void GraphBuilder::visitPtrToIntInst(PtrToIntInst& I) {
DSNode* N = getValueDest(I.getOperand(0)).getNode();
if(I.hasOneUse()) {
if(isa<ICmpInst>(*(I.use_begin()))) {
NumBoringIntToPtr++;
return;
}
}
if(I.hasOneUse()) {
Value *V = dyn_cast<Value>(*(I.use_begin()));
DenseSet<Value *> Seen;
while(V && V->hasOneUse() &&
Seen.insert(V).second) {
if(isa<LoadInst>(V))
break;
if(isa<StoreInst>(V))
break;
if(isa<CallInst>(V))
break;
V = dyn_cast<Value>(*(V->use_begin()));
}
if(isa<BranchInst>(V)){
NumBoringIntToPtr++;
return;
}
}
if(N)
N->setPtrToIntMarker();
}
/// Try to find redundant insertvalue instructions, like the following ones:
/// %0 = insertvalue { i8, i32 } undef, i8 %x, 0
/// %1 = insertvalue { i8, i32 } %0, i8 %y, 0
/// Here the second instruction inserts values at the same indices, as the
/// first one, making the first one redundant.
/// It should be transformed to:
/// %0 = insertvalue { i8, i32 } undef, i8 %y, 0
Instruction *InstCombiner::visitInsertValueInst(InsertValueInst &I) {
bool IsRedundant = false;
ArrayRef<unsigned int> FirstIndices = I.getIndices();
// If there is a chain of insertvalue instructions (each of them except the
// last one has only one use and it's another insertvalue insn from this
// chain), check if any of the 'children' uses the same indices as the first
// instruction. In this case, the first one is redundant.
Value *V = &I;
unsigned Depth = 0;
while (V->hasOneUse() && Depth < 10) {
User *U = V->user_back();
auto UserInsInst = dyn_cast<InsertValueInst>(U);
if (!UserInsInst || U->getOperand(0) != V)
break;
if (UserInsInst->getIndices() == FirstIndices) {
IsRedundant = true;
break;
}
V = UserInsInst;
Depth++;
}
if (IsRedundant)
return replaceInstUsesWith(I, I.getOperand(0));
return nullptr;
}
bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
BasicBlock *DefBB = I->getParent();
// If the result of a {s|z}ext and its source are both live out, rewrite all
// other uses of the source with result of extension.
Value *Src = I->getOperand(0);
if (Src->hasOneUse())
return false;
// Only do this xform if truncating is free.
if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
return false;
// Only safe to perform the optimization if the source is also defined in
// this block.
if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
return false;
bool DefIsLiveOut = false;
for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
UI != E; ++UI) {
Instruction *User = cast<Instruction>(*UI);
// Figure out which BB this ext is used in.
BasicBlock *UserBB = User->getParent();
if (UserBB == DefBB) continue;
DefIsLiveOut = true;
break;
}
if (!DefIsLiveOut)
return false;
// Make sure non of the uses are PHI nodes.
for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
UI != E; ++UI) {
Instruction *User = cast<Instruction>(*UI);
BasicBlock *UserBB = User->getParent();
if (UserBB == DefBB) continue;
// Be conservative. We don't want this xform to end up introducing
// reloads just before load / store instructions.
if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
return false;
}
// InsertedTruncs - Only insert one trunc in each block once.
DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
bool MadeChange = false;
for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
UI != E; ++UI) {
Use &TheUse = UI.getUse();
Instruction *User = cast<Instruction>(*UI);
// Figure out which BB this ext is used in.
BasicBlock *UserBB = User->getParent();
if (UserBB == DefBB) continue;
// Both src and def are live in this block. Rewrite the use.
Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
if (!InsertedTrunc) {
BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
}
// Replace a use of the {s|z}ext source with a use of the result.
TheUse = InsertedTrunc;
++NumExtUses;
MadeChange = true;
}
return MadeChange;
}
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;
}
// None of the following transforms are legal for volatile/atomic loads.
// FIXME: Some of it is okay for atomic loads; needs refactoring.
if (!LI.isSimple()) return nullptr;
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);
AAMDNodes AATags;
if (Value *AvailableVal =
FindAvailableLoadedValue(Op, LI.getParent(), BBI,
DefMaxInstsToScan, AA, &AATags)) {
if (LoadInst *NLI = dyn_cast<LoadInst>(AvailableVal)) {
unsigned KnownIDs[] = {
LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
LLVMContext::MD_noalias, LLVMContext::MD_range,
LLVMContext::MD_invariant_load, LLVMContext::MD_nonnull,
LLVMContext::MD_invariant_group, LLVMContext::MD_align,
LLVMContext::MD_dereferenceable,
LLVMContext::MD_dereferenceable_or_null};
combineMetadata(NLI, &LI, KnownIDs);
};
return ReplaceInstUsesWith(
LI, Builder->CreateBitOrPointerCast(AvailableVal, LI.getType(),
LI.getName() + ".cast"));
}
// load(gep null, ...) -> unreachable
if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
const Value *GEPI0 = GEPI->getOperand(0);
// TODO: Consider a target hook for valid address spaces for this xform.
if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
// 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.
new StoreInst(UndefValue::get(LI.getType()),
Constant::getNullValue(Op->getType()), &LI);
return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
}
}
// load null/undef -> unreachable
// TODO: Consider a target hook for valid address spaces for this xform.
if (isa<UndefValue>(Op) ||
(isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
// 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.
new StoreInst(UndefValue::get(LI.getType()),
Constant::getNullValue(Op->getType()), &LI);
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, SI) &&
isSafeToLoadUnconditionally(SI->getOperand(2), Align, 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");
V1->setAlignment(Align);
V2->setAlignment(Align);
return SelectInst::Create(SI->getCondition(), V1, V2);
}
// load (select (cond, null, P)) -> load P
if (isa<ConstantPointerNull>(SI->getOperand(1)) &&
LI.getPointerAddressSpace() == 0) {
LI.setOperand(0, SI->getOperand(2));
return &LI;
}
// load (select (cond, P, null)) -> load P
if (isa<ConstantPointerNull>(SI->getOperand(2)) &&
LI.getPointerAddressSpace() == 0) {
LI.setOperand(0, SI->getOperand(1));
return &LI;
}
}
}
return nullptr;
}
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);
// 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)) {
assert(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 or things that can modify memory.
if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
break;
}
// store X, null -> turns into 'unreachable' in SimplifyCFG
if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
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);
// The code below needs to be audited and adjusted for unordered atomics
if (!SI.isSimple())
return nullptr;
// 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;
}
Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
bool Changed = SimplifyAssociativeOrCommutative(I);
Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
if (Value *V = SimplifyAddInst(LHS, RHS, I.hasNoSignedWrap(),
I.hasNoUnsignedWrap(), TD))
return ReplaceInstUsesWith(I, V);
// (A*B)+(A*C) -> A*(B+C) etc
if (Value *V = SimplifyUsingDistributiveLaws(I))
return ReplaceInstUsesWith(I, V);
if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
// X + (signbit) --> X ^ signbit
const APInt &Val = CI->getValue();
if (Val.isSignBit())
return BinaryOperator::CreateXor(LHS, RHS);
// See if SimplifyDemandedBits can simplify this. This handles stuff like
// (X & 254)+1 -> (X&254)|1
if (SimplifyDemandedInstructionBits(I))
return &I;
// zext(bool) + C -> bool ? C + 1 : C
if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
if (ZI->getSrcTy()->isIntegerTy(1))
return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
Value *XorLHS = 0; ConstantInt *XorRHS = 0;
if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
const APInt &RHSVal = CI->getValue();
unsigned ExtendAmt = 0;
// If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
// If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
if (XorRHS->getValue() == -RHSVal) {
if (RHSVal.isPowerOf2())
ExtendAmt = TySizeBits - RHSVal.logBase2() - 1;
else if (XorRHS->getValue().isPowerOf2())
ExtendAmt = TySizeBits - XorRHS->getValue().logBase2() - 1;
}
if (ExtendAmt) {
APInt Mask = APInt::getHighBitsSet(TySizeBits, ExtendAmt);
if (!MaskedValueIsZero(XorLHS, Mask))
ExtendAmt = 0;
}
if (ExtendAmt) {
Constant *ShAmt = ConstantInt::get(I.getType(), ExtendAmt);
Value *NewShl = Builder->CreateShl(XorLHS, ShAmt, "sext");
return BinaryOperator::CreateAShr(NewShl, ShAmt);
}
}
}
if (isa<Constant>(RHS) && isa<PHINode>(LHS))
if (Instruction *NV = FoldOpIntoPhi(I))
return NV;
if (I.getType()->isIntegerTy(1))
return BinaryOperator::CreateXor(LHS, RHS);
// X + X --> X << 1
if (LHS == RHS) {
BinaryOperator *New =
BinaryOperator::CreateShl(LHS, ConstantInt::get(I.getType(), 1));
New->setHasNoSignedWrap(I.hasNoSignedWrap());
New->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
return New;
}
// -A + B --> B - A
// -A + -B --> -(A + B)
if (Value *LHSV = dyn_castNegVal(LHS)) {
if (Value *RHSV = dyn_castNegVal(RHS)) {
Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
return BinaryOperator::CreateNeg(NewAdd);
}
return BinaryOperator::CreateSub(RHS, LHSV);
}
// A + -B --> A - B
if (!isa<Constant>(RHS))
if (Value *V = dyn_castNegVal(RHS))
return BinaryOperator::CreateSub(LHS, V);
ConstantInt *C2;
if (Value *X = dyn_castFoldableMul(LHS, C2)) {
if (X == RHS) // X*C + X --> X * (C+1)
return BinaryOperator::CreateMul(RHS, AddOne(C2));
// X*C1 + X*C2 --> X * (C1+C2)
ConstantInt *C1;
if (X == dyn_castFoldableMul(RHS, C1))
return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
}
// X + X*C --> X * (C+1)
if (dyn_castFoldableMul(RHS, C2) == LHS)
return BinaryOperator::CreateMul(LHS, AddOne(C2));
// A+B --> A|B iff A and B have no bits set in common.
if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
APInt LHSKnownOne(IT->getBitWidth(), 0);
APInt LHSKnownZero(IT->getBitWidth(), 0);
ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
if (LHSKnownZero != 0) {
APInt RHSKnownOne(IT->getBitWidth(), 0);
APInt RHSKnownZero(IT->getBitWidth(), 0);
ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
// No bits in common -> bitwise or.
if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
return BinaryOperator::CreateOr(LHS, RHS);
}
}
// W*X + Y*Z --> W * (X+Z) iff W == Y
{
Value *W, *X, *Y, *Z;
if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
if (W != Y) {
if (W == Z) {
std::swap(Y, Z);
} else if (Y == X) {
std::swap(W, X);
} else if (X == Z) {
std::swap(Y, Z);
std::swap(W, X);
}
}
if (W == Y) {
Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName());
return BinaryOperator::CreateMul(W, NewAdd);
}
}
}
if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
Value *X = 0;
if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
return BinaryOperator::CreateSub(SubOne(CRHS), X);
// (X & FF00) + xx00 -> (X+xx00) & FF00
if (LHS->hasOneUse() &&
match(LHS, m_And(m_Value(X), m_ConstantInt(C2))) &&
CRHS->getValue() == (CRHS->getValue() & C2->getValue())) {
// See if all bits from the first bit set in the Add RHS up are included
// in the mask. First, get the rightmost bit.
const APInt &AddRHSV = CRHS->getValue();
// Form a mask of all bits from the lowest bit added through the top.
APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
// See if the and mask includes all of these bits.
APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
if (AddRHSHighBits == AddRHSHighBitsAnd) {
// Okay, the xform is safe. Insert the new add pronto.
Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
return BinaryOperator::CreateAnd(NewAdd, C2);
}
}
// Try to fold constant add into select arguments.
if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
if (Instruction *R = FoldOpIntoSelect(I, SI))
return R;
}
// add (select X 0 (sub n A)) A --> select X A n
{
SelectInst *SI = dyn_cast<SelectInst>(LHS);
Value *A = RHS;
if (!SI) {
SI = dyn_cast<SelectInst>(RHS);
A = LHS;
}
if (SI && SI->hasOneUse()) {
Value *TV = SI->getTrueValue();
Value *FV = SI->getFalseValue();
Value *N;
// Can we fold the add into the argument of the select?
// We check both true and false select arguments for a matching subtract.
if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
// Fold the add into the true select value.
return SelectInst::Create(SI->getCondition(), N, A);
if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
// Fold the add into the false select value.
return SelectInst::Create(SI->getCondition(), A, N);
}
}
// Check for (add (sext x), y), see if we can merge this into an
// integer add followed by a sext.
if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
// (add (sext x), cst) --> (sext (add x, cst'))
if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
Constant *CI =
ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
if (LHSConv->hasOneUse() &&
ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
// Insert the new, smaller add.
Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
CI, "addconv");
return new SExtInst(NewAdd, I.getType());
}
}
// (add (sext x), (sext y)) --> (sext (add int x, y))
if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
// Only do this if x/y have the same type, if at last one of them has a
// single use (so we don't increase the number of sexts), and if the
// integer add will not overflow.
if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
(LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
WillNotOverflowSignedAdd(LHSConv->getOperand(0),
RHSConv->getOperand(0))) {
// Insert the new integer add.
Value *NewAdd = Builder->CreateNSWAdd(LHSConv->getOperand(0),
RHSConv->getOperand(0), "addconv");
return new SExtInst(NewAdd, I.getType());
}
}
}
return Changed ? &I : 0;
}
Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
Value *Op = LI.getOperand(0);
// Attempt to improve the alignment.
if (DL) {
unsigned KnownAlign =
getOrEnforceKnownAlignment(Op, DL->getPrefTypeAlignment(LI.getType()),DL);
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);
}
// load (cast X) --> cast (load X) iff safe.
if (isa<CastInst>(Op))
if (Instruction *Res = InstCombineLoadCast(*this, LI, DL))
return Res;
// None of the following transforms are legal for volatile/atomic loads.
// FIXME: Some of it is okay for atomic loads; needs refactoring.
if (!LI.isSimple()) return nullptr;
// 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;
if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
return ReplaceInstUsesWith(LI, AvailableVal);
// load(gep null, ...) -> unreachable
if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
const Value *GEPI0 = GEPI->getOperand(0);
// TODO: Consider a target hook for valid address spaces for this xform.
if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
// 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.
new StoreInst(UndefValue::get(LI.getType()),
Constant::getNullValue(Op->getType()), &LI);
return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
}
}
// load null/undef -> unreachable
// TODO: Consider a target hook for valid address spaces for this xform.
if (isa<UndefValue>(Op) ||
(isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
// 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.
new StoreInst(UndefValue::get(LI.getType()),
Constant::getNullValue(Op->getType()), &LI);
return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
}
// Instcombine load (constantexpr_cast global) -> cast (load global)
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
if (CE->isCast())
if (Instruction *Res = InstCombineLoadCast(*this, LI, DL))
return Res;
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), SI, Align, DL) &&
isSafeToLoadUnconditionally(SI->getOperand(2), SI, Align, DL)) {
LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
SI->getOperand(1)->getName()+".val");
LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
SI->getOperand(2)->getName()+".val");
V1->setAlignment(Align);
V2->setAlignment(Align);
return SelectInst::Create(SI->getCondition(), V1, V2);
}
// load (select (cond, null, P)) -> load P
if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
if (C->isNullValue()) {
LI.setOperand(0, SI->getOperand(2));
return &LI;
}
// load (select (cond, P, null)) -> load P
if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
if (C->isNullValue()) {
LI.setOperand(0, SI->getOperand(1));
return &LI;
}
}
}
return nullptr;
}
/// If \param [in] BB has more than one predecessor that is a conditional
/// branch, attempt to use parallel and/or for the branch condition. \returns
/// true on success.
///
/// Before:
/// ......
/// %cmp10 = fcmp une float %tmp1, %tmp2
/// br i1 %cmp1, label %if.then, label %lor.rhs
///
/// lor.rhs:
/// ......
/// %cmp11 = fcmp une float %tmp3, %tmp4
/// br i1 %cmp11, label %if.then, label %ifend
///
/// if.end: // the merge block
/// ......
///
/// if.then: // has two predecessors, both of them contains conditional branch.
/// ......
/// br label %if.end;
///
/// After:
/// ......
/// %cmp10 = fcmp une float %tmp1, %tmp2
/// ......
/// %cmp11 = fcmp une float %tmp3, %tmp4
/// %cmp12 = or i1 %cmp10, %cmp11 // parallel-or mode.
/// br i1 %cmp12, label %if.then, label %ifend
///
/// if.end:
/// ......
///
/// if.then:
/// ......
/// br label %if.end;
///
/// Current implementation handles two cases.
/// Case 1: \param BB is on the else-path.
///
/// BB1
/// / |
/// BB2 |
/// / \ |
/// BB3 \ | where, BB1, BB2 contain conditional branches.
/// \ | / BB3 contains unconditional branch.
/// \ | / BB4 corresponds to \param BB which is also the merge.
/// BB => BB4
///
///
/// Corresponding source code:
///
/// if (a == b && c == d)
/// statement; // BB3
///
/// Case 2: \param BB BB is on the then-path.
///
/// BB1
/// / |
/// | BB2
/// \ / | where BB1, BB2 contain conditional branches.
/// BB => BB3 | BB3 contains unconditiona branch and corresponds
/// \ / to \param BB. BB4 is the merge.
/// BB4
///
/// Corresponding source code:
///
/// if (a == b || c == d)
/// statement; // BB3
///
/// In both cases, \param BB is the common successor of conditional branches.
/// In Case 1, \param BB (BB4) has an unconditional branch (BB3) as
/// its predecessor. In Case 2, \param BB (BB3) only has conditional branches
/// as its predecessors.
///
bool FlattenCFGOpt::FlattenParallelAndOr(BasicBlock *BB, IRBuilder<> &Builder,
Pass *P) {
PHINode *PHI = dyn_cast<PHINode>(BB->begin());
if (PHI)
return false; // For simplicity, avoid cases containing PHI nodes.
BasicBlock *LastCondBlock = NULL;
BasicBlock *FirstCondBlock = NULL;
BasicBlock *UnCondBlock = NULL;
int Idx = -1;
// Check predecessors of \param BB.
SmallPtrSet<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
for (SmallPtrSetIterator<BasicBlock *> PI = Preds.begin(), PE = Preds.end();
PI != PE; ++PI) {
BasicBlock *Pred = *PI;
BranchInst *PBI = dyn_cast<BranchInst>(Pred->getTerminator());
// All predecessors should terminate with a branch.
if (!PBI)
return false;
BasicBlock *PP = Pred->getSinglePredecessor();
if (PBI->isUnconditional()) {
// Case 1: Pred (BB3) is an unconditional block, it should
// have a single predecessor (BB2) that is also a predecessor
// of \param BB (BB4) and should not have address-taken.
// There should exist only one such unconditional
// branch among the predecessors.
if (UnCondBlock || !PP || (Preds.count(PP) == 0) ||
Pred->hasAddressTaken())
return false;
UnCondBlock = Pred;
continue;
}
// Only conditional branches are allowed beyond this point.
assert(PBI->isConditional());
// Condition's unique use should be the branch instruction.
Value *PC = PBI->getCondition();
if (!PC || !PC->hasOneUse())
return false;
if (PP && Preds.count(PP)) {
// These are internal condition blocks to be merged from, e.g.,
// BB2 in both cases.
// Should not be address-taken.
if (Pred->hasAddressTaken())
return false;
// Instructions in the internal condition blocks should be safe
// to hoist up.
for (BasicBlock::iterator BI = Pred->begin(), BE = PBI; BI != BE;) {
Instruction *CI = BI++;
if (isa<PHINode>(CI) || !isSafeToSpeculativelyExecute(CI))
return false;
}
} else {
// This is the condition block to be merged into, e.g. BB1 in
// both cases.
if (FirstCondBlock)
return false;
FirstCondBlock = Pred;
}
// Find whether BB is uniformly on the true (or false) path
// for all of its predecessors.
BasicBlock *PS1 = PBI->getSuccessor(0);
BasicBlock *PS2 = PBI->getSuccessor(1);
BasicBlock *PS = (PS1 == BB) ? PS2 : PS1;
int CIdx = (PS1 == BB) ? 0 : 1;
if (Idx == -1)
Idx = CIdx;
else if (CIdx != Idx)
return false;
// PS is the successor which is not BB. Check successors to identify
// the last conditional branch.
if (Preds.count(PS) == 0) {
// Case 2.
LastCondBlock = Pred;
} else {
// Case 1
BranchInst *BPS = dyn_cast<BranchInst>(PS->getTerminator());
if (BPS && BPS->isUnconditional()) {
// Case 1: PS(BB3) should be an unconditional branch.
LastCondBlock = Pred;
}
}
}
if (!FirstCondBlock || !LastCondBlock || (FirstCondBlock == LastCondBlock))
return false;
TerminatorInst *TBB = LastCondBlock->getTerminator();
BasicBlock *PS1 = TBB->getSuccessor(0);
BasicBlock *PS2 = TBB->getSuccessor(1);
BranchInst *PBI1 = dyn_cast<BranchInst>(PS1->getTerminator());
BranchInst *PBI2 = dyn_cast<BranchInst>(PS2->getTerminator());
// If PS1 does not jump into PS2, but PS2 jumps into PS1,
// attempt branch inversion.
if (!PBI1 || !PBI1->isUnconditional() ||
(PS1->getTerminator()->getSuccessor(0) != PS2)) {
// Check whether PS2 jumps into PS1.
if (!PBI2 || !PBI2->isUnconditional() ||
(PS2->getTerminator()->getSuccessor(0) != PS1))
return false;
// Do branch inversion.
BasicBlock *CurrBlock = LastCondBlock;
bool EverChanged = false;
while (1) {
BranchInst *BI = dyn_cast<BranchInst>(CurrBlock->getTerminator());
CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition());
CmpInst::Predicate Predicate = CI->getPredicate();
// Cannonicalize icmp_ne -> icmp_eq, fcmp_one -> fcmp_oeq
if ((Predicate == CmpInst::ICMP_NE) || (Predicate == CmpInst::FCMP_ONE)) {
CI->setPredicate(ICmpInst::getInversePredicate(Predicate));
BI->swapSuccessors();
EverChanged = true;
}
if (CurrBlock == FirstCondBlock)
break;
CurrBlock = CurrBlock->getSinglePredecessor();
}
return EverChanged;
}
// PS1 must have a conditional branch.
if (!PBI1 || !PBI1->isUnconditional())
return false;
// PS2 should not contain PHI node.
PHI = dyn_cast<PHINode>(PS2->begin());
if (PHI)
return false;
// Do the transformation.
BasicBlock *CB;
BranchInst *PBI = dyn_cast<BranchInst>(FirstCondBlock->getTerminator());
bool Iteration = true;
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
Value *PC = PBI->getCondition();
do {
CB = PBI->getSuccessor(1 - Idx);
// Delete the conditional branch.
FirstCondBlock->getInstList().pop_back();
FirstCondBlock->getInstList()
.splice(FirstCondBlock->end(), CB->getInstList());
PBI = cast<BranchInst>(FirstCondBlock->getTerminator());
Value *CC = PBI->getCondition();
// Merge conditions.
Builder.SetInsertPoint(PBI);
Value *NC;
if (Idx == 0)
// Case 2, use parallel or.
NC = Builder.CreateOr(PC, CC);
else
// Case 1, use parallel and.
NC = Builder.CreateAnd(PC, CC);
PBI->replaceUsesOfWith(CC, NC);
PC = NC;
if (CB == LastCondBlock)
Iteration = false;
// Remove internal conditional branches.
CB->dropAllReferences();
// make CB unreachable and let downstream to delete the block.
new UnreachableInst(CB->getContext(), CB);
} while (Iteration);
Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
DEBUG(dbgs() << "Use parallel and/or in:\n" << *FirstCondBlock);
return true;
}
Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
Value *Val = SI.getOperand(0);
Value *Ptr = SI.getOperand(1);
// If the RHS is an alloca with a single use, zapify the store, making the
// alloca dead.
// If the RHS is an alloca with a two uses, the other one being a
// llvm.dbg.declare, zapify the store and the declare, making the
// alloca dead. We must do this to prevent declares from affecting
// codegen.
if (!SI.isVolatile()) {
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);
if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
EraseInstFromFunction(*DI);
return EraseInstFromFunction(SI);
}
}
}
}
if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
EraseInstFromFunction(*DI);
return EraseInstFromFunction(SI);
}
}
// Attempt to improve the alignment.
if (TD) {
unsigned KnownAlign =
GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
unsigned StoreAlign = SI.getAlignment();
unsigned EffectiveStoreAlign = StoreAlign != 0 ? StoreAlign :
TD->getABITypeAlignment(Val->getType());
if (KnownAlign > EffectiveStoreAlign)
SI.setAlignment(KnownAlign);
else if (StoreAlign == 0)
SI.setAlignment(EffectiveStoreAlign);
}
// 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->isVolatile() &&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) &&
!SI.isVolatile())
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 or things that can modify memory.
if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
break;
}
if (SI.isVolatile()) return 0; // Don't hack volatile stores.
// store X, null -> turns into 'unreachable' in SimplifyCFG
if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
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 0; // Do not modify these!
}
// store undef, Ptr -> noop
if (isa<UndefValue>(Val))
return EraseInstFromFunction(SI);
// If the pointer destination is a cast, see if we can fold the cast into the
// source instead.
if (isa<CastInst>(Ptr))
if (Instruction *Res = InstCombineStoreToCast(*this, SI))
return Res;
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
if (CE->isCast())
if (Instruction *Res = InstCombineStoreToCast(*this, SI))
return Res;
// 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;
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 0; // xform done!
return 0;
}
/// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
/// instruction.
bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
SelectInst *SI = cast<SelectInst>(I.getOperand(1));
// div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
int NonNullOperand = -1;
if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
if (ST->isNullValue())
NonNullOperand = 2;
// div/rem X, (Cond ? Y : 0) -> div/rem X, Y
if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
if (ST->isNullValue())
NonNullOperand = 1;
if (NonNullOperand == -1)
return false;
Value *SelectCond = SI->getOperand(0);
// Change the div/rem to use 'Y' instead of the select.
I.setOperand(1, SI->getOperand(NonNullOperand));
// Okay, we know we replace the operand of the div/rem with 'Y' with no
// problem. However, the select, or the condition of the select may have
// multiple uses. Based on our knowledge that the operand must be non-zero,
// propagate the known value for the select into other uses of it, and
// propagate a known value of the condition into its other users.
// If the select and condition only have a single use, don't bother with this,
// early exit.
if (SI->use_empty() && SelectCond->hasOneUse())
return true;
// Scan the current block backward, looking for other uses of SI.
BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
while (BBI != BBFront) {
--BBI;
// If we found a call to a function, we can't assume it will return, so
// information from below it cannot be propagated above it.
if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
break;
// Replace uses of the select or its condition with the known values.
for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
I != E; ++I) {
if (*I == SI) {
*I = SI->getOperand(NonNullOperand);
Worklist.Add(BBI);
} else if (*I == SelectCond) {
*I = NonNullOperand == 1 ? ConstantInt::getTrue(BBI->getContext()) :
ConstantInt::getFalse(BBI->getContext());
Worklist.Add(BBI);
}
}
// If we past the instruction, quit looking for it.
if (&*BBI == SI)
SI = 0;
if (&*BBI == SelectCond)
SelectCond = 0;
// If we ran out of things to eliminate, break out of the loop.
if (SelectCond == 0 && SI == 0)
break;
}
return true;
}
/// LinearizeExprTree - Given an associative binary expression tree, traverse
/// all of the uses putting it into canonical form. This forces a left-linear
/// form of the expression (((a+b)+c)+d), and collects information about the
/// rank of the non-tree operands.
///
/// NOTE: These intentionally destroys the expression tree operands (turning
/// them into undef values) to reduce #uses of the values. This means that the
/// caller MUST use something like RewriteExprTree to put the values back in.
///
void Reassociate::LinearizeExprTree(BinaryOperator *I,
SmallVectorImpl<ValueEntry> &Ops) {
Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
unsigned Opcode = I->getOpcode();
// First step, linearize the expression if it is in ((A+B)+(C+D)) form.
BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
// If this is a multiply expression tree and it contains internal negations,
// transform them into multiplies by -1 so they can be reassociated.
if (I->getOpcode() == Instruction::Mul) {
if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap);
LHSBO = isReassociableOp(LHS, Opcode);
}
if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap);
RHSBO = isReassociableOp(RHS, Opcode);
}
}
if (!LHSBO) {
if (!RHSBO) {
// Neither the LHS or RHS as part of the tree, thus this is a leaf. As
// such, just remember these operands and their rank.
Ops.push_back(ValueEntry(getRank(LHS), LHS));
Ops.push_back(ValueEntry(getRank(RHS), RHS));
// Clear the leaves out.
I->setOperand(0, UndefValue::get(I->getType()));
I->setOperand(1, UndefValue::get(I->getType()));
return;
}
// Turn X+(Y+Z) -> (Y+Z)+X
std::swap(LHSBO, RHSBO);
std::swap(LHS, RHS);
bool Success = !I->swapOperands();
assert(Success && "swapOperands failed");
(void)Success;
MadeChange = true;
} else if (RHSBO) {
// Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the RHS is not
// part of the expression tree.
LinearizeExpr(I);
LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
RHS = I->getOperand(1);
RHSBO = 0;
}
// Okay, now we know that the LHS is a nested expression and that the RHS is
// not. Perform reassociation.
assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
// Move LHS right before I to make sure that the tree expression dominates all
// values.
LHSBO->moveBefore(I);
// Linearize the expression tree on the LHS.
LinearizeExprTree(LHSBO, Ops);
// Remember the RHS operand and its rank.
Ops.push_back(ValueEntry(getRank(RHS), RHS));
// Clear the RHS leaf out.
I->setOperand(1, UndefValue::get(I->getType()));
}
