#include "InstCombine.h"
#include "llvm/Intrinsics.h"
#include "llvm/Analysis/InstructionSimplify.h"
+#include "llvm/Transforms/Utils/CmpInstAnalysis.h"
+#include "llvm/Support/ConstantRange.h"
#include "llvm/Support/PatternMatch.h"
using namespace llvm;
using namespace PatternMatch;
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.
/// 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,
- InstCombiner::BuilderTy *Builder) {
- CmpInst::Predicate Pred;
- switch (Code) {
- default: assert(0 && "Illegal ICmp code!");
- case 0: // False.
- return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
- case 1: Pred = Sign ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
- case 2: Pred = ICmpInst::ICMP_EQ; break;
- case 3: Pred = Sign ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
- case 4: Pred = Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
- case 5: Pred = ICmpInst::ICMP_NE; break;
- case 6: Pred = Sign ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
- case 7: // True.
- return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
- }
- return Builder->CreateICmp(Pred, LHS, RHS);
+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
return Builder->CreateFCmp(Pred, LHS, RHS);
}
-/// 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));
-}
-
// OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
// the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
// guaranteed to be a binary operator.
/// 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
+/// 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.
Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
bool isSigned = LHS->isSigned() || RHS->isSigned();
- return getICmpValue(isSigned, Code, Op0, Op1, Builder);
+ return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
}
}
Value *NewOr = Builder->CreateOr(Val, Val2);
return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
}
+
+ // (icmp slt A, 0) & (icmp slt B, 0) --> (icmp slt (A&B), 0)
+ if (LHSCC == ICmpInst::ICMP_SLT && LHSCst->isZero()) {
+ Value *NewAnd = Builder->CreateAnd(Val, Val2);
+ return Builder->CreateICmp(LHSCC, NewAnd, LHSCst);
+ }
+
+ // (icmp sgt A, -1) & (icmp sgt B, -1) --> (icmp sgt (A|B), -1)
+ if (LHSCC == ICmpInst::ICMP_SGT && LHSCst->isAllOnesValue()) {
+ Value *NewOr = Builder->CreateOr(Val, Val2);
+ 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 == RHSCC && ICmpInst::isEquality(LHSCC) &&
+ LHS->hasOneUse() && RHS->hasOneUse()) {
+ Value *V;
+ ConstantInt *AndCst, *SmallCst = 0, *BigCst = 0;
+
+ // (trunc x) == C1 & (and x, CA) == C2
+ if (match(Val2, m_Trunc(m_Value(V))) &&
+ match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
+ SmallCst = RHSCst;
+ BigCst = LHSCst;
+ }
+ // (and x, CA) == C2 & (trunc x) == C1
+ 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);
+ }
+ }
+ }
+
+ // (X & C) == 0 & X > -1 -> (X & (C | SignBit)) == 0
+ if (LHS->hasOneUse() && RHS->hasOneUse() &&
+ ((LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero() &&
+ RHSCC == ICmpInst::ICMP_SGT && RHSCst->isAllOnesValue()) ||
+ (RHSCC == ICmpInst::ICMP_EQ && RHSCst->isZero() &&
+ LHSCC == ICmpInst::ICMP_SGT && LHSCst->isAllOnesValue()))) {
+ BinaryOperator *BO =
+ dyn_cast<BinaryOperator>(LHSCC == ICmpInst::ICMP_EQ ? Val : Val2);
+ ConstantInt *AndCst;
+ if (BO && match(BO, m_OneUse(m_And(m_Value(), m_ConstantInt(AndCst))))) {
+ APInt New = AndCst->getValue() | APInt::getSignBit(AndCst->getBitWidth());
+ BO->setOperand(1, ConstantInt::get(AndCst->getContext(), New));
+ return BO == Val ? LHS : RHS;
+ }
}
// From here on, we only handle:
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;
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 ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
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
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 ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
break;
case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
((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));
}
- // 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 == Op0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
- return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
}
// (A&((~A)|B)) -> A&B
// 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()) {
/// 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());
+ IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
if (!ITy || ITy->getBitWidth() % 16 ||
// ByteMask only allows up to 32-byte values.
ITy->getBitWidth() > 32*8)
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);
}
Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
bool isSigned = LHS->isSigned() || RHS->isSigned();
- return getICmpValue(isSigned, Code, Op0, Op1, Builder);
+ return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
}
}
Value *NewOr = Builder->CreateOr(Val, Val2);
return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
}
+
+ // (icmp slt A, 0) | (icmp slt B, 0) --> (icmp slt (A|B), 0)
+ if (LHSCC == ICmpInst::ICMP_SLT && LHSCst->isZero()) {
+ Value *NewOr = Builder->CreateOr(Val, Val2);
+ return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
+ }
+
+ // (icmp sgt A, -1) | (icmp sgt B, -1) --> (icmp sgt (A&B), -1)
+ if (LHSCC == ICmpInst::ICMP_SGT && LHSCst->isAllOnesValue()) {
+ Value *NewAnd = Builder->CreateAnd(Val, Val2);
+ return Builder->CreateICmp(LHSCC, NewAnd, LHSCst);
+ }
+ }
+
+ // (X & C) != 0 & X < 0 -> (X & (C | SignBit)) != 0
+ if (LHS->hasOneUse() && RHS->hasOneUse() &&
+ ((LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero() &&
+ RHSCC == ICmpInst::ICMP_SLT && RHSCst->isZero()) ||
+ (RHSCC == ICmpInst::ICMP_NE && RHSCst->isZero() &&
+ LHSCC == ICmpInst::ICMP_SLT && LHSCst->isZero()))) {
+ BinaryOperator *BO =
+ dyn_cast<BinaryOperator>(LHSCC == ICmpInst::ICMP_NE ? Val : Val2);
+ ConstantInt *AndCst;
+ if (BO && match(BO, m_OneUse(m_And(m_Value(), m_ConstantInt(AndCst))))) {
+ APInt New = AndCst->getValue() | APInt::getSignBit(AndCst->getBitWidth());
+ BO->setOperand(1, ConstantInt::get(AndCst->getContext(), New));
+ return BO == Val ? LHS : RHS;
+ }
}
// (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
return BinaryOperator::CreateNot(And);
}
+ // Canonicalize xor to the RHS.
+ if (match(Op0, m_Xor(m_Value(), m_Value())))
+ std::swap(Op0, Op1);
+
+ // A | ( A ^ B) -> A | B
+ // A | (~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 (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 (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
if (Value *Res = FoldOrOfICmps(LHS, RHS))
if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
CastInst *Op1C = dyn_cast<CastInst>(Op1);
if (Op1C && 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->isIntOrIntVectorTy()) {
Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
}
}
}
-
+
+ // 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
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"));
+ 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);
}
}
}
unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
bool isSigned = LHS->isSigned() || RHS->isSigned();
return ReplaceInstUsesWith(I,
- getICmpValue(isSigned, Code, Op0, Op1, Builder));
+ 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),