xref: /dports/devel/intel-graphics-compiler/intel-graphics-compiler-igc-1.0.9636/IGC/Compiler/Optimizer/IGCInstCombiner/7.0/InstCombineLoadStoreAlloca.cpp

/*========================== begin_copyright_notice ============================

Copyright (C) 2018-2021 Intel Corporation

SPDX-License-Identifier: MIT

============================= end_copyright_notice ===========================*/

/*========================== begin_copyright_notice ============================

This file is distributed under the University of Illinois Open Source License.
See LICENSE.TXT for details.

============================= end_copyright_notice ===========================*/

// This file implements the visit functions for load, store and alloca.

#include "common/LLVMWarningsPush.hpp"
#include "InstCombineInternal.h"
#include "llvm/ADT/MapVector.h"
#include "llvm/ADT/SmallString.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Loads.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/IR/ConstantRange.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/MDBuilder.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "Probe/Assertion.h"

using namespace llvm;
using namespace PatternMatch;
using namespace IGCombiner;

#define DEBUG_TYPE "instcombine"

STATISTIC(NumDeadStore, "Number of dead stores eliminated");
STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global");

/// pointsToConstantGlobal - Return true if V (possibly indirectly) points to
/// some part of a constant global variable.  This intentionally only accepts
/// constant expressions because we can't rewrite arbitrary instructions.
static bool pointsToConstantGlobal(Value* V) {
    if (GlobalVariable * GV = dyn_cast<GlobalVariable>(V))
        return GV->isConstant();

    if (ConstantExpr * CE = dyn_cast<ConstantExpr>(V)) {
        if (CE->getOpcode() == Instruction::BitCast ||
            CE->getOpcode() == Instruction::AddrSpaceCast ||
            CE->getOpcode() == Instruction::GetElementPtr)
            return pointsToConstantGlobal(CE->getOperand(0));
    }
    return false;
}

/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
/// pointer to an alloca.  Ignore any reads of the pointer, return false if we
/// see any stores or other unknown uses.  If we see pointer arithmetic, keep
/// track of whether it moves the pointer (with IsOffset) but otherwise traverse
/// the uses.  If we see a memcpy/memmove that targets an unoffseted pointer to
/// the alloca, and if the source pointer is a pointer to a constant global, we
/// can optimize this.
static bool
isOnlyCopiedFromConstantGlobal(Value* V, MemTransferInst*& TheCopy,
    SmallVectorImpl<Instruction*>& ToDelete) {
    // We track lifetime intrinsics as we encounter them.  If we decide to go
    // ahead and replace the value with the global, this lets the caller quickly
    // eliminate the markers.

    SmallVector<std::pair<Value*, bool>, 35> ValuesToInspect;
    ValuesToInspect.emplace_back(V, false);
    while (!ValuesToInspect.empty()) {
        auto ValuePair = ValuesToInspect.pop_back_val();
        const bool IsOffset = ValuePair.second;
        for (auto& U : ValuePair.first->uses()) {
            auto* I = cast<Instruction>(U.getUser());

            if (auto * LI = dyn_cast<LoadInst>(I)) {
                // Ignore non-volatile loads, they are always ok.
                if (!LI->isSimple()) return false;
                continue;
            }

            if (isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I)) {
                // If uses of the bitcast are ok, we are ok.
                ValuesToInspect.emplace_back(I, IsOffset);
                continue;
            }
            if (auto * GEP = dyn_cast<GetElementPtrInst>(I)) {
                // If the GEP has all zero indices, it doesn't offset the pointer. If it
                // doesn't, it does.
                ValuesToInspect.emplace_back(I, IsOffset || !GEP->hasAllZeroIndices());
                continue;
            }

            if (auto CS = CallSite(I)) {
                // If this is the function being called then we treat it like a load and
                // ignore it.
                if (CS.isCallee(&U))
                    continue;

                unsigned DataOpNo = CS.getDataOperandNo(&U);
                bool IsArgOperand = CS.isArgOperand(&U);

                // Inalloca arguments are clobbered by the call.
                if (IsArgOperand && CS.isInAllocaArgument(DataOpNo))
                    return false;

                // If this is a readonly/readnone call site, then we know it is just a
                // load (but one that potentially returns the value itself), so we can
                // ignore it if we know that the value isn't captured.
                if (CS.onlyReadsMemory() &&
                    (CS.getInstruction()->use_empty() || CS.doesNotCapture(DataOpNo)))
                    continue;

                // If this is being passed as a byval argument, the caller is making a
                // copy, so it is only a read of the alloca.
                if (IsArgOperand && CS.isByValArgument(DataOpNo))
                    continue;
            }

            // Lifetime intrinsics can be handled by the caller.
            if (IntrinsicInst * II = dyn_cast<IntrinsicInst>(I)) {
                if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
                    II->getIntrinsicID() == Intrinsic::lifetime_end) {
                    IGC_ASSERT_MESSAGE(II->use_empty(), "Lifetime markers have no result to use!");
                    ToDelete.push_back(II);
                    continue;
                }
            }

            // If this is isn't our memcpy/memmove, reject it as something we can't
            // handle.
            MemTransferInst* MI = dyn_cast<MemTransferInst>(I);
            if (!MI)
                return false;

            // If the transfer is using the alloca as a source of the transfer, then
            // ignore it since it is a load (unless the transfer is volatile).
            if (U.getOperandNo() == 1) {
                if (MI->isVolatile()) return false;
                continue;
            }

            // If we already have seen a copy, reject the second one.
            if (TheCopy) return false;

            // If the pointer has been offset from the start of the alloca, we can't
            // safely handle this.
            if (IsOffset) return false;

            // If the memintrinsic isn't using the alloca as the dest, reject it.
            if (U.getOperandNo() != 0) return false;

            // If the source of the memcpy/move is not a constant global, reject it.
            if (!pointsToConstantGlobal(MI->getSource()))
                return false;

            // Otherwise, the transform is safe.  Remember the copy instruction.
            TheCopy = MI;
        }
    }
    return true;
}

/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
/// modified by a copy from a constant global.  If we can prove this, we can
/// replace any uses of the alloca with uses of the global directly.
static MemTransferInst*
isOnlyCopiedFromConstantGlobal(AllocaInst* AI,
    SmallVectorImpl<Instruction*>& ToDelete) {
    MemTransferInst* TheCopy = nullptr;
    if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete))
        return TheCopy;
    return nullptr;
}

/// Returns true if V is dereferenceable for size of alloca.
static bool isDereferenceableForAllocaSize(const Value* V, const AllocaInst* AI,
    const DataLayout& DL) {
    if (AI->isArrayAllocation())
        return false;
    uint64_t AllocaSize = DL.getTypeStoreSize(AI->getAllocatedType());
    if (!AllocaSize)
        return false;
    return isDereferenceableAndAlignedPointer(V, AI->getAlignment(),
        APInt(64, AllocaSize), DL);
}

static Instruction* simplifyAllocaArraySize(InstCombiner& IC, AllocaInst& AI) {
    // Check for array size of 1 (scalar allocation).
    if (!AI.isArrayAllocation()) {
        // i32 1 is the canonical array size for scalar allocations.
        if (AI.getArraySize()->getType()->isIntegerTy(32))
            return nullptr;

        // Canonicalize it.
        Value* V = IC.Builder.getInt32(1);
        AI.setOperand(0, V);
        return &AI;
    }

    // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
    if (const ConstantInt * C = dyn_cast<ConstantInt>(AI.getArraySize())) {
        Type* NewTy = ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
        AllocaInst* New = IC.Builder.CreateAlloca(NewTy, nullptr, AI.getName());
        New->setAlignment(AI.getAlignment());

        // Scan to the end of the allocation instructions, to skip over a block of
        // allocas if possible...also skip interleaved debug info
        //
        BasicBlock::iterator It(New);
        while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It))
            ++It;

        // Now that I is pointing to the first non-allocation-inst in the block,
        // insert our getelementptr instruction...
        //
        Type* IdxTy = IC.getDataLayout().getIntPtrType(AI.getType());
        Value* NullIdx = Constant::getNullValue(IdxTy);
        Value* Idx[2] = { NullIdx, NullIdx };
        Instruction* GEP =
            GetElementPtrInst::CreateInBounds(New, Idx, New->getName() + ".sub");
        IC.InsertNewInstBefore(GEP, *It);

        // Now make everything use the getelementptr instead of the original
        // allocation.
        return IC.replaceInstUsesWith(AI, GEP);
    }

    if (isa<UndefValue>(AI.getArraySize()))
        return IC.replaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));

    // Ensure that the alloca array size argument has type intptr_t, so that
    // any casting is exposed early.
    Type* IntPtrTy = IC.getDataLayout().getIntPtrType(AI.getType());
    if (AI.getArraySize()->getType() != IntPtrTy) {
        Value* V = IC.Builder.CreateIntCast(AI.getArraySize(), IntPtrTy, false);
        AI.setOperand(0, V);
        return &AI;
    }

    return nullptr;
}

namespace {
    // If I and V are pointers in different address space, it is not allowed to
    // use replaceAllUsesWith since I and V have different types. A
    // non-target-specific transformation should not use addrspacecast on V since
    // the two address space may be disjoint depending on target.
    //
    // This class chases down uses of the old pointer until reaching the load
    // instructions, then replaces the old pointer in the load instructions with
    // the new pointer. If during the chasing it sees bitcast or GEP, it will
    // create new bitcast or GEP with the new pointer and use them in the load
    // instruction.
    class PointerReplacer {
    public:
        PointerReplacer(InstCombiner& IC) : IC(IC) {}
        void replacePointer(Instruction& I, Value* V);

    private:
        void findLoadAndReplace(Instruction& I);
        void replace(Instruction* I);
        Value* getReplacement(Value* I);

        SmallVector<Instruction*, 4> Path;
        MapVector<Value*, Value*> WorkMap;
        InstCombiner& IC;
    };
} // end anonymous namespace

void PointerReplacer::findLoadAndReplace(Instruction& I) {
    for (auto U : I.users()) {
        auto* Inst = dyn_cast<Instruction>(&*U);
        if (!Inst)
            return;
        LLVM_DEBUG(dbgs() << "Found pointer user: " << *U << '\n');
        if (isa<LoadInst>(Inst)) {
            for (auto P : Path)
                replace(P);
            replace(Inst);
        }
        else if (isa<GetElementPtrInst>(Inst) || isa<BitCastInst>(Inst)) {
            Path.push_back(Inst);
            findLoadAndReplace(*Inst);
            Path.pop_back();
        }
        else {
            return;
        }
    }
}

Value* PointerReplacer::getReplacement(Value* V) {
    auto Loc = WorkMap.find(V);
    if (Loc != WorkMap.end())
        return Loc->second;
    return nullptr;
}

void PointerReplacer::replace(Instruction* I) {
    if (getReplacement(I))
        return;

    if (auto * LT = dyn_cast<LoadInst>(I)) {
        auto* V = getReplacement(LT->getPointerOperand());
        IGC_ASSERT_MESSAGE(V, "Operand not replaced");
        auto* NewI = new LoadInst(V);
        NewI->takeName(LT);
        IC.InsertNewInstWith(NewI, *LT);
        IC.replaceInstUsesWith(*LT, NewI);
        WorkMap[LT] = NewI;
    }
    else if (auto * GEP = dyn_cast<GetElementPtrInst>(I)) {
        auto* V = getReplacement(GEP->getPointerOperand());
        IGC_ASSERT_MESSAGE(V, "Operand not replaced");
        SmallVector<Value*, 8> Indices;
        Indices.append(GEP->idx_begin(), GEP->idx_end());
        auto* NewI = GetElementPtrInst::Create(
            V->getType()->getPointerElementType(), V, Indices);
        IC.InsertNewInstWith(NewI, *GEP);
        NewI->takeName(GEP);
        WorkMap[GEP] = NewI;
    }
    else if (auto * BC = dyn_cast<BitCastInst>(I)) {
        auto* V = getReplacement(BC->getOperand(0));
        IGC_ASSERT_MESSAGE(V, "Operand not replaced");
        auto* NewT = PointerType::get(BC->getType()->getPointerElementType(),
            V->getType()->getPointerAddressSpace());
        auto* NewI = new BitCastInst(V, NewT);
        IC.InsertNewInstWith(NewI, *BC);
        NewI->takeName(BC);
        WorkMap[BC] = NewI;
    }
    else {
        IGC_ASSERT_EXIT_MESSAGE(0, "should never reach here");
    }
}

void PointerReplacer::replacePointer(Instruction& I, Value* V) {
    IGC_ASSERT(nullptr != V);
    IGC_ASSERT(nullptr != cast<PointerType>(I.getType()));
    IGC_ASSERT(nullptr != cast<PointerType>(V->getType()));
    IGC_ASSERT(cast<PointerType>(I.getType()) != cast<PointerType>(V->getType()));
    IGC_ASSERT(cast<PointerType>(I.getType())->getElementType() == cast<PointerType>(V->getType())->getElementType());
    WorkMap[&I] = V;
    findLoadAndReplace(I);
}

Instruction* InstCombiner::visitAllocaInst(AllocaInst& AI) {
    if (auto * I = simplifyAllocaArraySize(*this, AI))
        return I;

    if (AI.getAllocatedType()->isSized()) {
        // If the alignment is 0 (unspecified), assign it the preferred alignment.
        if (AI.getAlignment() == 0)
            AI.setAlignment(DL.getPrefTypeAlignment(AI.getAllocatedType()));

        // Move all alloca's of zero byte objects to the entry block and merge them
        // together.  Note that we only do this for alloca's, because malloc should
        // allocate and return a unique pointer, even for a zero byte allocation.
        if (DL.getTypeAllocSize(AI.getAllocatedType()) == 0) {
            // For a zero sized alloca there is no point in doing an array allocation.
            // This is helpful if the array size is a complicated expression not used
            // elsewhere.
            if (AI.isArrayAllocation()) {
                AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
                return &AI;
            }

            // Get the first instruction in the entry block.
            BasicBlock& EntryBlock = AI.getParent()->getParent()->getEntryBlock();
            Instruction* FirstInst = EntryBlock.getFirstNonPHIOrDbg();
            if (FirstInst != &AI) {
                // If the entry block doesn't start with a zero-size alloca then move
                // this one to the start of the entry block.  There is no problem with
                // dominance as the array size was forced to a constant earlier already.
                AllocaInst* EntryAI = dyn_cast<AllocaInst>(FirstInst);
                if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
                    DL.getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
                    AI.moveBefore(FirstInst);
                    return &AI;
                }

                // If the alignment of the entry block alloca is 0 (unspecified),
                // assign it the preferred alignment.
                if (EntryAI->getAlignment() == 0)
                    EntryAI->setAlignment(
                        DL.getPrefTypeAlignment(EntryAI->getAllocatedType()));
                // Replace this zero-sized alloca with the one at the start of the entry
                // block after ensuring that the address will be aligned enough for both
                // types.
                unsigned MaxAlign = std::max(EntryAI->getAlignment(),
                    AI.getAlignment());
                EntryAI->setAlignment(MaxAlign);
                if (AI.getType() != EntryAI->getType())
                    return new BitCastInst(EntryAI, AI.getType());
                return replaceInstUsesWith(AI, EntryAI);
            }
        }
    }

    if (AI.getAlignment()) {
        // Check to see if this allocation is only modified by a memcpy/memmove from
        // a constant global whose alignment is equal to or exceeds that of the
        // allocation.  If this is the case, we can change all users to use
        // the constant global instead.  This is commonly produced by the CFE by
        // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
        // is only subsequently read.
        SmallVector<Instruction*, 4> ToDelete;
        if (MemTransferInst * Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
            unsigned SourceAlign = getOrEnforceKnownAlignment(
                Copy->getSource(), AI.getAlignment(), DL, &AI, &AC, &DT);
            if (AI.getAlignment() <= SourceAlign &&
                isDereferenceableForAllocaSize(Copy->getSource(), &AI, DL)) {
                LLVM_DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
                LLVM_DEBUG(dbgs() << "  memcpy = " << *Copy << '\n');
                for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
                    eraseInstFromFunction(*ToDelete[i]);
                Constant* TheSrc = cast<Constant>(Copy->getSource());
                auto* SrcTy = TheSrc->getType();
                auto* DestTy = PointerType::get(AI.getType()->getPointerElementType(),
                    SrcTy->getPointerAddressSpace());
                Constant* Cast =
                    ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, DestTy);
                if (AI.getType()->getPointerAddressSpace() ==
                    SrcTy->getPointerAddressSpace()) {
                    Instruction* NewI = replaceInstUsesWith(AI, Cast);
                    eraseInstFromFunction(*Copy);
                    ++NumGlobalCopies;
                    return NewI;
                }
                else {
                    PointerReplacer PtrReplacer(*this);
                    PtrReplacer.replacePointer(AI, Cast);
                    ++NumGlobalCopies;
                }
            }
        }
    }

    // At last, use the generic allocation site handler to aggressively remove
    // unused allocas.
    return visitAllocSite(AI);
}

// Are we allowed to form a atomic load or store of this type?
static bool isSupportedAtomicType(Type* Ty) {
    return Ty->isIntOrPtrTy() || Ty->isFloatingPointTy();
}

/// Helper to combine a load to a new type.
///
/// This just does the work of combining a load to a new type. It handles
/// metadata, etc., and returns the new instruction. The \c NewTy should be the
/// loaded *value* type. This will convert it to a pointer, cast the operand to
/// that pointer type, load it, etc.
///
/// Note that this will create all of the instructions with whatever insert
/// point the \c InstCombiner currently is using.
static LoadInst* combineLoadToNewType(InstCombiner& IC, LoadInst& LI, Type* NewTy,
    const Twine& Suffix = "") {
    IGC_ASSERT_MESSAGE((!LI.isAtomic() || isSupportedAtomicType(NewTy)), "can't fold an atomic load to requested type");

    Value* Ptr = LI.getPointerOperand();
    unsigned AS = LI.getPointerAddressSpace();
    SmallVector<std::pair<unsigned, MDNode*>, 8> MD;
    LI.getAllMetadata(MD);

    Value* NewPtr = nullptr;
    if (!(match(Ptr, m_BitCast(m_Value(NewPtr))) &&
        NewPtr->getType()->getPointerElementType() == NewTy &&
        NewPtr->getType()->getPointerAddressSpace() == AS))
        NewPtr = IC.Builder.CreateBitCast(Ptr, NewTy->getPointerTo(AS));

    LoadInst* NewLoad = IC.Builder.CreateAlignedLoad(
        NewPtr, LI.getAlignment(), LI.isVolatile(), LI.getName() + Suffix);
    NewLoad->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
    MDBuilder MDB(NewLoad->getContext());
    for (const auto& MDPair : MD) {
        unsigned ID = MDPair.first;
        MDNode* N = MDPair.second;
        // Note, essentially every kind of metadata should be preserved here! This
        // routine is supposed to clone a load instruction changing *only its type*.
        // The only metadata it makes sense to drop is metadata which is invalidated
        // when the pointer type changes. This should essentially never be the case
        // in LLVM, but we explicitly switch over only known metadata to be
        // conservatively correct. If you are adding metadata to LLVM which pertains
        // to loads, you almost certainly want to add it here.
        switch (ID) {
        case LLVMContext::MD_dbg:
        case LLVMContext::MD_tbaa:
        case LLVMContext::MD_prof:
        case LLVMContext::MD_fpmath:
        case LLVMContext::MD_tbaa_struct:
        case LLVMContext::MD_invariant_load:
        case LLVMContext::MD_alias_scope:
        case LLVMContext::MD_noalias:
        case LLVMContext::MD_nontemporal:
        case LLVMContext::MD_mem_parallel_loop_access:
            // All of these directly apply.
            NewLoad->setMetadata(ID, N);
            break;

        case LLVMContext::MD_nonnull:
            copyNonnullMetadata(LI, N, *NewLoad);
            break;
        case LLVMContext::MD_align:
        case LLVMContext::MD_dereferenceable:
        case LLVMContext::MD_dereferenceable_or_null:
            // These only directly apply if the new type is also a pointer.
            if (NewTy->isPointerTy())
                NewLoad->setMetadata(ID, N);
            break;
        case LLVMContext::MD_range:
            copyRangeMetadata(IC.getDataLayout(), LI, N, *NewLoad);
            break;
        }
    }
    return NewLoad;
}

/// Combine a store to a new type.
///
/// Returns the newly created store instruction.
static StoreInst* combineStoreToNewValue(InstCombiner& IC, StoreInst& SI, Value* V) {
    IGC_ASSERT_MESSAGE((!SI.isAtomic() || isSupportedAtomicType(V->getType())), "can't fold an atomic store of requested type");

    Value* Ptr = SI.getPointerOperand();
    unsigned AS = SI.getPointerAddressSpace();
    SmallVector<std::pair<unsigned, MDNode*>, 8> MD;
    SI.getAllMetadata(MD);

    StoreInst* NewStore = IC.Builder.CreateAlignedStore(
        V, IC.Builder.CreateBitCast(Ptr, V->getType()->getPointerTo(AS)),
        SI.getAlignment(), SI.isVolatile());
    NewStore->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
    for (const auto& MDPair : MD) {
        unsigned ID = MDPair.first;
        MDNode* N = MDPair.second;
        // Note, essentially every kind of metadata should be preserved here! This
        // routine is supposed to clone a store instruction changing *only its
        // type*. The only metadata it makes sense to drop is metadata which is
        // invalidated when the pointer type changes. This should essentially
        // never be the case in LLVM, but we explicitly switch over only known
        // metadata to be conservatively correct. If you are adding metadata to
        // LLVM which pertains to stores, you almost certainly want to add it
        // here.
        switch (ID) {
        case LLVMContext::MD_dbg:
        case LLVMContext::MD_tbaa:
        case LLVMContext::MD_prof:
        case LLVMContext::MD_fpmath:
        case LLVMContext::MD_tbaa_struct:
        case LLVMContext::MD_alias_scope:
        case LLVMContext::MD_noalias:
        case LLVMContext::MD_nontemporal:
        case LLVMContext::MD_mem_parallel_loop_access:
            // All of these directly apply.
            NewStore->setMetadata(ID, N);
            break;

        case LLVMContext::MD_invariant_load:
        case LLVMContext::MD_nonnull:
        case LLVMContext::MD_range:
        case LLVMContext::MD_align:
        case LLVMContext::MD_dereferenceable:
        case LLVMContext::MD_dereferenceable_or_null:
            // These don't apply for stores.
            break;
        }
    }

    return NewStore;
}

/// Returns true if instruction represent minmax pattern like:
///   select ((cmp load V1, load V2), V1, V2).
static bool isMinMaxWithLoads(Value* V) {
    IGC_ASSERT_MESSAGE(V->getType()->isPointerTy(), "Expected pointer type.");
    // Ignore possible ty* to ixx* bitcast.
    V = peekThroughBitcast(V);
    // Check that select is select ((cmp load V1, load V2), V1, V2) - minmax
    // pattern.
    CmpInst::Predicate Pred;
    Instruction* L1;
    Instruction* L2;
    Value* LHS;
    Value* RHS;
    if (!match(V, m_Select(m_Cmp(Pred, m_Instruction(L1), m_Instruction(L2)),
        m_Value(LHS), m_Value(RHS))))
        return false;
    return (match(L1, m_Load(m_Specific(LHS))) &&
        match(L2, m_Load(m_Specific(RHS)))) ||
        (match(L1, m_Load(m_Specific(RHS))) &&
            match(L2, m_Load(m_Specific(LHS))));
}

/// Combine loads to match the type of their uses' value after looking
/// through intervening bitcasts.
///
/// The core idea here is that if the result of a load is used in an operation,
/// we should load the type most conducive to that operation. For example, when
/// loading an integer and converting that immediately to a pointer, we should
/// instead directly load a pointer.
///
/// However, this routine must never change the width of a load or the number of
/// loads as that would introduce a semantic change. This combine is expected to
/// be a semantic no-op which just allows loads to more closely model the types
/// of their consuming operations.
///
/// Currently, we also refuse to change the precise type used for an atomic load
/// or a volatile load. This is debatable, and might be reasonable to change
/// later. However, it is risky in case some backend or other part of LLVM is
/// relying on the exact type loaded to select appropriate atomic operations.
static Instruction* combineLoadToOperationType(InstCombiner& IC, LoadInst& LI) {
    // FIXME: We could probably with some care handle both volatile and ordered
    // atomic loads here but it isn't clear that this is important.
    if (!LI.isUnordered())
        return nullptr;

    if (LI.use_empty())
        return nullptr;

    // swifterror values can't be bitcasted.
    if (LI.getPointerOperand()->isSwiftError())
        return nullptr;

    Type* Ty = LI.getType();
    const DataLayout& DL = IC.getDataLayout();

    // Try to canonicalize loads which are only ever stored to operate over
    // integers instead of any other type. We only do this when the loaded type
    // is sized and has a size exactly the same as its store size and the store
    // size is a legal integer type.
    // Do not perform canonicalization if minmax pattern is found (to avoid
    // infinite loop).
    if (!Ty->isIntegerTy() && Ty->isSized() &&
        DL.isLegalInteger(DL.getTypeStoreSizeInBits(Ty)) &&
        DL.getTypeStoreSizeInBits(Ty) == DL.getTypeSizeInBits(Ty) &&
        !DL.isNonIntegralPointerType(Ty) &&
        !isMinMaxWithLoads(
            peekThroughBitcast(LI.getPointerOperand(), /*OneUseOnly=*/true))) {
        if (all_of(LI.users(), [&LI](User* U) {
            auto* SI = dyn_cast<StoreInst>(U);
            return SI && SI->getPointerOperand() != &LI &&
                !SI->getPointerOperand()->isSwiftError();
        })) {
            LoadInst* NewLoad = combineLoadToNewType(
                IC, LI,
                Type::getIntNTy(LI.getContext(), DL.getTypeStoreSizeInBits(Ty)));
            // Replace all the stores with stores of the newly loaded value.
            for (auto UI = LI.user_begin(), UE = LI.user_end(); UI != UE;) {
                auto* SI = cast<StoreInst>(*UI++);
                IC.Builder.SetInsertPoint(SI);
                combineStoreToNewValue(IC, *SI, NewLoad);
                IC.eraseInstFromFunction(*SI);
            }
            IGC_ASSERT_MESSAGE(LI.use_empty(), "Failed to remove all users of the load!");
            // Return the old load so the combiner can delete it safely.
            return &LI;
        }
    }

    // Fold away bit casts of the loaded value by loading the desired type.
    // We can do this for BitCastInsts as well as casts from and to pointer types,
    // as long as those are noops (i.e., the source or dest type have the same
    // bitwidth as the target's pointers).
    if (LI.hasOneUse())
        if (auto * CI = dyn_cast<CastInst>(LI.user_back()))
            if (CI->isNoopCast(DL))
                if (!LI.isAtomic() || isSupportedAtomicType(CI->getDestTy())) {
                    LoadInst* NewLoad = combineLoadToNewType(IC, LI, CI->getDestTy());
                    CI->replaceAllUsesWith(NewLoad);
                    IC.eraseInstFromFunction(*CI);
                    return &LI;
                }

    // FIXME: We should also canonicalize loads of vectors when their elements are
    // cast to other types.
    return nullptr;
}

static Instruction* unpackLoadToAggregate(InstCombiner& IC, LoadInst& LI) {
    // FIXME: We could probably with some care handle both volatile and atomic
    // stores here but it isn't clear that this is important.
    if (!LI.isSimple())
        return nullptr;

    Type* T = LI.getType();
    if (!T->isAggregateType())
        return nullptr;

    StringRef Name = LI.getName();
    IGC_ASSERT_MESSAGE(LI.getAlignment(), "Alignment must be set at this point");

    if (auto * ST = dyn_cast<StructType>(T)) {
        // If the struct only have one element, we unpack.
        auto NumElements = ST->getNumElements();
        if (NumElements == 1) {
            LoadInst* NewLoad = combineLoadToNewType(IC, LI, ST->getTypeAtIndex(0U),
                ".unpack");
            AAMDNodes AAMD;
            LI.getAAMetadata(AAMD);
            NewLoad->setAAMetadata(AAMD);
            return IC.replaceInstUsesWith(LI, IC.Builder.CreateInsertValue(
                UndefValue::get(T), NewLoad, 0, Name));
        }

        // We don't want to break loads with padding here as we'd loose
        // the knowledge that padding exists for the rest of the pipeline.
        const DataLayout& DL = IC.getDataLayout();
        auto* SL = DL.getStructLayout(ST);
        if (SL->hasPadding())
            return nullptr;

        auto Align = LI.getAlignment();
        if (!Align)
            Align = DL.getABITypeAlignment(ST);

        auto* Addr = LI.getPointerOperand();
        auto* IdxType = Type::getInt32Ty(T->getContext());
        auto* Zero = ConstantInt::get(IdxType, 0);

        Value* V = UndefValue::get(T);
        for (unsigned i = 0; i < NumElements; i++) {
            Value* Indices[2] = {
              Zero,
              ConstantInt::get(IdxType, i),
            };
            auto* Ptr = IC.Builder.CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices),
                Name + ".elt");
            auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
            auto* L = IC.Builder.CreateAlignedLoad(Ptr, EltAlign, Name + ".unpack");
            // Propagate AA metadata. It'll still be valid on the narrowed load.
            AAMDNodes AAMD;
            LI.getAAMetadata(AAMD);
            L->setAAMetadata(AAMD);
            V = IC.Builder.CreateInsertValue(V, L, i);
        }

        V->setName(Name);
        return IC.replaceInstUsesWith(LI, V);
    }

    if (auto * AT = dyn_cast<ArrayType>(T)) {
        auto* ET = AT->getElementType();
        auto NumElements = AT->getNumElements();
        if (NumElements == 1) {
            LoadInst* NewLoad = combineLoadToNewType(IC, LI, ET, ".unpack");
            AAMDNodes AAMD;
            LI.getAAMetadata(AAMD);
            NewLoad->setAAMetadata(AAMD);
            return IC.replaceInstUsesWith(LI, IC.Builder.CreateInsertValue(
                UndefValue::get(T), NewLoad, 0, Name));
        }

        // Bail out if the array is too large. Ideally we would like to optimize
        // arrays of arbitrary size but this has a terrible impact on compile time.
        // The threshold here is chosen arbitrarily, maybe needs a little bit of
        // tuning.
        if (NumElements > IC.MaxArraySizeForCombine)
            return nullptr;

        const DataLayout& DL = IC.getDataLayout();
        auto EltSize = DL.getTypeAllocSize(ET);
        auto Align = LI.getAlignment();
        if (!Align)
            Align = DL.getABITypeAlignment(T);

        auto* Addr = LI.getPointerOperand();
        auto* IdxType = Type::getInt64Ty(T->getContext());
        auto* Zero = ConstantInt::get(IdxType, 0);

        Value* V = UndefValue::get(T);
        uint64_t Offset = 0;
        for (uint64_t i = 0; i < NumElements; i++) {
            Value* Indices[2] = {
              Zero,
              ConstantInt::get(IdxType, i),
            };
            auto* Ptr = IC.Builder.CreateInBoundsGEP(AT, Addr, makeArrayRef(Indices),
                Name + ".elt");
            auto* L = IC.Builder.CreateAlignedLoad(Ptr, MinAlign(Align, Offset),
                Name + ".unpack");
            AAMDNodes AAMD;
            LI.getAAMetadata(AAMD);
            L->setAAMetadata(AAMD);
            V = IC.Builder.CreateInsertValue(V, L, i);
            Offset += EltSize;
        }

        V->setName(Name);
        return IC.replaceInstUsesWith(LI, V);
    }

    return nullptr;
}

// If we can determine that all possible objects pointed to by the provided
// pointer value are, not only dereferenceable, but also definitively less than
// or equal to the provided maximum size, then return true. Otherwise, return
// false (constant global values and allocas fall into this category).
//
// FIXME: This should probably live in ValueTracking (or similar).
static bool isObjectSizeLessThanOrEq(Value* V, uint64_t MaxSize,
    const DataLayout& DL) {
    SmallPtrSet<Value*, 4> Visited;
    SmallVector<Value*, 4> Worklist(1, V);

    do {
        Value* P = Worklist.pop_back_val();
        P = P->stripPointerCasts();

        if (!Visited.insert(P).second)
            continue;

        if (SelectInst * SI = dyn_cast<SelectInst>(P)) {
            Worklist.push_back(SI->getTrueValue());
            Worklist.push_back(SI->getFalseValue());
            continue;
        }

        if (PHINode * PN = dyn_cast<PHINode>(P)) {
            for (Value* IncValue : PN->incoming_values())
                Worklist.push_back(IncValue);
            continue;
        }

        if (GlobalAlias * GA = dyn_cast<GlobalAlias>(P)) {
            if (GA->isInterposable())
                return false;
            Worklist.push_back(GA->getAliasee());
            continue;
        }

        // If we know how big this object is, and it is less than MaxSize, continue
        // searching. Otherwise, return false.
        if (AllocaInst * AI = dyn_cast<AllocaInst>(P)) {
            if (!AI->getAllocatedType()->isSized())
                return false;

            ConstantInt* CS = dyn_cast<ConstantInt>(AI->getArraySize());
            if (!CS)
                return false;

            uint64_t TypeSize = DL.getTypeAllocSize(AI->getAllocatedType());
            // Make sure that, even if the multiplication below would wrap as an
            // uint64_t, we still do the right thing.
            if ((CS->getValue().zextOrSelf(128) * APInt(128, TypeSize)).ugt(MaxSize))
                return false;
            continue;
        }

        if (GlobalVariable * GV = dyn_cast<GlobalVariable>(P)) {
            if (!GV->hasDefinitiveInitializer() || !GV->isConstant())
                return false;

            uint64_t InitSize = DL.getTypeAllocSize(GV->getValueType());
            if (InitSize > MaxSize)
                return false;
            continue;
        }

        return false;
    } while (!Worklist.empty());

    return true;
}

// If we're indexing into an object of a known size, and the outer index is
// not a constant, but having any value but zero would lead to undefined
// behavior, replace it with zero.
//
// For example, if we have:
// @f.a = private unnamed_addr constant [1 x i32] [i32 12], align 4
// ...
// %arrayidx = getelementptr inbounds [1 x i32]* @f.a, i64 0, i64 %x
// ... = load i32* %arrayidx, align 4
// Then we know that we can replace %x in the GEP with i64 0.
//
// FIXME: We could fold any GEP index to zero that would cause UB if it were
// not zero. Currently, we only handle the first such index. Also, we could
// also search through non-zero constant indices if we kept track of the
// offsets those indices implied.
static bool canReplaceGEPIdxWithZero(InstCombiner& IC, GetElementPtrInst* GEPI,
    Instruction* MemI, unsigned& Idx) {
    if (GEPI->getNumOperands() < 2)
        return false;

    // Find the first non-zero index of a GEP. If all indices are zero, return
    // one past the last index.
    auto FirstNZIdx = [](const GetElementPtrInst* GEPI) {
        unsigned I = 1;
        for (unsigned IE = GEPI->getNumOperands(); I != IE; ++I) {
            Value* V = GEPI->getOperand(I);
            if (const ConstantInt * CI = dyn_cast<ConstantInt>(V))
                if (CI->isZero())
                    continue;

            break;
        }

        return I;
    };

    // Skip through initial 'zero' indices, and find the corresponding pointer
    // type. See if the next index is not a constant.
    Idx = FirstNZIdx(GEPI);
    if (Idx == GEPI->getNumOperands())
        return false;
    if (isa<Constant>(GEPI->getOperand(Idx)))
        return false;

    SmallVector<Value*, 4> Ops(GEPI->idx_begin(), GEPI->idx_begin() + Idx);
    Type* AllocTy =
        GetElementPtrInst::getIndexedType(GEPI->getSourceElementType(), Ops);
    if (!AllocTy || !AllocTy->isSized())
        return false;
    const DataLayout& DL = IC.getDataLayout();
    uint64_t TyAllocSize = DL.getTypeAllocSize(AllocTy);

    // If there are more indices after the one we might replace with a zero, make
    // sure they're all non-negative. If any of them are negative, the overall
    // address being computed might be before the base address determined by the
    // first non-zero index.
    auto IsAllNonNegative = [&]() {
        for (unsigned i = Idx + 1, e = GEPI->getNumOperands(); i != e; ++i) {
            KnownBits Known = IC.computeKnownBits(GEPI->getOperand(i), 0, MemI);
            if (Known.isNonNegative())
                continue;
            return false;
        }

        return true;
    };

    // FIXME: If the GEP is not inbounds, and there are extra indices after the
    // one we'll replace, those could cause the address computation to wrap
    // (rendering the IsAllNonNegative() check below insufficient). We can do
    // better, ignoring zero indices (and other indices we can prove small
    // enough not to wrap).
    if (Idx + 1 != GEPI->getNumOperands() && !GEPI->isInBounds())
        return false;

    // Note that isObjectSizeLessThanOrEq will return true only if the pointer is
    // also known to be dereferenceable.
    return isObjectSizeLessThanOrEq(GEPI->getOperand(0), TyAllocSize, DL) &&
        IsAllNonNegative();
}

// If we're indexing into an object with a variable index for the memory
// access, but the object has only one element, we can assume that the index
// will always be zero. If we replace the GEP, return it.
template <typename T>
static Instruction* replaceGEPIdxWithZero(InstCombiner& IC, Value* Ptr,
    T& MemI) {
    if (GetElementPtrInst * GEPI = dyn_cast<GetElementPtrInst>(Ptr)) {
        unsigned Idx;
        if (canReplaceGEPIdxWithZero(IC, GEPI, &MemI, Idx)) {
            Instruction* NewGEPI = GEPI->clone();
            NewGEPI->setOperand(Idx,
                ConstantInt::get(GEPI->getOperand(Idx)->getType(), 0));
            NewGEPI->insertBefore(GEPI);
            MemI.setOperand(MemI.getPointerOperandIndex(), NewGEPI);
            return NewGEPI;
        }
    }

    return nullptr;
}

static bool canSimplifyNullStoreOrGEP(StoreInst& SI) {
    if (NullPointerIsDefined(SI.getFunction(), SI.getPointerAddressSpace()))
        return false;

    auto* Ptr = SI.getPointerOperand();
    if (GetElementPtrInst * GEPI = dyn_cast<GetElementPtrInst>(Ptr))
        Ptr = GEPI->getOperand(0);
    return (isa<ConstantPointerNull>(Ptr) &&
        !NullPointerIsDefined(SI.getFunction(), SI.getPointerAddressSpace()));
}

static bool canSimplifyNullLoadOrGEP(LoadInst& LI, Value* Op) {
    if (GetElementPtrInst * GEPI = dyn_cast<GetElementPtrInst>(Op)) {
        const Value* GEPI0 = GEPI->getOperand(0);
        if (isa<ConstantPointerNull>(GEPI0) &&
            !NullPointerIsDefined(LI.getFunction(), GEPI->getPointerAddressSpace()))
            return true;
    }
    if (isa<UndefValue>(Op) ||
        (isa<ConstantPointerNull>(Op) &&
            !NullPointerIsDefined(LI.getFunction(), LI.getPointerAddressSpace())))
        return true;
    return false;
}

Instruction* InstCombiner::visitLoadInst(LoadInst& LI) {
    Value* Op = LI.getOperand(0);

    // Try to canonicalize the loaded type.
    if (Instruction * Res = combineLoadToOperationType(*this, LI))
        return Res;

    // Attempt to improve the alignment.
    unsigned KnownAlign = getOrEnforceKnownAlignment(
        Op, DL.getPrefTypeAlignment(LI.getType()), DL, &LI, &AC, &DT);
    unsigned LoadAlign = LI.getAlignment();
    unsigned EffectiveLoadAlign =
        LoadAlign != 0 ? LoadAlign : DL.getABITypeAlignment(LI.getType());

    if (KnownAlign > EffectiveLoadAlign)
        LI.setAlignment(KnownAlign);
    else if (LoadAlign == 0)
        LI.setAlignment(EffectiveLoadAlign);

    // Replace GEP indices if possible.
    if (Instruction * NewGEPI = replaceGEPIdxWithZero(*this, Op, LI)) {
        Worklist.Add(NewGEPI);
        return &LI;
    }

    if (Instruction * Res = unpackLoadToAggregate(*this, LI))
        return Res;

    // Do really simple store-to-load forwarding and load CSE, to catch cases
    // where there are several consecutive memory accesses to the same location,
    // separated by a few arithmetic operations.
    BasicBlock::iterator BBI(LI);
    bool IsLoadCSE = false;
    if (Value * AvailableVal = FindAvailableLoadedValue(
        &LI, LI.getParent(), BBI, DefMaxInstsToScan, AA, &IsLoadCSE)) {
        if (IsLoadCSE)
            combineMetadataForCSE(cast<LoadInst>(AvailableVal), &LI);

        return replaceInstUsesWith(
            LI, Builder.CreateBitOrPointerCast(AvailableVal, LI.getType(),
                LI.getName() + ".cast"));
    }

    // None of the following transforms are legal for volatile/ordered atomic
    // loads.  Most of them do apply for unordered atomics.
    if (!LI.isUnordered()) return nullptr;

    // load(gep null, ...) -> unreachable
    // load null/undef -> unreachable
    // TODO: Consider a target hook for valid address spaces for this xforms.
    if (canSimplifyNullLoadOrGEP(LI, Op)) {
        // Insert a new store to null instruction before the load to indicate
        // that this code is not reachable.  We do this instead of inserting
        // an unreachable instruction directly because we cannot modify the
        // CFG.
        StoreInst* SI = new StoreInst(UndefValue::get(LI.getType()),
            Constant::getNullValue(Op->getType()), &LI);
        SI->setDebugLoc(LI.getDebugLoc());
        return replaceInstUsesWith(LI, UndefValue::get(LI.getType()));
    }

    if (Op->hasOneUse()) {
        // Change select and PHI nodes to select values instead of addresses: this
        // helps alias analysis out a lot, allows many others simplifications, and
        // exposes redundancy in the code.
        //
        // Note that we cannot do the transformation unless we know that the
        // introduced loads cannot trap!  Something like this is valid as long as
        // the condition is always false: load (select bool %C, int* null, int* %G),
        // but it would not be valid if we transformed it to load from null
        // unconditionally.
        //
        if (SelectInst * SI = dyn_cast<SelectInst>(Op)) {
            // load (select (Cond, &V1, &V2))  --> select(Cond, load &V1, load &V2).
            unsigned Align = LI.getAlignment();
            if (isSafeToLoadUnconditionally(SI->getOperand(1), Align, DL, SI) &&
                isSafeToLoadUnconditionally(SI->getOperand(2), Align, DL, SI)) {
                LoadInst* V1 = Builder.CreateLoad(SI->getOperand(1),
                    SI->getOperand(1)->getName() + ".val");
                LoadInst* V2 = Builder.CreateLoad(SI->getOperand(2),
                    SI->getOperand(2)->getName() + ".val");
                IGC_ASSERT_MESSAGE(LI.isUnordered(), "implied by above");
                V1->setAlignment(Align);
                V1->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
                V2->setAlignment(Align);
                V2->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
                return SelectInst::Create(SI->getCondition(), V1, V2);
            }

            // load (select (cond, null, P)) -> load P
            if (isa<ConstantPointerNull>(SI->getOperand(1)) &&
                !NullPointerIsDefined(SI->getFunction(),
                    LI.getPointerAddressSpace())) {
                LI.setOperand(0, SI->getOperand(2));
                return &LI;
            }

            // load (select (cond, P, null)) -> load P
            if (isa<ConstantPointerNull>(SI->getOperand(2)) &&
                !NullPointerIsDefined(SI->getFunction(),
                    LI.getPointerAddressSpace())) {
                LI.setOperand(0, SI->getOperand(1));
                return &LI;
            }
        }
    }
    return nullptr;
}

/// Look for extractelement/insertvalue sequence that acts like a bitcast.
///
/// \returns underlying value that was "cast", or nullptr otherwise.
///
/// For example, if we have:
///
///     %E0 = extractelement <2 x double> %U, i32 0
///     %V0 = insertvalue [2 x double] undef, double %E0, 0
///     %E1 = extractelement <2 x double> %U, i32 1
///     %V1 = insertvalue [2 x double] %V0, double %E1, 1
///
/// and the layout of a <2 x double> is isomorphic to a [2 x double],
/// then %V1 can be safely approximated by a conceptual "bitcast" of %U.
/// Note that %U may contain non-undef values where %V1 has undef.
static Value* likeBitCastFromVector(InstCombiner& IC, Value* V) {
    Value* U = nullptr;
    while (auto * IV = dyn_cast<InsertValueInst>(V)) {
        auto* E = dyn_cast<ExtractElementInst>(IV->getInsertedValueOperand());
        if (!E)
            return nullptr;
        auto* W = E->getVectorOperand();
        if (!U)
            U = W;
        else if (U != W)
            return nullptr;
        auto* CI = dyn_cast<ConstantInt>(E->getIndexOperand());
        if (!CI || IV->getNumIndices() != 1 || CI->getZExtValue() != *IV->idx_begin())
            return nullptr;
        V = IV->getAggregateOperand();
    }
    if (!isa<UndefValue>(V) || !U)
        return nullptr;

    auto* UT = cast<VectorType>(U->getType());
    auto* VT = V->getType();
    // Check that types UT and VT are bitwise isomorphic.
    const auto& DL = IC.getDataLayout();
    if (DL.getTypeStoreSizeInBits(UT) != DL.getTypeStoreSizeInBits(VT)) {
        return nullptr;
    }
    if (auto * AT = dyn_cast<ArrayType>(VT)) {
        if (AT->getNumElements() != UT->getNumElements())
            return nullptr;
    }
    else {
        auto* ST = cast<StructType>(VT);
        if (ST->getNumElements() != UT->getNumElements())
            return nullptr;
        for (const auto* EltT : ST->elements()) {
            if (EltT != UT->getElementType())
                return nullptr;
        }
    }
    return U;
}

/// Combine stores to match the type of value being stored.
///
/// The core idea here is that the memory does not have any intrinsic type and
/// where we can we should match the type of a store to the type of value being
/// stored.
///
/// However, this routine must never change the width of a store or the number of
/// stores as that would introduce a semantic change. This combine is expected to
/// be a semantic no-op which just allows stores to more closely model the types
/// of their incoming values.
///
/// Currently, we also refuse to change the precise type used for an atomic or
/// volatile store. This is debatable, and might be reasonable to change later.
/// However, it is risky in case some backend or other part of LLVM is relying
/// on the exact type stored to select appropriate atomic operations.
///
/// \returns true if the store was successfully combined away. This indicates
/// the caller must erase the store instruction. We have to let the caller erase
/// the store instruction as otherwise there is no way to signal whether it was
/// combined or not: IC.EraseInstFromFunction returns a null pointer.
static bool combineStoreToValueType(InstCombiner& IC, StoreInst& SI) {
    // FIXME: We could probably with some care handle both volatile and ordered
    // atomic stores here but it isn't clear that this is important.
    if (!SI.isUnordered())
        return false;

    // swifterror values can't be bitcasted.
    if (SI.getPointerOperand()->isSwiftError())
        return false;

    Value* V = SI.getValueOperand();

    // Fold away bit casts of the stored value by storing the original type.
    if (auto * BC = dyn_cast<BitCastInst>(V)) {
        V = BC->getOperand(0);
        if (!SI.isAtomic() || isSupportedAtomicType(V->getType())) {
            combineStoreToNewValue(IC, SI, V);
            return true;
        }
    }

    if (Value * U = likeBitCastFromVector(IC, V))
        if (!SI.isAtomic() || isSupportedAtomicType(U->getType())) {
            combineStoreToNewValue(IC, SI, U);
            return true;
        }

    // FIXME: We should also canonicalize stores of vectors when their elements
    // are cast to other types.
    return false;
}

static bool unpackStoreToAggregate(InstCombiner& IC, StoreInst& SI) {
    // FIXME: We could probably with some care handle both volatile and atomic
    // stores here but it isn't clear that this is important.
    if (!SI.isSimple())
        return false;

    Value* V = SI.getValueOperand();
    Type* T = V->getType();

    if (!T->isAggregateType())
        return false;

    if (auto * ST = dyn_cast<StructType>(T)) {
        // If the struct only have one element, we unpack.
        unsigned Count = ST->getNumElements();
        if (Count == 1) {
            V = IC.Builder.CreateExtractValue(V, 0);
            combineStoreToNewValue(IC, SI, V);
            return true;
        }

        // We don't want to break loads with padding here as we'd loose
        // the knowledge that padding exists for the rest of the pipeline.
        const DataLayout& DL = IC.getDataLayout();
        auto* SL = DL.getStructLayout(ST);
        if (SL->hasPadding())
            return false;

        auto Align = SI.getAlignment();
        if (!Align)
            Align = DL.getABITypeAlignment(ST);

        SmallString<16> EltName = V->getName();
        EltName += ".elt";
        auto* Addr = SI.getPointerOperand();
        SmallString<16> AddrName = Addr->getName();
        AddrName += ".repack";

        auto* IdxType = Type::getInt32Ty(ST->getContext());
        auto* Zero = ConstantInt::get(IdxType, 0);
        for (unsigned i = 0; i < Count; i++) {
            Value* Indices[2] = {
              Zero,
              ConstantInt::get(IdxType, i),
            };
            auto* Ptr = IC.Builder.CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices),
                AddrName);
            auto* Val = IC.Builder.CreateExtractValue(V, i, EltName);
            auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
            llvm::Instruction* NS = IC.Builder.CreateAlignedStore(Val, Ptr, EltAlign);
            AAMDNodes AAMD;
            SI.getAAMetadata(AAMD);
            NS->setAAMetadata(AAMD);
        }

        return true;
    }

    if (auto * AT = dyn_cast<ArrayType>(T)) {
        // If the array only have one element, we unpack.
        auto NumElements = AT->getNumElements();
        if (NumElements == 1) {
            V = IC.Builder.CreateExtractValue(V, 0);
            combineStoreToNewValue(IC, SI, V);
            return true;
        }

        // Bail out if the array is too large. Ideally we would like to optimize
        // arrays of arbitrary size but this has a terrible impact on compile time.
        // The threshold here is chosen arbitrarily, maybe needs a little bit of
        // tuning.
        if (NumElements > IC.MaxArraySizeForCombine)
            return false;

        const DataLayout& DL = IC.getDataLayout();
        auto EltSize = DL.getTypeAllocSize(AT->getElementType());
        auto Align = SI.getAlignment();
        if (!Align)
            Align = DL.getABITypeAlignment(T);

        SmallString<16> EltName = V->getName();
        EltName += ".elt";
        auto* Addr = SI.getPointerOperand();
        SmallString<16> AddrName = Addr->getName();
        AddrName += ".repack";

        auto* IdxType = Type::getInt64Ty(T->getContext());
        auto* Zero = ConstantInt::get(IdxType, 0);

        uint64_t Offset = 0;
        for (uint64_t i = 0; i < NumElements; i++) {
            Value* Indices[2] = {
              Zero,
              ConstantInt::get(IdxType, i),
            };
            auto* Ptr = IC.Builder.CreateInBoundsGEP(AT, Addr, makeArrayRef(Indices),
                AddrName);
            auto* Val = IC.Builder.CreateExtractValue(V, i, EltName);
            auto EltAlign = MinAlign(Align, Offset);
            Instruction* NS = IC.Builder.CreateAlignedStore(Val, Ptr, EltAlign);
            AAMDNodes AAMD;
            SI.getAAMetadata(AAMD);
            NS->setAAMetadata(AAMD);
            Offset += EltSize;
        }

        return true;
    }

    return false;
}

/// equivalentAddressValues - Test if A and B will obviously have the same
/// value. This includes recognizing that %t0 and %t1 will have the same
/// value in code like this:
///   %t0 = getelementptr \@a, 0, 3
///   store i32 0, i32* %t0
///   %t1 = getelementptr \@a, 0, 3
///   %t2 = load i32* %t1
///
static bool equivalentAddressValues(Value* A, Value* B) {
    // Test if the values are trivially equivalent.
    if (A == B) return true;

    // Test if the values come form identical arithmetic instructions.
    // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
    // its only used to compare two uses within the same basic block, which
    // means that they'll always either have the same value or one of them
    // will have an undefined value.
    if (isa<BinaryOperator>(A) ||
        isa<CastInst>(A) ||
        isa<PHINode>(A) ||
        isa<GetElementPtrInst>(A))
        if (Instruction * BI = dyn_cast<Instruction>(B))
            if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
                return true;

    // Otherwise they may not be equivalent.
    return false;
}

/// Converts store (bitcast (load (bitcast (select ...)))) to
/// store (load (select ...)), where select is minmax:
/// select ((cmp load V1, load V2), V1, V2).
static bool removeBitcastsFromLoadStoreOnMinMax(InstCombiner& IC,
    StoreInst& SI) {
    // bitcast?
    if (!match(SI.getPointerOperand(), m_BitCast(m_Value())))
        return false;
    // load? integer?
    Value* LoadAddr;
    if (!match(SI.getValueOperand(), m_Load(m_BitCast(m_Value(LoadAddr)))))
        return false;
    auto* LI = cast<LoadInst>(SI.getValueOperand());
    if (!LI->getType()->isIntegerTy())
        return false;
    if (!isMinMaxWithLoads(LoadAddr))
        return false;

    if (!all_of(LI->users(), [LI, LoadAddr](User* U) {
        auto* SI = dyn_cast<StoreInst>(U);
        return SI && SI->getPointerOperand() != LI &&
            peekThroughBitcast(SI->getPointerOperand()) != LoadAddr &&
            !SI->getPointerOperand()->isSwiftError();
    }))
        return false;

    IC.Builder.SetInsertPoint(LI);
    LoadInst* NewLI = combineLoadToNewType(
        IC, *LI, LoadAddr->getType()->getPointerElementType());
    // Replace all the stores with stores of the newly loaded value.
    for (auto* UI : LI->users()) {
        auto* USI = cast<StoreInst>(UI);
        IC.Builder.SetInsertPoint(USI);
        combineStoreToNewValue(IC, *USI, NewLI);
    }
    IC.replaceInstUsesWith(*LI, UndefValue::get(LI->getType()));
    IC.eraseInstFromFunction(*LI);
    return true;
}

Instruction* InstCombiner::visitStoreInst(StoreInst& SI) {
    Value* Val = SI.getOperand(0);
    Value* Ptr = SI.getOperand(1);

    // Try to canonicalize the stored type.
    if (combineStoreToValueType(*this, SI))
        return eraseInstFromFunction(SI);

    // Attempt to improve the alignment.
    unsigned KnownAlign = getOrEnforceKnownAlignment(
        Ptr, DL.getPrefTypeAlignment(Val->getType()), DL, &SI, &AC, &DT);
    unsigned StoreAlign = SI.getAlignment();
    unsigned EffectiveStoreAlign =
        StoreAlign != 0 ? StoreAlign : DL.getABITypeAlignment(Val->getType());

    if (KnownAlign > EffectiveStoreAlign)
        SI.setAlignment(KnownAlign);
    else if (StoreAlign == 0)
        SI.setAlignment(EffectiveStoreAlign);

    // Try to canonicalize the stored type.
    if (unpackStoreToAggregate(*this, SI))
        return eraseInstFromFunction(SI);

    if (removeBitcastsFromLoadStoreOnMinMax(*this, SI))
        return eraseInstFromFunction(SI);

    // Replace GEP indices if possible.
    if (Instruction * NewGEPI = replaceGEPIdxWithZero(*this, Ptr, SI)) {
        Worklist.Add(NewGEPI);
        return &SI;
    }

    // Don't hack volatile/ordered stores.
    // FIXME: Some bits are legal for ordered atomic stores; needs refactoring.
    if (!SI.isUnordered()) return nullptr;

    // If the RHS is an alloca with a single use, zapify the store, making the
    // alloca dead.
    if (Ptr->hasOneUse()) {
        if (isa<AllocaInst>(Ptr))
            return eraseInstFromFunction(SI);
        if (GetElementPtrInst * GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
            if (isa<AllocaInst>(GEP->getOperand(0))) {
                if (GEP->getOperand(0)->hasOneUse())
                    return eraseInstFromFunction(SI);
            }
        }
    }

    // Do really simple DSE, to catch cases where there are several consecutive
    // stores to the same location, separated by a few arithmetic operations. This
    // situation often occurs with bitfield accesses.
    BasicBlock::iterator BBI(SI);
    for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
        --ScanInsts) {
        --BBI;
        // Don't count debug info directives, lest they affect codegen,
        // and we skip pointer-to-pointer bitcasts, which are NOPs.
        if (isa<DbgInfoIntrinsic>(BBI) ||
            (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
            ScanInsts++;
            continue;
        }

        if (StoreInst * PrevSI = dyn_cast<StoreInst>(BBI)) {
            // Prev store isn't volatile, and stores to the same location?
            if (PrevSI->isUnordered() && equivalentAddressValues(PrevSI->getOperand(1),
                SI.getOperand(1))) {
                ++NumDeadStore;
                ++BBI;
                eraseInstFromFunction(*PrevSI);
                continue;
            }
            break;
        }

        // If this is a load, we have to stop.  However, if the loaded value is from
        // the pointer we're loading and is producing the pointer we're storing,
        // then *this* store is dead (X = load P; store X -> P).
        if (LoadInst * LI = dyn_cast<LoadInst>(BBI)) {
            if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr)) {
                IGC_ASSERT_MESSAGE(SI.isUnordered(), "can't eliminate ordering operation");
                return eraseInstFromFunction(SI);
            }

            // Otherwise, this is a load from some other location.  Stores before it
            // may not be dead.
            break;
        }

        // Don't skip over loads, throws or things that can modify memory.
        if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory() || BBI->mayThrow())
            break;
    }

    // store X, null    -> turns into 'unreachable' in SimplifyCFG
    // store X, GEP(null, Y) -> turns into 'unreachable' in SimplifyCFG
    if (canSimplifyNullStoreOrGEP(SI)) {
        if (!isa<UndefValue>(Val)) {
            SI.setOperand(0, UndefValue::get(Val->getType()));
            if (Instruction * U = dyn_cast<Instruction>(Val))
                Worklist.Add(U);  // Dropped a use.
        }
        return nullptr;  // Do not modify these!
    }

    // store undef, Ptr -> noop
    if (isa<UndefValue>(Val))
        return eraseInstFromFunction(SI);

    // If this store is the last instruction in the basic block (possibly
    // excepting debug info instructions), and if the block ends with an
    // unconditional branch, try to move it to the successor block.
    BBI = SI.getIterator();
    do {
        ++BBI;
    } while (isa<DbgInfoIntrinsic>(BBI) ||
        (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
    if (BranchInst * BI = dyn_cast<BranchInst>(BBI))
        if (BI->isUnconditional())
            if (SimplifyStoreAtEndOfBlock(SI))
                return nullptr;  // xform done!

    return nullptr;
}

/// SimplifyStoreAtEndOfBlock - Turn things like:
///   if () { *P = v1; } else { *P = v2 }
/// into a phi node with a store in the successor.
///
/// Simplify things like:
///   *P = v1; if () { *P = v2; }
/// into a phi node with a store in the successor.
///
bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst& SI) {
    IGC_ASSERT_MESSAGE(SI.isUnordered(), "this code has not been auditted for volatile or ordered store case");

    BasicBlock* StoreBB = SI.getParent();

    // Check to see if the successor block has exactly two incoming edges.  If
    // so, see if the other predecessor contains a store to the same location.
    // if so, insert a PHI node (if needed) and move the stores down.
    BasicBlock* DestBB = StoreBB->getTerminator()->getSuccessor(0);

    // Determine whether Dest has exactly two predecessors and, if so, compute
    // the other predecessor.
    pred_iterator PI = pred_begin(DestBB);
    BasicBlock* P = *PI;
    BasicBlock* OtherBB = nullptr;

    if (P != StoreBB)
        OtherBB = P;

    if (++PI == pred_end(DestBB))
        return false;

    P = *PI;
    if (P != StoreBB) {
        if (OtherBB)
            return false;
        OtherBB = P;
    }
    if (++PI != pred_end(DestBB))
        return false;

    // Bail out if all the relevant blocks aren't distinct (this can happen,
    // for example, if SI is in an infinite loop)
    if (StoreBB == DestBB || OtherBB == DestBB)
        return false;

    // Verify that the other block ends in a branch and is not otherwise empty.
    BasicBlock::iterator BBI(OtherBB->getTerminator());
    BranchInst* OtherBr = dyn_cast<BranchInst>(BBI);
    if (!OtherBr || BBI == OtherBB->begin())
        return false;

    // If the other block ends in an unconditional branch, check for the 'if then
    // else' case.  there is an instruction before the branch.
    StoreInst* OtherStore = nullptr;
    if (OtherBr->isUnconditional()) {
        --BBI;
        // Skip over debugging info.
        while (isa<DbgInfoIntrinsic>(BBI) ||
            (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
            if (BBI == OtherBB->begin())
                return false;
            --BBI;
        }
        // If this isn't a store, isn't a store to the same location, or is not the
        // right kind of store, bail out.
        OtherStore = dyn_cast<StoreInst>(BBI);
        if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
            !SI.isSameOperationAs(OtherStore))
            return false;
    }
    else {
        // Otherwise, the other block ended with a conditional branch. If one of the
        // destinations is StoreBB, then we have the if/then case.
        if (OtherBr->getSuccessor(0) != StoreBB &&
            OtherBr->getSuccessor(1) != StoreBB)
            return false;

        // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
        // if/then triangle.  See if there is a store to the same ptr as SI that
        // lives in OtherBB.
        for (;; --BBI) {
            // Check to see if we find the matching store.
            if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
                if (OtherStore->getOperand(1) != SI.getOperand(1) ||
                    !SI.isSameOperationAs(OtherStore))
                    return false;
                break;
            }
            // If we find something that may be using or overwriting the stored
            // value, or if we run out of instructions, we can't do the xform.
            if (BBI->mayReadFromMemory() || BBI->mayThrow() ||
                BBI->mayWriteToMemory() || BBI == OtherBB->begin())
                return false;
        }

        // In order to eliminate the store in OtherBr, we have to
        // make sure nothing reads or overwrites the stored value in
        // StoreBB.
        for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
            // FIXME: This should really be AA driven.
            if (I->mayReadFromMemory() || I->mayThrow() || I->mayWriteToMemory())
                return false;
        }
    }

    // Insert a PHI node now if we need it.
    Value* MergedVal = OtherStore->getOperand(0);
    if (MergedVal != SI.getOperand(0)) {
        PHINode* PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
        PN->addIncoming(SI.getOperand(0), SI.getParent());
        PN->addIncoming(OtherStore->getOperand(0), OtherBB);
        MergedVal = InsertNewInstBefore(PN, DestBB->front());
    }

    // Advance to a place where it is safe to insert the new store and
    // insert it.
    BBI = DestBB->getFirstInsertionPt();
    StoreInst* NewSI = new StoreInst(MergedVal, SI.getOperand(1),
        SI.isVolatile(),
        SI.getAlignment(),
        SI.getOrdering(),
        SI.getSyncScopeID());
    InsertNewInstBefore(NewSI, *BBI);
    // The debug locations of the original instructions might differ; merge them.
    NewSI->applyMergedLocation(SI.getDebugLoc(), OtherStore->getDebugLoc());

    // If the two stores had AA tags, merge them.
    AAMDNodes AATags;
    SI.getAAMetadata(AATags);
    if (AATags) {
        OtherStore->getAAMetadata(AATags, /* Merge = */ true);
        NewSI->setAAMetadata(AATags);
    }

    // Nuke the old stores.
    eraseInstFromFunction(SI);
    eraseInstFromFunction(*OtherStore);
    return true;
}
#include "common/LLVMWarningsPop.hpp"