+/// getVShiftImm - Check if this is a valid build_vector for the immediate
+/// operand of a vector shift operation, where all the elements of the
+/// build_vector must have the same constant integer value.
+static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
+ // Ignore bit_converts.
+ while (Op.getOpcode() == ISD::BIT_CONVERT)
+ Op = Op.getOperand(0);
+ BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
+ APInt SplatBits, SplatUndef;
+ unsigned SplatBitSize;
+ bool HasAnyUndefs;
+ if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
+ HasAnyUndefs, ElementBits) ||
+ SplatBitSize > ElementBits)
+ return false;
+ Cnt = SplatBits.getSExtValue();
+ return true;
+}
+
+/// isVShiftLImm - Check if this is a valid build_vector for the immediate
+/// operand of a vector shift left operation. That value must be in the range:
+/// 0 <= Value < ElementBits for a left shift; or
+/// 0 <= Value <= ElementBits for a long left shift.
+static bool isVShiftLImm(SDValue Op, MVT VT, bool isLong, int64_t &Cnt) {
+ assert(VT.isVector() && "vector shift count is not a vector type");
+ unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
+ if (! getVShiftImm(Op, ElementBits, Cnt))
+ return false;
+ return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
+}
+
+/// isVShiftRImm - Check if this is a valid build_vector for the immediate
+/// operand of a vector shift right operation. For a shift opcode, the value
+/// is positive, but for an intrinsic the value count must be negative. The
+/// absolute value must be in the range:
+/// 1 <= |Value| <= ElementBits for a right shift; or
+/// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
+static bool isVShiftRImm(SDValue Op, MVT VT, bool isNarrow, bool isIntrinsic,
+ int64_t &Cnt) {
+ assert(VT.isVector() && "vector shift count is not a vector type");
+ unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
+ if (! getVShiftImm(Op, ElementBits, Cnt))
+ return false;
+ if (isIntrinsic)
+ Cnt = -Cnt;
+ return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
+}
+
+/// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
+static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
+ unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
+ switch (IntNo) {
+ default:
+ // Don't do anything for most intrinsics.
+ break;
+
+ // Vector shifts: check for immediate versions and lower them.
+ // Note: This is done during DAG combining instead of DAG legalizing because
+ // the build_vectors for 64-bit vector element shift counts are generally
+ // not legal, and it is hard to see their values after they get legalized to
+ // loads from a constant pool.
+ case Intrinsic::arm_neon_vshifts:
+ case Intrinsic::arm_neon_vshiftu:
+ case Intrinsic::arm_neon_vshiftls:
+ case Intrinsic::arm_neon_vshiftlu:
+ case Intrinsic::arm_neon_vshiftn:
+ case Intrinsic::arm_neon_vrshifts:
+ case Intrinsic::arm_neon_vrshiftu:
+ case Intrinsic::arm_neon_vrshiftn:
+ case Intrinsic::arm_neon_vqshifts:
+ case Intrinsic::arm_neon_vqshiftu:
+ case Intrinsic::arm_neon_vqshiftsu:
+ case Intrinsic::arm_neon_vqshiftns:
+ case Intrinsic::arm_neon_vqshiftnu:
+ case Intrinsic::arm_neon_vqshiftnsu:
+ case Intrinsic::arm_neon_vqrshiftns:
+ case Intrinsic::arm_neon_vqrshiftnu:
+ case Intrinsic::arm_neon_vqrshiftnsu: {
+ MVT VT = N->getOperand(1).getValueType();
+ int64_t Cnt;
+ unsigned VShiftOpc = 0;
+
+ switch (IntNo) {
+ case Intrinsic::arm_neon_vshifts:
+ case Intrinsic::arm_neon_vshiftu:
+ if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
+ VShiftOpc = ARMISD::VSHL;
+ break;
+ }
+ if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
+ VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
+ ARMISD::VSHRs : ARMISD::VSHRu);
+ break;
+ }
+ return SDValue();
+
+ case Intrinsic::arm_neon_vshiftls:
+ case Intrinsic::arm_neon_vshiftlu:
+ if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
+ break;
+ assert(0 && "invalid shift count for vshll intrinsic");
+ abort();
+
+ case Intrinsic::arm_neon_vrshifts:
+ case Intrinsic::arm_neon_vrshiftu:
+ if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
+ break;
+ return SDValue();
+
+ case Intrinsic::arm_neon_vqshifts:
+ case Intrinsic::arm_neon_vqshiftu:
+ if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
+ break;
+ return SDValue();
+
+ case Intrinsic::arm_neon_vqshiftsu:
+ if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
+ break;
+ assert(0 && "invalid shift count for vqshlu intrinsic");
+ abort();
+
+ case Intrinsic::arm_neon_vshiftn:
+ case Intrinsic::arm_neon_vrshiftn:
+ case Intrinsic::arm_neon_vqshiftns:
+ case Intrinsic::arm_neon_vqshiftnu:
+ case Intrinsic::arm_neon_vqshiftnsu:
+ case Intrinsic::arm_neon_vqrshiftns:
+ case Intrinsic::arm_neon_vqrshiftnu:
+ case Intrinsic::arm_neon_vqrshiftnsu:
+ // Narrowing shifts require an immediate right shift.
+ if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
+ break;
+ assert(0 && "invalid shift count for narrowing vector shift intrinsic");
+ abort();
+
+ default:
+ assert(0 && "unhandled vector shift");
+ }
+
+ switch (IntNo) {
+ case Intrinsic::arm_neon_vshifts:
+ case Intrinsic::arm_neon_vshiftu:
+ // Opcode already set above.
+ break;
+ case Intrinsic::arm_neon_vshiftls:
+ case Intrinsic::arm_neon_vshiftlu:
+ if (Cnt == VT.getVectorElementType().getSizeInBits())
+ VShiftOpc = ARMISD::VSHLLi;
+ else
+ VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
+ ARMISD::VSHLLs : ARMISD::VSHLLu);
+ break;
+ case Intrinsic::arm_neon_vshiftn:
+ VShiftOpc = ARMISD::VSHRN; break;
+ case Intrinsic::arm_neon_vrshifts:
+ VShiftOpc = ARMISD::VRSHRs; break;
+ case Intrinsic::arm_neon_vrshiftu:
+ VShiftOpc = ARMISD::VRSHRu; break;
+ case Intrinsic::arm_neon_vrshiftn:
+ VShiftOpc = ARMISD::VRSHRN; break;
+ case Intrinsic::arm_neon_vqshifts:
+ VShiftOpc = ARMISD::VQSHLs; break;
+ case Intrinsic::arm_neon_vqshiftu:
+ VShiftOpc = ARMISD::VQSHLu; break;
+ case Intrinsic::arm_neon_vqshiftsu:
+ VShiftOpc = ARMISD::VQSHLsu; break;
+ case Intrinsic::arm_neon_vqshiftns:
+ VShiftOpc = ARMISD::VQSHRNs; break;
+ case Intrinsic::arm_neon_vqshiftnu:
+ VShiftOpc = ARMISD::VQSHRNu; break;
+ case Intrinsic::arm_neon_vqshiftnsu:
+ VShiftOpc = ARMISD::VQSHRNsu; break;
+ case Intrinsic::arm_neon_vqrshiftns:
+ VShiftOpc = ARMISD::VQRSHRNs; break;
+ case Intrinsic::arm_neon_vqrshiftnu:
+ VShiftOpc = ARMISD::VQRSHRNu; break;
+ case Intrinsic::arm_neon_vqrshiftnsu:
+ VShiftOpc = ARMISD::VQRSHRNsu; break;
+ }
+
+ return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
+ N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
+ }
+
+ case Intrinsic::arm_neon_vshiftins: {
+ MVT VT = N->getOperand(1).getValueType();
+ int64_t Cnt;
+ unsigned VShiftOpc = 0;
+
+ if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
+ VShiftOpc = ARMISD::VSLI;
+ else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
+ VShiftOpc = ARMISD::VSRI;
+ else {
+ assert(0 && "invalid shift count for vsli/vsri intrinsic");
+ abort();
+ }
+
+ return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
+ N->getOperand(1), N->getOperand(2),
+ DAG.getConstant(Cnt, MVT::i32));
+ }
+
+ case Intrinsic::arm_neon_vqrshifts:
+ case Intrinsic::arm_neon_vqrshiftu:
+ // No immediate versions of these to check for.
+ break;
+ }
+
+ return SDValue();
+}
+
+/// PerformShiftCombine - Checks for immediate versions of vector shifts and
+/// lowers them. As with the vector shift intrinsics, this is done during DAG
+/// combining instead of DAG legalizing because the build_vectors for 64-bit
+/// vector element shift counts are generally not legal, and it is hard to see
+/// their values after they get legalized to loads from a constant pool.
+static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
+ const ARMSubtarget *ST) {
+ MVT VT = N->getValueType(0);
+
+ // Nothing to be done for scalar shifts.
+ if (! VT.isVector())
+ return SDValue();
+
+ assert(ST->hasNEON() && "unexpected vector shift");
+ int64_t Cnt;
+
+ switch (N->getOpcode()) {
+ default: assert(0 && "unexpected shift opcode");
+
+ case ISD::SHL:
+ if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
+ return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
+ DAG.getConstant(Cnt, MVT::i32));
+ break;
+
+ case ISD::SRA:
+ case ISD::SRL:
+ if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
+ unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
+ ARMISD::VSHRs : ARMISD::VSHRu);
+ return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
+ DAG.getConstant(Cnt, MVT::i32));
+ }
+ }
+ return SDValue();
+}
+
+/// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
+/// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
+static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
+ const ARMSubtarget *ST) {
+ SDValue N0 = N->getOperand(0);
+
+ // Check for sign- and zero-extensions of vector extract operations of 8-
+ // and 16-bit vector elements. NEON supports these directly. They are
+ // handled during DAG combining because type legalization will promote them
+ // to 32-bit types and it is messy to recognize the operations after that.
+ if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
+ SDValue Vec = N0.getOperand(0);
+ SDValue Lane = N0.getOperand(1);
+ MVT VT = N->getValueType(0);
+ MVT EltVT = N0.getValueType();
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+
+ if (VT == MVT::i32 &&
+ (EltVT == MVT::i8 || EltVT == MVT::i16) &&
+ TLI.isTypeLegal(Vec.getValueType())) {
+
+ unsigned Opc = 0;
+ switch (N->getOpcode()) {
+ default: assert(0 && "unexpected opcode");
+ case ISD::SIGN_EXTEND:
+ Opc = ARMISD::VGETLANEs;
+ break;
+ case ISD::ZERO_EXTEND:
+ case ISD::ANY_EXTEND:
+ Opc = ARMISD::VGETLANEu;
+ break;
+ }
+ return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
+ }
+ }
+
+ return SDValue();
+}
+