diff options
Diffstat (limited to 'lib/Target/Hexagon')
21 files changed, 1627 insertions, 2959 deletions
diff --git a/lib/Target/Hexagon/Hexagon.td b/lib/Target/Hexagon/Hexagon.td index 8a5ee40..af1c56b 100644 --- a/lib/Target/Hexagon/Hexagon.td +++ b/lib/Target/Hexagon/Hexagon.td @@ -89,7 +89,7 @@ def getPredOpcode : InstrMapping { // def getPredNewOpcode : InstrMapping { let FilterClass = "PredNewRel"; - let RowFields = ["BaseOpcode", "PredSense", "isNVStore"]; + let RowFields = ["BaseOpcode", "PredSense", "isNVStore", "isBrTaken"]; let ColFields = ["PNewValue"]; let KeyCol = [""]; let ValueCols = [["new"]]; diff --git a/lib/Target/Hexagon/HexagonCFGOptimizer.cpp b/lib/Target/Hexagon/HexagonCFGOptimizer.cpp index d4078ad..b6022ca 100644 --- a/lib/Target/Hexagon/HexagonCFGOptimizer.cpp +++ b/lib/Target/Hexagon/HexagonCFGOptimizer.cpp @@ -52,8 +52,8 @@ private: char HexagonCFGOptimizer::ID = 0; static bool IsConditionalBranch(int Opc) { - return (Opc == Hexagon::JMP_c) || (Opc == Hexagon::JMP_cNot) - || (Opc == Hexagon::JMP_cdnPt) || (Opc == Hexagon::JMP_cdnNotPt); + return (Opc == Hexagon::JMP_t) || (Opc == Hexagon::JMP_f) + || (Opc == Hexagon::JMP_tnew_t) || (Opc == Hexagon::JMP_fnew_t); } @@ -68,20 +68,20 @@ HexagonCFGOptimizer::InvertAndChangeJumpTarget(MachineInstr* MI, const HexagonInstrInfo *QII = QTM.getInstrInfo(); int NewOpcode = 0; switch(MI->getOpcode()) { - case Hexagon::JMP_c: - NewOpcode = Hexagon::JMP_cNot; + case Hexagon::JMP_t: + NewOpcode = Hexagon::JMP_f; break; - case Hexagon::JMP_cNot: - NewOpcode = Hexagon::JMP_c; + case Hexagon::JMP_f: + NewOpcode = Hexagon::JMP_t; break; - case Hexagon::JMP_cdnPt: - NewOpcode = Hexagon::JMP_cdnNotPt; + case Hexagon::JMP_tnew_t: + NewOpcode = Hexagon::JMP_fnew_t; break; - case Hexagon::JMP_cdnNotPt: - NewOpcode = Hexagon::JMP_cdnPt; + case Hexagon::JMP_fnew_t: + NewOpcode = Hexagon::JMP_tnew_t; break; default: @@ -156,8 +156,8 @@ bool HexagonCFGOptimizer::runOnMachineFunction(MachineFunction &Fn) { // The target of the unconditional branch must be JumpAroundTarget. // TODO: If not, we should not invert the unconditional branch. MachineBasicBlock* CondBranchTarget = NULL; - if ((MI->getOpcode() == Hexagon::JMP_c) || - (MI->getOpcode() == Hexagon::JMP_cNot)) { + if ((MI->getOpcode() == Hexagon::JMP_t) || + (MI->getOpcode() == Hexagon::JMP_f)) { CondBranchTarget = MI->getOperand(1).getMBB(); } diff --git a/lib/Target/Hexagon/HexagonFrameLowering.cpp b/lib/Target/Hexagon/HexagonFrameLowering.cpp index d6a9329..de993ee8 100644 --- a/lib/Target/Hexagon/HexagonFrameLowering.cpp +++ b/lib/Target/Hexagon/HexagonFrameLowering.cpp @@ -189,7 +189,7 @@ void HexagonFrameLowering::emitEpilogue(MachineFunction &MF, // Replace 'jumpr r31' instruction with dealloc_return for V4 and higher // versions. - if (STI.hasV4TOps() && MBBI->getOpcode() == Hexagon::JMPR + if (STI.hasV4TOps() && MBBI->getOpcode() == Hexagon::JMPret && !DisableDeallocRet) { // Remove jumpr node. MBB.erase(MBBI); diff --git a/lib/Target/Hexagon/HexagonHardwareLoops.cpp b/lib/Target/Hexagon/HexagonHardwareLoops.cpp index 62aed13..d002788 100644 --- a/lib/Target/Hexagon/HexagonHardwareLoops.cpp +++ b/lib/Target/Hexagon/HexagonHardwareLoops.cpp @@ -541,12 +541,6 @@ CountValue *HexagonHardwareLoops::getLoopTripCount(MachineLoop *L, case Hexagon::CMPEQrr: Cmp = !Negated ? Comparison::EQ : Comparison::NE; break; - case Hexagon::CMPLTrr: - Cmp = !Negated ? Comparison::LTs : Comparison::GEs; - break; - case Hexagon::CMPLTUrr: - Cmp = !Negated ? Comparison::LTu : Comparison::GEu; - break; case Hexagon::CMPGTUri: case Hexagon::CMPGTUrr: Cmp = !Negated ? Comparison::GTu : Comparison::LEu; @@ -701,7 +695,7 @@ CountValue *HexagonHardwareLoops::computeCount(MachineLoop *Loop, // If the induction variable bump is not a power of 2, quit. // Othwerise we'd need a general integer division. - if (!isPowerOf2_64(abs(IVBump))) + if (!isPowerOf2_64(abs64(IVBump))) return 0; MachineBasicBlock *PH = Loop->getLoopPreheader(); @@ -1125,8 +1119,8 @@ bool HexagonHardwareLoops::convertToHardwareLoop(MachineLoop *L) { // The loop ends with either: // - a conditional branch followed by an unconditional branch, or // - a conditional branch to the loop start. - if (LastI->getOpcode() == Hexagon::JMP_c || - LastI->getOpcode() == Hexagon::JMP_cNot) { + if (LastI->getOpcode() == Hexagon::JMP_t || + LastI->getOpcode() == Hexagon::JMP_f) { // Delete one and change/add an uncond. branch to out of the loop. MachineBasicBlock *BranchTarget = LastI->getOperand(1).getMBB(); LastI = LastMBB->erase(LastI); @@ -1430,7 +1424,6 @@ MachineBasicBlock *HexagonHardwareLoops::createPreheaderForLoop( return 0; typedef MachineBasicBlock::instr_iterator instr_iterator; - typedef MachineBasicBlock::pred_iterator pred_iterator; // Verify that all existing predecessors have analyzable branches // (or no branches at all). diff --git a/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp b/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp index 3a1c48b..ba6c100 100644 --- a/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp +++ b/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp @@ -113,6 +113,46 @@ public: SDNode *SelectAdd(SDNode *N); bool isConstExtProfitable(SDNode *N) const; +// XformMskToBitPosU5Imm - Returns the bit position which +// the single bit 32 bit mask represents. +// Used in Clr and Set bit immediate memops. +SDValue XformMskToBitPosU5Imm(uint32_t Imm) { + int32_t bitPos; + bitPos = Log2_32(Imm); + assert(bitPos >= 0 && bitPos < 32 && + "Constant out of range for 32 BitPos Memops"); + return CurDAG->getTargetConstant(bitPos, MVT::i32); +} + +// XformMskToBitPosU4Imm - Returns the bit position which the single bit 16 bit +// mask represents. Used in Clr and Set bit immediate memops. +SDValue XformMskToBitPosU4Imm(uint16_t Imm) { + return XformMskToBitPosU5Imm(Imm); +} + +// XformMskToBitPosU3Imm - Returns the bit position which the single bit 8 bit +// mask represents. Used in Clr and Set bit immediate memops. +SDValue XformMskToBitPosU3Imm(uint8_t Imm) { + return XformMskToBitPosU5Imm(Imm); +} + +// Return true if there is exactly one bit set in V, i.e., if V is one of the +// following integers: 2^0, 2^1, ..., 2^31. +bool ImmIsSingleBit(uint32_t v) const { + uint32_t c = CountPopulation_64(v); + // Only return true if we counted 1 bit. + return c == 1; +} + +// XformM5ToU5Imm - Return a target constant with the specified value, of type +// i32 where the negative literal is transformed into a positive literal for +// use in -= memops. +inline SDValue XformM5ToU5Imm(signed Imm) { + assert( (Imm >= -31 && Imm <= -1) && "Constant out of range for Memops"); + return CurDAG->getTargetConstant( - Imm, MVT::i32); +} + + // XformU7ToU7M1Imm - Return a target constant decremented by 1, in range // [1..128], used in cmpb.gtu instructions. inline SDValue XformU7ToU7M1Imm(signed Imm) { @@ -120,6 +160,17 @@ inline SDValue XformU7ToU7M1Imm(signed Imm) { return CurDAG->getTargetConstant(Imm - 1, MVT::i8); } +// XformS8ToS8M1Imm - Return a target constant decremented by 1. +inline SDValue XformSToSM1Imm(signed Imm) { + return CurDAG->getTargetConstant(Imm - 1, MVT::i32); +} + +// XformU8ToU8M1Imm - Return a target constant decremented by 1. +inline SDValue XformUToUM1Imm(unsigned Imm) { + assert((Imm >= 1) && "Cannot decrement unsigned int less than 1"); + return CurDAG->getTargetConstant(Imm - 1, MVT::i32); +} + // Include the pieces autogenerated from the target description. #include "HexagonGenDAGISel.inc" }; @@ -657,7 +708,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, DebugLoc dl) { // Build post increment store. SDNode* Result = CurDAG->getMachineNode(Opcode, dl, MVT::i32, - MVT::Other, Ops, 4); + MVT::Other, Ops); MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); MemOp[0] = ST->getMemOperand(); cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1); @@ -683,8 +734,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, DebugLoc dl) { // Build regular store. SDValue TargetConstVal = CurDAG->getTargetConstant(Val, MVT::i32); - SDNode* Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops, - 4); + SDNode* Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops); // Build splitted incriment instruction. SDNode* Result_2 = CurDAG->getMachineNode(Hexagon::ADD_ri, dl, MVT::i32, Base, @@ -740,7 +790,7 @@ SDNode *HexagonDAGToDAGISel::SelectBaseOffsetStore(StoreSDNode *ST, Value, Chain}; // build indexed store SDNode* Result = CurDAG->getMachineNode(Opcode, dl, - MVT::Other, Ops, 4); + MVT::Other, Ops); MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); MemOp[0] = ST->getMemOperand(); cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1); @@ -1190,8 +1240,7 @@ SDNode *HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) { } EVT ReturnValueVT = N->getValueType(0); SDNode *Result = CurDAG->getMachineNode(IntrinsicWithPred, dl, - ReturnValueVT, - Ops.data(), Ops.size()); + ReturnValueVT, Ops); ReplaceUses(N, Result); return Result; } diff --git a/lib/Target/Hexagon/HexagonISelLowering.cpp b/lib/Target/Hexagon/HexagonISelLowering.cpp index 0a8b1af..0e5b8dc 100644 --- a/lib/Target/Hexagon/HexagonISelLowering.cpp +++ b/lib/Target/Hexagon/HexagonISelLowering.cpp @@ -1002,14 +1002,6 @@ HexagonTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { return FrameAddr; } - -SDValue HexagonTargetLowering::LowerMEMBARRIER(SDValue Op, - SelectionDAG& DAG) const { - DebugLoc dl = Op.getDebugLoc(); - return DAG.getNode(HexagonISD::BARRIER, dl, MVT::Other, Op.getOperand(0)); -} - - SDValue HexagonTargetLowering::LowerATOMIC_FENCE(SDValue Op, SelectionDAG& DAG) const { DebugLoc dl = Op.getDebugLoc(); @@ -1361,7 +1353,6 @@ HexagonTargetLowering::HexagonTargetLowering(HexagonTargetMachine } - setOperationAction(ISD::BRIND, MVT::Other, Expand); if (EmitJumpTables) { setOperationAction(ISD::BR_JT, MVT::Other, Custom); } else { @@ -1370,12 +1361,13 @@ HexagonTargetLowering::HexagonTargetLowering(HexagonTargetMachine // Increase jump tables cutover to 5, was 4. setMinimumJumpTableEntries(5); + setOperationAction(ISD::BR_CC, MVT::Other, Expand); setOperationAction(ISD::BR_CC, MVT::f32, Expand); setOperationAction(ISD::BR_CC, MVT::f64, Expand); setOperationAction(ISD::BR_CC, MVT::i1, Expand); setOperationAction(ISD::BR_CC, MVT::i32, Expand); + setOperationAction(ISD::BR_CC, MVT::i64, Expand); - setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom); setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom); setOperationAction(ISD::FSIN , MVT::f64, Expand); @@ -1442,7 +1434,7 @@ HexagonTargetLowering::HexagonTargetLowering(HexagonTargetMachine setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand); setOperationAction(ISD::EHSELECTION, MVT::i32, Expand); - setOperationAction(ISD::EH_RETURN, MVT::Other, Expand); + setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); if (TM.getSubtargetImpl()->isSubtargetV2()) { setExceptionPointerRegister(Hexagon::R20); @@ -1497,6 +1489,7 @@ HexagonTargetLowering::getTargetNodeName(unsigned Opcode) const { case HexagonISD::RET_FLAG: return "HexagonISD::RET_FLAG"; case HexagonISD::BR_JT: return "HexagonISD::BR_JT"; case HexagonISD::TC_RETURN: return "HexagonISD::TC_RETURN"; + case HexagonISD::EH_RETURN: return "HexagonISD::EH_RETURN"; } } @@ -1518,16 +1511,43 @@ bool HexagonTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const { } SDValue +HexagonTargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const { + SDValue Chain = Op.getOperand(0); + SDValue Offset = Op.getOperand(1); + SDValue Handler = Op.getOperand(2); + DebugLoc dl = Op.getDebugLoc(); + + // Mark function as containing a call to EH_RETURN. + HexagonMachineFunctionInfo *FuncInfo = + DAG.getMachineFunction().getInfo<HexagonMachineFunctionInfo>(); + FuncInfo->setHasEHReturn(); + + unsigned OffsetReg = Hexagon::R28; + + SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), + DAG.getRegister(Hexagon::R30, getPointerTy()), + DAG.getIntPtrConstant(4)); + Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(), + false, false, 0); + Chain = DAG.getCopyToReg(Chain, dl, OffsetReg, Offset); + + // Not needed we already use it as explict input to EH_RETURN. + // MF.getRegInfo().addLiveOut(OffsetReg); + + return DAG.getNode(HexagonISD::EH_RETURN, dl, MVT::Other, Chain); +} + +SDValue HexagonTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { switch (Op.getOpcode()) { default: llvm_unreachable("Should not custom lower this!"); case ISD::ConstantPool: return LowerConstantPool(Op, DAG); + case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG); // Frame & Return address. Currently unimplemented. case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); case ISD::GlobalTLSAddress: llvm_unreachable("TLS not implemented for Hexagon."); - case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG); case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG); case ISD::GlobalAddress: return LowerGLOBALADDRESS(Op, DAG); case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); diff --git a/lib/Target/Hexagon/HexagonISelLowering.h b/lib/Target/Hexagon/HexagonISelLowering.h index 3279cc6..bb1acc1 100644 --- a/lib/Target/Hexagon/HexagonISelLowering.h +++ b/lib/Target/Hexagon/HexagonISelLowering.h @@ -62,7 +62,8 @@ namespace llvm { WrapperShuffEH, WrapperShuffOB, WrapperShuffOH, - TC_RETURN + TC_RETURN, + EH_RETURN }; } @@ -101,6 +102,7 @@ namespace llvm { SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const; SDValue LowerINLINEASM(SDValue Op, SelectionDAG &DAG) const; SDValue LowerEH_LABEL(SDValue Op, SelectionDAG &DAG) const; + SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const; SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, @@ -122,7 +124,6 @@ namespace llvm { SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const; SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; - SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG& DAG) const; SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG& DAG) const; SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; diff --git a/lib/Target/Hexagon/HexagonInstrInfo.cpp b/lib/Target/Hexagon/HexagonInstrInfo.cpp index 96a252e..e0beab0 100644 --- a/lib/Target/Hexagon/HexagonInstrInfo.cpp +++ b/lib/Target/Hexagon/HexagonInstrInfo.cpp @@ -23,6 +23,7 @@ #include "llvm/CodeGen/MachineMemOperand.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/PseudoSourceValue.h" +#include "llvm/Support/Debug.h" #include "llvm/Support/MathExtras.h" #define GET_INSTRINFO_CTOR #define GET_INSTRMAP_INFO @@ -118,16 +119,16 @@ HexagonInstrInfo::InsertBranch(MachineBasicBlock &MBB,MachineBasicBlock *TBB, DebugLoc DL) const{ int BOpc = Hexagon::JMP; - int BccOpc = Hexagon::JMP_c; + int BccOpc = Hexagon::JMP_t; assert(TBB && "InsertBranch must not be told to insert a fallthrough"); int regPos = 0; // Check if ReverseBranchCondition has asked to reverse this branch // If we want to reverse the branch an odd number of times, we want - // JMP_cNot. + // JMP_f. if (!Cond.empty() && Cond[0].isImm() && Cond[0].getImm() == 0) { - BccOpc = Hexagon::JMP_cNot; + BccOpc = Hexagon::JMP_f; regPos = 1; } @@ -174,8 +175,8 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, FBB = NULL; // If the block has no terminators, it just falls into the block after it. - MachineBasicBlock::iterator I = MBB.end(); - if (I == MBB.begin()) + MachineBasicBlock::instr_iterator I = MBB.instr_end(); + if (I == MBB.instr_begin()) return false; // A basic block may looks like this: @@ -194,13 +195,24 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, --I; if (I->isEHLabel()) return true; - } while (I != MBB.begin()); + } while (I != MBB.instr_begin()); - I = MBB.end(); + I = MBB.instr_end(); --I; while (I->isDebugValue()) { - if (I == MBB.begin()) + if (I == MBB.instr_begin()) + return false; + --I; + } + + // Delete the JMP if it's equivalent to a fall-through. + if (AllowModify && I->getOpcode() == Hexagon::JMP && + MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) { + DEBUG(dbgs()<< "\nErasing the jump to successor block\n";); + I->eraseFromParent(); + I = MBB.instr_end(); + if (I == MBB.instr_begin()) return false; --I; } @@ -209,23 +221,42 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, // Get the last instruction in the block. MachineInstr *LastInst = I; + MachineInstr *SecondLastInst = NULL; + // Find one more terminator if present. + do { + if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(I)) { + if (!SecondLastInst) + SecondLastInst = I; + else + // This is a third branch. + return true; + } + if (I == MBB.instr_begin()) + break; + --I; + } while(I); + + int LastOpcode = LastInst->getOpcode(); + + bool LastOpcodeHasJMP_c = PredOpcodeHasJMP_c(LastOpcode); + bool LastOpcodeHasNot = PredOpcodeHasNot(LastOpcode); // If there is only one terminator instruction, process it. - if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) { - if (LastInst->getOpcode() == Hexagon::JMP) { + if (LastInst && !SecondLastInst) { + if (LastOpcode == Hexagon::JMP) { TBB = LastInst->getOperand(0).getMBB(); return false; } - if (LastInst->getOpcode() == Hexagon::JMP_c) { - // Block ends with fall-through true condbranch. - TBB = LastInst->getOperand(1).getMBB(); + if (LastOpcode == Hexagon::ENDLOOP0) { + TBB = LastInst->getOperand(0).getMBB(); Cond.push_back(LastInst->getOperand(0)); return false; } - if (LastInst->getOpcode() == Hexagon::JMP_cNot) { - // Block ends with fall-through false condbranch. + if (LastOpcodeHasJMP_c) { TBB = LastInst->getOperand(1).getMBB(); - Cond.push_back(MachineOperand::CreateImm(0)); + if (LastOpcodeHasNot) { + Cond.push_back(MachineOperand::CreateImm(0)); + } Cond.push_back(LastInst->getOperand(0)); return false; } @@ -233,29 +264,14 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, return true; } - // Get the instruction before it if it's a terminator. - MachineInstr *SecondLastInst = I; - - // If there are three terminators, we don't know what sort of block this is. - if (SecondLastInst && I != MBB.begin() && - isUnpredicatedTerminator(--I)) - return true; - - // If the block ends with Hexagon::BRCOND and Hexagon:JMP, handle it. - if (((SecondLastInst->getOpcode() == Hexagon::BRCOND) || - (SecondLastInst->getOpcode() == Hexagon::JMP_c)) && - LastInst->getOpcode() == Hexagon::JMP) { - TBB = SecondLastInst->getOperand(1).getMBB(); - Cond.push_back(SecondLastInst->getOperand(0)); - FBB = LastInst->getOperand(0).getMBB(); - return false; - } + int SecLastOpcode = SecondLastInst->getOpcode(); - // If the block ends with Hexagon::JMP_cNot and Hexagon:JMP, handle it. - if ((SecondLastInst->getOpcode() == Hexagon::JMP_cNot) && - LastInst->getOpcode() == Hexagon::JMP) { + bool SecLastOpcodeHasJMP_c = PredOpcodeHasJMP_c(SecLastOpcode); + bool SecLastOpcodeHasNot = PredOpcodeHasNot(SecLastOpcode); + if (SecLastOpcodeHasJMP_c && (LastOpcode == Hexagon::JMP)) { TBB = SecondLastInst->getOperand(1).getMBB(); - Cond.push_back(MachineOperand::CreateImm(0)); + if (SecLastOpcodeHasNot) + Cond.push_back(MachineOperand::CreateImm(0)); Cond.push_back(SecondLastInst->getOperand(0)); FBB = LastInst->getOperand(0).getMBB(); return false; @@ -263,8 +279,7 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, // If the block ends with two Hexagon:JMPs, handle it. The second one is not // executed, so remove it. - if (SecondLastInst->getOpcode() == Hexagon::JMP && - LastInst->getOpcode() == Hexagon::JMP) { + if (SecLastOpcode == Hexagon::JMP && LastOpcode == Hexagon::JMP) { TBB = SecondLastInst->getOperand(0).getMBB(); I = LastInst; if (AllowModify) @@ -272,6 +287,15 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, return false; } + // If the block ends with an ENDLOOP, and JMP, handle it. + if (SecLastOpcode == Hexagon::ENDLOOP0 && + LastOpcode == Hexagon::JMP) { + TBB = SecondLastInst->getOperand(0).getMBB(); + Cond.push_back(SecondLastInst->getOperand(0)); + FBB = LastInst->getOperand(0).getMBB(); + return false; + } + // Otherwise, can't handle this. return true; } @@ -279,8 +303,8 @@ bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, unsigned HexagonInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { int BOpc = Hexagon::JMP; - int BccOpc = Hexagon::JMP_c; - int BccOpcNot = Hexagon::JMP_cNot; + int BccOpc = Hexagon::JMP_t; + int BccOpcNot = Hexagon::JMP_f; MachineBasicBlock::iterator I = MBB.end(); if (I == MBB.begin()) return 0; @@ -325,8 +349,6 @@ bool HexagonInstrInfo::analyzeCompare(const MachineInstr *MI, case Hexagon::CMPGTUrr: case Hexagon::CMPGTri: case Hexagon::CMPGTrr: - case Hexagon::CMPLTUrr: - case Hexagon::CMPLTrr: SrcReg = MI->getOperand(1).getReg(); Mask = ~0; break; @@ -366,8 +388,6 @@ bool HexagonInstrInfo::analyzeCompare(const MachineInstr *MI, case Hexagon::CMPhEQrr_xor_V4: case Hexagon::CMPhGTUrr_V4: case Hexagon::CMPhGTrr_shl_V4: - case Hexagon::CMPLTUrr: - case Hexagon::CMPLTrr: SrcReg2 = MI->getOperand(2).getReg(); return true; @@ -537,6 +557,15 @@ MachineInstr *HexagonInstrInfo::foldMemoryOperandImpl(MachineFunction &MF, return(0); } +MachineInstr* +HexagonInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF, + int FrameIx, uint64_t Offset, + const MDNode *MDPtr, + DebugLoc DL) const { + MachineInstrBuilder MIB = BuildMI(MF, DL, get(Hexagon::DBG_VALUE)) + .addImm(0).addImm(Offset).addMetadata(MDPtr); + return &*MIB; +} unsigned HexagonInstrInfo::createVR(MachineFunction* MF, MVT VT) const { @@ -737,11 +766,6 @@ bool HexagonInstrInfo::isNewValueStore(const MachineInstr *MI) const { case Hexagon::STrib_abs_cdnPt_nv_V4: case Hexagon::STrib_abs_cNotPt_nv_V4: case Hexagon::STrib_abs_cdnNotPt_nv_V4: - case Hexagon::STrib_imm_abs_nv_V4: - case Hexagon::STrib_imm_abs_cPt_nv_V4: - case Hexagon::STrib_imm_abs_cdnPt_nv_V4: - case Hexagon::STrib_imm_abs_cNotPt_nv_V4: - case Hexagon::STrib_imm_abs_cdnNotPt_nv_V4: // Store Halfword case Hexagon::STrih_nv_V4: @@ -775,11 +799,6 @@ bool HexagonInstrInfo::isNewValueStore(const MachineInstr *MI) const { case Hexagon::STrih_abs_cdnPt_nv_V4: case Hexagon::STrih_abs_cNotPt_nv_V4: case Hexagon::STrih_abs_cdnNotPt_nv_V4: - case Hexagon::STrih_imm_abs_nv_V4: - case Hexagon::STrih_imm_abs_cPt_nv_V4: - case Hexagon::STrih_imm_abs_cdnPt_nv_V4: - case Hexagon::STrih_imm_abs_cNotPt_nv_V4: - case Hexagon::STrih_imm_abs_cdnNotPt_nv_V4: // Store Word case Hexagon::STriw_nv_V4: @@ -813,11 +832,6 @@ bool HexagonInstrInfo::isNewValueStore(const MachineInstr *MI) const { case Hexagon::STriw_abs_cdnPt_nv_V4: case Hexagon::STriw_abs_cNotPt_nv_V4: case Hexagon::STriw_abs_cdnNotPt_nv_V4: - case Hexagon::STriw_imm_abs_nv_V4: - case Hexagon::STriw_imm_abs_cPt_nv_V4: - case Hexagon::STriw_imm_abs_cdnPt_nv_V4: - case Hexagon::STriw_imm_abs_cNotPt_nv_V4: - case Hexagon::STriw_imm_abs_cdnNotPt_nv_V4: return true; } } @@ -994,9 +1008,6 @@ bool HexagonInstrInfo::isPredicable(MachineInstr *MI) const { case Hexagon::ZXTB: case Hexagon::ZXTH: return Subtarget.hasV4TOps(); - - case Hexagon::JMPR: - return false; } return true; @@ -1033,10 +1044,10 @@ unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const { case Hexagon::TFRI_cNotPt: return Hexagon::TFRI_cPt; - case Hexagon::JMP_c: - return Hexagon::JMP_cNot; - case Hexagon::JMP_cNot: - return Hexagon::JMP_c; + case Hexagon::JMP_t: + return Hexagon::JMP_f; + case Hexagon::JMP_f: + return Hexagon::JMP_t; case Hexagon::ADD_ri_cPt: return Hexagon::ADD_ri_cNotPt; @@ -1104,10 +1115,10 @@ unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const { return Hexagon::ZXTH_cPt_V4; - case Hexagon::JMPR_cPt: - return Hexagon::JMPR_cNotPt; - case Hexagon::JMPR_cNotPt: - return Hexagon::JMPR_cPt; + case Hexagon::JMPR_t: + return Hexagon::JMPR_f; + case Hexagon::JMPR_f: + return Hexagon::JMPR_t; // V4 indexed+scaled load. case Hexagon::LDrid_indexed_shl_cPt_V4: @@ -1490,8 +1501,8 @@ getMatchingCondBranchOpcode(int Opc, bool invertPredicate) const { return !invertPredicate ? Hexagon::TFRI_cPt : Hexagon::TFRI_cNotPt; case Hexagon::JMP: - return !invertPredicate ? Hexagon::JMP_c : - Hexagon::JMP_cNot; + return !invertPredicate ? Hexagon::JMP_t : + Hexagon::JMP_f; case Hexagon::JMP_EQrrPt_nv_V4: return !invertPredicate ? Hexagon::JMP_EQrrPt_nv_V4 : Hexagon::JMP_EQrrNotPt_nv_V4; @@ -1521,8 +1532,8 @@ getMatchingCondBranchOpcode(int Opc, bool invertPredicate) const { Hexagon::ZXTH_cNotPt_V4; case Hexagon::JMPR: - return !invertPredicate ? Hexagon::JMPR_cPt : - Hexagon::JMPR_cNotPt; + return !invertPredicate ? Hexagon::JMPR_t : + Hexagon::JMPR_f; // V4 indexed+scaled load. case Hexagon::LDrid_indexed_shl_V4: @@ -1821,11 +1832,15 @@ PredicateInstruction(MachineInstr *MI, // It is better to have an assert here to check this. But I don't know how // to write this assert because findFirstPredOperandIdx() would return -1 if (oper < -1) oper = -1; + MI->getOperand(oper+1).ChangeToRegister(PredMO.getReg(), PredMO.isDef(), - PredMO.isImplicit(), PredMO.isKill(), + PredMO.isImplicit(), false, PredMO.isDead(), PredMO.isUndef(), PredMO.isDebug()); + MachineRegisterInfo &RegInfo = MI->getParent()->getParent()->getRegInfo(); + RegInfo.clearKillFlags(PredMO.getReg()); + if (hasGAOpnd) { unsigned int i; @@ -1881,6 +1896,13 @@ bool HexagonInstrInfo::isPredicated(const MachineInstr *MI) const { return ((F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask); } +bool HexagonInstrInfo::isPredicatedNew(const MachineInstr *MI) const { + const uint64_t F = MI->getDesc().TSFlags; + + assert(isPredicated(MI)); + return ((F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask); +} + bool HexagonInstrInfo::DefinesPredicate(MachineInstr *MI, std::vector<MachineOperand> &Pred) const { @@ -1991,46 +2013,28 @@ isValidOffset(const int Opcode, const int Offset) const { return (Offset >= Hexagon_ADDI_OFFSET_MIN) && (Offset <= Hexagon_ADDI_OFFSET_MAX); - case Hexagon::MEMw_ADDi_indexed_MEM_V4 : - case Hexagon::MEMw_SUBi_indexed_MEM_V4 : - case Hexagon::MEMw_ADDr_indexed_MEM_V4 : - case Hexagon::MEMw_SUBr_indexed_MEM_V4 : - case Hexagon::MEMw_ANDr_indexed_MEM_V4 : - case Hexagon::MEMw_ORr_indexed_MEM_V4 : - case Hexagon::MEMw_ADDi_MEM_V4 : - case Hexagon::MEMw_SUBi_MEM_V4 : - case Hexagon::MEMw_ADDr_MEM_V4 : - case Hexagon::MEMw_SUBr_MEM_V4 : - case Hexagon::MEMw_ANDr_MEM_V4 : - case Hexagon::MEMw_ORr_MEM_V4 : + case Hexagon::MemOPw_ADDi_V4 : + case Hexagon::MemOPw_SUBi_V4 : + case Hexagon::MemOPw_ADDr_V4 : + case Hexagon::MemOPw_SUBr_V4 : + case Hexagon::MemOPw_ANDr_V4 : + case Hexagon::MemOPw_ORr_V4 : return (0 <= Offset && Offset <= 255); - case Hexagon::MEMh_ADDi_indexed_MEM_V4 : - case Hexagon::MEMh_SUBi_indexed_MEM_V4 : - case Hexagon::MEMh_ADDr_indexed_MEM_V4 : - case Hexagon::MEMh_SUBr_indexed_MEM_V4 : - case Hexagon::MEMh_ANDr_indexed_MEM_V4 : - case Hexagon::MEMh_ORr_indexed_MEM_V4 : - case Hexagon::MEMh_ADDi_MEM_V4 : - case Hexagon::MEMh_SUBi_MEM_V4 : - case Hexagon::MEMh_ADDr_MEM_V4 : - case Hexagon::MEMh_SUBr_MEM_V4 : - case Hexagon::MEMh_ANDr_MEM_V4 : - case Hexagon::MEMh_ORr_MEM_V4 : + case Hexagon::MemOPh_ADDi_V4 : + case Hexagon::MemOPh_SUBi_V4 : + case Hexagon::MemOPh_ADDr_V4 : + case Hexagon::MemOPh_SUBr_V4 : + case Hexagon::MemOPh_ANDr_V4 : + case Hexagon::MemOPh_ORr_V4 : return (0 <= Offset && Offset <= 127); - case Hexagon::MEMb_ADDi_indexed_MEM_V4 : - case Hexagon::MEMb_SUBi_indexed_MEM_V4 : - case Hexagon::MEMb_ADDr_indexed_MEM_V4 : - case Hexagon::MEMb_SUBr_indexed_MEM_V4 : - case Hexagon::MEMb_ANDr_indexed_MEM_V4 : - case Hexagon::MEMb_ORr_indexed_MEM_V4 : - case Hexagon::MEMb_ADDi_MEM_V4 : - case Hexagon::MEMb_SUBi_MEM_V4 : - case Hexagon::MEMb_ADDr_MEM_V4 : - case Hexagon::MEMb_SUBr_MEM_V4 : - case Hexagon::MEMb_ANDr_MEM_V4 : - case Hexagon::MEMb_ORr_MEM_V4 : + case Hexagon::MemOPb_ADDi_V4 : + case Hexagon::MemOPb_SUBi_V4 : + case Hexagon::MemOPb_ADDr_V4 : + case Hexagon::MemOPb_SUBr_V4 : + case Hexagon::MemOPb_ANDr_V4 : + case Hexagon::MemOPb_ORr_V4 : return (0 <= Offset && Offset <= 63); // LDri_pred and STriw_pred are pseudo operations, so it has to take offset of @@ -2086,44 +2090,33 @@ isMemOp(const MachineInstr *MI) const { switch (MI->getOpcode()) { default: return false; - case Hexagon::MEMw_ADDi_indexed_MEM_V4 : - case Hexagon::MEMw_SUBi_indexed_MEM_V4 : - case Hexagon::MEMw_ADDr_indexed_MEM_V4 : - case Hexagon::MEMw_SUBr_indexed_MEM_V4 : - case Hexagon::MEMw_ANDr_indexed_MEM_V4 : - case Hexagon::MEMw_ORr_indexed_MEM_V4 : - case Hexagon::MEMw_ADDi_MEM_V4 : - case Hexagon::MEMw_SUBi_MEM_V4 : - case Hexagon::MEMw_ADDr_MEM_V4 : - case Hexagon::MEMw_SUBr_MEM_V4 : - case Hexagon::MEMw_ANDr_MEM_V4 : - case Hexagon::MEMw_ORr_MEM_V4 : - case Hexagon::MEMh_ADDi_indexed_MEM_V4 : - case Hexagon::MEMh_SUBi_indexed_MEM_V4 : - case Hexagon::MEMh_ADDr_indexed_MEM_V4 : - case Hexagon::MEMh_SUBr_indexed_MEM_V4 : - case Hexagon::MEMh_ANDr_indexed_MEM_V4 : - case Hexagon::MEMh_ORr_indexed_MEM_V4 : - case Hexagon::MEMh_ADDi_MEM_V4 : - case Hexagon::MEMh_SUBi_MEM_V4 : - case Hexagon::MEMh_ADDr_MEM_V4 : - case Hexagon::MEMh_SUBr_MEM_V4 : - case Hexagon::MEMh_ANDr_MEM_V4 : - case Hexagon::MEMh_ORr_MEM_V4 : - case Hexagon::MEMb_ADDi_indexed_MEM_V4 : - case Hexagon::MEMb_SUBi_indexed_MEM_V4 : - case Hexagon::MEMb_ADDr_indexed_MEM_V4 : - case Hexagon::MEMb_SUBr_indexed_MEM_V4 : - case Hexagon::MEMb_ANDr_indexed_MEM_V4 : - case Hexagon::MEMb_ORr_indexed_MEM_V4 : - case Hexagon::MEMb_ADDi_MEM_V4 : - case Hexagon::MEMb_SUBi_MEM_V4 : - case Hexagon::MEMb_ADDr_MEM_V4 : - case Hexagon::MEMb_SUBr_MEM_V4 : - case Hexagon::MEMb_ANDr_MEM_V4 : - case Hexagon::MEMb_ORr_MEM_V4 : - return true; + case Hexagon::MemOPw_ADDi_V4 : + case Hexagon::MemOPw_SUBi_V4 : + case Hexagon::MemOPw_ADDr_V4 : + case Hexagon::MemOPw_SUBr_V4 : + case Hexagon::MemOPw_ANDr_V4 : + case Hexagon::MemOPw_ORr_V4 : + case Hexagon::MemOPh_ADDi_V4 : + case Hexagon::MemOPh_SUBi_V4 : + case Hexagon::MemOPh_ADDr_V4 : + case Hexagon::MemOPh_SUBr_V4 : + case Hexagon::MemOPh_ANDr_V4 : + case Hexagon::MemOPh_ORr_V4 : + case Hexagon::MemOPb_ADDi_V4 : + case Hexagon::MemOPb_SUBi_V4 : + case Hexagon::MemOPb_ADDr_V4 : + case Hexagon::MemOPb_SUBr_V4 : + case Hexagon::MemOPb_ANDr_V4 : + case Hexagon::MemOPb_ORr_V4 : + case Hexagon::MemOPb_SETBITi_V4: + case Hexagon::MemOPh_SETBITi_V4: + case Hexagon::MemOPw_SETBITi_V4: + case Hexagon::MemOPb_CLRBITi_V4: + case Hexagon::MemOPh_CLRBITi_V4: + case Hexagon::MemOPw_CLRBITi_V4: + return true; } + return false; } @@ -2142,14 +2135,10 @@ bool HexagonInstrInfo::isNewValueJumpCandidate(const MachineInstr *MI) const { default: return false; case Hexagon::CMPEQrr: case Hexagon::CMPEQri: - case Hexagon::CMPLTrr: case Hexagon::CMPGTrr: case Hexagon::CMPGTri: - case Hexagon::CMPLTUrr: case Hexagon::CMPGTUrr: case Hexagon::CMPGTUri: - case Hexagon::CMPGEri: - case Hexagon::CMPGEUri: return true; } } @@ -2382,6 +2371,13 @@ isConditionalStore (const MachineInstr* MI) const { } } +// Returns true, if any one of the operands is a dot new +// insn, whether it is predicated dot new or register dot new. +bool HexagonInstrInfo::isDotNewInst (const MachineInstr* MI) const { + return (isNewValueInst(MI) || + (isPredicated(MI) && isPredicatedNew(MI))); +} + unsigned HexagonInstrInfo::getAddrMode(const MachineInstr* MI) const { const uint64_t F = MI->getDesc().TSFlags; @@ -2476,6 +2472,34 @@ bool HexagonInstrInfo::isConstExtended(MachineInstr *MI) const { return (ImmValue < MinValue || ImmValue > MaxValue); } +// Returns the opcode to use when converting MI, which is a conditional jump, +// into a conditional instruction which uses the .new value of the predicate. +// We also use branch probabilities to add a hint to the jump. +int +HexagonInstrInfo::getDotNewPredJumpOp(MachineInstr *MI, + const + MachineBranchProbabilityInfo *MBPI) const { + + // We assume that block can have at most two successors. + bool taken = false; + MachineBasicBlock *Src = MI->getParent(); + MachineOperand *BrTarget = &MI->getOperand(1); + MachineBasicBlock *Dst = BrTarget->getMBB(); + + const BranchProbability Prediction = MBPI->getEdgeProbability(Src, Dst); + if (Prediction >= BranchProbability(1,2)) + taken = true; + + switch (MI->getOpcode()) { + case Hexagon::JMP_t: + return taken ? Hexagon::JMP_tnew_t : Hexagon::JMP_tnew_nt; + case Hexagon::JMP_f: + return taken ? Hexagon::JMP_fnew_t : Hexagon::JMP_fnew_nt; + + default: + llvm_unreachable("Unexpected jump instruction."); + } +} // Returns true if a particular operand is extendable for an instruction. bool HexagonInstrInfo::isOperandExtended(const MachineInstr *MI, unsigned short OperandNum) const { @@ -2580,3 +2604,18 @@ short HexagonInstrInfo::getNonExtOpcode (const MachineInstr *MI) const { } return -1; } + +bool HexagonInstrInfo::PredOpcodeHasJMP_c(Opcode_t Opcode) const { + return (Opcode == Hexagon::JMP_t) || + (Opcode == Hexagon::JMP_f) || + (Opcode == Hexagon::JMP_tnew_t) || + (Opcode == Hexagon::JMP_fnew_t) || + (Opcode == Hexagon::JMP_tnew_nt) || + (Opcode == Hexagon::JMP_fnew_nt); +} + +bool HexagonInstrInfo::PredOpcodeHasNot(Opcode_t Opcode) const { + return (Opcode == Hexagon::JMP_f) || + (Opcode == Hexagon::JMP_fnew_t) || + (Opcode == Hexagon::JMP_fnew_nt); +} diff --git a/lib/Target/Hexagon/HexagonInstrInfo.h b/lib/Target/Hexagon/HexagonInstrInfo.h index d2f059a..e0bec04 100644 --- a/lib/Target/Hexagon/HexagonInstrInfo.h +++ b/lib/Target/Hexagon/HexagonInstrInfo.h @@ -16,9 +16,9 @@ #include "HexagonRegisterInfo.h" #include "MCTargetDesc/HexagonBaseInfo.h" -#include "llvm/Target/TargetFrameLowering.h" #include "llvm/Target/TargetInstrInfo.h" - +#include "llvm/Target/TargetFrameLowering.h" +#include "llvm/CodeGen/MachineBranchProbabilityInfo.h" #define GET_INSTRINFO_HEADER #include "HexagonGenInstrInfo.inc" @@ -28,6 +28,8 @@ namespace llvm { class HexagonInstrInfo : public HexagonGenInstrInfo { const HexagonRegisterInfo RI; const HexagonSubtarget& Subtarget; + typedef unsigned Opcode_t; + public: explicit HexagonInstrInfo(HexagonSubtarget &ST); @@ -127,6 +129,7 @@ public: const BranchProbability &Probability) const; virtual bool isPredicated(const MachineInstr *MI) const; + virtual bool isPredicatedNew(const MachineInstr *MI) const; virtual bool DefinesPredicate(MachineInstr *MI, std::vector<MachineOperand> &Pred) const; virtual bool @@ -140,6 +143,11 @@ public: isProfitableToDupForIfCvt(MachineBasicBlock &MBB,unsigned NumCycles, const BranchProbability &Probability) const; + virtual MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF, + int FrameIx, + uint64_t Offset, + const MDNode *MDPtr, + DebugLoc DL) const; virtual DFAPacketizer* CreateTargetScheduleState(const TargetMachine *TM, const ScheduleDAG *DAG) const; @@ -170,6 +178,7 @@ public: bool isConditionalLoad (const MachineInstr* MI) const; bool isConditionalStore(const MachineInstr* MI) const; bool isNewValueInst(const MachineInstr* MI) const; + bool isDotNewInst(const MachineInstr* MI) const; bool isDeallocRet(const MachineInstr *MI) const; unsigned getInvertedPredicatedOpcode(const int Opc) const; bool isExtendable(const MachineInstr* MI) const; @@ -182,6 +191,8 @@ public: void immediateExtend(MachineInstr *MI) const; bool isConstExtended(MachineInstr *MI) const; + int getDotNewPredJumpOp(MachineInstr *MI, + const MachineBranchProbabilityInfo *MBPI) const; unsigned getAddrMode(const MachineInstr* MI) const; bool isOperandExtended(const MachineInstr *MI, unsigned short OperandNum) const; @@ -190,6 +201,9 @@ public: int getMaxValue(const MachineInstr *MI) const; bool NonExtEquivalentExists (const MachineInstr *MI) const; short getNonExtOpcode(const MachineInstr *MI) const; + bool PredOpcodeHasJMP_c(Opcode_t Opcode) const; + bool PredOpcodeHasNot(Opcode_t Opcode) const; + private: int getMatchingCondBranchOpcode(int Opc, bool sense) const; diff --git a/lib/Target/Hexagon/HexagonInstrInfo.td b/lib/Target/Hexagon/HexagonInstrInfo.td index d7bab20..2a4b17b 100644 --- a/lib/Target/Hexagon/HexagonInstrInfo.td +++ b/lib/Target/Hexagon/HexagonInstrInfo.td @@ -14,6 +14,8 @@ include "HexagonInstrFormats.td" include "HexagonOperands.td" +//===----------------------------------------------------------------------===// + // Multi-class for logical operators. multiclass ALU32_rr_ri<string OpcStr, SDNode OpNode> { def rr : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c), @@ -34,12 +36,6 @@ multiclass CMP64_rr<string OpcStr, PatFrag OpNode> { [(set (i1 PredRegs:$dst), (OpNode (i64 DoubleRegs:$b), (i64 DoubleRegs:$c)))]>; } -multiclass CMP32_rr<string OpcStr, PatFrag OpNode> { - def rr : ALU32_rr<(outs PredRegs:$dst), (ins IntRegs:$b, IntRegs:$c), - !strconcat("$dst = ", !strconcat(OpcStr, "($b, $c)")), - [(set (i1 PredRegs:$dst), - (OpNode (i32 IntRegs:$b), (i32 IntRegs:$c)))]>; -} multiclass CMP32_rr_ri_s10<string OpcStr, string CextOp, PatFrag OpNode> { let CextOpcode = CextOp in { @@ -75,14 +71,6 @@ multiclass CMP32_rr_ri_u9<string OpcStr, string CextOp, PatFrag OpNode> { } } -multiclass CMP32_ri_u8<string OpcStr, PatFrag OpNode> { -let isExtendable = 1, opExtendable = 2, isExtentSigned = 0, opExtentBits = 8 in - def ri : ALU32_ri<(outs PredRegs:$dst), (ins IntRegs:$b, u8Ext:$c), - !strconcat("$dst = ", !strconcat(OpcStr, "($b, #$c)")), - [(set (i1 PredRegs:$dst), (OpNode (i32 IntRegs:$b), - u8ExtPred:$c))]>; -} - multiclass CMP32_ri_s8<string OpcStr, PatFrag OpNode> { let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 8 in def ri : ALU32_ri<(outs PredRegs:$dst), (ins IntRegs:$b, s8Ext:$c), @@ -95,22 +83,30 @@ let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 8 in //===----------------------------------------------------------------------===// // ALU32/ALU (Instructions with register-register form) //===----------------------------------------------------------------------===// -multiclass ALU32_Pbase<string mnemonic, bit isNot, - bit isPredNew> { +def SDTHexagonI64I32I32 : SDTypeProfile<1, 2, + [SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>; + +def HexagonWrapperCombineII : + SDNode<"HexagonISD::WrapperCombineII", SDTHexagonI64I32I32>; + +def HexagonWrapperCombineRR : + SDNode<"HexagonISD::WrapperCombineRR", SDTHexagonI64I32I32>; - let PNewValue = !if(isPredNew, "new", "") in - def NAME : ALU32_rr<(outs IntRegs:$dst), +multiclass ALU32_Pbase<string mnemonic, RegisterClass RC, bit isNot, + bit isPredNew> { + let isPredicatedNew = isPredNew in + def NAME : ALU32_rr<(outs RC:$dst), (ins PredRegs:$src1, IntRegs:$src2, IntRegs: $src3), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew,".new) $dst = ", ") $dst = ")#mnemonic#"($src2, $src3)", []>; } -multiclass ALU32_Pred<string mnemonic, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { - defm _c#NAME : ALU32_Pbase<mnemonic, PredNot, 0>; +multiclass ALU32_Pred<string mnemonic, RegisterClass RC, bit PredNot> { + let isPredicatedFalse = PredNot in { + defm _c#NAME : ALU32_Pbase<mnemonic, RC, PredNot, 0>; // Predicate new - defm _cdn#NAME : ALU32_Pbase<mnemonic, PredNot, 1>; + defm _cdn#NAME : ALU32_Pbase<mnemonic, RC, PredNot, 1>; } } @@ -125,8 +121,8 @@ multiclass ALU32_base<string mnemonic, string CextOp, SDNode OpNode> { (i32 IntRegs:$src2)))]>; let neverHasSideEffects = 1, isPredicated = 1 in { - defm Pt : ALU32_Pred<mnemonic, 0>; - defm NotPt : ALU32_Pred<mnemonic, 1>; + defm Pt : ALU32_Pred<mnemonic, IntRegs, 0>; + defm NotPt : ALU32_Pred<mnemonic, IntRegs, 1>; } } } @@ -140,11 +136,42 @@ let isCommutable = 1 in { defm SUB_rr : ALU32_base<"sub", "SUB", sub>, ImmRegRel, PredNewRel; +// Combines the two integer registers SRC1 and SRC2 into a double register. +let isPredicable = 1 in +class T_Combine : ALU32_rr<(outs DoubleRegs:$dst), + (ins IntRegs:$src1, IntRegs:$src2), + "$dst = combine($src1, $src2)", + [(set (i64 DoubleRegs:$dst), + (i64 (HexagonWrapperCombineRR (i32 IntRegs:$src1), + (i32 IntRegs:$src2))))]>; + +multiclass Combine_base { + let BaseOpcode = "combine" in { + def NAME : T_Combine; + let neverHasSideEffects = 1, isPredicated = 1 in { + defm Pt : ALU32_Pred<"combine", DoubleRegs, 0>; + defm NotPt : ALU32_Pred<"combine", DoubleRegs, 1>; + } + } +} + +defm COMBINE_rr : Combine_base, PredNewRel; + +// Combines the two immediates SRC1 and SRC2 into a double register. +class COMBINE_imm<Operand imm1, Operand imm2, PatLeaf pat1, PatLeaf pat2> : + ALU32_ii<(outs DoubleRegs:$dst), (ins imm1:$src1, imm2:$src2), + "$dst = combine(#$src1, #$src2)", + [(set (i64 DoubleRegs:$dst), + (i64 (HexagonWrapperCombineII (i32 pat1:$src1), (i32 pat2:$src2))))]>; + +let isExtendable = 1, opExtendable = 1, isExtentSigned = 1, opExtentBits = 8 in +def COMBINE_Ii : COMBINE_imm<s8Ext, s8Imm, s8ExtPred, s8ImmPred>; + //===----------------------------------------------------------------------===// // ALU32/ALU (ADD with register-immediate form) //===----------------------------------------------------------------------===// multiclass ALU32ri_Pbase<string mnemonic, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : ALU32_ri<(outs IntRegs:$dst), (ins PredRegs:$src1, IntRegs:$src2, s8Ext: $src3), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew,".new) $dst = ", @@ -153,7 +180,7 @@ multiclass ALU32ri_Pbase<string mnemonic, bit isNot, bit isPredNew> { } multiclass ALU32ri_Pred<string mnemonic, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ALU32ri_Pbase<mnemonic, PredNot, 0>; // Predicate new defm _cdn#NAME : ALU32ri_Pbase<mnemonic, PredNot, 1>; @@ -189,11 +216,6 @@ def OR_ri : ALU32_ri<(outs IntRegs:$dst), [(set (i32 IntRegs:$dst), (or (i32 IntRegs:$src1), s10ExtPred:$src2))]>, ImmRegRel; -def NOT_rr : ALU32_rr<(outs IntRegs:$dst), - (ins IntRegs:$src1), - "$dst = not($src1)", - [(set (i32 IntRegs:$dst), (not (i32 IntRegs:$src1)))]>; - let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 10, InputType = "imm", CextOpcode = "AND" in def AND_ri : ALU32_ri<(outs IntRegs:$dst), @@ -201,10 +223,7 @@ def AND_ri : ALU32_ri<(outs IntRegs:$dst), "$dst = and($src1, #$src2)", [(set (i32 IntRegs:$dst), (and (i32 IntRegs:$src1), s10ExtPred:$src2))]>, ImmRegRel; -// Negate. -def NEG : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = neg($src1)", - [(set (i32 IntRegs:$dst), (ineg (i32 IntRegs:$src1)))]>; + // Nop. let neverHasSideEffects = 1 in def NOP : ALU32_rr<(outs), (ins), @@ -220,15 +239,21 @@ def SUB_ri : ALU32_ri<(outs IntRegs:$dst), [(set IntRegs:$dst, (sub s10ExtPred:$src1, IntRegs:$src2))]>, ImmRegRel; +// Rd = not(Rs) gets mapped to Rd=sub(#-1, Rs). +def : Pat<(not (i32 IntRegs:$src1)), + (SUB_ri -1, (i32 IntRegs:$src1))>; + +// Rd = neg(Rs) gets mapped to Rd=sub(#0, Rs). +// Pattern definition for 'neg' was not necessary. multiclass TFR_Pred<bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { def _c#NAME : ALU32_rr<(outs IntRegs:$dst), (ins PredRegs:$src1, IntRegs:$src2), !if(PredNot, "if (!$src1", "if ($src1")#") $dst = $src2", []>; // Predicate new - let PNewValue = "new" in + let isPredicatedNew = 1 in def _cdn#NAME : ALU32_rr<(outs IntRegs:$dst), (ins PredRegs:$src1, IntRegs:$src2), !if(PredNot, "if (!$src1", "if ($src1")#".new) $dst = $src2", @@ -274,10 +299,10 @@ class T_TFR64_Pred<bit PredNot, bit isPredNew> } multiclass TFR64_Pred<bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { def _c#NAME : T_TFR64_Pred<PredNot, 0>; - let PNewValue = "new" in + let isPredicatedNew = 1 in def _cdn#NAME : T_TFR64_Pred<PredNot, 1>; // Predicate new } } @@ -309,14 +334,14 @@ multiclass TFR64_base<string BaseName> { } multiclass TFRI_Pred<bit PredNot> { - let isMoveImm = 1, PredSense = !if(PredNot, "false", "true") in { + let isMoveImm = 1, isPredicatedFalse = PredNot in { def _c#NAME : ALU32_ri<(outs IntRegs:$dst), (ins PredRegs:$src1, s12Ext:$src2), !if(PredNot, "if (!$src1", "if ($src1")#") $dst = #$src2", []>; // Predicate new - let PNewValue = "new" in + let isPredicatedNew = 1 in def _cdn#NAME : ALU32_rr<(outs IntRegs:$dst), (ins PredRegs:$src1, s12Ext:$src2), !if(PredNot, "if (!$src1", "if ($src1")#".new) $dst = #$src2", @@ -359,52 +384,6 @@ def TFCR : CRInst<(outs CRRegs:$dst), (ins IntRegs:$src1), // ALU32/PERM + //===----------------------------------------------------------------------===// -// Combine. - -def SDTHexagonI64I32I32 : SDTypeProfile<1, 2, - [SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>; - -def HexagonWrapperCombineII : - SDNode<"HexagonISD::WrapperCombineII", SDTHexagonI64I32I32>; -def HexagonWrapperCombineRR : - SDNode<"HexagonISD::WrapperCombineRR", SDTHexagonI64I32I32>; - -// Combines the two integer registers SRC1 and SRC2 into a double register. -let isPredicable = 1 in -def COMBINE_rr : ALU32_rr<(outs DoubleRegs:$dst), (ins IntRegs:$src1, - IntRegs:$src2), - "$dst = combine($src1, $src2)", - [(set (i64 DoubleRegs:$dst), - (i64 (HexagonWrapperCombineRR (i32 IntRegs:$src1), - (i32 IntRegs:$src2))))]>; - -// Rd=combine(Rt.[HL], Rs.[HL]) -class COMBINE_halves<string A, string B>: ALU32_rr<(outs IntRegs:$dst), - (ins IntRegs:$src1, - IntRegs:$src2), - "$dst = combine($src1."# A #", $src2."# B #")", []>; - -let isPredicable = 1 in { - def COMBINE_hh : COMBINE_halves<"H", "H">; - def COMBINE_hl : COMBINE_halves<"H", "L">; - def COMBINE_lh : COMBINE_halves<"L", "H">; - def COMBINE_ll : COMBINE_halves<"L", "L">; -} - -def : Pat<(i32 (trunc (i64 (srl (i64 DoubleRegs:$a), (i32 16))))), - (COMBINE_lh (EXTRACT_SUBREG (i64 DoubleRegs:$a), subreg_hireg), - (EXTRACT_SUBREG (i64 DoubleRegs:$a), subreg_loreg))>; - -// Combines the two immediates SRC1 and SRC2 into a double register. -class COMBINE_imm<Operand imm1, Operand imm2, PatLeaf pat1, PatLeaf pat2> : - ALU32_ii<(outs DoubleRegs:$dst), (ins imm1:$src1, imm2:$src2), - "$dst = combine(#$src1, #$src2)", - [(set (i64 DoubleRegs:$dst), - (i64 (HexagonWrapperCombineII (i32 pat1:$src1), (i32 pat2:$src2))))]>; - -let isExtendable = 1, opExtendable = 1, isExtentSigned = 1, opExtentBits = 8 in -def COMBINE_Ii : COMBINE_imm<s8Ext, s8Imm, s8ExtPred, s8ImmPred>; - // Mux. def VMUX_prr64 : ALU64_rr<(outs DoubleRegs:$dst), (ins PredRegs:$src1, DoubleRegs:$src2, @@ -446,38 +425,58 @@ def MUX_ii : ALU32_ii<(outs IntRegs:$dst), (ins PredRegs:$src1, s8Ext:$src2, s8ExtPred:$src2, s8ImmPred:$src3)))]>; -// Shift halfword. -let isPredicable = 1 in -def ASLH : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = aslh($src1)", - [(set (i32 IntRegs:$dst), (shl 16, (i32 IntRegs:$src1)))]>; +// ALU32 - aslh, asrh, sxtb, sxth, zxtb, zxth +multiclass ALU32_2op_Pbase<string mnemonic, bit isNot, bit isPredNew> { + let isPredicatedNew = isPredNew in + def NAME : ALU32Inst<(outs IntRegs:$dst), + (ins PredRegs:$src1, IntRegs:$src2), + !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew,".new) $dst = ", + ") $dst = ")#mnemonic#"($src2)">, + Requires<[HasV4T]>; +} -let isPredicable = 1 in -def ASRH : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = asrh($src1)", - [(set (i32 IntRegs:$dst), (sra 16, (i32 IntRegs:$src1)))]>; +multiclass ALU32_2op_Pred<string mnemonic, bit PredNot> { + let isPredicatedFalse = PredNot in { + defm _c#NAME : ALU32_2op_Pbase<mnemonic, PredNot, 0>; + // Predicate new + defm _cdn#NAME : ALU32_2op_Pbase<mnemonic, PredNot, 1>; + } +} -// Sign extend. -let isPredicable = 1 in -def SXTB : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = sxtb($src1)", - [(set (i32 IntRegs:$dst), (sext_inreg (i32 IntRegs:$src1), i8))]>; +multiclass ALU32_2op_base<string mnemonic> { + let BaseOpcode = mnemonic in { + let isPredicable = 1, neverHasSideEffects = 1 in + def NAME : ALU32Inst<(outs IntRegs:$dst), + (ins IntRegs:$src1), + "$dst = "#mnemonic#"($src1)">; -let isPredicable = 1 in -def SXTH : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = sxth($src1)", - [(set (i32 IntRegs:$dst), (sext_inreg (i32 IntRegs:$src1), i16))]>; - -// Zero extend. -let isPredicable = 1, neverHasSideEffects = 1 in -def ZXTB : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = zxtb($src1)", - []>; + let Predicates = [HasV4T], validSubTargets = HasV4SubT, isPredicated = 1, + neverHasSideEffects = 1 in { + defm Pt_V4 : ALU32_2op_Pred<mnemonic, 0>; + defm NotPt_V4 : ALU32_2op_Pred<mnemonic, 1>; + } + } +} + +defm ASLH : ALU32_2op_base<"aslh">, PredNewRel; +defm ASRH : ALU32_2op_base<"asrh">, PredNewRel; +defm SXTB : ALU32_2op_base<"sxtb">, PredNewRel; +defm SXTH : ALU32_2op_base<"sxth">, PredNewRel; +defm ZXTB : ALU32_2op_base<"zxtb">, PredNewRel; +defm ZXTH : ALU32_2op_base<"zxth">, PredNewRel; + +def : Pat <(shl (i32 IntRegs:$src1), (i32 16)), + (ASLH IntRegs:$src1)>; + +def : Pat <(sra (i32 IntRegs:$src1), (i32 16)), + (ASRH IntRegs:$src1)>; + +def : Pat <(sext_inreg (i32 IntRegs:$src1), i8), + (SXTB IntRegs:$src1)>; + +def : Pat <(sext_inreg (i32 IntRegs:$src1), i16), + (SXTH IntRegs:$src1)>; -let isPredicable = 1, neverHasSideEffects = 1 in -def ZXTH : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), - "$dst = zxth($src1)", - []>; //===----------------------------------------------------------------------===// // ALU32/PERM - //===----------------------------------------------------------------------===// @@ -487,39 +486,24 @@ def ZXTH : ALU32_rr<(outs IntRegs:$dst), (ins IntRegs:$src1), // ALU32/PRED + //===----------------------------------------------------------------------===// -// Conditional combine. -let neverHasSideEffects = 1, isPredicated = 1 in -def COMBINE_rr_cPt : ALU32_rr<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3), - "if ($src1) $dst = combine($src2, $src3)", - []>; - -let neverHasSideEffects = 1, isPredicated = 1 in -def COMBINE_rr_cNotPt : ALU32_rr<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3), - "if (!$src1) $dst = combine($src2, $src3)", - []>; - -let neverHasSideEffects = 1, isPredicated = 1 in -def COMBINE_rr_cdnPt : ALU32_rr<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3), - "if ($src1.new) $dst = combine($src2, $src3)", - []>; - -let neverHasSideEffects = 1, isPredicated = 1 in -def COMBINE_rr_cdnNotPt : ALU32_rr<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3), - "if (!$src1.new) $dst = combine($src2, $src3)", - []>; - // Compare. defm CMPGTU : CMP32_rr_ri_u9<"cmp.gtu", "CMPGTU", setugt>, ImmRegRel; defm CMPGT : CMP32_rr_ri_s10<"cmp.gt", "CMPGT", setgt>, ImmRegRel; -defm CMPLT : CMP32_rr<"cmp.lt", setlt>; -defm CMPLTU : CMP32_rr<"cmp.ltu", setult>; defm CMPEQ : CMP32_rr_ri_s10<"cmp.eq", "CMPEQ", seteq>, ImmRegRel; -defm CMPGE : CMP32_ri_s8<"cmp.ge", setge>; -defm CMPGEU : CMP32_ri_u8<"cmp.geu", setuge>; + +// SDNode for converting immediate C to C-1. +def DEC_CONST_SIGNED : SDNodeXForm<imm, [{ + // Return the byte immediate const-1 as an SDNode. + int32_t imm = N->getSExtValue(); + return XformSToSM1Imm(imm); +}]>; + +// SDNode for converting immediate C to C-1. +def DEC_CONST_UNSIGNED : SDNodeXForm<imm, [{ + // Return the byte immediate const-1 as an SDNode. + uint32_t imm = N->getZExtValue(); + return XformUToUM1Imm(imm); +}]>; def CTLZ_rr : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1), "$dst = cl0($src1)", @@ -753,112 +737,153 @@ def XOR_pp : SInst<(outs PredRegs:$dst), (ins PredRegs:$src1, PredRegs:$src2), // CR - //===----------------------------------------------------------------------===// +def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone, + [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>; +def eh_return: SDNode<"HexagonISD::EH_RETURN", SDTNone, + [SDNPHasChain]>; -//===----------------------------------------------------------------------===// -// J + -//===----------------------------------------------------------------------===// -// Jump to address. -let isBranch = 1, isTerminator=1, isBarrier = 1, isPredicable = 1 in { - def JMP : JInst< (outs), - (ins brtarget:$offset), - "jump $offset", - [(br bb:$offset)]>; -} +def SDHexagonBR_JT: SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>; +def HexagonBR_JT: SDNode<"HexagonISD::BR_JT", SDHexagonBR_JT, [SDNPHasChain]>; -// if (p0) jump -let isBranch = 1, isTerminator=1, Defs = [PC], - isPredicated = 1 in { - def JMP_c : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if ($src) jump $offset", - [(brcond (i1 PredRegs:$src), bb:$offset)]>; -} +let InputType = "imm", isBarrier = 1, isPredicable = 1, +Defs = [PC], isExtendable = 1, opExtendable = 0, isExtentSigned = 1, +opExtentBits = 24 in +class T_JMP <dag InsDag, list<dag> JumpList = []> + : JInst<(outs), InsDag, + "jump $dst" , JumpList> { + bits<24> dst; + + let IClass = 0b0101; + + let Inst{27-25} = 0b100; + let Inst{24-16} = dst{23-15}; + let Inst{13-1} = dst{14-2}; +} + +let InputType = "imm", isExtendable = 1, opExtendable = 1, isExtentSigned = 1, +Defs = [PC], isPredicated = 1, opExtentBits = 17 in +class T_JMP_c <bit PredNot, bit isPredNew, bit isTaken>: + JInst<(outs ), (ins PredRegs:$src, brtarget:$dst), + !if(PredNot, "if (!$src", "if ($src")# + !if(isPredNew, ".new) ", ") ")#"jump"# + !if(isPredNew, !if(isTaken, ":t ", ":nt "), " ")#"$dst"> { + + let isBrTaken = !if(isPredNew, !if(isTaken, "true", "false"), ""); + let isPredicatedFalse = PredNot; + let isPredicatedNew = isPredNew; + bits<2> src; + bits<17> dst; + + let IClass = 0b0101; + + let Inst{27-24} = 0b1100; + let Inst{21} = PredNot; + let Inst{12} = !if(isPredNew, isTaken, zero); + let Inst{11} = isPredNew; + let Inst{9-8} = src; + let Inst{23-22} = dst{16-15}; + let Inst{20-16} = dst{14-10}; + let Inst{13} = dst{9}; + let Inst{7-1} = dst{8-2}; + } -// if (!p0) jump -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def JMP_cNot : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if (!$src) jump $offset", - []>; +let isBarrier = 1, Defs = [PC], isPredicable = 1, InputType = "reg" in +class T_JMPr<dag InsDag = (ins IntRegs:$dst)> + : JRInst<(outs ), InsDag, + "jumpr $dst" , + []> { + bits<5> dst; + + let IClass = 0b0101; + let Inst{27-21} = 0b0010100; + let Inst{20-16} = dst; } -let isTerminator = 1, isBranch = 1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def BRCOND : JInst < (outs), (ins PredRegs:$pred, brtarget:$dst), - "if ($pred) jump $dst", - []>; +let Defs = [PC], isPredicated = 1, InputType = "reg" in +class T_JMPr_c <bit PredNot, bit isPredNew, bit isTaken>: + JRInst <(outs ), (ins PredRegs:$src, IntRegs:$dst), + !if(PredNot, "if (!$src", "if ($src")# + !if(isPredNew, ".new) ", ") ")#"jumpr"# + !if(isPredNew, !if(isTaken, ":t ", ":nt "), " ")#"$dst"> { + + let isBrTaken = !if(isPredNew, !if(isTaken, "true", "false"), ""); + let isPredicatedFalse = PredNot; + let isPredicatedNew = isPredNew; + bits<2> src; + bits<5> dst; + + let IClass = 0b0101; + + let Inst{27-22} = 0b001101; + let Inst{21} = PredNot; + let Inst{20-16} = dst; + let Inst{12} = !if(isPredNew, isTaken, zero); + let Inst{11} = isPredNew; + let Inst{9-8} = src; + let Predicates = !if(isPredNew, [HasV3T], [HasV2T]); + let validSubTargets = !if(isPredNew, HasV3SubT, HasV2SubT); } -// Jump to address conditioned on new predicate. -// if (p0) jump:t -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def JMP_cdnPt : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if ($src.new) jump:t $offset", - []>; +multiclass JMP_Pred<bit PredNot> { + def _#NAME : T_JMP_c<PredNot, 0, 0>; + // Predicate new + def _#NAME#new_t : T_JMP_c<PredNot, 1, 1>; // taken + def _#NAME#new_nt : T_JMP_c<PredNot, 1, 0>; // not taken } -// if (!p0) jump:t -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def JMP_cdnNotPt : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if (!$src.new) jump:t $offset", - []>; +multiclass JMP_base<string BaseOp> { + let BaseOpcode = BaseOp in { + def NAME : T_JMP<(ins brtarget:$dst), [(br bb:$dst)]>; + defm t : JMP_Pred<0>; + defm f : JMP_Pred<1>; + } } -// Not taken. -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def JMP_cdnPnt : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if ($src.new) jump:nt $offset", - []>; +multiclass JMPR_Pred<bit PredNot> { + def NAME: T_JMPr_c<PredNot, 0, 0>; + // Predicate new + def NAME#new_tV3 : T_JMPr_c<PredNot, 1, 1>; // taken + def NAME#new_ntV3 : T_JMPr_c<PredNot, 1, 0>; // not taken } -// Not taken. -let isBranch = 1, isTerminator=1, neverHasSideEffects = 1, Defs = [PC], - isPredicated = 1 in { - def JMP_cdnNotPnt : JInst< (outs), - (ins PredRegs:$src, brtarget:$offset), - "if (!$src.new) jump:nt $offset", - []>; +multiclass JMPR_base<string BaseOp> { + let BaseOpcode = BaseOp in { + def NAME : T_JMPr; + defm _t : JMPR_Pred<0>; + defm _f : JMPR_Pred<1>; + } } -//===----------------------------------------------------------------------===// -// J - -//===----------------------------------------------------------------------===// -//===----------------------------------------------------------------------===// -// JR + -//===----------------------------------------------------------------------===// -def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone, - [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>; +let isTerminator = 1, neverHasSideEffects = 1 in { +let isBranch = 1 in +defm JMP : JMP_base<"JMP">, PredNewRel; -// Jump to address from register. -let isPredicable =1, isReturn = 1, isTerminator = 1, isBarrier = 1, - Defs = [PC], Uses = [R31] in { - def JMPR: JRInst<(outs), (ins), - "jumpr r31", - [(retflag)]>; -} +let isBranch = 1, isIndirectBranch = 1 in +defm JMPR : JMPR_base<"JMPr">, PredNewRel; -// Jump to address from register. -let isReturn = 1, isTerminator = 1, isBarrier = 1, isPredicated = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cPt: JRInst<(outs), (ins PredRegs:$src1), - "if ($src1) jumpr r31", - []>; +let isReturn = 1, isCodeGenOnly = 1 in +defm JMPret : JMPR_base<"JMPret">, PredNewRel; } -// Jump to address from register. -let isReturn = 1, isTerminator = 1, isBarrier = 1, isPredicated = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cNotPt: JRInst<(outs), (ins PredRegs:$src1), - "if (!$src1) jumpr r31", - []>; -} +def : Pat<(retflag), + (JMPret (i32 R31))>; + +def : Pat <(brcond (i1 PredRegs:$src1), bb:$offset), + (JMP_t (i1 PredRegs:$src1), bb:$offset)>; + +// A return through builtin_eh_return. +let isReturn = 1, isTerminator = 1, isBarrier = 1, neverHasSideEffects = 1, +isCodeGenOnly = 1, Defs = [PC], Uses = [R28], isPredicable = 0 in +def EH_RETURN_JMPR : T_JMPr; + +def : Pat<(eh_return), + (EH_RETURN_JMPR (i32 R31))>; + +def : Pat<(HexagonBR_JT (i32 IntRegs:$dst)), + (JMPR (i32 IntRegs:$dst))>; + +def : Pat<(brind (i32 IntRegs:$dst)), + (JMPR (i32 IntRegs:$dst))>; //===----------------------------------------------------------------------===// // JR - @@ -871,7 +896,7 @@ let isReturn = 1, isTerminator = 1, isBarrier = 1, isPredicated = 1, // Load -- MEMri operand multiclass LD_MEMri_Pbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : LDInst2<(outs RC:$dst), (ins PredRegs:$src1, MEMri:$addr), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -880,7 +905,7 @@ multiclass LD_MEMri_Pbase<string mnemonic, RegisterClass RC, } multiclass LD_MEMri_Pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : LD_MEMri_Pbase<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : LD_MEMri_Pbase<mnemonic, RC, PredNot, 1>; @@ -937,7 +962,7 @@ def : Pat < (i64 (load ADDRriS11_3:$addr)), // Load - Base with Immediate offset addressing mode multiclass LD_Idxd_Pbase<string mnemonic, RegisterClass RC, Operand predImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : LDInst2<(outs RC:$dst), (ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -947,7 +972,7 @@ multiclass LD_Idxd_Pbase<string mnemonic, RegisterClass RC, Operand predImmOp, multiclass LD_Idxd_Pred<string mnemonic, RegisterClass RC, Operand predImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : LD_Idxd_Pbase<mnemonic, RC, predImmOp, PredNot, 0>; // Predicate new defm _cdn#NAME : LD_Idxd_Pbase<mnemonic, RC, predImmOp, PredNot, 1>; @@ -1009,20 +1034,6 @@ def : Pat < (i64 (load (add IntRegs:$src1, s11_3ExtPred:$offset))), (LDrid_indexed IntRegs:$src1, s11_3ExtPred:$offset) >; } -let neverHasSideEffects = 1 in -def LDrid_GP : LDInst2<(outs DoubleRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memd(#$global+$offset)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDd_GP : LDInst2<(outs DoubleRegs:$dst), - (ins globaladdress:$global), - "$dst = memd(#$global)", - []>, - Requires<[NoV4T]>; - //===----------------------------------------------------------------------===// // Post increment load // Make sure that in post increment load, the first operand is always the post @@ -1031,7 +1042,7 @@ def LDd_GP : LDInst2<(outs DoubleRegs:$dst), multiclass LD_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : LDInst2PI<(outs RC:$dst, IntRegs:$dst2), (ins PredRegs:$src1, IntRegs:$src2, ImmOp:$offset), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1042,7 +1053,7 @@ multiclass LD_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp, multiclass LD_PostInc_Pred<string mnemonic, RegisterClass RC, Operand ImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : LD_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 0>; // Predicate new let Predicates = [HasV4T], validSubTargets = HasV4SubT in @@ -1095,27 +1106,6 @@ let AddedComplexity = 20 in def : Pat < (i32 (extloadi8 (add IntRegs:$src1, s11_0ImmPred:$offset))), (i32 (LDrib_indexed IntRegs:$src1, s11_0ImmPred:$offset)) >; -let neverHasSideEffects = 1 in -def LDrib_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memb(#$global+$offset)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDb_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst = memb(#$global)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDub_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst = memub(#$global)", - []>, - Requires<[NoV4T]>; - def : Pat < (i32 (extloadi16 ADDRriS11_1:$addr)), (i32 (LDrih ADDRriS11_1:$addr))>; @@ -1123,27 +1113,6 @@ let AddedComplexity = 20 in def : Pat < (i32 (extloadi16 (add IntRegs:$src1, s11_1ImmPred:$offset))), (i32 (LDrih_indexed IntRegs:$src1, s11_1ImmPred:$offset)) >; -let neverHasSideEffects = 1 in -def LDrih_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memh(#$global+$offset)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDh_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst = memh(#$global)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDuh_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst = memuh(#$global)", - []>, - Requires<[NoV4T]>; - let AddedComplexity = 10 in def : Pat < (i32 (zextloadi1 ADDRriS11_0:$addr)), (i32 (LDriub ADDRriS11_0:$addr))>; @@ -1152,21 +1121,6 @@ let AddedComplexity = 20 in def : Pat < (i32 (zextloadi1 (add IntRegs:$src1, s11_0ImmPred:$offset))), (i32 (LDriub_indexed IntRegs:$src1, s11_0ImmPred:$offset))>; -let neverHasSideEffects = 1 in -def LDriub_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memub(#$global+$offset)", - []>, - Requires<[NoV4T]>; - -// Load unsigned halfword. -let neverHasSideEffects = 1 in -def LDriuh_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memuh(#$global+$offset)", - []>, - Requires<[NoV4T]>; - // Load predicate. let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 13, isPseudo = 1, Defs = [R10,R11,D5], neverHasSideEffects = 1 in @@ -1175,21 +1129,6 @@ def LDriw_pred : LDInst2<(outs PredRegs:$dst), "Error; should not emit", []>; -// Indexed load. -let neverHasSideEffects = 1 in -def LDriw_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global, u16Imm:$offset), - "$dst = memw(#$global+$offset)", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def LDw_GP : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst = memw(#$global)", - []>, - Requires<[NoV4T]>; - // Deallocate stack frame. let Defs = [R29, R30, R31], Uses = [R29], neverHasSideEffects = 1 in { def DEALLOCFRAME : LDInst2<(outs), (ins), @@ -1423,35 +1362,15 @@ def SUBri_acc : MInst_acc<(outs IntRegs: $dst), (ins IntRegs:$src1, // ST + //===----------------------------------------------------------------------===// /// -/// Assumptions::: ****** DO NOT IGNORE ******** -/// 1. Make sure that in post increment store, the zero'th operand is always the -/// post increment operand. -/// 2. Make sure that the store value operand(Rt/Rtt) in a store is always the -/// last operand. -/// // Store doubleword. -let neverHasSideEffects = 1 in -def STrid_GP : STInst2<(outs), - (ins globaladdress:$global, u16Imm:$offset, DoubleRegs:$src), - "memd(#$global+$offset) = $src", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def STd_GP : STInst2<(outs), - (ins globaladdress:$global, DoubleRegs:$src), - "memd(#$global) = $src", - []>, - Requires<[NoV4T]>; - //===----------------------------------------------------------------------===// // Post increment store //===----------------------------------------------------------------------===// multiclass ST_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : STInst2PI<(outs IntRegs:$dst), (ins PredRegs:$src1, IntRegs:$src2, ImmOp:$offset, RC:$src3), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1462,7 +1381,7 @@ multiclass ST_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp, multiclass ST_PostInc_Pred<string mnemonic, RegisterClass RC, Operand ImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME# : ST_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 0>; // Predicate new let Predicates = [HasV4T], validSubTargets = HasV4SubT in @@ -1516,7 +1435,7 @@ def : Pat<(post_store (i64 DoubleRegs:$src1), IntRegs:$src2, //===----------------------------------------------------------------------===// multiclass ST_MEMri_Pbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : STInst2<(outs), (ins PredRegs:$src1, MEMri:$addr, RC: $src2), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1525,7 +1444,7 @@ multiclass ST_MEMri_Pbase<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_MEMri_Pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_MEMri_Pbase<mnemonic, RC, PredNot, 0>; // Predicate new @@ -1582,7 +1501,7 @@ def : Pat<(store (i64 DoubleRegs:$src1), ADDRriS11_3:$addr), //===----------------------------------------------------------------------===// multiclass ST_Idxd_Pbase<string mnemonic, RegisterClass RC, Operand predImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : STInst2<(outs), (ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3, RC: $src4), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1592,7 +1511,7 @@ multiclass ST_Idxd_Pbase<string mnemonic, RegisterClass RC, Operand predImmOp, multiclass ST_Idxd_Pred<string mnemonic, RegisterClass RC, Operand predImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true"), isPredicated = 1 in { + let isPredicatedFalse = PredNot, isPredicated = 1 in { defm _c#NAME : ST_Idxd_Pbase<mnemonic, RC, predImmOp, PredNot, 0>; // Predicate new @@ -1655,36 +1574,6 @@ def : Pat<(store (i64 DoubleRegs:$src1), (add IntRegs:$src2, (i64 DoubleRegs:$src1))>; } -// memb(gp+#u16:0)=Rt -let neverHasSideEffects = 1 in -def STrib_GP : STInst2<(outs), - (ins globaladdress:$global, u16Imm:$offset, IntRegs:$src), - "memb(#$global+$offset) = $src", - []>, - Requires<[NoV4T]>; - -// memb(#global)=Rt -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def STb_GP : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memb(#$global) = $src", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1 in -def STrih_GP : STInst2<(outs), - (ins globaladdress:$global, u16Imm:$offset, IntRegs:$src), - "memh(#$global+$offset) = $src", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def STh_GP : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memh(#$global) = $src", - []>, - Requires<[NoV4T]>; - // memh(Rx++#s4:1)=Rt.H // Store word. @@ -1695,20 +1584,6 @@ def STriw_pred : STInst2<(outs), "Error; should not emit", []>; -let neverHasSideEffects = 1 in -def STriw_GP : STInst2<(outs), - (ins globaladdress:$global, u16Imm:$offset, IntRegs:$src), - "memw(#$global+$offset) = $src", - []>, - Requires<[NoV4T]>; - -let neverHasSideEffects = 1, validSubTargets = NoV4SubT in -def STw_GP : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memw(#$global) = $src", - []>, - Requires<[NoV4T]>; - // Allocate stack frame. let Defs = [R29, R30], Uses = [R31, R30], neverHasSideEffects = 1 in { def ALLOCFRAME : STInst2<(outs), @@ -2152,20 +2027,18 @@ let isCall = 1, neverHasSideEffects = 1, []>; } -// Tail Calls. -let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1 in { - def TCRETURNtg : JInst<(outs), (ins calltarget:$dst), - "jump $dst // TAILCALL", []>; -} -let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1 in { - def TCRETURNtext : JInst<(outs), (ins calltarget:$dst), - "jump $dst // TAILCALL", []>; -} -let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1 in { - def TCRETURNR : JInst<(outs), (ins IntRegs:$dst), - "jumpr $dst // TAILCALL", []>; +// Indirect tail-call. +let isCodeGenOnly = 1, isCall = 1, isReturn = 1 in +def TCRETURNR : T_JMPr; + +// Direct tail-calls. +let isCall = 1, isReturn = 1, isBarrier = 1, isPredicable = 0, +isTerminator = 1, isCodeGenOnly = 1 in { + def TCRETURNtg : T_JMP<(ins calltarget:$dst)>; + def TCRETURNtext : T_JMP<(ins calltarget:$dst)>; } + // Map call instruction. def : Pat<(call (i32 IntRegs:$dst)), (CALLR (i32 IntRegs:$dst))>, Requires<[HasV2TOnly]>; @@ -2183,68 +2056,26 @@ def : Pat<(HexagonTCRet (i32 IntRegs:$dst)), // Atomic load and store support // 8 bit atomic load -def : Pat<(atomic_load_8 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDub_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_load_8 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (i32 (LDriub_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - def : Pat<(atomic_load_8 ADDRriS11_0:$src1), (i32 (LDriub ADDRriS11_0:$src1))>; def : Pat<(atomic_load_8 (add (i32 IntRegs:$src1), s11_0ImmPred:$offset)), (i32 (LDriub_indexed (i32 IntRegs:$src1), s11_0ImmPred:$offset))>; - - // 16 bit atomic load -def : Pat<(atomic_load_16 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDuh_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_load_16 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (i32 (LDriuh_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - def : Pat<(atomic_load_16 ADDRriS11_1:$src1), (i32 (LDriuh ADDRriS11_1:$src1))>; def : Pat<(atomic_load_16 (add (i32 IntRegs:$src1), s11_1ImmPred:$offset)), (i32 (LDriuh_indexed (i32 IntRegs:$src1), s11_1ImmPred:$offset))>; - - -// 32 bit atomic load -def : Pat<(atomic_load_32 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDw_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_load_32 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (i32 (LDriw_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - def : Pat<(atomic_load_32 ADDRriS11_2:$src1), (i32 (LDriw ADDRriS11_2:$src1))>; def : Pat<(atomic_load_32 (add (i32 IntRegs:$src1), s11_2ImmPred:$offset)), (i32 (LDriw_indexed (i32 IntRegs:$src1), s11_2ImmPred:$offset))>; - // 64 bit atomic load -def : Pat<(atomic_load_64 (HexagonCONST32_GP tglobaladdr:$global)), - (i64 (LDd_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_load_64 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (i64 (LDrid_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - def : Pat<(atomic_load_64 ADDRriS11_3:$src1), (i64 (LDrid ADDRriS11_3:$src1))>; @@ -2252,30 +2083,6 @@ def : Pat<(atomic_load_64 (add (i32 IntRegs:$src1), s11_3ImmPred:$offset)), (i64 (LDrid_indexed (i32 IntRegs:$src1), s11_3ImmPred:$offset))>; -// 64 bit atomic store -def : Pat<(atomic_store_64 (HexagonCONST32_GP tglobaladdr:$global), - (i64 DoubleRegs:$src1)), - (STd_GP tglobaladdr:$global, (i64 DoubleRegs:$src1))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_store_64 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset), - (i64 DoubleRegs:$src1)), - (STrid_GP tglobaladdr:$global, u16ImmPred:$offset, - (i64 DoubleRegs:$src1))>, Requires<[NoV4T]>; - -// 8 bit atomic store -def : Pat<(atomic_store_8 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STb_GP tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_store_8 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset), - (i32 IntRegs:$src1)), - (STrib_GP tglobaladdr:$global, u16ImmPred:$offset, - (i32 IntRegs:$src1))>, Requires<[NoV4T]>; - def : Pat<(atomic_store_8 ADDRriS11_0:$src2, (i32 IntRegs:$src1)), (STrib ADDRriS11_0:$src2, (i32 IntRegs:$src1))>; @@ -2285,18 +2092,6 @@ def : Pat<(atomic_store_8 (add (i32 IntRegs:$src2), s11_0ImmPred:$offset), (i32 IntRegs:$src1))>; -// 16 bit atomic store -def : Pat<(atomic_store_16 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STh_GP tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_store_16 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset), - (i32 IntRegs:$src1)), - (STrih_GP tglobaladdr:$global, u16ImmPred:$offset, - (i32 IntRegs:$src1))>, Requires<[NoV4T]>; - def : Pat<(atomic_store_16 ADDRriS11_1:$src2, (i32 IntRegs:$src1)), (STrih ADDRriS11_1:$src2, (i32 IntRegs:$src1))>; @@ -2305,20 +2100,6 @@ def : Pat<(atomic_store_16 (i32 IntRegs:$src1), (STrih_indexed (i32 IntRegs:$src2), s11_1ImmPred:$offset, (i32 IntRegs:$src1))>; - -// 32 bit atomic store -def : Pat<(atomic_store_32 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STw_GP tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -def : Pat<(atomic_store_32 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset), - (i32 IntRegs:$src1)), - (STriw_GP tglobaladdr:$global, u16ImmPred:$offset, - (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - def : Pat<(atomic_store_32 ADDRriS11_2:$src2, (i32 IntRegs:$src1)), (STriw ADDRriS11_2:$src2, (i32 IntRegs:$src1))>; @@ -2354,10 +2135,11 @@ def : Pat <(add (i1 PredRegs:$src1), -1), // Map from p0 = setlt(r0, r1) r2 = mux(p0, r3, r4) => // p0 = cmp.lt(r0, r1), r0 = mux(p0, r2, r1). +// cmp.lt(r0, r1) -> cmp.gt(r1, r0) def : Pat <(select (i1 (setlt (i32 IntRegs:$src1), (i32 IntRegs:$src2))), (i32 IntRegs:$src3), (i32 IntRegs:$src4)), - (i32 (TFR_condset_rr (CMPLTrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), + (i32 (TFR_condset_rr (CMPGTrr (i32 IntRegs:$src2), (i32 IntRegs:$src1)), (i32 IntRegs:$src4), (i32 IntRegs:$src3)))>, Requires<[HasV2TOnly]>; @@ -2375,210 +2157,27 @@ def : Pat <(select (not (i1 PredRegs:$src1)), s12ImmPred:$src2, // Map from p0 = pnot(p0); r0 = mux(p0, r1, #i) // => r0 = TFR_condset_ir(p0, #i, r1) -def : Pat <(select (not PredRegs:$src1), IntRegs:$src2, s12ImmPred:$src3), +def : Pat <(select (not (i1 PredRegs:$src1)), IntRegs:$src2, s12ImmPred:$src3), (i32 (TFR_condset_ir (i1 PredRegs:$src1), s12ImmPred:$src3, (i32 IntRegs:$src2)))>; // Map from p0 = pnot(p0); if (p0) jump => if (!p0) jump. -def : Pat <(brcond (not PredRegs:$src1), bb:$offset), - (JMP_cNot (i1 PredRegs:$src1), bb:$offset)>; +def : Pat <(brcond (not (i1 PredRegs:$src1)), bb:$offset), + (JMP_f (i1 PredRegs:$src1), bb:$offset)>; // Map from p2 = pnot(p2); p1 = and(p0, p2) => p1 = and(p0, !p2). -def : Pat <(and PredRegs:$src1, (not PredRegs:$src2)), +def : Pat <(and (i1 PredRegs:$src1), (not (i1 PredRegs:$src2))), (i1 (AND_pnotp (i1 PredRegs:$src1), (i1 PredRegs:$src2)))>; -// Map from store(globaladdress + x) -> memd(#foo + x). -let AddedComplexity = 100 in -def : Pat <(store (i64 DoubleRegs:$src1), - (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (STrid_GP tglobaladdr:$global, u16ImmPred:$offset, - (i64 DoubleRegs:$src1))>, Requires<[NoV4T]>; - -// Map from store(globaladdress) -> memd(#foo). -let AddedComplexity = 100 in -def : Pat <(store (i64 DoubleRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STd_GP tglobaladdr:$global, (i64 DoubleRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress + x) -> memw(#foo + x). -let AddedComplexity = 100 in -def : Pat <(store (i32 IntRegs:$src1), - (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (STriw_GP tglobaladdr:$global, u16ImmPred:$offset, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress) -> memw(#foo + 0). -let AddedComplexity = 100 in -def : Pat <(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), - (STriw_GP tglobaladdr:$global, 0, (i32 IntRegs:$src1))>; - -// Map from store(globaladdress) -> memw(#foo). -let AddedComplexity = 100 in -def : Pat <(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), - (STriw_GP tglobaladdr:$global, 0, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress + x) -> memh(#foo + x). -let AddedComplexity = 100 in -def : Pat <(truncstorei16 (i32 IntRegs:$src1), - (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (STrih_GP tglobaladdr:$global, u16ImmPred:$offset, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress) -> memh(#foo). -let AddedComplexity = 100 in -def : Pat <(truncstorei16 (i32 IntRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STh_GP tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress + x) -> memb(#foo + x). -let AddedComplexity = 100 in -def : Pat <(truncstorei8 (i32 IntRegs:$src1), - (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset)), - (STrib_GP tglobaladdr:$global, u16ImmPred:$offset, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from store(globaladdress) -> memb(#foo). -let AddedComplexity = 100 in -def : Pat <(truncstorei8 (i32 IntRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STb_GP tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memw(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (load (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDriw_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memw(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDw_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memd(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i64 (load (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i64 (LDrid_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memw(#foo + 0). -let AddedComplexity = 100 in -def : Pat <(i64 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i64 (LDd_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd. -let AddedComplexity = 100 in -def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i1 (TFR_PdRs (i32 (LDb_GP tglobaladdr:$global))))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memh(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi16 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDrih_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memh(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDrih_GP tglobaladdr:$global, 0))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memuh(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi16 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDriuh_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memuh(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDriuh_GP tglobaladdr:$global, 0))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memh(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDh_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memuh(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDuh_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memb(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi8 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDrib_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memb(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi8 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDrib_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress + x) -> memub(#foo + x). -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi8 (add (HexagonCONST32_GP tglobaladdr:$global), - u16ImmPred:$offset))), - (i32 (LDriub_GP tglobaladdr:$global, u16ImmPred:$offset))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memb(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// Map from load(globaladdress) -> memb(#foo). -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; -// Map from load(globaladdress) -> memub(#foo). let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDub_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -// When the Interprocedural Global Variable optimizer realizes that a -// certain global variable takes only two constant values, it shrinks the -// global to a boolean. Catch those loads here in the following 3 patterns. -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP tglobaladdr:$global))>, - Requires<[NoV4T]>; - -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDub_GP tglobaladdr:$global))>, +def : Pat <(i64 (zextloadi1 (HexagonCONST32 tglobaladdr:$global))), + (i64 (COMBINE_rr (TFRI 0), + (LDriub_indexed (CONST32_set tglobaladdr:$global), 0)))>, Requires<[NoV4T]>; // Map from i1 loads to 32 bits. This assumes that the i1* is byte aligned. +let AddedComplexity = 10 in def : Pat <(i32 (zextloadi1 ADDRriS11_0:$addr)), (i32 (AND_rr (i32 (LDrib ADDRriS11_0:$addr)), (TFRI 0x1)))>; @@ -2597,43 +2196,46 @@ def : Pat <(i64 (sext_inreg (i64 DoubleRegs:$src1), i8)), subreg_loreg))))))>; // We want to prevent emitting pnot's as much as possible. -// Map brcond with an unsupported setcc to a JMP_cNot. +// Map brcond with an unsupported setcc to a JMP_f. def : Pat <(brcond (i1 (setne (i32 IntRegs:$src1), (i32 IntRegs:$src2))), bb:$offset), - (JMP_cNot (CMPEQrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), + (JMP_f (CMPEQrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), bb:$offset)>; def : Pat <(brcond (i1 (setne (i32 IntRegs:$src1), s10ImmPred:$src2)), bb:$offset), - (JMP_cNot (CMPEQri (i32 IntRegs:$src1), s10ImmPred:$src2), bb:$offset)>; + (JMP_f (CMPEQri (i32 IntRegs:$src1), s10ImmPred:$src2), bb:$offset)>; def : Pat <(brcond (i1 (setne (i1 PredRegs:$src1), (i1 -1))), bb:$offset), - (JMP_cNot (i1 PredRegs:$src1), bb:$offset)>; + (JMP_f (i1 PredRegs:$src1), bb:$offset)>; def : Pat <(brcond (i1 (setne (i1 PredRegs:$src1), (i1 0))), bb:$offset), - (JMP_c (i1 PredRegs:$src1), bb:$offset)>; + (JMP_t (i1 PredRegs:$src1), bb:$offset)>; +// cmp.lt(Rs, Imm) -> !cmp.ge(Rs, Imm) -> !cmp.gt(Rs, Imm-1) def : Pat <(brcond (i1 (setlt (i32 IntRegs:$src1), s8ImmPred:$src2)), bb:$offset), - (JMP_cNot (CMPGEri (i32 IntRegs:$src1), s8ImmPred:$src2), bb:$offset)>; + (JMP_f (CMPGTri (i32 IntRegs:$src1), + (DEC_CONST_SIGNED s8ImmPred:$src2)), bb:$offset)>; +// cmp.lt(r0, r1) -> cmp.gt(r1, r0) def : Pat <(brcond (i1 (setlt (i32 IntRegs:$src1), (i32 IntRegs:$src2))), bb:$offset), - (JMP_c (CMPLTrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), bb:$offset)>; + (JMP_t (CMPGTrr (i32 IntRegs:$src2), (i32 IntRegs:$src1)), bb:$offset)>; def : Pat <(brcond (i1 (setuge (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), bb:$offset), - (JMP_cNot (CMPGTU64rr (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1)), + (JMP_f (CMPGTU64rr (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1)), bb:$offset)>; def : Pat <(brcond (i1 (setule (i32 IntRegs:$src1), (i32 IntRegs:$src2))), bb:$offset), - (JMP_cNot (CMPGTUrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), + (JMP_f (CMPGTUrr (i32 IntRegs:$src1), (i32 IntRegs:$src2)), bb:$offset)>; def : Pat <(brcond (i1 (setule (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), bb:$offset), - (JMP_cNot (CMPGTU64rr (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2)), + (JMP_f (CMPGTU64rr (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2)), bb:$offset)>; // Map from a 64-bit select to an emulated 64-bit mux. @@ -2694,12 +2296,6 @@ def : Pat<(truncstorei32 (i64 DoubleRegs:$src), ADDRriS11_0:$addr), def : Pat<(store (i1 -1), ADDRriS11_2:$addr), (STrib ADDRriS11_2:$addr, (TFRI 1))>; -let AddedComplexity = 100 in -// Map from i1 = constant<-1>; memw(CONST32(#foo)) = i1 -> r0 = 1; -// memw(#foo) = r0 -def : Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)), - (STb_GP tglobaladdr:$global, (TFRI 1))>, - Requires<[NoV4T]>; // Map from i1 = constant<-1>; store i1 -> r0 = 1; store r0. def : Pat<(store (i1 -1), ADDRriS11_2:$addr), @@ -2717,8 +2313,8 @@ def : Pat<(i64 (anyext (i32 IntRegs:$src1))), // Map cmple -> cmpgt. // rs <= rt -> !(rs > rt). -def : Pat<(i1 (setle (i32 IntRegs:$src1), s10ImmPred:$src2)), - (i1 (NOT_p (CMPGTri (i32 IntRegs:$src1), s10ImmPred:$src2)))>; +def : Pat<(i1 (setle (i32 IntRegs:$src1), s10ExtPred:$src2)), + (i1 (NOT_p (CMPGTri (i32 IntRegs:$src1), s10ExtPred:$src2)))>; // rs <= rt -> !(rs > rt). def : Pat<(i1 (setle (i32 IntRegs:$src1), (i32 IntRegs:$src2))), @@ -2731,8 +2327,8 @@ def : Pat<(i1 (setle (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), // Map cmpne -> cmpeq. // Hexagon_TODO: We should improve on this. // rs != rt -> !(rs == rt). -def : Pat <(i1 (setne (i32 IntRegs:$src1), s10ImmPred:$src2)), - (i1 (NOT_p(i1 (CMPEQri (i32 IntRegs:$src1), s10ImmPred:$src2))))>; +def : Pat <(i1 (setne (i32 IntRegs:$src1), s10ExtPred:$src2)), + (i1 (NOT_p(i1 (CMPEQri (i32 IntRegs:$src1), s10ExtPred:$src2))))>; // Map cmpne(Rs) -> !cmpeqe(Rs). // rs != rt -> !(rs == rt). @@ -2754,8 +2350,9 @@ def : Pat <(i1 (setne (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), def : Pat <(i1 (setge (i32 IntRegs:$src1), (i32 IntRegs:$src2))), (i1 (NOT_p (i1 (CMPGTrr (i32 IntRegs:$src2), (i32 IntRegs:$src1)))))>; -def : Pat <(i1 (setge (i32 IntRegs:$src1), s8ImmPred:$src2)), - (i1 (CMPGEri (i32 IntRegs:$src1), s8ImmPred:$src2))>; +// cmpge(Rs, Imm) -> cmpgt(Rs, Imm-1) +def : Pat <(i1 (setge (i32 IntRegs:$src1), s8ExtPred:$src2)), + (i1 (CMPGTri (i32 IntRegs:$src1), (DEC_CONST_SIGNED s8ExtPred:$src2)))>; // Map cmpge(Rss, Rtt) -> !cmpgt(Rtt, Rss). // rss >= rtt -> !(rtt > rss). @@ -2764,9 +2361,10 @@ def : Pat <(i1 (setge (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), (i64 DoubleRegs:$src1)))))>; // Map cmplt(Rs, Imm) -> !cmpge(Rs, Imm). +// !cmpge(Rs, Imm) -> !cmpgt(Rs, Imm-1). // rs < rt -> !(rs >= rt). -def : Pat <(i1 (setlt (i32 IntRegs:$src1), s8ImmPred:$src2)), - (i1 (NOT_p (CMPGEri (i32 IntRegs:$src1), s8ImmPred:$src2)))>; +def : Pat <(i1 (setlt (i32 IntRegs:$src1), s8ExtPred:$src2)), + (i1 (NOT_p (CMPGTri (i32 IntRegs:$src1), (DEC_CONST_SIGNED s8ExtPred:$src2))))>; // Map cmplt(Rs, Rt) -> cmpgt(Rt, Rs). // rs < rt -> rt > rs. @@ -2790,13 +2388,17 @@ def : Pat <(i1 (setult (i32 IntRegs:$src1), (i32 IntRegs:$src2))), def : Pat <(i1 (setult (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), (i1 (CMPGTU64rr (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1)))>; -// Generate cmpgeu(Rs, #u8) -def : Pat <(i1 (setuge (i32 IntRegs:$src1), u8ImmPred:$src2)), - (i1 (CMPGEUri (i32 IntRegs:$src1), u8ImmPred:$src2))>; +// Generate cmpgeu(Rs, #0) -> cmpeq(Rs, Rs) +def : Pat <(i1 (setuge (i32 IntRegs:$src1), 0)), + (i1 (CMPEQrr (i32 IntRegs:$src1), (i32 IntRegs:$src1)))>; + +// Generate cmpgeu(Rs, #u8) -> cmpgtu(Rs, #u8 -1) +def : Pat <(i1 (setuge (i32 IntRegs:$src1), u8ExtPred:$src2)), + (i1 (CMPGTUri (i32 IntRegs:$src1), (DEC_CONST_UNSIGNED u8ExtPred:$src2)))>; // Generate cmpgtu(Rs, #u9) -def : Pat <(i1 (setugt (i32 IntRegs:$src1), u9ImmPred:$src2)), - (i1 (CMPGTUri (i32 IntRegs:$src1), u9ImmPred:$src2))>; +def : Pat <(i1 (setugt (i32 IntRegs:$src1), u9ExtPred:$src2)), + (i1 (CMPGTUri (i32 IntRegs:$src1), u9ExtPred:$src2))>; // Map from Rs >= Rt -> !(Rt > Rs). // rs >= rt -> !(rt > rs). @@ -2808,7 +2410,7 @@ def : Pat <(i1 (setuge (i32 IntRegs:$src1), (i32 IntRegs:$src2))), def : Pat <(i1 (setuge (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))), (i1 (NOT_p (CMPGTU64rr (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1))))>; -// Map from cmpleu(Rs, Rs) -> !cmpgtu(Rs, Rs). +// Map from cmpleu(Rs, Rt) -> !cmpgtu(Rs, Rt). // Map from (Rs <= Rt) -> !(Rs > Rt). def : Pat <(i1 (setule (i32 IntRegs:$src1), (i32 IntRegs:$src2))), (i1 (NOT_p (CMPGTUrr (i32 IntRegs:$src1), (i32 IntRegs:$src2))))>; @@ -2904,6 +2506,13 @@ def: Pat <(i64 (zextloadi32 ADDRriS11_2:$src1)), (i64 (COMBINE_rr (TFRI 0), (LDriw ADDRriS11_2:$src1)))>, Requires<[NoV4T]>; +let AddedComplexity = 100 in +def: Pat <(i64 (zextloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))), + (i64 (COMBINE_rr (TFRI 0), (LDriw_indexed IntRegs:$src1, + s11_2ExtPred:$offset)))>, + Requires<[NoV4T]>; + +let AddedComplexity = 10 in def: Pat <(i32 (zextloadi1 ADDRriS11_0:$src1)), (i32 (LDriw ADDRriS11_0:$src1))>; @@ -2920,6 +2529,48 @@ def : Pat <(i64 (anyext (i1 PredRegs:$src1))), (i64 (SXTW (i32 (MUX_ii (i1 PredRegs:$src1), 1, 0))))>; +let AddedComplexity = 100 in +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zextloadi32 (i32 (add IntRegs:$src2, + s11_2ExtPred:$offset2)))))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + (LDriw_indexed IntRegs:$src2, + s11_2ExtPred:$offset2)))>; + +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zextloadi32 ADDRriS11_2:$srcLow)))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + (LDriw ADDRriS11_2:$srcLow)))>; + +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zext (i32 IntRegs:$srcLow))))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + IntRegs:$srcLow))>; + +let AddedComplexity = 100 in +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zextloadi32 (i32 (add IntRegs:$src2, + s11_2ExtPred:$offset2)))))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + (LDriw_indexed IntRegs:$src2, + s11_2ExtPred:$offset2)))>; + +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zextloadi32 ADDRriS11_2:$srcLow)))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + (LDriw ADDRriS11_2:$srcLow)))>; + +def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh), + (i32 32))), + (i64 (zext (i32 IntRegs:$srcLow))))), + (i64 (COMBINE_rr (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg), + IntRegs:$srcLow))>; + // Any extended 64-bit load. // anyext i32 -> i64 def: Pat <(i64 (extloadi32 ADDRriS11_2:$src1)), @@ -3054,19 +2705,6 @@ let AddedComplexity = 100 in def : Pat<(i32 (sext_inreg (Hexagon_ARGEXTEND (i32 IntRegs:$src1)), i16)), (COPY (i32 IntRegs:$src1))>; -def SDHexagonBR_JT: SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>; -def HexagonBR_JT: SDNode<"HexagonISD::BR_JT", SDHexagonBR_JT, [SDNPHasChain]>; - -let isBranch=1, isIndirectBranch=1, isTerminator=1, isBarrier = 1 in -def BR_JT : JRInst<(outs), (ins IntRegs:$src), - "jumpr $src", - [(HexagonBR_JT (i32 IntRegs:$src))]>; - -let isBranch=1, isIndirectBranch=1, isTerminator=1 in -def BRIND : JRInst<(outs), (ins IntRegs:$src), - "jumpr $src", - [(brind (i32 IntRegs:$src))]>; - def HexagonWrapperJT: SDNode<"HexagonISD::WrapperJT", SDTIntUnaryOp>; def : Pat<(HexagonWrapperJT tjumptable:$dst), diff --git a/lib/Target/Hexagon/HexagonInstrInfoV3.td b/lib/Target/Hexagon/HexagonInstrInfoV3.td index 157ab3d..7e75554 100644 --- a/lib/Target/Hexagon/HexagonInstrInfoV3.td +++ b/lib/Target/Hexagon/HexagonInstrInfoV3.td @@ -11,6 +11,11 @@ // //===----------------------------------------------------------------------===// +def callv3 : SDNode<"HexagonISD::CALLv3", SDT_SPCall, + [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>; + +def callv3nr : SDNode<"HexagonISD::CALLv3nr", SDT_SPCall, + [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>; //===----------------------------------------------------------------------===// // J + @@ -40,41 +45,6 @@ let isCall = 1, neverHasSideEffects = 1, []>, Requires<[HasV3TOnly]>; } - -// Jump to address from register -// if(p?.new) jumpr:t r? -let isReturn = 1, isTerminator = 1, isBarrier = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cdnPt_V3: JRInst<(outs), (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) jumpr:t $src2", - []>, Requires<[HasV3T]>; -} - -// if (!p?.new) jumpr:t r? -let isReturn = 1, isTerminator = 1, isBarrier = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cdnNotPt_V3: JRInst<(outs), (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) jumpr:t $src2", - []>, Requires<[HasV3T]>; -} - -// Not taken. -// if(p?.new) jumpr:nt r? -let isReturn = 1, isTerminator = 1, isBarrier = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cdnPnt: JRInst<(outs), (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) jumpr:nt $src2", - []>, Requires<[HasV3T]>; -} - -// if (!p?.new) jumpr:nt r? -let isReturn = 1, isTerminator = 1, isBarrier = 1, - Defs = [PC], Uses = [R31] in { - def JMPR_cdnNotPnt: JRInst<(outs), (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) jumpr:nt $src2", - []>, Requires<[HasV3T]>; -} - //===----------------------------------------------------------------------===// // JR - //===----------------------------------------------------------------------===// diff --git a/lib/Target/Hexagon/HexagonInstrInfoV4.td b/lib/Target/Hexagon/HexagonInstrInfoV4.td index 1d0643d..744efe8 100644 --- a/lib/Target/Hexagon/HexagonInstrInfoV4.td +++ b/lib/Target/Hexagon/HexagonInstrInfoV4.td @@ -95,164 +95,6 @@ def NumUsesBelowThresCONST32 : PatFrag<(ops node:$addr), //===----------------------------------------------------------------------===// // ALU32 + //===----------------------------------------------------------------------===// - -// Shift halfword. - -let isPredicated = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in { -def ASLH_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = aslh($src2)", - []>, - Requires<[HasV4T]>; - -def ASLH_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = aslh($src2)", - []>, - Requires<[HasV4T]>; - -def ASLH_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = aslh($src2)", - []>, - Requires<[HasV4T]>; - -def ASLH_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = aslh($src2)", - []>, - Requires<[HasV4T]>; - -def ASRH_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = asrh($src2)", - []>, - Requires<[HasV4T]>; - -def ASRH_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = asrh($src2)", - []>, - Requires<[HasV4T]>; - -def ASRH_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = asrh($src2)", - []>, - Requires<[HasV4T]>; - -def ASRH_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = asrh($src2)", - []>, - Requires<[HasV4T]>; -} - -// Sign extend. - -let isPredicated = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in { -def SXTB_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = sxtb($src2)", - []>, - Requires<[HasV4T]>; - -def SXTB_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = sxtb($src2)", - []>, - Requires<[HasV4T]>; - -def SXTB_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = sxtb($src2)", - []>, - Requires<[HasV4T]>; - -def SXTB_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = sxtb($src2)", - []>, - Requires<[HasV4T]>; - - -def SXTH_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = sxth($src2)", - []>, - Requires<[HasV4T]>; - -def SXTH_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = sxth($src2)", - []>, - Requires<[HasV4T]>; - -def SXTH_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = sxth($src2)", - []>, - Requires<[HasV4T]>; - -def SXTH_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = sxth($src2)", - []>, - Requires<[HasV4T]>; -} - -// Zero exten. - -let neverHasSideEffects = 1, isPredicated = 1, validSubTargets = HasV4SubT in { -def ZXTB_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = zxtb($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTB_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = zxtb($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTB_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = zxtb($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTB_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = zxtb($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTH_cPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1) $dst = zxth($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTH_cNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1) $dst = zxth($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTH_cdnPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if ($src1.new) $dst = zxth($src2)", - []>, - Requires<[HasV4T]>; - -def ZXTH_cdnNotPt_V4 : ALU32_rr<(outs IntRegs:$dst), - (ins PredRegs:$src1, IntRegs:$src2), - "if (!$src1.new) $dst = zxth($src2)", - []>, - Requires<[HasV4T]>; -} - // Generate frame index addresses. let neverHasSideEffects = 1, isReMaterializable = 1, isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT in @@ -367,105 +209,31 @@ def COMBINE_iI_V4 : ALU32_ii<(outs DoubleRegs:$dst), //===----------------------------------------------------------------------===// // LD + //===----------------------------------------------------------------------===// -// -// These absolute set addressing mode instructions accept immediate as -// an operand. We have duplicated these patterns to take global address. - +//===----------------------------------------------------------------------===// +// Template class for load instructions with Absolute set addressing mode. +//===----------------------------------------------------------------------===// let isExtended = 1, opExtendable = 2, neverHasSideEffects = 1, -validSubTargets = HasV4SubT in { -def LDrid_abs_setimm_V4 : LDInst2<(outs DoubleRegs:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = memd($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memb(Re=#U6) -def LDrib_abs_setimm_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = memb($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memh(Re=#U6) -def LDrih_abs_setimm_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = memh($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memub(Re=#U6) -def LDriub_abs_setimm_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = memub($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memuh(Re=#U6) -def LDriuh_abs_setimm_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), +validSubTargets = HasV4SubT in +class T_LD_abs_set<string mnemonic, RegisterClass RC>: + LDInst2<(outs RC:$dst1, IntRegs:$dst2), (ins u0AlwaysExt:$addr), - "$dst1 = memuh($dst2=##$addr)", + "$dst1 = "#mnemonic#"($dst2=##$addr)", []>, Requires<[HasV4T]>; -// Rd=memw(Re=#U6) -def LDriw_abs_setimm_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = memw($dst2=##$addr)", - []>, - Requires<[HasV4T]>; -} +def LDrid_abs_set_V4 : T_LD_abs_set <"memd", DoubleRegs>; +def LDrib_abs_set_V4 : T_LD_abs_set <"memb", IntRegs>; +def LDriub_abs_set_V4 : T_LD_abs_set <"memub", IntRegs>; +def LDrih_abs_set_V4 : T_LD_abs_set <"memh", IntRegs>; +def LDriw_abs_set_V4 : T_LD_abs_set <"memw", IntRegs>; +def LDriuh_abs_set_V4 : T_LD_abs_set <"memuh", IntRegs>; -// Following patterns are defined for absolute set addressing mode -// instruction which take global address as operand. -let isExtended = 1, opExtendable = 2, neverHasSideEffects = 1, -validSubTargets = HasV4SubT in { -def LDrid_abs_set_V4 : LDInst2<(outs DoubleRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memd($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memb(Re=#U6) -def LDrib_abs_set_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memb($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memh(Re=#U6) -def LDrih_abs_set_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memh($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memub(Re=#U6) -def LDriub_abs_set_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memub($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memuh(Re=#U6) -def LDriuh_abs_set_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memuh($dst2=##$addr)", - []>, - Requires<[HasV4T]>; - -// Rd=memw(Re=#U6) -def LDriw_abs_set_V4 : LDInst2<(outs IntRegs:$dst1, IntRegs:$dst2), - (ins globaladdressExt:$addr), - "$dst1 = memw($dst2=##$addr)", - []>, - Requires<[HasV4T]>; -} // multiclass for load instructions with base + register offset // addressing mode multiclass ld_idxd_shl_pbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : LDInst2<(outs RC:$dst), (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3, u2Imm:$offset), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -474,7 +242,7 @@ multiclass ld_idxd_shl_pbase<string mnemonic, RegisterClass RC, bit isNot, } multiclass ld_idxd_shl_pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ld_idxd_shl_pbase<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : ld_idxd_shl_pbase<mnemonic, RC, PredNot, 1>; @@ -596,329 +364,6 @@ def : Pat <(i32 (load (add IntRegs:$src1, IntRegs:$src2))), Requires<[HasV4T]>; } -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDd_GP_V4 : LDInst2<(outs DoubleRegs:$dst), - (ins globaladdress:$global), - "$dst=memd(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rtt=memd(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDd_GP_cPt_V4 : LDInst2<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memd(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rtt=memd(##global) -def LDd_GP_cNotPt_V4 : LDInst2<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memd(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rtt=memd(##global) -def LDd_GP_cdnPt_V4 : LDInst2<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memd(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rtt=memd(##global) -def LDd_GP_cdnNotPt_V4 : LDInst2<(outs DoubleRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memd(##$global)", - []>, - Requires<[HasV4T]>; -} - -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDb_GP_V4 : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst=memb(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memb(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDb_GP_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memb(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memb(##global) -def LDb_GP_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memb(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memb(##global) -def LDb_GP_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memb(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memb(##global) -def LDb_GP_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memb(##$global)", - []>, - Requires<[HasV4T]>; -} - -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDub_GP_V4 : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst=memub(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memub(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDub_GP_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memub(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rt=memub(##global) -def LDub_GP_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memub(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memub(##global) -def LDub_GP_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memub(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rt=memub(##global) -def LDub_GP_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memub(##$global)", - []>, - Requires<[HasV4T]>; -} - -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDh_GP_V4 : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst=memh(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memh(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDh_GP_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memh(##global) -def LDh_GP_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memh(##global) -def LDh_GP_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memh(##global) -def LDh_GP_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memh(##$global)", - []>, - Requires<[HasV4T]>; -} - -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDuh_GP_V4 : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst=memuh(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memuh(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDuh_GP_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memuh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memuh(##global) -def LDuh_GP_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memuh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memuh(##global) -def LDuh_GP_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memuh(##$global)", - []>, - Requires<[HasV4T]>; - -// if (!Pv) Rt=memuh(##global) -def LDuh_GP_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memuh(##$global)", - []>, - Requires<[HasV4T]>; -} - -let isPredicable = 1, neverHasSideEffects = 1, validSubTargets = HasV4SubT in -def LDw_GP_V4 : LDInst2<(outs IntRegs:$dst), - (ins globaladdress:$global), - "$dst=memw(#$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memw(##global) -let neverHasSideEffects = 1, isPredicated = 1, isExtended = 1, opExtendable = 2, -validSubTargets = HasV4SubT in { -def LDw_GP_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1) $dst=memw(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rt=memw(##global) -def LDw_GP_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1) $dst=memw(##$global)", - []>, - Requires<[HasV4T]>; - -// if (Pv) Rt=memw(##global) -def LDw_GP_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if ($src1.new) $dst=memw(##$global)", - []>, - Requires<[HasV4T]>; - - -// if (!Pv) Rt=memw(##global) -def LDw_GP_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$global), - "if (!$src1.new) $dst=memw(##$global)", - []>, - Requires<[HasV4T]>; -} - - -def : Pat <(atomic_load_64 (HexagonCONST32_GP tglobaladdr:$global)), - (i64 (LDd_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -def : Pat <(atomic_load_32 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDw_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -def : Pat <(atomic_load_16 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDuh_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -def : Pat <(atomic_load_8 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (LDub_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memw(#foo + 0) -let AddedComplexity = 100 in -def : Pat <(i64 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i64 (LDd_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd -let AddedComplexity = 100 in -def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i1 (TFR_PdRs (i32 (LDb_GP_V4 tglobaladdr:$global))))>, - Requires<[HasV4T]>; - -// When the Interprocedural Global Variable optimizer realizes that a certain -// global variable takes only two constant values, it shrinks the global to -// a boolean. Catch those loads here in the following 3 patterns. -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memb(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memb(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDb_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDub_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memub(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDub_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memh(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDh_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memh(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDh_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memuh(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDuh_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress) -> memw(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (LDw_GP_V4 tglobaladdr:$global))>, - Requires<[HasV4T]>; - // zext i1->i64 def : Pat <(i64 (zext (i1 PredRegs:$src1))), (i64 (COMBINE_Ir_V4 0, (MUX_ii (i1 PredRegs:$src1), 1, 0)))>, @@ -1008,78 +453,29 @@ def: Pat <(i64 (extloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))), // ST + //===----------------------------------------------------------------------===// /// -/// Assumptions::: ****** DO NOT IGNORE ******** -/// 1. Make sure that in post increment store, the zero'th operand is always the -/// post increment operand. -/// 2. Make sure that the store value operand(Rt/Rtt) in a store is always the -/// last operand. -/// - -// memd(Re=#U)=Rtt -let isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT in { -def STrid_abs_setimm_V4 : STInst2<(outs IntRegs:$dst1), - (ins DoubleRegs:$src1, u0AlwaysExt:$src2), - "memd($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; - -// memb(Re=#U)=Rs -def STrib_abs_setimm_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, u0AlwaysExt:$src2), - "memb($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; - -// memh(Re=#U)=Rs -def STrih_abs_setimm_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, u0AlwaysExt:$src2), - "memh($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; - -// memw(Re=#U)=Rs -def STriw_abs_setimm_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, u0AlwaysExt:$src2), - "memw($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; -} - -// memd(Re=#U)=Rtt -let isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT in { -def STrid_abs_set_V4 : STInst2<(outs IntRegs:$dst1), - (ins DoubleRegs:$src1, globaladdressExt:$src2), - "memd($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; - -// memb(Re=#U)=Rs -def STrib_abs_set_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, globaladdressExt:$src2), - "memb($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; - -// memh(Re=#U)=Rs -def STrih_abs_set_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, globaladdressExt:$src2), - "memh($dst1=##$src2) = $src1", +//===----------------------------------------------------------------------===// +// Template class for store instructions with Absolute set addressing mode. +//===----------------------------------------------------------------------===// +let isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT in +class T_ST_abs_set<string mnemonic, RegisterClass RC>: + STInst2<(outs IntRegs:$dst1), + (ins RC:$src1, u0AlwaysExt:$src2), + mnemonic#"($dst1=##$src2) = $src1", []>, Requires<[HasV4T]>; -// memw(Re=#U)=Rs -def STriw_abs_set_V4 : STInst2<(outs IntRegs:$dst1), - (ins IntRegs:$src1, globaladdressExt:$src2), - "memw($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; -} +def STrid_abs_set_V4 : T_ST_abs_set <"memd", DoubleRegs>; +def STrib_abs_set_V4 : T_ST_abs_set <"memb", IntRegs>; +def STrih_abs_set_V4 : T_ST_abs_set <"memh", IntRegs>; +def STriw_abs_set_V4 : T_ST_abs_set <"memw", IntRegs>; +//===----------------------------------------------------------------------===// // multiclass for store instructions with base + register offset addressing // mode +//===----------------------------------------------------------------------===// multiclass ST_Idxd_shl_Pbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : STInst2<(outs), (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3, u2Imm:$src4, RC:$src5), @@ -1090,7 +486,7 @@ multiclass ST_Idxd_shl_Pbase<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_Idxd_shl_Pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Idxd_shl_Pbase<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Idxd_shl_Pbase<mnemonic, RC, PredNot, 1>; @@ -1118,7 +514,7 @@ multiclass ST_Idxd_shl<string mnemonic, string CextOp, RegisterClass RC> { // addressing mode. multiclass ST_Idxd_shl_Pbase_nv<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_nv_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, IntRegs:$src2, IntRegs:$src3, u2Imm:$src4, RC:$src5), @@ -1129,7 +525,7 @@ multiclass ST_Idxd_shl_Pbase_nv<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_Idxd_shl_Pred_nv<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Idxd_shl_Pbase_nv<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Idxd_shl_Pbase_nv<mnemonic, RC, PredNot, 1>; @@ -1192,17 +588,59 @@ def : Pat<(store (i64 DoubleRegs:$src4), u2ImmPred:$src3, DoubleRegs:$src4)>; } -// memd(Ru<<#u2+#U6)=Rtt -let isExtended = 1, opExtendable = 2, AddedComplexity = 10, -validSubTargets = HasV4SubT in -def STrid_shl_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, DoubleRegs:$src4), - "memd($src1<<#$src2+#$src3) = $src4", - [(store (i64 DoubleRegs:$src4), +let isExtended = 1, opExtendable = 2 in +class T_ST_LongOff <string mnemonic, PatFrag stOp, RegisterClass RC, ValueType VT> : + STInst<(outs), + (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, RC:$src4), + mnemonic#"($src1<<#$src2+##$src3) = $src4", + [(stOp (VT RC:$src4), (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2), u0AlwaysExtPred:$src3))]>, Requires<[HasV4T]>; +let isExtended = 1, opExtendable = 2, mayStore = 1, isNVStore = 1 in +class T_ST_LongOff_nv <string mnemonic> : + NVInst_V4<(outs), + (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), + mnemonic#"($src1<<#$src2+##$src3) = $src4.new", + []>, + Requires<[HasV4T]>; + +multiclass ST_LongOff <string mnemonic, string BaseOp, PatFrag stOp> { + let BaseOpcode = BaseOp#"_shl" in { + let isNVStorable = 1 in + def NAME#_V4 : T_ST_LongOff<mnemonic, stOp, IntRegs, i32>; + + def NAME#_nv_V4 : T_ST_LongOff_nv<mnemonic>; + } +} + +let AddedComplexity = 10, validSubTargets = HasV4SubT in { + def STrid_shl_V4 : T_ST_LongOff<"memd", store, DoubleRegs, i64>; + defm STrib_shl : ST_LongOff <"memb", "STrib", truncstorei8>, NewValueRel; + defm STrih_shl : ST_LongOff <"memh", "Strih", truncstorei16>, NewValueRel; + defm STriw_shl : ST_LongOff <"memw", "STriw", store>, NewValueRel; +} + +let AddedComplexity = 40 in +multiclass T_ST_LOff_Pats <InstHexagon I, RegisterClass RC, ValueType VT, + PatFrag stOp> { + def : Pat<(stOp (VT RC:$src4), + (add (shl IntRegs:$src1, u2ImmPred:$src2), + (NumUsesBelowThresCONST32 tglobaladdr:$src3))), + (I IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3, RC:$src4)>; + + def : Pat<(stOp (VT RC:$src4), + (add IntRegs:$src1, + (NumUsesBelowThresCONST32 tglobaladdr:$src3))), + (I IntRegs:$src1, 0, tglobaladdr:$src3, RC:$src4)>; +} + +defm : T_ST_LOff_Pats<STrid_shl_V4, DoubleRegs, i64, store>; +defm : T_ST_LOff_Pats<STriw_shl_V4, IntRegs, i32, store>; +defm : T_ST_LOff_Pats<STrib_shl_V4, IntRegs, i32, truncstorei8>; +defm : T_ST_LOff_Pats<STrih_shl_V4, IntRegs, i32, truncstorei16>; + // memd(Rx++#s4:3)=Rtt // memd(Rx++#s4:3:circ(Mu))=Rtt // memd(Rx++I:circ(Mu))=Rtt @@ -1222,7 +660,7 @@ def STrid_shl_V4 : STInst<(outs), //===----------------------------------------------------------------------===// multiclass ST_Imm_Pbase<string mnemonic, Operand OffsetOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : STInst2<(outs), (ins PredRegs:$src1, IntRegs:$src2, OffsetOp:$src3, s6Ext:$src4), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1232,7 +670,7 @@ multiclass ST_Imm_Pbase<string mnemonic, Operand OffsetOp, bit isNot, } multiclass ST_Imm_Pred<string mnemonic, Operand OffsetOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Imm_Pbase<mnemonic, OffsetOp, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Imm_Pbase<mnemonic, OffsetOp, PredNot, 1>; @@ -1280,17 +718,6 @@ def : Pat <(truncstorei8 s8ExtPred:$src2, (i32 IntRegs:$src1)), (STrib_imm_V4 IntRegs:$src1, 0, s8ExtPred:$src2)>, Requires<[HasV4T]>; -// memb(Ru<<#u2+#U6)=Rt -let isExtended = 1, opExtendable = 2, AddedComplexity = 10, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STrib_shl_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memb($src1<<#$src2+#$src3) = $src4", - [(truncstorei8 (i32 IntRegs:$src4), - (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2), - u0AlwaysExtPred:$src3))]>, - Requires<[HasV4T]>; - // memb(Rx++#s4:0:circ(Mu))=Rt // memb(Rx++I:circ(Mu))=Rt // memb(Rx++Mu)=Rt @@ -1311,17 +738,6 @@ def : Pat <(truncstorei16 s8ExtPred:$src2, (i32 IntRegs:$src1)), // TODO: needs to be implemented. // memh(Ru<<#u2+#U6)=Rt.H -// memh(Ru<<#u2+#U6)=Rt -let isExtended = 1, opExtendable = 2, AddedComplexity = 10, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STrih_shl_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memh($src1<<#$src2+#$src3) = $src4", - [(truncstorei16 (i32 IntRegs:$src4), - (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2), - u0AlwaysExtPred:$src3))]>, - Requires<[HasV4T]>; - // memh(Rx++#s4:1:circ(Mu))=Rt.H // memh(Rx++#s4:1:circ(Mu))=Rt // memh(Rx++I:circ(Mu))=Rt.H @@ -1358,241 +774,11 @@ def : Pat <(store s8ExtPred:$src2, (i32 IntRegs:$src1)), (STriw_imm_V4 IntRegs:$src1, 0, s8ExtPred:$src2)>, Requires<[HasV4T]>; -// memw(Ru<<#u2+#U6)=Rt -let isExtended = 1, opExtendable = 2, AddedComplexity = 10, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STriw_shl_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memw($src1<<#$src2+#$src3) = $src4", - [(store (i32 IntRegs:$src4), - (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2), - u0AlwaysExtPred:$src3))]>, - Requires<[HasV4T]>; - // memw(Rx++#s4:2)=Rt // memw(Rx++#s4:2:circ(Mu))=Rt // memw(Rx++I:circ(Mu))=Rt // memw(Rx++Mu)=Rt // memw(Rx++Mu:brev)=Rt -// memw(gp+#u16:2)=Rt - - -// memd(#global)=Rtt -let isPredicable = 1, mayStore = 1, neverHasSideEffects = 1, -validSubTargets = HasV4SubT in -def STd_GP_V4 : STInst2<(outs), - (ins globaladdress:$global, DoubleRegs:$src), - "memd(#$global) = $src", - []>, - Requires<[HasV4T]>; - -// if (Pv) memd(##global) = Rtt -let mayStore = 1, neverHasSideEffects = 1, isPredicated = 1, -isExtended = 1, opExtendable = 1, validSubTargets = HasV4SubT in { -def STd_GP_cPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, DoubleRegs:$src2), - "if ($src1) memd(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memd(##global) = Rtt -def STd_GP_cNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, DoubleRegs:$src2), - "if (!$src1) memd(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (Pv) memd(##global) = Rtt -def STd_GP_cdnPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, DoubleRegs:$src2), - "if ($src1.new) memd(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memd(##global) = Rtt -def STd_GP_cdnNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, DoubleRegs:$src2), - "if (!$src1.new) memd(##$global) = $src2", - []>, - Requires<[HasV4T]>; -} - -// memb(#global)=Rt -let isPredicable = 1, neverHasSideEffects = 1, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STb_GP_V4 : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memb(#$global) = $src", - []>, - Requires<[HasV4T]>; - -// if (Pv) memb(##global) = Rt -let neverHasSideEffects = 1, isPredicated = 1, isNVStorable = 1, -isExtended = 1, opExtendable = 1, validSubTargets = HasV4SubT in { -def STb_GP_cPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memb(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memb(##global) = Rt -def STb_GP_cNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memb(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (Pv) memb(##global) = Rt -def STb_GP_cdnPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memb(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memb(##global) = Rt -def STb_GP_cdnNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memb(##$global) = $src2", - []>, - Requires<[HasV4T]>; -} - -// memh(#global)=Rt -let isPredicable = 1, neverHasSideEffects = 1, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STh_GP_V4 : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memh(#$global) = $src", - []>, - Requires<[HasV4T]>; - -// if (Pv) memh(##global) = Rt -let neverHasSideEffects = 1, isPredicated = 1, isNVStorable = 1, -isExtended = 1, opExtendable = 1, validSubTargets = HasV4SubT in { -def STh_GP_cPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memh(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memh(##global) = Rt -def STh_GP_cNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memh(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (Pv) memh(##global) = Rt -def STh_GP_cdnPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memh(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memh(##global) = Rt -def STh_GP_cdnNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memh(##$global) = $src2", - []>, - Requires<[HasV4T]>; -} - -// memw(#global)=Rt -let isPredicable = 1, neverHasSideEffects = 1, isNVStorable = 1, -validSubTargets = HasV4SubT in -def STw_GP_V4 : STInst2<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memw(#$global) = $src", - []>, - Requires<[HasV4T]>; - -// if (Pv) memw(##global) = Rt -let neverHasSideEffects = 1, isPredicated = 1, isNVStorable = 1, -isExtended = 1, opExtendable = 1, validSubTargets = HasV4SubT in { -def STw_GP_cPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memw(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memw(##global) = Rt -def STw_GP_cNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memw(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (Pv) memw(##global) = Rt -def STw_GP_cdnPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memw(##$global) = $src2", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memw(##global) = Rt -def STw_GP_cdnNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memw(##$global) = $src2", - []>, - Requires<[HasV4T]>; -} - -// 64 bit atomic store -def : Pat <(atomic_store_64 (HexagonCONST32_GP tglobaladdr:$global), - (i64 DoubleRegs:$src1)), - (STd_GP_V4 tglobaladdr:$global, (i64 DoubleRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress) -> memd(#foo) -let AddedComplexity = 100 in -def : Pat <(store (i64 DoubleRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STd_GP_V4 tglobaladdr:$global, (i64 DoubleRegs:$src1))>, - Requires<[HasV4T]>; - -// 8 bit atomic store -def : Pat < (atomic_store_8 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STb_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress) -> memb(#foo) -let AddedComplexity = 100 in -def : Pat<(truncstorei8 (i32 IntRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STb_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from "i1 = constant<-1>; memw(CONST32(#foo)) = i1" -// to "r0 = 1; memw(#foo) = r0" -let AddedComplexity = 100 in -def : Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)), - (STb_GP_V4 tglobaladdr:$global, (TFRI 1))>, - Requires<[HasV4T]>; - -def : Pat<(atomic_store_16 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STh_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress) -> memh(#foo) -let AddedComplexity = 100 in -def : Pat<(truncstorei16 (i32 IntRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (STh_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// 32 bit atomic store -def : Pat<(atomic_store_32 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (STw_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress) -> memw(#foo) -let AddedComplexity = 100 in -def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), - (STw_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; //===----------------------------------------------------------------------=== // ST - @@ -1607,7 +793,7 @@ def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), // multiclass ST_Idxd_Pbase_nv<string mnemonic, RegisterClass RC, Operand predImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_nv_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3, RC: $src4), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1618,7 +804,7 @@ multiclass ST_Idxd_Pbase_nv<string mnemonic, RegisterClass RC, multiclass ST_Idxd_Pred_nv<string mnemonic, RegisterClass RC, Operand predImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Idxd_Pbase_nv<mnemonic, RC, predImmOp, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Idxd_Pbase_nv<mnemonic, RC, predImmOp, PredNot, 1>; @@ -1660,7 +846,7 @@ let addrMode = BaseImmOffset, validSubTargets = HasV4SubT in { // and MEMri operand. multiclass ST_MEMri_Pbase_nv<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_nv_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, MEMri:$addr, RC: $src2), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1670,7 +856,7 @@ multiclass ST_MEMri_Pbase_nv<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_MEMri_Pred_nv<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_MEMri_Pbase_nv<mnemonic, RC, PredNot, 0>; // Predicate new @@ -1706,15 +892,6 @@ mayStore = 1 in { defm STriw: ST_MEMri_nv<"memw", "STriw", IntRegs, 13, 8>, AddrModeRel; } -// memb(Ru<<#u2+#U6)=Nt.new -let isExtended = 1, opExtendable = 2, mayStore = 1, AddedComplexity = 10, -isNVStore = 1, validSubTargets = HasV4SubT in -def STrib_shl_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memb($src1<<#$src2+#$src3) = $src4.new", - []>, - Requires<[HasV4T]>; - //===----------------------------------------------------------------------===// // Post increment store // mem[bhwd](Rx++#s4:[0123])=Nt.new @@ -1722,7 +899,7 @@ def STrib_shl_nv_V4 : NVInst_V4<(outs), multiclass ST_PostInc_Pbase_nv<string mnemonic, RegisterClass RC, Operand ImmOp, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_nv_V4 : NVInstPI_V4<(outs IntRegs:$dst), (ins PredRegs:$src1, IntRegs:$src2, ImmOp:$offset, RC:$src3), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", @@ -1734,7 +911,7 @@ multiclass ST_PostInc_Pbase_nv<string mnemonic, RegisterClass RC, Operand ImmOp, multiclass ST_PostInc_Pred_nv<string mnemonic, RegisterClass RC, Operand ImmOp, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_PostInc_Pbase_nv<mnemonic, RC, ImmOp, PredNot, 0>; // Predicate new let Predicates = [HasV4T], validSubTargets = HasV4SubT in @@ -1772,146 +949,15 @@ defm POST_STwri: ST_PostInc_nv <"memw", "STriw", IntRegs, s4_2Imm>, AddrModeRel; // memb(Rx++I:circ(Mu))=Nt.new // memb(Rx++Mu)=Nt.new // memb(Rx++Mu:brev)=Nt.new - -// memb(#global)=Nt.new -let mayStore = 1, neverHasSideEffects = 1 in -def STb_GP_nv_V4 : NVInst_V4<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memb(#$global) = $src.new", - []>, - Requires<[HasV4T]>; - -// memh(Ru<<#u2+#U6)=Nt.new -let isExtended = 1, opExtendable = 2, mayStore = 1, AddedComplexity = 10, -isNVStore = 1, validSubTargets = HasV4SubT in -def STrih_shl_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memh($src1<<#$src2+#$src3) = $src4.new", - []>, - Requires<[HasV4T]>; - // memh(Rx++#s4:1:circ(Mu))=Nt.new // memh(Rx++I:circ(Mu))=Nt.new // memh(Rx++Mu)=Nt.new // memh(Rx++Mu:brev)=Nt.new -// memh(#global)=Nt.new -let mayStore = 1, neverHasSideEffects = 1 in -def STh_GP_nv_V4 : NVInst_V4<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memh(#$global) = $src.new", - []>, - Requires<[HasV4T]>; - -// memw(Ru<<#u2+#U6)=Nt.new -let isExtended = 1, opExtendable = 2, mayStore = 1, AddedComplexity = 10, -isNVStore = 1, validSubTargets = HasV4SubT in -def STriw_shl_nv_V4 : NVInst_V4<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - "memw($src1<<#$src2+#$src3) = $src4.new", - []>, - Requires<[HasV4T]>; - // memw(Rx++#s4:2:circ(Mu))=Nt.new // memw(Rx++I:circ(Mu))=Nt.new // memw(Rx++Mu)=Nt.new // memw(Rx++Mu:brev)=Nt.new -// memw(gp+#u16:2)=Nt.new - -let mayStore = 1, neverHasSideEffects = 1, isNVStore = 1, -validSubTargets = HasV4SubT in -def STw_GP_nv_V4 : NVInst_V4<(outs), - (ins globaladdress:$global, IntRegs:$src), - "memw(#$global) = $src.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memb(##global) = Rt -let mayStore = 1, neverHasSideEffects = 1, isNVStore = 1, -isExtended = 1, opExtendable = 1, validSubTargets = HasV4SubT in { -def STb_GP_cPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memb(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memb(##global) = Rt -def STb_GP_cNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memb(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memb(##global) = Rt -def STb_GP_cdnPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memb(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memb(##global) = Rt -def STb_GP_cdnNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memb(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memh(##global) = Rt -def STh_GP_cPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memh(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memh(##global) = Rt -def STh_GP_cNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memh(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memh(##global) = Rt -def STh_GP_cdnPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memh(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memh(##global) = Rt -def STh_GP_cdnNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memh(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memw(##global) = Rt -def STw_GP_cPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1) memw(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memw(##global) = Rt -def STw_GP_cNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1) memw(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (Pv) memw(##global) = Rt -def STw_GP_cdnPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if ($src1.new) memw(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; - -// if (!Pv) memw(##global) = Rt -def STw_GP_cdnNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdress:$global, IntRegs:$src2), - "if (!$src1.new) memw(##$global) = $src2.new", - []>, - Requires<[HasV4T]>; -} //===----------------------------------------------------------------------===// // NV/ST - @@ -2658,414 +1704,367 @@ def LSRd_rr_xor_V4 : MInst_acc<(outs DoubleRegs:$dst), // MEMOP: Word, Half, Byte //===----------------------------------------------------------------------===// +def MEMOPIMM : SDNodeXForm<imm, [{ + // Call the transformation function XformM5ToU5Imm to get the negative + // immediate's positive counterpart. + int32_t imm = N->getSExtValue(); + return XformM5ToU5Imm(imm); +}]>; + +def MEMOPIMM_HALF : SDNodeXForm<imm, [{ + // -1 .. -31 represented as 65535..65515 + // assigning to a short restores our desired signed value. + // Call the transformation function XformM5ToU5Imm to get the negative + // immediate's positive counterpart. + int16_t imm = N->getSExtValue(); + return XformM5ToU5Imm(imm); +}]>; + +def MEMOPIMM_BYTE : SDNodeXForm<imm, [{ + // -1 .. -31 represented as 255..235 + // assigning to a char restores our desired signed value. + // Call the transformation function XformM5ToU5Imm to get the negative + // immediate's positive counterpart. + int8_t imm = N->getSExtValue(); + return XformM5ToU5Imm(imm); +}]>; + +def SETMEMIMM : SDNodeXForm<imm, [{ + // Return the bit position we will set [0-31]. + // As an SDNode. + int32_t imm = N->getSExtValue(); + return XformMskToBitPosU5Imm(imm); +}]>; + +def CLRMEMIMM : SDNodeXForm<imm, [{ + // Return the bit position we will clear [0-31]. + // As an SDNode. + // we bit negate the value first + int32_t imm = ~(N->getSExtValue()); + return XformMskToBitPosU5Imm(imm); +}]>; + +def SETMEMIMM_SHORT : SDNodeXForm<imm, [{ + // Return the bit position we will set [0-15]. + // As an SDNode. + int16_t imm = N->getSExtValue(); + return XformMskToBitPosU4Imm(imm); +}]>; + +def CLRMEMIMM_SHORT : SDNodeXForm<imm, [{ + // Return the bit position we will clear [0-15]. + // As an SDNode. + // we bit negate the value first + int16_t imm = ~(N->getSExtValue()); + return XformMskToBitPosU4Imm(imm); +}]>; + +def SETMEMIMM_BYTE : SDNodeXForm<imm, [{ + // Return the bit position we will set [0-7]. + // As an SDNode. + int8_t imm = N->getSExtValue(); + return XformMskToBitPosU3Imm(imm); +}]>; + +def CLRMEMIMM_BYTE : SDNodeXForm<imm, [{ + // Return the bit position we will clear [0-7]. + // As an SDNode. + // we bit negate the value first + int8_t imm = ~(N->getSExtValue()); + return XformMskToBitPosU3Imm(imm); +}]>; + //===----------------------------------------------------------------------===// -// MEMOP: Word -// -// Implemented: -// MEMw_ADDi_indexed_V4 : memw(Rs+#u6:2)+=#U5 -// MEMw_SUBi_indexed_V4 : memw(Rs+#u6:2)-=#U5 -// MEMw_ADDr_indexed_V4 : memw(Rs+#u6:2)+=Rt -// MEMw_SUBr_indexed_V4 : memw(Rs+#u6:2)-=Rt -// MEMw_CLRr_indexed_V4 : memw(Rs+#u6:2)&=Rt -// MEMw_SETr_indexed_V4 : memw(Rs+#u6:2)|=Rt -// MEMw_ADDi_V4 : memw(Rs+#u6:2)+=#U5 -// MEMw_SUBi_V4 : memw(Rs+#u6:2)-=#U5 -// MEMw_ADDr_V4 : memw(Rs+#u6:2)+=Rt -// MEMw_SUBr_V4 : memw(Rs+#u6:2)-=Rt -// MEMw_CLRr_V4 : memw(Rs+#u6:2)&=Rt -// MEMw_SETr_V4 : memw(Rs+#u6:2)|=Rt -// -// Not implemented: -// MEMw_CLRi_indexed_V4 : memw(Rs+#u6:2)=clrbit(#U5) -// MEMw_SETi_indexed_V4 : memw(Rs+#u6:2)=setbit(#U5) -// MEMw_CLRi_V4 : memw(Rs+#u6:2)=clrbit(#U5) -// MEMw_SETi_V4 : memw(Rs+#u6:2)=setbit(#U5) +// Template class for MemOp instructions with the register value. //===----------------------------------------------------------------------===// +class MemOp_rr_base <string opc, bits<2> opcBits, Operand ImmOp, + string memOp, bits<2> memOpBits> : + MEMInst_V4<(outs), + (ins IntRegs:$base, ImmOp:$offset, IntRegs:$delta), + opc#"($base+#$offset)"#memOp#"$delta", + []>, + Requires<[HasV4T, UseMEMOP]> { + + bits<5> base; + bits<5> delta; + bits<32> offset; + bits<6> offsetBits; // memb - u6:0 , memh - u6:1, memw - u6:2 + + let offsetBits = !if (!eq(opcBits, 0b00), offset{5-0}, + !if (!eq(opcBits, 0b01), offset{6-1}, + !if (!eq(opcBits, 0b10), offset{7-2},0))); + + let IClass = 0b0011; + let Inst{27-24} = 0b1110; + let Inst{22-21} = opcBits; + let Inst{20-16} = base; + let Inst{13} = 0b0; + let Inst{12-7} = offsetBits; + let Inst{6-5} = memOpBits; + let Inst{4-0} = delta; +} +//===----------------------------------------------------------------------===// +// Template class for MemOp instructions with the immediate value. +//===----------------------------------------------------------------------===// +class MemOp_ri_base <string opc, bits<2> opcBits, Operand ImmOp, + string memOp, bits<2> memOpBits> : + MEMInst_V4 <(outs), + (ins IntRegs:$base, ImmOp:$offset, u5Imm:$delta), + opc#"($base+#$offset)"#memOp#"#$delta" + #!if(memOpBits{1},")", ""), // clrbit, setbit - include ')' + []>, + Requires<[HasV4T, UseMEMOP]> { + bits<5> base; + bits<5> delta; + bits<32> offset; + bits<6> offsetBits; // memb - u6:0 , memh - u6:1, memw - u6:2 -// memw(Rs+#u6:2) += #U5 -let AddedComplexity = 30 in -def MEMw_ADDi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, u5Imm:$addend), - "memw($base+#$offset) += #$addend", - []>, - Requires<[HasV4T, UseMEMOP]>; + let offsetBits = !if (!eq(opcBits, 0b00), offset{5-0}, + !if (!eq(opcBits, 0b01), offset{6-1}, + !if (!eq(opcBits, 0b10), offset{7-2},0))); -// memw(Rs+#u6:2) -= #U5 -let AddedComplexity = 30 in -def MEMw_SUBi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, u5Imm:$subend), - "memw($base+#$offset) -= #$subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) += Rt -let AddedComplexity = 30 in -def MEMw_ADDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, IntRegs:$addend), - "memw($base+#$offset) += $addend", - [(store (add (load (add (i32 IntRegs:$base), u6_2ImmPred:$offset)), - (i32 IntRegs:$addend)), - (add (i32 IntRegs:$base), u6_2ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) -= Rt -let AddedComplexity = 30 in -def MEMw_SUBr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, IntRegs:$subend), - "memw($base+#$offset) -= $subend", - [(store (sub (load (add (i32 IntRegs:$base), u6_2ImmPred:$offset)), - (i32 IntRegs:$subend)), - (add (i32 IntRegs:$base), u6_2ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) &= Rt -let AddedComplexity = 30 in -def MEMw_ANDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, IntRegs:$andend), - "memw($base+#$offset) &= $andend", - [(store (and (load (add (i32 IntRegs:$base), u6_2ImmPred:$offset)), - (i32 IntRegs:$andend)), - (add (i32 IntRegs:$base), u6_2ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) |= Rt -let AddedComplexity = 30 in -def MEMw_ORr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_2Imm:$offset, IntRegs:$orend), - "memw($base+#$offset) |= $orend", - [(store (or (load (add (i32 IntRegs:$base), u6_2ImmPred:$offset)), - (i32 IntRegs:$orend)), - (add (i32 IntRegs:$base), u6_2ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) += #U5 -let AddedComplexity = 30 in -def MEMw_ADDi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$addend), - "memw($addr) += $addend", - []>, - Requires<[HasV4T, UseMEMOP]>; + let IClass = 0b0011; + let Inst{27-24} = 0b1111; + let Inst{22-21} = opcBits; + let Inst{20-16} = base; + let Inst{13} = 0b0; + let Inst{12-7} = offsetBits; + let Inst{6-5} = memOpBits; + let Inst{4-0} = delta; +} -// memw(Rs+#u6:2) -= #U5 -let AddedComplexity = 30 in -def MEMw_SUBi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$subend), - "memw($addr) -= $subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) += Rt -let AddedComplexity = 30 in -def MEMw_ADDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$addend), - "memw($addr) += $addend", - [(store (add (load ADDRriU6_2:$addr), (i32 IntRegs:$addend)), - ADDRriU6_2:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) -= Rt -let AddedComplexity = 30 in -def MEMw_SUBr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$subend), - "memw($addr) -= $subend", - [(store (sub (load ADDRriU6_2:$addr), (i32 IntRegs:$subend)), - ADDRriU6_2:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) &= Rt -let AddedComplexity = 30 in -def MEMw_ANDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$andend), - "memw($addr) &= $andend", - [(store (and (load ADDRriU6_2:$addr), (i32 IntRegs:$andend)), - ADDRriU6_2:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memw(Rs+#u6:2) |= Rt -let AddedComplexity = 30 in -def MEMw_ORr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$orend), - "memw($addr) |= $orend", - [(store (or (load ADDRriU6_2:$addr), (i32 IntRegs:$orend)), - ADDRriU6_2:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; +// multiclass to define MemOp instructions with register operand. +multiclass MemOp_rr<string opc, bits<2> opcBits, Operand ImmOp> { + def _ADD#NAME#_V4 : MemOp_rr_base <opc, opcBits, ImmOp, " += ", 0b00>; // add + def _SUB#NAME#_V4 : MemOp_rr_base <opc, opcBits, ImmOp, " -= ", 0b01>; // sub + def _AND#NAME#_V4 : MemOp_rr_base <opc, opcBits, ImmOp, " &= ", 0b10>; // and + def _OR#NAME#_V4 : MemOp_rr_base <opc, opcBits, ImmOp, " |= ", 0b11>; // or +} + +// multiclass to define MemOp instructions with immediate Operand. +multiclass MemOp_ri<string opc, bits<2> opcBits, Operand ImmOp> { + def _ADD#NAME#_V4 : MemOp_ri_base <opc, opcBits, ImmOp, " += ", 0b00 >; + def _SUB#NAME#_V4 : MemOp_ri_base <opc, opcBits, ImmOp, " -= ", 0b01 >; + def _CLRBIT#NAME#_V4 : MemOp_ri_base<opc, opcBits, ImmOp, " =clrbit(", 0b10>; + def _SETBIT#NAME#_V4 : MemOp_ri_base<opc, opcBits, ImmOp, " =setbit(", 0b11>; +} + +multiclass MemOp_base <string opc, bits<2> opcBits, Operand ImmOp> { + defm r : MemOp_rr <opc, opcBits, ImmOp>; + defm i : MemOp_ri <opc, opcBits, ImmOp>; +} + +// Define MemOp instructions. +let isExtendable = 1, opExtendable = 1, isExtentSigned = 0, +validSubTargets =HasV4SubT in { + let opExtentBits = 6, accessSize = ByteAccess in + defm MemOPb : MemOp_base <"memb", 0b00, u6_0Ext>; + + let opExtentBits = 7, accessSize = HalfWordAccess in + defm MemOPh : MemOp_base <"memh", 0b01, u6_1Ext>; + + let opExtentBits = 8, accessSize = WordAccess in + defm MemOPw : MemOp_base <"memw", 0b10, u6_2Ext>; +} //===----------------------------------------------------------------------===// -// MEMOP: Halfword -// -// Implemented: -// MEMh_ADDi_indexed_V4 : memw(Rs+#u6:2)+=#U5 -// MEMh_SUBi_indexed_V4 : memw(Rs+#u6:2)-=#U5 -// MEMh_ADDr_indexed_V4 : memw(Rs+#u6:2)+=Rt -// MEMh_SUBr_indexed_V4 : memw(Rs+#u6:2)-=Rt -// MEMh_CLRr_indexed_V4 : memw(Rs+#u6:2)&=Rt -// MEMh_SETr_indexed_V4 : memw(Rs+#u6:2)|=Rt -// MEMh_ADDi_V4 : memw(Rs+#u6:2)+=#U5 -// MEMh_SUBi_V4 : memw(Rs+#u6:2)-=#U5 -// MEMh_ADDr_V4 : memw(Rs+#u6:2)+=Rt -// MEMh_SUBr_V4 : memw(Rs+#u6:2)-=Rt -// MEMh_CLRr_V4 : memw(Rs+#u6:2)&=Rt -// MEMh_SETr_V4 : memw(Rs+#u6:2)|=Rt -// -// Not implemented: -// MEMh_CLRi_indexed_V4 : memw(Rs+#u6:2)=clrbit(#U5) -// MEMh_SETi_indexed_V4 : memw(Rs+#u6:2)=setbit(#U5) -// MEMh_CLRi_V4 : memw(Rs+#u6:2)=clrbit(#U5) -// MEMh_SETi_V4 : memw(Rs+#u6:2)=setbit(#U5) +// Multiclass to define 'Def Pats' for ALU operations on the memory +// Here value used for the ALU operation is an immediate value. +// mem[bh](Rs+#0) += #U5 +// mem[bh](Rs+#u6) += #U5 //===----------------------------------------------------------------------===// +multiclass MemOpi_u5Pats <PatFrag ldOp, PatFrag stOp, PatLeaf ExtPred, + InstHexagon MI, SDNode OpNode> { + let AddedComplexity = 180 in + def : Pat < (stOp (OpNode (ldOp IntRegs:$addr), u5ImmPred:$addend), + IntRegs:$addr), + (MI IntRegs:$addr, #0, u5ImmPred:$addend )>; -// memh(Rs+#u6:1) += #U5 -let AddedComplexity = 30 in -def MEMh_ADDi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, u5Imm:$addend), - "memh($base+#$offset) += $addend", - []>, - Requires<[HasV4T, UseMEMOP]>; + let AddedComplexity = 190 in + def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, ExtPred:$offset)), + u5ImmPred:$addend), + (add IntRegs:$base, ExtPred:$offset)), + (MI IntRegs:$base, ExtPred:$offset, u5ImmPred:$addend)>; +} -// memh(Rs+#u6:1) -= #U5 -let AddedComplexity = 30 in -def MEMh_SUBi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, u5Imm:$subend), - "memh($base+#$offset) -= $subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) += Rt -let AddedComplexity = 30 in -def MEMh_ADDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, IntRegs:$addend), - "memh($base+#$offset) += $addend", - [(truncstorei16 (add (sextloadi16 (add (i32 IntRegs:$base), - u6_1ImmPred:$offset)), - (i32 IntRegs:$addend)), - (add (i32 IntRegs:$base), u6_1ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) -= Rt -let AddedComplexity = 30 in -def MEMh_SUBr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, IntRegs:$subend), - "memh($base+#$offset) -= $subend", - [(truncstorei16 (sub (sextloadi16 (add (i32 IntRegs:$base), - u6_1ImmPred:$offset)), - (i32 IntRegs:$subend)), - (add (i32 IntRegs:$base), u6_1ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) &= Rt -let AddedComplexity = 30 in -def MEMh_ANDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, IntRegs:$andend), - "memh($base+#$offset) += $andend", - [(truncstorei16 (and (sextloadi16 (add (i32 IntRegs:$base), - u6_1ImmPred:$offset)), - (i32 IntRegs:$andend)), - (add (i32 IntRegs:$base), u6_1ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) |= Rt -let AddedComplexity = 30 in -def MEMh_ORr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_1Imm:$offset, IntRegs:$orend), - "memh($base+#$offset) |= $orend", - [(truncstorei16 (or (sextloadi16 (add (i32 IntRegs:$base), - u6_1ImmPred:$offset)), - (i32 IntRegs:$orend)), - (add (i32 IntRegs:$base), u6_1ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) += #U5 -let AddedComplexity = 30 in -def MEMh_ADDi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$addend), - "memh($addr) += $addend", - []>, - Requires<[HasV4T, UseMEMOP]>; +multiclass MemOpi_u5ALUOp<PatFrag ldOp, PatFrag stOp, PatLeaf ExtPred, + InstHexagon addMI, InstHexagon subMI> { + defm : MemOpi_u5Pats<ldOp, stOp, ExtPred, addMI, add>; + defm : MemOpi_u5Pats<ldOp, stOp, ExtPred, subMI, sub>; +} -// memh(Rs+#u6:1) -= #U5 -let AddedComplexity = 30 in -def MEMh_SUBi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$subend), - "memh($addr) -= $subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) += Rt -let AddedComplexity = 30 in -def MEMh_ADDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$addend), - "memh($addr) += $addend", - [(truncstorei16 (add (sextloadi16 ADDRriU6_1:$addr), - (i32 IntRegs:$addend)), ADDRriU6_1:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) -= Rt -let AddedComplexity = 30 in -def MEMh_SUBr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$subend), - "memh($addr) -= $subend", - [(truncstorei16 (sub (sextloadi16 ADDRriU6_1:$addr), - (i32 IntRegs:$subend)), ADDRriU6_1:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) &= Rt -let AddedComplexity = 30 in -def MEMh_ANDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$andend), - "memh($addr) &= $andend", - [(truncstorei16 (and (sextloadi16 ADDRriU6_1:$addr), - (i32 IntRegs:$andend)), ADDRriU6_1:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memh(Rs+#u6:1) |= Rt -let AddedComplexity = 30 in -def MEMh_ORr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$orend), - "memh($addr) |= $orend", - [(truncstorei16 (or (sextloadi16 ADDRriU6_1:$addr), - (i32 IntRegs:$orend)), ADDRriU6_1:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; +multiclass MemOpi_u5ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > { + // Half Word + defm : MemOpi_u5ALUOp <ldOpHalf, truncstorei16, u6_1ExtPred, + MemOPh_ADDi_V4, MemOPh_SUBi_V4>; + // Byte + defm : MemOpi_u5ALUOp <ldOpByte, truncstorei8, u6ExtPred, + MemOPb_ADDi_V4, MemOPb_SUBi_V4>; +} + +let Predicates = [HasV4T, UseMEMOP] in { + defm : MemOpi_u5ExtType<zextloadi8, zextloadi16>; // zero extend + defm : MemOpi_u5ExtType<sextloadi8, sextloadi16>; // sign extend + defm : MemOpi_u5ExtType<extloadi8, extloadi16>; // any extend + // Word + defm : MemOpi_u5ALUOp <load, store, u6_2ExtPred, MemOPw_ADDi_V4, + MemOPw_SUBi_V4>; +} //===----------------------------------------------------------------------===// -// MEMOP: Byte -// -// Implemented: -// MEMb_ADDi_indexed_V4 : memb(Rs+#u6:0)+=#U5 -// MEMb_SUBi_indexed_V4 : memb(Rs+#u6:0)-=#U5 -// MEMb_ADDr_indexed_V4 : memb(Rs+#u6:0)+=Rt -// MEMb_SUBr_indexed_V4 : memb(Rs+#u6:0)-=Rt -// MEMb_CLRr_indexed_V4 : memb(Rs+#u6:0)&=Rt -// MEMb_SETr_indexed_V4 : memb(Rs+#u6:0)|=Rt -// MEMb_ADDi_V4 : memb(Rs+#u6:0)+=#U5 -// MEMb_SUBi_V4 : memb(Rs+#u6:0)-=#U5 -// MEMb_ADDr_V4 : memb(Rs+#u6:0)+=Rt -// MEMb_SUBr_V4 : memb(Rs+#u6:0)-=Rt -// MEMb_CLRr_V4 : memb(Rs+#u6:0)&=Rt -// MEMb_SETr_V4 : memb(Rs+#u6:0)|=Rt -// -// Not implemented: -// MEMb_CLRi_indexed_V4 : memb(Rs+#u6:0)=clrbit(#U5) -// MEMb_SETi_indexed_V4 : memb(Rs+#u6:0)=setbit(#U5) -// MEMb_CLRi_V4 : memb(Rs+#u6:0)=clrbit(#U5) -// MEMb_SETi_V4 : memb(Rs+#u6:0)=setbit(#U5) +// multiclass to define 'Def Pats' for ALU operations on the memory. +// Here value used for the ALU operation is a negative value. +// mem[bh](Rs+#0) += #m5 +// mem[bh](Rs+#u6) += #m5 //===----------------------------------------------------------------------===// -// memb(Rs+#u6:0) += #U5 -let AddedComplexity = 30 in -def MEMb_ADDi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, u5Imm:$addend), - "memb($base+#$offset) += $addend", - []>, - Requires<[HasV4T, UseMEMOP]>; +multiclass MemOpi_m5Pats <PatFrag ldOp, PatFrag stOp, PatLeaf extPred, + PatLeaf immPred, ComplexPattern addrPred, + SDNodeXForm xformFunc, InstHexagon MI> { + let AddedComplexity = 190 in + def : Pat <(stOp (add (ldOp IntRegs:$addr), immPred:$subend), + IntRegs:$addr), + (MI IntRegs:$addr, #0, (xformFunc immPred:$subend) )>; -// memb(Rs+#u6:0) -= #U5 -let AddedComplexity = 30 in -def MEMb_SUBi_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, u5Imm:$subend), - "memb($base+#$offset) -= $subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) += Rt -let AddedComplexity = 30 in -def MEMb_ADDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, IntRegs:$addend), - "memb($base+#$offset) += $addend", - [(truncstorei8 (add (sextloadi8 (add (i32 IntRegs:$base), - u6_0ImmPred:$offset)), - (i32 IntRegs:$addend)), - (add (i32 IntRegs:$base), u6_0ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) -= Rt -let AddedComplexity = 30 in -def MEMb_SUBr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, IntRegs:$subend), - "memb($base+#$offset) -= $subend", - [(truncstorei8 (sub (sextloadi8 (add (i32 IntRegs:$base), - u6_0ImmPred:$offset)), - (i32 IntRegs:$subend)), - (add (i32 IntRegs:$base), u6_0ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) &= Rt -let AddedComplexity = 30 in -def MEMb_ANDr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, IntRegs:$andend), - "memb($base+#$offset) += $andend", - [(truncstorei8 (and (sextloadi8 (add (i32 IntRegs:$base), - u6_0ImmPred:$offset)), - (i32 IntRegs:$andend)), - (add (i32 IntRegs:$base), u6_0ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) |= Rt -let AddedComplexity = 30 in -def MEMb_ORr_indexed_MEM_V4 : MEMInst_V4<(outs), - (ins IntRegs:$base, u6_0Imm:$offset, IntRegs:$orend), - "memb($base+#$offset) |= $orend", - [(truncstorei8 (or (sextloadi8 (add (i32 IntRegs:$base), - u6_0ImmPred:$offset)), - (i32 IntRegs:$orend)), - (add (i32 IntRegs:$base), u6_0ImmPred:$offset))]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) += #U5 -let AddedComplexity = 30 in -def MEMb_ADDi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$addend), - "memb($addr) += $addend", - []>, - Requires<[HasV4T, UseMEMOP]>; + let AddedComplexity = 195 in + def : Pat<(stOp (add (ldOp (add IntRegs:$base, extPred:$offset)), + immPred:$subend), + (add IntRegs:$base, extPred:$offset)), + (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$subend))>; +} -// memb(Rs+#u6:0) -= #U5 -let AddedComplexity = 30 in -def MEMb_SUBi_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, u5Imm:$subend), - "memb($addr) -= $subend", - []>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) += Rt -let AddedComplexity = 30 in -def MEMb_ADDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$addend), - "memb($addr) += $addend", - [(truncstorei8 (add (sextloadi8 ADDRriU6_0:$addr), - (i32 IntRegs:$addend)), ADDRriU6_0:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) -= Rt -let AddedComplexity = 30 in -def MEMb_SUBr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$subend), - "memb($addr) -= $subend", - [(truncstorei8 (sub (sextloadi8 ADDRriU6_0:$addr), - (i32 IntRegs:$subend)), ADDRriU6_0:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) &= Rt -let AddedComplexity = 30 in -def MEMb_ANDr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$andend), - "memb($addr) &= $andend", - [(truncstorei8 (and (sextloadi8 ADDRriU6_0:$addr), - (i32 IntRegs:$andend)), ADDRriU6_0:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; - -// memb(Rs+#u6:0) |= Rt -let AddedComplexity = 30 in -def MEMb_ORr_MEM_V4 : MEMInst_V4<(outs), - (ins MEMri:$addr, IntRegs:$orend), - "memb($addr) |= $orend", - [(truncstorei8 (or (sextloadi8 ADDRriU6_0:$addr), - (i32 IntRegs:$orend)), ADDRriU6_0:$addr)]>, - Requires<[HasV4T, UseMEMOP]>; +multiclass MemOpi_m5ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > { + // Half Word + defm : MemOpi_m5Pats <ldOpHalf, truncstorei16, u6_1ExtPred, m5HImmPred, + ADDRriU6_1, MEMOPIMM_HALF, MemOPh_SUBi_V4>; + // Byte + defm : MemOpi_m5Pats <ldOpByte, truncstorei8, u6ExtPred, m5BImmPred, + ADDRriU6_0, MEMOPIMM_BYTE, MemOPb_SUBi_V4>; +} +let Predicates = [HasV4T, UseMEMOP] in { + defm : MemOpi_m5ExtType<zextloadi8, zextloadi16>; // zero extend + defm : MemOpi_m5ExtType<sextloadi8, sextloadi16>; // sign extend + defm : MemOpi_m5ExtType<extloadi8, extloadi16>; // any extend + + // Word + defm : MemOpi_m5Pats <load, store, u6_2ExtPred, m5ImmPred, + ADDRriU6_2, MEMOPIMM, MemOPw_SUBi_V4>; +} + +//===----------------------------------------------------------------------===// +// Multiclass to define 'def Pats' for bit operations on the memory. +// mem[bhw](Rs+#0) = [clrbit|setbit](#U5) +// mem[bhw](Rs+#u6) = [clrbit|setbit](#U5) +//===----------------------------------------------------------------------===// + +multiclass MemOpi_bitPats <PatFrag ldOp, PatFrag stOp, PatLeaf immPred, + PatLeaf extPred, ComplexPattern addrPred, + SDNodeXForm xformFunc, InstHexagon MI, SDNode OpNode> { + + // mem[bhw](Rs+#u6:[012]) = [clrbit|setbit](#U5) + let AddedComplexity = 250 in + def : Pat<(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)), + immPred:$bitend), + (add IntRegs:$base, extPred:$offset)), + (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$bitend))>; + + // mem[bhw](Rs+#0) = [clrbit|setbit](#U5) + let AddedComplexity = 225 in + def : Pat <(stOp (OpNode (ldOp addrPred:$addr), immPred:$bitend), + addrPred:$addr), + (MI IntRegs:$addr, #0, (xformFunc immPred:$bitend))>; +} + +multiclass MemOpi_bitExtType<PatFrag ldOpByte, PatFrag ldOpHalf > { + // Byte - clrbit + defm : MemOpi_bitPats<ldOpByte, truncstorei8, Clr3ImmPred, u6ExtPred, + ADDRriU6_0, CLRMEMIMM_BYTE, MemOPb_CLRBITi_V4, and>; + // Byte - setbit + defm : MemOpi_bitPats<ldOpByte, truncstorei8, Set3ImmPred, u6ExtPred, + ADDRriU6_0, SETMEMIMM_BYTE, MemOPb_SETBITi_V4, or>; + // Half Word - clrbit + defm : MemOpi_bitPats<ldOpHalf, truncstorei16, Clr4ImmPred, u6_1ExtPred, + ADDRriU6_1, CLRMEMIMM_SHORT, MemOPh_CLRBITi_V4, and>; + // Half Word - setbit + defm : MemOpi_bitPats<ldOpHalf, truncstorei16, Set4ImmPred, u6_1ExtPred, + ADDRriU6_1, SETMEMIMM_SHORT, MemOPh_SETBITi_V4, or>; +} + +let Predicates = [HasV4T, UseMEMOP] in { + // mem[bh](Rs+#0) = [clrbit|setbit](#U5) + // mem[bh](Rs+#u6:[01]) = [clrbit|setbit](#U5) + defm : MemOpi_bitExtType<zextloadi8, zextloadi16>; // zero extend + defm : MemOpi_bitExtType<sextloadi8, sextloadi16>; // sign extend + defm : MemOpi_bitExtType<extloadi8, extloadi16>; // any extend + + // memw(Rs+#0) = [clrbit|setbit](#U5) + // memw(Rs+#u6:2) = [clrbit|setbit](#U5) + defm : MemOpi_bitPats<load, store, Clr5ImmPred, u6_2ExtPred, ADDRriU6_2, + CLRMEMIMM, MemOPw_CLRBITi_V4, and>; + defm : MemOpi_bitPats<load, store, Set5ImmPred, u6_2ExtPred, ADDRriU6_2, + SETMEMIMM, MemOPw_SETBITi_V4, or>; +} + +//===----------------------------------------------------------------------===// +// Multiclass to define 'def Pats' for ALU operations on the memory +// where addend is a register. +// mem[bhw](Rs+#0) [+-&|]= Rt +// mem[bhw](Rs+#U6:[012]) [+-&|]= Rt +//===----------------------------------------------------------------------===// + +multiclass MemOpr_Pats <PatFrag ldOp, PatFrag stOp, ComplexPattern addrPred, + PatLeaf extPred, InstHexagon MI, SDNode OpNode> { + let AddedComplexity = 141 in + // mem[bhw](Rs+#0) [+-&|]= Rt + def : Pat <(stOp (OpNode (ldOp addrPred:$addr), (i32 IntRegs:$addend)), + addrPred:$addr), + (MI IntRegs:$addr, #0, (i32 IntRegs:$addend) )>; + + // mem[bhw](Rs+#U6:[012]) [+-&|]= Rt + let AddedComplexity = 150 in + def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)), + (i32 IntRegs:$orend)), + (add IntRegs:$base, extPred:$offset)), + (MI IntRegs:$base, extPred:$offset, (i32 IntRegs:$orend) )>; +} + +multiclass MemOPr_ALUOp<PatFrag ldOp, PatFrag stOp, + ComplexPattern addrPred, PatLeaf extPred, + InstHexagon addMI, InstHexagon subMI, + InstHexagon andMI, InstHexagon orMI > { + + defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, addMI, add>; + defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, subMI, sub>; + defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, andMI, and>; + defm : MemOpr_Pats <ldOp, stOp, addrPred, extPred, orMI, or>; +} + +multiclass MemOPr_ExtType<PatFrag ldOpByte, PatFrag ldOpHalf > { + // Half Word + defm : MemOPr_ALUOp <ldOpHalf, truncstorei16, ADDRriU6_1, u6_1ExtPred, + MemOPh_ADDr_V4, MemOPh_SUBr_V4, + MemOPh_ANDr_V4, MemOPh_ORr_V4>; + // Byte + defm : MemOPr_ALUOp <ldOpByte, truncstorei8, ADDRriU6_0, u6ExtPred, + MemOPb_ADDr_V4, MemOPb_SUBr_V4, + MemOPb_ANDr_V4, MemOPb_ORr_V4>; +} + +// Define 'def Pats' for MemOps with register addend. +let Predicates = [HasV4T, UseMEMOP] in { + // Byte, Half Word + defm : MemOPr_ExtType<zextloadi8, zextloadi16>; // zero extend + defm : MemOPr_ExtType<sextloadi8, sextloadi16>; // sign extend + defm : MemOPr_ExtType<extloadi8, extloadi16>; // any extend + // Word + defm : MemOPr_ALUOp <load, store, ADDRriU6_2, u6_2ExtPred, MemOPw_ADDr_V4, + MemOPw_SUBr_V4, MemOPw_ANDr_V4, MemOPw_ORr_V4 >; +} //===----------------------------------------------------------------------===// // XTYPE/PRED + @@ -3146,7 +2145,7 @@ def CMPbEQri_V4 : MInst<(outs PredRegs:$dst), def : Pat <(brcond (i1 (setne (and (i32 IntRegs:$src1), 255), u8ImmPred:$src2)), bb:$offset), - (JMP_cNot (CMPbEQri_V4 (i32 IntRegs:$src1), u8ImmPred:$src2), + (JMP_f (CMPbEQri_V4 (i32 IntRegs:$src1), u8ImmPred:$src2), bb:$offset)>, Requires<[HasV4T]>; @@ -3629,9 +2628,9 @@ let isReturn = 1, isTerminator = 1, multiclass ST_Abs_Predbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_V4 : STInst2<(outs), - (ins PredRegs:$src1, globaladdressExt:$absaddr, RC: $src2), + (ins PredRegs:$src1, u0AlwaysExt:$absaddr, RC: $src2), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", ") ")#mnemonic#"(##$absaddr) = $src2", []>, @@ -3639,7 +2638,7 @@ multiclass ST_Abs_Predbase<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_Abs_Pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Abs_Predbase<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Abs_Predbase<mnemonic, RC, PredNot, 1>; @@ -3651,7 +2650,7 @@ multiclass ST_Abs<string mnemonic, string CextOp, RegisterClass RC> { let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in { let opExtendable = 0, isPredicable = 1 in def NAME#_V4 : STInst2<(outs), - (ins globaladdressExt:$absaddr, RC:$src), + (ins u0AlwaysExt:$absaddr, RC:$src), mnemonic#"(##$absaddr) = $src", []>, Requires<[HasV4T]>; @@ -3665,9 +2664,9 @@ multiclass ST_Abs<string mnemonic, string CextOp, RegisterClass RC> { multiclass ST_Abs_Predbase_nv<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME#_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, globaladdressExt:$absaddr, RC: $src2), + (ins PredRegs:$src1, u0AlwaysExt:$absaddr, RC: $src2), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", ") ")#mnemonic#"(##$absaddr) = $src2.new", []>, @@ -3675,7 +2674,7 @@ multiclass ST_Abs_Predbase_nv<string mnemonic, RegisterClass RC, bit isNot, } multiclass ST_Abs_Pred_nv<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : ST_Abs_Predbase_nv<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : ST_Abs_Predbase_nv<mnemonic, RC, PredNot, 1>; @@ -3687,7 +2686,7 @@ multiclass ST_Abs_nv<string mnemonic, string CextOp, RegisterClass RC> { let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in { let opExtendable = 0, isPredicable = 1 in def NAME#_nv_V4 : NVInst_V4<(outs), - (ins globaladdressExt:$absaddr, RC:$src), + (ins u0AlwaysExt:$absaddr, RC:$src), mnemonic#"(##$absaddr) = $src.new", []>, Requires<[HasV4T]>; @@ -3700,16 +2699,19 @@ multiclass ST_Abs_nv<string mnemonic, string CextOp, RegisterClass RC> { } let addrMode = Absolute in { + let accessSize = ByteAccess in defm STrib_abs : ST_Abs<"memb", "STrib", IntRegs>, ST_Abs_nv<"memb", "STrib", IntRegs>, AddrModeRel; + let accessSize = HalfWordAccess in defm STrih_abs : ST_Abs<"memh", "STrih", IntRegs>, ST_Abs_nv<"memh", "STrih", IntRegs>, AddrModeRel; + let accessSize = WordAccess in defm STriw_abs : ST_Abs<"memw", "STriw", IntRegs>, ST_Abs_nv<"memw", "STriw", IntRegs>, AddrModeRel; - let isNVStorable = 0 in + let accessSize = DoubleWordAccess, isNVStorable = 0 in defm STrid_abs : ST_Abs<"memd", "STrid", DoubleRegs>, AddrModeRel; } @@ -3730,11 +2732,115 @@ def : Pat<(store (i64 DoubleRegs:$src1), (STrid_abs_V4 tglobaladdr: $absaddr, DoubleRegs: $src1)>; } +//===----------------------------------------------------------------------===// +// multiclass for store instructions with GP-relative addressing mode. +// mem[bhwd](#global)=Rt +// if ([!]Pv[.new]) mem[bhwd](##global) = Rt +//===----------------------------------------------------------------------===// +let mayStore = 1, isNVStorable = 1 in +multiclass ST_GP<string mnemonic, string BaseOp, RegisterClass RC> { + let BaseOpcode = BaseOp, isPredicable = 1 in + def NAME#_V4 : STInst2<(outs), + (ins globaladdress:$global, RC:$src), + mnemonic#"(#$global) = $src", + []>; + + // When GP-relative instructions are predicated, their addressing mode is + // changed to absolute and they are always constant extended. + let BaseOpcode = BaseOp, isExtended = 1, opExtendable = 1, + isPredicated = 1 in { + defm Pt : ST_Abs_Pred <mnemonic, RC, 0>; + defm NotPt : ST_Abs_Pred <mnemonic, RC, 1>; + } +} + +let mayStore = 1, isNVStore = 1 in +multiclass ST_GP_nv<string mnemonic, string BaseOp, RegisterClass RC> { + let BaseOpcode = BaseOp, isPredicable = 1 in + def NAME#_nv_V4 : NVInst_V4<(outs), + (ins u0AlwaysExt:$global, RC:$src), + mnemonic#"(#$global) = $src.new", + []>, + Requires<[HasV4T]>; + + // When GP-relative instructions are predicated, their addressing mode is + // changed to absolute and they are always constant extended. + let BaseOpcode = BaseOp, isExtended = 1, opExtendable = 1, + isPredicated = 1 in { + defm Pt : ST_Abs_Pred_nv<mnemonic, RC, 0>; + defm NotPt : ST_Abs_Pred_nv<mnemonic, RC, 1>; + } +} + +let validSubTargets = HasV4SubT, neverHasSideEffects = 1 in { + let isNVStorable = 0 in + defm STd_GP : ST_GP <"memd", "STd_GP", DoubleRegs>, PredNewRel; + + defm STb_GP : ST_GP<"memb", "STb_GP", IntRegs>, + ST_GP_nv<"memb", "STb_GP", IntRegs>, NewValueRel; + defm STh_GP : ST_GP<"memh", "STh_GP", IntRegs>, + ST_GP_nv<"memh", "STh_GP", IntRegs>, NewValueRel; + defm STw_GP : ST_GP<"memw", "STw_GP", IntRegs>, + ST_GP_nv<"memw", "STw_GP", IntRegs>, NewValueRel; +} + +// 64 bit atomic store +def : Pat <(atomic_store_64 (HexagonCONST32_GP tglobaladdr:$global), + (i64 DoubleRegs:$src1)), + (STd_GP_V4 tglobaladdr:$global, (i64 DoubleRegs:$src1))>, + Requires<[HasV4T]>; + +// Map from store(globaladdress) -> memd(#foo) +let AddedComplexity = 100 in +def : Pat <(store (i64 DoubleRegs:$src1), + (HexagonCONST32_GP tglobaladdr:$global)), + (STd_GP_V4 tglobaladdr:$global, (i64 DoubleRegs:$src1))>; + +// 8 bit atomic store +def : Pat < (atomic_store_8 (HexagonCONST32_GP tglobaladdr:$global), + (i32 IntRegs:$src1)), + (STb_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +// Map from store(globaladdress) -> memb(#foo) +let AddedComplexity = 100 in +def : Pat<(truncstorei8 (i32 IntRegs:$src1), + (HexagonCONST32_GP tglobaladdr:$global)), + (STb_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +// Map from "i1 = constant<-1>; memw(CONST32(#foo)) = i1" +// to "r0 = 1; memw(#foo) = r0" +let AddedComplexity = 100 in +def : Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)), + (STb_GP_V4 tglobaladdr:$global, (TFRI 1))>; + +def : Pat<(atomic_store_16 (HexagonCONST32_GP tglobaladdr:$global), + (i32 IntRegs:$src1)), + (STh_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +// Map from store(globaladdress) -> memh(#foo) +let AddedComplexity = 100 in +def : Pat<(truncstorei16 (i32 IntRegs:$src1), + (HexagonCONST32_GP tglobaladdr:$global)), + (STh_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +// 32 bit atomic store +def : Pat<(atomic_store_32 (HexagonCONST32_GP tglobaladdr:$global), + (i32 IntRegs:$src1)), + (STw_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +// Map from store(globaladdress) -> memw(#foo) +let AddedComplexity = 100 in +def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), + (STw_GP_V4 tglobaladdr:$global, (i32 IntRegs:$src1))>; + +//===----------------------------------------------------------------------===// +// Multiclass for the load instructions with absolute addressing mode. +//===----------------------------------------------------------------------===// multiclass LD_Abs_Predbase<string mnemonic, RegisterClass RC, bit isNot, bit isPredNew> { - let PNewValue = !if(isPredNew, "new", "") in + let isPredicatedNew = isPredNew in def NAME : LDInst2<(outs RC:$dst), - (ins PredRegs:$src1, globaladdressExt:$absaddr), + (ins PredRegs:$src1, u0AlwaysExt:$absaddr), !if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ", ") ")#"$dst = "#mnemonic#"(##$absaddr)", []>, @@ -3742,7 +2848,7 @@ multiclass LD_Abs_Predbase<string mnemonic, RegisterClass RC, bit isNot, } multiclass LD_Abs_Pred<string mnemonic, RegisterClass RC, bit PredNot> { - let PredSense = !if(PredNot, "false", "true") in { + let isPredicatedFalse = PredNot in { defm _c#NAME : LD_Abs_Predbase<mnemonic, RC, PredNot, 0>; // Predicate new defm _cdn#NAME : LD_Abs_Predbase<mnemonic, RC, PredNot, 1>; @@ -3754,7 +2860,7 @@ multiclass LD_Abs<string mnemonic, string CextOp, RegisterClass RC> { let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in { let opExtendable = 1, isPredicable = 1 in def NAME#_V4 : LDInst2<(outs RC:$dst), - (ins globaladdressExt:$absaddr), + (ins u0AlwaysExt:$absaddr), "$dst = "#mnemonic#"(##$absaddr)", []>, Requires<[HasV4T]>; @@ -3767,33 +2873,138 @@ multiclass LD_Abs<string mnemonic, string CextOp, RegisterClass RC> { } let addrMode = Absolute in { + let accessSize = ByteAccess in { defm LDrib_abs : LD_Abs<"memb", "LDrib", IntRegs>, AddrModeRel; defm LDriub_abs : LD_Abs<"memub", "LDriub", IntRegs>, AddrModeRel; + } + let accessSize = HalfWordAccess in { defm LDrih_abs : LD_Abs<"memh", "LDrih", IntRegs>, AddrModeRel; defm LDriuh_abs : LD_Abs<"memuh", "LDriuh", IntRegs>, AddrModeRel; + } + let accessSize = WordAccess in defm LDriw_abs : LD_Abs<"memw", "LDriw", IntRegs>, AddrModeRel; + + let accessSize = DoubleWordAccess in defm LDrid_abs : LD_Abs<"memd", "LDrid", DoubleRegs>, AddrModeRel; } -let Predicates = [HasV4T], AddedComplexity = 30 in +let Predicates = [HasV4T], AddedComplexity = 30 in { def : Pat<(i32 (load (HexagonCONST32 tglobaladdr:$absaddr))), (LDriw_abs_V4 tglobaladdr: $absaddr)>; -let Predicates = [HasV4T], AddedComplexity=30 in def : Pat<(i32 (sextloadi8 (HexagonCONST32 tglobaladdr:$absaddr))), (LDrib_abs_V4 tglobaladdr:$absaddr)>; -let Predicates = [HasV4T], AddedComplexity=30 in def : Pat<(i32 (zextloadi8 (HexagonCONST32 tglobaladdr:$absaddr))), (LDriub_abs_V4 tglobaladdr:$absaddr)>; -let Predicates = [HasV4T], AddedComplexity=30 in def : Pat<(i32 (sextloadi16 (HexagonCONST32 tglobaladdr:$absaddr))), (LDrih_abs_V4 tglobaladdr:$absaddr)>; -let Predicates = [HasV4T], AddedComplexity=30 in def : Pat<(i32 (zextloadi16 (HexagonCONST32 tglobaladdr:$absaddr))), (LDriuh_abs_V4 tglobaladdr:$absaddr)>; +} + +//===----------------------------------------------------------------------===// +// multiclass for load instructions with GP-relative addressing mode. +// Rx=mem[bhwd](##global) +// if ([!]Pv[.new]) Rx=mem[bhwd](##global) +//===----------------------------------------------------------------------===// +let neverHasSideEffects = 1, validSubTargets = HasV4SubT in +multiclass LD_GP<string mnemonic, string BaseOp, RegisterClass RC> { + let BaseOpcode = BaseOp in { + let isPredicable = 1 in + def NAME#_V4 : LDInst2<(outs RC:$dst), + (ins globaladdress:$global), + "$dst = "#mnemonic#"(#$global)", + []>; + + let isExtended = 1, opExtendable = 2, isPredicated = 1 in { + defm Pt_V4 : LD_Abs_Pred<mnemonic, RC, 0>; + defm NotPt_V4 : LD_Abs_Pred<mnemonic, RC, 1>; + } + } +} + +defm LDd_GP : LD_GP<"memd", "LDd_GP", DoubleRegs>, PredNewRel; +defm LDb_GP : LD_GP<"memb", "LDb_GP", IntRegs>, PredNewRel; +defm LDub_GP : LD_GP<"memub", "LDub_GP", IntRegs>, PredNewRel; +defm LDh_GP : LD_GP<"memh", "LDh_GP", IntRegs>, PredNewRel; +defm LDuh_GP : LD_GP<"memuh", "LDuh_GP", IntRegs>, PredNewRel; +defm LDw_GP : LD_GP<"memw", "LDw_GP", IntRegs>, PredNewRel; + +def : Pat <(atomic_load_64 (HexagonCONST32_GP tglobaladdr:$global)), + (i64 (LDd_GP_V4 tglobaladdr:$global))>; + +def : Pat <(atomic_load_32 (HexagonCONST32_GP tglobaladdr:$global)), + (i32 (LDw_GP_V4 tglobaladdr:$global))>; + +def : Pat <(atomic_load_16 (HexagonCONST32_GP tglobaladdr:$global)), + (i32 (LDuh_GP_V4 tglobaladdr:$global))>; + +def : Pat <(atomic_load_8 (HexagonCONST32_GP tglobaladdr:$global)), + (i32 (LDub_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memw(#foo + 0) +let AddedComplexity = 100 in +def : Pat <(i64 (load (HexagonCONST32_GP tglobaladdr:$global))), + (i64 (LDd_GP_V4 tglobaladdr:$global))>; + +// Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd +let AddedComplexity = 100 in +def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))), + (i1 (TFR_PdRs (i32 (LDb_GP_V4 tglobaladdr:$global))))>; + +// When the Interprocedural Global Variable optimizer realizes that a certain +// global variable takes only two constant values, it shrinks the global to +// a boolean. Catch those loads here in the following 3 patterns. +let AddedComplexity = 100 in +def : Pat <(i32 (extloadi1 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDb_GP_V4 tglobaladdr:$global))>; + +let AddedComplexity = 100 in +def : Pat <(i32 (sextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDb_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memb(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (extloadi8 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDb_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memb(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (sextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDb_GP_V4 tglobaladdr:$global))>; + +let AddedComplexity = 100 in +def : Pat <(i32 (zextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDub_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memub(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (zextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDub_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memh(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (extloadi16 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDh_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memh(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDh_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memuh(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDuh_GP_V4 tglobaladdr:$global))>; + +// Map from load(globaladdress) -> memw(#foo) +let AddedComplexity = 100 in +def : Pat <(i32 (load (HexagonCONST32_GP tglobaladdr:$global))), + (i32 (LDw_GP_V4 tglobaladdr:$global))>; + // Transfer global address into a register let AddedComplexity=50, isMoveImm = 1, isReMaterializable = 1 in @@ -3842,19 +3053,21 @@ def : Pat<(HexagonCONST32_GP tglobaladdr:$src1), // Load - Indirect with long offset: These instructions take global address // as an operand -let AddedComplexity = 10 in +let isExtended = 1, opExtendable = 3, AddedComplexity = 40, +validSubTargets = HasV4SubT in def LDrid_ind_lo_V4 : LDInst<(outs DoubleRegs:$dst), - (ins IntRegs:$src1, u2Imm:$src2, globaladdress:$offset), + (ins IntRegs:$src1, u2Imm:$src2, globaladdressExt:$offset), "$dst=memd($src1<<#$src2+##$offset)", [(set (i64 DoubleRegs:$dst), (load (add (shl IntRegs:$src1, u2ImmPred:$src2), (HexagonCONST32 tglobaladdr:$offset))))]>, Requires<[HasV4T]>; -let AddedComplexity = 10 in +let AddedComplexity = 40 in multiclass LD_indirect_lo<string OpcStr, PatFrag OpNode> { +let isExtended = 1, opExtendable = 3, validSubTargets = HasV4SubT in def _lo_V4 : LDInst<(outs IntRegs:$dst), - (ins IntRegs:$src1, u2Imm:$src2, globaladdress:$offset), + (ins IntRegs:$src1, u2Imm:$src2, globaladdressExt:$offset), !strconcat("$dst = ", !strconcat(OpcStr, "($src1<<#$src2+##$offset)")), [(set IntRegs:$dst, @@ -3865,202 +3078,53 @@ multiclass LD_indirect_lo<string OpcStr, PatFrag OpNode> { defm LDrib_ind : LD_indirect_lo<"memb", sextloadi8>; defm LDriub_ind : LD_indirect_lo<"memub", zextloadi8>; +defm LDriub_ind_anyext : LD_indirect_lo<"memub", extloadi8>; defm LDrih_ind : LD_indirect_lo<"memh", sextloadi16>; defm LDriuh_ind : LD_indirect_lo<"memuh", zextloadi16>; +defm LDriuh_ind_anyext : LD_indirect_lo<"memuh", extloadi16>; defm LDriw_ind : LD_indirect_lo<"memw", load>; -// Store - Indirect with long offset: These instructions take global address -// as an operand -let AddedComplexity = 10 in -def STrid_ind_lo_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, globaladdress:$src3, - DoubleRegs:$src4), - "memd($src1<<#$src2+#$src3) = $src4", - [(store (i64 DoubleRegs:$src4), - (add (shl IntRegs:$src1, u2ImmPred:$src2), - (HexagonCONST32 tglobaladdr:$src3)))]>, - Requires<[HasV4T]>; - -let AddedComplexity = 10 in -multiclass ST_indirect_lo<string OpcStr, PatFrag OpNode> { - def _lo_V4 : STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, globaladdress:$src3, - IntRegs:$src4), - !strconcat(OpcStr, "($src1<<#$src2+##$src3) = $src4"), - [(OpNode (i32 IntRegs:$src4), - (add (shl IntRegs:$src1, u2ImmPred:$src2), - (HexagonCONST32 tglobaladdr:$src3)))]>, - Requires<[HasV4T]>; -} - -defm STrib_ind : ST_indirect_lo<"memb", truncstorei8>; -defm STrih_ind : ST_indirect_lo<"memh", truncstorei16>; -defm STriw_ind : ST_indirect_lo<"memw", store>; - -// Store - absolute addressing mode: These instruction take constant -// value as the extended operand. -multiclass ST_absimm<string OpcStr> { -let isExtended = 1, opExtendable = 0, isPredicable = 1, -validSubTargets = HasV4SubT in - def _abs_V4 : STInst2<(outs), - (ins u0AlwaysExt:$src1, IntRegs:$src2), - !strconcat(OpcStr, "(##$src1) = $src2"), - []>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 1, isPredicated = 1, -validSubTargets = HasV4SubT in { - def _abs_cPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if ($src1)", !strconcat(OpcStr, "(##$src2) = $src3")), - []>, - Requires<[HasV4T]>; - - def _abs_cNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if (!$src1)", !strconcat(OpcStr, "(##$src2) = $src3")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if ($src1.new)", - !strconcat(OpcStr, "(##$src2) = $src3")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnNotPt_V4 : STInst2<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if (!$src1.new)", - !strconcat(OpcStr, "(##$src2) = $src3")), - []>, - Requires<[HasV4T]>; -} - -let isExtended = 1, opExtendable = 0, mayStore = 1, isNVStore = 1, -validSubTargets = HasV4SubT in - def _abs_nv_V4 : NVInst_V4<(outs), - (ins u0AlwaysExt:$src1, IntRegs:$src2), - !strconcat(OpcStr, "(##$src1) = $src2.new"), - []>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 1, mayStore = 1, isPredicated = 1, -isNVStore = 1, validSubTargets = HasV4SubT in { - def _abs_cPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if ($src1)", - !strconcat(OpcStr, "(##$src2) = $src3.new")), - []>, - Requires<[HasV4T]>; - - def _abs_cNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if (!$src1)", - !strconcat(OpcStr, "(##$src2) = $src3.new")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if ($src1.new)", - !strconcat(OpcStr, "(##$src2) = $src3.new")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnNotPt_nv_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, u0AlwaysExt:$src2, IntRegs:$src3), - !strconcat("if (!$src1.new)", - !strconcat(OpcStr, "(##$src2) = $src3.new")), - []>, - Requires<[HasV4T]>; -} -} +let AddedComplexity = 40 in +def : Pat <(i32 (sextloadi8 (add IntRegs:$src1, + (NumUsesBelowThresCONST32 tglobaladdr:$offset)))), + (i32 (LDrib_ind_lo_V4 IntRegs:$src1, 0, tglobaladdr:$offset))>, + Requires<[HasV4T]>; -defm STrib_imm : ST_absimm<"memb">; -defm STrih_imm : ST_absimm<"memh">; -defm STriw_imm : ST_absimm<"memw">; +let AddedComplexity = 40 in +def : Pat <(i32 (zextloadi8 (add IntRegs:$src1, + (NumUsesBelowThresCONST32 tglobaladdr:$offset)))), + (i32 (LDriub_ind_lo_V4 IntRegs:$src1, 0, tglobaladdr:$offset))>, + Requires<[HasV4T]>; let Predicates = [HasV4T], AddedComplexity = 30 in { def : Pat<(truncstorei8 (i32 IntRegs:$src1), u0AlwaysExtPred:$src2), - (STrib_imm_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; + (STrib_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; def : Pat<(truncstorei16 (i32 IntRegs:$src1), u0AlwaysExtPred:$src2), - (STrih_imm_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; + (STrih_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; def : Pat<(store (i32 IntRegs:$src1), u0AlwaysExtPred:$src2), - (STriw_imm_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; -} - -// Load - absolute addressing mode: These instruction take constant -// value as the extended operand - -multiclass LD_absimm<string OpcStr> { -let isExtended = 1, opExtendable = 1, isPredicable = 1, -validSubTargets = HasV4SubT in - def _abs_V4 : LDInst2<(outs IntRegs:$dst), - (ins u0AlwaysExt:$src), - !strconcat("$dst = ", - !strconcat(OpcStr, "(##$src)")), - []>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 2, isPredicated = 1, -validSubTargets = HasV4SubT in { - def _abs_cPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, u0AlwaysExt:$src2), - !strconcat("if ($src1) $dst = ", - !strconcat(OpcStr, "(##$src2)")), - []>, - Requires<[HasV4T]>; - - def _abs_cNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, u0AlwaysExt:$src2), - !strconcat("if (!$src1) $dst = ", - !strconcat(OpcStr, "(##$src2)")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, u0AlwaysExt:$src2), - !strconcat("if ($src1.new) $dst = ", - !strconcat(OpcStr, "(##$src2)")), - []>, - Requires<[HasV4T]>; - - def _abs_cdnNotPt_V4 : LDInst2<(outs IntRegs:$dst), - (ins PredRegs:$src1, u0AlwaysExt:$src2), - !strconcat("if (!$src1.new) $dst = ", - !strconcat(OpcStr, "(##$src2)")), - []>, - Requires<[HasV4T]>; + (STriw_abs_V4 u0AlwaysExtPred:$src2, IntRegs: $src1)>; } -} - -defm LDrib_imm : LD_absimm<"memb">; -defm LDriub_imm : LD_absimm<"memub">; -defm LDrih_imm : LD_absimm<"memh">; -defm LDriuh_imm : LD_absimm<"memuh">; -defm LDriw_imm : LD_absimm<"memw">; let Predicates = [HasV4T], AddedComplexity = 30 in { def : Pat<(i32 (load u0AlwaysExtPred:$src)), - (LDriw_imm_abs_V4 u0AlwaysExtPred:$src)>; + (LDriw_abs_V4 u0AlwaysExtPred:$src)>; def : Pat<(i32 (sextloadi8 u0AlwaysExtPred:$src)), - (LDrib_imm_abs_V4 u0AlwaysExtPred:$src)>; + (LDrib_abs_V4 u0AlwaysExtPred:$src)>; def : Pat<(i32 (zextloadi8 u0AlwaysExtPred:$src)), - (LDriub_imm_abs_V4 u0AlwaysExtPred:$src)>; + (LDriub_abs_V4 u0AlwaysExtPred:$src)>; def : Pat<(i32 (sextloadi16 u0AlwaysExtPred:$src)), - (LDrih_imm_abs_V4 u0AlwaysExtPred:$src)>; + (LDrih_abs_V4 u0AlwaysExtPred:$src)>; def : Pat<(i32 (zextloadi16 u0AlwaysExtPred:$src)), - (LDriuh_imm_abs_V4 u0AlwaysExtPred:$src)>; + (LDriuh_abs_V4 u0AlwaysExtPred:$src)>; } -// Indexed store double word - global address. +// Indexed store word - global address. // memw(Rs+#u6:2)=#S8 let AddedComplexity = 10 in def STriw_offset_ext_V4 : STInst<(outs), diff --git a/lib/Target/Hexagon/HexagonMCInst.h b/lib/Target/Hexagon/HexagonMCInst.h deleted file mode 100644 index e16636e..0000000 --- a/lib/Target/Hexagon/HexagonMCInst.h +++ /dev/null @@ -1,41 +0,0 @@ -//===- HexagonMCInst.h - Hexagon sub-class of MCInst ----------------------===// -// -// The LLVM Compiler Infrastructure -// -// This file is distributed under the University of Illinois Open Source -// License. See LICENSE.TXT for details. -// -//===----------------------------------------------------------------------===// -// -// This class extends MCInst to allow some VLIW annotation. -// -//===----------------------------------------------------------------------===// - -#ifndef HEXAGONMCINST_H -#define HEXAGONMCINST_H - -#include "llvm/CodeGen/MachineInstr.h" -#include "llvm/MC/MCInst.h" - -namespace llvm { - class HexagonMCInst: public MCInst { - // Packet start and end markers - unsigned startPacket: 1, endPacket: 1; - const MachineInstr *MachineI; - public: - explicit HexagonMCInst(): MCInst(), - startPacket(0), endPacket(0) {} - - const MachineInstr* getMI() const { return MachineI; } - - void setMI(const MachineInstr *MI) { MachineI = MI; } - - bool isStartPacket() const { return (startPacket); } - bool isEndPacket() const { return (endPacket); } - - void setStartPacket(bool yes) { startPacket = yes; } - void setEndPacket(bool yes) { endPacket = yes; } - }; -} - -#endif diff --git a/lib/Target/Hexagon/HexagonMachineFunctionInfo.h b/lib/Target/Hexagon/HexagonMachineFunctionInfo.h index 0318c51..bd7b26a 100644 --- a/lib/Target/Hexagon/HexagonMachineFunctionInfo.h +++ b/lib/Target/Hexagon/HexagonMachineFunctionInfo.h @@ -29,15 +29,18 @@ class HexagonMachineFunctionInfo : public MachineFunctionInfo { std::vector<MachineInstr*> AllocaAdjustInsts; int VarArgsFrameIndex; bool HasClobberLR; + bool HasEHReturn; std::map<const MachineInstr*, unsigned> PacketInfo; public: - HexagonMachineFunctionInfo() : SRetReturnReg(0), HasClobberLR(0) {} + HexagonMachineFunctionInfo() : SRetReturnReg(0), HasClobberLR(0), + HasEHReturn(false) {} HexagonMachineFunctionInfo(MachineFunction &MF) : SRetReturnReg(0), - HasClobberLR(0) {} + HasClobberLR(0), + HasEHReturn(false) {} unsigned getSRetReturnReg() const { return SRetReturnReg; } void setSRetReturnReg(unsigned Reg) { SRetReturnReg = Reg; } @@ -69,6 +72,8 @@ public: void setHasClobberLR(bool v) { HasClobberLR = v; } bool hasClobberLR() const { return HasClobberLR; } + bool hasEHReturn() const { return HasEHReturn; }; + void setHasEHReturn(bool H = true) { HasEHReturn = H; }; }; } // End llvm namespace diff --git a/lib/Target/Hexagon/HexagonNewValueJump.cpp b/lib/Target/Hexagon/HexagonNewValueJump.cpp index cd3d289..72af876 100644 --- a/lib/Target/Hexagon/HexagonNewValueJump.cpp +++ b/lib/Target/Hexagon/HexagonNewValueJump.cpp @@ -68,6 +68,7 @@ namespace { HexagonNewValueJump() : MachineFunctionPass(ID) { } virtual void getAnalysisUsage(AnalysisUsage &AU) const { + AU.addRequired<MachineBranchProbabilityInfo>(); MachineFunctionPass::getAnalysisUsage(AU); } @@ -78,6 +79,8 @@ namespace { virtual bool runOnMachineFunction(MachineFunction &Fn); private: + /// \brief A handle to the branch probability pass. + const MachineBranchProbabilityInfo *MBPI; }; @@ -208,19 +211,15 @@ static bool canCompareBeNewValueJump(const HexagonInstrInfo *QII, // range specified by the arch. if (!secondReg) { int64_t v = MI->getOperand(2).getImm(); - if (MI->getOpcode() == Hexagon::CMPGEri || - (MI->getOpcode() == Hexagon::CMPGEUri && v > 0)) - --v; if (!(isUInt<5>(v) || ((MI->getOpcode() == Hexagon::CMPEQri || - MI->getOpcode() == Hexagon::CMPGTri || - MI->getOpcode() == Hexagon::CMPGEri) && + MI->getOpcode() == Hexagon::CMPGTri) && (v == -1)))) return false; } - unsigned cmpReg1, cmpOp2 = 0; // cmpOp2 assignment silences compiler warning. + unsigned cmpReg1, cmpOp2; cmpReg1 = MI->getOperand(1).getReg(); if (secondReg) { @@ -271,58 +270,58 @@ static bool canCompareBeNewValueJump(const HexagonInstrInfo *QII, // Given a compare operator, return a matching New Value Jump // compare operator. Make sure that MI here is included in // HexagonInstrInfo.cpp::isNewValueJumpCandidate -static unsigned getNewValueJumpOpcode(const MachineInstr *MI, int reg, - bool secondRegNewified) { +static unsigned getNewValueJumpOpcode(MachineInstr *MI, int reg, + bool secondRegNewified, + MachineBasicBlock *jmpTarget, + const MachineBranchProbabilityInfo + *MBPI) { + bool taken = false; + MachineBasicBlock *Src = MI->getParent(); + const BranchProbability Prediction = + MBPI->getEdgeProbability(Src, jmpTarget); + + if (Prediction >= BranchProbability(1,2)) + taken = true; + switch (MI->getOpcode()) { case Hexagon::CMPEQrr: - return Hexagon::JMP_EQrrPt_nv_V4; + return taken ? Hexagon::JMP_EQrrPt_nv_V4 : Hexagon::JMP_EQrrPnt_nv_V4; case Hexagon::CMPEQri: { if (reg >= 0) - return Hexagon::JMP_EQriPt_nv_V4; + return taken ? Hexagon::JMP_EQriPt_nv_V4 : Hexagon::JMP_EQriPnt_nv_V4; else - return Hexagon::JMP_EQriPtneg_nv_V4; + return taken ? Hexagon::JMP_EQriPtneg_nv_V4 + : Hexagon::JMP_EQriPntneg_nv_V4; } - case Hexagon::CMPLTrr: case Hexagon::CMPGTrr: { if (secondRegNewified) - return Hexagon::JMP_GTrrdnPt_nv_V4; + return taken ? Hexagon::JMP_GTrrdnPt_nv_V4 + : Hexagon::JMP_GTrrdnPnt_nv_V4; else - return Hexagon::JMP_GTrrPt_nv_V4; - } - - case Hexagon::CMPGEri: { - if (reg >= 1) - return Hexagon::JMP_GTriPt_nv_V4; - else - return Hexagon::JMP_GTriPtneg_nv_V4; + return taken ? Hexagon::JMP_GTrrPt_nv_V4 + : Hexagon::JMP_GTrrPnt_nv_V4; } case Hexagon::CMPGTri: { if (reg >= 0) - return Hexagon::JMP_GTriPt_nv_V4; + return taken ? Hexagon::JMP_GTriPt_nv_V4 : Hexagon::JMP_GTriPnt_nv_V4; else - return Hexagon::JMP_GTriPtneg_nv_V4; + return taken ? Hexagon::JMP_GTriPtneg_nv_V4 + : Hexagon::JMP_GTriPntneg_nv_V4; } - case Hexagon::CMPLTUrr: case Hexagon::CMPGTUrr: { if (secondRegNewified) - return Hexagon::JMP_GTUrrdnPt_nv_V4; + return taken ? Hexagon::JMP_GTUrrdnPt_nv_V4 + : Hexagon::JMP_GTUrrdnPnt_nv_V4; else - return Hexagon::JMP_GTUrrPt_nv_V4; + return taken ? Hexagon::JMP_GTUrrPt_nv_V4 : Hexagon::JMP_GTUrrPnt_nv_V4; } case Hexagon::CMPGTUri: - return Hexagon::JMP_GTUriPt_nv_V4; - - case Hexagon::CMPGEUri: { - if (reg == 0) - return Hexagon::JMP_EQrrPt_nv_V4; - else - return Hexagon::JMP_GTUriPt_nv_V4; - } + return taken ? Hexagon::JMP_GTUriPt_nv_V4 : Hexagon::JMP_GTUriPnt_nv_V4; default: llvm_unreachable("Could not find matching New Value Jump instruction."); @@ -346,6 +345,7 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { QII = static_cast<const HexagonInstrInfo *>(MF.getTarget().getInstrInfo()); QRI = static_cast<const HexagonRegisterInfo *>(MF.getTarget().getRegisterInfo()); + MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); if (!QRI->Subtarget.hasV4TOps() || DisableNewValueJumps) { @@ -393,12 +393,12 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { DEBUG(dbgs() << "Instr: "; MI->dump(); dbgs() << "\n"); if (!foundJump && - (MI->getOpcode() == Hexagon::JMP_c || - MI->getOpcode() == Hexagon::JMP_cNot || - MI->getOpcode() == Hexagon::JMP_cdnPt || - MI->getOpcode() == Hexagon::JMP_cdnPnt || - MI->getOpcode() == Hexagon::JMP_cdnNotPt || - MI->getOpcode() == Hexagon::JMP_cdnNotPnt)) { + (MI->getOpcode() == Hexagon::JMP_t || + MI->getOpcode() == Hexagon::JMP_f || + MI->getOpcode() == Hexagon::JMP_tnew_t || + MI->getOpcode() == Hexagon::JMP_tnew_nt || + MI->getOpcode() == Hexagon::JMP_fnew_t || + MI->getOpcode() == Hexagon::JMP_fnew_nt)) { // This is where you would insert your compare and // instr that feeds compare jmpPos = MII; @@ -434,9 +434,9 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { jmpTarget = MI->getOperand(1).getMBB(); foundJump = true; - if (MI->getOpcode() == Hexagon::JMP_cNot || - MI->getOpcode() == Hexagon::JMP_cdnNotPt || - MI->getOpcode() == Hexagon::JMP_cdnNotPnt) { + if (MI->getOpcode() == Hexagon::JMP_f || + MI->getOpcode() == Hexagon::JMP_fnew_t || + MI->getOpcode() == Hexagon::JMP_fnew_nt) { invertPredicate = true; } continue; @@ -525,10 +525,8 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { if (isSecondOpReg) { // In case of CMPLT, or CMPLTU, or EQ with the second register // to newify, swap the operands. - if (cmpInstr->getOpcode() == Hexagon::CMPLTrr || - cmpInstr->getOpcode() == Hexagon::CMPLTUrr || - (cmpInstr->getOpcode() == Hexagon::CMPEQrr && - feederReg == (unsigned) cmpOp2)) { + if (cmpInstr->getOpcode() == Hexagon::CMPEQrr && + feederReg == (unsigned) cmpOp2) { unsigned tmp = cmpReg1; bool tmpIsKill = MO1IsKill; cmpReg1 = cmpOp2; @@ -582,42 +580,34 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { assert((QII->isNewValueJumpCandidate(cmpInstr)) && "This compare is not a New Value Jump candidate."); unsigned opc = getNewValueJumpOpcode(cmpInstr, cmpOp2, - isSecondOpNewified); + isSecondOpNewified, + jmpTarget, MBPI); if (invertPredicate) opc = QII->getInvertedPredicatedOpcode(opc); - // Manage the conversions from CMPGEUri to either CMPEQrr - // or CMPGTUri properly. See Arch spec for CMPGEUri instructions. - // This has to be after the getNewValueJumpOpcode function call as - // second operand of the compare could be modified in this logic. - if (cmpInstr->getOpcode() == Hexagon::CMPGEUri) { - if (cmpOp2 == 0) { - cmpOp2 = cmpReg1; - MO2IsKill = MO1IsKill; - isSecondOpReg = true; - } else - --cmpOp2; - } - - // Manage the conversions from CMPGEri to CMPGTUri properly. - // See Arch spec for CMPGEri instructions. - if (cmpInstr->getOpcode() == Hexagon::CMPGEri) - --cmpOp2; - - if (isSecondOpReg) { + if (isSecondOpReg) NewMI = BuildMI(*MBB, jmpPos, dl, QII->get(opc)) .addReg(cmpReg1, getKillRegState(MO1IsKill)) .addReg(cmpOp2, getKillRegState(MO2IsKill)) .addMBB(jmpTarget); - } - else { + + else if ((cmpInstr->getOpcode() == Hexagon::CMPEQri || + cmpInstr->getOpcode() == Hexagon::CMPGTri) && + cmpOp2 == -1 ) + // Corresponding new-value compare jump instructions don't have the + // operand for -1 immediate value. + NewMI = BuildMI(*MBB, jmpPos, dl, + QII->get(opc)) + .addReg(cmpReg1, getKillRegState(MO1IsKill)) + .addMBB(jmpTarget); + + else NewMI = BuildMI(*MBB, jmpPos, dl, QII->get(opc)) .addReg(cmpReg1, getKillRegState(MO1IsKill)) .addImm(cmpOp2) .addMBB(jmpTarget); - } assert(NewMI && "New Value Jump Instruction Not created!"); if (cmpInstr->getOperand(0).isReg() && diff --git a/lib/Target/Hexagon/HexagonPeephole.cpp b/lib/Target/Hexagon/HexagonPeephole.cpp index 576f1d7..6c4eb7e 100644 --- a/lib/Target/Hexagon/HexagonPeephole.cpp +++ b/lib/Target/Hexagon/HexagonPeephole.cpp @@ -73,6 +73,10 @@ static cl::opt<bool> DisableOptSZExt("disable-hexagon-optszext", cl::Hidden, cl::ZeroOrMore, cl::init(false), cl::desc("Disable Optimization of Sign/Zero Extends")); +static cl::opt<bool> DisableOptExtTo64("disable-hexagon-opt-ext-to-64", + cl::Hidden, cl::ZeroOrMore, cl::init(false), + cl::desc("Disable Optimization of extensions to i64.")); + namespace { struct HexagonPeephole : public MachineFunctionPass { const HexagonInstrInfo *QII; @@ -142,6 +146,21 @@ bool HexagonPeephole::runOnMachineFunction(MachineFunction &MF) { } } + // Look for %vreg170<def> = COMBINE_ir_V4 (0, %vreg169) + // %vreg170:DoublRegs, %vreg169:IntRegs + if (!DisableOptExtTo64 && + MI->getOpcode () == Hexagon::COMBINE_Ir_V4) { + assert (MI->getNumOperands() == 3); + MachineOperand &Dst = MI->getOperand(0); + MachineOperand &Src1 = MI->getOperand(1); + MachineOperand &Src2 = MI->getOperand(2); + if (Src1.getImm() != 0) + continue; + unsigned DstReg = Dst.getReg(); + unsigned SrcReg = Src2.getReg(); + PeepholeMap[DstReg] = SrcReg; + } + // Look for this sequence below // %vregDoubleReg1 = LSRd_ri %vregDoubleReg0, 32 // %vregIntReg = COPY %vregDoubleReg1:subreg_loreg. diff --git a/lib/Target/Hexagon/HexagonRegisterInfo.cpp b/lib/Target/Hexagon/HexagonRegisterInfo.cpp index f947dfc..d8b4e2f 100644 --- a/lib/Target/Hexagon/HexagonRegisterInfo.cpp +++ b/lib/Target/Hexagon/HexagonRegisterInfo.cpp @@ -14,25 +14,26 @@ #include "HexagonRegisterInfo.h" #include "Hexagon.h" -#include "HexagonMachineFunctionInfo.h" #include "HexagonSubtarget.h" #include "HexagonTargetMachine.h" +#include "HexagonMachineFunctionInfo.h" #include "llvm/ADT/BitVector.h" #include "llvm/ADT/STLExtras.h" -#include "llvm/CodeGen/MachineFrameInfo.h" +#include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFunctionPass.h" -#include "llvm/CodeGen/MachineInstrBuilder.h" +#include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/CodeGen/PseudoSourceValue.h" #include "llvm/CodeGen/RegisterScavenging.h" #include "llvm/IR/Function.h" #include "llvm/IR/Type.h" #include "llvm/MC/MachineLocation.h" -#include "llvm/Support/CommandLine.h" -#include "llvm/Support/ErrorHandling.h" #include "llvm/Target/TargetInstrInfo.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetOptions.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/ErrorHandling.h" using namespace llvm; @@ -215,28 +216,41 @@ void HexagonRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, MI.getOperand(FIOperandNum).ChangeToRegister(resReg, false, false,true); MI.getOperand(FIOperandNum+1).ChangeToImmediate(0); } else if (TII.isMemOp(&MI)) { - unsigned resReg = HEXAGON_RESERVED_REG_1; - if (!MFI.hasVarSizedObjects() && - TII.isValidOffset(MI.getOpcode(), (FrameSize+Offset))) { - MI.getOperand(FIOperandNum).ChangeToRegister(getStackRegister(), - false, false, true); - MI.getOperand(FIOperandNum+1).ChangeToImmediate(FrameSize+Offset); - } else if (!TII.isValidOffset(Hexagon::ADD_ri, Offset)) { - BuildMI(*MI.getParent(), II, MI.getDebugLoc(), - TII.get(Hexagon::CONST32_Int_Real), resReg).addImm(Offset); - BuildMI(*MI.getParent(), II, MI.getDebugLoc(), - TII.get(Hexagon::ADD_rr), - resReg).addReg(FrameReg).addReg(resReg); - MI.getOperand(FIOperandNum).ChangeToRegister(resReg, false, false, - true); - MI.getOperand(FIOperandNum+1).ChangeToImmediate(0); + // use the constant extender if the instruction provides it + // and we are V4TOps. + if (Subtarget.hasV4TOps()) { + if (TII.isConstExtended(&MI)) { + MI.getOperand(FIOperandNum).ChangeToRegister(FrameReg, false); + MI.getOperand(FIOperandNum+1).ChangeToImmediate(Offset); + TII.immediateExtend(&MI); + } else { + llvm_unreachable("Need to implement for memops"); + } } else { - BuildMI(*MI.getParent(), II, MI.getDebugLoc(), - TII.get(Hexagon::ADD_ri), - resReg).addReg(FrameReg).addImm(Offset); - MI.getOperand(FIOperandNum).ChangeToRegister(resReg, false, false, - true); - MI.getOperand(FIOperandNum+1).ChangeToImmediate(0); + // Only V3 and older instructions here. + unsigned ResReg = HEXAGON_RESERVED_REG_1; + if (!MFI.hasVarSizedObjects() && + TII.isValidOffset(MI.getOpcode(), (FrameSize+Offset))) { + MI.getOperand(FIOperandNum).ChangeToRegister(getStackRegister(), + false, false, false); + MI.getOperand(FIOperandNum+1).ChangeToImmediate(FrameSize+Offset); + } else if (!TII.isValidOffset(Hexagon::ADD_ri, Offset)) { + BuildMI(*MI.getParent(), II, MI.getDebugLoc(), + TII.get(Hexagon::CONST32_Int_Real), ResReg).addImm(Offset); + BuildMI(*MI.getParent(), II, MI.getDebugLoc(), + TII.get(Hexagon::ADD_rr), ResReg).addReg(FrameReg). + addReg(ResReg); + MI.getOperand(FIOperandNum).ChangeToRegister(ResReg, false, false, + true); + MI.getOperand(FIOperandNum+1).ChangeToImmediate(0); + } else { + BuildMI(*MI.getParent(), II, MI.getDebugLoc(), + TII.get(Hexagon::ADD_ri), ResReg).addReg(FrameReg). + addImm(Offset); + MI.getOperand(FIOperandNum).ChangeToRegister(ResReg, false, false, + true); + MI.getOperand(FIOperandNum+1).ChangeToImmediate(0); + } } } else { unsigned dstReg = MI.getOperand(0).getReg(); diff --git a/lib/Target/Hexagon/HexagonSubtarget.cpp b/lib/Target/Hexagon/HexagonSubtarget.cpp index 4bacb8f..07d5ce1 100644 --- a/lib/Target/Hexagon/HexagonSubtarget.cpp +++ b/lib/Target/Hexagon/HexagonSubtarget.cpp @@ -29,8 +29,16 @@ EnableV3("enable-hexagon-v3", cl::Hidden, static cl::opt<bool> EnableMemOps( "enable-hexagon-memops", - cl::Hidden, cl::ZeroOrMore, cl::ValueDisallowed, - cl::desc("Generate V4 memop instructions.")); + cl::Hidden, cl::ZeroOrMore, cl::ValueDisallowed, cl::init(true), + cl::desc( + "Generate V4 MEMOP in code generation for Hexagon target")); + +static cl::opt<bool> +DisableMemOps( + "disable-hexagon-memops", + cl::Hidden, cl::ZeroOrMore, cl::ValueDisallowed, cl::init(false), + cl::desc( + "Do not generate V4 MEMOP in code generation for Hexagon target")); static cl::opt<bool> EnableIEEERndNear( @@ -64,7 +72,10 @@ HexagonSubtarget::HexagonSubtarget(StringRef TT, StringRef CPU, StringRef FS): // Initialize scheduling itinerary for the specified CPU. InstrItins = getInstrItineraryForCPU(CPUString); - if (EnableMemOps) + // UseMemOps on by default unless disabled explicitly + if (DisableMemOps) + UseMemOps = false; + else if (EnableMemOps) UseMemOps = true; else UseMemOps = false; diff --git a/lib/Target/Hexagon/HexagonTargetMachine.cpp b/lib/Target/Hexagon/HexagonTargetMachine.cpp index d9fef3e..ce45c62 100644 --- a/lib/Target/Hexagon/HexagonTargetMachine.cpp +++ b/lib/Target/Hexagon/HexagonTargetMachine.cpp @@ -35,6 +35,10 @@ opt<bool> DisableHexagonMISched("disable-hexagon-misched", cl::Hidden, cl::ZeroOrMore, cl::init(false), cl::desc("Disable Hexagon MI Scheduling")); +static cl::opt<bool> DisableHexagonCFGOpt("disable-hexagon-cfgopt", + cl::Hidden, cl::ZeroOrMore, cl::init(false), + cl::desc("Disable Hexagon CFG Optimization")); + /// HexagonTargetMachineModule - Note that this is used on hosts that /// cannot link in a library unless there are references into the /// library. In particular, it seems that it is not possible to get @@ -75,19 +79,20 @@ HexagonTargetMachine::HexagonTargetMachine(const Target &T, StringRef TT, TSInfo(*this), FrameLowering(Subtarget), InstrItins(&Subtarget.getInstrItineraryData()) { - setMCUseCFI(false); + setMCUseCFI(false); } // addPassesForOptimizations - Allow the backend (target) to add Target // Independent Optimization passes to the Pass Manager. bool HexagonTargetMachine::addPassesForOptimizations(PassManagerBase &PM) { - - PM.add(createConstantPropagationPass()); - PM.add(createLoopSimplifyPass()); - PM.add(createDeadCodeEliminationPass()); - PM.add(createConstantPropagationPass()); - PM.add(createLoopUnrollPass()); - PM.add(createLoopStrengthReducePass()); + if (getOptLevel() != CodeGenOpt::None) { + PM.add(createConstantPropagationPass()); + PM.add(createLoopSimplifyPass()); + PM.add(createDeadCodeEliminationPass()); + PM.add(createConstantPropagationPass()); + PM.add(createLoopUnrollPass()); + PM.add(createLoopStrengthReducePass()); + } return true; } @@ -121,38 +126,45 @@ TargetPassConfig *HexagonTargetMachine::createPassConfig(PassManagerBase &PM) { } bool HexagonPassConfig::addInstSelector() { - addPass(createHexagonRemoveExtendOps(getHexagonTargetMachine())); + + if (getOptLevel() != CodeGenOpt::None) + addPass(createHexagonRemoveExtendOps(getHexagonTargetMachine())); + addPass(createHexagonISelDag(getHexagonTargetMachine(), getOptLevel())); - addPass(createHexagonPeephole()); + + if (getOptLevel() != CodeGenOpt::None) + addPass(createHexagonPeephole()); + return false; } bool HexagonPassConfig::addPreRegAlloc() { - if (!DisableHardwareLoops) { + if (!DisableHardwareLoops && getOptLevel() != CodeGenOpt::None) addPass(createHexagonHardwareLoops()); - } return false; } bool HexagonPassConfig::addPostRegAlloc() { - addPass(createHexagonCFGOptimizer(getHexagonTargetMachine())); + if (!DisableHexagonCFGOpt && getOptLevel() != CodeGenOpt::None) + addPass(createHexagonCFGOptimizer(getHexagonTargetMachine())); return true; } bool HexagonPassConfig::addPreSched2() { - addPass(&IfConverterID); + if (getOptLevel() != CodeGenOpt::None) + addPass(&IfConverterID); return true; } bool HexagonPassConfig::addPreEmitPass() { - if (!DisableHardwareLoops) { + if (!DisableHardwareLoops && getOptLevel() != CodeGenOpt::None) addPass(createHexagonFixupHwLoops()); - } - addPass(createHexagonNewValueJump()); + if (getOptLevel() != CodeGenOpt::None) + addPass(createHexagonNewValueJump()); // Expand Spill code for predicate registers. addPass(createHexagonExpandPredSpillCode(getHexagonTargetMachine())); @@ -161,7 +173,8 @@ bool HexagonPassConfig::addPreEmitPass() { addPass(createHexagonSplitTFRCondSets(getHexagonTargetMachine())); // Create Packets. - addPass(createHexagonPacketizer()); + if (getOptLevel() != CodeGenOpt::None) + addPass(createHexagonPacketizer()); return false; } diff --git a/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp b/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp index 866beb1..e592df9 100644 --- a/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp +++ b/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp @@ -48,19 +48,35 @@ #include "HexagonMachineFunctionInfo.h" #include <map> +#include <vector> using namespace llvm; +static cl::opt<bool> PacketizeVolatiles("hexagon-packetize-volatiles", + cl::ZeroOrMore, cl::Hidden, cl::init(true), + cl::desc("Allow non-solo packetization of volatile memory references")); + +extern cl::opt<bool> ScheduleInlineAsm; +extern cl::opt<bool> CountDeadOutput; + +namespace llvm { + void initializeHexagonPacketizerPass(PassRegistry&); +} + + namespace { class HexagonPacketizer : public MachineFunctionPass { public: static char ID; - HexagonPacketizer() : MachineFunctionPass(ID) {} + HexagonPacketizer() : MachineFunctionPass(ID) { + initializeHexagonPacketizerPass(*PassRegistry::getPassRegistry()); + } void getAnalysisUsage(AnalysisUsage &AU) const { AU.setPreservesCFG(); AU.addRequired<MachineDominatorTree>(); + AU.addRequired<MachineBranchProbabilityInfo>(); AU.addPreserved<MachineDominatorTree>(); AU.addRequired<MachineLoopInfo>(); AU.addPreserved<MachineLoopInfo>(); @@ -96,10 +112,17 @@ namespace { // schedule this instruction. bool FoundSequentialDependence; + /// \brief A handle to the branch probability pass. + const MachineBranchProbabilityInfo *MBPI; + + // Track MIs with ignored dependece. + std::vector<MachineInstr*> IgnoreDepMIs; + public: // Ctor. HexagonPacketizerList(MachineFunction &MF, MachineLoopInfo &MLI, - MachineDominatorTree &MDT); + MachineDominatorTree &MDT, + const MachineBranchProbabilityInfo *MBPI); // initPacketizerState - initialize some internal flags. void initPacketizerState(); @@ -123,20 +146,20 @@ namespace { private: bool IsCallDependent(MachineInstr* MI, SDep::Kind DepType, unsigned DepReg); bool PromoteToDotNew(MachineInstr* MI, SDep::Kind DepType, - MachineBasicBlock::iterator &MII, - const TargetRegisterClass* RC); + MachineBasicBlock::iterator &MII, + const TargetRegisterClass* RC); bool CanPromoteToDotNew(MachineInstr* MI, SUnit* PacketSU, - unsigned DepReg, - std::map <MachineInstr*, SUnit*> MIToSUnit, - MachineBasicBlock::iterator &MII, - const TargetRegisterClass* RC); + unsigned DepReg, + std::map <MachineInstr*, SUnit*> MIToSUnit, + MachineBasicBlock::iterator &MII, + const TargetRegisterClass* RC); bool CanPromoteToNewValue(MachineInstr* MI, SUnit* PacketSU, - unsigned DepReg, - std::map <MachineInstr*, SUnit*> MIToSUnit, - MachineBasicBlock::iterator &MII); + unsigned DepReg, + std::map <MachineInstr*, SUnit*> MIToSUnit, + MachineBasicBlock::iterator &MII); bool CanPromoteToNewValueStore(MachineInstr* MI, MachineInstr* PacketMI, - unsigned DepReg, - std::map <MachineInstr*, SUnit*> MIToSUnit); + unsigned DepReg, + std::map <MachineInstr*, SUnit*> MIToSUnit); bool DemoteToDotOld(MachineInstr* MI); bool ArePredicatesComplements(MachineInstr* MI1, MachineInstr* MI2, std::map <MachineInstr*, SUnit*> MIToSUnit); @@ -149,23 +172,34 @@ namespace { bool canReserveResourcesForConstExt(MachineInstr *MI); void reserveResourcesForConstExt(MachineInstr* MI); bool isNewValueInst(MachineInstr* MI); - bool isDotNewInst(MachineInstr* MI); }; } +INITIALIZE_PASS_BEGIN(HexagonPacketizer, "packets", "Hexagon Packetizer", + false, false) +INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) +INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) +INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) +INITIALIZE_PASS_END(HexagonPacketizer, "packets", "Hexagon Packetizer", + false, false) + + // HexagonPacketizerList Ctor. HexagonPacketizerList::HexagonPacketizerList( - MachineFunction &MF, MachineLoopInfo &MLI,MachineDominatorTree &MDT) + MachineFunction &MF, MachineLoopInfo &MLI,MachineDominatorTree &MDT, + const MachineBranchProbabilityInfo *MBPI) : VLIWPacketizerList(MF, MLI, MDT, true){ + this->MBPI = MBPI; } bool HexagonPacketizer::runOnMachineFunction(MachineFunction &Fn) { const TargetInstrInfo *TII = Fn.getTarget().getInstrInfo(); MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>(); MachineDominatorTree &MDT = getAnalysis<MachineDominatorTree>(); - + const MachineBranchProbabilityInfo *MBPI = + &getAnalysis<MachineBranchProbabilityInfo>(); // Instantiate the packetizer. - HexagonPacketizerList Packetizer(Fn, MLI, MDT); + HexagonPacketizerList Packetizer(Fn, MLI, MDT, MBPI); // DFA state table should not be empty. assert(Packetizer.getResourceTracker() && "Empty DFA table!"); @@ -711,8 +745,10 @@ static int GetDotNewOp(const int opc) { } // Return .new predicate version for an instruction -static int GetDotNewPredOp(const int opc) { - switch (opc) { +static int GetDotNewPredOp(MachineInstr *MI, + const MachineBranchProbabilityInfo *MBPI, + const HexagonInstrInfo *QII) { + switch (MI->getOpcode()) { default: llvm_unreachable("Unknown .new type"); // Conditional stores // Store byte conditionally @@ -858,17 +894,15 @@ static int GetDotNewPredOp(const int opc) { return Hexagon::STw_GP_cdnNotPt_V4; // Condtional Jumps - case Hexagon::JMP_c: - return Hexagon::JMP_cdnPt; - - case Hexagon::JMP_cNot: - return Hexagon::JMP_cdnNotPt; + case Hexagon::JMP_t: + case Hexagon::JMP_f: + return QII->getDotNewPredJumpOp(MI, MBPI); - case Hexagon::JMPR_cPt: - return Hexagon::JMPR_cdnPt_V3; + case Hexagon::JMPR_t: + return Hexagon::JMPR_tnew_tV3; - case Hexagon::JMPR_cNotPt: - return Hexagon::JMPR_cdnNotPt_V3; + case Hexagon::JMPR_f: + return Hexagon::JMPR_fnew_tV3; // Conditional Transfers case Hexagon::TFR_cPt: @@ -1262,7 +1296,7 @@ bool HexagonPacketizerList::PromoteToDotNew(MachineInstr* MI, int NewOpcode; if (RC == &Hexagon::PredRegsRegClass) - NewOpcode = GetDotNewPredOp(MI->getOpcode()); + NewOpcode = GetDotNewPredOp(MI, MBPI, QII); else NewOpcode = GetDotNewOp(MI->getOpcode()); MI->setDesc(QII->get(NewOpcode)); @@ -1307,17 +1341,17 @@ static int GetDotOldOp(const int opc) { case Hexagon::TFRI_cdnNotPt: return Hexagon::TFRI_cNotPt; - case Hexagon::JMP_cdnPt: - return Hexagon::JMP_c; + case Hexagon::JMP_tnew_t: + return Hexagon::JMP_t; - case Hexagon::JMP_cdnNotPt: - return Hexagon::JMP_cNot; + case Hexagon::JMP_fnew_t: + return Hexagon::JMP_f; - case Hexagon::JMPR_cdnPt_V3: - return Hexagon::JMPR_cPt; + case Hexagon::JMPR_tnew_tV3: + return Hexagon::JMPR_t; - case Hexagon::JMPR_cdnNotPt_V3: - return Hexagon::JMPR_cNotPt; + case Hexagon::JMPR_fnew_tV3: + return Hexagon::JMPR_f; // Load double word @@ -1913,7 +1947,7 @@ static bool GetPredicateSense(MachineInstr* MI, case Hexagon::STrih_imm_cdnPt_V4 : case Hexagon::STriw_imm_cPt_V4 : case Hexagon::STriw_imm_cdnPt_V4 : - case Hexagon::JMP_cdnPt : + case Hexagon::JMP_tnew_t : case Hexagon::LDrid_cPt : case Hexagon::LDrid_cdnPt : case Hexagon::LDrid_indexed_cPt : @@ -2052,7 +2086,7 @@ static bool GetPredicateSense(MachineInstr* MI, case Hexagon::STrih_imm_cdnNotPt_V4 : case Hexagon::STriw_imm_cNotPt_V4 : case Hexagon::STriw_imm_cdnNotPt_V4 : - case Hexagon::JMP_cdnNotPt : + case Hexagon::JMP_fnew_t : case Hexagon::LDrid_cNotPt : case Hexagon::LDrid_cdnNotPt : case Hexagon::LDrid_indexed_cNotPt : @@ -2154,172 +2188,6 @@ static bool GetPredicateSense(MachineInstr* MI, return false; } -bool HexagonPacketizerList::isDotNewInst(MachineInstr* MI) { - const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; - if (QII->isNewValueInst(MI)) - return true; - - switch (MI->getOpcode()) { - case Hexagon::TFR_cdnNotPt: - case Hexagon::TFR_cdnPt: - case Hexagon::TFRI_cdnNotPt: - case Hexagon::TFRI_cdnPt: - case Hexagon::LDrid_cdnPt : - case Hexagon::LDrid_cdnNotPt : - case Hexagon::LDrid_indexed_cdnPt : - case Hexagon::LDrid_indexed_cdnNotPt : - case Hexagon::POST_LDrid_cdnPt_V4 : - case Hexagon::POST_LDrid_cdnNotPt_V4 : - case Hexagon::LDriw_cdnPt : - case Hexagon::LDriw_cdnNotPt : - case Hexagon::LDriw_indexed_cdnPt : - case Hexagon::LDriw_indexed_cdnNotPt : - case Hexagon::POST_LDriw_cdnPt_V4 : - case Hexagon::POST_LDriw_cdnNotPt_V4 : - case Hexagon::LDrih_cdnPt : - case Hexagon::LDrih_cdnNotPt : - case Hexagon::LDrih_indexed_cdnPt : - case Hexagon::LDrih_indexed_cdnNotPt : - case Hexagon::POST_LDrih_cdnPt_V4 : - case Hexagon::POST_LDrih_cdnNotPt_V4 : - case Hexagon::LDrib_cdnPt : - case Hexagon::LDrib_cdnNotPt : - case Hexagon::LDrib_indexed_cdnPt : - case Hexagon::LDrib_indexed_cdnNotPt : - case Hexagon::POST_LDrib_cdnPt_V4 : - case Hexagon::POST_LDrib_cdnNotPt_V4 : - case Hexagon::LDriuh_cdnPt : - case Hexagon::LDriuh_cdnNotPt : - case Hexagon::LDriuh_indexed_cdnPt : - case Hexagon::LDriuh_indexed_cdnNotPt : - case Hexagon::POST_LDriuh_cdnPt_V4 : - case Hexagon::POST_LDriuh_cdnNotPt_V4 : - case Hexagon::LDriub_cdnPt : - case Hexagon::LDriub_cdnNotPt : - case Hexagon::LDriub_indexed_cdnPt : - case Hexagon::LDriub_indexed_cdnNotPt : - case Hexagon::POST_LDriub_cdnPt_V4 : - case Hexagon::POST_LDriub_cdnNotPt_V4 : - - case Hexagon::LDrid_indexed_shl_cdnPt_V4 : - case Hexagon::LDrid_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDrib_indexed_shl_cdnPt_V4 : - case Hexagon::LDrib_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriub_indexed_shl_cdnPt_V4 : - case Hexagon::LDriub_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDrih_indexed_shl_cdnPt_V4 : - case Hexagon::LDrih_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriuh_indexed_shl_cdnPt_V4 : - case Hexagon::LDriuh_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriw_indexed_shl_cdnPt_V4 : - case Hexagon::LDriw_indexed_shl_cdnNotPt_V4 : - -// Coditional add - case Hexagon::ADD_ri_cdnPt: - case Hexagon::ADD_ri_cdnNotPt: - case Hexagon::ADD_rr_cdnPt: - case Hexagon::ADD_rr_cdnNotPt: - - // Conditional logical operations - case Hexagon::XOR_rr_cdnPt : - case Hexagon::XOR_rr_cdnNotPt : - case Hexagon::AND_rr_cdnPt : - case Hexagon::AND_rr_cdnNotPt : - case Hexagon::OR_rr_cdnPt : - case Hexagon::OR_rr_cdnNotPt : - - // Conditonal subtract - case Hexagon::SUB_rr_cdnPt : - case Hexagon::SUB_rr_cdnNotPt : - - // Conditional combine - case Hexagon::COMBINE_rr_cdnPt : - case Hexagon::COMBINE_rr_cdnNotPt : - - // Conditional shift operations - case Hexagon::ASLH_cdnPt_V4: - case Hexagon::ASLH_cdnNotPt_V4: - case Hexagon::ASRH_cdnPt_V4: - case Hexagon::ASRH_cdnNotPt_V4: - case Hexagon::SXTB_cdnPt_V4: - case Hexagon::SXTB_cdnNotPt_V4: - case Hexagon::SXTH_cdnPt_V4: - case Hexagon::SXTH_cdnNotPt_V4: - case Hexagon::ZXTB_cdnPt_V4: - case Hexagon::ZXTB_cdnNotPt_V4: - case Hexagon::ZXTH_cdnPt_V4: - case Hexagon::ZXTH_cdnNotPt_V4: - - // Conditional stores - case Hexagon::STrib_imm_cdnPt_V4 : - case Hexagon::STrib_imm_cdnNotPt_V4 : - case Hexagon::STrib_cdnPt_V4 : - case Hexagon::STrib_cdnNotPt_V4 : - case Hexagon::STrib_indexed_cdnPt_V4 : - case Hexagon::STrib_indexed_cdnNotPt_V4 : - case Hexagon::POST_STbri_cdnPt_V4 : - case Hexagon::POST_STbri_cdnNotPt_V4 : - case Hexagon::STrib_indexed_shl_cdnPt_V4 : - case Hexagon::STrib_indexed_shl_cdnNotPt_V4 : - - // Store doubleword conditionally - case Hexagon::STrid_indexed_cdnPt_V4 : - case Hexagon::STrid_indexed_cdnNotPt_V4 : - case Hexagon::STrid_indexed_shl_cdnPt_V4 : - case Hexagon::STrid_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_STdri_cdnPt_V4 : - case Hexagon::POST_STdri_cdnNotPt_V4 : - - // Store halfword conditionally - case Hexagon::STrih_cdnPt_V4 : - case Hexagon::STrih_cdnNotPt_V4 : - case Hexagon::STrih_indexed_cdnPt_V4 : - case Hexagon::STrih_indexed_cdnNotPt_V4 : - case Hexagon::STrih_imm_cdnPt_V4 : - case Hexagon::STrih_imm_cdnNotPt_V4 : - case Hexagon::STrih_indexed_shl_cdnPt_V4 : - case Hexagon::STrih_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_SThri_cdnPt_V4 : - case Hexagon::POST_SThri_cdnNotPt_V4 : - - // Store word conditionally - case Hexagon::STriw_cdnPt_V4 : - case Hexagon::STriw_cdnNotPt_V4 : - case Hexagon::STriw_indexed_cdnPt_V4 : - case Hexagon::STriw_indexed_cdnNotPt_V4 : - case Hexagon::STriw_imm_cdnPt_V4 : - case Hexagon::STriw_imm_cdnNotPt_V4 : - case Hexagon::STriw_indexed_shl_cdnPt_V4 : - case Hexagon::STriw_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_STwri_cdnPt_V4 : - case Hexagon::POST_STwri_cdnNotPt_V4 : - - case Hexagon::LDd_GP_cdnPt_V4: - case Hexagon::LDd_GP_cdnNotPt_V4: - case Hexagon::LDb_GP_cdnPt_V4: - case Hexagon::LDb_GP_cdnNotPt_V4: - case Hexagon::LDub_GP_cdnPt_V4: - case Hexagon::LDub_GP_cdnNotPt_V4: - case Hexagon::LDh_GP_cdnPt_V4: - case Hexagon::LDh_GP_cdnNotPt_V4: - case Hexagon::LDuh_GP_cdnPt_V4: - case Hexagon::LDuh_GP_cdnNotPt_V4: - case Hexagon::LDw_GP_cdnPt_V4: - case Hexagon::LDw_GP_cdnNotPt_V4: - - case Hexagon::STd_GP_cdnPt_V4: - case Hexagon::STd_GP_cdnNotPt_V4: - case Hexagon::STb_GP_cdnPt_V4: - case Hexagon::STb_GP_cdnNotPt_V4: - case Hexagon::STh_GP_cdnPt_V4: - case Hexagon::STh_GP_cdnNotPt_V4: - case Hexagon::STw_GP_cdnPt_V4: - case Hexagon::STw_GP_cdnNotPt_V4: - return true; - } - return false; -} - static MachineOperand& GetPostIncrementOperand(MachineInstr *MI, const HexagonInstrInfo *QII) { assert(QII->isPostIncrement(MI) && "Not a post increment operation."); @@ -2490,7 +2358,7 @@ bool HexagonPacketizerList::CanPromoteToNewValueStore( MachineInstr *MI, // sense, i.e, either both should be negated or both should be none negated. if (( predRegNumDst != predRegNumSrc) || - isDotNewInst(PacketMI) != isDotNewInst(MI) || + QII->isDotNewInst(PacketMI) != QII->isDotNewInst(MI) || GetPredicateSense(MI, QII) != GetPredicateSense(PacketMI, QII)) { return false; } @@ -2600,8 +2468,9 @@ bool HexagonPacketizerList::CanPromoteToDotNew( MachineInstr *MI, MachineBasicBlock::iterator &MII, const TargetRegisterClass* RC ) { - // already a dot new instruction - if (isDotNewInst(MI) && !IsNewifyStore(MI)) + const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; + // Already a dot new instruction. + if (QII->isDotNewInst(MI) && !IsNewifyStore(MI)) return false; if (!isNewifiable(MI)) @@ -2616,7 +2485,6 @@ bool HexagonPacketizerList::CanPromoteToDotNew( MachineInstr *MI, else { // Create a dot new machine instruction to see if resources can be // allocated. If not, bail out now. - const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; int NewOpcode = GetDotNewOp(MI->getOpcode()); const MCInstrDesc &desc = QII->get(NewOpcode); DebugLoc dl; @@ -2759,7 +2627,7 @@ bool HexagonPacketizerList::ArePredicatesComplements (MachineInstr* MI1, // !p0 is not complimentary to p0.new return ((MI1->getOperand(1).getReg() == MI2->getOperand(1).getReg()) && (GetPredicateSense(MI1, QII) != GetPredicateSense(MI2, QII)) && - (isDotNewInst(MI1) == isDotNewInst(MI2))); + (QII->isDotNewInst(MI1) == QII->isDotNewInst(MI2))); } // initPacketizerState - Initialize packetizer flags diff --git a/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp b/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp index 86f75d1..3deb8d1 100644 --- a/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp +++ b/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp @@ -31,6 +31,7 @@ HexagonMCAsmInfo::HexagonMCAsmInfo(const Target &T, StringRef TT) { AscizDirective = "\t.string\t"; WeakRefDirective = "\t.weak\t"; + SupportsDebugInformation = true; UsesELFSectionDirectiveForBSS = true; ExceptionsType = ExceptionHandling::DwarfCFI; } |