// Return a mask with Count low bits set.
static uint64_t allOnes(unsigned int Count) {
- return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1;
+ assert(Count <= 64);
+ if (Count > 63)
+ return UINT64_MAX;
+ return (uint64_t(1) << Count) - 1;
}
// Represents operands 2 to 5 of the ROTATE AND ... SELECTED BITS operation
// Used by SystemZOperands.td to create integer constants.
inline SDValue getImm(const SDNode *Node, uint64_t Imm) const {
- return CurDAG->getTargetConstant(Imm, Node->getValueType(0));
+ return CurDAG->getTargetConstant(Imm, SDLoc(Node), Node->getValueType(0));
}
const SystemZTargetMachine &getTargetMachine() const {
Addr, Base, Disp, Index);
}
+ // Try to match Addr as an address with a base, 12-bit displacement
+ // and index, where the index is element Elem of a vector.
+ // Return true on success, storing the base, displacement and vector
+ // in Base, Disp and Index respectively.
+ bool selectBDVAddr12Only(SDValue Addr, SDValue Elem, SDValue &Base,
+ SDValue &Disp, SDValue &Index) const;
+
// Check whether (or Op (and X InsertMask)) is effectively an insertion
// of X into bits InsertMask of some Y != Op. Return true if so and
// set Op to that Y.
SDNode *splitLargeImmediate(unsigned Opcode, SDNode *Node, SDValue Op0,
uint64_t UpperVal, uint64_t LowerVal);
+ // Try to use gather instruction Opcode to implement vector insertion N.
+ SDNode *tryGather(SDNode *N, unsigned Opcode);
+
+ // Try to use scatter instruction Opcode to implement store Store.
+ SDNode *tryScatter(StoreSDNode *Store, unsigned Opcode);
+
// Return true if Load and Store are loads and stores of the same size
// and are guaranteed not to overlap. Such operations can be implemented
// using block (SS-format) instructions.
// Override SelectionDAGISel.
SDNode *Select(SDNode *Node) override;
- bool SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
+ bool SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
std::vector<SDValue> &OutOps) override;
// Include the pieces autogenerated from the target description.
}
// Lower the displacement to a TargetConstant.
- Disp = CurDAG->getTargetConstant(AM.Disp, VT);
+ Disp = CurDAG->getTargetConstant(AM.Disp, SDLoc(Base), VT);
}
void SystemZDAGToDAGISel::getAddressOperands(const SystemZAddressingMode &AM,
return true;
}
+bool SystemZDAGToDAGISel::selectBDVAddr12Only(SDValue Addr, SDValue Elem,
+ SDValue &Base,
+ SDValue &Disp,
+ SDValue &Index) const {
+ SDValue Regs[2];
+ if (selectBDXAddr12Only(Addr, Regs[0], Disp, Regs[1]) &&
+ Regs[0].getNode() && Regs[1].getNode()) {
+ for (unsigned int I = 0; I < 2; ++I) {
+ Base = Regs[I];
+ Index = Regs[1 - I];
+ // We can't tell here whether the index vector has the right type
+ // for the access; the caller needs to do that instead.
+ if (Index.getOpcode() == ISD::ZERO_EXTEND)
+ Index = Index.getOperand(0);
+ if (Index.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+ Index.getOperand(1) == Elem) {
+ Index = Index.getOperand(0);
+ return true;
+ }
+ }
+ }
+ return false;
+}
+
bool SystemZDAGToDAGISel::detectOrAndInsertion(SDValue &Op,
uint64_t InsertMask) const {
// We're only interested in cases where the insertion is into some operand
RxSBG.Input = N.getOperand(0);
return true;
}
-
+
case ISD::ANY_EXTEND:
// Bits above the extended operand are don't-care.
RxSBG.Input = N.getOperand(0);
return true;
}
// Fall through.
-
+
case ISD::SIGN_EXTEND: {
// Check that the extension bits are don't-care (i.e. are masked out
// by the final mask).
}
SDNode *SystemZDAGToDAGISel::tryRISBGZero(SDNode *N) {
+ SDLoc DL(N);
EVT VT = N->getValueType(0);
+ if (!VT.isInteger() || VT.getSizeInBits() > 64)
+ return nullptr;
RxSBGOperands RISBG(SystemZ::RISBG, SDValue(N, 0));
unsigned Count = 0;
while (expandRxSBG(RISBG))
// Force the new mask into the DAG, since it may include known-one bits.
auto *MaskN = cast<ConstantSDNode>(N->getOperand(1).getNode());
if (MaskN->getZExtValue() != RISBG.Mask) {
- SDValue NewMask = CurDAG->getConstant(RISBG.Mask, VT);
+ SDValue NewMask = CurDAG->getConstant(RISBG.Mask, DL, VT);
N = CurDAG->UpdateNodeOperands(N, N->getOperand(0), NewMask);
return SelectCode(N);
}
return nullptr;
}
- }
+ }
+
+ // If the RISBG operands require no rotation and just masks the bottom
+ // 8/16 bits, attempt to convert this to a LLC zero extension.
+ if (RISBG.Rotate == 0 && (RISBG.Mask == 0xff || RISBG.Mask == 0xffff)) {
+ unsigned OpCode = (RISBG.Mask == 0xff ? SystemZ::LLGCR : SystemZ::LLGHR);
+ if (VT == MVT::i32) {
+ if (Subtarget->hasHighWord())
+ OpCode = (RISBG.Mask == 0xff ? SystemZ::LLCRMux : SystemZ::LLHRMux);
+ else
+ OpCode = (RISBG.Mask == 0xff ? SystemZ::LLCR : SystemZ::LLHR);
+ }
+
+ SDValue In = convertTo(DL, VT, RISBG.Input);
+ N = CurDAG->getMachineNode(OpCode, DL, VT, In);
+ return convertTo(DL, VT, SDValue(N, 0)).getNode();
+ }
unsigned Opcode = SystemZ::RISBG;
+ // Prefer RISBGN if available, since it does not clobber CC.
+ if (Subtarget->hasMiscellaneousExtensions())
+ Opcode = SystemZ::RISBGN;
EVT OpcodeVT = MVT::i64;
if (VT == MVT::i32 && Subtarget->hasHighWord()) {
Opcode = SystemZ::RISBMux;
RISBG.End &= 31;
}
SDValue Ops[5] = {
- getUNDEF(SDLoc(N), OpcodeVT),
- convertTo(SDLoc(N), OpcodeVT, RISBG.Input),
- CurDAG->getTargetConstant(RISBG.Start, MVT::i32),
- CurDAG->getTargetConstant(RISBG.End | 128, MVT::i32),
- CurDAG->getTargetConstant(RISBG.Rotate, MVT::i32)
+ getUNDEF(DL, OpcodeVT),
+ convertTo(DL, OpcodeVT, RISBG.Input),
+ CurDAG->getTargetConstant(RISBG.Start, DL, MVT::i32),
+ CurDAG->getTargetConstant(RISBG.End | 128, DL, MVT::i32),
+ CurDAG->getTargetConstant(RISBG.Rotate, DL, MVT::i32)
};
- N = CurDAG->getMachineNode(Opcode, SDLoc(N), OpcodeVT, Ops);
- return convertTo(SDLoc(N), VT, SDValue(N, 0)).getNode();
+ N = CurDAG->getMachineNode(Opcode, DL, OpcodeVT, Ops);
+ return convertTo(DL, VT, SDValue(N, 0)).getNode();
}
SDNode *SystemZDAGToDAGISel::tryRxSBG(SDNode *N, unsigned Opcode) {
+ SDLoc DL(N);
+ EVT VT = N->getValueType(0);
+ if (!VT.isInteger() || VT.getSizeInBits() > 64)
+ return nullptr;
// Try treating each operand of N as the second operand of the RxSBG
// and see which goes deepest.
RxSBGOperands RxSBG[] = {
// See whether we can avoid an AND in the first operand by converting
// ROSBG to RISBG.
- if (Opcode == SystemZ::ROSBG && detectOrAndInsertion(Op0, RxSBG[I].Mask))
+ if (Opcode == SystemZ::ROSBG && detectOrAndInsertion(Op0, RxSBG[I].Mask)) {
Opcode = SystemZ::RISBG;
-
- EVT VT = N->getValueType(0);
+ // Prefer RISBGN if available, since it does not clobber CC.
+ if (Subtarget->hasMiscellaneousExtensions())
+ Opcode = SystemZ::RISBGN;
+ }
+
SDValue Ops[5] = {
- convertTo(SDLoc(N), MVT::i64, Op0),
- convertTo(SDLoc(N), MVT::i64, RxSBG[I].Input),
- CurDAG->getTargetConstant(RxSBG[I].Start, MVT::i32),
- CurDAG->getTargetConstant(RxSBG[I].End, MVT::i32),
- CurDAG->getTargetConstant(RxSBG[I].Rotate, MVT::i32)
+ convertTo(DL, MVT::i64, Op0),
+ convertTo(DL, MVT::i64, RxSBG[I].Input),
+ CurDAG->getTargetConstant(RxSBG[I].Start, DL, MVT::i32),
+ CurDAG->getTargetConstant(RxSBG[I].End, DL, MVT::i32),
+ CurDAG->getTargetConstant(RxSBG[I].Rotate, DL, MVT::i32)
};
- N = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i64, Ops);
- return convertTo(SDLoc(N), VT, SDValue(N, 0)).getNode();
+ N = CurDAG->getMachineNode(Opcode, DL, MVT::i64, Ops);
+ return convertTo(DL, VT, SDValue(N, 0)).getNode();
}
SDNode *SystemZDAGToDAGISel::splitLargeImmediate(unsigned Opcode, SDNode *Node,
uint64_t LowerVal) {
EVT VT = Node->getValueType(0);
SDLoc DL(Node);
- SDValue Upper = CurDAG->getConstant(UpperVal, VT);
+ SDValue Upper = CurDAG->getConstant(UpperVal, DL, VT);
if (Op0.getNode())
Upper = CurDAG->getNode(Opcode, DL, VT, Op0, Upper);
Upper = SDValue(Select(Upper.getNode()), 0);
- SDValue Lower = CurDAG->getConstant(LowerVal, VT);
+ SDValue Lower = CurDAG->getConstant(LowerVal, DL, VT);
SDValue Or = CurDAG->getNode(Opcode, DL, VT, Upper, Lower);
return Or.getNode();
}
+SDNode *SystemZDAGToDAGISel::tryGather(SDNode *N, unsigned Opcode) {
+ SDValue ElemV = N->getOperand(2);
+ auto *ElemN = dyn_cast<ConstantSDNode>(ElemV);
+ if (!ElemN)
+ return 0;
+
+ unsigned Elem = ElemN->getZExtValue();
+ EVT VT = N->getValueType(0);
+ if (Elem >= VT.getVectorNumElements())
+ return 0;
+
+ auto *Load = dyn_cast<LoadSDNode>(N->getOperand(1));
+ if (!Load || !Load->hasOneUse())
+ return 0;
+ if (Load->getMemoryVT().getSizeInBits() !=
+ Load->getValueType(0).getSizeInBits())
+ return 0;
+
+ SDValue Base, Disp, Index;
+ if (!selectBDVAddr12Only(Load->getBasePtr(), ElemV, Base, Disp, Index) ||
+ Index.getValueType() != VT.changeVectorElementTypeToInteger())
+ return 0;
+
+ SDLoc DL(Load);
+ SDValue Ops[] = {
+ N->getOperand(0), Base, Disp, Index,
+ CurDAG->getTargetConstant(Elem, DL, MVT::i32), Load->getChain()
+ };
+ SDNode *Res = CurDAG->getMachineNode(Opcode, DL, VT, MVT::Other, Ops);
+ ReplaceUses(SDValue(Load, 1), SDValue(Res, 1));
+ return Res;
+}
+
+SDNode *SystemZDAGToDAGISel::tryScatter(StoreSDNode *Store, unsigned Opcode) {
+ SDValue Value = Store->getValue();
+ if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
+ return 0;
+ if (Store->getMemoryVT().getSizeInBits() !=
+ Value.getValueType().getSizeInBits())
+ return 0;
+
+ SDValue ElemV = Value.getOperand(1);
+ auto *ElemN = dyn_cast<ConstantSDNode>(ElemV);
+ if (!ElemN)
+ return 0;
+
+ SDValue Vec = Value.getOperand(0);
+ EVT VT = Vec.getValueType();
+ unsigned Elem = ElemN->getZExtValue();
+ if (Elem >= VT.getVectorNumElements())
+ return 0;
+
+ SDValue Base, Disp, Index;
+ if (!selectBDVAddr12Only(Store->getBasePtr(), ElemV, Base, Disp, Index) ||
+ Index.getValueType() != VT.changeVectorElementTypeToInteger())
+ return 0;
+
+ SDLoc DL(Store);
+ SDValue Ops[] = {
+ Vec, Base, Disp, Index, CurDAG->getTargetConstant(Elem, DL, MVT::i32),
+ Store->getChain()
+ };
+ return CurDAG->getMachineNode(Opcode, DL, MVT::Other, Ops);
+}
+
bool SystemZDAGToDAGISel::canUseBlockOperation(StoreSDNode *Store,
LoadSDNode *Load) const {
// Check that the two memory operands have the same size.
if (V1 == V2 && End1 == End2)
return false;
- return !AA->alias(AliasAnalysis::Location(V1, End1, Load->getAAInfo()),
- AliasAnalysis::Location(V2, End2, Store->getAAInfo()));
+ return !AA->alias(MemoryLocation(V1, End1, Load->getAAInfo()),
+ MemoryLocation(V2, End2, Store->getAAInfo()));
}
bool SystemZDAGToDAGISel::storeLoadCanUseMVC(SDNode *N) const {
uint64_t ConstCCMask =
cast<ConstantSDNode>(CCMask.getNode())->getZExtValue();
// Invert the condition.
- CCMask = CurDAG->getConstant(ConstCCValid ^ ConstCCMask,
+ CCMask = CurDAG->getConstant(ConstCCValid ^ ConstCCMask, SDLoc(Node),
CCMask.getValueType());
SDValue Op4 = Node->getOperand(4);
Node = CurDAG->UpdateNodeOperands(Node, Op1, Op0, CCValid, CCMask, Op4);
}
break;
}
+
+ case ISD::INSERT_VECTOR_ELT: {
+ EVT VT = Node->getValueType(0);
+ unsigned ElemBitSize = VT.getVectorElementType().getSizeInBits();
+ if (ElemBitSize == 32)
+ ResNode = tryGather(Node, SystemZ::VGEF);
+ else if (ElemBitSize == 64)
+ ResNode = tryGather(Node, SystemZ::VGEG);
+ break;
+ }
+
+ case ISD::STORE: {
+ auto *Store = cast<StoreSDNode>(Node);
+ unsigned ElemBitSize = Store->getValue().getValueType().getSizeInBits();
+ if (ElemBitSize == 32)
+ ResNode = tryScatter(Store, SystemZ::VSCEF);
+ else if (ElemBitSize == 64)
+ ResNode = tryScatter(Store, SystemZ::VSCEG);
+ break;
+ }
}
// Select the default instruction
bool SystemZDAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDValue &Op,
- char ConstraintCode,
+ unsigned ConstraintID,
std::vector<SDValue> &OutOps) {
- assert(ConstraintCode == 'm' && "Unexpected constraint code");
- // Accept addresses with short displacements, which are compatible
- // with Q, R, S and T. But keep the index operand for future expansion.
- SDValue Base, Disp, Index;
- if (!selectBDXAddr(SystemZAddressingMode::FormBD,
- SystemZAddressingMode::Disp12Only,
- Op, Base, Disp, Index))
- return true;
- OutOps.push_back(Base);
- OutOps.push_back(Disp);
- OutOps.push_back(Index);
- return false;
+ switch(ConstraintID) {
+ default:
+ llvm_unreachable("Unexpected asm memory constraint");
+ case InlineAsm::Constraint_i:
+ case InlineAsm::Constraint_m:
+ case InlineAsm::Constraint_Q:
+ case InlineAsm::Constraint_R:
+ case InlineAsm::Constraint_S:
+ case InlineAsm::Constraint_T:
+ // Accept addresses with short displacements, which are compatible
+ // with Q, R, S and T. But keep the index operand for future expansion.
+ SDValue Base, Disp, Index;
+ if (selectBDXAddr(SystemZAddressingMode::FormBD,
+ SystemZAddressingMode::Disp12Only,
+ Op, Base, Disp, Index)) {
+ OutOps.push_back(Base);
+ OutOps.push_back(Disp);
+ OutOps.push_back(Index);
+ return false;
+ }
+ break;
+ }
+ return true;
}