SDValue ARMSelectionDAGInfo::EmitTargetCodeForMemcpy(
SelectionDAG &DAG, const SDLoc &dl, SDValue Chain, SDValue Dst, SDValue Src,
SDValue Size, unsigned Align, bool isVolatile, bool AlwaysInline,
MachinePointerInfo DstPtrInfo, MachinePointerInfo SrcPtrInfo) const {
const ARMSubtarget &Subtarget =
DAG.getMachineFunction().getSubtarget<ARMSubtarget>();
// Do repeated 4-byte loads and stores. To be improved.
// This requires 4-byte alignment.
if ((Align & 3) != 0)
return SDValue();
// This requires the copy size to be a constant, preferably
// within a subtarget-specific limit.
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
if (!ConstantSize)
return EmitSpecializedLibcall(DAG, dl, Chain, Dst, Src, Size, Align,
RTLIB::MEMCPY);
uint64_t SizeVal = ConstantSize->getZExtValue();
if (!AlwaysInline && SizeVal > Subtarget.getMaxInlineSizeThreshold())
return EmitSpecializedLibcall(DAG, dl, Chain, Dst, Src, Size, Align,
RTLIB::MEMCPY);
unsigned BytesLeft = SizeVal & 3;
unsigned NumMemOps = SizeVal >> 2;
unsigned EmittedNumMemOps = 0;
EVT VT = MVT::i32;
unsigned VTSize = 4;
unsigned i = 0;
// Emit a maximum of 4 loads in Thumb1 since we have fewer registers
const unsigned MaxLoadsInLDM = Subtarget.isThumb1Only() ? 4 : 6;
SDValue TFOps[6];
SDValue Loads[6];
uint64_t SrcOff = 0, DstOff = 0;
// FIXME: We should invent a VMEMCPY pseudo-instruction that lowers to
// VLDM/VSTM and make this code emit it when appropriate. This would reduce
// pressure on the general purpose registers. However this seems harder to map
// onto the register allocator's view of the world.
// The number of MEMCPY pseudo-instructions to emit. We use up to
// MaxLoadsInLDM registers per mcopy, which will get lowered into ldm/stm
// later on. This is a lower bound on the number of MEMCPY operations we must
// emit.
unsigned NumMEMCPYs = (NumMemOps + MaxLoadsInLDM - 1) / MaxLoadsInLDM;
// Code size optimisation: do not inline memcpy if expansion results in
// more instructions than the libary call.
if (NumMEMCPYs > 1 && DAG.getMachineFunction().getFunction().optForMinSize()) {
return SDValue();
}
SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other, MVT::Glue);
for (unsigned I = 0; I != NumMEMCPYs; ++I) {
// Evenly distribute registers among MEMCPY operations to reduce register
// pressure.
unsigned NextEmittedNumMemOps = NumMemOps * (I + 1) / NumMEMCPYs;
unsigned NumRegs = NextEmittedNumMemOps - EmittedNumMemOps;
Dst = DAG.getNode(ARMISD::MEMCPY, dl, VTs, Chain, Dst, Src,
DAG.getConstant(NumRegs, dl, MVT::i32));
Src = Dst.getValue(1);
Chain = Dst.getValue(2);
DstPtrInfo = DstPtrInfo.getWithOffset(NumRegs * VTSize);
SrcPtrInfo = SrcPtrInfo.getWithOffset(NumRegs * VTSize);
EmittedNumMemOps = NextEmittedNumMemOps;
}
if (BytesLeft == 0)
return Chain;
// Issue loads / stores for the trailing (1 - 3) bytes.
auto getRemainingValueType = [](unsigned BytesLeft) {
return (BytesLeft >= 2) ? MVT::i16 : MVT::i8;
};
auto getRemainingSize = [](unsigned BytesLeft) {
return (BytesLeft >= 2) ? 2 : 1;
};
unsigned BytesLeftSave = BytesLeft;
i = 0;
while (BytesLeft) {
VT = getRemainingValueType(BytesLeft);
VTSize = getRemainingSize(BytesLeft);
Loads[i] = DAG.getLoad(VT, dl, Chain,
DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
DAG.getConstant(SrcOff, dl, MVT::i32)),
SrcPtrInfo.getWithOffset(SrcOff));
TFOps[i] = Loads[i].getValue(1);
++i;
SrcOff += VTSize;
BytesLeft -= VTSize;
}
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
makeArrayRef(TFOps, i));
i = 0;
BytesLeft = BytesLeftSave;
while (BytesLeft) {
VT = getRemainingValueType(BytesLeft);
VTSize = getRemainingSize(BytesLeft);
TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
DAG.getConstant(DstOff, dl, MVT::i32)),
DstPtrInfo.getWithOffset(DstOff));
++i;
DstOff += VTSize;
BytesLeft -= VTSize;
}
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
makeArrayRef(TFOps, i));
}
SDValue
X86SelectionDAGInfo::EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl,
SDValue Chain, SDValue Dst, SDValue Src,
SDValue Size, unsigned Align,
bool isVolatile, bool AlwaysInline,
MachinePointerInfo DstPtrInfo,
MachinePointerInfo SrcPtrInfo) const {
// This requires the copy size to be a constant, preferably
// within a subtarget-specific limit.
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
if (!ConstantSize)
return SDValue();
uint64_t SizeVal = ConstantSize->getZExtValue();
if (!AlwaysInline && SizeVal > Subtarget->getMaxInlineSizeThreshold())
return SDValue();
/// If not DWORD aligned, it is more efficient to call the library. However
/// if calling the library is not allowed (AlwaysInline), then soldier on as
/// the code generated here is better than the long load-store sequence we
/// would otherwise get.
if (!AlwaysInline && (Align & 3) != 0)
return SDValue();
// If to a segment-relative address space, use the default lowering.
if (DstPtrInfo.getAddrSpace() >= 256 ||
SrcPtrInfo.getAddrSpace() >= 256)
return SDValue();
// ESI might be used as a base pointer, in that case we can't simply overwrite
// the register. Fall back to generic code.
const X86RegisterInfo *TRI =
static_cast<const X86RegisterInfo *>(DAG.getTarget().getRegisterInfo());
if (TRI->hasBasePointer(DAG.getMachineFunction()) &&
TRI->getBaseRegister() == X86::ESI)
return SDValue();
MVT AVT;
if (Align & 1)
AVT = MVT::i8;
else if (Align & 2)
AVT = MVT::i16;
else if (Align & 4)
// DWORD aligned
AVT = MVT::i32;
else
// QWORD aligned
AVT = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
unsigned UBytes = AVT.getSizeInBits() / 8;
unsigned CountVal = SizeVal / UBytes;
SDValue Count = DAG.getIntPtrConstant(CountVal);
unsigned BytesLeft = SizeVal % UBytes;
SDValue InFlag;
Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
X86::ECX,
Count, InFlag);
InFlag = Chain.getValue(1);
Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
X86::EDI,
Dst, InFlag);
InFlag = Chain.getValue(1);
Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
X86::ESI,
Src, InFlag);
InFlag = Chain.getValue(1);
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops);
SmallVector<SDValue, 4> Results;
Results.push_back(RepMovs);
if (BytesLeft) {
// Handle the last 1 - 7 bytes.
unsigned Offset = SizeVal - BytesLeft;
EVT DstVT = Dst.getValueType();
EVT SrcVT = Src.getValueType();
EVT SizeVT = Size.getValueType();
Results.push_back(DAG.getMemcpy(Chain, dl,
DAG.getNode(ISD::ADD, dl, DstVT, Dst,
DAG.getConstant(Offset, DstVT)),
DAG.getNode(ISD::ADD, dl, SrcVT, Src,
DAG.getConstant(Offset, SrcVT)),
DAG.getConstant(BytesLeft, SizeVT),
Align, isVolatile, AlwaysInline,
DstPtrInfo.getWithOffset(Offset),
SrcPtrInfo.getWithOffset(Offset)));
}
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Results);
}
SDValue
X86SelectionDAGInfo::EmitTargetCodeForMemset(SelectionDAG &DAG, SDLoc dl,
SDValue Chain,
SDValue Dst, SDValue Src,
SDValue Size, unsigned Align,
bool isVolatile,
MachinePointerInfo DstPtrInfo) const {
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
// If to a segment-relative address space, use the default lowering.
if (DstPtrInfo.getAddrSpace() >= 256)
return SDValue();
// If not DWORD aligned or size is more than the threshold, call the library.
// The libc version is likely to be faster for these cases. It can use the
// address value and run time information about the CPU.
if ((Align & 3) != 0 ||
!ConstantSize ||
ConstantSize->getZExtValue() >
Subtarget->getMaxInlineSizeThreshold()) {
// Check to see if there is a specialized entry-point for memory zeroing.
ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
if (const char *bzeroEntry = V &&
V->isNullValue() ? Subtarget->getBZeroEntry() : nullptr) {
EVT IntPtr = TLI.getPointerTy();
Type *IntPtrTy = getDataLayout()->getIntPtrType(*DAG.getContext());
TargetLowering::ArgListTy Args;
TargetLowering::ArgListEntry Entry;
Entry.Node = Dst;
Entry.Ty = IntPtrTy;
Args.push_back(Entry);
Entry.Node = Size;
Args.push_back(Entry);
TargetLowering::CallLoweringInfo CLI(DAG);
CLI.setDebugLoc(dl).setChain(Chain)
.setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
DAG.getExternalSymbol(bzeroEntry, IntPtr), &Args, 0)
.setDiscardResult();
std::pair<SDValue,SDValue> CallResult = TLI.LowerCallTo(CLI);
return CallResult.second;
}
// Otherwise have the target-independent code call memset.
return SDValue();
}
uint64_t SizeVal = ConstantSize->getZExtValue();
SDValue InFlag;
EVT AVT;
SDValue Count;
ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
unsigned BytesLeft = 0;
bool TwoRepStos = false;
if (ValC) {
unsigned ValReg;
uint64_t Val = ValC->getZExtValue() & 255;
// If the value is a constant, then we can potentially use larger sets.
switch (Align & 3) {
case 2: // WORD aligned
AVT = MVT::i16;
ValReg = X86::AX;
Val = (Val << 8) | Val;
break;
case 0: // DWORD aligned
AVT = MVT::i32;
ValReg = X86::EAX;
Val = (Val << 8) | Val;
Val = (Val << 16) | Val;
if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
AVT = MVT::i64;
ValReg = X86::RAX;
Val = (Val << 32) | Val;
}
break;
default: // Byte aligned
AVT = MVT::i8;
ValReg = X86::AL;
Count = DAG.getIntPtrConstant(SizeVal);
break;
}
if (AVT.bitsGT(MVT::i8)) {
unsigned UBytes = AVT.getSizeInBits() / 8;
Count = DAG.getIntPtrConstant(SizeVal / UBytes);
BytesLeft = SizeVal % UBytes;
}
Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
InFlag);
InFlag = Chain.getValue(1);
} else {
AVT = MVT::i8;
Count = DAG.getIntPtrConstant(SizeVal);
Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
InFlag = Chain.getValue(1);
}
Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
X86::ECX,
Count, InFlag);
InFlag = Chain.getValue(1);
Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
X86::EDI,
Dst, InFlag);
InFlag = Chain.getValue(1);
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops);
if (TwoRepStos) {
InFlag = Chain.getValue(1);
Count = Size;
EVT CVT = Count.getValueType();
SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
X86::ECX,
Left, InFlag);
InFlag = Chain.getValue(1);
Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops);
} else if (BytesLeft) {
// Handle the last 1 - 7 bytes.
unsigned Offset = SizeVal - BytesLeft;
EVT AddrVT = Dst.getValueType();
EVT SizeVT = Size.getValueType();
Chain = DAG.getMemset(Chain, dl,
DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
DAG.getConstant(Offset, AddrVT)),
Src,
DAG.getConstant(BytesLeft, SizeVT),
Align, isVolatile, DstPtrInfo.getWithOffset(Offset));
}
// TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
return Chain;
}
void DAGTypeLegalizer::ExpandRes_BITCAST(SDNode *N, SDValue &Lo, SDValue &Hi) {
EVT OutVT = N->getValueType(0);
EVT NOutVT = TLI.getTypeToTransformTo(*DAG.getContext(), OutVT);
SDValue InOp = N->getOperand(0);
EVT InVT = InOp.getValueType();
SDLoc dl(N);
// Handle some special cases efficiently.
switch (getTypeAction(InVT)) {
case TargetLowering::TypeLegal:
case TargetLowering::TypePromoteInteger:
break;
case TargetLowering::TypePromoteFloat:
llvm_unreachable("Bitcast of a promotion-needing float should never need"
"expansion");
case TargetLowering::TypeSoftenFloat:
// Convert the integer operand instead.
SplitInteger(GetSoftenedFloat(InOp), Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeExpandInteger:
case TargetLowering::TypeExpandFloat:
// Convert the expanded pieces of the input.
GetExpandedOp(InOp, Lo, Hi);
if (TLI.hasBigEndianPartOrdering(InVT) !=
TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeSplitVector:
GetSplitVector(InOp, Lo, Hi);
if (TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeScalarizeVector:
// Convert the element instead.
SplitInteger(BitConvertToInteger(GetScalarizedVector(InOp)), Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeWidenVector: {
assert(!(InVT.getVectorNumElements() & 1) && "Unsupported BITCAST");
InOp = GetWidenedVector(InOp);
EVT LoVT, HiVT;
std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(InVT);
std::tie(Lo, Hi) = DAG.SplitVector(InOp, dl, LoVT, HiVT);
if (TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
}
}
if (InVT.isVector() && OutVT.isInteger()) {
// Handle cases like i64 = BITCAST v1i64 on x86, where the operand
// is legal but the result is not.
unsigned NumElems = 2;
EVT ElemVT = NOutVT;
EVT NVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, NumElems);
// If <ElemVT * N> is not a legal type, try <ElemVT/2 * (N*2)>.
while (!isTypeLegal(NVT)) {
unsigned NewSizeInBits = ElemVT.getSizeInBits() / 2;
// If the element size is smaller than byte, bail.
if (NewSizeInBits < 8)
break;
NumElems *= 2;
ElemVT = EVT::getIntegerVT(*DAG.getContext(), NewSizeInBits);
NVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, NumElems);
}
if (isTypeLegal(NVT)) {
SDValue CastInOp = DAG.getNode(ISD::BITCAST, dl, NVT, InOp);
SmallVector<SDValue, 8> Vals;
for (unsigned i = 0; i < NumElems; ++i)
Vals.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ElemVT,
CastInOp, DAG.getConstant(i, dl,
TLI.getVectorIdxTy())));
// Build Lo, Hi pair by pairing extracted elements if needed.
unsigned Slot = 0;
for (unsigned e = Vals.size(); e - Slot > 2; Slot += 2, e += 1) {
// Each iteration will BUILD_PAIR two nodes and append the result until
// there are only two nodes left, i.e. Lo and Hi.
SDValue LHS = Vals[Slot];
SDValue RHS = Vals[Slot + 1];
if (TLI.isBigEndian())
std::swap(LHS, RHS);
Vals.push_back(DAG.getNode(ISD::BUILD_PAIR, dl,
EVT::getIntegerVT(
*DAG.getContext(),
LHS.getValueType().getSizeInBits() << 1),
LHS, RHS));
}
Lo = Vals[Slot++];
Hi = Vals[Slot++];
if (TLI.isBigEndian())
std::swap(Lo, Hi);
return;
}
}
// Lower the bit-convert to a store/load from the stack.
assert(NOutVT.isByteSized() && "Expanded type not byte sized!");
// Create the stack frame object. Make sure it is aligned for both
// the source and expanded destination types.
unsigned Alignment =
TLI.getDataLayout()->getPrefTypeAlignment(NOutVT.
getTypeForEVT(*DAG.getContext()));
SDValue StackPtr = DAG.CreateStackTemporary(InVT, Alignment);
int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
// Emit a store to the stack slot.
SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, InOp, StackPtr, PtrInfo,
false, false, 0);
// Load the first half from the stack slot.
Lo = DAG.getLoad(NOutVT, dl, Store, StackPtr, PtrInfo,
false, false, false, 0);
// Increment the pointer to the other half.
unsigned IncrementSize = NOutVT.getSizeInBits() / 8;
StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
DAG.getConstant(IncrementSize, dl,
StackPtr.getValueType()));
// Load the second half from the stack slot.
Hi = DAG.getLoad(NOutVT, dl, Store, StackPtr,
PtrInfo.getWithOffset(IncrementSize), false,
false, false, MinAlign(Alignment, IncrementSize));
// Handle endianness of the load.
if (TLI.hasBigEndianPartOrdering(OutVT))
std::swap(Lo, Hi);
}
SDValue SystemZSelectionDAGInfo::
EmitTargetCodeForMemset(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
SDValue Dst, SDValue Byte, SDValue Size,
unsigned Align, bool IsVolatile,
MachinePointerInfo DstPtrInfo) const {
EVT PtrVT = Dst.getValueType();
if (IsVolatile)
return SDValue();
if (ConstantSDNode *CSize = dyn_cast<ConstantSDNode>(Size)) {
uint64_t Bytes = CSize->getZExtValue();
if (Bytes == 0)
return SDValue();
if (ConstantSDNode *CByte = dyn_cast<ConstantSDNode>(Byte)) {
// Handle cases that can be done using at most two of
// MVI, MVHI, MVHHI and MVGHI. The latter two can only be
// used if ByteVal is all zeros or all ones; in other casees,
// we can move at most 2 halfwords.
uint64_t ByteVal = CByte->getZExtValue();
if (ByteVal == 0 || ByteVal == 255 ?
Bytes <= 16 && CountPopulation_64(Bytes) <= 2 :
Bytes <= 4) {
unsigned Size1 = Bytes == 16 ? 8 : 1 << findLastSet(Bytes);
unsigned Size2 = Bytes - Size1;
SDValue Chain1 = memsetStore(DAG, DL, Chain, Dst, ByteVal, Size1,
Align, DstPtrInfo);
if (Size2 == 0)
return Chain1;
Dst = DAG.getNode(ISD::ADD, DL, PtrVT, Dst,
DAG.getConstant(Size1, PtrVT));
DstPtrInfo = DstPtrInfo.getWithOffset(Size1);
SDValue Chain2 = memsetStore(DAG, DL, Chain, Dst, ByteVal, Size2,
std::min(Align, Size1), DstPtrInfo);
return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chain1, Chain2);
}
} else {
// Handle one and two bytes using STC.
if (Bytes <= 2) {
SDValue Chain1 = DAG.getStore(Chain, DL, Byte, Dst, DstPtrInfo,
false, false, Align);
if (Bytes == 1)
return Chain1;
SDValue Dst2 = DAG.getNode(ISD::ADD, DL, PtrVT, Dst,
DAG.getConstant(1, PtrVT));
SDValue Chain2 = DAG.getStore(Chain, DL, Byte, Dst2,
DstPtrInfo.getWithOffset(1),
false, false, 1);
return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chain1, Chain2);
}
}
assert(Bytes >= 2 && "Should have dealt with 0- and 1-byte cases already");
// Copy the byte to the first location and then use MVC to copy
// it to the rest.
Chain = DAG.getStore(Chain, DL, Byte, Dst, DstPtrInfo,
false, false, Align);
SDValue DstPlus1 = DAG.getNode(ISD::ADD, DL, PtrVT, Dst,
DAG.getConstant(1, PtrVT));
return emitMVC(DAG, DL, Chain, DstPlus1, Dst, Bytes - 1);
}
return SDValue();
}
SDValue
ARMSelectionDAGInfo::EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl,
SDValue Chain,
SDValue Dst, SDValue Src,
SDValue Size, unsigned Align,
bool isVolatile, bool AlwaysInline,
MachinePointerInfo DstPtrInfo,
MachinePointerInfo SrcPtrInfo) const {
// Do repeated 4-byte loads and stores. To be improved.
// This requires 4-byte alignment.
if ((Align & 3) != 0)
return SDValue();
// This requires the copy size to be a constant, preferably
// within a subtarget-specific limit.
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
if (!ConstantSize)
return SDValue();
uint64_t SizeVal = ConstantSize->getZExtValue();
if (!AlwaysInline && SizeVal > Subtarget->getMaxInlineSizeThreshold())
return SDValue();
unsigned BytesLeft = SizeVal & 3;
unsigned NumMemOps = SizeVal >> 2;
unsigned EmittedNumMemOps = 0;
EVT VT = MVT::i32;
unsigned VTSize = 4;
unsigned i = 0;
const unsigned MAX_LOADS_IN_LDM = 6;
SDValue TFOps[MAX_LOADS_IN_LDM];
SDValue Loads[MAX_LOADS_IN_LDM];
uint64_t SrcOff = 0, DstOff = 0;
// Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
// same number of stores. The loads and stores will get combined into
// ldm/stm later on.
while (EmittedNumMemOps < NumMemOps) {
for (i = 0;
i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
Loads[i] = DAG.getLoad(VT, dl, Chain,
DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
DAG.getConstant(SrcOff, MVT::i32)),
SrcPtrInfo.getWithOffset(SrcOff), isVolatile,
false, false, 0);
TFOps[i] = Loads[i].getValue(1);
SrcOff += VTSize;
}
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
for (i = 0;
i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
DAG.getConstant(DstOff, MVT::i32)),
DstPtrInfo.getWithOffset(DstOff),
isVolatile, false, 0);
DstOff += VTSize;
}
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
EmittedNumMemOps += i;
}
if (BytesLeft == 0)
return Chain;
// Issue loads / stores for the trailing (1 - 3) bytes.
unsigned BytesLeftSave = BytesLeft;
i = 0;
while (BytesLeft) {
if (BytesLeft >= 2) {
VT = MVT::i16;
VTSize = 2;
} else {
VT = MVT::i8;
VTSize = 1;
}
Loads[i] = DAG.getLoad(VT, dl, Chain,
DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
DAG.getConstant(SrcOff, MVT::i32)),
SrcPtrInfo.getWithOffset(SrcOff),
false, false, false, 0);
TFOps[i] = Loads[i].getValue(1);
++i;
SrcOff += VTSize;
BytesLeft -= VTSize;
}
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
i = 0;
BytesLeft = BytesLeftSave;
while (BytesLeft) {
if (BytesLeft >= 2) {
VT = MVT::i16;
VTSize = 2;
} else {
VT = MVT::i8;
VTSize = 1;
}
TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
DAG.getConstant(DstOff, MVT::i32)),
DstPtrInfo.getWithOffset(DstOff), false, false, 0);
++i;
DstOff += VTSize;
BytesLeft -= VTSize;
}
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
}
SDValue X86SelectionDAGInfo::EmitTargetCodeForMemcpy(
SelectionDAG &DAG, const SDLoc &dl, SDValue Chain, SDValue Dst, SDValue Src,
SDValue Size, unsigned Align, bool isVolatile, bool AlwaysInline,
MachinePointerInfo DstPtrInfo, MachinePointerInfo SrcPtrInfo) const {
// This requires the copy size to be a constant, preferably
// within a subtarget-specific limit.
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
const X86Subtarget &Subtarget =
DAG.getMachineFunction().getSubtarget<X86Subtarget>();
if (!ConstantSize)
return SDValue();
RepMovsRepeats Repeats(ConstantSize->getZExtValue());
if (!AlwaysInline && Repeats.Size > Subtarget.getMaxInlineSizeThreshold())
return SDValue();
/// If not DWORD aligned, it is more efficient to call the library. However
/// if calling the library is not allowed (AlwaysInline), then soldier on as
/// the code generated here is better than the long load-store sequence we
/// would otherwise get.
if (!AlwaysInline && (Align & 3) != 0)
return SDValue();
// If to a segment-relative address space, use the default lowering.
if (DstPtrInfo.getAddrSpace() >= 256 ||
SrcPtrInfo.getAddrSpace() >= 256)
return SDValue();
// If the base register might conflict with our physical registers, bail out.
const MCPhysReg ClobberSet[] = {X86::RCX, X86::RSI, X86::RDI,
X86::ECX, X86::ESI, X86::EDI};
if (isBaseRegConflictPossible(DAG, ClobberSet))
return SDValue();
// If the target has enhanced REPMOVSB, then it's at least as fast to use
// REP MOVSB instead of REP MOVS{W,D,Q}, and it avoids having to handle
// BytesLeft.
if (!Subtarget.hasERMSB() && !(Align & 1)) {
if (Align & 2)
// WORD aligned
Repeats.AVT = MVT::i16;
else if (Align & 4)
// DWORD aligned
Repeats.AVT = MVT::i32;
else
// QWORD aligned
Repeats.AVT = Subtarget.is64Bit() ? MVT::i64 : MVT::i32;
if (Repeats.BytesLeft() > 0 &&
DAG.getMachineFunction().getFunction().optForMinSize()) {
// When agressively optimizing for size, avoid generating the code to
// handle BytesLeft.
Repeats.AVT = MVT::i8;
}
}
SDValue InFlag;
Chain = DAG.getCopyToReg(Chain, dl, Subtarget.is64Bit() ? X86::RCX : X86::ECX,
DAG.getIntPtrConstant(Repeats.Count(), dl), InFlag);
InFlag = Chain.getValue(1);
Chain = DAG.getCopyToReg(Chain, dl, Subtarget.is64Bit() ? X86::RDI : X86::EDI,
Dst, InFlag);
InFlag = Chain.getValue(1);
Chain = DAG.getCopyToReg(Chain, dl, Subtarget.is64Bit() ? X86::RSI : X86::ESI,
Src, InFlag);
InFlag = Chain.getValue(1);
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue Ops[] = { Chain, DAG.getValueType(Repeats.AVT), InFlag };
SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops);
SmallVector<SDValue, 4> Results;
Results.push_back(RepMovs);
if (Repeats.BytesLeft()) {
// Handle the last 1 - 7 bytes.
unsigned Offset = Repeats.Size - Repeats.BytesLeft();
EVT DstVT = Dst.getValueType();
EVT SrcVT = Src.getValueType();
EVT SizeVT = Size.getValueType();
Results.push_back(DAG.getMemcpy(Chain, dl,
DAG.getNode(ISD::ADD, dl, DstVT, Dst,
DAG.getConstant(Offset, dl,
DstVT)),
DAG.getNode(ISD::ADD, dl, SrcVT, Src,
DAG.getConstant(Offset, dl,
SrcVT)),
DAG.getConstant(Repeats.BytesLeft(), dl,
SizeVT),
Align, isVolatile, AlwaysInline, false,
DstPtrInfo.getWithOffset(Offset),
SrcPtrInfo.getWithOffset(Offset)));
}
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Results);
}
SDValue
X86SelectionDAGInfo::EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl,
SDValue Chain, SDValue Dst, SDValue Src,
SDValue Size, unsigned Align,
bool isVolatile, bool AlwaysInline,
MachinePointerInfo DstPtrInfo,
MachinePointerInfo SrcPtrInfo) const {
// This requires the copy size to be a constant, preferably
// within a subtarget-specific limit.
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
if (!ConstantSize)
return SDValue();
uint64_t SizeVal = ConstantSize->getZExtValue();
if (!AlwaysInline && SizeVal > Subtarget->getMaxInlineSizeThreshold())
return SDValue();
/// If not DWORD aligned, it is more efficient to call the library. However
/// if calling the library is not allowed (AlwaysInline), then soldier on as
/// the code generated here is better than the long load-store sequence we
/// would otherwise get.
if (!AlwaysInline && (Align & 3) != 0)
return SDValue();
// If to a segment-relative address space, use the default lowering.
if (DstPtrInfo.getAddrSpace() >= 256 ||
SrcPtrInfo.getAddrSpace() >= 256)
return SDValue();
// If ESI is used as a base pointer, we must preserve it when doing rep movs.
const X86RegisterInfo *TRI =
static_cast<const X86RegisterInfo *>(DAG.getTarget().getRegisterInfo());
bool PreserveESI = TRI->hasBasePointer(DAG.getMachineFunction()) &&
TRI->getBaseRegister() == X86::ESI;
MVT AVT;
if (Align & 1)
AVT = MVT::i8;
else if (Align & 2)
AVT = MVT::i16;
else if (Align & 4)
// DWORD aligned
AVT = MVT::i32;
else
// QWORD aligned
AVT = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
unsigned UBytes = AVT.getSizeInBits() / 8;
unsigned CountVal = SizeVal / UBytes;
SDValue Count = DAG.getIntPtrConstant(CountVal);
unsigned BytesLeft = SizeVal % UBytes;
if (PreserveESI) {
// Save ESI to a physical register. (We cannot use a virtual register
// because if it is spilled we wouldn't be able to reload it.)
// We don't glue this because the register dependencies are explicit.
Chain = DAG.getCopyToReg(Chain, dl, X86::EDX,
DAG.getRegister(X86::ESI, MVT::i32));
}
SDValue InGlue(0, 0);
Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
X86::ECX,
Count, InGlue);
InGlue = Chain.getValue(1);
Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
X86::EDI,
Dst, InGlue);
InGlue = Chain.getValue(1);
Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
X86::ESI,
Src, InGlue);
InGlue = Chain.getValue(1);
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue Ops[] = { Chain, DAG.getValueType(AVT), InGlue };
// FIXME: Make X86rep_movs explicitly use FCX, RDI, RSI instead of glue.
SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
array_lengthof(Ops));
if (PreserveESI) {
InGlue = RepMovs.getValue(1);
RepMovs = DAG.getCopyToReg(RepMovs, dl, X86::ESI,
DAG.getRegister(X86::EDX, MVT::i32), InGlue);
}
SmallVector<SDValue, 4> Results;
Results.push_back(RepMovs);
if (BytesLeft) {
// Handle the last 1 - 7 bytes.
unsigned Offset = SizeVal - BytesLeft;
EVT DstVT = Dst.getValueType();
EVT SrcVT = Src.getValueType();
EVT SizeVT = Size.getValueType();
Results.push_back(DAG.getMemcpy(Chain, dl,
DAG.getNode(ISD::ADD, dl, DstVT, Dst,
DAG.getConstant(Offset, DstVT)),
DAG.getNode(ISD::ADD, dl, SrcVT, Src,
DAG.getConstant(Offset, SrcVT)),
DAG.getConstant(BytesLeft, SizeVT),
Align, isVolatile, AlwaysInline,
DstPtrInfo.getWithOffset(Offset),
SrcPtrInfo.getWithOffset(Offset)));
}
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
&Results[0], Results.size());
}
void DAGTypeLegalizer::ExpandRes_BITCAST(SDNode *N, SDValue &Lo, SDValue &Hi) {
EVT OutVT = N->getValueType(0);
EVT NOutVT = TLI.getTypeToTransformTo(*DAG.getContext(), OutVT);
SDValue InOp = N->getOperand(0);
EVT InVT = InOp.getValueType();
DebugLoc dl = N->getDebugLoc();
// Handle some special cases efficiently.
switch (getTypeAction(InVT)) {
default:
assert(false && "Unknown type action!");
case TargetLowering::TypeLegal:
case TargetLowering::TypePromoteInteger:
break;
case TargetLowering::TypeSoftenFloat:
// Convert the integer operand instead.
SplitInteger(GetSoftenedFloat(InOp), Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeExpandInteger:
case TargetLowering::TypeExpandFloat:
// Convert the expanded pieces of the input.
GetExpandedOp(InOp, Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeSplitVector:
GetSplitVector(InOp, Lo, Hi);
if (TLI.isBigEndian())
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeScalarizeVector:
// Convert the element instead.
SplitInteger(BitConvertToInteger(GetScalarizedVector(InOp)), Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
case TargetLowering::TypeWidenVector: {
assert(!(InVT.getVectorNumElements() & 1) && "Unsupported BITCAST");
InOp = GetWidenedVector(InOp);
EVT InNVT = EVT::getVectorVT(*DAG.getContext(), InVT.getVectorElementType(),
InVT.getVectorNumElements()/2);
Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, InOp,
DAG.getIntPtrConstant(0));
Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InNVT, InOp,
DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
if (TLI.isBigEndian())
std::swap(Lo, Hi);
Lo = DAG.getNode(ISD::BITCAST, dl, NOutVT, Lo);
Hi = DAG.getNode(ISD::BITCAST, dl, NOutVT, Hi);
return;
}
}
if (InVT.isVector() && OutVT.isInteger()) {
// Handle cases like i64 = BITCAST v1i64 on x86, where the operand
// is legal but the result is not.
EVT NVT = EVT::getVectorVT(*DAG.getContext(), NOutVT, 2);
if (isTypeLegal(NVT)) {
SDValue CastInOp = DAG.getNode(ISD::BITCAST, dl, NVT, InOp);
Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NOutVT, CastInOp,
DAG.getIntPtrConstant(0));
Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NOutVT, CastInOp,
DAG.getIntPtrConstant(1));
if (TLI.isBigEndian())
std::swap(Lo, Hi);
return;
}
}
// Lower the bit-convert to a store/load from the stack.
assert(NOutVT.isByteSized() && "Expanded type not byte sized!");
// Create the stack frame object. Make sure it is aligned for both
// the source and expanded destination types.
unsigned Alignment =
TLI.getTargetData()->getPrefTypeAlignment(NOutVT.
getTypeForEVT(*DAG.getContext()));
SDValue StackPtr = DAG.CreateStackTemporary(InVT, Alignment);
int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
// Emit a store to the stack slot.
SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, InOp, StackPtr, PtrInfo,
false, false, 0);
// Load the first half from the stack slot.
Lo = DAG.getLoad(NOutVT, dl, Store, StackPtr, PtrInfo, false, false, 0);
// Increment the pointer to the other half.
unsigned IncrementSize = NOutVT.getSizeInBits() / 8;
StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
DAG.getIntPtrConstant(IncrementSize));
// Load the second half from the stack slot.
Hi = DAG.getLoad(NOutVT, dl, Store, StackPtr,
PtrInfo.getWithOffset(IncrementSize), false,
false, MinAlign(Alignment, IncrementSize));
// Handle endianness of the load.
if (TLI.isBigEndian())
std::swap(Lo, Hi);
}
SDValue X86SelectionDAGInfo::EmitTargetCodeForMemset(
SelectionDAG &DAG, const SDLoc &dl, SDValue Chain, SDValue Dst, SDValue Val,
SDValue Size, unsigned Align, bool isVolatile,
MachinePointerInfo DstPtrInfo) const {
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
const X86Subtarget &Subtarget =
DAG.getMachineFunction().getSubtarget<X86Subtarget>();
#ifndef NDEBUG
// If the base register might conflict with our physical registers, bail out.
const MCPhysReg ClobberSet[] = {X86::RCX, X86::RAX, X86::RDI,
X86::ECX, X86::EAX, X86::EDI};
assert(!isBaseRegConflictPossible(DAG, ClobberSet));
#endif
// If to a segment-relative address space, use the default lowering.
if (DstPtrInfo.getAddrSpace() >= 256)
return SDValue();
// If not DWORD aligned or size is more than the threshold, call the library.
// The libc version is likely to be faster for these cases. It can use the
// address value and run time information about the CPU.
if ((Align & 3) != 0 || !ConstantSize ||
ConstantSize->getZExtValue() > Subtarget.getMaxInlineSizeThreshold()) {
// Check to see if there is a specialized entry-point for memory zeroing.
ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Val);
if (const char *bzeroName = (ValC && ValC->isNullValue())
? DAG.getTargetLoweringInfo().getLibcallName(RTLIB::BZERO)
: nullptr) {
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
EVT IntPtr = TLI.getPointerTy(DAG.getDataLayout());
Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(*DAG.getContext());
TargetLowering::ArgListTy Args;
TargetLowering::ArgListEntry Entry;
Entry.Node = Dst;
Entry.Ty = IntPtrTy;
Args.push_back(Entry);
Entry.Node = Size;
Args.push_back(Entry);
TargetLowering::CallLoweringInfo CLI(DAG);
CLI.setDebugLoc(dl)
.setChain(Chain)
.setLibCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
DAG.getExternalSymbol(bzeroName, IntPtr),
std::move(Args))
.setDiscardResult();
std::pair<SDValue,SDValue> CallResult = TLI.LowerCallTo(CLI);
return CallResult.second;
}
// Otherwise have the target-independent code call memset.
return SDValue();
}
uint64_t SizeVal = ConstantSize->getZExtValue();
SDValue InFlag;
EVT AVT;
SDValue Count;
ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Val);
unsigned BytesLeft = 0;
if (ValC) {
unsigned ValReg;
uint64_t Val = ValC->getZExtValue() & 255;
// If the value is a constant, then we can potentially use larger sets.
switch (Align & 3) {
case 2: // WORD aligned
AVT = MVT::i16;
ValReg = X86::AX;
Val = (Val << 8) | Val;
break;
case 0: // DWORD aligned
AVT = MVT::i32;
ValReg = X86::EAX;
Val = (Val << 8) | Val;
Val = (Val << 16) | Val;
if (Subtarget.is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
AVT = MVT::i64;
ValReg = X86::RAX;
Val = (Val << 32) | Val;
}
break;
default: // Byte aligned
AVT = MVT::i8;
ValReg = X86::AL;
Count = DAG.getIntPtrConstant(SizeVal, dl);
break;
}
if (AVT.bitsGT(MVT::i8)) {
unsigned UBytes = AVT.getSizeInBits() / 8;
Count = DAG.getIntPtrConstant(SizeVal / UBytes, dl);
BytesLeft = SizeVal % UBytes;
}
Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, dl, AVT),
InFlag);
InFlag = Chain.getValue(1);
} else {
AVT = MVT::i8;
Count = DAG.getIntPtrConstant(SizeVal, dl);
Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Val, InFlag);
InFlag = Chain.getValue(1);
}
bool Use64BitRegs = Subtarget.isTarget64BitLP64();
Chain = DAG.getCopyToReg(Chain, dl, Use64BitRegs ? X86::RCX : X86::ECX,
Count, InFlag);
InFlag = Chain.getValue(1);
Chain = DAG.getCopyToReg(Chain, dl, Use64BitRegs ? X86::RDI : X86::EDI,
Dst, InFlag);
InFlag = Chain.getValue(1);
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops);
if (BytesLeft) {
// Handle the last 1 - 7 bytes.
unsigned Offset = SizeVal - BytesLeft;
EVT AddrVT = Dst.getValueType();
EVT SizeVT = Size.getValueType();
Chain = DAG.getMemset(Chain, dl,
DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
DAG.getConstant(Offset, dl, AddrVT)),
Val,
DAG.getConstant(BytesLeft, dl, SizeVT),
Align, isVolatile, false,
DstPtrInfo.getWithOffset(Offset));
}
// TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
return Chain;
}
