//===----------------------------------------------------------------------===//
#include "InstCombine.h"
-#include "llvm/Intrinsics.h"
#include "llvm/Analysis/InstructionSimplify.h"
+#include "llvm/IR/Intrinsics.h"
+#include "llvm/Support/ConstantRange.h"
#include "llvm/Support/PatternMatch.h"
+#include "llvm/Transforms/Utils/CmpInstAnalysis.h"
using namespace llvm;
using namespace PatternMatch;
/// AddOne - Add one to a ConstantInt.
-static Constant *AddOne(Constant *C) {
- return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
+static Constant *AddOne(ConstantInt *C) {
+ return ConstantInt::get(C->getContext(), C->getValue() + 1);
}
/// SubOne - Subtract one from a ConstantInt.
static Constant *SubOne(ConstantInt *C) {
// ~(~(X)) -> X.
if (BinaryOperator::isNot(V))
return true;
-
+
// Constants can be considered to be not'ed values.
if (isa<ConstantInt>(V))
return true;
-
+
// Compares can be inverted if they have a single use.
if (CmpInst *CI = dyn_cast<CmpInst>(V))
return CI->hasOneUse();
-
+
return false;
}
if (!isFreeToInvert(Operand))
return Operand;
}
-
+
// Constants can be considered to be not'ed values...
if (ConstantInt *C = dyn_cast<ConstantInt>(V))
return ConstantInt::get(C->getType(), ~C->getValue());
return 0;
}
-
-/// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
-/// are carefully arranged to allow folding of expressions such as:
-///
-/// (A < B) | (A > B) --> (A != B)
-///
-/// Note that this is only valid if the first and second predicates have the
-/// same sign. Is illegal to do: (A u< B) | (A s> B)
-///
-/// Three bits are used to represent the condition, as follows:
-/// 0 A > B
-/// 1 A == B
-/// 2 A < B
-///
-/// <=> Value Definition
-/// 000 0 Always false
-/// 001 1 A > B
-/// 010 2 A == B
-/// 011 3 A >= B
-/// 100 4 A < B
-/// 101 5 A != B
-/// 110 6 A <= B
-/// 111 7 Always true
-///
-static unsigned getICmpCode(const ICmpInst *ICI) {
- switch (ICI->getPredicate()) {
- // False -> 0
- case ICmpInst::ICMP_UGT: return 1; // 001
- case ICmpInst::ICMP_SGT: return 1; // 001
- case ICmpInst::ICMP_EQ: return 2; // 010
- case ICmpInst::ICMP_UGE: return 3; // 011
- case ICmpInst::ICMP_SGE: return 3; // 011
- case ICmpInst::ICMP_ULT: return 4; // 100
- case ICmpInst::ICMP_SLT: return 4; // 100
- case ICmpInst::ICMP_NE: return 5; // 101
- case ICmpInst::ICMP_ULE: return 6; // 110
- case ICmpInst::ICMP_SLE: return 6; // 110
- // True -> 7
- default:
- llvm_unreachable("Invalid ICmp predicate!");
- return 0;
- }
-}
-
/// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
/// predicate into a three bit mask. It also returns whether it is an ordered
/// predicate by reference.
default:
// Not expecting FCMP_FALSE and FCMP_TRUE;
llvm_unreachable("Unexpected FCmp predicate!");
- return 0;
}
}
-/// getICmpValue - This is the complement of getICmpCode, which turns an
-/// opcode and two operands into either a constant true or false, or a brand
+/// getNewICmpValue - This is the complement of getICmpCode, which turns an
+/// opcode and two operands into either a constant true or false, or a brand
/// new ICmp instruction. The sign is passed in to determine which kind
/// of predicate to use in the new icmp instruction.
-static Value *getICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS) {
- switch (Code) {
- default: assert(0 && "Illegal ICmp code!");
- case 0:
- return ConstantInt::getFalse(LHS->getContext());
- case 1:
- if (Sign)
- return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
- return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
- case 2:
- return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
- case 3:
- if (Sign)
- return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
- return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
- case 4:
- if (Sign)
- return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
- return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
- case 5:
- return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
- case 6:
- if (Sign)
- return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
- return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
- case 7:
- return ConstantInt::getTrue(LHS->getContext());
- }
+static Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
+ InstCombiner::BuilderTy *Builder) {
+ ICmpInst::Predicate NewPred;
+ if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred))
+ return NewConstant;
+ return Builder->CreateICmp(NewPred, LHS, RHS);
}
/// getFCmpValue - This is the complement of getFCmpCode, which turns an
/// opcode and two operands into either a FCmp instruction. isordered is passed
/// in to determine which kind of predicate to use in the new fcmp instruction.
static Value *getFCmpValue(bool isordered, unsigned code,
- Value *LHS, Value *RHS) {
+ Value *LHS, Value *RHS,
+ InstCombiner::BuilderTy *Builder) {
+ CmpInst::Predicate Pred;
switch (code) {
default: llvm_unreachable("Illegal FCmp code!");
- case 0:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
- case 1:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
- case 2:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
- case 3:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
- case 4:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
- case 5:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
- case 6:
- if (isordered)
- return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
- else
- return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
- case 7: return ConstantInt::getTrue(LHS->getContext());
+ case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break;
+ case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break;
+ case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break;
+ case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break;
+ case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break;
+ case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break;
+ case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
+ case 7:
+ if (!isordered) return ConstantInt::getTrue(LHS->getContext());
+ Pred = FCmpInst::FCMP_ORD; break;
}
-}
-
-/// PredicatesFoldable - Return true if both predicates match sign or if at
-/// least one of them is an equality comparison (which is signless).
-static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
- return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
- (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
- (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
+ return Builder->CreateFCmp(Pred, LHS, RHS);
}
// OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
}
break;
case Instruction::Or:
- if (Together == AndRHS) // (X | C) & C --> C
- return ReplaceInstUsesWith(TheAnd, AndRHS);
-
- if (Op->hasOneUse() && Together != OpRHS) {
- // (X | C1) & C2 --> (X | (C1&C2)) & C2
- Value *Or = Builder->CreateOr(X, Together);
- Or->takeName(Op);
- return BinaryOperator::CreateAnd(Or, AndRHS);
+ if (Op->hasOneUse()){
+ if (Together != OpRHS) {
+ // (X | C1) & C2 --> (X | (C1&C2)) & C2
+ Value *Or = Builder->CreateOr(X, Together);
+ Or->takeName(Op);
+ return BinaryOperator::CreateAnd(Or, AndRHS);
+ }
+
+ ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
+ if (TogetherCI && !TogetherCI->isZero()){
+ // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
+ // NOTE: This reduces the number of bits set in the & mask, which
+ // can expose opportunities for store narrowing.
+ Together = ConstantExpr::getXor(AndRHS, Together);
+ Value *And = Builder->CreateAnd(X, Together);
+ And->takeName(Op);
+ return BinaryOperator::CreateOr(And, OpRHS);
+ }
}
+
break;
case Instruction::Add:
if (Op->hasOneUse()) {
// Adding a one to a single bit bit-field should be turned into an XOR
// of the bit. First thing to check is to see if this AND is with a
// single bit constant.
- const APInt &AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
+ const APInt &AndRHSV = AndRHS->getValue();
// If there is only one bit set.
if (AndRHSV.isPowerOf2()) {
// Ok, at this point, we know that we are masking the result of the
// ADD down to exactly one bit. If the constant we are adding has
// no bits set below this bit, then we can eliminate the ADD.
- const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
+ const APInt& AddRHS = OpRHS->getValue();
// Check to see if any bits below the one bit set in AndRHSV are set.
if ((AddRHS & (AndRHSV-1)) == 0) {
uint32_t BitWidth = AndRHS->getType()->getBitWidth();
uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
- ConstantInt *CI = ConstantInt::get(AndRHS->getContext(),
- AndRHS->getValue() & ShlMask);
+ ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShlMask);
- if (CI->getValue() == ShlMask) {
- // Masking out bits that the shift already masks
+ if (CI->getValue() == ShlMask)
+ // Masking out bits that the shift already masks.
return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
- } else if (CI != AndRHS) { // Reducing bits set in and.
+
+ if (CI != AndRHS) { // Reducing bits set in and.
TheAnd.setOperand(1, CI);
return &TheAnd;
}
uint32_t BitWidth = AndRHS->getType()->getBitWidth();
uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
- ConstantInt *CI = ConstantInt::get(Op->getContext(),
- AndRHS->getValue() & ShrMask);
+ ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShrMask);
- if (CI->getValue() == ShrMask) {
- // Masking out bits that the shift already masks.
+ if (CI->getValue() == ShrMask)
+ // Masking out bits that the shift already masks.
return ReplaceInstUsesWith(TheAnd, Op);
- } else if (CI != AndRHS) {
+
+ if (CI != AndRHS) {
TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
return &TheAnd;
}
uint32_t BitWidth = AndRHS->getType()->getBitWidth();
uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
- Constant *C = ConstantInt::get(Op->getContext(),
- AndRHS->getValue() & ShrMask);
+ Constant *C = Builder->getInt(AndRHS->getValue() & ShrMask);
if (C == AndRHS) { // Masking out bits shifted in.
// (Val ashr C1) & C2 -> (Val lshr C1) & C2
// Make the argument unsigned.
return 0;
}
-
-/// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
-/// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
-/// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
+/// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
+/// (V < Lo || V >= Hi). In practice, we emit the more efficient
+/// (V-Lo) \<u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
/// whether to treat the V, Lo and HI as signed or not. IB is the location to
/// insert new instructions.
-Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
- bool isSigned, bool Inside,
- Instruction &IB) {
- assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
+Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
+ bool isSigned, bool Inside) {
+ assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
"Lo is not <= Hi in range emission code!");
-
+
if (Inside) {
if (Lo == Hi) // Trivially false.
- return new ICmpInst(ICmpInst::ICMP_NE, V, V);
+ return Builder->getFalse();
// V >= Min && V < Hi --> V < Hi
if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
- ICmpInst::Predicate pred = (isSigned ?
+ ICmpInst::Predicate pred = (isSigned ?
ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
- return new ICmpInst(pred, V, Hi);
+ return Builder->CreateICmp(pred, V, Hi);
}
// Emit V-Lo <u Hi-Lo
Constant *NegLo = ConstantExpr::getNeg(Lo);
Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
- return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
+ return Builder->CreateICmpULT(Add, UpperBound);
}
if (Lo == Hi) // Trivially true.
- return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
+ return Builder->getTrue();
// V < Min || V >= Hi -> V > Hi-1
Hi = SubOne(cast<ConstantInt>(Hi));
if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
- ICmpInst::Predicate pred = (isSigned ?
+ ICmpInst::Predicate pred = (isSigned ?
ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
- return new ICmpInst(pred, V, Hi);
+ return Builder->CreateICmp(pred, V, Hi);
}
// Emit V-Lo >u Hi-1-Lo
ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
- return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
+ return Builder->CreateICmpUGT(Add, LowerBound);
}
// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
// look for the first zero bit after the run of ones
MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
// look for the first non-zero bit
- ME = V.getActiveBits();
+ ME = V.getActiveBits();
return true;
}
/// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
/// where isSub determines whether the operator is a sub. If we can fold one of
/// the following xforms:
-///
+///
/// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
/// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
/// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
case Instruction::And:
if (ConstantExpr::getAnd(N, Mask) == Mask) {
// If the AndRHS is a power of two minus one (0+1+), this is simple.
- if ((Mask->getValue().countLeadingZeros() +
- Mask->getValue().countPopulation()) ==
+ if ((Mask->getValue().countLeadingZeros() +
+ Mask->getValue().countPopulation()) ==
Mask->getValue().getBitWidth())
break;
case Instruction::Or:
case Instruction::Xor:
// If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
- if ((Mask->getValue().countLeadingZeros() +
+ if ((Mask->getValue().countLeadingZeros() +
Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
&& ConstantExpr::getAnd(N, Mask)->isNullValue())
break;
return 0;
}
-
+
if (isSub)
return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
}
+/// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
+/// One of A and B is considered the mask, the other the value. This is
+/// described as the "AMask" or "BMask" part of the enum. If the enum
+/// contains only "Mask", then both A and B can be considered masks.
+/// If A is the mask, then it was proven, that (A & C) == C. This
+/// is trivial if C == A, or C == 0. If both A and C are constants, this
+/// proof is also easy.
+/// For the following explanations we assume that A is the mask.
+/// The part "AllOnes" declares, that the comparison is true only
+/// if (A & B) == A, or all bits of A are set in B.
+/// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
+/// The part "AllZeroes" declares, that the comparison is true only
+/// if (A & B) == 0, or all bits of A are cleared in B.
+/// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
+/// The part "Mixed" declares, that (A & B) == C and C might or might not
+/// contain any number of one bits and zero bits.
+/// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
+/// The Part "Not" means, that in above descriptions "==" should be replaced
+/// by "!=".
+/// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
+/// If the mask A contains a single bit, then the following is equivalent:
+/// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
+/// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
+enum MaskedICmpType {
+ FoldMskICmp_AMask_AllOnes = 1,
+ FoldMskICmp_AMask_NotAllOnes = 2,
+ FoldMskICmp_BMask_AllOnes = 4,
+ FoldMskICmp_BMask_NotAllOnes = 8,
+ FoldMskICmp_Mask_AllZeroes = 16,
+ FoldMskICmp_Mask_NotAllZeroes = 32,
+ FoldMskICmp_AMask_Mixed = 64,
+ FoldMskICmp_AMask_NotMixed = 128,
+ FoldMskICmp_BMask_Mixed = 256,
+ FoldMskICmp_BMask_NotMixed = 512
+};
+
+/// return the set of pattern classes (from MaskedICmpType)
+/// that (icmp SCC (A & B), C) satisfies
+static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
+ ICmpInst::Predicate SCC)
+{
+ ConstantInt *ACst = dyn_cast<ConstantInt>(A);
+ ConstantInt *BCst = dyn_cast<ConstantInt>(B);
+ ConstantInt *CCst = dyn_cast<ConstantInt>(C);
+ bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
+ bool icmp_abit = (ACst != 0 && !ACst->isZero() &&
+ ACst->getValue().isPowerOf2());
+ bool icmp_bbit = (BCst != 0 && !BCst->isZero() &&
+ BCst->getValue().isPowerOf2());
+ unsigned result = 0;
+ if (CCst != 0 && CCst->isZero()) {
+ // if C is zero, then both A and B qualify as mask
+ result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
+ FoldMskICmp_Mask_AllZeroes |
+ FoldMskICmp_AMask_Mixed |
+ FoldMskICmp_BMask_Mixed)
+ : (FoldMskICmp_Mask_NotAllZeroes |
+ FoldMskICmp_Mask_NotAllZeroes |
+ FoldMskICmp_AMask_NotMixed |
+ FoldMskICmp_BMask_NotMixed));
+ if (icmp_abit)
+ result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
+ FoldMskICmp_AMask_NotMixed)
+ : (FoldMskICmp_AMask_AllOnes |
+ FoldMskICmp_AMask_Mixed));
+ if (icmp_bbit)
+ result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
+ FoldMskICmp_BMask_NotMixed)
+ : (FoldMskICmp_BMask_AllOnes |
+ FoldMskICmp_BMask_Mixed));
+ return result;
+ }
+ if (A == C) {
+ result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
+ FoldMskICmp_AMask_Mixed)
+ : (FoldMskICmp_AMask_NotAllOnes |
+ FoldMskICmp_AMask_NotMixed));
+ if (icmp_abit)
+ result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
+ FoldMskICmp_AMask_NotMixed)
+ : (FoldMskICmp_Mask_AllZeroes |
+ FoldMskICmp_AMask_Mixed));
+ } else if (ACst != 0 && CCst != 0 &&
+ ConstantExpr::getAnd(ACst, CCst) == CCst) {
+ result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
+ : FoldMskICmp_AMask_NotMixed);
+ }
+ if (B == C) {
+ result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
+ FoldMskICmp_BMask_Mixed)
+ : (FoldMskICmp_BMask_NotAllOnes |
+ FoldMskICmp_BMask_NotMixed));
+ if (icmp_bbit)
+ result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
+ FoldMskICmp_BMask_NotMixed)
+ : (FoldMskICmp_Mask_AllZeroes |
+ FoldMskICmp_BMask_Mixed));
+ } else if (BCst != 0 && CCst != 0 &&
+ ConstantExpr::getAnd(BCst, CCst) == CCst) {
+ result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
+ : FoldMskICmp_BMask_NotMixed);
+ }
+ return result;
+}
+
+/// decomposeBitTestICmp - Decompose an icmp into the form ((X & Y) pred Z)
+/// if possible. The returned predicate is either == or !=. Returns false if
+/// decomposition fails.
+static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred,
+ Value *&X, Value *&Y, Value *&Z) {
+ // X < 0 is equivalent to (X & SignBit) != 0.
+ if (I->getPredicate() == ICmpInst::ICMP_SLT)
+ if (ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)))
+ if (C->isZero()) {
+ X = I->getOperand(0);
+ Y = ConstantInt::get(I->getContext(),
+ APInt::getSignBit(C->getBitWidth()));
+ Pred = ICmpInst::ICMP_NE;
+ Z = C;
+ return true;
+ }
+
+ // X > -1 is equivalent to (X & SignBit) == 0.
+ if (I->getPredicate() == ICmpInst::ICMP_SGT)
+ if (ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)))
+ if (C->isAllOnesValue()) {
+ X = I->getOperand(0);
+ Y = ConstantInt::get(I->getContext(),
+ APInt::getSignBit(C->getBitWidth()));
+ Pred = ICmpInst::ICMP_EQ;
+ Z = ConstantInt::getNullValue(C->getType());
+ return true;
+ }
+
+ return false;
+}
+
+/// foldLogOpOfMaskedICmpsHelper:
+/// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
+/// return the set of pattern classes (from MaskedICmpType)
+/// that both LHS and RHS satisfy
+static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
+ Value*& B, Value*& C,
+ Value*& D, Value*& E,
+ ICmpInst *LHS, ICmpInst *RHS,
+ ICmpInst::Predicate &LHSCC,
+ ICmpInst::Predicate &RHSCC) {
+ if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
+ // vectors are not (yet?) supported
+ if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
+
+ // Here comes the tricky part:
+ // LHS might be of the form L11 & L12 == X, X == L21 & L22,
+ // and L11 & L12 == L21 & L22. The same goes for RHS.
+ // Now we must find those components L** and R**, that are equal, so
+ // that we can extract the parameters A, B, C, D, and E for the canonical
+ // above.
+ Value *L1 = LHS->getOperand(0);
+ Value *L2 = LHS->getOperand(1);
+ Value *L11,*L12,*L21,*L22;
+ // Check whether the icmp can be decomposed into a bit test.
+ if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) {
+ L21 = L22 = L1 = 0;
+ } else {
+ // Look for ANDs in the LHS icmp.
+ if (match(L1, m_And(m_Value(L11), m_Value(L12)))) {
+ if (!match(L2, m_And(m_Value(L21), m_Value(L22))))
+ L21 = L22 = 0;
+ } else {
+ if (!match(L2, m_And(m_Value(L11), m_Value(L12))))
+ return 0;
+ std::swap(L1, L2);
+ L21 = L22 = 0;
+ }
+ }
+
+ // Bail if LHS was a icmp that can't be decomposed into an equality.
+ if (!ICmpInst::isEquality(LHSCC))
+ return 0;
+
+ Value *R1 = RHS->getOperand(0);
+ Value *R2 = RHS->getOperand(1);
+ Value *R11,*R12;
+ bool ok = false;
+ if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) {
+ if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
+ A = R11; D = R12;
+ } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
+ A = R12; D = R11;
+ } else {
+ return 0;
+ }
+ E = R2; R1 = 0; ok = true;
+ } else if (match(R1, m_And(m_Value(R11), m_Value(R12)))) {
+ if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
+ A = R11; D = R12; E = R2; ok = true;
+ } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
+ A = R12; D = R11; E = R2; ok = true;
+ }
+ }
+
+ // Bail if RHS was a icmp that can't be decomposed into an equality.
+ if (!ICmpInst::isEquality(RHSCC))
+ return 0;
+
+ // Look for ANDs in on the right side of the RHS icmp.
+ if (!ok && match(R2, m_And(m_Value(R11), m_Value(R12)))) {
+ if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
+ A = R11; D = R12; E = R1; ok = true;
+ } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
+ A = R12; D = R11; E = R1; ok = true;
+ } else {
+ return 0;
+ }
+ }
+ if (!ok)
+ return 0;
+
+ if (L11 == A) {
+ B = L12; C = L2;
+ } else if (L12 == A) {
+ B = L11; C = L2;
+ } else if (L21 == A) {
+ B = L22; C = L1;
+ } else if (L22 == A) {
+ B = L21; C = L1;
+ }
+
+ unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC);
+ unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC);
+ return left_type & right_type;
+}
+/// foldLogOpOfMaskedICmps:
+/// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
+/// into a single (icmp(A & X) ==/!= Y)
+static Value* foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS,
+ ICmpInst::Predicate NEWCC,
+ llvm::InstCombiner::BuilderTy* Builder) {
+ Value *A = 0, *B = 0, *C = 0, *D = 0, *E = 0;
+ ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
+ unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS,
+ LHSCC, RHSCC);
+ if (mask == 0) return 0;
+ assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) &&
+ "foldLogOpOfMaskedICmpsHelper must return an equality predicate.");
+
+ if (NEWCC == ICmpInst::ICMP_NE)
+ mask >>= 1; // treat "Not"-states as normal states
+
+ if (mask & FoldMskICmp_Mask_AllZeroes) {
+ // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
+ // -> (icmp eq (A & (B|D)), 0)
+ Value* newOr = Builder->CreateOr(B, D);
+ Value* newAnd = Builder->CreateAnd(A, newOr);
+ // we can't use C as zero, because we might actually handle
+ // (icmp ne (A & B), B) & (icmp ne (A & D), D)
+ // with B and D, having a single bit set
+ Value* zero = Constant::getNullValue(A->getType());
+ return Builder->CreateICmp(NEWCC, newAnd, zero);
+ }
+ if (mask & FoldMskICmp_BMask_AllOnes) {
+ // (icmp eq (A & B), B) & (icmp eq (A & D), D)
+ // -> (icmp eq (A & (B|D)), (B|D))
+ Value* newOr = Builder->CreateOr(B, D);
+ Value* newAnd = Builder->CreateAnd(A, newOr);
+ return Builder->CreateICmp(NEWCC, newAnd, newOr);
+ }
+ if (mask & FoldMskICmp_AMask_AllOnes) {
+ // (icmp eq (A & B), A) & (icmp eq (A & D), A)
+ // -> (icmp eq (A & (B&D)), A)
+ Value* newAnd1 = Builder->CreateAnd(B, D);
+ Value* newAnd = Builder->CreateAnd(A, newAnd1);
+ return Builder->CreateICmp(NEWCC, newAnd, A);
+ }
+ if (mask & FoldMskICmp_BMask_Mixed) {
+ // (icmp eq (A & B), C) & (icmp eq (A & D), E)
+ // We already know that B & C == C && D & E == E.
+ // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
+ // C and E, which are shared by both the mask B and the mask D, don't
+ // contradict, then we can transform to
+ // -> (icmp eq (A & (B|D)), (C|E))
+ // Currently, we only handle the case of B, C, D, and E being constant.
+ ConstantInt *BCst = dyn_cast<ConstantInt>(B);
+ if (BCst == 0) return 0;
+ ConstantInt *DCst = dyn_cast<ConstantInt>(D);
+ if (DCst == 0) return 0;
+ // we can't simply use C and E, because we might actually handle
+ // (icmp ne (A & B), B) & (icmp eq (A & D), D)
+ // with B and D, having a single bit set
+
+ ConstantInt *CCst = dyn_cast<ConstantInt>(C);
+ if (CCst == 0) return 0;
+ if (LHSCC != NEWCC)
+ CCst = dyn_cast<ConstantInt>( ConstantExpr::getXor(BCst, CCst) );
+ ConstantInt *ECst = dyn_cast<ConstantInt>(E);
+ if (ECst == 0) return 0;
+ if (RHSCC != NEWCC)
+ ECst = dyn_cast<ConstantInt>( ConstantExpr::getXor(DCst, ECst) );
+ ConstantInt* MCst = dyn_cast<ConstantInt>(
+ ConstantExpr::getAnd(ConstantExpr::getAnd(BCst, DCst),
+ ConstantExpr::getXor(CCst, ECst)) );
+ // if there is a conflict we should actually return a false for the
+ // whole construct
+ if (!MCst->isZero())
+ return 0;
+ Value *newOr1 = Builder->CreateOr(B, D);
+ Value *newOr2 = ConstantExpr::getOr(CCst, ECst);
+ Value *newAnd = Builder->CreateAnd(A, newOr1);
+ return Builder->CreateICmp(NEWCC, newAnd, newOr2);
+ }
+ return 0;
+}
+
/// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
-Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
- ICmpInst *LHS, ICmpInst *RHS) {
+Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
// (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
bool isSigned = LHS->isSigned() || RHS->isSigned();
- Value *RV = getICmpValue(isSigned, Code, Op0, Op1);
- if (Instruction *I = dyn_cast<Instruction>(RV))
- return I;
- // Otherwise, it's a constant boolean value.
- return ReplaceInstUsesWith(I, RV);
+ return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
}
}
-
+
+ // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
+ if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_EQ, Builder))
+ return V;
+
// This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
if (LHSCst == 0 || RHSCst == 0) return 0;
-
+
if (LHSCst == RHSCst && LHSCC == RHSCC) {
// (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
// where C is a power of 2
if (LHSCC == ICmpInst::ICMP_ULT &&
LHSCst->getValue().isPowerOf2()) {
Value *NewOr = Builder->CreateOr(Val, Val2);
- return new ICmpInst(LHSCC, NewOr, LHSCst);
+ return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
}
-
+
// (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
Value *NewOr = Builder->CreateOr(Val, Val2);
- return new ICmpInst(LHSCC, NewOr, LHSCst);
+ return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
+ }
+ }
+
+ // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
+ // where CMAX is the all ones value for the truncated type,
+ // iff the lower bits of C2 and CA are zero.
+ if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC &&
+ LHS->hasOneUse() && RHS->hasOneUse()) {
+ Value *V;
+ ConstantInt *AndCst, *SmallCst = 0, *BigCst = 0;
+
+ // (trunc x) == C1 & (and x, CA) == C2
+ // (and x, CA) == C2 & (trunc x) == C1
+ if (match(Val2, m_Trunc(m_Value(V))) &&
+ match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
+ SmallCst = RHSCst;
+ BigCst = LHSCst;
+ } else if (match(Val, m_Trunc(m_Value(V))) &&
+ match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
+ SmallCst = LHSCst;
+ BigCst = RHSCst;
+ }
+
+ if (SmallCst && BigCst) {
+ unsigned BigBitSize = BigCst->getType()->getBitWidth();
+ unsigned SmallBitSize = SmallCst->getType()->getBitWidth();
+
+ // Check that the low bits are zero.
+ APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
+ if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) {
+ Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue());
+ APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue();
+ Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N);
+ return Builder->CreateICmp(LHSCC, NewAnd, NewVal);
+ }
}
}
-
+
// From here on, we only handle:
// (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
if (Val != Val2) return 0;
-
+
// ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
return 0;
-
+
+ // Make a constant range that's the intersection of the two icmp ranges.
+ // If the intersection is empty, we know that the result is false.
+ ConstantRange LHSRange =
+ ConstantRange::makeICmpRegion(LHSCC, LHSCst->getValue());
+ ConstantRange RHSRange =
+ ConstantRange::makeICmpRegion(RHSCC, RHSCst->getValue());
+
+ if (LHSRange.intersectWith(RHSRange).isEmptySet())
+ return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
+
// We can't fold (ugt x, C) & (sgt x, C2).
if (!PredicatesFoldable(LHSCC, RHSCC))
return 0;
-
+
// Ensure that the larger constant is on the RHS.
bool ShouldSwap;
if (CmpInst::isSigned(LHSCC) ||
- (ICmpInst::isEquality(LHSCC) &&
+ (ICmpInst::isEquality(LHSCC) &&
CmpInst::isSigned(RHSCC)))
ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
else
ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
-
+
if (ShouldSwap) {
std::swap(LHS, RHS);
std::swap(LHSCst, RHSCst);
// At this point, we know we have two icmp instructions
// comparing a value against two constants and and'ing the result
// together. Because of the above check, we know that we only have
- // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
- // (from the icmp folding check above), that the two constants
+ // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
+ // (from the icmp folding check above), that the two constants
// are not equal and that the larger constant is on the RHS
assert(LHSCst != RHSCst && "Compares not folded above?");
case ICmpInst::ICMP_EQ:
switch (RHSCC) {
default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
- case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
- case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
- return ReplaceInstUsesWith(I, LHS);
+ return LHS;
}
case ICmpInst::ICMP_NE:
switch (RHSCC) {
default: llvm_unreachable("Unknown integer condition code!");
case ICmpInst::ICMP_ULT:
if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
- return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
+ return Builder->CreateICmpULT(Val, LHSCst);
break; // (X != 13 & X u< 15) -> no change
case ICmpInst::ICMP_SLT:
if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
- return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
+ return Builder->CreateICmpSLT(Val, LHSCst);
break; // (X != 13 & X s< 15) -> no change
case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
- return ReplaceInstUsesWith(I, RHS);
+ return RHS;
case ICmpInst::ICMP_NE:
if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
Constant *AddCST = ConstantExpr::getNeg(LHSCst);
Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
- return new ICmpInst(ICmpInst::ICMP_UGT, Add,
- ConstantInt::get(Add->getType(), 1));
+ return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1));
}
break; // (X != 13 & X != 15) -> no change
}
default: llvm_unreachable("Unknown integer condition code!");
case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
+ return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
break;
case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
- return ReplaceInstUsesWith(I, LHS);
+ return LHS;
case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
break;
}
case ICmpInst::ICMP_SLT:
switch (RHSCC) {
default: llvm_unreachable("Unknown integer condition code!");
- case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
- case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
break;
case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
- return ReplaceInstUsesWith(I, LHS);
+ return LHS;
case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
break;
}
default: llvm_unreachable("Unknown integer condition code!");
case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
- return ReplaceInstUsesWith(I, RHS);
+ return RHS;
case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
break;
case ICmpInst::ICMP_NE:
if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
- return new ICmpInst(LHSCC, Val, RHSCst);
+ return Builder->CreateICmp(LHSCC, Val, RHSCst);
break; // (X u> 13 & X != 15) -> no change
case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
- return InsertRangeTest(Val, AddOne(LHSCst),
- RHSCst, false, true, I);
+ return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
break;
}
default: llvm_unreachable("Unknown integer condition code!");
case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
- return ReplaceInstUsesWith(I, RHS);
+ return RHS;
case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
break;
case ICmpInst::ICMP_NE:
if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
- return new ICmpInst(LHSCC, Val, RHSCst);
+ return Builder->CreateICmp(LHSCC, Val, RHSCst);
break; // (X s> 13 & X != 15) -> no change
case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
- return InsertRangeTest(Val, AddOne(LHSCst),
- RHSCst, true, true, I);
+ return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
break;
}
break;
}
-
+
return 0;
}
-Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
- FCmpInst *RHS) {
-
+/// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of
+/// instcombine, this returns a Value which should already be inserted into the
+/// function.
+Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
RHS->getPredicate() == FCmpInst::FCMP_ORD) {
+ if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType())
+ return 0;
+
// (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
// If either of the constants are nans, then the whole thing returns
// false.
if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
- return new FCmpInst(FCmpInst::FCMP_ORD,
- LHS->getOperand(0), RHS->getOperand(0));
+ return Builder->getFalse();
+ return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
}
-
+
// Handle vector zeros. This occurs because the canonical form of
// "fcmp ord x,x" is "fcmp ord x, 0".
if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
isa<ConstantAggregateZero>(RHS->getOperand(1)))
- return new FCmpInst(FCmpInst::FCMP_ORD,
- LHS->getOperand(0), RHS->getOperand(0));
+ return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
return 0;
}
-
+
Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
-
-
+
+
if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
// Swap RHS operands to match LHS.
Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
std::swap(Op1LHS, Op1RHS);
}
-
+
if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
// Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
if (Op0CC == Op1CC)
- return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
-
+ return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
+ return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
if (Op0CC == FCmpInst::FCMP_TRUE)
- return ReplaceInstUsesWith(I, RHS);
+ return RHS;
if (Op1CC == FCmpInst::FCMP_TRUE)
- return ReplaceInstUsesWith(I, LHS);
-
+ return LHS;
+
bool Op0Ordered;
bool Op1Ordered;
unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
+ // uno && ord -> false
+ if (Op0Pred == 0 && Op1Pred == 0 && Op0Ordered != Op1Ordered)
+ return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
if (Op1Pred == 0) {
std::swap(LHS, RHS);
std::swap(Op0Pred, Op1Pred);
std::swap(Op0Ordered, Op1Ordered);
}
if (Op0Pred == 0) {
- // uno && ueq -> uno && (uno || eq) -> ueq
+ // uno && ueq -> uno && (uno || eq) -> uno
// ord && olt -> ord && (ord && lt) -> olt
- if (Op0Ordered == Op1Ordered)
- return ReplaceInstUsesWith(I, RHS);
-
+ if (!Op0Ordered && (Op0Ordered == Op1Ordered))
+ return LHS;
+ if (Op0Ordered && (Op0Ordered == Op1Ordered))
+ return RHS;
+
// uno && oeq -> uno && (ord && eq) -> false
- // uno && ord -> false
if (!Op0Ordered)
- return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
+ return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
// ord && ueq -> ord && (uno || eq) -> oeq
- return cast<Instruction>(getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS));
+ return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
}
}
Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
- bool Changed = SimplifyCommutative(I);
+ bool Changed = SimplifyAssociativeOrCommutative(I);
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
if (Value *V = SimplifyAndInst(Op0, Op1, TD))
return ReplaceInstUsesWith(I, V);
- // See if we can simplify any instructions used by the instruction whose sole
+ // (A|B)&(A|C) -> A|(B&C) etc
+ if (Value *V = SimplifyUsingDistributiveLaws(I))
+ return ReplaceInstUsesWith(I, V);
+
+ // See if we can simplify any instructions used by the instruction whose sole
// purpose is to compute bits we don't care about.
if (SimplifyDemandedInstructionBits(I))
- return &I;
+ return &I;
if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
const APInt &AndRHSMask = AndRHS->getValue();
- APInt NotAndRHS(~AndRHSMask);
// Optimize a variety of ((val OP C1) & C2) combinations...
if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
switch (Op0I->getOpcode()) {
default: break;
case Instruction::Xor:
- case Instruction::Or:
+ case Instruction::Or: {
// If the mask is only needed on one incoming arm, push it up.
if (!Op0I->hasOneUse()) break;
-
+
+ APInt NotAndRHS(~AndRHSMask);
if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
// Not masking anything out for the LHS, move to RHS.
Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
}
break;
+ }
case Instruction::Add:
// ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
// ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
// (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
// has 1's for all bits that the subtraction with A might affect.
- if (Op0I->hasOneUse()) {
+ if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
uint32_t BitWidth = AndRHSMask.getBitWidth();
uint32_t Zeros = AndRHSMask.countLeadingZeros();
APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
- ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
- if (!(A && A->isZero()) && // avoid infinite recursion.
- MaskedValueIsZero(Op0LHS, Mask)) {
+ if (MaskedValueIsZero(Op0LHS, Mask)) {
Value *NewNeg = Builder->CreateNeg(Op0RHS);
return BinaryOperator::CreateAnd(NewNeg, AndRHS);
}
if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
return Res;
- } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
- // If this is an integer truncation or change from signed-to-unsigned, and
- // if the source is an and/or with immediate, transform it. This
- // frequently occurs for bitfield accesses.
- if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
- if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
- CastOp->getNumOperands() == 2)
- if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1))){
- if (CastOp->getOpcode() == Instruction::And) {
- // Change: and (cast (and X, C1) to T), C2
- // into : and (cast X to T), trunc_or_bitcast(C1)&C2
- // This will fold the two constants together, which may allow
- // other simplifications.
- Value *NewCast = Builder->CreateTruncOrBitCast(
- CastOp->getOperand(0), I.getType(),
- CastOp->getName()+".shrunk");
- // trunc_or_bitcast(C1)&C2
- Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
- C3 = ConstantExpr::getAnd(C3, AndRHS);
- return BinaryOperator::CreateAnd(NewCast, C3);
- } else if (CastOp->getOpcode() == Instruction::Or) {
- // Change: and (cast (or X, C1) to T), C2
- // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
- Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
- if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
- // trunc(C1)&C2
- return ReplaceInstUsesWith(I, AndRHS);
- }
- }
+ }
+
+ // If this is an integer truncation, and if the source is an 'and' with
+ // immediate, transform it. This frequently occurs for bitfield accesses.
+ {
+ Value *X = 0; ConstantInt *YC = 0;
+ if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
+ // Change: and (trunc (and X, YC) to T), C2
+ // into : and (trunc X to T), trunc(YC) & C2
+ // This will fold the two constants together, which may allow
+ // other simplifications.
+ Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
+ Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
+ C3 = ConstantExpr::getAnd(C3, AndRHS);
+ return BinaryOperator::CreateAnd(NewCast, C3);
}
}
match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
((A == C && B == D) || (A == D && B == C)))
return BinaryOperator::CreateXor(A, B);
-
+
// ~(A&B) & (A|B) -> A^B
if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
((A == C && B == D) || (A == D && B == C)))
return BinaryOperator::CreateXor(A, B);
-
- if (Op0->hasOneUse() &&
- match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
- if (A == Op1) { // (A^B)&A -> A&(A^B)
- I.swapOperands(); // Simplify below
- std::swap(Op0, Op1);
- } else if (B == Op1) { // (A^B)&B -> B&(B^A)
- cast<BinaryOperator>(Op0)->swapOperands();
- I.swapOperands(); // Simplify below
- std::swap(Op0, Op1);
+
+ // A&(A^B) => A & ~B
+ {
+ Value *tmpOp0 = Op0;
+ Value *tmpOp1 = Op1;
+ if (Op0->hasOneUse() &&
+ match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
+ if (A == Op1 || B == Op1 ) {
+ tmpOp1 = Op0;
+ tmpOp0 = Op1;
+ // Simplify below
+ }
}
- }
- if (Op1->hasOneUse() &&
- match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
- if (B == Op0) { // B&(A^B) -> B&(B^A)
- cast<BinaryOperator>(Op1)->swapOperands();
- std::swap(A, B);
+ if (tmpOp1->hasOneUse() &&
+ match(tmpOp1, m_Xor(m_Value(A), m_Value(B)))) {
+ if (B == tmpOp0) {
+ std::swap(A, B);
+ }
+ // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if
+ // A is originally -1 (or a vector of -1 and undefs), then we enter
+ // an endless loop. By checking that A is non-constant we ensure that
+ // we will never get to the loop.
+ if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
+ return BinaryOperator::CreateAnd(A, Builder->CreateNot(B));
}
- if (A == Op0) // A&(A^B) -> A & ~B
- return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
}
// (A&((~A)|B)) -> A&B
match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
return BinaryOperator::CreateAnd(A, Op0);
}
-
+
if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
- if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
- return Res;
-
+ if (Value *Res = FoldAndOfICmps(LHS, RHS))
+ return ReplaceInstUsesWith(I, Res);
+
// If and'ing two fcmp, try combine them into one.
if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
- if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
- return Res;
-
-
+ if (Value *Res = FoldAndOfFCmps(LHS, RHS))
+ return ReplaceInstUsesWith(I, Res);
+
+
// fold (and (cast A), (cast B)) -> (cast (and A, B))
if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
- const Type *SrcTy = Op0C->getOperand(0)->getType();
+ Type *SrcTy = Op0C->getOperand(0)->getType();
if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
SrcTy == Op1C->getOperand(0)->getType() &&
SrcTy->isIntOrIntVectorTy()) {
Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
-
+
// Only do this if the casts both really cause code to be generated.
if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
}
-
+
// If this is and(cast(icmp), cast(icmp)), try to fold this even if the
// cast is otherwise not optimizable. This happens for vector sexts.
if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
- if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS)) {
- InsertNewInstBefore(Res, I);
+ if (Value *Res = FoldAndOfICmps(LHS, RHS))
return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
- }
-
+
// If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
// cast is otherwise not optimizable. This happens for vector sexts.
if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
- if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS)) {
- InsertNewInstBefore(Res, I);
+ if (Value *Res = FoldAndOfFCmps(LHS, RHS))
return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
- }
}
}
-
+
// (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
- if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
+ if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
SI0->getOperand(1) == SI1->getOperand(1) &&
(SI0->hasOneUse() || SI1->hasOneUse())) {
Value *NewOp =
Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
SI0->getName());
- return BinaryOperator::Create(SI1->getOpcode(), NewOp,
+ return BinaryOperator::Create(SI1->getOpcode(), NewOp,
SI1->getOperand(1));
}
}
+ {
+ Value *X = 0;
+ bool OpsSwapped = false;
+ // Canonicalize SExt or Not to the LHS
+ if (match(Op1, m_SExt(m_Value())) ||
+ match(Op1, m_Not(m_Value()))) {
+ std::swap(Op0, Op1);
+ OpsSwapped = true;
+ }
+
+ // Fold (and (sext bool to A), B) --> (select bool, B, 0)
+ if (match(Op0, m_SExt(m_Value(X))) &&
+ X->getType()->getScalarType()->isIntegerTy(1)) {
+ Value *Zero = Constant::getNullValue(Op1->getType());
+ return SelectInst::Create(X, Op1, Zero);
+ }
+
+ // Fold (and ~(sext bool to A), B) --> (select bool, 0, B)
+ if (match(Op0, m_Not(m_SExt(m_Value(X)))) &&
+ X->getType()->getScalarType()->isIntegerTy(1)) {
+ Value *Zero = Constant::getNullValue(Op0->getType());
+ return SelectInst::Create(X, Zero, Op1);
+ }
+
+ if (OpsSwapped)
+ std::swap(Op0, Op1);
+ }
+
return Changed ? &I : 0;
}
CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
ByteValues);
}
-
+
// If this is a logical shift by a constant multiple of 8, recurse with
// OverallLeftShift and ByteMask adjusted.
if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
- unsigned ShAmt =
+ unsigned ShAmt =
cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
// Ensure the shift amount is defined and of a byte value.
if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
if (OverallLeftShift >= (int)ByteValues.size()) return true;
if (OverallLeftShift <= -(int)ByteValues.size()) return true;
- return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
+ return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
ByteValues);
}
unsigned NumBytes = ByteValues.size();
APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
-
+
for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
// If this byte is masked out by a later operation, we don't care what
// the and mask is.
if ((ByteMask & (1 << i)) == 0)
continue;
-
+
// If the AndMask is all zeros for this byte, clear the bit.
APInt MaskB = AndMask & Byte;
if (MaskB == 0) {
ByteMask &= ~(1U << i);
continue;
}
-
+
// If the AndMask is not all ones for this byte, it's not a bytezap.
if (MaskB != Byte)
return true;
// Otherwise, this byte is kept.
}
- return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
+ return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
ByteValues);
}
}
-
+
// Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
// the input value to the bswap. Some observations: 1) if more than one byte
// is demanded from this input, then it could not be successfully assembled
// into a byteswap. At least one of the two bytes would not be aligned with
// their ultimate destination.
if (!isPowerOf2_32(ByteMask)) return true;
- unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
-
+ unsigned InputByteNo = countTrailingZeros(ByteMask);
+
// 2) The input and ultimate destinations must line up: if byte 3 of an i32
// is demanded, it needs to go into byte 0 of the result. This means that the
// byte needs to be shifted until it lands in the right byte bucket. The
// part of the value (e.g. byte 3) then it must be shifted right. If from the
// low part, it must be shifted left.
unsigned DestByteNo = InputByteNo + OverallLeftShift;
- if (InputByteNo < ByteValues.size()/2) {
- if (ByteValues.size()-1-DestByteNo != InputByteNo)
- return true;
- } else {
- if (ByteValues.size()-1-DestByteNo != InputByteNo)
- return true;
- }
-
+ if (ByteValues.size()-1-DestByteNo != InputByteNo)
+ return true;
+
// If the destination byte value is already defined, the values are or'd
// together, which isn't a bswap (unless it's an or of the same bits).
if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
/// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
/// If so, insert the new bswap intrinsic and return it.
Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
- const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
- if (!ITy || ITy->getBitWidth() % 16 ||
+ IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
+ if (!ITy || ITy->getBitWidth() % 16 ||
// ByteMask only allows up to 32-byte values.
- ITy->getBitWidth() > 32*8)
+ ITy->getBitWidth() > 32*8)
return 0; // Can only bswap pairs of bytes. Can't do vectors.
-
+
/// ByteValues - For each byte of the result, we keep track of which value
/// defines each byte.
SmallVector<Value*, 8> ByteValues;
ByteValues.resize(ITy->getBitWidth()/8);
-
+
// Try to find all the pieces corresponding to the bswap.
uint32_t ByteMask = ~0U >> (32-ByteValues.size());
if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
return 0;
-
+
// Check to see if all of the bytes come from the same value.
Value *V = ByteValues[0];
if (V == 0) return 0; // Didn't find a byte? Must be zero.
-
+
// Check to make sure that all of the bytes come from the same value.
for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
if (ByteValues[i] != V)
return 0;
- const Type *Tys[] = { ITy };
Module *M = I.getParent()->getParent()->getParent();
- Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
+ Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
return CallInst::Create(F, V);
}
return SelectInst::Create(Cond, C, B);
if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
return SelectInst::Create(Cond, C, B);
-
+
// ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
return SelectInst::Create(Cond, C, D);
}
/// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
-Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
- ICmpInst *LHS, ICmpInst *RHS) {
+Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
// (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
bool isSigned = LHS->isSigned() || RHS->isSigned();
- Value *RV = getICmpValue(isSigned, Code, Op0, Op1);
- if (Instruction *I = dyn_cast<Instruction>(RV))
- return I;
- // Otherwise, it's a constant boolean value.
- return ReplaceInstUsesWith(I, RV);
+ return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
}
}
-
- // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
+
+ // handle (roughly):
+ // (icmp ne (A & B), C) | (icmp ne (A & D), E)
+ if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_NE, Builder))
+ return V;
+
Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
+
+ if (LHS->hasOneUse() || RHS->hasOneUse()) {
+ // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
+ // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
+ Value *A = 0, *B = 0;
+ if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) {
+ B = Val;
+ if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1))
+ A = Val2;
+ else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2)
+ A = RHS->getOperand(1);
+ }
+ // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
+ // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
+ else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
+ B = Val2;
+ if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1))
+ A = Val;
+ else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val)
+ A = LHS->getOperand(1);
+ }
+ if (A && B)
+ return Builder->CreateICmp(
+ ICmpInst::ICMP_UGE,
+ Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
+ }
+
+ // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
if (LHSCst == 0 || RHSCst == 0) return 0;
- // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
- if (LHSCst == RHSCst && LHSCC == RHSCC &&
- LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
- Value *NewOr = Builder->CreateOr(Val, Val2);
- return new ICmpInst(LHSCC, NewOr, LHSCst);
+ if (LHSCst == RHSCst && LHSCC == RHSCC) {
+ // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
+ if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
+ Value *NewOr = Builder->CreateOr(Val, Val2);
+ return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
+ }
+ }
+
+ // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
+ // iff C2 + CA == C1.
+ if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
+ ConstantInt *AddCst;
+ if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
+ if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
+ return Builder->CreateICmpULE(Val, LHSCst);
}
-
+
// From here on, we only handle:
// (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
if (Val != Val2) return 0;
-
+
// ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
return 0;
-
+
// We can't fold (ugt x, C) | (sgt x, C2).
if (!PredicatesFoldable(LHSCC, RHSCC))
return 0;
-
+
// Ensure that the larger constant is on the RHS.
bool ShouldSwap;
if (CmpInst::isSigned(LHSCC) ||
- (ICmpInst::isEquality(LHSCC) &&
+ (ICmpInst::isEquality(LHSCC) &&
CmpInst::isSigned(RHSCC)))
ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
else
ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
-
+
if (ShouldSwap) {
std::swap(LHS, RHS);
std::swap(LHSCst, RHSCst);
std::swap(LHSCC, RHSCC);
}
-
+
// At this point, we know we have two icmp instructions
// comparing a value against two constants and or'ing the result
// together. Because of the above check, we know that we only have
switch (RHSCC) {
default: llvm_unreachable("Unknown integer condition code!");
case ICmpInst::ICMP_EQ:
+ if (LHS->getOperand(0) == RHS->getOperand(0)) {
+ // if LHSCst and RHSCst differ only by one bit:
+ // (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1
+ assert(LHSCst->getValue().ule(LHSCst->getValue()));
+
+ APInt Xor = LHSCst->getValue() ^ RHSCst->getValue();
+ if (Xor.isPowerOf2()) {
+ Value *NegCst = Builder->getInt(~Xor);
+ Value *And = Builder->CreateAnd(LHS->getOperand(0), NegCst);
+ return Builder->CreateICmp(ICmpInst::ICMP_EQ, And, LHSCst);
+ }
+ }
+
if (LHSCst == SubOne(RHSCst)) {
// (X == 13 | X == 14) -> X-13 <u 2
Constant *AddCST = ConstantExpr::getNeg(LHSCst);
Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
- return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
+ return Builder->CreateICmpULT(Add, AddCST);
}
+
break; // (X == 13 | X == 15) -> no change
case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
- return ReplaceInstUsesWith(I, RHS);
+ return RHS;
}
break;
case ICmpInst::ICMP_NE:
case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
- return ReplaceInstUsesWith(I, LHS);
+ return LHS;
case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
- return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
+ return Builder->getTrue();
}
- break;
case ICmpInst::ICMP_ULT:
switch (RHSCC) {
default: llvm_unreachable("Unknown integer condition code!");
// If RHSCst is [us]MAXINT, it is always false. Not handling
// this can cause overflow.
if (RHSCst->isMaxValue(false))
- return ReplaceInstUsesWith(I, LHS);
- return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
- false, false, I);
+ return LHS;
+ return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
break;
case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
- return ReplaceInstUsesWith(I, RHS);
+ return RHS;
case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
break;
}
// If RHSCst is [us]MAXINT, it is always false. Not handling
// this can cause overflow.
if (RHSCst->isMaxValue(true))
- return ReplaceInstUsesWith(I, LHS);
- return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
- true, false, I);
+ return LHS;
+ return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
break;
case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
- return ReplaceInstUsesWith(I, RHS);
+ return RHS;
case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
break;
}
default: llvm_unreachable("Unknown integer condition code!");
case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
- return ReplaceInstUsesWith(I, LHS);
+ return LHS;
case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
break;
case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
- return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
+ return Builder->getTrue();
case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
break;
}
default: llvm_unreachable("Unknown integer condition code!");
case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
- return ReplaceInstUsesWith(I, LHS);
+ return LHS;
case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
break;
case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
- return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
+ return Builder->getTrue();
case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
break;
}
return 0;
}
-Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
- FCmpInst *RHS) {
+/// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of
+/// instcombine, this returns a Value which should already be inserted into the
+/// function.
+Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
- RHS->getPredicate() == FCmpInst::FCMP_UNO &&
+ RHS->getPredicate() == FCmpInst::FCMP_UNO &&
LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
// If either of the constants are nans, then the whole thing returns
// true.
if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
- return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
-
+ return Builder->getTrue();
+
// Otherwise, no need to compare the two constants, compare the
// rest.
- return new FCmpInst(FCmpInst::FCMP_UNO,
- LHS->getOperand(0), RHS->getOperand(0));
+ return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
}
-
+
// Handle vector zeros. This occurs because the canonical form of
// "fcmp uno x,x" is "fcmp uno x, 0".
if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
isa<ConstantAggregateZero>(RHS->getOperand(1)))
- return new FCmpInst(FCmpInst::FCMP_UNO,
- LHS->getOperand(0), RHS->getOperand(0));
-
+ return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
+
return 0;
}
-
+
Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
-
+
if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
// Swap RHS operands to match LHS.
Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
// Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
if (Op0CC == Op1CC)
- return new FCmpInst((FCmpInst::Predicate)Op0CC,
- Op0LHS, Op0RHS);
+ return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
- return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
+ return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
if (Op0CC == FCmpInst::FCMP_FALSE)
- return ReplaceInstUsesWith(I, RHS);
+ return RHS;
if (Op1CC == FCmpInst::FCMP_FALSE)
- return ReplaceInstUsesWith(I, LHS);
+ return LHS;
bool Op0Ordered;
bool Op1Ordered;
unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
if (Op0Ordered == Op1Ordered) {
// If both are ordered or unordered, return a new fcmp with
// or'ed predicates.
- Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS);
- if (Instruction *I = dyn_cast<Instruction>(RV))
- return I;
- // Otherwise, it's a constant boolean value...
- return ReplaceInstUsesWith(I, RV);
+ return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
}
}
return 0;
/// ((A | B) & C1) | (B & C2)
///
/// into:
-///
+///
/// (A & C1) | B
///
/// when the XOR of the two constants is "all ones" (-1).
}
Instruction *InstCombiner::visitOr(BinaryOperator &I) {
- bool Changed = SimplifyCommutative(I);
+ bool Changed = SimplifyAssociativeOrCommutative(I);
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
if (Value *V = SimplifyOrInst(Op0, Op1, TD))
return ReplaceInstUsesWith(I, V);
-
-
- // See if we can simplify any instructions used by the instruction whose sole
+
+ // (A&B)|(A&C) -> A&(B|C) etc
+ if (Value *V = SimplifyUsingDistributiveLaws(I))
+ return ReplaceInstUsesWith(I, V);
+
+ // See if we can simplify any instructions used by the instruction whose sole
// purpose is to compute bits we don't care about.
if (SimplifyDemandedInstructionBits(I))
return &I;
if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
ConstantInt *C1 = 0; Value *X = 0;
// (X & C1) | C2 --> (X | C2) & (C1|C2)
+ // iff (C1 & C2) == 0.
if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
+ (RHS->getValue() & C1->getValue()) != 0 &&
Op0->hasOneUse()) {
Value *Or = Builder->CreateOr(X, RHS);
Or->takeName(Op0);
- return BinaryOperator::CreateAnd(Or,
- ConstantInt::get(I.getContext(),
- RHS->getValue() | C1->getValue()));
+ return BinaryOperator::CreateAnd(Or,
+ Builder->getInt(RHS->getValue() | C1->getValue()));
}
// (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
Value *Or = Builder->CreateOr(X, RHS);
Or->takeName(Op0);
return BinaryOperator::CreateXor(Or,
- ConstantInt::get(I.getContext(),
- C1->getValue() & ~RHS->getValue()));
+ Builder->getInt(C1->getValue() & ~RHS->getValue()));
}
// Try to fold constant and into select arguments.
if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
if (Instruction *R = FoldOpIntoSelect(I, SI))
return R;
+
if (isa<PHINode>(Op0))
if (Instruction *NV = FoldOpIntoPhi(I))
return NV;
// (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
if (match(Op0, m_Or(m_Value(), m_Value())) ||
match(Op1, m_Or(m_Value(), m_Value())) ||
- (match(Op0, m_Shift(m_Value(), m_Value())) &&
- match(Op1, m_Shift(m_Value(), m_Value())))) {
+ (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
+ match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
if (Instruction *BSwap = MatchBSwap(I))
return BSwap;
}
-
+
// (X^C)|Y -> (X|Y)^C iff Y&C == 0
if (Op0->hasOneUse() &&
match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
Value *C = 0, *D = 0;
if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
match(Op1, m_And(m_Value(B), m_Value(D)))) {
- Value *V1 = 0, *V2 = 0, *V3 = 0;
+ Value *V1 = 0, *V2 = 0;
C1 = dyn_cast<ConstantInt>(C);
C2 = dyn_cast<ConstantInt>(D);
if (C1 && C2) { // (A & C1)|(B & C2)
return ReplaceInstUsesWith(I, B);
}
}
-
+
if ((C1->getValue() & C2->getValue()) == 0) {
// ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
// iff (C1&C2) == 0 and (N&~C1) == 0
((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
(V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
return BinaryOperator::CreateAnd(A,
- ConstantInt::get(A->getContext(),
- C1->getValue()|C2->getValue()));
+ Builder->getInt(C1->getValue()|C2->getValue()));
// Or commutes, try both ways.
if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
(V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
return BinaryOperator::CreateAnd(B,
- ConstantInt::get(B->getContext(),
- C1->getValue()|C2->getValue()));
-
+ Builder->getInt(C1->getValue()|C2->getValue()));
+
// ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
// iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
ConstantInt *C3 = 0, *C4 = 0;
(C4->getValue() & ~C2->getValue()) == 0) {
V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
return BinaryOperator::CreateAnd(V2,
- ConstantInt::get(B->getContext(),
- C1->getValue()|C2->getValue()));
+ Builder->getInt(C1->getValue()|C2->getValue()));
}
}
}
-
- // Check to see if we have any common things being and'ed. If so, find the
- // terms for V1 & (V2|V3).
- if (Op0->hasOneUse() || Op1->hasOneUse()) {
- V1 = 0;
- if (A == B) // (A & C)|(A & D) == A & (C|D)
- V1 = A, V2 = C, V3 = D;
- else if (A == D) // (A & C)|(B & A) == A & (B|C)
- V1 = A, V2 = B, V3 = C;
- else if (C == B) // (A & C)|(C & D) == C & (A|D)
- V1 = C, V2 = A, V3 = D;
- else if (C == D) // (A & C)|(B & C) == C & (A|B)
- V1 = C, V2 = A, V3 = B;
-
- if (V1) {
- Value *Or = Builder->CreateOr(V2, V3, "tmp");
- return BinaryOperator::CreateAnd(V1, Or);
- }
- }
// (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants.
// Don't do this for vector select idioms, the code generator doesn't handle
if ((match(A, m_Not(m_Specific(B))) &&
match(D, m_Not(m_Specific(C)))))
return BinaryOperator::CreateXor(C, B);
+
+ // ((A|B)&1)|(B&-2) -> (A&1) | B
+ if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
+ match(A, m_Or(m_Specific(B), m_Value(V1)))) {
+ Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
+ if (Ret) return Ret;
+ }
+ // (B&-2)|((A|B)&1) -> (A&1) | B
+ if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
+ match(B, m_Or(m_Value(V1), m_Specific(A)))) {
+ Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
+ if (Ret) return Ret;
+ }
}
-
+
// (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
- if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
+ if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
SI0->getOperand(1) == SI1->getOperand(1) &&
(SI0->hasOneUse() || SI1->hasOneUse())) {
Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
SI0->getName());
- return BinaryOperator::Create(SI1->getOpcode(), NewOp,
+ return BinaryOperator::Create(SI1->getOpcode(), NewOp,
SI1->getOperand(1));
}
}
- // ((A|B)&1)|(B&-2) -> (A&1) | B
- if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
- match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
- Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
- if (Ret) return Ret;
- }
- // (B&-2)|((A|B)&1) -> (A&1) | B
- if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
- match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
- Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
- if (Ret) return Ret;
- }
-
// (~A | ~B) == (~(A & B)) - De Morgan's Law
if (Value *Op0NotVal = dyn_castNotVal(Op0))
if (Value *Op1NotVal = dyn_castNotVal(Op1))
return BinaryOperator::CreateNot(And);
}
+ // Canonicalize xor to the RHS.
+ bool SwappedForXor = false;
+ if (match(Op0, m_Xor(m_Value(), m_Value()))) {
+ std::swap(Op0, Op1);
+ SwappedForXor = true;
+ }
+
+ // A | ( A ^ B) -> A | B
+ // A | (~A ^ B) -> A | ~B
+ // (A & B) | (A ^ B)
+ if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
+ if (Op0 == A || Op0 == B)
+ return BinaryOperator::CreateOr(A, B);
+
+ if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
+ match(Op0, m_And(m_Specific(B), m_Specific(A))))
+ return BinaryOperator::CreateOr(A, B);
+
+ if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
+ Value *Not = Builder->CreateNot(B, B->getName()+".not");
+ return BinaryOperator::CreateOr(Not, Op0);
+ }
+ if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
+ Value *Not = Builder->CreateNot(A, A->getName()+".not");
+ return BinaryOperator::CreateOr(Not, Op0);
+ }
+ }
+
+ // A | ~(A | B) -> A | ~B
+ // A | ~(A ^ B) -> A | ~B
+ if (match(Op1, m_Not(m_Value(A))))
+ if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
+ if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
+ Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
+ B->getOpcode() == Instruction::Xor)) {
+ Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
+ B->getOperand(0);
+ Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
+ return BinaryOperator::CreateOr(Not, Op0);
+ }
+
+ if (SwappedForXor)
+ std::swap(Op0, Op1);
+
if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
- if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
- return Res;
-
+ if (Value *Res = FoldOrOfICmps(LHS, RHS))
+ return ReplaceInstUsesWith(I, Res);
+
// (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
- if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
- return Res;
-
+ if (Value *Res = FoldOrOfFCmps(LHS, RHS))
+ return ReplaceInstUsesWith(I, Res);
+
// fold (or (cast A), (cast B)) -> (cast (or A, B))
if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
- if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
- if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
- const Type *SrcTy = Op0C->getOperand(0)->getType();
- if (SrcTy == Op1C->getOperand(0)->getType() &&
- SrcTy->isIntOrIntVectorTy()) {
- Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
-
- if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
- // Only do this if the casts both really cause code to be
- // generated.
- ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
- ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
- Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
- return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
- }
-
- // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
- // cast is otherwise not optimizable. This happens for vector sexts.
- if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
- if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
- if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS)) {
- InsertNewInstBefore(Res, I);
- return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
- }
-
- // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
- // cast is otherwise not optimizable. This happens for vector sexts.
- if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
- if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
- if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS)) {
- InsertNewInstBefore(Res, I);
- return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
- }
+ CastInst *Op1C = dyn_cast<CastInst>(Op1);
+ if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
+ Type *SrcTy = Op0C->getOperand(0)->getType();
+ if (SrcTy == Op1C->getOperand(0)->getType() &&
+ SrcTy->isIntOrIntVectorTy()) {
+ Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
+
+ if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
+ // Only do this if the casts both really cause code to be
+ // generated.
+ ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
+ ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
+ Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
+ return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
}
+
+ // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
+ // cast is otherwise not optimizable. This happens for vector sexts.
+ if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
+ if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
+ if (Value *Res = FoldOrOfICmps(LHS, RHS))
+ return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
+
+ // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
+ // cast is otherwise not optimizable. This happens for vector sexts.
+ if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
+ if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
+ if (Value *Res = FoldOrOfFCmps(LHS, RHS))
+ return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
}
+ }
+ }
+
+ // or(sext(A), B) -> A ? -1 : B where A is an i1
+ // or(A, sext(B)) -> B ? -1 : A where B is an i1
+ if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
+ return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
+ if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
+ return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
+
+ // Note: If we've gotten to the point of visiting the outer OR, then the
+ // inner one couldn't be simplified. If it was a constant, then it won't
+ // be simplified by a later pass either, so we try swapping the inner/outer
+ // ORs in the hopes that we'll be able to simplify it this way.
+ // (X|C) | V --> (X|V) | C
+ if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
+ match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
+ Value *Inner = Builder->CreateOr(A, Op1);
+ Inner->takeName(Op0);
+ return BinaryOperator::CreateOr(Inner, C1);
+ }
+
+ // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
+ // Since this OR statement hasn't been optimized further yet, we hope
+ // that this transformation will allow the new ORs to be optimized.
+ {
+ Value *X = 0, *Y = 0;
+ if (Op0->hasOneUse() && Op1->hasOneUse() &&
+ match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
+ match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
+ Value *orTrue = Builder->CreateOr(A, C);
+ Value *orFalse = Builder->CreateOr(B, D);
+ return SelectInst::Create(X, orTrue, orFalse);
+ }
}
-
+
return Changed ? &I : 0;
}
Instruction *InstCombiner::visitXor(BinaryOperator &I) {
- bool Changed = SimplifyCommutative(I);
+ bool Changed = SimplifyAssociativeOrCommutative(I);
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
- if (isa<UndefValue>(Op1)) {
- if (isa<UndefValue>(Op0))
- // Handle undef ^ undef -> 0 special case. This is a common
- // idiom (misuse).
- return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
- return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
- }
+ if (Value *V = SimplifyXorInst(Op0, Op1, TD))
+ return ReplaceInstUsesWith(I, V);
+
+ // (A&B)^(A&C) -> A&(B^C) etc
+ if (Value *V = SimplifyUsingDistributiveLaws(I))
+ return ReplaceInstUsesWith(I, V);
- // xor X, X = 0
- if (Op0 == Op1)
- return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
-
- // See if we can simplify any instructions used by the instruction whose sole
+ // See if we can simplify any instructions used by the instruction whose sole
// purpose is to compute bits we don't care about.
if (SimplifyDemandedInstructionBits(I))
return &I;
- if (I.getType()->isVectorTy())
- if (isa<ConstantAggregateZero>(Op1))
- return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
// Is this a ~ operation?
if (Value *NotOp = dyn_castNotVal(&I)) {
if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
- if (Op0I->getOpcode() == Instruction::And ||
+ if (Op0I->getOpcode() == Instruction::And ||
Op0I->getOpcode() == Instruction::Or) {
// ~(~X & Y) --> (X | ~Y) - De Morgan's Law
// ~(~X | Y) === (X & ~Y) - De Morgan's Law
return BinaryOperator::CreateOr(Op0NotVal, NotY);
return BinaryOperator::CreateAnd(Op0NotVal, NotY);
}
-
+
// ~(X & Y) --> (~X | ~Y) - De Morgan's Law
// ~(X | Y) === (~X & ~Y) - De Morgan's Law
- if (isFreeToInvert(Op0I->getOperand(0)) &&
+ if (isFreeToInvert(Op0I->getOperand(0)) &&
isFreeToInvert(Op0I->getOperand(1))) {
Value *NotX =
Builder->CreateNot(Op0I->getOperand(0), "notlhs");
}
}
}
-
-
+
+
if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
- if (RHS->isOne() && Op0->hasOneUse()) {
+ if (RHS->isOne() && Op0->hasOneUse())
// xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
- if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
- return new ICmpInst(ICI->getInversePredicate(),
- ICI->getOperand(0), ICI->getOperand(1));
-
- if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
- return new FCmpInst(FCI->getInversePredicate(),
- FCI->getOperand(0), FCI->getOperand(1));
- }
+ if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
+ return CmpInst::Create(CI->getOpcode(),
+ CI->getInversePredicate(),
+ CI->getOperand(0), CI->getOperand(1));
// fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
if (CI->hasOneUse() && Op0C->hasOneUse()) {
Instruction::CastOps Opcode = Op0C->getOpcode();
if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
- (RHS == ConstantExpr::getCast(Opcode,
- ConstantInt::getTrue(I.getContext()),
+ (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(),
Op0C->getDestTy()))) {
CI->setPredicate(CI->getInversePredicate());
return CastInst::Create(Opcode, CI, Op0C->getType());
ConstantInt::get(I.getType(), 1));
return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
}
-
+
if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
if (Op0I->getOpcode() == Instruction::Add) {
// ~(X-c) --> (-c-1)-X
Op0I->getOperand(0));
} else if (RHS->getValue().isSignBit()) {
// (X + C) ^ signbit -> (X + C + signbit)
- Constant *C = ConstantInt::get(I.getContext(),
- RHS->getValue() + Op0CI->getValue());
+ Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue());
return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
}
// Anything in both C1 and C2 is known to be zero, remove it from
// NewRHS.
Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
- NewRHS = ConstantExpr::getAnd(NewRHS,
+ NewRHS = ConstantExpr::getAnd(NewRHS,
ConstantExpr::getNot(CommonBits));
Worklist.Add(Op0I);
I.setOperand(0, Op0I->getOperand(0));
I.setOperand(1, NewRHS);
return &I;
}
+ } else if (Op0I->getOpcode() == Instruction::LShr) {
+ // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
+ // E1 = "X ^ C1"
+ BinaryOperator *E1;
+ ConstantInt *C1;
+ if (Op0I->hasOneUse() &&
+ (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
+ E1->getOpcode() == Instruction::Xor &&
+ (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
+ // fold (C1 >> C2) ^ C3
+ ConstantInt *C2 = Op0CI, *C3 = RHS;
+ APInt FoldConst = C1->getValue().lshr(C2->getValue());
+ FoldConst ^= C3->getValue();
+ // Prepare the two operands.
+ Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2);
+ Opnd0->takeName(Op0I);
+ cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
+ Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
+
+ return BinaryOperator::CreateXor(Opnd0, FoldVal);
+ }
}
}
}
return NV;
}
- if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
- if (X == Op1)
- return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
-
- if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
- if (X == Op0)
- return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
-
-
BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
if (Op1I) {
Value *A, *B;
I.swapOperands(); // Simplified below.
std::swap(Op0, Op1);
}
- } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
- return ReplaceInstUsesWith(I, B); // A^(A^B) == B
- } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
- return ReplaceInstUsesWith(I, A); // A^(B^A) == B
- } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
+ } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
Op1I->hasOneUse()){
if (A == Op0) { // A^(A&B) -> A^(B&A)
Op1I->swapOperands();
}
}
}
-
+
BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
if (Op0I) {
Value *A, *B;
if (A == Op1) // (B|A)^B == (A|B)^B
std::swap(A, B);
if (B == Op1) // (A|B)^B == A & ~B
- return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
- } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
- return ReplaceInstUsesWith(I, B); // (A^B)^A == B
- } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
- return ReplaceInstUsesWith(I, A); // (B^A)^A == B
- } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
+ return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
+ } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
Op0I->hasOneUse()){
if (A == Op1) // (A&B)^A -> (B&A)^A
std::swap(A, B);
if (B == Op1 && // (B&A)^A == ~B & A
!isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
- return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
+ return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
}
}
}
-
+
// (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
- if (Op0I && Op1I && Op0I->isShift() &&
- Op0I->getOpcode() == Op1I->getOpcode() &&
+ if (Op0I && Op1I && Op0I->isShift() &&
+ Op0I->getOpcode() == Op1I->getOpcode() &&
Op0I->getOperand(1) == Op1I->getOperand(1) &&
- (Op1I->hasOneUse() || Op1I->hasOneUse())) {
+ (Op0I->hasOneUse() || Op1I->hasOneUse())) {
Value *NewOp =
Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
Op0I->getName());
- return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
+ return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
Op1I->getOperand(1));
}
-
+
if (Op0I && Op1I) {
Value *A, *B, *C, *D;
// (A & B)^(A | B) -> A ^ B
if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
- if ((A == C && B == D) || (A == D && B == C))
+ if ((A == C && B == D) || (A == D && B == C))
return BinaryOperator::CreateXor(A, B);
}
// (A | B)^(A & B) -> A ^ B
if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
match(Op1I, m_And(m_Value(C), m_Value(D)))) {
- if ((A == C && B == D) || (A == D && B == C))
+ if ((A == C && B == D) || (A == D && B == C))
return BinaryOperator::CreateXor(A, B);
}
-
- // (A & B)^(C & D)
- if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
- match(Op0I, m_And(m_Value(A), m_Value(B))) &&
- match(Op1I, m_And(m_Value(C), m_Value(D)))) {
- // (X & Y)^(X & Y) -> (Y^Z) & X
- Value *X = 0, *Y = 0, *Z = 0;
- if (A == C)
- X = A, Y = B, Z = D;
- else if (A == D)
- X = A, Y = B, Z = C;
- else if (B == C)
- X = B, Y = A, Z = D;
- else if (B == D)
- X = B, Y = A, Z = C;
-
- if (X) {
- Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
- return BinaryOperator::CreateAnd(NewOp, X);
- }
- }
}
-
+
// (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
bool isSigned = LHS->isSigned() || RHS->isSigned();
- Value *RV = getICmpValue(isSigned, Code, Op0, Op1);
- if (Instruction *I = dyn_cast<Instruction>(RV))
- return I;
- // Otherwise, it's a constant boolean value.
- return ReplaceInstUsesWith(I, RV);
+ return ReplaceInstUsesWith(I,
+ getNewICmpValue(isSigned, Code, Op0, Op1,
+ Builder));
}
}
if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
- const Type *SrcTy = Op0C->getOperand(0)->getType();
+ Type *SrcTy = Op0C->getOperand(0)->getType();
if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
// Only do this if the casts both really cause code to be generated.
- ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
+ ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
I.getType()) &&
- ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
+ ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
I.getType())) {
Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
Op1C->getOperand(0), I.getName());