#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/ValueTracking.h"
-#include "llvm/DataLayout.h"
-#include "llvm/GlobalAlias.h"
-#include "llvm/Operator.h"
+#include "llvm/Analysis/MemoryBuiltins.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/GlobalAlias.h"
+#include "llvm/IR/Operator.h"
#include "llvm/Support/ConstantRange.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/PatternMatch.h"
/// accumulates the total constant offset applied in the returned constant. It
/// returns 0 if V is not a pointer, and returns the constant '0' if there are
/// no constant offsets applied.
-static Constant *stripAndComputeConstantOffsets(const DataLayout &TD,
+///
+/// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
+/// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
+/// folding.
+static Constant *stripAndComputeConstantOffsets(const DataLayout *TD,
Value *&V) {
- if (!V->getType()->isPointerTy())
- return 0;
+ assert(V->getType()->getScalarType()->isPointerTy());
+
+ // Without DataLayout, just be conservative for now. Theoretically, more could
+ // be done in this case.
+ if (!TD)
+ return ConstantInt::get(IntegerType::get(V->getContext(), 64), 0);
- unsigned IntPtrWidth = TD.getPointerSizeInBits();
+ unsigned IntPtrWidth = TD->getPointerSizeInBits();
APInt Offset = APInt::getNullValue(IntPtrWidth);
// Even though we don't look through PHI nodes, we could be called on an
Visited.insert(V);
do {
if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
- if (!GEP->isInBounds() || !GEP->accumulateConstantOffset(TD, Offset))
+ if (!GEP->isInBounds() || !GEP->accumulateConstantOffset(*TD, Offset))
break;
V = GEP->getPointerOperand();
} else if (Operator::getOpcode(V) == Instruction::BitCast) {
} else {
break;
}
- assert(V->getType()->isPointerTy() && "Unexpected operand type!");
+ assert(V->getType()->getScalarType()->isPointerTy() &&
+ "Unexpected operand type!");
} while (Visited.insert(V));
- Type *IntPtrTy = TD.getIntPtrType(V->getContext());
- return ConstantInt::get(IntPtrTy, Offset);
+ Type *IntPtrTy = TD->getIntPtrType(V->getContext());
+ Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
+ if (V->getType()->isVectorTy())
+ return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
+ OffsetIntPtr);
+ return OffsetIntPtr;
}
/// \brief Compute the constant difference between two pointer values.
/// If the difference is not a constant, returns zero.
-static Constant *computePointerDifference(const DataLayout &TD,
+static Constant *computePointerDifference(const DataLayout *TD,
Value *LHS, Value *RHS) {
Constant *LHSOffset = stripAndComputeConstantOffsets(TD, LHS);
- if (!LHSOffset)
- return 0;
Constant *RHSOffset = stripAndComputeConstantOffsets(TD, RHS);
- if (!RHSOffset)
- return 0;
// If LHS and RHS are not related via constant offsets to the same base
// value, there is nothing we can do here.
return W;
// Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
- if (Q.TD && match(Op0, m_PtrToInt(m_Value(X))) &&
+ if (match(Op0, m_PtrToInt(m_Value(X))) &&
match(Op1, m_PtrToInt(m_Value(Y))))
- if (Constant *Result = computePointerDifference(*Q.TD, X, Y))
+ if (Constant *Result = computePointerDifference(Q.TD, X, Y))
return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
// Mul distributes over Sub. Try some generic simplifications based on this.
if (Value *V = SimplifyShift(Instruction::LShr, Op0, Op1, Q, MaxRecurse))
return V;
+ // X >> X -> 0
+ if (Op0 == Op1)
+ return Constant::getNullValue(Op0->getType());
+
// undef >>l X -> 0
if (match(Op0, m_Undef()))
return Constant::getNullValue(Op0->getType());
if (Value *V = SimplifyShift(Instruction::AShr, Op0, Op1, Q, MaxRecurse))
return V;
+ // X >> X -> 0
+ if (Op0 == Op1)
+ return Constant::getNullValue(Op0->getType());
+
// all ones >>a X -> all ones
if (match(Op0, m_AllOnes()))
return Op0;
// A & (-A) = A if A is a power of two or zero.
if (match(Op0, m_Neg(m_Specific(Op1))) ||
match(Op1, m_Neg(m_Specific(Op0)))) {
- if (isPowerOfTwo(Op0, /*OrZero*/true))
+ if (isKnownToBeAPowerOfTwo(Op0, /*OrZero*/true))
return Op0;
- if (isPowerOfTwo(Op1, /*OrZero*/true))
+ if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/true))
return Op1;
}
return 0;
}
-static Constant *computePointerICmp(const DataLayout &TD,
+// A significant optimization not implemented here is assuming that alloca
+// addresses are not equal to incoming argument values. They don't *alias*,
+// as we say, but that doesn't mean they aren't equal, so we take a
+// conservative approach.
+//
+// This is inspired in part by C++11 5.10p1:
+// "Two pointers of the same type compare equal if and only if they are both
+// null, both point to the same function, or both represent the same
+// address."
+//
+// This is pretty permissive.
+//
+// It's also partly due to C11 6.5.9p6:
+// "Two pointers compare equal if and only if both are null pointers, both are
+// pointers to the same object (including a pointer to an object and a
+// subobject at its beginning) or function, both are pointers to one past the
+// last element of the same array object, or one is a pointer to one past the
+// end of one array object and the other is a pointer to the start of a
+// different array object that happens to immediately follow the first array
+// object in the address space.)
+//
+// C11's version is more restrictive, however there's no reason why an argument
+// couldn't be a one-past-the-end value for a stack object in the caller and be
+// equal to the beginning of a stack object in the callee.
+//
+// If the C and C++ standards are ever made sufficiently restrictive in this
+// area, it may be possible to update LLVM's semantics accordingly and reinstate
+// this optimization.
+static Constant *computePointerICmp(const DataLayout *TD,
+ const TargetLibraryInfo *TLI,
CmpInst::Predicate Pred,
Value *LHS, Value *RHS) {
+ // First, skip past any trivial no-ops.
+ LHS = LHS->stripPointerCasts();
+ RHS = RHS->stripPointerCasts();
+
+ // A non-null pointer is not equal to a null pointer.
+ if (llvm::isKnownNonNull(LHS) && isa<ConstantPointerNull>(RHS) &&
+ (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
+ return ConstantInt::get(GetCompareTy(LHS),
+ !CmpInst::isTrueWhenEqual(Pred));
+
// We can only fold certain predicates on pointer comparisons.
switch (Pred) {
default:
break;
}
+ // Strip off any constant offsets so that we can reason about them.
+ // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
+ // here and compare base addresses like AliasAnalysis does, however there are
+ // numerous hazards. AliasAnalysis and its utilities rely on special rules
+ // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
+ // doesn't need to guarantee pointer inequality when it says NoAlias.
Constant *LHSOffset = stripAndComputeConstantOffsets(TD, LHS);
- if (!LHSOffset)
- return 0;
Constant *RHSOffset = stripAndComputeConstantOffsets(TD, RHS);
- if (!RHSOffset)
- return 0;
- // If LHS and RHS are not related via constant offsets to the same base
- // value, there is nothing we can do here.
- if (LHS != RHS)
- return 0;
+ // If LHS and RHS are related via constant offsets to the same base
+ // value, we can replace it with an icmp which just compares the offsets.
+ if (LHS == RHS)
+ return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
+
+ // Various optimizations for (in)equality comparisons.
+ if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
+ // Different non-empty allocations that exist at the same time have
+ // different addresses (if the program can tell). Global variables always
+ // exist, so they always exist during the lifetime of each other and all
+ // allocas. Two different allocas usually have different addresses...
+ //
+ // However, if there's an @llvm.stackrestore dynamically in between two
+ // allocas, they may have the same address. It's tempting to reduce the
+ // scope of the problem by only looking at *static* allocas here. That would
+ // cover the majority of allocas while significantly reducing the likelihood
+ // of having an @llvm.stackrestore pop up in the middle. However, it's not
+ // actually impossible for an @llvm.stackrestore to pop up in the middle of
+ // an entry block. Also, if we have a block that's not attached to a
+ // function, we can't tell if it's "static" under the current definition.
+ // Theoretically, this problem could be fixed by creating a new kind of
+ // instruction kind specifically for static allocas. Such a new instruction
+ // could be required to be at the top of the entry block, thus preventing it
+ // from being subject to a @llvm.stackrestore. Instcombine could even
+ // convert regular allocas into these special allocas. It'd be nifty.
+ // However, until then, this problem remains open.
+ //
+ // So, we'll assume that two non-empty allocas have different addresses
+ // for now.
+ //
+ // With all that, if the offsets are within the bounds of their allocations
+ // (and not one-past-the-end! so we can't use inbounds!), and their
+ // allocations aren't the same, the pointers are not equal.
+ //
+ // Note that it's not necessary to check for LHS being a global variable
+ // address, due to canonicalization and constant folding.
+ if (isa<AllocaInst>(LHS) &&
+ (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
+ ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
+ ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
+ uint64_t LHSSize, RHSSize;
+ if (LHSOffsetCI && RHSOffsetCI &&
+ getObjectSize(LHS, LHSSize, TD, TLI) &&
+ getObjectSize(RHS, RHSSize, TD, TLI)) {
+ const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
+ const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
+ if (!LHSOffsetValue.isNegative() &&
+ !RHSOffsetValue.isNegative() &&
+ LHSOffsetValue.ult(LHSSize) &&
+ RHSOffsetValue.ult(RHSSize)) {
+ return ConstantInt::get(GetCompareTy(LHS),
+ !CmpInst::isTrueWhenEqual(Pred));
+ }
+ }
+
+ // Repeat the above check but this time without depending on DataLayout
+ // or being able to compute a precise size.
+ if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
+ !cast<PointerType>(RHS->getType())->isEmptyTy() &&
+ LHSOffset->isNullValue() &&
+ RHSOffset->isNullValue())
+ return ConstantInt::get(GetCompareTy(LHS),
+ !CmpInst::isTrueWhenEqual(Pred));
+ }
+ }
- return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
+ // Otherwise, fail.
+ return 0;
}
/// SimplifyICmpInst - Given operands for an ICmpInst, see if we can
}
}
- // icmp <object*>, <object*/null> - Different identified objects have
- // different addresses (unless null), and what's more the address of an
- // identified local is never equal to another argument (again, barring null).
- // Note that generalizing to the case where LHS is a global variable address
- // or null is pointless, since if both LHS and RHS are constants then we
- // already constant folded the compare, and if only one of them is then we
- // moved it to RHS already.
- Value *LHSPtr = LHS->stripPointerCasts();
- Value *RHSPtr = RHS->stripPointerCasts();
- if (LHSPtr == RHSPtr)
- return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
-
- // Be more aggressive about stripping pointer adjustments when checking a
- // comparison of an alloca address to another object. We can rip off all
- // inbounds GEP operations, even if they are variable.
- LHSPtr = LHSPtr->stripInBoundsOffsets();
- if (llvm::isIdentifiedObject(LHSPtr)) {
- RHSPtr = RHSPtr->stripInBoundsOffsets();
- if (llvm::isKnownNonNull(LHSPtr) || llvm::isKnownNonNull(RHSPtr)) {
- // If both sides are different identified objects, they aren't equal
- // unless they're null.
- if (LHSPtr != RHSPtr && llvm::isIdentifiedObject(RHSPtr) &&
- Pred == CmpInst::ICMP_EQ)
- return ConstantInt::get(ITy, false);
-
- // A local identified object (alloca or noalias call) can't equal any
- // incoming argument, unless they're both null or they belong to
- // different functions. The latter happens during inlining.
- if (Instruction *LHSInst = dyn_cast<Instruction>(LHSPtr))
- if (Argument *RHSArg = dyn_cast<Argument>(RHSPtr))
- if (LHSInst->getParent()->getParent() == RHSArg->getParent() &&
- Pred == CmpInst::ICMP_EQ)
- return ConstantInt::get(ITy, false);
- }
-
- // Assume that the constant null is on the right.
- if (llvm::isKnownNonNull(LHSPtr) && isa<ConstantPointerNull>(RHSPtr)) {
- if (Pred == CmpInst::ICMP_EQ)
- return ConstantInt::get(ITy, false);
- else if (Pred == CmpInst::ICMP_NE)
- return ConstantInt::get(ITy, true);
- }
- } else if (Argument *LHSArg = dyn_cast<Argument>(LHSPtr)) {
- RHSPtr = RHSPtr->stripInBoundsOffsets();
- // An alloca can't be equal to an argument unless they come from separate
- // functions via inlining.
- if (AllocaInst *RHSInst = dyn_cast<AllocaInst>(RHSPtr)) {
- if (LHSArg->getParent() == RHSInst->getParent()->getParent()) {
- if (Pred == CmpInst::ICMP_EQ)
- return ConstantInt::get(ITy, false);
- else if (Pred == CmpInst::ICMP_NE)
- return ConstantInt::get(ITy, true);
- }
- }
- }
-
// If we are comparing with zero then try hard since this is a common case.
if (match(RHS, m_Zero())) {
bool LHSKnownNonNegative, LHSKnownNegative;
}
}
+ // icmp pred (urem X, Y), Y
if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
bool KnownNonNegative, KnownNegative;
switch (Pred) {
break;
case ICmpInst::ICMP_SGT:
case ICmpInst::ICMP_SGE:
- ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.TD);
+ ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.TD);
if (!KnownNonNegative)
break;
// fall-through
return getFalse(ITy);
case ICmpInst::ICMP_SLT:
case ICmpInst::ICMP_SLE:
- ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.TD);
+ ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.TD);
if (!KnownNonNegative)
break;
// fall-through
return getTrue(ITy);
}
}
+
+ // icmp pred X, (urem Y, X)
if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
bool KnownNonNegative, KnownNegative;
switch (Pred) {
break;
case ICmpInst::ICMP_SGT:
case ICmpInst::ICMP_SGE:
- ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.TD);
+ ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.TD);
if (!KnownNonNegative)
break;
// fall-through
return getTrue(ITy);
case ICmpInst::ICMP_SLT:
case ICmpInst::ICMP_SLE:
- ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.TD);
+ ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.TD);
if (!KnownNonNegative)
break;
// fall-through
// Simplify comparisons of related pointers using a powerful, recursive
// GEP-walk when we have target data available..
- if (Q.TD && LHS->getType()->isPointerTy() && RHS->getType()->isPointerTy())
- if (Constant *C = computePointerICmp(*Q.TD, Pred, LHS, RHS))
+ if (LHS->getType()->isPointerTy())
+ if (Constant *C = computePointerICmp(Q.TD, Q.TLI, Pred, LHS, RHS))
return C;
if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
RecursionLimit);
}
-static Value *SimplifyCallInst(CallInst *CI, const Query &) {
- // call undef -> undef
- if (isa<UndefValue>(CI->getCalledValue()))
- return UndefValue::get(CI->getType());
+static bool IsIdempotent(Intrinsic::ID ID) {
+ switch (ID) {
+ default: return false;
+
+ // Unary idempotent: f(f(x)) = f(x)
+ case Intrinsic::fabs:
+ case Intrinsic::floor:
+ case Intrinsic::ceil:
+ case Intrinsic::trunc:
+ case Intrinsic::rint:
+ case Intrinsic::nearbyint:
+ return true;
+ }
+}
+
+template <typename IterTy>
+static Value *SimplifyIntrinsic(Intrinsic::ID IID, IterTy ArgBegin, IterTy ArgEnd,
+ const Query &Q, unsigned MaxRecurse) {
+ // Perform idempotent optimizations
+ if (!IsIdempotent(IID))
+ return 0;
+
+ // Unary Ops
+ if (std::distance(ArgBegin, ArgEnd) == 1)
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin))
+ if (II->getIntrinsicID() == IID)
+ return II;
return 0;
}
+template <typename IterTy>
+static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
+ const Query &Q, unsigned MaxRecurse) {
+ Type *Ty = V->getType();
+ if (PointerType *PTy = dyn_cast<PointerType>(Ty))
+ Ty = PTy->getElementType();
+ FunctionType *FTy = cast<FunctionType>(Ty);
+
+ // call undef -> undef
+ if (isa<UndefValue>(V))
+ return UndefValue::get(FTy->getReturnType());
+
+ Function *F = dyn_cast<Function>(V);
+ if (!F)
+ return 0;
+
+ if (unsigned IID = F->getIntrinsicID())
+ if (Value *Ret =
+ SimplifyIntrinsic((Intrinsic::ID) IID, ArgBegin, ArgEnd, Q, MaxRecurse))
+ return Ret;
+
+ if (!canConstantFoldCallTo(F))
+ return 0;
+
+ SmallVector<Constant *, 4> ConstantArgs;
+ ConstantArgs.reserve(ArgEnd - ArgBegin);
+ for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
+ Constant *C = dyn_cast<Constant>(*I);
+ if (!C)
+ return 0;
+ ConstantArgs.push_back(C);
+ }
+
+ return ConstantFoldCall(F, ConstantArgs, Q.TLI);
+}
+
+Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
+ User::op_iterator ArgEnd, const DataLayout *TD,
+ const TargetLibraryInfo *TLI,
+ const DominatorTree *DT) {
+ return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(TD, TLI, DT),
+ RecursionLimit);
+}
+
+Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
+ const DataLayout *TD, const TargetLibraryInfo *TLI,
+ const DominatorTree *DT) {
+ return ::SimplifyCall(V, Args.begin(), Args.end(), Query(TD, TLI, DT),
+ RecursionLimit);
+}
+
/// SimplifyInstruction - See if we can compute a simplified version of this
/// instruction. If not, this returns null.
Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout *TD,
case Instruction::PHI:
Result = SimplifyPHINode(cast<PHINode>(I), Query (TD, TLI, DT));
break;
- case Instruction::Call:
- Result = SimplifyCallInst(cast<CallInst>(I), Query (TD, TLI, DT));
+ case Instruction::Call: {
+ CallSite CS(cast<CallInst>(I));
+ Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
+ TD, TLI, DT);
break;
+ }
case Instruction::Trunc:
Result = SimplifyTruncInst(I->getOperand(0), I->getType(), TD, TLI, DT);
break;