//===----------------------------------------------------------------------===//
#include "InstCombine.h"
-#include "llvm/IntrinsicInst.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/MemoryBuiltins.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/IntrinsicInst.h"
#include "llvm/Support/ConstantRange.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/PatternMatch.h"
+#include "llvm/Target/TargetLibraryInfo.h"
using namespace llvm;
using namespace PatternMatch;
/// If we can't emit an optimized form for this expression, this returns null.
///
static Value *EvaluateGEPOffsetExpression(User *GEP, InstCombiner &IC) {
- TargetData &TD = *IC.getTargetData();
+ DataLayout &TD = *IC.getDataLayout();
gep_type_iterator GTI = gep_type_begin(GEP);
// Check to see if this gep only has a single variable index. If so, and if
Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
ICmpInst::Predicate Cond,
Instruction &I) {
+ // Don't transform signed compares of GEPs into index compares. Even if the
+ // GEP is inbounds, the final add of the base pointer can have signed overflow
+ // and would change the result of the icmp.
+ // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
+ // the maximum signed value for the pointer type.
+ if (ICmpInst::isSigned(Cond))
+ return 0;
+
// Look through bitcasts.
if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
RHS = BCI->getOperand(0);
// of the high bits truncated out of x are known.
unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
- APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
- ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
+ ComputeMaskedBits(LHSI->getOperand(0), KnownZero, KnownOne);
// If all the high bits are known, we can do this xform.
if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
// Pull in the high bits from known-ones set.
APInt NewRHS = RHS->getValue().zext(SrcBits);
- NewRHS |= KnownOne;
+ NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits-DstBits);
return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
ConstantInt::get(ICI.getContext(), NewRHS));
}
ICI.setOperand(0, NewAnd);
return &ICI;
}
+
+ // Replace ((X & AndCST) > RHSV) with ((X & AndCST) != 0), if any
+ // bit set in (X & AndCST) will produce a result greater than RHSV.
+ if (ICI.getPredicate() == ICmpInst::ICMP_UGT) {
+ unsigned NTZ = AndCST->getValue().countTrailingZeros();
+ if ((NTZ < AndCST->getBitWidth()) &&
+ APInt::getOneBitSet(AndCST->getBitWidth(), NTZ).ugt(RHSV))
+ return new ICmpInst(ICmpInst::ICMP_NE, LHSI,
+ Constant::getNullValue(RHS->getType()));
+ }
}
// Try to optimize things like "A[i]&42 == 0" to index computations.
return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
And, Constant::getNullValue(And->getType()));
}
+
+ // Transform (icmp pred iM (shl iM %v, N), CI)
+ // -> (icmp pred i(M-N) (trunc %v iM to i(N-N)), (trunc (CI>>N))
+ // Transform the shl to a trunc if (trunc (CI>>N)) has no loss.
+ // This enables to get rid of the shift in favor of a trunc which can be
+ // free on the target. It has the additional benefit of comparing to a
+ // smaller constant, which will be target friendly.
+ unsigned Amt = ShAmt->getLimitedValue(TypeBits-1);
+ if (Amt != 0 && RHSV.countTrailingZeros() >= Amt) {
+ Type *NTy = IntegerType::get(ICI.getContext(), TypeBits - Amt);
+ Constant *NCI = ConstantExpr::getTrunc(
+ ConstantExpr::getAShr(RHS,
+ ConstantInt::get(RHS->getType(), Amt)),
+ NTy);
+ return new ICmpInst(ICI.getPredicate(),
+ Builder->CreateTrunc(LHSI->getOperand(0), NTy),
+ NCI);
+ }
+
break;
}
// Try not to increase register pressure.
BO0->hasOneUse() && BO1->hasOneUse()) {
// Determine Y and Z in the form icmp (X+Y), (X+Z).
- Value *Y = (A == C || A == D) ? B : A;
- Value *Z = (C == A || C == B) ? D : C;
+ Value *Y, *Z;
+ if (A == C) {
+ // C + B == C + D -> B == D
+ Y = B;
+ Z = D;
+ } else if (A == D) {
+ // D + B == C + D -> B == C
+ Y = B;
+ Z = C;
+ } else if (B == C) {
+ // A + C == C + D -> A == D
+ Y = A;
+ Z = D;
+ } else {
+ assert(B == D);
+ // A + D == C + D -> A == C
+ Y = A;
+ Z = C;
+ }
return new ICmpInst(Pred, Y, Z);
}
}
}
+ // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
+ // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
+ ConstantInt *Cst1;
+ if ((Op0->hasOneUse() &&
+ match(Op0, m_ZExt(m_Value(A))) &&
+ match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
+ (Op1->hasOneUse() &&
+ match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
+ match(Op1, m_ZExt(m_Value(A))))) {
+ APInt Pow2 = Cst1->getValue() + 1;
+ if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
+ Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
+ return new ICmpInst(I.getPredicate(), A,
+ Builder->CreateTrunc(B, A->getType()));
+ }
+
// Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
// "icmp (and X, mask), cst"
uint64_t ShAmt = 0;
- ConstantInt *Cst1;
if (Op0->hasOneUse() &&
match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A),
m_ConstantInt(ShAmt))))) &&
case ICmpInst::ICMP_UGE:
// (float)int >= -4.4 --> true
// (float)int >= 4.4 --> int > 4
- if (!RHS.isNegative())
+ if (RHS.isNegative())
return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
Pred = ICmpInst::ICMP_UGT;
break;
if (!RHSF)
break;
- // We can't convert a PPC double double.
- if (RHSF->getType()->isPPC_FP128Ty())
- break;
-
const fltSemantics *Sem;
// FIXME: This shouldn't be here.
if (LHSExt->getSrcTy()->isHalfTy())
Sem = &APFloat::IEEEquad;
else if (LHSExt->getSrcTy()->isX86_FP80Ty())
Sem = &APFloat::x87DoubleExtended;
+ else if (LHSExt->getSrcTy()->isPPC_FP128Ty())
+ Sem = &APFloat::PPCDoubleDouble;
else
break;
return Res;
}
break;
+ case Instruction::Call: {
+ CallInst *CI = cast<CallInst>(LHSI);
+ LibFunc::Func Func;
+ // Various optimization for fabs compared with zero.
+ if (RHSC->isNullValue() && CI->getCalledFunction() &&
+ TLI->getLibFunc(CI->getCalledFunction()->getName(), Func) &&
+ TLI->has(Func)) {
+ if (Func == LibFunc::fabs || Func == LibFunc::fabsf ||
+ Func == LibFunc::fabsl) {
+ switch (I.getPredicate()) {
+ default: break;
+ // fabs(x) < 0 --> false
+ case FCmpInst::FCMP_OLT:
+ return ReplaceInstUsesWith(I, Builder->getFalse());
+ // fabs(x) > 0 --> x != 0
+ case FCmpInst::FCMP_OGT:
+ return new FCmpInst(FCmpInst::FCMP_ONE, CI->getArgOperand(0),
+ RHSC);
+ // fabs(x) <= 0 --> x == 0
+ case FCmpInst::FCMP_OLE:
+ return new FCmpInst(FCmpInst::FCMP_OEQ, CI->getArgOperand(0),
+ RHSC);
+ // fabs(x) >= 0 --> !isnan(x)
+ case FCmpInst::FCMP_OGE:
+ return new FCmpInst(FCmpInst::FCMP_ORD, CI->getArgOperand(0),
+ RHSC);
+ // fabs(x) == 0 --> x == 0
+ // fabs(x) != 0 --> x != 0
+ case FCmpInst::FCMP_OEQ:
+ case FCmpInst::FCMP_UEQ:
+ case FCmpInst::FCMP_ONE:
+ case FCmpInst::FCMP_UNE:
+ return new FCmpInst(I.getPredicate(), CI->getArgOperand(0),
+ RHSC);
+ }
+ }
+ }
+ }
}
}
projects / oota-llvm.git / blobdiff