addRegisterClass(MVT::i32, &SystemZ::GRX32BitRegClass);
else
addRegisterClass(MVT::i32, &SystemZ::GR32BitRegClass);
- addRegisterClass(MVT::i64, &SystemZ::GR64BitRegClass);
- addRegisterClass(MVT::f32, &SystemZ::FP32BitRegClass);
- addRegisterClass(MVT::f64, &SystemZ::FP64BitRegClass);
+ addRegisterClass(MVT::i64, &SystemZ::GR64BitRegClass);
+ if (Subtarget.hasVector()) {
+ addRegisterClass(MVT::f32, &SystemZ::VR32BitRegClass);
+ addRegisterClass(MVT::f64, &SystemZ::VR64BitRegClass);
+ } else {
+ addRegisterClass(MVT::f32, &SystemZ::FP32BitRegClass);
+ addRegisterClass(MVT::f64, &SystemZ::FP64BitRegClass);
+ }
addRegisterClass(MVT::f128, &SystemZ::FP128BitRegClass);
if (Subtarget.hasVector()) {
addRegisterClass(MVT::v8i16, &SystemZ::VR128BitRegClass);
addRegisterClass(MVT::v4i32, &SystemZ::VR128BitRegClass);
addRegisterClass(MVT::v2i64, &SystemZ::VR128BitRegClass);
+ addRegisterClass(MVT::v4f32, &SystemZ::VR128BitRegClass);
+ addRegisterClass(MVT::v2f64, &SystemZ::VR128BitRegClass);
}
// Compute derived properties from the register classes
if (isTypeLegal(VT)) {
// These operations are legal for anything that can be stored in a
// vector register, even if there is no native support for the format
- // as such.
+ // as such. In particular, we can do these for v4f32 even though there
+ // are no specific instructions for that format.
setOperationAction(ISD::LOAD, VT, Legal);
setOperationAction(ISD::STORE, VT, Legal);
setOperationAction(ISD::VSELECT, VT, Legal);
// Convert a GPR scalar to a vector by inserting it into element 0.
setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
+ // Use a series of unpacks for extensions.
+ setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Custom);
+ setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Custom);
+
// Detect shifts by a scalar amount and convert them into
// V*_BY_SCALAR.
setOperationAction(ISD::SHL, VT, Custom);
}
}
+ if (Subtarget.hasVector()) {
+ // There should be no need to check for float types other than v2f64
+ // since <2 x f32> isn't a legal type.
+ setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
+ setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
+ setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
+ setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
+ }
+
// Handle floating-point types.
for (unsigned I = MVT::FIRST_FP_VALUETYPE;
I <= MVT::LAST_FP_VALUETYPE;
}
}
+ // Handle floating-point vector types.
+ if (Subtarget.hasVector()) {
+ // Scalar-to-vector conversion is just a subreg.
+ setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Legal);
+ setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f64, Legal);
+
+ // Some insertions and extractions can be done directly but others
+ // need to go via integers.
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
+
+ // These operations have direct equivalents.
+ setOperationAction(ISD::FADD, MVT::v2f64, Legal);
+ setOperationAction(ISD::FNEG, MVT::v2f64, Legal);
+ setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
+ setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
+ setOperationAction(ISD::FMA, MVT::v2f64, Legal);
+ setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
+ setOperationAction(ISD::FABS, MVT::v2f64, Legal);
+ setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
+ setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
+ setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
+ setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
+ setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
+ setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
+ setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
+ }
+
// We have fused multiply-addition for f32 and f64 but not f128.
setOperationAction(ISD::FMA, MVT::f32, Legal);
setOperationAction(ISD::FMA, MVT::f64, Legal);
// We have 64-bit FPR<->GPR moves, but need special handling for
// 32-bit forms.
- setOperationAction(ISD::BITCAST, MVT::i32, Custom);
- setOperationAction(ISD::BITCAST, MVT::f32, Custom);
+ if (!Subtarget.hasVector()) {
+ setOperationAction(ISD::BITCAST, MVT::i32, Custom);
+ setOperationAction(ISD::BITCAST, MVT::f32, Custom);
+ }
// VASTART and VACOPY need to deal with the SystemZ-specific varargs
// structure, but VAEND is a no-op.
setTargetDAGCombine(ISD::SIGN_EXTEND);
setTargetDAGCombine(ISD::STORE);
setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
+ setTargetDAGCombine(ISD::FP_ROUND);
// Handle intrinsics.
setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
+ setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
// We want to use MVC in preference to even a single load/store pair.
MaxStoresPerMemcpy = 0;
return true;
}
+// We do not yet support 128-bit single-element vector types. If the user
+// attempts to use such types as function argument or return type, prefer
+// to error out instead of emitting code violating the ABI.
+static void VerifyVectorType(MVT VT, EVT ArgVT) {
+ if (ArgVT.isVector() && !VT.isVector())
+ report_fatal_error("Unsupported vector argument or return type");
+}
+
+static void VerifyVectorTypes(const SmallVectorImpl<ISD::InputArg> &Ins) {
+ for (unsigned i = 0; i < Ins.size(); ++i)
+ VerifyVectorType(Ins[i].VT, Ins[i].ArgVT);
+}
+
+static void VerifyVectorTypes(const SmallVectorImpl<ISD::OutputArg> &Outs) {
+ for (unsigned i = 0; i < Outs.size(); ++i)
+ VerifyVectorType(Outs[i].VT, Outs[i].ArgVT);
+}
+
// Value is a value that has been passed to us in the location described by VA
// (and so has type VA.getLocVT()). Convert Value to VA.getValVT(), chaining
// any loads onto Chain.
else if (VA.getLocInfo() == CCValAssign::Indirect)
Value = DAG.getLoad(VA.getValVT(), DL, Chain, Value,
MachinePointerInfo(), false, false, false, 0);
- else
+ else if (VA.getLocInfo() == CCValAssign::BCvt) {
+ // If this is a short vector argument loaded from the stack,
+ // extend from i64 to full vector size and then bitcast.
+ assert(VA.getLocVT() == MVT::i64);
+ assert(VA.getValVT().isVector());
+ Value = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i64,
+ Value, DAG.getUNDEF(MVT::i64));
+ Value = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Value);
+ } else
assert(VA.getLocInfo() == CCValAssign::Full && "Unsupported getLocInfo");
return Value;
}
return DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Value);
case CCValAssign::AExt:
return DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Value);
+ case CCValAssign::BCvt:
+ // If this is a short vector argument to be stored to the stack,
+ // bitcast to v2i64 and then extract first element.
+ assert(VA.getLocVT() == MVT::i64);
+ assert(VA.getValVT().isVector());
+ Value = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Value);
+ return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VA.getLocVT(), Value,
+ DAG.getConstant(0, DL, MVT::i32));
case CCValAssign::Full:
return Value;
default:
auto *TFL =
static_cast<const SystemZFrameLowering *>(Subtarget.getFrameLowering());
+ // Detect unsupported vector argument types.
+ if (Subtarget.hasVector())
+ VerifyVectorTypes(Ins);
+
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
SystemZCCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
case MVT::v8i16:
case MVT::v4i32:
case MVT::v2i64:
+ case MVT::v4f32:
+ case MVT::v2f64:
RC = &SystemZ::VR128BitRegClass;
break;
}
MachineFunction &MF = DAG.getMachineFunction();
EVT PtrVT = getPointerTy();
+ // Detect unsupported vector argument and return types.
+ if (Subtarget.hasVector()) {
+ VerifyVectorTypes(Outs);
+ VerifyVectorTypes(Ins);
+ }
+
// Analyze the operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
SystemZCCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
SDLoc DL, SelectionDAG &DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
+ // Detect unsupported vector return types.
+ if (Subtarget.hasVector())
+ VerifyVectorTypes(Outs);
+
// Assign locations to each returned value.
SmallVector<CCValAssign, 16> RetLocs;
CCState RetCCInfo(CallConv, IsVarArg, MF, RetLocs, *DAG.getContext());
}
}
+// Return true if Op is an intrinsic node without chain that returns the
+// CC value as its final argument. Provide the associated SystemZISD
+// opcode and the mask of valid CC values if so.
+static bool isIntrinsicWithCC(SDValue Op, unsigned &Opcode, unsigned &CCValid) {
+ unsigned Id = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
+ switch (Id) {
+ case Intrinsic::s390_vpkshs:
+ case Intrinsic::s390_vpksfs:
+ case Intrinsic::s390_vpksgs:
+ Opcode = SystemZISD::PACKS_CC;
+ CCValid = SystemZ::CCMASK_VCMP;
+ return true;
+
+ case Intrinsic::s390_vpklshs:
+ case Intrinsic::s390_vpklsfs:
+ case Intrinsic::s390_vpklsgs:
+ Opcode = SystemZISD::PACKLS_CC;
+ CCValid = SystemZ::CCMASK_VCMP;
+ return true;
+
+ case Intrinsic::s390_vceqbs:
+ case Intrinsic::s390_vceqhs:
+ case Intrinsic::s390_vceqfs:
+ case Intrinsic::s390_vceqgs:
+ Opcode = SystemZISD::VICMPES;
+ CCValid = SystemZ::CCMASK_VCMP;
+ return true;
+
+ case Intrinsic::s390_vchbs:
+ case Intrinsic::s390_vchhs:
+ case Intrinsic::s390_vchfs:
+ case Intrinsic::s390_vchgs:
+ Opcode = SystemZISD::VICMPHS;
+ CCValid = SystemZ::CCMASK_VCMP;
+ return true;
+
+ case Intrinsic::s390_vchlbs:
+ case Intrinsic::s390_vchlhs:
+ case Intrinsic::s390_vchlfs:
+ case Intrinsic::s390_vchlgs:
+ Opcode = SystemZISD::VICMPHLS;
+ CCValid = SystemZ::CCMASK_VCMP;
+ return true;
+
+ case Intrinsic::s390_vtm:
+ Opcode = SystemZISD::VTM;
+ CCValid = SystemZ::CCMASK_VCMP;
+ return true;
+
+ case Intrinsic::s390_vfaebs:
+ case Intrinsic::s390_vfaehs:
+ case Intrinsic::s390_vfaefs:
+ Opcode = SystemZISD::VFAE_CC;
+ CCValid = SystemZ::CCMASK_ANY;
+ return true;
+
+ case Intrinsic::s390_vfaezbs:
+ case Intrinsic::s390_vfaezhs:
+ case Intrinsic::s390_vfaezfs:
+ Opcode = SystemZISD::VFAEZ_CC;
+ CCValid = SystemZ::CCMASK_ANY;
+ return true;
+
+ case Intrinsic::s390_vfeebs:
+ case Intrinsic::s390_vfeehs:
+ case Intrinsic::s390_vfeefs:
+ Opcode = SystemZISD::VFEE_CC;
+ CCValid = SystemZ::CCMASK_ANY;
+ return true;
+
+ case Intrinsic::s390_vfeezbs:
+ case Intrinsic::s390_vfeezhs:
+ case Intrinsic::s390_vfeezfs:
+ Opcode = SystemZISD::VFEEZ_CC;
+ CCValid = SystemZ::CCMASK_ANY;
+ return true;
+
+ case Intrinsic::s390_vfenebs:
+ case Intrinsic::s390_vfenehs:
+ case Intrinsic::s390_vfenefs:
+ Opcode = SystemZISD::VFENE_CC;
+ CCValid = SystemZ::CCMASK_ANY;
+ return true;
+
+ case Intrinsic::s390_vfenezbs:
+ case Intrinsic::s390_vfenezhs:
+ case Intrinsic::s390_vfenezfs:
+ Opcode = SystemZISD::VFENEZ_CC;
+ CCValid = SystemZ::CCMASK_ANY;
+ return true;
+
+ case Intrinsic::s390_vistrbs:
+ case Intrinsic::s390_vistrhs:
+ case Intrinsic::s390_vistrfs:
+ Opcode = SystemZISD::VISTR_CC;
+ CCValid = SystemZ::CCMASK_0 | SystemZ::CCMASK_3;
+ return true;
+
+ case Intrinsic::s390_vstrcbs:
+ case Intrinsic::s390_vstrchs:
+ case Intrinsic::s390_vstrcfs:
+ Opcode = SystemZISD::VSTRC_CC;
+ CCValid = SystemZ::CCMASK_ANY;
+ return true;
+
+ case Intrinsic::s390_vstrczbs:
+ case Intrinsic::s390_vstrczhs:
+ case Intrinsic::s390_vstrczfs:
+ Opcode = SystemZISD::VSTRCZ_CC;
+ CCValid = SystemZ::CCMASK_ANY;
+ return true;
+
+ case Intrinsic::s390_vfcedbs:
+ Opcode = SystemZISD::VFCMPES;
+ CCValid = SystemZ::CCMASK_VCMP;
+ return true;
+
+ case Intrinsic::s390_vfchdbs:
+ Opcode = SystemZISD::VFCMPHS;
+ CCValid = SystemZ::CCMASK_VCMP;
+ return true;
+
+ case Intrinsic::s390_vfchedbs:
+ Opcode = SystemZISD::VFCMPHES;
+ CCValid = SystemZ::CCMASK_VCMP;
+ return true;
+
+ case Intrinsic::s390_vftcidb:
+ Opcode = SystemZISD::VFTCI;
+ CCValid = SystemZ::CCMASK_VCMP;
+ return true;
+
+ default:
+ return false;
+ }
+}
+
// Emit an intrinsic with chain with a glued value instead of its CC result.
static SDValue emitIntrinsicWithChainAndGlue(SelectionDAG &DAG, SDValue Op,
unsigned Opcode) {
return Intr;
}
+// Emit an intrinsic with a glued value instead of its CC result.
+static SDValue emitIntrinsicWithGlue(SelectionDAG &DAG, SDValue Op,
+ unsigned Opcode) {
+ // Copy all operands except the intrinsic ID.
+ unsigned NumOps = Op.getNumOperands();
+ SmallVector<SDValue, 6> Ops;
+ Ops.reserve(NumOps - 1);
+ for (unsigned I = 1; I < NumOps; ++I)
+ Ops.push_back(Op.getOperand(I));
+
+ if (Op->getNumValues() == 1)
+ return DAG.getNode(Opcode, SDLoc(Op), MVT::Glue, Ops);
+ assert(Op->getNumValues() == 2 && "Expected exactly one non-CC result");
+ SDVTList RawVTs = DAG.getVTList(Op->getValueType(0), MVT::Glue);
+ return DAG.getNode(Opcode, SDLoc(Op), RawVTs, Ops);
+}
+
// CC is a comparison that will be implemented using an integer or
// floating-point comparison. Return the condition code mask for
// a branch on true. In the integer case, CCMASK_CMP_UO is set for
CmpOp0.getResNo() == 0 && CmpOp0->hasNUsesOfValue(1, 0) &&
isIntrinsicWithCCAndChain(CmpOp0, Opcode, CCValid))
return getIntrinsicCmp(DAG, Opcode, CmpOp0, CCValid, Constant, Cond);
+ if (CmpOp0.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
+ CmpOp0.getResNo() == CmpOp0->getNumValues() - 1 &&
+ isIntrinsicWithCC(CmpOp0, Opcode, CCValid))
+ return getIntrinsicCmp(DAG, Opcode, CmpOp0, CCValid, Constant, Cond);
}
Comparison C(CmpOp0, CmpOp1);
C.CCMask = CCMaskForCondCode(Cond);
case ISD::INTRINSIC_W_CHAIN:
Op = emitIntrinsicWithChainAndGlue(DAG, C.Op0, C.Opcode);
break;
+ case ISD::INTRINSIC_WO_CHAIN:
+ Op = emitIntrinsicWithGlue(DAG, C.Op0, C.Opcode);
+ break;
default:
llvm_unreachable("Invalid comparison operands");
}
return Result;
}
-// Return the SystemZISD vector comparison operation for CC, or 0 if it cannot
-// be done directly.
-static unsigned getVectorComparison(ISD::CondCode CC) {
+// Return the SystemISD vector comparison operation for CC, or 0 if it cannot
+// be done directly. IsFP is true if CC is for a floating-point rather than
+// integer comparison.
+static unsigned getVectorComparison(ISD::CondCode CC, bool IsFP) {
switch (CC) {
+ case ISD::SETOEQ:
case ISD::SETEQ:
- return SystemZISD::VICMPE;
+ return IsFP ? SystemZISD::VFCMPE : SystemZISD::VICMPE;
+
+ case ISD::SETOGE:
+ case ISD::SETGE:
+ return IsFP ? SystemZISD::VFCMPHE : 0;
+ case ISD::SETOGT:
case ISD::SETGT:
- return SystemZISD::VICMPH;
+ return IsFP ? SystemZISD::VFCMPH : SystemZISD::VICMPH;
case ISD::SETUGT:
- return SystemZISD::VICMPHL;
+ return IsFP ? 0 : SystemZISD::VICMPHL;
default:
return 0;
// Return the SystemZISD vector comparison operation for CC or its inverse,
// or 0 if neither can be done directly. Indicate in Invert whether the
-// result is for the inverse of CC.
-static unsigned getVectorComparisonOrInvert(ISD::CondCode CC, bool &Invert) {
- if (unsigned Opcode = getVectorComparison(CC)) {
+// result is for the inverse of CC. IsFP is true if CC is for a
+// floating-point rather than integer comparison.
+static unsigned getVectorComparisonOrInvert(ISD::CondCode CC, bool IsFP,
+ bool &Invert) {
+ if (unsigned Opcode = getVectorComparison(CC, IsFP)) {
Invert = false;
return Opcode;
}
- CC = ISD::getSetCCInverse(CC, true);
- if (unsigned Opcode = getVectorComparison(CC)) {
+ CC = ISD::getSetCCInverse(CC, !IsFP);
+ if (unsigned Opcode = getVectorComparison(CC, IsFP)) {
Invert = true;
return Opcode;
}
return 0;
}
+// Return a v2f64 that contains the extended form of elements Start and Start+1
+// of v4f32 value Op.
+static SDValue expandV4F32ToV2F64(SelectionDAG &DAG, int Start, SDLoc DL,
+ SDValue Op) {
+ int Mask[] = { Start, -1, Start + 1, -1 };
+ Op = DAG.getVectorShuffle(MVT::v4f32, DL, Op, DAG.getUNDEF(MVT::v4f32), Mask);
+ return DAG.getNode(SystemZISD::VEXTEND, DL, MVT::v2f64, Op);
+}
+
+// Build a comparison of vectors CmpOp0 and CmpOp1 using opcode Opcode,
+// producing a result of type VT.
+static SDValue getVectorCmp(SelectionDAG &DAG, unsigned Opcode, SDLoc DL,
+ EVT VT, SDValue CmpOp0, SDValue CmpOp1) {
+ // There is no hardware support for v4f32, so extend the vector into
+ // two v2f64s and compare those.
+ if (CmpOp0.getValueType() == MVT::v4f32) {
+ SDValue H0 = expandV4F32ToV2F64(DAG, 0, DL, CmpOp0);
+ SDValue L0 = expandV4F32ToV2F64(DAG, 2, DL, CmpOp0);
+ SDValue H1 = expandV4F32ToV2F64(DAG, 0, DL, CmpOp1);
+ SDValue L1 = expandV4F32ToV2F64(DAG, 2, DL, CmpOp1);
+ SDValue HRes = DAG.getNode(Opcode, DL, MVT::v2i64, H0, H1);
+ SDValue LRes = DAG.getNode(Opcode, DL, MVT::v2i64, L0, L1);
+ return DAG.getNode(SystemZISD::PACK, DL, VT, HRes, LRes);
+ }
+ return DAG.getNode(Opcode, DL, VT, CmpOp0, CmpOp1);
+}
+
// Lower a vector comparison of type CC between CmpOp0 and CmpOp1, producing
// an integer mask of type VT.
static SDValue lowerVectorSETCC(SelectionDAG &DAG, SDLoc DL, EVT VT,
ISD::CondCode CC, SDValue CmpOp0,
SDValue CmpOp1) {
+ bool IsFP = CmpOp0.getValueType().isFloatingPoint();
bool Invert = false;
SDValue Cmp;
- // It doesn't really matter whether we try the inversion or the swap first,
- // since there are no cases where both work.
- if (unsigned Opcode = getVectorComparisonOrInvert(CC, Invert))
- Cmp = DAG.getNode(Opcode, DL, VT, CmpOp0, CmpOp1);
- else {
- CC = ISD::getSetCCSwappedOperands(CC);
- if (unsigned Opcode = getVectorComparisonOrInvert(CC, Invert))
- Cmp = DAG.getNode(Opcode, DL, VT, CmpOp1, CmpOp0);
- else
- llvm_unreachable("Unhandled comparison");
+ switch (CC) {
+ // Handle tests for order using (or (ogt y x) (oge x y)).
+ case ISD::SETUO:
+ Invert = true;
+ case ISD::SETO: {
+ assert(IsFP && "Unexpected integer comparison");
+ SDValue LT = getVectorCmp(DAG, SystemZISD::VFCMPH, DL, VT, CmpOp1, CmpOp0);
+ SDValue GE = getVectorCmp(DAG, SystemZISD::VFCMPHE, DL, VT, CmpOp0, CmpOp1);
+ Cmp = DAG.getNode(ISD::OR, DL, VT, LT, GE);
+ break;
+ }
+
+ // Handle <> tests using (or (ogt y x) (ogt x y)).
+ case ISD::SETUEQ:
+ Invert = true;
+ case ISD::SETONE: {
+ assert(IsFP && "Unexpected integer comparison");
+ SDValue LT = getVectorCmp(DAG, SystemZISD::VFCMPH, DL, VT, CmpOp1, CmpOp0);
+ SDValue GT = getVectorCmp(DAG, SystemZISD::VFCMPH, DL, VT, CmpOp0, CmpOp1);
+ Cmp = DAG.getNode(ISD::OR, DL, VT, LT, GT);
+ break;
+ }
+
+ // Otherwise a single comparison is enough. It doesn't really
+ // matter whether we try the inversion or the swap first, since
+ // there are no cases where both work.
+ default:
+ if (unsigned Opcode = getVectorComparisonOrInvert(CC, IsFP, Invert))
+ Cmp = getVectorCmp(DAG, Opcode, DL, VT, CmpOp0, CmpOp1);
+ else {
+ CC = ISD::getSetCCSwappedOperands(CC);
+ if (unsigned Opcode = getVectorComparisonOrInvert(CC, IsFP, Invert))
+ Cmp = getVectorCmp(DAG, Opcode, DL, VT, CmpOp1, CmpOp0);
+ else
+ llvm_unreachable("Unhandled comparison");
+ }
+ break;
}
if (Invert) {
SDValue Mask = DAG.getNode(SystemZISD::BYTE_MASK, DL, MVT::v16i8,
return SDValue();
}
+SDValue
+SystemZTargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
+ SelectionDAG &DAG) const {
+ unsigned Opcode, CCValid;
+ if (isIntrinsicWithCC(Op, Opcode, CCValid)) {
+ SDValue Glued = emitIntrinsicWithGlue(DAG, Op, Opcode);
+ SDValue CC = getCCResult(DAG, Glued.getNode());
+ if (Op->getNumValues() == 1)
+ return CC;
+ assert(Op->getNumValues() == 2 && "Expected a CC and non-CC result");
+ return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), Op->getVTList(),
+ Glued, CC);
+ }
+
+ unsigned Id = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
+ switch (Id) {
+ case Intrinsic::s390_vpdi:
+ return DAG.getNode(SystemZISD::PERMUTE_DWORDS, SDLoc(Op), Op.getValueType(),
+ Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
+
+ case Intrinsic::s390_vperm:
+ return DAG.getNode(SystemZISD::PERMUTE, SDLoc(Op), Op.getValueType(),
+ Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
+
+ case Intrinsic::s390_vuphb:
+ case Intrinsic::s390_vuphh:
+ case Intrinsic::s390_vuphf:
+ return DAG.getNode(SystemZISD::UNPACK_HIGH, SDLoc(Op), Op.getValueType(),
+ Op.getOperand(1));
+
+ case Intrinsic::s390_vuplhb:
+ case Intrinsic::s390_vuplhh:
+ case Intrinsic::s390_vuplhf:
+ return DAG.getNode(SystemZISD::UNPACKL_HIGH, SDLoc(Op), Op.getValueType(),
+ Op.getOperand(1));
+
+ case Intrinsic::s390_vuplb:
+ case Intrinsic::s390_vuplhw:
+ case Intrinsic::s390_vuplf:
+ return DAG.getNode(SystemZISD::UNPACK_LOW, SDLoc(Op), Op.getValueType(),
+ Op.getOperand(1));
+
+ case Intrinsic::s390_vupllb:
+ case Intrinsic::s390_vupllh:
+ case Intrinsic::s390_vupllf:
+ return DAG.getNode(SystemZISD::UNPACKL_LOW, SDLoc(Op), Op.getValueType(),
+ Op.getOperand(1));
+
+ case Intrinsic::s390_vsumb:
+ case Intrinsic::s390_vsumh:
+ case Intrinsic::s390_vsumgh:
+ case Intrinsic::s390_vsumgf:
+ case Intrinsic::s390_vsumqf:
+ case Intrinsic::s390_vsumqg:
+ return DAG.getNode(SystemZISD::VSUM, SDLoc(Op), Op.getValueType(),
+ Op.getOperand(1), Op.getOperand(2));
+ }
+
+ return SDValue();
+}
+
namespace {
// Says that SystemZISD operation Opcode can be used to perform the equivalent
// of a VPERM with permute vector Bytes. If Opcode takes three operands,
return DAG.getNode(ISD::BITCAST, DL, VT, Op);
}
+// Return true if the given BUILD_VECTOR is a scalar-to-vector conversion.
+static bool isScalarToVector(SDValue Op) {
+ for (unsigned I = 1, E = Op.getNumOperands(); I != E; ++I)
+ if (Op.getOperand(I).getOpcode() != ISD::UNDEF)
+ return false;
+ return true;
+}
+
+// Return a vector of type VT that contains Value in the first element.
+// The other elements don't matter.
+static SDValue buildScalarToVector(SelectionDAG &DAG, SDLoc DL, EVT VT,
+ SDValue Value) {
+ // If we have a constant, replicate it to all elements and let the
+ // BUILD_VECTOR lowering take care of it.
+ if (Value.getOpcode() == ISD::Constant ||
+ Value.getOpcode() == ISD::ConstantFP) {
+ SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Value);
+ return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Ops);
+ }
+ if (Value.getOpcode() == ISD::UNDEF)
+ return DAG.getUNDEF(VT);
+ return DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Value);
+}
+
+// Return a vector of type VT in which Op0 is in element 0 and Op1 is in
+// element 1. Used for cases in which replication is cheap.
+static SDValue buildMergeScalars(SelectionDAG &DAG, SDLoc DL, EVT VT,
+ SDValue Op0, SDValue Op1) {
+ if (Op0.getOpcode() == ISD::UNDEF) {
+ if (Op1.getOpcode() == ISD::UNDEF)
+ return DAG.getUNDEF(VT);
+ return DAG.getNode(SystemZISD::REPLICATE, DL, VT, Op1);
+ }
+ if (Op1.getOpcode() == ISD::UNDEF)
+ return DAG.getNode(SystemZISD::REPLICATE, DL, VT, Op0);
+ return DAG.getNode(SystemZISD::MERGE_HIGH, DL, VT,
+ buildScalarToVector(DAG, DL, VT, Op0),
+ buildScalarToVector(DAG, DL, VT, Op1));
+}
+
// Extend GPR scalars Op0 and Op1 to doublewords and return a v2i64
// vector for them.
static SDValue joinDwords(SelectionDAG &DAG, SDLoc DL, SDValue Op0,
if (VT == MVT::v2i64)
return joinDwords(DAG, DL, Elems[0], Elems[1]);
+ // Use a 64-bit merge high to combine two doubles.
+ if (VT == MVT::v2f64)
+ return buildMergeScalars(DAG, DL, VT, Elems[0], Elems[1]);
+
+ // Build v4f32 values directly from the FPRs:
+ //
+ // <Axxx> <Bxxx> <Cxxxx> <Dxxx>
+ // V V VMRHF
+ // <ABxx> <CDxx>
+ // V VMRHG
+ // <ABCD>
+ if (VT == MVT::v4f32) {
+ SDValue Op01 = buildMergeScalars(DAG, DL, VT, Elems[0], Elems[1]);
+ SDValue Op23 = buildMergeScalars(DAG, DL, VT, Elems[2], Elems[3]);
+ // Avoid unnecessary undefs by reusing the other operand.
+ if (Op01.getOpcode() == ISD::UNDEF)
+ Op01 = Op23;
+ else if (Op23.getOpcode() == ISD::UNDEF)
+ Op23 = Op01;
+ // Merging identical replications is a no-op.
+ if (Op01.getOpcode() == SystemZISD::REPLICATE && Op01 == Op23)
+ return Op01;
+ Op01 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Op01);
+ Op23 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Op23);
+ SDValue Op = DAG.getNode(SystemZISD::MERGE_HIGH,
+ DL, MVT::v2i64, Op01, Op23);
+ return DAG.getNode(ISD::BITCAST, DL, VT, Op);
+ }
+
// Collect the constant terms.
SmallVector<SDValue, SystemZ::VectorBytes> Constants(NumElements, SDValue());
SmallVector<bool, SystemZ::VectorBytes> Done(NumElements, false);
if (Res.getNode())
return Res;
+ // Detect SCALAR_TO_VECTOR conversions.
+ if (isOperationLegal(ISD::SCALAR_TO_VECTOR, VT) && isScalarToVector(Op))
+ return buildScalarToVector(DAG, DL, VT, Op.getOperand(0));
+
// Otherwise use buildVector to build the vector up from GPRs.
unsigned NumElements = Op.getNumOperands();
SmallVector<SDValue, SystemZ::VectorBytes> Ops(NumElements);
Op.getOperand(0), DAG.getConstant(0, DL, MVT::i32));
}
+SDValue SystemZTargetLowering::lowerINSERT_VECTOR_ELT(SDValue Op,
+ SelectionDAG &DAG) const {
+ // Handle insertions of floating-point values.
+ SDLoc DL(Op);
+ SDValue Op0 = Op.getOperand(0);
+ SDValue Op1 = Op.getOperand(1);
+ SDValue Op2 = Op.getOperand(2);
+ EVT VT = Op.getValueType();
+
+ // Insertions into constant indices of a v2f64 can be done using VPDI.
+ // However, if the inserted value is a bitcast or a constant then it's
+ // better to use GPRs, as below.
+ if (VT == MVT::v2f64 &&
+ Op1.getOpcode() != ISD::BITCAST &&
+ Op1.getOpcode() != ISD::ConstantFP &&
+ Op2.getOpcode() == ISD::Constant) {
+ uint64_t Index = dyn_cast<ConstantSDNode>(Op2)->getZExtValue();
+ unsigned Mask = VT.getVectorNumElements() - 1;
+ if (Index <= Mask)
+ return Op;
+ }
+
+ // Otherwise bitcast to the equivalent integer form and insert via a GPR.
+ MVT IntVT = MVT::getIntegerVT(VT.getVectorElementType().getSizeInBits());
+ MVT IntVecVT = MVT::getVectorVT(IntVT, VT.getVectorNumElements());
+ SDValue Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, IntVecVT,
+ DAG.getNode(ISD::BITCAST, DL, IntVecVT, Op0),
+ DAG.getNode(ISD::BITCAST, DL, IntVT, Op1), Op2);
+ return DAG.getNode(ISD::BITCAST, DL, VT, Res);
+}
+
+SDValue
+SystemZTargetLowering::lowerEXTRACT_VECTOR_ELT(SDValue Op,
+ SelectionDAG &DAG) const {
+ // Handle extractions of floating-point values.
+ SDLoc DL(Op);
+ SDValue Op0 = Op.getOperand(0);
+ SDValue Op1 = Op.getOperand(1);
+ EVT VT = Op.getValueType();
+ EVT VecVT = Op0.getValueType();
+
+ // Extractions of constant indices can be done directly.
+ if (auto *CIndexN = dyn_cast<ConstantSDNode>(Op1)) {
+ uint64_t Index = CIndexN->getZExtValue();
+ unsigned Mask = VecVT.getVectorNumElements() - 1;
+ if (Index <= Mask)
+ return Op;
+ }
+
+ // Otherwise bitcast to the equivalent integer form and extract via a GPR.
+ MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits());
+ MVT IntVecVT = MVT::getVectorVT(IntVT, VecVT.getVectorNumElements());
+ SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, IntVT,
+ DAG.getNode(ISD::BITCAST, DL, IntVecVT, Op0), Op1);
+ return DAG.getNode(ISD::BITCAST, DL, VT, Res);
+}
+
+SDValue
+SystemZTargetLowering::lowerExtendVectorInreg(SDValue Op, SelectionDAG &DAG,
+ unsigned UnpackHigh) const {
+ SDValue PackedOp = Op.getOperand(0);
+ EVT OutVT = Op.getValueType();
+ EVT InVT = PackedOp.getValueType();
+ unsigned ToBits = OutVT.getVectorElementType().getSizeInBits();
+ unsigned FromBits = InVT.getVectorElementType().getSizeInBits();
+ do {
+ FromBits *= 2;
+ EVT OutVT = MVT::getVectorVT(MVT::getIntegerVT(FromBits),
+ SystemZ::VectorBits / FromBits);
+ PackedOp = DAG.getNode(UnpackHigh, SDLoc(PackedOp), OutVT, PackedOp);
+ } while (FromBits != ToBits);
+ return PackedOp;
+}
+
SDValue SystemZTargetLowering::lowerShift(SDValue Op, SelectionDAG &DAG,
unsigned ByScalar) const {
// Look for cases where a vector shift can use the *_BY_SCALAR form.
return lowerPREFETCH(Op, DAG);
case ISD::INTRINSIC_W_CHAIN:
return lowerINTRINSIC_W_CHAIN(Op, DAG);
+ case ISD::INTRINSIC_WO_CHAIN:
+ return lowerINTRINSIC_WO_CHAIN(Op, DAG);
case ISD::BUILD_VECTOR:
return lowerBUILD_VECTOR(Op, DAG);
case ISD::VECTOR_SHUFFLE:
return lowerVECTOR_SHUFFLE(Op, DAG);
case ISD::SCALAR_TO_VECTOR:
return lowerSCALAR_TO_VECTOR(Op, DAG);
+ case ISD::INSERT_VECTOR_ELT:
+ return lowerINSERT_VECTOR_ELT(Op, DAG);
+ case ISD::EXTRACT_VECTOR_ELT:
+ return lowerEXTRACT_VECTOR_ELT(Op, DAG);
+ case ISD::SIGN_EXTEND_VECTOR_INREG:
+ return lowerExtendVectorInreg(Op, DAG, SystemZISD::UNPACK_HIGH);
+ case ISD::ZERO_EXTEND_VECTOR_INREG:
+ return lowerExtendVectorInreg(Op, DAG, SystemZISD::UNPACKL_HIGH);
case ISD::SHL:
return lowerShift(Op, DAG, SystemZISD::VSHL_BY_SCALAR);
case ISD::SRL:
const char *SystemZTargetLowering::getTargetNodeName(unsigned Opcode) const {
#define OPCODE(NAME) case SystemZISD::NAME: return "SystemZISD::" #NAME
- switch (Opcode) {
+ switch ((SystemZISD::NodeType)Opcode) {
+ case SystemZISD::FIRST_NUMBER: break;
OPCODE(RET_FLAG);
OPCODE(CALL);
OPCODE(SIBCALL);
OPCODE(PERMUTE_DWORDS);
OPCODE(PERMUTE);
OPCODE(PACK);
+ OPCODE(PACKS_CC);
+ OPCODE(PACKLS_CC);
+ OPCODE(UNPACK_HIGH);
+ OPCODE(UNPACKL_HIGH);
+ OPCODE(UNPACK_LOW);
+ OPCODE(UNPACKL_LOW);
OPCODE(VSHL_BY_SCALAR);
OPCODE(VSRL_BY_SCALAR);
OPCODE(VSRA_BY_SCALAR);
OPCODE(VICMPE);
OPCODE(VICMPH);
OPCODE(VICMPHL);
+ OPCODE(VICMPES);
+ OPCODE(VICMPHS);
+ OPCODE(VICMPHLS);
+ OPCODE(VFCMPE);
+ OPCODE(VFCMPH);
+ OPCODE(VFCMPHE);
+ OPCODE(VFCMPES);
+ OPCODE(VFCMPHS);
+ OPCODE(VFCMPHES);
+ OPCODE(VFTCI);
+ OPCODE(VEXTEND);
+ OPCODE(VROUND);
+ OPCODE(VTM);
+ OPCODE(VFAE_CC);
+ OPCODE(VFAEZ_CC);
+ OPCODE(VFEE_CC);
+ OPCODE(VFEEZ_CC);
+ OPCODE(VFENE_CC);
+ OPCODE(VFENEZ_CC);
+ OPCODE(VISTR_CC);
+ OPCODE(VSTRC_CC);
+ OPCODE(VSTRCZ_CC);
OPCODE(ATOMIC_SWAPW);
OPCODE(ATOMIC_LOADW_ADD);
OPCODE(ATOMIC_LOADW_SUB);
}
}
}
+ if (Opcode == SystemZISD::MERGE_HIGH ||
+ Opcode == SystemZISD::MERGE_LOW) {
+ SDValue Op0 = N->getOperand(0);
+ SDValue Op1 = N->getOperand(1);
+ if (Op0.getOpcode() == ISD::BITCAST)
+ Op0 = Op0.getOperand(0);
+ if (Op0.getOpcode() == SystemZISD::BYTE_MASK &&
+ cast<ConstantSDNode>(Op0.getOperand(0))->getZExtValue() == 0) {
+ // (z_merge_* 0, 0) -> 0. This is mostly useful for using VLLEZF
+ // for v4f32.
+ if (Op1 == N->getOperand(0))
+ return Op1;
+ // (z_merge_? 0, X) -> (z_unpackl_? 0, X).
+ EVT VT = Op1.getValueType();
+ unsigned ElemBytes = VT.getVectorElementType().getStoreSize();
+ if (ElemBytes <= 4) {
+ Opcode = (Opcode == SystemZISD::MERGE_HIGH ?
+ SystemZISD::UNPACKL_HIGH : SystemZISD::UNPACKL_LOW);
+ EVT InVT = VT.changeVectorElementTypeToInteger();
+ EVT OutVT = MVT::getVectorVT(MVT::getIntegerVT(ElemBytes * 16),
+ SystemZ::VectorBytes / ElemBytes / 2);
+ if (VT != InVT) {
+ Op1 = DAG.getNode(ISD::BITCAST, SDLoc(N), InVT, Op1);
+ DCI.AddToWorklist(Op1.getNode());
+ }
+ SDValue Op = DAG.getNode(Opcode, SDLoc(N), OutVT, Op1);
+ DCI.AddToWorklist(Op.getNode());
+ return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
+ }
+ }
+ }
// If we have (truncstoreiN (extract_vector_elt X, Y), Z) then it is better
// for the extraction to be done on a vMiN value, so that we can use VSTE.
// If X has wider elements then convert it to:
N->getOperand(0) == N->getOperand(1))
return DAG.getNode(SystemZISD::REPLICATE, SDLoc(N), N->getValueType(0),
N->getOperand(0));
+ // (fround (extract_vector_elt X 0))
+ // (fround (extract_vector_elt X 1)) ->
+ // (extract_vector_elt (VROUND X) 0)
+ // (extract_vector_elt (VROUND X) 1)
+ //
+ // This is a special case since the target doesn't really support v2f32s.
+ if (Opcode == ISD::FP_ROUND) {
+ SDValue Op0 = N->getOperand(0);
+ if (N->getValueType(0) == MVT::f32 &&
+ Op0.hasOneUse() &&
+ Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+ Op0.getOperand(0).getValueType() == MVT::v2f64 &&
+ Op0.getOperand(1).getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue() == 0) {
+ SDValue Vec = Op0.getOperand(0);
+ for (auto *U : Vec->uses()) {
+ if (U != Op0.getNode() &&
+ U->hasOneUse() &&
+ U->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+ U->getOperand(0) == Vec &&
+ U->getOperand(1).getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(U->getOperand(1))->getZExtValue() == 1) {
+ SDValue OtherRound = SDValue(*U->use_begin(), 0);
+ if (OtherRound.getOpcode() == ISD::FP_ROUND &&
+ OtherRound.getOperand(0) == SDValue(U, 0) &&
+ OtherRound.getValueType() == MVT::f32) {
+ SDValue VRound = DAG.getNode(SystemZISD::VROUND, SDLoc(N),
+ MVT::v4f32, Vec);
+ DCI.AddToWorklist(VRound.getNode());
+ SDValue Extract1 =
+ DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(U), MVT::f32,
+ VRound, DAG.getConstant(2, SDLoc(U), MVT::i32));
+ DCI.AddToWorklist(Extract1.getNode());
+ DAG.ReplaceAllUsesOfValueWith(OtherRound, Extract1);
+ SDValue Extract0 =
+ DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op0), MVT::f32,
+ VRound, DAG.getConstant(0, SDLoc(Op0), MVT::i32));
+ return Extract0;
+ }
+ }
+ }
+ }
+ }
return SDValue();
}