#define DEBUG_TYPE "legalizedag"
+namespace {
+
+struct FloatSignAsInt;
+
//===----------------------------------------------------------------------===//
/// This takes an arbitrary SelectionDAG as input and
/// hacks on it until the target machine can handle it. This involves
/// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
/// will attempt merge setcc and brc instructions into brcc's.
///
-namespace {
class SelectionDAGLegalize {
const TargetMachine &TM;
const TargetLowering &TLI;
SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
SmallVectorImpl<SDValue> &Results);
- SDValue ExpandFCOPYSIGN(SDNode *Node);
+ void getSignAsIntValue(FloatSignAsInt &State, SDLoc DL, SDValue Value) const;
+ SDValue modifySignAsInt(const FloatSignAsInt &State, SDLoc DL,
+ SDValue NewIntValue) const;
+ SDValue ExpandFCOPYSIGN(SDNode *Node) const;
+ SDValue ExpandFABS(SDNode *Node) const;
SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
SDLoc dl);
SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
SDLoc dl);
+ SDValue ExpandBITREVERSE(SDValue Op, SDLoc dl);
SDValue ExpandBSWAP(SDValue Op, SDLoc dl);
SDValue ExpandBitCount(unsigned Opc, SDValue Op, SDLoc dl);
SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
- std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
-
- void ExpandNode(SDNode *Node);
+ // if ExpandNode returns false, LegalizeOp falls back to ConvertNodeToLibcall
+ bool ExpandNode(SDNode *Node);
+ void ConvertNodeToLibcall(SDNode *Node);
void PromoteNode(SDNode *Node);
public:
case ISD::SETCC:
case ISD::BR_CC: {
unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
- Node->getOpcode() == ISD::SETCC ? 2 : 1;
+ Node->getOpcode() == ISD::SETCC ? 2 :
+ Node->getOpcode() == ISD::SETCCE ? 3 : 1;
unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
ISD::CondCode CCCode =
}
// FALL THROUGH
case TargetLowering::Expand:
- ExpandNode(Node);
+ if (ExpandNode(Node))
+ return;
+ // FALL THROUGH
+ case TargetLowering::LibCall:
+ ConvertNodeToLibcall(Node);
return;
case TargetLowering::Promote:
PromoteNode(Node);
false, false, false, 0);
}
-SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
- SDLoc dl(Node);
- SDValue Tmp1 = Node->getOperand(0);
- SDValue Tmp2 = Node->getOperand(1);
-
- // Get the sign bit of the RHS. First obtain a value that has the same
- // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
- SDValue SignBit;
- EVT FloatVT = Tmp2.getValueType();
- EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
+namespace {
+/// Keeps track of state when getting the sign of a floating-point value as an
+/// integer.
+struct FloatSignAsInt {
+ EVT FloatVT;
+ SDValue Chain;
+ SDValue FloatPtr;
+ SDValue IntPtr;
+ MachinePointerInfo IntPointerInfo;
+ MachinePointerInfo FloatPointerInfo;
+ SDValue IntValue;
+ APInt SignMask;
+};
+}
+
+/// Bitcast a floating-point value to an integer value. Only bitcast the part
+/// containing the sign bit if the target has no integer value capable of
+/// holding all bits of the floating-point value.
+void SelectionDAGLegalize::getSignAsIntValue(FloatSignAsInt &State,
+ SDLoc DL, SDValue Value) const {
+ EVT FloatVT = Value.getValueType();
+ unsigned NumBits = FloatVT.getSizeInBits();
+ State.FloatVT = FloatVT;
+ EVT IVT = EVT::getIntegerVT(*DAG.getContext(), NumBits);
+ // Convert to an integer of the same size.
if (TLI.isTypeLegal(IVT)) {
- // Convert to an integer with the same sign bit.
- SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
+ State.IntValue = DAG.getNode(ISD::BITCAST, DL, IVT, Value);
+ State.SignMask = APInt::getSignBit(NumBits);
+ return;
+ }
+
+ auto &DataLayout = DAG.getDataLayout();
+ // Store the float to memory, then load the sign part out as an integer.
+ MVT LoadTy = TLI.getRegisterType(*DAG.getContext(), MVT::i8);
+ // First create a temporary that is aligned for both the load and store.
+ SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
+ int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
+ // Then store the float to it.
+ State.FloatPtr = StackPtr;
+ MachineFunction &MF = DAG.getMachineFunction();
+ State.FloatPointerInfo = MachinePointerInfo::getFixedStack(MF, FI);
+ State.Chain = DAG.getStore(DAG.getEntryNode(), DL, Value, State.FloatPtr,
+ State.FloatPointerInfo, false, false, 0);
+
+ SDValue IntPtr;
+ if (DataLayout.isBigEndian()) {
+ assert(FloatVT.isByteSized() && "Unsupported floating point type!");
+ // Load out a legal integer with the same sign bit as the float.
+ IntPtr = StackPtr;
+ State.IntPointerInfo = State.FloatPointerInfo;
} else {
- auto &DL = DAG.getDataLayout();
- // Store the float to memory, then load the sign part out as an integer.
- MVT LoadTy = TLI.getPointerTy(DL);
- // First create a temporary that is aligned for both the load and store.
- SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
- // Then store the float to it.
- SDValue Ch =
- DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
- false, false, 0);
- if (DL.isBigEndian()) {
- assert(FloatVT.isByteSized() && "Unsupported floating point type!");
- // Load out a legal integer with the same sign bit as the float.
- SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
- false, false, false, 0);
- } else { // Little endian
- SDValue LoadPtr = StackPtr;
- // The float may be wider than the integer we are going to load. Advance
- // the pointer so that the loaded integer will contain the sign bit.
- unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
- unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
- LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(), LoadPtr,
- DAG.getConstant(ByteOffset, dl,
- LoadPtr.getValueType()));
- // Load a legal integer containing the sign bit.
- SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
- false, false, false, 0);
- // Move the sign bit to the top bit of the loaded integer.
- unsigned BitShift = LoadTy.getSizeInBits() -
- (FloatVT.getSizeInBits() - 8 * ByteOffset);
- assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
- if (BitShift)
- SignBit = DAG.getNode(
- ISD::SHL, dl, LoadTy, SignBit,
- DAG.getConstant(BitShift, dl,
- TLI.getShiftAmountTy(SignBit.getValueType(), DL)));
- }
+ // Advance the pointer so that the loaded byte will contain the sign bit.
+ unsigned ByteOffset = (FloatVT.getSizeInBits() / 8) - 1;
+ IntPtr = DAG.getNode(ISD::ADD, DL, StackPtr.getValueType(), StackPtr,
+ DAG.getConstant(ByteOffset, DL, StackPtr.getValueType()));
+ State.IntPointerInfo = MachinePointerInfo::getFixedStack(MF, FI,
+ ByteOffset);
+ }
+
+ State.IntPtr = IntPtr;
+ State.IntValue = DAG.getExtLoad(ISD::EXTLOAD, DL, LoadTy, State.Chain,
+ IntPtr, State.IntPointerInfo, MVT::i8,
+ false, false, false, 0);
+ State.SignMask = APInt::getOneBitSet(LoadTy.getSizeInBits(), 7);
+}
+
+/// Replace the integer value produced by getSignAsIntValue() with a new value
+/// and cast the result back to a floating-point type.
+SDValue SelectionDAGLegalize::modifySignAsInt(const FloatSignAsInt &State,
+ SDLoc DL, SDValue NewIntValue) const {
+ if (!State.Chain)
+ return DAG.getNode(ISD::BITCAST, DL, State.FloatVT, NewIntValue);
+
+ // Override the part containing the sign bit in the value stored on the stack.
+ SDValue Chain = DAG.getTruncStore(State.Chain, DL, NewIntValue, State.IntPtr,
+ State.IntPointerInfo, MVT::i8, false, false,
+ 0);
+ return DAG.getLoad(State.FloatVT, DL, Chain, State.FloatPtr,
+ State.FloatPointerInfo, false, false, false, 0);
+}
+
+SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode *Node) const {
+ SDLoc DL(Node);
+ SDValue Mag = Node->getOperand(0);
+ SDValue Sign = Node->getOperand(1);
+
+ // Get sign bit into an integer value.
+ FloatSignAsInt SignAsInt;
+ getSignAsIntValue(SignAsInt, DL, Sign);
+
+ EVT IntVT = SignAsInt.IntValue.getValueType();
+ SDValue SignMask = DAG.getConstant(SignAsInt.SignMask, DL, IntVT);
+ SDValue SignBit = DAG.getNode(ISD::AND, DL, IntVT, SignAsInt.IntValue,
+ SignMask);
+
+ // If FABS is legal transform FCOPYSIGN(x, y) => sign(x) ? -FABS(x) : FABS(X)
+ EVT FloatVT = Mag.getValueType();
+ if (TLI.isOperationLegalOrCustom(ISD::FABS, FloatVT) &&
+ TLI.isOperationLegalOrCustom(ISD::FNEG, FloatVT)) {
+ SDValue AbsValue = DAG.getNode(ISD::FABS, DL, FloatVT, Mag);
+ SDValue NegValue = DAG.getNode(ISD::FNEG, DL, FloatVT, AbsValue);
+ SDValue Cond = DAG.getSetCC(DL, getSetCCResultType(IntVT), SignBit,
+ DAG.getConstant(0, DL, IntVT), ISD::SETNE);
+ return DAG.getSelect(DL, FloatVT, Cond, NegValue, AbsValue);
}
- // Now get the sign bit proper, by seeing whether the value is negative.
- SignBit = DAG.getSetCC(dl, getSetCCResultType(SignBit.getValueType()),
- SignBit,
- DAG.getConstant(0, dl, SignBit.getValueType()),
- ISD::SETLT);
- // Get the absolute value of the result.
- SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
- // Select between the nabs and abs value based on the sign bit of
- // the input.
- return DAG.getSelect(dl, AbsVal.getValueType(), SignBit,
- DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
- AbsVal);
+
+ // Transform values to integer, copy the sign bit and transform back.
+ FloatSignAsInt MagAsInt;
+ getSignAsIntValue(MagAsInt, DL, Mag);
+ assert(SignAsInt.SignMask == MagAsInt.SignMask);
+ SDValue ClearSignMask = DAG.getConstant(~SignAsInt.SignMask, DL, IntVT);
+ SDValue ClearedSign = DAG.getNode(ISD::AND, DL, IntVT, MagAsInt.IntValue,
+ ClearSignMask);
+ SDValue CopiedSign = DAG.getNode(ISD::OR, DL, IntVT, ClearedSign, SignBit);
+
+ return modifySignAsInt(MagAsInt, DL, CopiedSign);
+}
+
+SDValue SelectionDAGLegalize::ExpandFABS(SDNode *Node) const {
+ SDLoc DL(Node);
+ SDValue Value = Node->getOperand(0);
+
+ // Transform FABS(x) => FCOPYSIGN(x, 0.0) if FCOPYSIGN is legal.
+ EVT FloatVT = Value.getValueType();
+ if (TLI.isOperationLegalOrCustom(ISD::FCOPYSIGN, FloatVT)) {
+ SDValue Zero = DAG.getConstantFP(0.0, DL, FloatVT);
+ return DAG.getNode(ISD::FCOPYSIGN, DL, FloatVT, Value, Zero);
+ }
+
+ // Transform value to integer, clear the sign bit and transform back.
+ FloatSignAsInt ValueAsInt;
+ getSignAsIntValue(ValueAsInt, DL, Value);
+ EVT IntVT = ValueAsInt.IntValue.getValueType();
+ SDValue ClearSignMask = DAG.getConstant(~ValueAsInt.SignMask, DL, IntVT);
+ SDValue ClearedSign = DAG.getNode(ISD::AND, DL, IntVT, ValueAsInt.IntValue,
+ ClearSignMask);
+ return modifySignAsInt(ValueAsInt, DL, ClearedSign);
}
void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
}
+/// Open code the operations for BITREVERSE.
+SDValue SelectionDAGLegalize::ExpandBITREVERSE(SDValue Op, SDLoc dl) {
+ EVT VT = Op.getValueType();
+ EVT SHVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
+ unsigned Sz = VT.getScalarSizeInBits();
+
+ SDValue Tmp, Tmp2;
+ Tmp = DAG.getConstant(0, dl, VT);
+ for (unsigned I = 0, J = Sz-1; I < Sz; ++I, --J) {
+ if (I < J)
+ Tmp2 =
+ DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(J - I, dl, SHVT));
+ else
+ Tmp2 =
+ DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(I - J, dl, SHVT));
+
+ APInt Shift(Sz, 1);
+ Shift = Shift.shl(J);
+ Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Shift, dl, VT));
+ Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp, Tmp2);
+ }
+
+ return Tmp;
+}
+
/// Open code the operations for BSWAP of the specified operation.
SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, SDLoc dl) {
EVT VT = Op.getValueType();
}
}
-std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
- unsigned Opc = Node->getOpcode();
- MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
- RTLIB::Libcall LC = RTLIB::getATOMIC(Opc, VT);
- assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpected atomic op or value type!");
-
- return ExpandChainLibCall(LC, Node, false);
-}
-
-void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
+bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
SmallVector<SDValue, 8> Results;
SDLoc dl(Node);
SDValue Tmp1, Tmp2, Tmp3, Tmp4;
Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
Results.push_back(Tmp1);
break;
+ case ISD::BITREVERSE:
+ Results.push_back(ExpandBITREVERSE(Node->getOperand(0), dl));
+ break;
case ISD::BSWAP:
Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
break;
Results.push_back(DAG.getConstant(0, dl, MVT::i32));
Results.push_back(Node->getOperand(0));
break;
- case ISD::ATOMIC_FENCE: {
- // If the target didn't lower this, lower it to '__sync_synchronize()' call
- // FIXME: handle "fence singlethread" more efficiently.
- TargetLowering::ArgListTy Args;
-
- TargetLowering::CallLoweringInfo CLI(DAG);
- CLI.setDebugLoc(dl)
- .setChain(Node->getOperand(0))
- .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
- DAG.getExternalSymbol("__sync_synchronize",
- TLI.getPointerTy(DAG.getDataLayout())),
- std::move(Args), 0);
-
- std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
-
- Results.push_back(CallResult.second);
- break;
- }
case ISD::ATOMIC_LOAD: {
// There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
SDValue Zero = DAG.getConstant(0, dl, Node->getValueType(0));
Results.push_back(Swap.getValue(1));
break;
}
- // By default, atomic intrinsics are marked Legal and lowered. Targets
- // which don't support them directly, however, may want libcalls, in which
- // case they mark them Expand, and we get here.
- case ISD::ATOMIC_SWAP:
- case ISD::ATOMIC_LOAD_ADD:
- case ISD::ATOMIC_LOAD_SUB:
- case ISD::ATOMIC_LOAD_AND:
- case ISD::ATOMIC_LOAD_OR:
- case ISD::ATOMIC_LOAD_XOR:
- case ISD::ATOMIC_LOAD_NAND:
- case ISD::ATOMIC_LOAD_MIN:
- case ISD::ATOMIC_LOAD_MAX:
- case ISD::ATOMIC_LOAD_UMIN:
- case ISD::ATOMIC_LOAD_UMAX:
- case ISD::ATOMIC_CMP_SWAP: {
- std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
- Results.push_back(Tmp.first);
- Results.push_back(Tmp.second);
- break;
- }
case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
// Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and
// splits out the success value as a comparison. Expanding the resulting
}
break;
}
- case ISD::TRAP: {
- // If this operation is not supported, lower it to 'abort()' call
- TargetLowering::ArgListTy Args;
- TargetLowering::CallLoweringInfo CLI(DAG);
- CLI.setDebugLoc(dl)
- .setChain(Node->getOperand(0))
- .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
- DAG.getExternalSymbol("abort",
- TLI.getPointerTy(DAG.getDataLayout())),
- std::move(Args), 0);
- std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
-
- Results.push_back(CallResult.second);
- break;
- }
case ISD::FP_ROUND:
case ISD::BITCAST:
Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
Node->getOperand(0));
Results.push_back(Tmp1);
break;
- case ISD::FABS: {
- // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
- EVT VT = Node->getValueType(0);
- Tmp1 = Node->getOperand(0);
- Tmp2 = DAG.getConstantFP(0.0, dl, VT);
- Tmp2 = DAG.getSetCC(dl, getSetCCResultType(Tmp1.getValueType()),
- Tmp1, Tmp2, ISD::SETUGT);
- Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
- Tmp1 = DAG.getSelect(dl, VT, Tmp2, Tmp1, Tmp3);
- Results.push_back(Tmp1);
+ case ISD::FABS:
+ Results.push_back(ExpandFABS(Node));
break;
- }
case ISD::SMIN:
case ISD::SMAX:
case ISD::UMIN:
break;
}
- case ISD::FMINNUM:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::FMIN_F32, RTLIB::FMIN_F64,
- RTLIB::FMIN_F80, RTLIB::FMIN_F128,
- RTLIB::FMIN_PPCF128));
- break;
- case ISD::FMAXNUM:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::FMAX_F32, RTLIB::FMAX_F64,
- RTLIB::FMAX_F80, RTLIB::FMAX_F128,
- RTLIB::FMAX_PPCF128));
- break;
- case ISD::FSQRT:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
- RTLIB::SQRT_F80, RTLIB::SQRT_F128,
- RTLIB::SQRT_PPCF128));
- break;
case ISD::FSIN:
case ISD::FCOS: {
EVT VT = Node->getValueType(0);
- bool isSIN = Node->getOpcode() == ISD::FSIN;
// Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
// fcos which share the same operand and both are used.
if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
&& useSinCos(Node)) {
SDVTList VTs = DAG.getVTList(VT, VT);
Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
- if (!isSIN)
+ if (Node->getOpcode() == ISD::FCOS)
Tmp1 = Tmp1.getValue(1);
Results.push_back(Tmp1);
- } else if (isSIN) {
- Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
- RTLIB::SIN_F80, RTLIB::SIN_F128,
- RTLIB::SIN_PPCF128));
- } else {
- Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
- RTLIB::COS_F80, RTLIB::COS_F128,
- RTLIB::COS_PPCF128));
}
break;
}
- case ISD::FSINCOS:
- // Expand into sincos libcall.
- ExpandSinCosLibCall(Node, Results);
- break;
- case ISD::FLOG:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
- RTLIB::LOG_F80, RTLIB::LOG_F128,
- RTLIB::LOG_PPCF128));
- break;
- case ISD::FLOG2:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
- RTLIB::LOG2_F80, RTLIB::LOG2_F128,
- RTLIB::LOG2_PPCF128));
- break;
- case ISD::FLOG10:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
- RTLIB::LOG10_F80, RTLIB::LOG10_F128,
- RTLIB::LOG10_PPCF128));
- break;
- case ISD::FEXP:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
- RTLIB::EXP_F80, RTLIB::EXP_F128,
- RTLIB::EXP_PPCF128));
- break;
- case ISD::FEXP2:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
- RTLIB::EXP2_F80, RTLIB::EXP2_F128,
- RTLIB::EXP2_PPCF128));
- break;
- case ISD::FTRUNC:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
- RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
- RTLIB::TRUNC_PPCF128));
- break;
- case ISD::FFLOOR:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
- RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
- RTLIB::FLOOR_PPCF128));
- break;
- case ISD::FCEIL:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
- RTLIB::CEIL_F80, RTLIB::CEIL_F128,
- RTLIB::CEIL_PPCF128));
- break;
- case ISD::FRINT:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
- RTLIB::RINT_F80, RTLIB::RINT_F128,
- RTLIB::RINT_PPCF128));
- break;
- case ISD::FNEARBYINT:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
- RTLIB::NEARBYINT_F64,
- RTLIB::NEARBYINT_F80,
- RTLIB::NEARBYINT_F128,
- RTLIB::NEARBYINT_PPCF128));
- break;
- case ISD::FROUND:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
- RTLIB::ROUND_F64,
- RTLIB::ROUND_F80,
- RTLIB::ROUND_F128,
- RTLIB::ROUND_PPCF128));
- break;
- case ISD::FPOWI:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
- RTLIB::POWI_F80, RTLIB::POWI_F128,
- RTLIB::POWI_PPCF128));
- break;
- case ISD::FPOW:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
- RTLIB::POW_F80, RTLIB::POW_F128,
- RTLIB::POW_PPCF128));
- break;
- case ISD::FDIV:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
- RTLIB::DIV_F80, RTLIB::DIV_F128,
- RTLIB::DIV_PPCF128));
- break;
- case ISD::FREM:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
- RTLIB::REM_F80, RTLIB::REM_F128,
- RTLIB::REM_PPCF128));
- break;
- case ISD::FMA:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
- RTLIB::FMA_F80, RTLIB::FMA_F128,
- RTLIB::FMA_PPCF128));
- break;
case ISD::FMAD:
llvm_unreachable("Illegal fmad should never be formed");
- case ISD::FADD:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::ADD_F32, RTLIB::ADD_F64,
- RTLIB::ADD_F80, RTLIB::ADD_F128,
- RTLIB::ADD_PPCF128));
- break;
- case ISD::FMUL:
- Results.push_back(ExpandFPLibCall(Node, RTLIB::MUL_F32, RTLIB::MUL_F64,
- RTLIB::MUL_F80, RTLIB::MUL_F128,
- RTLIB::MUL_PPCF128));
- break;
- case ISD::FP16_TO_FP: {
- if (Node->getValueType(0) == MVT::f32) {
- Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
- break;
+ case ISD::FP16_TO_FP:
+ if (Node->getValueType(0) != MVT::f32) {
+ // We can extend to types bigger than f32 in two steps without changing
+ // the result. Since "f16 -> f32" is much more commonly available, give
+ // CodeGen the option of emitting that before resorting to a libcall.
+ SDValue Res =
+ DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
+ Results.push_back(
+ DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
}
-
- // We can extend to types bigger than f32 in two steps without changing the
- // result. Since "f16 -> f32" is much more commonly available, give CodeGen
- // the option of emitting that before resorting to a libcall.
- SDValue Res =
- DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
- Results.push_back(
- DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
break;
- }
- case ISD::FP_TO_FP16: {
+ case ISD::FP_TO_FP16:
if (!TLI.useSoftFloat() && TM.Options.UnsafeFPMath) {
SDValue Op = Node->getOperand(0);
MVT SVT = Op.getSimpleValueType();
DAG.getIntPtrConstant(0, dl));
Results.push_back(
DAG.getNode(ISD::FP_TO_FP16, dl, MVT::i16, FloatVal));
- break;
}
}
-
- RTLIB::Libcall LC =
- RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
- assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
- Results.push_back(ExpandLibCall(LC, Node, false));
break;
- }
case ISD::ConstantFP: {
ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
// Check to see if this FP immediate is already legal.
Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1, Flags);
Results.push_back(Tmp1);
- } else {
- Results.push_back(ExpandFPLibCall(Node, RTLIB::SUB_F32, RTLIB::SUB_F64,
- RTLIB::SUB_F80, RTLIB::SUB_F128,
- RTLIB::SUB_PPCF128));
}
break;
}
if (TLI.isOperationLegalOrCustom(DivRemOpc, VT)) {
SDVTList VTs = DAG.getVTList(VT, VT);
Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
+ Results.push_back(Tmp1);
} else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
// X % Y -> X-X/Y*Y
Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
- } else if (isSigned)
- Tmp1 = ExpandIntLibCall(Node, true,
- RTLIB::SREM_I8,
- RTLIB::SREM_I16, RTLIB::SREM_I32,
- RTLIB::SREM_I64, RTLIB::SREM_I128);
- else
- Tmp1 = ExpandIntLibCall(Node, false,
- RTLIB::UREM_I8,
- RTLIB::UREM_I16, RTLIB::UREM_I32,
- RTLIB::UREM_I64, RTLIB::UREM_I128);
- Results.push_back(Tmp1);
+ Results.push_back(Tmp1);
+ }
break;
}
case ISD::UDIV:
bool isSigned = Node->getOpcode() == ISD::SDIV;
unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
EVT VT = Node->getValueType(0);
- SDVTList VTs = DAG.getVTList(VT, VT);
- if (TLI.isOperationLegalOrCustom(DivRemOpc, VT))
+ if (TLI.isOperationLegalOrCustom(DivRemOpc, VT)) {
+ SDVTList VTs = DAG.getVTList(VT, VT);
Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
Node->getOperand(1));
- else if (isSigned)
- Tmp1 = ExpandIntLibCall(Node, true,
- RTLIB::SDIV_I8,
- RTLIB::SDIV_I16, RTLIB::SDIV_I32,
- RTLIB::SDIV_I64, RTLIB::SDIV_I128);
- else
- Tmp1 = ExpandIntLibCall(Node, false,
- RTLIB::UDIV_I8,
- RTLIB::UDIV_I16, RTLIB::UDIV_I32,
- RTLIB::UDIV_I64, RTLIB::UDIV_I128);
- Results.push_back(Tmp1);
+ Results.push_back(Tmp1);
+ }
break;
}
case ISD::MULHU:
Results.push_back(Tmp1.getValue(1));
break;
}
- case ISD::SDIVREM:
- case ISD::UDIVREM:
- // Expand into divrem libcall
- ExpandDivRemLibCall(Node, Results);
- break;
case ISD::MUL: {
EVT VT = Node->getValueType(0);
SDVTList VTs = DAG.getVTList(VT, VT);
TLI.getShiftAmountTy(HalfType, DAG.getDataLayout()));
Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
- break;
}
-
- Tmp1 = ExpandIntLibCall(Node, false,
- RTLIB::MUL_I8,
- RTLIB::MUL_I16, RTLIB::MUL_I32,
- RTLIB::MUL_I64, RTLIB::MUL_I128);
- Results.push_back(Tmp1);
break;
}
case ISD::SADDO:
break;
}
+ // Replace the original node with the legalized result.
+ if (Results.empty())
+ return false;
+
+ ReplaceNode(Node, Results.data());
+ return true;
+}
+
+void SelectionDAGLegalize::ConvertNodeToLibcall(SDNode *Node) {
+ SmallVector<SDValue, 8> Results;
+ SDLoc dl(Node);
+ SDValue Tmp1, Tmp2, Tmp3, Tmp4;
+ unsigned Opc = Node->getOpcode();
+ switch (Opc) {
+ case ISD::ATOMIC_FENCE: {
+ // If the target didn't lower this, lower it to '__sync_synchronize()' call
+ // FIXME: handle "fence singlethread" more efficiently.
+ TargetLowering::ArgListTy Args;
+
+ TargetLowering::CallLoweringInfo CLI(DAG);
+ CLI.setDebugLoc(dl)
+ .setChain(Node->getOperand(0))
+ .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
+ DAG.getExternalSymbol("__sync_synchronize",
+ TLI.getPointerTy(DAG.getDataLayout())),
+ std::move(Args), 0);
+
+ std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
+
+ Results.push_back(CallResult.second);
+ break;
+ }
+ // By default, atomic intrinsics are marked Legal and lowered. Targets
+ // which don't support them directly, however, may want libcalls, in which
+ // case they mark them Expand, and we get here.
+ case ISD::ATOMIC_SWAP:
+ case ISD::ATOMIC_LOAD_ADD:
+ case ISD::ATOMIC_LOAD_SUB:
+ case ISD::ATOMIC_LOAD_AND:
+ case ISD::ATOMIC_LOAD_OR:
+ case ISD::ATOMIC_LOAD_XOR:
+ case ISD::ATOMIC_LOAD_NAND:
+ case ISD::ATOMIC_LOAD_MIN:
+ case ISD::ATOMIC_LOAD_MAX:
+ case ISD::ATOMIC_LOAD_UMIN:
+ case ISD::ATOMIC_LOAD_UMAX:
+ case ISD::ATOMIC_CMP_SWAP: {
+ MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
+ RTLIB::Libcall LC = RTLIB::getATOMIC(Opc, VT);
+ assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpected atomic op or value type!");
+
+ std::pair<SDValue, SDValue> Tmp = ExpandChainLibCall(LC, Node, false);
+ Results.push_back(Tmp.first);
+ Results.push_back(Tmp.second);
+ break;
+ }
+ case ISD::TRAP: {
+ // If this operation is not supported, lower it to 'abort()' call
+ TargetLowering::ArgListTy Args;
+ TargetLowering::CallLoweringInfo CLI(DAG);
+ CLI.setDebugLoc(dl)
+ .setChain(Node->getOperand(0))
+ .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
+ DAG.getExternalSymbol("abort",
+ TLI.getPointerTy(DAG.getDataLayout())),
+ std::move(Args), 0);
+ std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
+
+ Results.push_back(CallResult.second);
+ break;
+ }
+ case ISD::FMINNUM:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::FMIN_F32, RTLIB::FMIN_F64,
+ RTLIB::FMIN_F80, RTLIB::FMIN_F128,
+ RTLIB::FMIN_PPCF128));
+ break;
+ case ISD::FMAXNUM:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::FMAX_F32, RTLIB::FMAX_F64,
+ RTLIB::FMAX_F80, RTLIB::FMAX_F128,
+ RTLIB::FMAX_PPCF128));
+ break;
+ case ISD::FSQRT:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
+ RTLIB::SQRT_F80, RTLIB::SQRT_F128,
+ RTLIB::SQRT_PPCF128));
+ break;
+ case ISD::FSIN:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
+ RTLIB::SIN_F80, RTLIB::SIN_F128,
+ RTLIB::SIN_PPCF128));
+ break;
+ case ISD::FCOS:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
+ RTLIB::COS_F80, RTLIB::COS_F128,
+ RTLIB::COS_PPCF128));
+ break;
+ case ISD::FSINCOS:
+ // Expand into sincos libcall.
+ ExpandSinCosLibCall(Node, Results);
+ break;
+ case ISD::FLOG:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
+ RTLIB::LOG_F80, RTLIB::LOG_F128,
+ RTLIB::LOG_PPCF128));
+ break;
+ case ISD::FLOG2:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
+ RTLIB::LOG2_F80, RTLIB::LOG2_F128,
+ RTLIB::LOG2_PPCF128));
+ break;
+ case ISD::FLOG10:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
+ RTLIB::LOG10_F80, RTLIB::LOG10_F128,
+ RTLIB::LOG10_PPCF128));
+ break;
+ case ISD::FEXP:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
+ RTLIB::EXP_F80, RTLIB::EXP_F128,
+ RTLIB::EXP_PPCF128));
+ break;
+ case ISD::FEXP2:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
+ RTLIB::EXP2_F80, RTLIB::EXP2_F128,
+ RTLIB::EXP2_PPCF128));
+ break;
+ case ISD::FTRUNC:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
+ RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
+ RTLIB::TRUNC_PPCF128));
+ break;
+ case ISD::FFLOOR:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
+ RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
+ RTLIB::FLOOR_PPCF128));
+ break;
+ case ISD::FCEIL:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
+ RTLIB::CEIL_F80, RTLIB::CEIL_F128,
+ RTLIB::CEIL_PPCF128));
+ break;
+ case ISD::FRINT:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
+ RTLIB::RINT_F80, RTLIB::RINT_F128,
+ RTLIB::RINT_PPCF128));
+ break;
+ case ISD::FNEARBYINT:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
+ RTLIB::NEARBYINT_F64,
+ RTLIB::NEARBYINT_F80,
+ RTLIB::NEARBYINT_F128,
+ RTLIB::NEARBYINT_PPCF128));
+ break;
+ case ISD::FROUND:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
+ RTLIB::ROUND_F64,
+ RTLIB::ROUND_F80,
+ RTLIB::ROUND_F128,
+ RTLIB::ROUND_PPCF128));
+ break;
+ case ISD::FPOWI:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
+ RTLIB::POWI_F80, RTLIB::POWI_F128,
+ RTLIB::POWI_PPCF128));
+ break;
+ case ISD::FPOW:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
+ RTLIB::POW_F80, RTLIB::POW_F128,
+ RTLIB::POW_PPCF128));
+ break;
+ case ISD::FDIV:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
+ RTLIB::DIV_F80, RTLIB::DIV_F128,
+ RTLIB::DIV_PPCF128));
+ break;
+ case ISD::FREM:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
+ RTLIB::REM_F80, RTLIB::REM_F128,
+ RTLIB::REM_PPCF128));
+ break;
+ case ISD::FMA:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
+ RTLIB::FMA_F80, RTLIB::FMA_F128,
+ RTLIB::FMA_PPCF128));
+ break;
+ case ISD::FADD:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::ADD_F32, RTLIB::ADD_F64,
+ RTLIB::ADD_F80, RTLIB::ADD_F128,
+ RTLIB::ADD_PPCF128));
+ break;
+ case ISD::FMUL:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::MUL_F32, RTLIB::MUL_F64,
+ RTLIB::MUL_F80, RTLIB::MUL_F128,
+ RTLIB::MUL_PPCF128));
+ break;
+ case ISD::FP16_TO_FP:
+ if (Node->getValueType(0) == MVT::f32) {
+ Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
+ }
+ break;
+ case ISD::FP_TO_FP16: {
+ RTLIB::Libcall LC =
+ RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
+ assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
+ Results.push_back(ExpandLibCall(LC, Node, false));
+ break;
+ }
+ case ISD::FSUB:
+ Results.push_back(ExpandFPLibCall(Node, RTLIB::SUB_F32, RTLIB::SUB_F64,
+ RTLIB::SUB_F80, RTLIB::SUB_F128,
+ RTLIB::SUB_PPCF128));
+ break;
+ case ISD::SREM:
+ Results.push_back(ExpandIntLibCall(Node, true,
+ RTLIB::SREM_I8,
+ RTLIB::SREM_I16, RTLIB::SREM_I32,
+ RTLIB::SREM_I64, RTLIB::SREM_I128));
+ break;
+ case ISD::UREM:
+ Results.push_back(ExpandIntLibCall(Node, false,
+ RTLIB::UREM_I8,
+ RTLIB::UREM_I16, RTLIB::UREM_I32,
+ RTLIB::UREM_I64, RTLIB::UREM_I128));
+ break;
+ case ISD::SDIV:
+ Results.push_back(ExpandIntLibCall(Node, true,
+ RTLIB::SDIV_I8,
+ RTLIB::SDIV_I16, RTLIB::SDIV_I32,
+ RTLIB::SDIV_I64, RTLIB::SDIV_I128));
+ break;
+ case ISD::UDIV:
+ Results.push_back(ExpandIntLibCall(Node, false,
+ RTLIB::UDIV_I8,
+ RTLIB::UDIV_I16, RTLIB::UDIV_I32,
+ RTLIB::UDIV_I64, RTLIB::UDIV_I128));
+ break;
+ case ISD::SDIVREM:
+ case ISD::UDIVREM:
+ // Expand into divrem libcall
+ ExpandDivRemLibCall(Node, Results);
+ break;
+ case ISD::MUL:
+ Results.push_back(ExpandIntLibCall(Node, false,
+ RTLIB::MUL_I8,
+ RTLIB::MUL_I16, RTLIB::MUL_I32,
+ RTLIB::MUL_I64, RTLIB::MUL_I128));
+ break;
+ }
+
// Replace the original node with the legalized result.
if (!Results.empty())
ReplaceNode(Node, Results.data());
}
+// Determine the vector type to use in place of an original scalar element when
+// promoting equally sized vectors.
+static MVT getPromotedVectorElementType(const TargetLowering &TLI,
+ MVT EltVT, MVT NewEltVT) {
+ unsigned OldEltsPerNewElt = EltVT.getSizeInBits() / NewEltVT.getSizeInBits();
+ MVT MidVT = MVT::getVectorVT(NewEltVT, OldEltsPerNewElt);
+ assert(TLI.isTypeLegal(MidVT) && "unexpected");
+ return MidVT;
+}
+
void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
SmallVector<SDValue, 8> Results;
MVT OVT = Node->getSimpleValueType(0);
if (Node->getOpcode() == ISD::UINT_TO_FP ||
Node->getOpcode() == ISD::SINT_TO_FP ||
- Node->getOpcode() == ISD::SETCC) {
+ Node->getOpcode() == ISD::SETCC ||
+ Node->getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
+ Node->getOpcode() == ISD::INSERT_VECTOR_ELT) {
OVT = Node->getOperand(0).getSimpleValueType();
}
if (Node->getOpcode() == ISD::BR_CC)
Tmp2, DAG.getIntPtrConstant(0, dl)));
break;
}
+ case ISD::BUILD_VECTOR: {
+ MVT EltVT = OVT.getVectorElementType();
+ MVT NewEltVT = NVT.getVectorElementType();
+
+ // Handle bitcasts to a different vector type with the same total bit size
+ //
+ // e.g. v2i64 = build_vector i64:x, i64:y => v4i32
+ // =>
+ // v4i32 = concat_vectors (v2i32 (bitcast i64:x)), (v2i32 (bitcast i64:y))
+
+ assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() &&
+ "Invalid promote type for build_vector");
+ assert(NewEltVT.bitsLT(EltVT) && "not handled");
+
+ MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
+
+ SmallVector<SDValue, 8> NewOps;
+ for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I) {
+ SDValue Op = Node->getOperand(I);
+ NewOps.push_back(DAG.getNode(ISD::BITCAST, SDLoc(Op), MidVT, Op));
+ }
+
+ SDLoc SL(Node);
+ SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SL, NVT, NewOps);
+ SDValue CvtVec = DAG.getNode(ISD::BITCAST, SL, OVT, Concat);
+ Results.push_back(CvtVec);
+ break;
+ }
+ case ISD::EXTRACT_VECTOR_ELT: {
+ MVT EltVT = OVT.getVectorElementType();
+ MVT NewEltVT = NVT.getVectorElementType();
+
+ // Handle bitcasts to a different vector type with the same total bit size.
+ //
+ // e.g. v2i64 = extract_vector_elt x:v2i64, y:i32
+ // =>
+ // v4i32:castx = bitcast x:v2i64
+ //
+ // i64 = bitcast
+ // (v2i32 build_vector (i32 (extract_vector_elt castx, (2 * y))),
+ // (i32 (extract_vector_elt castx, (2 * y + 1)))
+ //
+
+ assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() &&
+ "Invalid promote type for extract_vector_elt");
+ assert(NewEltVT.bitsLT(EltVT) && "not handled");
+
+ MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
+ unsigned NewEltsPerOldElt = MidVT.getVectorNumElements();
+
+ SDValue Idx = Node->getOperand(1);
+ EVT IdxVT = Idx.getValueType();
+ SDLoc SL(Node);
+ SDValue Factor = DAG.getConstant(NewEltsPerOldElt, SL, IdxVT);
+ SDValue NewBaseIdx = DAG.getNode(ISD::MUL, SL, IdxVT, Idx, Factor);
+
+ SDValue CastVec = DAG.getNode(ISD::BITCAST, SL, NVT, Node->getOperand(0));
+
+ SmallVector<SDValue, 8> NewOps;
+ for (unsigned I = 0; I < NewEltsPerOldElt; ++I) {
+ SDValue IdxOffset = DAG.getConstant(I, SL, IdxVT);
+ SDValue TmpIdx = DAG.getNode(ISD::ADD, SL, IdxVT, NewBaseIdx, IdxOffset);
+
+ SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, NewEltVT,
+ CastVec, TmpIdx);
+ NewOps.push_back(Elt);
+ }
+
+ SDValue NewVec = DAG.getNode(ISD::BUILD_VECTOR, SL, MidVT, NewOps);
+
+ Results.push_back(DAG.getNode(ISD::BITCAST, SL, EltVT, NewVec));
+ break;
+ }
+ case ISD::INSERT_VECTOR_ELT: {
+ MVT EltVT = OVT.getVectorElementType();
+ MVT NewEltVT = NVT.getVectorElementType();
+
+ // Handle bitcasts to a different vector type with the same total bit size
+ //
+ // e.g. v2i64 = insert_vector_elt x:v2i64, y:i64, z:i32
+ // =>
+ // v4i32:castx = bitcast x:v2i64
+ // v2i32:casty = bitcast y:i64
+ //
+ // v2i64 = bitcast
+ // (v4i32 insert_vector_elt
+ // (v4i32 insert_vector_elt v4i32:castx,
+ // (extract_vector_elt casty, 0), 2 * z),
+ // (extract_vector_elt casty, 1), (2 * z + 1))
+
+ assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() &&
+ "Invalid promote type for insert_vector_elt");
+ assert(NewEltVT.bitsLT(EltVT) && "not handled");
+
+ MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
+ unsigned NewEltsPerOldElt = MidVT.getVectorNumElements();
+
+ SDValue Val = Node->getOperand(1);
+ SDValue Idx = Node->getOperand(2);
+ EVT IdxVT = Idx.getValueType();
+ SDLoc SL(Node);
+
+ SDValue Factor = DAG.getConstant(NewEltsPerOldElt, SDLoc(), IdxVT);
+ SDValue NewBaseIdx = DAG.getNode(ISD::MUL, SL, IdxVT, Idx, Factor);
+
+ SDValue CastVec = DAG.getNode(ISD::BITCAST, SL, NVT, Node->getOperand(0));
+ SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, MidVT, Val);
+
+ SDValue NewVec = CastVec;
+ for (unsigned I = 0; I < NewEltsPerOldElt; ++I) {
+ SDValue IdxOffset = DAG.getConstant(I, SL, IdxVT);
+ SDValue InEltIdx = DAG.getNode(ISD::ADD, SL, IdxVT, NewBaseIdx, IdxOffset);
+
+ SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, NewEltVT,
+ CastVal, IdxOffset);
+
+ NewVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, SL, NVT,
+ NewVec, Elt, InEltIdx);
+ }
+
+ Results.push_back(DAG.getNode(ISD::BITCAST, SL, OVT, NewVec));
+ break;
+ }
+ case ISD::SCALAR_TO_VECTOR: {
+ MVT EltVT = OVT.getVectorElementType();
+ MVT NewEltVT = NVT.getVectorElementType();
+
+ // Handle bitcasts to different vector type with the smae total bit size.
+ //
+ // e.g. v2i64 = scalar_to_vector x:i64
+ // =>
+ // concat_vectors (v2i32 bitcast x:i64), (v2i32 undef)
+ //
+
+ MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT);
+ SDValue Val = Node->getOperand(0);
+ SDLoc SL(Node);
+
+ SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, MidVT, Val);
+ SDValue Undef = DAG.getUNDEF(MidVT);
+
+ SmallVector<SDValue, 8> NewElts;
+ NewElts.push_back(CastVal);
+ for (unsigned I = 1, NElts = OVT.getVectorNumElements(); I != NElts; ++I)
+ NewElts.push_back(Undef);
+
+ SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SL, NVT, NewElts);
+ SDValue CvtVec = DAG.getNode(ISD::BITCAST, SL, OVT, Concat);
+ Results.push_back(CvtVec);
+ break;
+ }
}
// Replace the original node with the legalized result.
projects / oota-llvm.git / blobdiff