//===----------------------------------------------------------------------===//
#include "InstCombine.h"
-#include "llvm/IntrinsicInst.h"
+#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/MemoryBuiltins.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/IntrinsicInst.h"
#include "llvm/Support/ConstantRange.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/PatternMatch.h"
+#include "llvm/Target/TargetLibraryInfo.h"
using namespace llvm;
using namespace PatternMatch;
static bool HasAddOverflow(ConstantInt *Result,
ConstantInt *In1, ConstantInt *In2,
bool IsSigned) {
- if (IsSigned)
- if (In2->getValue().isNegative())
- return Result->getValue().sgt(In1->getValue());
- else
- return Result->getValue().slt(In1->getValue());
- else
+ if (!IsSigned)
return Result->getValue().ult(In1->getValue());
+
+ if (In2->isNegative())
+ return Result->getValue().sgt(In1->getValue());
+ return Result->getValue().slt(In1->getValue());
}
/// AddWithOverflow - Compute Result = In1+In2, returning true if the result
Constant *In2, bool IsSigned = false) {
Result = ConstantExpr::getAdd(In1, In2);
- if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
+ if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
if (HasAddOverflow(ExtractElement(Result, Idx),
static bool HasSubOverflow(ConstantInt *Result,
ConstantInt *In1, ConstantInt *In2,
bool IsSigned) {
- if (IsSigned)
- if (In2->getValue().isNegative())
- return Result->getValue().slt(In1->getValue());
- else
- return Result->getValue().sgt(In1->getValue());
- else
+ if (!IsSigned)
return Result->getValue().ugt(In1->getValue());
+
+ if (In2->isNegative())
+ return Result->getValue().slt(In1->getValue());
+
+ return Result->getValue().sgt(In1->getValue());
}
/// SubWithOverflow - Compute Result = In1-In2, returning true if the result
Constant *In2, bool IsSigned = false) {
Result = ConstantExpr::getSub(In1, In2);
- if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
+ if (VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
if (HasSubOverflow(ExtractElement(Result, Idx),
case ICmpInst::ICMP_UGT:
// True if LHS u> RHS and RHS == high-bit-mask - 1
TrueIfSigned = true;
- return RHS->getValue() ==
- APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
- case ICmpInst::ICMP_UGE:
+ return RHS->isMaxValue(true);
+ case ICmpInst::ICMP_UGE:
// True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
TrueIfSigned = true;
return RHS->getValue().isSignBit();
return (~CI->getValue() + 1).isPowerOf2();
}
-/// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
+/// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
/// set of known zero and one bits, compute the maximum and minimum values that
/// could have the specified known zero and known one bits, returning them in
/// min/max.
// bit if it is unknown.
Min = KnownOne;
Max = KnownOne|UnknownBits;
-
+
if (UnknownBits.isNegative()) { // Sign bit is unknown
Min.setBit(Min.getBitWidth()-1);
Max.clearBit(Max.getBitWidth()-1);
KnownZero.getBitWidth() == Max.getBitWidth() &&
"Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
APInt UnknownBits = ~(KnownZero|KnownOne);
-
+
// The minimum value is when the unknown bits are all zeros.
Min = KnownOne;
// The maximum value is when the unknown bits are all ones.
CmpInst &ICI, ConstantInt *AndCst) {
// We need TD information to know the pointer size unless this is inbounds.
if (!GEP->isInBounds() && TD == 0) return 0;
+
+ Constant *Init = GV->getInitializer();
+ if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
+ return 0;
- ConstantArray *Init = dyn_cast<ConstantArray>(GV->getInitializer());
- if (Init == 0 || Init->getNumOperands() > 1024) return 0;
-
+ uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
+ if (ArrayElementCount > 1024) return 0; // Don't blow up on huge arrays.
+
// There are many forms of this optimization we can handle, for now, just do
// the simple index into a single-dimensional array.
//
// type they index. Collect the indices. This is typically for arrays of
// structs.
SmallVector<unsigned, 4> LaterIndices;
-
- const Type *EltTy = cast<ArrayType>(Init->getType())->getElementType();
+
+ Type *EltTy = Init->getType()->getArrayElementType();
for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
if (Idx == 0) return 0; // Variable index.
-
+
uint64_t IdxVal = Idx->getZExtValue();
if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
-
- if (const StructType *STy = dyn_cast<StructType>(EltTy))
+
+ if (StructType *STy = dyn_cast<StructType>(EltTy))
EltTy = STy->getElementType(IdxVal);
- else if (
const ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
+ else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
if (IdxVal >= ATy->getNumElements()) return 0;
EltTy = ATy->getElementType();
} else {
return 0; // Unknown type.
}
-
+
LaterIndices.push_back(IdxVal);
}
-
+
enum { Overdefined = -3, Undefined = -2 };
// Variables for our state machines.
-
+
// FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
// "i == 47 | i == 87", where 47 is the first index the condition is true for,
// and 87 is the second (and last) index. FirstTrueElement is -2 when
// FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
// form "i != 47 & i != 87". Same state transitions as for true elements.
int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
-
+
/// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
/// define a state machine that triggers for ranges of values that the index
/// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
/// index in the range (inclusive). We use -2 for undefined here because we
/// use relative comparisons and don't want 0-1 to match -1.
int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
-
+
// MagicBitvector - This is a magic bitvector where we set a bit if the
// comparison is true for element 'i'. If there are 64 elements or less in
// the array, this will fully represent all the comparison results.
uint64_t MagicBitvector = 0;
-
-
+
+
// Scan the array and see if one of our patterns matches.
Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
- for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) {
- Constant *Elt = Init->getOperand(i);
-
+ for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
+ Constant *Elt = Init->getAggregateElement(i);
+ if (Elt == 0) return 0;
+
// If this is indexing an array of structures, get the structure element.
if (!LaterIndices.empty())
- Elt = ConstantExpr::getExtractValue(Elt, LaterIndices.data(),
- LaterIndices.size());
-
+ Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
+
// If the element is masked, handle it.
if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
-
+
// Find out if the comparison would be true or false for the i'th element.
Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
- CompareRHS, TD);
+ CompareRHS, TD, TLI);
// If the result is undef for this element, ignore it.
if (isa<UndefValue>(C)) {
// Extend range state machines to cover this element in case there is an
FalseRangeEnd = i;
continue;
}
-
+
// If we can't compute the result for any of the elements, we have to give
// up evaluating the entire conditional.
if (!isa<ConstantInt>(C)) return 0;
-
+
// Otherwise, we know if the comparison is true or false for this element,
// update our state machines.
bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
-
+
// State machine for single/double/range index comparison.
if (IsTrueForElt) {
// Update the TrueElement state machine.
SecondTrueElement = i;
else
SecondTrueElement = Overdefined;
-
+
// Update range state machine.
if (TrueRangeEnd == (int)i-1)
TrueRangeEnd = i;
SecondFalseElement = i;
else
SecondFalseElement = Overdefined;
-
+
// Update range state machine.
if (FalseRangeEnd == (int)i-1)
FalseRangeEnd = i;
FalseRangeEnd = Overdefined;
}
}
-
-
+
+
// If this element is in range, update our magic bitvector.
if (i < 64 && IsTrueForElt)
MagicBitvector |= 1ULL << i;
-
+
// If all of our states become overdefined, bail out early. Since the
// predicate is expensive, only check it every 8 elements. This is only
// really useful for really huge arrays.
if (!GEP->isInBounds() &&
Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
-
+
// If the comparison is only true for one or two elements, emit direct
// comparisons.
if (SecondTrueElement != Overdefined) {
// None true -> false.
if (FirstTrueElement == Undefined)
return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext()));
-
+
Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
-
+
// True for one element -> 'i == 47'.
if (SecondTrueElement == Undefined)
return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
-
+
// True for two elements -> 'i == 47 | i == 72'.
Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
// None false -> true.
if (FirstFalseElement == Undefined)
return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext()));
-
+
Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
// False for one element -> 'i != 47'.
if (SecondFalseElement == Undefined)
return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
-
+
// False for two elements -> 'i != 47 & i != 72'.
Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
return BinaryOperator::CreateAnd(C1, C2);
}
-
+
// If the comparison can be replaced with a range comparison for the elements
// where it is true, emit the range check.
if (TrueRangeEnd != Overdefined) {
assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
-
+
// Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
if (FirstTrueElement) {
Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
Idx = Builder->CreateAdd(Idx, Offs);
}
-
+
Value *End = ConstantInt::get(Idx->getType(),
TrueRangeEnd-FirstTrueElement+1);
return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
}
-
+
// False range check.
if (FalseRangeEnd != Overdefined) {
assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
Idx = Builder->CreateAdd(Idx, Offs);
}
-
+
Value *End = ConstantInt::get(Idx->getType(),
FalseRangeEnd-FirstFalseElement);
return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
}
-
-
+
+
// If a 32-bit or 64-bit magic bitvector captures the entire comparison state
// of this load, replace it with computation that does:
// ((magic_cst >> i) & 1) != 0
- if (Init->getNumOperands() <= 32 ||
- (TD && Init->getNumOperands() <= 64 && TD->isLegalInteger(64))) {
- const Type *Ty;
- if (Init->getNumOperands() <= 32)
+ if (ArrayElementCount <= 32 ||
+ (TD && ArrayElementCount <= 64 && TD->isLegalInteger(64))) {
+ Type *Ty;
+ if (ArrayElementCount <= 32)
Ty = Type::getInt32Ty(Init->getContext());
else
Ty = Type::getInt64Ty(Init->getContext());
V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
}
-
+
return 0;
}
/// to generate the first by knowing that pointer arithmetic doesn't overflow.
///
/// If we can't emit an optimized form for this expression, this returns null.
-///
-static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
- InstCombiner &IC) {
- TargetData &TD = *IC.getTargetData();
+///
+static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
+ DataLayout &TD = *IC.getDataLayout();
gep_type_iterator GTI = gep_type_begin(GEP);
-
+
// Check to see if this gep only has a single variable index. If so, and if
// any constant indices are a multiple of its scale, then we can compute this
// in terms of the scale of the variable index. For example, if the GEP
if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
// Compute the aggregate offset of constant indices.
if (CI->isZero()) continue;
-
+
// Handle a struct index, which adds its field offset to the pointer.
- if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
+ if (StructType *STy = dyn_cast<StructType>(*GTI)) {
Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
} else {
uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
break;
}
}
-
+
// If there are no variable indices, we must have a constant offset, just
// evaluate it the general way.
if (i == e) return 0;
-
+
Value *VariableIdx = GEP->getOperand(i);
// Determine the scale factor of the variable element. For example, this is
// 4 if the variable index is into an array of i32.
uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
-
+
// Verify that there are no other variable indices. If so, emit the hard way.
for (++i, ++GTI; i != e; ++i, ++GTI) {
ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
if (!CI) return 0;
-
+
// Compute the aggregate offset of constant indices.
if (CI->isZero()) continue;
-
+
// Handle a struct index, which adds its field offset to the pointer.
- if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
+ if (StructType *STy = dyn_cast<StructType>(*GTI)) {
Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
} else {
uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
Offset += Size*CI->getSExtValue();
}
}
-
+
// Okay, we know we have a single variable index, which must be a
// pointer/array/vector index. If there is no offset, life is simple, return
// the index.
// Cast to intptrty in case a truncation occurs. If an extension is needed,
// we don't need to bother extending: the extension won't affect where the
// computation crosses zero.
- if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
- VariableIdx = new TruncInst(VariableIdx,
- TD.getIntPtrType(VariableIdx->getContext()),
- VariableIdx->getName(), &I);
+ if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
+ Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
+ VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy);
+ }
return VariableIdx;
}
-
+
// Otherwise, there is an index. The computation we will do will be modulo
// the pointer size, so get it.
uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
-
+
Offset &= PtrSizeMask;
VariableScale &= PtrSizeMask;
-
+
// To do this transformation, any constant index must be a multiple of the
// variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
// but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
int64_t NewOffs = Offset / (int64_t)VariableScale;
if (Offset != NewOffs*(int64_t)VariableScale)
return 0;
-
+
// Okay, we can do this evaluation. Start by converting the index to intptr.
- const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
+ Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
if (VariableIdx->getType() != IntPtrTy)
- VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
- true /*SExt*/,
- VariableIdx->getName(), &I);
+ VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy,
+ true /*Signed*/);
Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
- return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
+ return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset");
}
/// FoldGEPICmp - Fold comparisons between a GEP instruction and something
Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
ICmpInst::Predicate Cond,
Instruction &I) {
+ // Don't transform signed compares of GEPs into index compares. Even if the
+ // GEP is inbounds, the final add of the base pointer can have signed overflow
+ // and would change the result of the icmp.
+ // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
+ // the maximum signed value for the pointer type.
+ if (ICmpInst::isSigned(Cond))
+ return 0;
+
// Look through bitcasts.
if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
RHS = BCI->getOperand(0);
// This transformation (ignoring the base and scales) is valid because we
// know pointers can't overflow since the gep is inbounds. See if we can
// output an optimized form.
- Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
-
+ Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this);
+
// If not, synthesize the offset the hard way.
if (Offset == 0)
Offset = EmitGEPOffset(GEPLHS);
return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
GEPLHS->getOperand(0), GEPRHS->getOperand(0));
+ // If we're comparing GEPs with two base pointers that only differ in type
+ // and both GEPs have only constant indices or just one use, then fold
+ // the compare with the adjusted indices.
+ if (TD && GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
+ (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
+ (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
+ PtrBase->stripPointerCasts() ==
+ GEPRHS->getOperand(0)->stripPointerCasts()) {
+ Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond),
+ EmitGEPOffset(GEPLHS),
+ EmitGEPOffset(GEPRHS));
+ return ReplaceInstUsesWith(I, Cmp);
+ }
+
// Otherwise, the base pointers are different and the indices are
// different, bail out.
return 0;
if (AllZeros)
return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
+ bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
// If the GEPs only differ by one index, compare it.
unsigned NumDifferences = 0; // Keep track of # differences.
ConstantInt::get(Type::getInt1Ty(I.getContext()),
ICmpInst::isTrueWhenEqual(Cond)));
- else if (NumDifferences == 1) {
+ else if (NumDifferences == 1 && GEPsInBounds) {
Value *LHSV = GEPLHS->getOperand(DiffOperand);
Value *RHSV = GEPRHS->getOperand(DiffOperand);
// Make sure we do a signed comparison here.
// Only lower this if the icmp is the only user of the GEP or if we expect
// the result to fold to a constant!
if (TD &&
+ GEPsInBounds &&
(isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
(isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
// ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
}
-
+
// (X+4) == X -> false.
if (Pred == ICmpInst::ICMP_EQ)
return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
// From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
// so the values can never be equal. Similarly for all other "or equals"
// operators.
-
+
// (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
// (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
// (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
- Value *R =
+ Value *R =
ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
}
-
+
// (X+1) >u X --> X <u (0-1) --> X != 255
// (X+2) >u X --> X <u (0-2) --> X <u 254
// (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
-
+
unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
ConstantInt *SMax = ConstantInt::get(X->getContext(),
APInt::getSignedMaxValue(BitWidth));
// (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
-
+
// (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
// (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
// (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
// (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
// (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
// (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
-
+
assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
ConstantInt *DivRHS) {
ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
const APInt &CmpRHSV = CmpRHS->getValue();
-
- // FIXME: If the operand types don't match the type of the divide
+
+ // FIXME: If the operand types don't match the type of the divide
// then don't attempt this transform. The code below doesn't have the
// logic to deal with a signed divide and an unsigned compare (and
- // vice versa). This is because (x /s C1) <s C2 produces different
+ // vice versa). This is because (x /s C1) <s C2 produces different
// results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
- // (x /u C1) <u C2. Simply casting the operands and result won't
- // work. :( The if statement below tests that condition and bails
+ // (x /u C1) <u C2. Simply casting the operands and result won't
+ // work. :( The if statement below tests that condition and bails
// if it finds it.
bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
}
// Compute Prod = CI * DivRHS. We are essentially solving an equation
- // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
- // C2 (CI). By solving for X we can turn this into a range check
- // instead of computing a divide.
+ // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
+ // C2 (CI). By solving for X we can turn this into a range check
+ // instead of computing a divide.
Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
// Determine if the product overflows by seeing if the product is
// not equal to the divide. Make sure we do the same kind of divide
- // as in the LHS instruction that we're folding.
+ // as in the LHS instruction that we're folding.
bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
/// If the division is known to be exact, then there is no remainder from the
/// divide, so the covered range size is unit, otherwise it is the divisor.
ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
-
+
// Figure out the interval that is being checked. For example, a comparison
- // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
+ // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
// Compute this interval based on the constants involved and the signedness of
// the compare/divide. This computes a half-open interval, keeping track of
// whether either value in the interval overflows. After analysis each
// to the same result value.
HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
}
-
+
} else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
if (CmpRHSV == 0) { // (X / pos) op 0
// Can't overflow. e.g. X/2 op 0 --> [-1, 2)
LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
}
}
- } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
+ } else if (DivRHS->isNegative()) { // Divisor is < 0.
if (DivI->isExact())
RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
if (CmpRHSV == 0) { // (X / neg) op 0
if (!HiOverflow)
HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
}
-
+
// Dividing by a negative swaps the condition. LT <-> GT
Pred = ICmpInst::getSwappedPredicate(Pred);
}
Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
ConstantInt *ShAmt) {
const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
-
+
// Check that the shift amount is in range. If not, don't perform
// undefined shifts. When the shift is visited it will be
// simplified.
uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
if (ShAmtVal >= TypeBits || ShAmtVal == 0)
return 0;
-
+
if (!ICI.isEquality()) {
// If we have an unsigned comparison and an ashr, we can't simplify this.
// Similarly for signed comparisons with lshr.
if (ICI.isSigned() != (Shr->getOpcode() == Instruction::AShr))
return 0;
-
- // Otherwise, all lshr and all exact ashr's are equivalent to a udiv/sdiv by
- // a power of 2. Since we already have logic to simplify these, transform
- // to div and then simplify the resultant comparison.
+
+ // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv
+ // by a power of 2. Since we already have logic to simplify these,
+ // transform to div and then simplify the resultant comparison.
if (Shr->getOpcode() == Instruction::AShr &&
- !Shr->isExact())
+ (!Shr->isExact() || ShAmtVal == TypeBits - 1))
return 0;
-
+
// Revisit the shift (to delete it).
Worklist.Add(Shr);
-
+
Constant *DivCst =
ConstantInt::get(Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal));
-
+
Value *Tmp =
Shr->getOpcode() == Instruction::AShr ?
Builder->CreateSDiv(Shr->getOperand(0), DivCst, "", Shr->isExact()) :
Builder->CreateUDiv(Shr->getOperand(0), DivCst, "", Shr->isExact());
-
+
ICI.setOperand(0, Tmp);
-
+
// If the builder folded the binop, just return it.
BinaryOperator *TheDiv = dyn_cast<BinaryOperator>(Tmp);
if (TheDiv == 0)
return &ICI;
-
+
// Otherwise, fold this div/compare.
assert(TheDiv->getOpcode() == Instruction::SDiv ||
TheDiv->getOpcode() == Instruction::UDiv);
-
+
Instruction *Res = FoldICmpDivCst(ICI, TheDiv, cast<ConstantInt>(DivCst));
assert(Res && "This div/cst should have folded!");
return Res;
}
-
-
+
+
// If we are comparing against bits always shifted out, the
// comparison cannot succeed.
APInt Comp = CmpRHSV << ShAmtVal;
Comp = Comp.lshr(ShAmtVal);
else
Comp = Comp.ashr(ShAmtVal);
-
+
if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
IsICMP_NE);
return ReplaceInstUsesWith(ICI, Cst);
}
-
+
// Otherwise, check to see if the bits shifted out are known to be zero.
// If so, we can compare against the unshifted value:
// (X & 4) >> 1 == 2 --> (X & 4) == 4.
if (Shr->hasOneUse() && Shr->isExact())
return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
-
+
if (Shr->hasOneUse()) {
// Otherwise strength reduce the shift into an and.
APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
-
+
Value *And = Builder->CreateAnd(Shr->getOperand(0),
Mask, Shr->getName()+".mask");
return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
Instruction *LHSI,
ConstantInt *RHS) {
const APInt &RHSV = RHS->getValue();
-
+
switch (LHSI->getOpcode()) {
case Instruction::Trunc:
if (ICI.isEquality() && LHSI->hasOneUse()) {
// of the high bits truncated out of x are known.
unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
- APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
- ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
-
+ ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
+
// If all the high bits are known, we can do this xform.
if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
// Pull in the high bits from known-ones set.
APInt NewRHS = RHS->getValue().zext(SrcBits);
- NewRHS |= KnownOne;
+ NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
ConstantInt::get(ICI.getContext(), NewRHS));
}
}
break;
-
+
case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
// If this is a comparison that tests the signbit (X < 0) or (x > -1),
if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
(ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
Value *CompareVal = LHSI->getOperand(0);
-
+
// If the sign bit of the XorCST is not set, there is no change to
// the operation, just stop using the Xor.
- if (!XorCST->getValue().isNegative()) {
+ if (!XorCST->isNegative()) {
ICI.setOperand(0, CompareVal);
Worklist.Add(LHSI);
return &ICI;
}
-
+
// Was the old condition true if the operand is positive?
bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
-
+
// If so, the new one isn't.
isTrueIfPositive ^= true;
-
+
if (isTrueIfPositive)
return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
SubOne(RHS));
}
// (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
- if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
+ if (!ICI.isEquality() && XorCST->isMaxValue(true)) {
const APInt &NotSignBit = XorCST->getValue();
ICmpInst::Predicate Pred = ICI.isSigned()
? ICI.getUnsignedPredicate()
if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
LHSI->getOperand(0)->hasOneUse()) {
ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
-
+
// If the LHS is an AND of a truncating cast, we can widen the
// and/compare to be the input width without changing the value
// produced, eliminating a cast.
if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
// We can do this transformation if either the AND constant does not
- // have its sign bit set or if it is an equality comparison.
+ // have its sign bit set or if it is an equality comparison.
// Extending a relational comparison when we're checking the sign
// bit would not work.
- if (Cast->hasOneUse() &&
- (ICI.isEquality() ||
- (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
- uint32_t BitWidth =
- cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
- APInt NewCST = AndCST->getValue().zext(BitWidth);
- APInt NewCI = RHSV.zext(BitWidth);
- Value *NewAnd =
+ if (ICI.isEquality() ||
+ (!AndCST->isNegative() && RHSV.isNonNegative())) {
+ Value *NewAnd =
Builder->CreateAnd(Cast->getOperand(0),
- ConstantInt::get(ICI.getContext(), NewCST),
- LHSI->getName());
+ ConstantExpr::getZExt(AndCST, Cast->getSrcTy()));
+ NewAnd->takeName(LHSI);
return new ICmpInst(ICI.getPredicate(), NewAnd,
- ConstantInt::get(ICI.getContext(), NewCI));
+ ConstantExpr::getZExt(RHS, Cast->getSrcTy()));
}
}
-
+
+ // If the LHS is an AND of a zext, and we have an equality compare, we can
+ // shrink the and/compare to the smaller type, eliminating the cast.
+ if (ZExtInst *Cast = dyn_cast<ZExtInst>(LHSI->getOperand(0))) {
+ IntegerType *Ty = cast<IntegerType>(Cast->getSrcTy());
+ // Make sure we don't compare the upper bits, SimplifyDemandedBits
+ // should fold the icmp to true/false in that case.
+ if (ICI.isEquality() && RHSV.getActiveBits() <= Ty->getBitWidth()) {
+ Value *NewAnd =
+ Builder->CreateAnd(Cast->getOperand(0),
+ ConstantExpr::getTrunc(AndCST, Ty));
+ NewAnd->takeName(LHSI);
+ return new ICmpInst(ICI.getPredicate(), NewAnd,
+ ConstantExpr::getTrunc(RHS, Ty));
+ }
+ }
+
// If this is: (X >> C1) & C2 != C3 (where any shift and any compare
// could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
// happens a LOT in code produced by the C front-end, for bitfield
BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
if (Shift && !Shift->isShift())
Shift = 0;
-
+
ConstantInt *ShAmt;
ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
- const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
- const Type *AndTy = AndCST->getType(); // Type of the and.
-
+ Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
+ Type *AndTy = AndCST->getType(); // Type of the and.
+
// We can fold this as long as we can't shift unknown bits
// into the mask. This can only happen with signed shift
// rights, as they sign-extend.
// of the bits shifted in could be tested after the mask.
uint32_t TyBits = Ty->getPrimitiveSizeInBits();
int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
-
+
uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
- if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
+ if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
AndCST->getValue()) == 0)
CanFold = true;
}
-
+
if (CanFold) {
Constant *NewCst;
if (Shift->getOpcode() == Instruction::Shl)
NewCst = ConstantExpr::getLShr(RHS, ShAmt);
else
NewCst = ConstantExpr::getShl(RHS, ShAmt);
-
+
// Check to see if we are shifting out any of the bits being
// compared.
if (ConstantExpr::get(Shift->getOpcode(),
}
}
}
-
+
// Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
// preferable because it allows the C<<Y expression to be hoisted out
// of a loop if Y is invariant and X is not.
// Compute C << Y.
Value *NS;
if (Shift->getOpcode() == Instruction::LShr) {
- NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
+ NS = Builder->CreateShl(AndCST, Shift->getOperand(1));
} else {
// Insert a logical shift.
- NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
+ NS = Builder->CreateLShr(AndCST, Shift->getOperand(1));
}
-
+
// Compute X & (C << Y).
- Value *NewAnd =
+ Value *NewAnd =
Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
-
+
ICI.setOperand(0, NewAnd);
return &ICI;
}
+
+ // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any
+ // bit set in (X & AndCST) will produce a result greater than RHSV.
+ if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
+ unsigned NTZ = AndCST->getValue().countTrailingZeros();
+ if ((NTZ < AndCST->getBitWidth()) &&
+ APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV))
+ return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
+ Constant::getNullValue(RHS->getType()));
+ }
}
-
+
// Try to optimize things like "A[i]&42 == 0" to index computations.
if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
if (GetElementPtrInst *GEP =
}
break;
}
-
+
case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
if (!ShAmt) break;
-
+
uint32_t TypeBits = RHSV.getBitWidth();
-
+
// Check that the shift amount is in range. If not, don't perform
// undefined shifts. When the shift is visited it will be
// simplified.
if (ShAmt->uge(TypeBits))
break;
-
+
if (ICI.isEquality()) {
// If we are comparing against bits always shifted out, the
// comparison cannot succeed.
ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
return ReplaceInstUsesWith(ICI, Cst);
}
-
+
// If the shift is NUW, then it is just shifting out zeros, no need for an
// AND.
if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
ConstantExpr::getLShr(RHS, ShAmt));
-
+
if (LHSI->hasOneUse()) {
// Otherwise strength reduce the shift into an and.
uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
Constant *Mask =
- ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
+ ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
TypeBits-ShAmtVal));
-
+
Value *And =
Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
return new ICmpInst(ICI.getPredicate(), And,
ConstantExpr::getLShr(RHS, ShAmt));
}
}
-
+
// Otherwise, if this is a comparison of the sign bit, simplify to and/test.
bool TrueIfSigned = false;
if (LHSI->hasOneUse() &&
isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
// (X << 31) <s 0 --> (X&1) != 0
Constant *Mask = ConstantInt::get(LHSI->getOperand(0)->getType(),
- APInt::getOneBitSet(TypeBits,
+ APInt::getOneBitSet(TypeBits,
TypeBits-ShAmt->getZExtValue()-1));
Value *And =
Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
And, Constant::getNullValue(And->getType()));
}
+
+ // Transform (icmp pred iM (shl iM %v, N), CI)
+ // -> (icmp pred i(M-N) (trunc %v iM to i(N-N)), (trunc (CI>>N))
+ // Transform the shl to a trunc if (trunc (CI>>N)) has no loss.
+ // This enables to get rid of the shift in favor of a trunc which can be
+ // free on the target. It has the additional benefit of comparing to a
+ // smaller constant, which will be target friendly.
+ unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
+ if (Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
+ Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
+ Constant *NCI = ConstantExpr::getTrunc(
+ ConstantExpr::getAShr(RHS,
+ ConstantInt::get(RHS->getType(), Amt)),
+ NTy);
+ return new ICmpInst(ICI.getPredicate(),
+ Builder->CreateTrunc(LHSI->getOperand(0), NTy),
+ NCI);
+ }
+
break;
}
-
+
case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
case Instruction::AShr: {
// Handle equality comparisons of shift-by-constant.
}
break;
}
-
+
case Instruction::SDiv:
case Instruction::UDiv:
// Fold: icmp pred ([us]div X, C1), C2 -> range test
- // Fold this div into the comparison, producing a range check.
- // Determine, based on the divide type, what the range is being
- // checked. If there is an overflow on the low or high side, remember
+ // Fold this div into the comparison, producing a range check.
+ // Determine, based on the divide type, what the range is being
+ // checked. If there is an overflow on the low or high side, remember
// it, otherwise compute the range [low, hi) bounding the new value.
// See: InsertRangeTest above for the kinds of replacements possible.
if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
}
break;
}
-
+
// Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
if (ICI.isEquality()) {
bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
-
- // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
+
+ // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
// the second operand is a constant, simplify a bit.
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
switch (BO->getOpcode()) {
// Replace ((add A, B) != 0) with (A != -B) if A or B is
// efficiently invertible, or if the add has just this one use.
Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
-
+
if (Value *NegVal = dyn_castNegVal(BOp1))
return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
if (Value *NegVal = dyn_castNegVal(BOp0))
case Instruction::Xor:
// For the xor case, we can xor two constants together, eliminating
// the explicit xor.
- if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
- return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
+ if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
+ return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
ConstantExpr::getXor(RHS, BOC));
-
- // FALLTHROUGH
+ } else if (RHSV == 0) {
+ // Replace ((xor A, B) != 0) with (A != B)
+ return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
+ BO->getOperand(1));
+ }
+ break;
case Instruction::Sub:
- // Replace (([sub|xor] A, B) != 0) with (A != B)
- if (RHSV == 0)
+ // Replace ((sub A, B) != C) with (B != A-C) if A & C are constants.
+ if (ConstantInt *BOp0C = dyn_cast<ConstantInt>(BO->getOperand(0))) {
+ if (BO->hasOneUse())
+ return new ICmpInst(ICI.getPredicate(), BO->getOperand(1),
+ ConstantExpr::getSub(BOp0C, RHS));
+ } else if (RHSV == 0) {
+ // Replace ((sub A, B) != 0) with (A != B)
return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
BO->getOperand(1));
+ }
break;
-
case Instruction::Or:
// If bits are being or'd in that are not present in the constant we
// are comparing against, then the comparison could never succeed!
Constant *NotCI = ConstantExpr::getNot(RHS);
if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
return ReplaceInstUsesWith(ICI,
- ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
+ ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
isICMP_NE));
}
break;
-
+
case Instruction::And:
if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
// If bits are being compared against that are and'd out, then the
return ReplaceInstUsesWith(ICI,
ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
isICMP_NE));
-
+
// If we have ((X & C) == C), turn it into ((X & C) != 0).
if (RHS == BOC && RHSV.isPowerOf2())
return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
ICmpInst::ICMP_NE, LHSI,
Constant::getNullValue(RHS->getType()));
-
+
+ // Don't perform the following transforms if the AND has multiple uses
+ if (!BO->hasOneUse())
+ break;
+
// Replace (and X, (1 << size(X)-1) != 0) with x s< 0
if (BOC->getValue().isSignBit()) {
Value *X = BO->getOperand(0);
Constant *Zero = Constant::getNullValue(X->getType());
- ICmpInst::Predicate pred = isICMP_NE ?
+ ICmpInst::Predicate pred = isICMP_NE ?
ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
return new ICmpInst(pred, X, Zero);
}
-
+
// ((X & ~7) == 0) --> X < 8
if (RHSV == 0 && isHighOnes(BOC)) {
Value *X = BO->getOperand(0);
Constant *NegX = ConstantExpr::getNeg(BOC);
- ICmpInst::Predicate pred = isICMP_NE ?
+ ICmpInst::Predicate pred = isICMP_NE ?
ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
return new ICmpInst(pred, X, NegX);
}
Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
Value *LHSCIOp = LHSCI->getOperand(0);
- const Type *SrcTy = LHSCIOp->getType();
- const Type *DestTy = LHSCI->getType();
+ Type *SrcTy = LHSCIOp->getType();
+ Type *DestTy = LHSCI->getType();
Value *RHSCIOp;
- // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
+ // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
// integer type is the same size as the pointer type.
if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
TD->getPointerSizeInBits() ==
if (RHSOp)
return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
}
-
+
// The code below only handles extension cast instructions, so far.
// Enforce this.
if (LHSCI->getOpcode() != Instruction::ZExt &&
if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
// Not an extension from the same type?
RHSCIOp = CI->getOperand(0);
- if (RHSCIOp->getType() != LHSCIOp->getType())
+ if (RHSCIOp->getType() != LHSCIOp->getType())
return 0;
-
+
// If the signedness of the two casts doesn't agree (i.e. one is a sext
// and the other is a zext), then we can't handle this.
if (CI->getOpcode() != LHSCI->getOpcode())
return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
}
- // The re-extended constant changed so the constant cannot be represented
+ // The re-extended constant changed so the constant cannot be represented
// in the shorter type. Consequently, we cannot emit a simple comparison.
// All the cases that fold to true or false will have already been handled
// by SimplifyICmpInst, so only deal with the tricky case.
// llvm.sadd.with.overflow. To do this, we have to replace the original add
// with a narrower add, and discard the add-with-constant that is part of the
// range check (if we can't eliminate it, this isn't profitable).
-
+
// In order to eliminate the add-with-constant, the compare can be its only
// use.
Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
if (!AddWithCst->hasOneUse()) return 0;
-
+
// If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
if (!CI2->getValue().isPowerOf2()) return 0;
unsigned NewWidth = CI2->getValue().countTrailingZeros();
if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
-
+
// The width of the new add formed is 1 more than the bias.
++NewWidth;
-
+
// Check to see that CI1 is an all-ones value with NewWidth bits.
if (CI1->getBitWidth() == NewWidth ||
CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
return 0;
-
- // In order to replace the original add with a narrower
+
+ // This is only really a signed overflow check if the inputs have been
+ // sign-extended; check for that condition. For example, if CI2 is 2^31 and
+ // the operands of the add are 64 bits wide, we need at least 33 sign bits.
+ unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
+ if (IC.ComputeNumSignBits(A) < NeededSignBits ||
+ IC.ComputeNumSignBits(B) < NeededSignBits)
+ return 0;
+
+ // In order to replace the original add with a narrower
// llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
// and truncates that discard the high bits of the add. Verify that this is
// the case.
for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
UI != E; ++UI) {
if (*UI == AddWithCst) continue;
-
+
// Only accept truncates for now. We would really like a nice recursive
// predicate like SimplifyDemandedBits, but which goes downwards the use-def
// chain to see which bits of a value are actually demanded. If the
if (TI == 0 ||
TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
}
-
+
// If the pattern matches, truncate the inputs to the narrower type and
// use the sadd_with_overflow intrinsic to efficiently compute both the
// result and the overflow bit.
Module *M = I.getParent()->getParent()->getParent();
-
- const Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
+
+ Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
- &NewType, 1);
+ NewType);
InstCombiner::BuilderTy *Builder = IC.Builder;
-
+
// Put the new code above the original add, in case there are any uses of the
// add between the add and the compare.
Builder->SetInsertPoint(OrigAdd);
-
+
Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
-
+
// The inner add was the result of the narrow add, zero extended to the
// wider type. Replace it with the result computed by the intrinsic.
IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
-
+
// The original icmp gets replaced with the overflow value.
return ExtractValueInst::Create(Call, 1, "sadd.overflow");
}
// Don't bother doing this transformation for pointers, don't do it for
// vectors.
if (!isa<IntegerType>(OrigAddV->getType())) return 0;
-
+
// If the add is a constant expr, then we don't bother transforming it.
Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
if (OrigAdd == 0) return 0;
-
+
Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
-
+
// Put the new code above the original add, in case there are any uses of the
// add between the add and the compare.
InstCombiner::BuilderTy *Builder = IC.Builder;
Builder->SetInsertPoint(OrigAdd);
Module *M = I.getParent()->getParent()->getParent();
- const Type *Ty = LHS->getType();
- Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, &Ty,1);
+ Type *Ty = LHS->getType();
+ Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
Value *Add = Builder->CreateExtractValue(Call, 0);
unsigned BitWidth, bool isSignCheck) {
if (isSignCheck)
return APInt::getSignBit(BitWidth);
-
+
ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
if (!CI) return APInt::getAllOnesValue(BitWidth);
const APInt &RHS = CI->getValue();
-
+
switch (I.getPredicate()) {
- // For a UGT comparison, we don't care about any bits that
+ // For a UGT comparison, we don't care about any bits that
// correspond to the trailing ones of the comparand. The value of these
// bits doesn't impact the outcome of the comparison, because any value
// greater than the RHS must differ in a bit higher than these due to carry.
APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
return ~lowBitsSet;
}
-
+
// Similarly, for a ULT comparison, we don't care about the trailing zeros.
// Any value less than the RHS must differ in a higher bit because of carries.
case ICmpInst::ICMP_ULT: {
APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
return ~lowBitsSet;
}
-
+
default:
return APInt::getAllOnesValue(BitWidth);
}
-
+
}
Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
bool Changed = false;
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
-
+
/// Orders the operands of the compare so that they are listed from most
/// complex to least complex. This puts constants before unary operators,
/// before binary operators.
std::swap(Op0, Op1);
Changed = true;
}
-
+
if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
return ReplaceInstUsesWith(I, V);
-
- const Type *Ty = Op0->getType();
+
+ // comparing -val or val with non-zero is the same as just comparing val
+ // ie, abs(val) != 0 -> val != 0
+ if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero()))
+ {
+ Value *Cond, *SelectTrue, *SelectFalse;
+ if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
+ m_Value(SelectFalse)))) {
+ if (Value *V = dyn_castNegVal(SelectTrue)) {
+ if (V == SelectFalse)
+ return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
+ }
+ else if (Value *V = dyn_castNegVal(SelectFalse)) {
+ if (V == SelectTrue)
+ return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
+ }
+ }
+ }
+
+ Type *Ty = Op0->getType();
// icmp's with boolean values can always be turned into bitwise operations
if (Ty->isIntegerTy(1)) {
BitWidth = Ty->getScalarSizeInBits();
else if (TD) // Pointers require TD info to get their size.
BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
-
+
bool isSignBit = false;
// See if we are doing a comparison with a constant.
if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
Value *A = 0, *B = 0;
-
+
// Match the following pattern, which is a common idiom when writing
// overflow-safe integer arithmetic function. The source performs an
// addition in wider type, and explicitly checks for overflow using
// sadd_with_overflow intrinsic.
//
// TODO: This could probably be generalized to handle other overflow-safe
- // operations if we worked out the formulas to compute the appropriate
+ // operations if we worked out the formulas to compute the appropriate
// magic constants.
- //
+ //
// sum = a + b
// if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
{
if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
return Res;
}
-
+
// (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
if (I.isEquality() && CI->isZero() &&
match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
// (icmp cond A B) if cond is equality
return new ICmpInst(I.getPredicate(), A, B);
}
-
+
// If we have an icmp le or icmp ge instruction, turn it into the
// appropriate icmp lt or icmp gt instruction. This allows us to rely on
// them being folded in the code below. The SimplifyICmpInst code has
return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
ConstantInt::get(CI->getContext(), CI->getValue()-1));
}
-
+
// If this comparison is a normal comparison, it demands all
// bits, if it is a sign bit comparison, it only demands the sign bit.
bool UnusedBit;
case ICmpInst::ICMP_EQ: {
if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
-
+
// If all bits are known zero except for one, then we know at most one
// bit is set. If the comparison is against zero, then this is a check
// to see if *that* bit is set.
if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
LHSC->getValue() != Op0KnownZeroInverted)
LHS = Op0;
-
+
// If the LHS is 1 << x, and we know the result is a power of 2 like 8,
// then turn "((1 << x)&8) == 0" into "x != 3".
Value *X = 0;
return new ICmpInst(ICmpInst::ICMP_NE, X,
ConstantInt::get(X->getType(), CmpVal));
}
-
+
// If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
// then turn "((8 >>u x)&1) == 0" into "x != 3".
const APInt *CI;
ConstantInt::get(X->getType(),
CI->countTrailingZeros()));
}
-
+
break;
}
case ICmpInst::ICMP_NE: {
if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
-
+
// If all bits are known zero except for one, then we know at most one
// bit is set. If the comparison is against zero, then this is a check
// to see if *that* bit is set.
if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
LHSC->getValue() != Op0KnownZeroInverted)
LHS = Op0;
-
+
// If the LHS is 1 << x, and we know the result is a power of 2 like 8,
// then turn "((1 << x)&8) != 0" into "x == 3".
Value *X = 0;
return new ICmpInst(ICmpInst::ICMP_EQ, X,
ConstantInt::get(X->getType(), CmpVal));
}
-
+
// If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
// then turn "((8 >>u x)&1) != 0" into "x == 3".
const APInt *CI;
ConstantInt::get(X->getType(),
CI->countTrailingZeros()));
}
-
+
break;
}
case ICmpInst::ICMP_ULT:
// See if we are doing a comparison between a constant and an instruction that
// can be folded into the comparison.
if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
- // Since the RHS is a ConstantInt (CI), if the left hand side is an
- // instruction, see if that instruction also has constants so that the
- // instruction can be folded into the icmp
+ // Since the RHS is a ConstantInt (CI), if the left hand side is an
+ // instruction, see if that instruction also has constants so that the
+ // instruction can be folded into the icmp
if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
return Res;
case Instruction::IntToPtr:
// icmp pred inttoptr(X), null -> icmp pred X, 0
if (RHSC->isNullValue() && TD &&
- TD->getIntPtrType(RHSC->getContext()) ==
+ TD->getIntPtrType(RHSC->getContext()) ==
LHSI->getOperand(0)->getType())
return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
Constant::getNullValue(LHSI->getOperand(0)->getType()));
// values. If the ptr->ptr cast can be stripped off both arguments, we do so
// now.
if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
- if (Op0->getType()->isPointerTy() &&
- (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
+ if (Op0->getType()->isPointerTy() &&
+ (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
// We keep moving the cast from the left operand over to the right
// operand, where it can often be eliminated completely.
Op0 = CI->getOperand(0);
return new ICmpInst(I.getPredicate(), Op0, Op1);
}
}
-
+
if (isa<CastInst>(Op0)) {
// Handle the special case of: icmp (cast bool to X), <cst>
// This comes up when you have code like
// Try not to increase register pressure.
BO0->hasOneUse() && BO1->hasOneUse()) {
// Determine Y and Z in the form icmp (X+Y), (X+Z).
- Value *Y = (A == C || A == D) ? B : A;
- Value *Z = (C == A || C == B) ? D : C;
+ Value *Y, *Z;
+ if (A == C) {
+ // C + B == C + D -> B == D
+ Y = B;
+ Z = D;
+ } else if (A == D) {
+ // D + B == C + D -> B == C
+ Y = B;
+ Z = C;
+ } else if (B == C) {
+ // A + C == C + D -> A == D
+ Y = A;
+ Z = D;
+ } else {
+ assert(B == D);
+ // A + D == C + D -> A == C
+ Y = A;
+ Z = C;
+ }
return new ICmpInst(Pred, Y, Z);
}
return new ICmpInst(Pred, BO0->getOperand(0),
BO1->getOperand(0));
}
-
- if (CI->getValue().isMaxSignedValue()) {
+
+ if (CI->isMaxValue(true)) {
ICmpInst::Predicate Pred = I.isSigned()
? I.getUnsignedPredicate()
: I.getSignedPredicate();
// Mask = -1 >> count-trailing-zeros(Cst).
if (!CI->isZero() && !CI->isOne()) {
const APInt &AP = CI->getValue();
- ConstantInt *Mask = ConstantInt::get(I.getContext(),
+ ConstantInt *Mask = ConstantInt::get(I.getContext(),
APInt::getLowBitsSet(AP.getBitWidth(),
AP.getBitWidth() -
AP.countTrailingZeros()));
// fall-through
case Instruction::SDiv:
case Instruction::AShr:
- if (!BO0->isExact() && !BO1->isExact())
+ if (!BO0->isExact() || !BO1->isExact())
break;
return new ICmpInst(I.getPredicate(), BO0->getOperand(0),
BO1->getOperand(0));
}
}
}
-
+
{ Value *A, *B;
// ~x < ~y --> y < x
// ~x < cst --> ~cst < x
// (a+b) <u a --> llvm.uadd.with.overflow.
// (a+b) <u b --> llvm.uadd.with.overflow.
if (I.getPredicate() == ICmpInst::ICMP_ULT &&
- match(Op0, m_Add(m_Value(A), m_Value(B))) &&
+ match(Op0, m_Add(m_Value(A), m_Value(B))) &&
(Op1 == A || Op1 == B))
if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
return R;
-
+
// a >u (a+b) --> llvm.uadd.with.overflow.
// b >u (a+b) --> llvm.uadd.with.overflow.
if (I.getPredicate() == ICmpInst::ICMP_UGT &&
if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
return R;
}
-
+
if (I.isEquality()) {
Value *A, *B, *C, *D;
match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
Constant *NC = ConstantInt::get(I.getContext(),
C1->getValue() ^ C2->getValue());
- Value *Xor = Builder->CreateXor(C, NC, "tmp");
+ Value *Xor = Builder->CreateXor(C, NC);
return new ICmpInst(I.getPredicate(), A, Xor);
}
-
+
// A^B == A^D -> B == D
if (A == C) return new ICmpInst(I.getPredicate(), B, D);
if (A == D) return new ICmpInst(I.getPredicate(), B, C);
if (B == D) return new ICmpInst(I.getPredicate(), A, C);
}
}
-
+
if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
(A == Op0 || B == Op0)) {
// A == (A^B) -> B == 0
}
// (X&Z) == (Y&Z) -> (X^Y) & Z == 0
- if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
+ if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
Value *X = 0, *Y = 0, *Z = 0;
-
+
if (A == C) {
X = B; Y = D; Z = A;
} else if (A == D) {
} else if (B == D) {
X = A; Y = C; Z = B;
}
-
+
if (X) { // Build (X^Y) & Z
- Op1 = Builder->CreateXor(X, Y, "tmp");
- Op1 = Builder->CreateAnd(Op1, Z, "tmp");
+ Op1 = Builder->CreateXor(X, Y);
+ Op1 = Builder->CreateAnd(Op1, Z);
I.setOperand(0, Op1);
I.setOperand(1, Constant::getNullValue(Op1->getType()));
return &I;
}
}
-
+
+ // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
+ // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
+ ConstantInt *Cst1;
+ if ((Op0->hasOneUse() &&
+ match(Op0, m_ZExt(m_Value(A))) &&
+ match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
+ (Op1->hasOneUse() &&
+ match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
+ match(Op1, m_ZExt(m_Value(A))))) {
+ APInt Pow2 = Cst1->getValue() + 1;
+ if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
+ Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
+ return new ICmpInst(I.getPredicate(), A,
+ Builder->CreateTrunc(B, A->getType()));
+ }
+
// Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
// "icmp (and X, mask), cst"
uint64_t ShAmt = 0;
- ConstantInt *Cst1;
if (Op0->hasOneUse() &&
match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
m_ConstantInt(ShAmt))))) &&
// when it exposes other optimizations.
!A->hasOneUse()) {
unsigned ASize =cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
-
+
if (ShAmt < ASize) {
APInt MaskV =
APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
MaskV <<= ShAmt;
-
+
APInt CmpV = Cst1->getValue().zext(ASize);
CmpV <<= ShAmt;
-
+
Value *Mask = Builder->CreateAnd(A, Builder->getInt(MaskV));
return new ICmpInst(I.getPredicate(), Mask, Builder->getInt(CmpV));
}
}
}
-
+
{
Value *X; ConstantInt *Cst;
// icmp X+Cst, X
Constant *RHSC) {
if (!isa<ConstantFP>(RHSC)) return 0;
const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
-
+
// Get the width of the mantissa. We don't want to hack on conversions that
// might lose information from the integer, e.g. "i64 -> float"
int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
if (MantissaWidth == -1) return 0; // Unknown.
-
+
// Check to see that the input is converted from an integer type that is small
// enough that preserves all bits. TODO: check here for "known" sign bits.
// This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
-
+
// If this is a uitofp instruction, we need an extra bit to hold the sign.
bool LHSUnsigned = isa<UIToFPInst>(LHSI);
if (LHSUnsigned)
++InputSize;
-
+
// If the conversion would lose info, don't hack on this.
if ((int)InputSize > MantissaWidth)
return 0;
-
+
// Otherwise, we can potentially simplify the comparison. We know that it
// will always come through as an integer value and we know the constant is
// not a NAN (it would have been previously simplified).
assert(!RHS.isNaN() && "NaN comparison not already folded!");
-
+
ICmpInst::Predicate Pred;
switch (I.getPredicate()) {
default: llvm_unreachable("Unexpected predicate!");
case FCmpInst::FCMP_UNO:
return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
}
-
- const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
-
+
+ IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
+
// Now we know that the APFloat is a normal number, zero or inf.
-
+
// See if the FP constant is too large for the integer. For example,
// comparing an i8 to 300.0.
unsigned IntWidth = IntTy->getScalarSizeInBits();
-
+
if (!LHSUnsigned) {
// If the RHS value is > SignedMax, fold the comparison. This handles +INF
// and large values.
return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
}
}
-
+
if (!LHSUnsigned) {
// See if the RHS value is < SignedMin.
APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
}
+ } else {
+ // See if the RHS value is < UnsignedMin.
+ APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
+ SMin.convertFromAPInt(APInt::getMinValue(IntWidth), true,
+ APFloat::rmNearestTiesToEven);
+ if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // umin > 12312.0
+ if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
+ Pred == ICmpInst::ICMP_UGE)
+ return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
+ return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
+ }
}
// Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
case ICmpInst::ICMP_UGE:
// (float)int >= -4.4 --> true
// (float)int >= 4.4 --> int > 4
- if (!RHS.isNegative())
+ if (RHS.isNegative())
return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
Pred = ICmpInst::ICMP_UGT;
break;
Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
bool Changed = false;
-
+
/// Orders the operands of the compare so that they are listed from most
/// complex to least complex. This puts constants before unary operators,
/// before binary operators.
}
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
-
+
if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
return ReplaceInstUsesWith(I, V);
I.setPredicate(FCmpInst::FCMP_UNO);
I.setOperand(1, Constant::getNullValue(Op0->getType()));
return &I;
-
+
case FCmpInst::FCMP_ORD: // True if ordered (no nans)
case FCmpInst::FCMP_OEQ: // True if ordered and equal
case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
return &I;
}
}
-
+
// Handle fcmp with constant RHS
if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
if (!RHSF)
break;
- // We can't convert a PPC double double.
- if (RHSF->getType()->isPPC_FP128Ty())
- break;
-
const fltSemantics *Sem;
// FIXME: This shouldn't be here.
- if (LHSExt->getSrcTy()->isFloatTy())
+ if (LHSExt->getSrcTy()->isHalfTy())
+ Sem = &APFloat::IEEEhalf;
+ else if (LHSExt->getSrcTy()->isFloatTy())
Sem = &APFloat::IEEEsingle;
else if (LHSExt->getSrcTy()->isDoubleTy())
Sem = &APFloat::IEEEdouble;
Sem = &APFloat::IEEEquad;
else if (LHSExt->getSrcTy()->isX86_FP80Ty())
Sem = &APFloat::x87DoubleExtended;
+ else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
+ Sem = &APFloat::PPCDoubleDouble;
else
break;
APFloat F = RHSF->getValueAPF();
F.convert(*Sem, APFloat::rmNearestTiesToEven, &Lossy);
- // Avoid lossy conversions and denormals.
+ // Avoid lossy conversions and denormals. Zero is a special case
+ // that's OK to convert.
+ APFloat Fabs = F;
+ Fabs.clearSign();
if (!Lossy &&
- F.compare(APFloat::getSmallestNormalized(*Sem)) !=
- APFloat::cmpLessThan)
+ ((Fabs.compare(APFloat::getSmallestNormalized(*Sem)) !=
+ APFloat::cmpLessThan) || Fabs.isZero()))
+
return new FCmpInst(I.getPredicate(), LHSExt->getOperand(0),
ConstantFP::get(RHSC->getContext(), F));
break;
return Res;
}
break;
+ case Instruction::Call: {
+ CallInst *CI = cast<CallInst>(LHSI);
+ LibFunc::Func Func;
+ // Various optimization for fabs compared with zero.
+ if (RHSC->isNullValue() && CI->getCalledFunction() &&
+ TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
+ TLI->has(Func)) {
+ if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
+ Func == LibFunc::fabsl) {
+ switch (I.getPredicate()) {
+ default: break;
+ // fabs(x) < 0 --> false
+ case FCmpInst::FCMP_OLT:
+ return ReplaceInstUsesWith(I, Builder->getFalse());
+ // fabs(x) > 0 --> x != 0
+ case FCmpInst::FCMP_OGT:
+ return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
+ RHSC);
+ // fabs(x) <= 0 --> x == 0
+ case FCmpInst::FCMP_OLE:
+ return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
+ RHSC);
+ // fabs(x) >= 0 --> !isnan(x)
+ case FCmpInst::FCMP_OGE:
+ return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
+ RHSC);
+ // fabs(x) == 0 --> x == 0
+ // fabs(x) != 0 --> x != 0
+ case FCmpInst::FCMP_OEQ:
+ case FCmpInst::FCMP_UEQ:
+ case FCmpInst::FCMP_ONE:
+ case FCmpInst::FCMP_UNE:
+ return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
+ RHSC);
+ }
+ }
+ }
+ }
}
}
projects / oota-llvm.git / blobdiff