assert(MVT::isInteger(VT) && "Invalid VT!");
unsigned VTBits = MVT::getSizeInBits(VT);
unsigned Tmp, Tmp2;
+ unsigned FirstAnswer = 1;
if (Depth == 6)
return 1; // Limit search depth.
case ISD::AND:
case ISD::OR:
case ISD::XOR: // NOT is handled here.
- // Logical binary ops preserve the number of sign bits.
+ // Logical binary ops preserve the number of sign bits at the worst.
Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1);
- if (Tmp == 1) return 1; // Early out.
- Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
- return std::min(Tmp, Tmp2);
+ if (Tmp != 1) {
+ Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1);
+ FirstAnswer = std::min(Tmp, Tmp2);
+ // We computed what we know about the sign bits as our first
+ // answer. Now proceed to the generic code that uses
+ // ComputeMaskedBits, and pick whichever answer is better.
+ }
+ break;
case ISD::SELECT:
Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1);
Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
Op.getOpcode() == ISD::INTRINSIC_VOID) {
unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth);
- if (NumBits > 1) return NumBits;
+ if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits);
}
// Finally, if we can prove that the top bits of the result are 0's or 1's,
Mask = KnownOne;
} else {
// Nothing known.
- return 1;
+ return FirstAnswer;
}
// Okay, we know that the sign bit in Mask is set. Use CLZ to determine
Mask <<= Mask.getBitWidth()-VTBits;
// Return # leading zeros. We use 'min' here in case Val was zero before
// shifting. We don't want to return '64' as for an i32 "0".
- return std::min(VTBits, Mask.countLeadingZeros());
+ return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros()));
}
unsigned NumElems = PermMask.getNumOperands();
SDOperand V = (Idx < NumElems) ? N->getOperand(0) : N->getOperand(1);
Idx %= NumElems;
- if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
- return (Idx == 0)
- ? V.getOperand(0) : getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
+
+ if (V.getOpcode() == ISD::BIT_CONVERT) {
+ V = V.getOperand(0);
+ if (MVT::getVectorNumElements(V.getValueType()) != NumElems)
+ return SDOperand();
}
+ if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
+ return (Idx == 0) ? V.getOperand(0)
+ : getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
+ if (V.getOpcode() == ISD::BUILD_VECTOR)
+ return V.getOperand(Idx);
if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
SDOperand Elt = PermMask.getOperand(Idx);
if (Elt.getOpcode() == ISD::UNDEF)
const Value *SrcSV, uint64_t SrcSVOff){
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
- // Expand memcpy to a series of store ops if the size operand falls below
- // a certain threshold.
+ // Expand memcpy to a series of load and store ops if the size operand falls
+ // below a certain threshold.
std::vector<MVT::ValueType> MemOps;
uint64_t Limit = -1;
if (!AlwaysInline)
&OutChains[0], OutChains.size());
}
+static SDOperand getMemmoveLoadsAndStores(SelectionDAG &DAG,
+ SDOperand Chain, SDOperand Dst,
+ SDOperand Src, uint64_t Size,
+ unsigned Align, bool AlwaysInline,
+ const Value *DstSV, uint64_t DstSVOff,
+ const Value *SrcSV, uint64_t SrcSVOff){
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+
+ // Expand memmove to a series of load and store ops if the size operand falls
+ // below a certain threshold.
+ std::vector<MVT::ValueType> MemOps;
+ uint64_t Limit = -1;
+ if (!AlwaysInline)
+ Limit = TLI.getMaxStoresPerMemmove();
+ unsigned DstAlign = Align; // Destination alignment can change.
+ if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign,
+ DAG, TLI))
+ return SDOperand();
+
+ std::string Str;
+ uint64_t SrcOff = 0, DstOff = 0;
+
+ SmallVector<SDOperand, 8> LoadValues;
+ SmallVector<SDOperand, 8> LoadChains;
+ SmallVector<SDOperand, 8> OutChains;
+ unsigned NumMemOps = MemOps.size();
+ for (unsigned i = 0; i < NumMemOps; i++) {
+ MVT::ValueType VT = MemOps[i];
+ unsigned VTSize = MVT::getSizeInBits(VT) / 8;
+ SDOperand Value, Store;
+
+ Value = DAG.getLoad(VT, Chain,
+ getMemBasePlusOffset(Src, SrcOff, DAG),
+ SrcSV, SrcSVOff + SrcOff, false, Align);
+ LoadValues.push_back(Value);
+ LoadChains.push_back(Value.getValue(1));
+ SrcOff += VTSize;
+ }
+ Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
+ &LoadChains[0], LoadChains.size());
+ OutChains.clear();
+ for (unsigned i = 0; i < NumMemOps; i++) {
+ MVT::ValueType VT = MemOps[i];
+ unsigned VTSize = MVT::getSizeInBits(VT) / 8;
+ SDOperand Value, Store;
+
+ Store = DAG.getStore(Chain, LoadValues[i],
+ getMemBasePlusOffset(Dst, DstOff, DAG),
+ DstSV, DstSVOff + DstOff, false, DstAlign);
+ OutChains.push_back(Store);
+ DstOff += VTSize;
+ }
+
+ return DAG.getNode(ISD::TokenFactor, MVT::Other,
+ &OutChains[0], OutChains.size());
+}
+
static SDOperand getMemsetStores(SelectionDAG &DAG,
SDOperand Chain, SDOperand Dst,
SDOperand Src, uint64_t Size,
const Value *DstSV, uint64_t DstSVOff,
const Value *SrcSV, uint64_t SrcSVOff) {
- // TODO: Optimize small memmove cases with simple loads and stores,
- // ensuring that all loads precede all stores. This can cause severe
- // register pressure, so targets should be careful with the size limit.
+ // Check to see if we should lower the memmove to loads and stores first.
+ // For cases within the target-specified limits, this is the best choice.
+ ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
+ if (ConstantSize) {
+ // Memmove with size zero? Just return the original chain.
+ if (ConstantSize->isNullValue())
+ return Chain;
+
+ SDOperand Result =
+ getMemmoveLoadsAndStores(*this, Chain, Dst, Src, ConstantSize->getValue(),
+ Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
+ if (Result.Val)
+ return Result;
+ }
// Then check to see if we should lower the memmove with target-specific
// code. If the target chooses to do this, this is the next best.
}
bool Indexed = AM != ISD::UNINDEXED;
- assert(Indexed || Offset.getOpcode() == ISD::UNDEF &&
+ assert((Indexed || Offset.getOpcode() == ISD::UNDEF) &&
"Unindexed load with an offset!");
SDVTList VTs = Indexed ?
projects / oota-llvm.git / blobdiff