+ return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
+}
+
+/// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
+/// element has been zero/sign-extended, depending on the isSigned parameter,
+/// from an integer type half its size.
+static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
+ bool isSigned) {
+ // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
+ EVT VT = N->getValueType(0);
+ if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
+ SDNode *BVN = N->getOperand(0).getNode();
+ if (BVN->getValueType(0) != MVT::v4i32 ||
+ BVN->getOpcode() != ISD::BUILD_VECTOR)
+ return false;
+ unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
+ unsigned HiElt = 1 - LoElt;
+ ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
+ ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
+ ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
+ ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
+ if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
+ return false;
+ if (isSigned) {
+ if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
+ Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
+ return true;
+ } else {
+ if (Hi0->isNullValue() && Hi1->isNullValue())
+ return true;
+ }
+ return false;
+ }
+
+ if (N->getOpcode() != ISD::BUILD_VECTOR)
+ return false;
+
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
+ SDNode *Elt = N->getOperand(i).getNode();
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
+ unsigned EltSize = VT.getVectorElementType().getSizeInBits();
+ unsigned HalfSize = EltSize / 2;
+ if (isSigned) {
+ int64_t SExtVal = C->getSExtValue();
+ if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
+ return false;
+ } else {
+ if ((C->getZExtValue() >> HalfSize) != 0)
+ return false;
+ }
+ continue;
+ }
+ return false;
+ }
+
+ return true;
+}
+
+/// isSignExtended - Check if a node is a vector value that is sign-extended
+/// or a constant BUILD_VECTOR with sign-extended elements.
+static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
+ if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
+ return true;
+ if (isExtendedBUILD_VECTOR(N, DAG, true))
+ return true;
+ return false;
+}
+
+/// isZeroExtended - Check if a node is a vector value that is zero-extended
+/// or a constant BUILD_VECTOR with zero-extended elements.
+static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
+ if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
+ return true;
+ if (isExtendedBUILD_VECTOR(N, DAG, false))
+ return true;
+ return false;
+}
+
+/// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
+/// load, or BUILD_VECTOR with extended elements, return the unextended value.
+static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
+ if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
+ return N->getOperand(0);
+ if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
+ return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
+ LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
+ LD->isNonTemporal(), LD->getAlignment());
+ // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
+ // have been legalized as a BITCAST from v4i32.
+ if (N->getOpcode() == ISD::BITCAST) {
+ SDNode *BVN = N->getOperand(0).getNode();
+ assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
+ BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
+ unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
+ return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
+ BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
+ }
+ // Construct a new BUILD_VECTOR with elements truncated to half the size.
+ assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
+ EVT VT = N->getValueType(0);
+ unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
+ unsigned NumElts = VT.getVectorNumElements();
+ MVT TruncVT = MVT::getIntegerVT(EltSize);
+ SmallVector<SDValue, 8> Ops;
+ for (unsigned i = 0; i != NumElts; ++i) {
+ ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
+ const APInt &CInt = C->getAPIntValue();
+ Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
+ }
+ return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
+ MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
+}
+
+static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
+ // Multiplications are only custom-lowered for 128-bit vectors so that
+ // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
+ EVT VT = Op.getValueType();
+ assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
+ SDNode *N0 = Op.getOperand(0).getNode();
+ SDNode *N1 = Op.getOperand(1).getNode();
+ unsigned NewOpc = 0;
+ if (isSignExtended(N0, DAG) && isSignExtended(N1, DAG))
+ NewOpc = ARMISD::VMULLs;
+ else if (isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG))
+ NewOpc = ARMISD::VMULLu;
+ else if (VT == MVT::v2i64)
+ // Fall through to expand this. It is not legal.
+ return SDValue();
+ else
+ // Other vector multiplications are legal.
+ return Op;
+
+ // Legalize to a VMULL instruction.
+ DebugLoc DL = Op.getDebugLoc();
+ SDValue Op0 = SkipExtension(N0, DAG);
+ SDValue Op1 = SkipExtension(N1, DAG);
+
+ assert(Op0.getValueType().is64BitVector() &&
+ Op1.getValueType().is64BitVector() &&
+ "unexpected types for extended operands to VMULL");
+ return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
+}
+
+static SDValue
+LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
+ // Convert to float
+ // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
+ // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
+ X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
+ Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
+ X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
+ Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
+ // Get reciprocal estimate.
+ // float4 recip = vrecpeq_f32(yf);
+ Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
+ // Because char has a smaller range than uchar, we can actually get away
+ // without any newton steps. This requires that we use a weird bias
+ // of 0xb000, however (again, this has been exhaustively tested).
+ // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
+ X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
+ X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
+ Y = DAG.getConstant(0xb000, MVT::i32);
+ Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
+ X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
+ X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
+ // Convert back to short.
+ X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
+ X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
+ return X;
+}
+
+static SDValue
+LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
+ SDValue N2;
+ // Convert to float.
+ // float4 yf = vcvt_f32_s32(vmovl_s16(y));
+ // float4 xf = vcvt_f32_s32(vmovl_s16(x));
+ N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
+ N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
+ N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
+ N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
+
+ // Use reciprocal estimate and one refinement step.
+ // float4 recip = vrecpeq_f32(yf);
+ // recip *= vrecpsq_f32(yf, recip);
+ N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
+ N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
+ N1, N2);
+ N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
+ // Because short has a smaller range than ushort, we can actually get away
+ // with only a single newton step. This requires that we use a weird bias
+ // of 89, however (again, this has been exhaustively tested).
+ // float4 result = as_float4(as_int4(xf*recip) + 89);
+ N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
+ N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
+ N1 = DAG.getConstant(89, MVT::i32);
+ N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
+ N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
+ N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
+ // Convert back to integer and return.
+ // return vmovn_s32(vcvt_s32_f32(result));
+ N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
+ N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
+ return N0;
+}
+
+static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
+ EVT VT = Op.getValueType();
+ assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
+ "unexpected type for custom-lowering ISD::SDIV");
+
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue N0 = Op.getOperand(0);
+ SDValue N1 = Op.getOperand(1);
+ SDValue N2, N3;
+
+ if (VT == MVT::v8i8) {
+ N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
+ N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
+
+ N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
+ DAG.getIntPtrConstant(4));
+ N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
+ DAG.getIntPtrConstant(4));
+ N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
+ DAG.getIntPtrConstant(0));
+ N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
+ DAG.getIntPtrConstant(0));
+
+ N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
+ N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
+
+ N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
+ N0 = LowerCONCAT_VECTORS(N0, DAG);
+
+ N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
+ return N0;
+ }
+ return LowerSDIV_v4i16(N0, N1, dl, DAG);
+}
+
+static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
+ EVT VT = Op.getValueType();
+ assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
+ "unexpected type for custom-lowering ISD::UDIV");
+
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue N0 = Op.getOperand(0);
+ SDValue N1 = Op.getOperand(1);
+ SDValue N2, N3;
+
+ if (VT == MVT::v8i8) {
+ N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
+ N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
+
+ N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
+ DAG.getIntPtrConstant(4));
+ N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
+ DAG.getIntPtrConstant(4));
+ N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
+ DAG.getIntPtrConstant(0));
+ N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
+ DAG.getIntPtrConstant(0));
+
+ N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
+ N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
+
+ N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
+ N0 = LowerCONCAT_VECTORS(N0, DAG);
+
+ N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
+ DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
+ N0);
+ return N0;
+ }
+
+ // v4i16 sdiv ... Convert to float.
+ // float4 yf = vcvt_f32_s32(vmovl_u16(y));
+ // float4 xf = vcvt_f32_s32(vmovl_u16(x));
+ N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
+ N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
+ N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
+ N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
+
+ // Use reciprocal estimate and two refinement steps.
+ // float4 recip = vrecpeq_f32(yf);
+ // recip *= vrecpsq_f32(yf, recip);
+ // recip *= vrecpsq_f32(yf, recip);
+ N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
+ N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
+ N1, N2);
+ N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
+ N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
+ DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
+ N1, N2);
+ N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
+ // Simply multiplying by the reciprocal estimate can leave us a few ulps
+ // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
+ // and that it will never cause us to return an answer too large).
+ // float4 result = as_float4(as_int4(xf*recip) + 89);
+ N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
+ N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
+ N1 = DAG.getConstant(2, MVT::i32);
+ N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
+ N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
+ N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
+ // Convert back to integer and return.
+ // return vmovn_u32(vcvt_s32_f32(result));
+ N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
+ N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
+ return N0;