinaryOperator *NewBO;
  if (OpNo == 0) {
    NewBO = BinaryOperator::Create(NewOp, NextInChain, TheOther, "", IP);
  } else {
    NewBO = BinaryOperator::Create(NewOp, TheOther, NextInChain, "", IP);
  }
  NewBO->takeName(BO);
  return NewBO;
}

/// Analyze XOR instruction to extract disjoint constant bits for address
/// folding
///
/// This function identifies bits in an XOR constant operand that are disjoint
/// from the base operand's known set bits. For these disjoint bits, XOR behaves
/// identically to addition, allowing us to extract them as constant offsets
/// that can be folded into addressing modes.
///
/// Transformation: `Base ^ Const` becomes `(Base ^ NonDisjointBits) +
/// DisjointBits` where DisjointBits = Const & KnownZeros(Base)
///
/// Example with ptr having known-zero low bit:
///   Original: `xor %ptr, 3`    ; 3 = 0b11
///   Analysis: DisjointBits = 3 & KnownZeros(%ptr) = 0b11 & 0b01 = 0b01
///   Result:   `(xor %ptr, 2) + 1` where 1 can be folded into address mode
///
/// \param XorInst The XOR binary operator to analyze
/// \return APInt containing the disjoint bits that can be extracted as offset,
///         or zero if no disjoint bits exist
APInt ConstantOffsetExtractor::extractDisjointBitsFromXor(
    BinaryOperator *XorInst) {
  assert(XorInst && XorInst->getOpcode() == Instruction::Xor &&
         "Expected XOR instruction");

  const unsigned BitWidth = XorInst->getType()->getScalarSizeInBits();
  Value *BaseOperand;
  ConstantInt *XorConstant;

  // Match pattern: xor BaseOperand, Constant.
  if (!match(XorInst, m_Xor(m_Value(BaseOperand), m_ConstantInt(XorConstant))))
    return APInt::getZero(BitWidth);

  // Compute known bits for the base operand.
  const SimplifyQuery SQ(DL);
  const KnownBits BaseKnownBits = computeKnownBits(BaseOperand, SQ);
  const APInt &ConstantValue = XorConstant->getValue();

  // Identify disjoint bits: constant bits that are known zero in base.
  const APInt DisjointBits = ConstantValue & BaseKnownBits.Zero;

  // Early exit if no disjoint bits found.
  if (DisjointBits.isZero())
    return APInt::getZero(BitWidth);

  // Compute the remaining non-disjoint bits that stay in the XOR.
  const APInt NonDisjointBits = ConstantValue & ~DisjointBits;

  // FIXME: Enhance XOR constant extraction to handle nested binary operations.
  // Currently we only extract disjoint bits from the immediate XOR constant,
  // but we could recursively process cases like:
  //   xor (add %base, C1), C2  ->  add %base, (C1 ^ disjoint_bits(C2))
  // This requires careful analysis to ensure the transformation preserves
  // semantics, particularly around sign extension and overflow behavior.

  // Add the non-disjoint constant to the user chain for later transformation
  // This will replace the original constant in the XOR with the new
  // constant.
  UserChain.push_back(ConstantInt::get(XorInst->getType(), NonDisjointBits));
  return DisjointBits;
}

/// A helper function to check if reassociating through an entry in the user
/// chain would invalidate the GEP's nuw flag.
static bool allowsPreservingNUW(const User *U) {
  if (const BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
    // Binary operations need to be effectively add nuw.
    auto Opcode = BO->getOpcode();
    if (Opcode == BinaryOperator::Or) {
      // Ors are only considered here if they are disjoint. The addition that
      // they represent in this case is NUW.
      assert(cast<PossiblyDisjointInst>(BO)->isDisjoint());
      return true;
    }
    return Opcode == BinaryOperator::Add && BO->hasNoUnsignedWrap();
  }
  // UserChain can only contain ConstantInt, CastInst, or BinaryOperator.
  // Among the possible CastInsts, only trunc without nuw is a problem: If it
  // is distributed through an add nuw, wrapping may occur:
  // "add nuw trunc(a), trunc(b)" is more poisonous than "trunc(add nuw a, b)"
  if (const TruncInst *TI = dyn_cast<TruncInst>(U))
    return TI->hasNoUnsignedWrap();
  return isa<CastInst>(U) || isa<ConstantInt>(U);
}

Value *ConstantOffsetExtractor::Extract(Value *Idx, GetElementPtrInst *GEP,
                                        User *&UserChainTail,
                                        bool &PreservesNUW) {
  ConstantOffsetExtractor Extractor(GEP->getIterator());
  // Find a non-zero constant offset first.
  APInt ConstantOffset =
      Extractor.find(Idx, /* SignExtended */ false, /* ZeroExtended */ false,
                     GEP->isInBounds());
  if (ConstantOffset == 0) {
    UserChainTail = nullptr;
    PreservesNUW = true;
    return nullptr;
  }

  PreservesNUW = all_of(Extractor.UserChain, allowsPreservingNUW);

  // Separates the constant offset from the GEP index.
  Value *IdxWithoutConstOffset = Extractor.rebuildWithoutConstOffset();
  UserChainTail = Extractor.UserChain.back();
  return IdxWithoutConstOffset;
}

int64_t ConstantOffsetExtractor::Find(Value *Idx, GetElementPtrInst *GEP) {
  // If Idx is an index of an inbound GEP, Idx is guaranteed to be non-negative.
  return ConstantOffsetExtractor(GEP->getIterator())
      .find(Idx, /* SignExtended */ false, /* ZeroExtended */ false,
            GEP->isInBounds())
      .getSExtValue();
}

bool SeparateConstOffsetFromGEP::canonicalizeArrayIndicesToIndexSize(
    GetElementPtrInst *GEP) {
  bool Changed = false;
  Type *PtrIdxTy = DL->getIndexType(GEP->getType());
  gep_type_iterator GTI = gep_type_begin(*GEP);
  for (User::op_iterator I = GEP->op_begin() + 1, E = GEP->op_end();
       I != E; ++I, ++GTI) {
    // Skip struct member indices which must be i32.
    if (GTI.isSequential()) {
      if ((*I)->getType() != PtrIdxTy) {
        *I = CastInst::CreateIntegerCast(*I, PtrIdxTy, true, "idxprom",
                                         GEP->getIterator());
        Changed = true;
      }
    }
  }
  return Changed;
}

int64_t
SeparateConstOffsetFromGEP::accumulateByteOffset(GetElementPtrInst *GEP,
                                                 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 (GTI.isSequential()) {
      // Constant offsets of scalable types are not really constant.
      if (GTI.getIndexedType()->isScalableTy())
        continue;

      // Tries to extract a constant offset from this GEP index.
      int64_t ConstantOffset =
          ConstantOffsetExtractor::Find(GEP->getOperand(I), GEP);
      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 * GTI.getSequentialElementStride(*DL);
      }
    } else if (LowerGEP) {
      StructType *StTy = GTI.getStructType();
      uint64_t Field = cast<ConstantInt>(GEP->getOperand(I))->getZExtValue();
      // Skip field 0 as the offset is always 0.
      if (Field != 0) {
        NeedsExtraction = true;
        AccumulativeByteOffset +=
            DL->getStructLayout(StTy)->getElementOffset(Field);
      }
    }
  }
  return AccumulativeByteOffset;
}

void SeparateConstOffsetFromGEP::lowerToSingleIndexGEPs(
    GetElementPtrInst *Variadic, int64_t AccumulativeByteOffset) {
  IRBuilder<> Builder(Variadic);
  Type *PtrIndexTy = DL->getIndexType(Variadic->getType());

  Value *ResultPtr = Variadic->getOperand(0);
  Loop *L = LI->getLoopFor(Variadic->getParent());
  // Check if the base is not loop invariant or used more than once.
  bool isSwapCandidate =
      L && L->isLoopInvariant(ResultPtr) &&
      !hasMoreThanOneUseInLoop(ResultPtr, L);
  Value *FirstResult = nullptr;

  gep_type_iterator GTI = gep_type_begin(*Variadic);
  // Create an ugly GEP for each sequential index. We don't create GEPs for
  // structure indices, as they are accumulated in the constant offset index.
  for (unsigned I = 1, E = Variadic->getNumOperands(); I != E; ++I, ++GTI) {
    if (GTI.isSequential()) {
      Value *Idx = Variadic->getOperand(I);
      // Skip zero indices.
      if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx))
        if (CI->isZero())
          continue;

      APInt ElementSize = APInt(PtrIndexTy->getIntegerBitWidth(),
                                GTI.getSequentialElementStride(*DL));
      // Scale the index by element size.
      if (ElementSize != 1) {
        if (ElementSize.isPowerOf2()) {
          Idx = Builder.CreateShl(
              Idx, ConstantInt::get(PtrIndexTy, ElementSize.logBase2()));
        } else {
          Idx =
              Builder.CreateMul(Idx, ConstantInt::get(PtrIndexTy, ElementSize));
        }
      }
      // Create an ugly GEP with a single index for each index.
      ResultPtr = Builder.CreatePtrAdd(ResultPtr, Idx, "uglygep");
      if (FirstResult == nullptr)
        FirstResult = ResultPtr;
    }
  }

  // Create a GEP with the constant offset index.
  if (AccumulativeByteOffset != 0) {
    Value *Offset = ConstantInt::get(PtrIndexTy, AccumulativeByteOffset);
    ResultPtr = Builder.CreatePtrAdd(ResultPtr, Offset, "uglygep");
  } else
    isSwapCandidate = false;

  // If we created a GEP with constant index, and the base is loop invariant,
  // then we swap the first one with it, so LICM can move constant GEP out
  // later.
  auto *FirstGEP = dyn_cast_or_null<GetElementPtrInst>(FirstResult);
  auto *SecondGEP = dyn_cast<GetElementPtrInst>(ResultPtr);
  if (isSwapCandidate && isLegalToSwapOperand(FirstGEP, SecondGEP, L))
    swapGEPOperand(FirstGEP, SecondGEP);

  Variadic->replaceAllUsesWith(ResultPtr);
  Variadic->eraseFromParent();
}

bool SeparateConstOffsetFromGEP::reorderGEP(GetElementPtrInst *GEP,
                                            TargetTransformInfo &TTI) {
  auto PtrGEP = dyn_cast<GetElementPtrInst>(GEP->getPointerOperand());
  if (!PtrGEP)
    return false;

  bool NestedNeedsExtraction;
  int64_t NestedByteOffset =
      accumulateByteOffset(PtrGEP, NestedNeedsExtraction);
  if (!NestedNeedsExtraction)
    return false;

  unsigned AddrSpace = PtrGEP->getPointerAddressSpace();
  if (!TTI.isLegalAddressingMode(GEP->getResultElementType(),
                                 /*BaseGV=*/nullptr, NestedByteOffset,
                                 /*HasBaseReg=*/true, /*Scale=*/0, AddrSpace))
    return false;

  bool GEPInBounds = GEP->isInBounds();
  bool PtrGEPInBounds = PtrGEP->isInBounds();
  bool IsChainInBounds = GEPInBounds && PtrGEPInBounds;
  if (IsChainInBounds) {
    auto IsKnownNonNegative = [this](Value *V) {
      return isKnownNonNegative(V, *DL);
    };
    IsChainInBounds &= all_of(GEP->indices(), IsKnownNonNegative);
    if (IsChainInBounds)
      IsChainInBounds &= all_of(PtrGEP->indices(), IsKnownNonNegative);
  }

  IRBuilder<> Builder(GEP);
  // For trivial GEP chains, we can swap the indices.
  Value *NewSrc = Builder.CreateGEP(
      GEP->getSourceElementType(), PtrGEP->getPointerOperand(),
      SmallVector<Value *, 4>(GEP->indices()), "", IsChainInBounds);
  Value *NewGEP = Builder.CreateGEP(PtrGEP->getSourceElementType(), NewSrc,
                                    SmallVector<Value *, 4>(PtrGEP->indices()),
                                    "", IsChainInBounds);
  GEP->replaceAllUsesWith(NewGEP);
  RecursivelyDeleteTriviallyDeadInstructions(GEP);
  return true;
}

bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
  // Skip vector GEPs.
  if (GEP->getType()->isVectorTy())
    return false;

  // If the base of this GEP is a ptradd of a constant, lets pass the constant
  // along. This ensures that when we have a chain of GEPs the constant
  // offset from each is accumulated.
  Value *NewBase;
  const APInt *BaseOffset;
  const bool ExtractBase =
      match(GEP->getPointerOperand(),
            m_PtrAdd(m_Value(NewBase), m_APInt(BaseOffset)));

  const int64_t BaseByteOffset = ExtractBase ? BaseOffset->getSExtValue() : 0;

  // The backend can already nicely handle the case where all indices are
  // constant.
  if (GEP->hasAllConstantIndices() && !ExtractBase)
    return false;

  bool Changed = canonicalizeArrayIndicesToIndexSize(GEP);

  bool NeedsExtraction;
  int64_t AccumulativeByteOffset =
      BaseByteOffset + accumulateByteOffset(GEP, NeedsExtraction);

  TargetTransformInfo &TTI = GetTTI(*GEP->getFunction());

  if (!NeedsExtraction && !ExtractBase) {
    Changed |= reorderGEP(GEP, TTI);
    return Changed;
  }

  // If LowerGEP is disabled, 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.
  // If LowerGEP is enabled, even the extracted constant offset can not match
  // the addressing mode, we can still do optimizations to other lowered parts
  // of variable indices. Therefore, we don't check for addressing modes in that
  // case.
  if (!LowerGEP) {
    unsigned AddrSpace = GEP->getPointerAddressSpace();
    if (!TTI.isLegalAddressingMode(GEP->getResultElementType(),
                                   /*BaseGV=*/nullptr, AccumulativeByteOffset,
                                   /*HasBaseReg=*/true, /*Scale=*/0,
                                   AddrSpace)) {
      return Changed;
    }
  }

  // Track information for preserving GEP flags.
  bool AllOffsetsNonNegative = AccumulativeByteOffset >= 0;
  bool AllNUWPreserved = GEP->hasNoUnsignedWrap();
  bool NewGEPInBounds = GEP->isInBounds();
  bool NewGEPNUSW = GEP->hasNoUnsignedSignedWrap();

  // Remove the constant offset in each sequential index. The resultant GEP
  // computes the variadic base.
  // Notice that we don't remove struct field indices here. If LowerGEP is
  // disabled, a structure index is not accumulated and we still use the old
  // one. If LowerGEP is enabled, a structure index is accumulated in the
  // constant offset. LowerToSingleIndexGEPs will later handle the constant
  // offset and won't need a new structure index.
  gep_type_iterator GTI = gep_type_begin(*GEP);
  for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
    if (GTI.isSequential()) {
      // Constant offsets of scalable types are not really constant.
      if (GTI.getIndexedType()->isScalableTy())
        continue;

      // Splits this GEP index into a variadic part and a constant offset, and
      // uses the variadic part as the new index.
      Value *OldIdx = GEP->getOperand(I);
      User *UserChainTail;
      bool PreservesNUW;
      Value *NewIdx = ConstantOffsetExtractor::Extract(
          OldIdx, GEP, UserChainTail, PreservesNUW);
      if (NewIdx != nullptr) {
        // Switches to the index with the constant offset removed.
        GEP->setOperand(I, NewIdx);
        // After switching to the new index, we can garbage-collect UserChain
        // and the old index if they are not used.
        RecursivelyDeleteTriviallyDeadInstructions(UserChainTail);
        RecursivelyDeleteTriviallyDeadInstructions(OldIdx);
        AllOffsetsNonNegative =
            AllOffsetsNonNegative && isKnownNonNegative(NewIdx, *DL);
        AllNUWPreserved &= PreservesNUW;
      }
    }
  }
  if (ExtractBase) {
    GEPOperator *Base = cast<GEPOperator>(GEP->getPointerOperand());
    AllNUWPreserved &= Base->hasNoUnsignedWrap();
    NewGEPInBounds &= Base->isInBounds();
    NewGEPNUSW &= Base->hasNoUnsignedSignedWrap();
    AllOffsetsNonNegative &= BaseByteOffset >= 0;

    GEP->setOperand(0, NewBase);
    RecursivelyDeleteTriviallyDeadInstructions(Base);
  }

  // Clear the inbounds attribute because the new index may be off-bound.
  // e.g.,
  //
  //   b     = add i64 a, 5
  //   addr  = gep inbounds float, float* p, i64 b
  //
  // is transformed to:
  //
  //   addr2 = gep float, float* p, i64 a ; inbounds removed
  //   addr  = gep float, float* addr2, i64 5 ; inbounds removed
  //
  // 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.
  GEPNoWrapFlags NewGEPFlags = GEPNoWrapFlags::none();

  // If the initial GEP was inbounds/nusw and all variable indices and the
  // accumulated offsets are non-negative, they can be added in any order and
  // the intermediate results are in bounds and don't overflow in a nusw sense.
  // So, we can preserve the inbounds/nusw flag for both GEPs.
  bool CanPreserveInBoundsNUSW = AllOffsetsNonNegative;

  // If the initial GEP was NUW and all operations that we reassociate were NUW
  // additions, the resulting GEPs are also NUW.
  if (AllNUWPreserved) {
    NewGEPFlags |= GEPNoWrapFlags::noUnsignedWrap();
    // If the initial GEP additionally had NUSW (or inbounds, which implies
    // NUSW), we know that the indices in the initial GEP must all have their
    // signbit not set. For indices that are the result of NUW adds, the
    // add-operands therefore also don't have their signbit set. Therefore, all
    // indices of the resulting GEPs are non-negative -> we can preserve
    // the inbounds/nusw flag.
    CanPreserveInBoundsNUSW |= NewGEPNUSW;
  }

  if (CanPreserveInBoundsNUSW) {
    if (NewGEPInBounds)
      NewGEPFlags |= GEPNoWrapFlags::inBounds();
    else if (NewGEPNUSW)
      NewGEPFlags |= GEPNoWrapFlags::noUnsignedSignedWrap();
  }

  GEP->setNoWrapFlags(NewGEPFlags);

  // Lowers a GEP to GEPs with a single index.
  if (LowerGEP) {
    lowerToSingleIndexGEPs(GEP, AccumulativeByteOffset);
    return true;
  }

  // No need to create another GEP if the accumulative byte offset is 0.
  if (AccumulativeByteOffset == 0)
    return true;

  // Offsets the base with the accumulative byte offset.
  //
  //   %gep                        ; the base
  //   ... %gep ...
  //
  // => add the offset
  //
  //   %gep2                       ; clone of %gep
  //   %new.gep = gep i8, %gep2, %offset
  //   %gep                        ; will be removed
  //   ... %gep ...
  //
  // => replace all uses of %gep with %new.gep and remove %gep
  //
  //   %gep2                       ; clone of %gep
  //   %new.gep = gep i8, %gep2, %offset
  //   ... %new.gep ...
  Instruction *NewGEP = GEP->clone();
  NewGEP->insertBefore(GEP->getIterator());

  Type *PtrIdxTy = DL->getIndexType(GEP->getType());
  IRBuilder<> Builder(GEP);
  NewGEP = cast<Instruction>(Builder.CreatePtrAdd(
      NewGEP, ConstantInt::get(PtrIdxTy, AccumulativeByteOffset, true),
      GEP->getName(), NewGEPFlags));
  NewGEP->copyMetadata(*GEP);

  GEP->replaceAllUsesWith(NewGEP);
  GEP->eraseFromParent();

  return true;
}

bool SeparateConstOffsetFromGEPLegacyPass::runOnFunction(Function &F) {
  if (skipFunction(F))
    return false;
  auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
  auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
  auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
  auto GetTTI = [this](Function &F) -> TargetTransformInfo & {
    return this->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
  };
  SeparateConstOffsetFromGEP Impl(DT, LI, TLI, GetTTI, LowerGEP);
  return Impl.run(F);
}

bool SeparateConstOffsetFromGEP::run(Function &F) {
  if (DisableSeparateConstOffsetFromGEP)
    return false;

  DL = &F.getDataLayout();
  bool Changed = false;

  ReversePostOrderTraversal<Function *> RPOT(&F);
  for (BasicBlock *B : RPOT) {
    if (!DT->isReachableFromEntry(B))
      continue;

    for (Instruction &I : llvm::make_early_inc_range(*B))
      if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I))
        Changed |= splitGEP(GEP);
    // No need to split GEP ConstantExprs because all its indices are constant
    // already.
  }

  Changed |= reuniteExts(F);

  if (VerifyNoDeadCode)
    verifyNoDeadCode(F);

  return Changed;
}

Instruction *SeparateConstOffsetFromGEP::findClosestMatchingDominator(
    ExprKey Key, Instruction *Dominatee,
    DenseMap<ExprKey, SmallVector<Instruction *, 2>> &DominatingExprs) {
  auto Pos = DominatingExprs.find(Key);
  if (Pos == DominatingExprs.end())
    return nullptr;

  auto &Candidates = Pos->second;
  // Because we process the basic blocks in pre-order of the dominator tree, a
  // candidate that doesn't dominate the current instruction won't dominate any
  // future instruction either. Therefore, we pop it out of the stack. This
  // optimization makes the algorithm O(n).
  while (!Candidates.empty()) {
    Instruction *Candidate = Candidates.back();
    if (DT->dominates(Candidate, Dominatee))
      return Candidate;
    Candidates.pop_back();
  }
  return nullptr;
}

bool SeparateConstOffsetFromGEP::reuniteExts(Instruction *I) {
  if (!I->getType()->isIntOrIntVectorTy())
    return false;

  //   Dom: LHS+RHS
  //   I: sext(LHS)+sext(RHS)
  // If Dom can't sign overflow and Dom dominates I, optimize I to sext(Dom).
  // TODO: handle zext
  Value *LHS = nullptr, *RHS = nullptr;
  if (match(I, m_Add(m_SExt(m_Value(LHS)), m_SExt(m_Value(RHS))))) {
    if (LHS->getType() == RHS->getType()) {
      ExprKey Key = createNormalizedCommutablePair(LHS, RHS);
      if (auto *Dom = findClosestMatchingDominator(Key, I, DominatingAdds)) {
        Instruction *NewSExt =
            new SExtInst(Dom, I->getType(), "", I->getIterator());
        NewSExt->takeName(I);
        I->replaceAllUsesWith(NewSExt);
        NewSExt->setDebugLoc(I->getDebugLoc());
        RecursivelyDeleteTriviallyDeadInstructions(I);
        return true;
      }
    }
  } else if (match(I, m_Sub(m_SExt(m_Value(LHS)), m_SExt(m_Value(RHS))))) {
    if (LHS->getType() == RHS->getType()) {
      if (auto *Dom =
              findClosestMatchingDominator({LHS, RHS}, I, DominatingSubs)) {
        Instruction *NewSExt =
            new SExtInst(Dom, I->getType(), "", I->getIterator());
        NewSExt->takeName(I);
        I->replaceAllUsesWith(NewSExt);
        NewSExt->setDebugLoc(I->getDebugLoc());
        RecursivelyDeleteTriviallyDeadInstructions(I);
        return true;
      }
    }
  }

  // Add I to DominatingExprs if it's an add/sub that can't sign overflow.
  if (match(I, m_NSWAdd(m_Value(LHS), m_Value(RHS)))) {
    if (programUndefinedIfPoison(I)) {
      ExprKey Key = createNormalizedCommutablePair(LHS, RHS);
      DominatingAdds[Key].push_back(I);
    }
  } else if (match(I, m_NSWSub(m_Value(LHS), m_Value(RHS)))) {
    if (programUndefinedIfPoison(I))
      DominatingSubs[{LHS, RHS}].push_back(I);
  }
  return false;
}

bool SeparateConstOffsetFromGEP::reuniteExts(Function &F) {
  bool Changed = false;
  DominatingAdds.clear();
  DominatingSubs.clear();
  for (const auto Node : depth_first(DT)) {
    BasicBlock *BB = Node->getBlock();
    for (Instruction &I : llvm::make_early_inc_range(*BB))
      Changed |= reuniteExts(&I);
  }
  return Changed;
}

void SeparateConstOffsetFromGEP::verifyNoDeadCode(Function &F) {
  for (BasicBlock &B : F) {
    for (Instruction &I : B) {
      if (isInstructionTriviallyDead(&I)) {
        std::string ErrMessage;
        raw_string_ostream RSO(ErrMessage);
        RSO << "Dead instruction detected!\n" << I << "\n";
        llvm_unreachable(RSO.str().c_str());
      }
    }
  }
}

bool SeparateConstOffsetFromGEP::isLegalToSwapOperand(
    GetElementPtrInst *FirstGEP, GetElementPtrInst *SecondGEP, Loop *CurLoop) {
  if (!FirstGEP || !FirstGEP->hasOneUse())
    return false;

  if (!SecondGEP || FirstGEP->getParent() != SecondGEP->getParent())
    return false;

  if (FirstGEP == SecondGEP)
    return false;

  unsigned FirstNum = FirstGEP->getNumOperands();
  unsigned SecondNum = SecondGEP->getNumOperands();
  // Give up if the number of operands are not 2.
  if (FirstNum != SecondNum || FirstNum != 2)
    return false;

  Value *FirstBase = FirstGEP->getOperand(0);
  Value *SecondBase = SecondGEP->getOperand(0);
  Value *FirstOffset = FirstGEP->getOperand(1);
  // Give up if the index of the first GEP is loop invariant.
  if (CurLoop->isLoopInvariant(FirstOffset))
    return false;

  // Give up if base doesn't have same type.
  if (FirstBase->getType() != SecondBase->getType())
    return false;

  Instruction *FirstOffsetDef = dyn_cast<Instruction>(FirstOffset);

  // Check if the second operand of first GEP has constant coefficient.
  // For an example, for the following code,  we won't gain anything by
  // hoisting the second GEP out because the second GEP can be folded away.
  //   %scevgep.sum.ur159 = add i64 %idxprom48.ur, 256
  //   %67 = shl i64 %scevgep.sum.ur159, 2
  //   %uglygep160 = getelementptr i8* %65, i64 %67
  //   %uglygep161 = getelementptr i8* %uglygep160, i64 -1024

  // Skip constant shift instruction which may be generated by Splitting GEPs.
  if (FirstOffsetDef && FirstOffsetDef->isShift() &&
      isa<ConstantInt>(FirstOffsetDef->getOperand(1)))
    FirstOffsetDef = dyn_cast<Instruction>(FirstOffsetDef->getOperand(0));

  // Give up if FirstOffsetDef is an Add or Sub with constant.
  // Because it may not profitable at all due to constant folding.
  if (FirstOffsetDef)
    if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FirstOffsetDef)) {
      unsigned opc = BO->getOpcode();
      if ((opc == Instruction::Add || opc == Instruction::Sub) &&
          (isa<ConstantInt>(BO->getOperand(0)) ||
           isa<ConstantInt>(BO->getOperand(1))))
        return false;
    }
  return true;
}

bool SeparateConstOffsetFromGEP::hasMoreThanOneUseInLoop(Value *V, Loop *L) {
  // TODO: Could look at uses of globals, but we need to make sure we are
  // looking at the correct function.
  if (isa<Constant>(V))
    return false;

  int UsesInLoop = 0;
  for (User *U : V->users()) {
    if (Instruction *User = dyn_cast<Instruction>(U))
      if (L->contains(User))
        if (++UsesInLoop > 1)
          return true;
  }
  return false;
}

void SeparateConstOffsetFromGEP::swapGEPOperand(GetElementPtrInst *First,
                                                GetElementPtrInst *Second) {
  Value *Offset1 = First->getOperand(1);
  Value *Offset2 = Second->getOperand(1);
  First->setOperand(1, Offset2);
  Second->setOperand(1, Offset1);

  // We changed p+o+c to p+c+o, p+c may not be inbound anymore.
  const DataLayout &DAL = First->getDataLayout();
  APInt Offset(DAL.getIndexSizeInBits(
                   cast<PointerType>(First->getType())->getAddressSpace()),
               0);
  Value *NewBase =
      First->stripAndAccumulateInBoundsConstantOffsets(DAL, Offset);
  uint64_t ObjectSize;
  if (!getObjectSize(NewBase, ObjectSize, DAL, TLI) ||
     Offset.ugt(ObjectSize)) {
    // TODO(gep_nowrap): Make flag preservation more precise.
    First->setNoWrapFlags(GEPNoWrapFlags::none());
    Second->setNoWrapFlags(GEPNoWrapFlags::none());
  } else
    First->setIsInBounds(true);
}

void SeparateConstOffsetFromGEPPass::printPipeline(
    raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
  static_cast<PassInfoMixin<SeparateConstOffsetFromGEPPass> *>(this)
      ->printPipeline(OS, MapClassName2PassName);
  OS << '<';
  if (LowerGEP)
    OS << "lower-gep";
  OS << '>';
}

PreservedAnalyses
SeparateConstOffsetFromGEPPass::run(Function &F, FunctionAnalysisManager &AM) {
  auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
  auto *LI = &AM.getResult<LoopAnalysis>(F);
  auto *TLI = &AM.getResult<TargetLibraryAnalysis>(F);
  auto GetTTI = [&AM](Function &F) -> TargetTransformInfo & {
    return AM.getResult<TargetIRAnalysis>(F);
  };
  SeparateConstOffsetFromGEP Impl(DT, LI, TLI, GetTTI, LowerGEP);
  if (!Impl.run(F))
    return PreservedAnalyses::all();
  PreservedAnalyses PA;
  PA.preserveSet<CFGAnalyses>();
  return PA;
}