#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/MemoryBuiltins.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/DataLayout.h"
+#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Support/ConstantRange.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/PatternMatch.h"
// order the state machines in complexity of the generated code.
Value *Idx = GEP->getOperand(2);
+ unsigned AS = GEP->getPointerAddressSpace();
// If the index is larger than the pointer size of the target, truncate the
// index down like the GEP would do implicitly. We don't have to do this for
// an inbounds GEP because the index can't be out of range.
if (!GEP->isInBounds() &&
- Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
+ Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits(AS))
Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
// If the comparison is only true for one or two elements, emit direct
/// 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
}
}
+ unsigned AS = cast<GetElementPtrInst>(GEP)->getPointerAddressSpace();
// Okay, we know we have a single variable index, which must be a
// pointer/array/vector index. If there is no offset, life is simple, return
// the index.
- unsigned IntPtrWidth = TD.getPointerSizeInBits();
+ unsigned IntPtrWidth = TD.getPointerSizeInBits(AS);
if (Offset == 0) {
// Cast to intptrty in case a truncation occurs. If an extension is needed,
// we don't need to bother extending: the extension won't affect where the
// 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 minimum signed value for the pointer type.
+ // the maximum signed value for the pointer type.
if (ICmpInst::isSigned(Cond))
return 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));
}
// Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
// integer type is the same size as the pointer type.
if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
- TD->getPointerSizeInBits() ==
+ TD->getPointerSizeInBits(
+ cast<PtrToIntInst>(LHSCI)->getPointerAddressSpace()) ==
cast<IntegerType>(DestTy)->getBitWidth()) {
Value *RHSOp = 0;
if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
}
}
+ // 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;
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