| 1 | //===-- HexagonISelLowering.h - Hexagon DAG Lowering Interface --*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file defines the interfaces that Hexagon uses to lower LLVM code into a |
| 10 | // selection DAG. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef LLVM_LIB_TARGET_HEXAGON_HEXAGONISELLOWERING_H |
| 15 | #define LLVM_LIB_TARGET_HEXAGON_HEXAGONISELLOWERING_H |
| 16 | |
| 17 | #include "Hexagon.h" |
| 18 | #include "MCTargetDesc/HexagonMCTargetDesc.h" |
| 19 | #include "llvm/ADT/StringRef.h" |
| 20 | #include "llvm/CodeGen/ISDOpcodes.h" |
| 21 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
| 22 | #include "llvm/CodeGen/TargetLowering.h" |
| 23 | #include "llvm/CodeGen/ValueTypes.h" |
| 24 | #include "llvm/CodeGenTypes/MachineValueType.h" |
| 25 | #include "llvm/IR/CallingConv.h" |
| 26 | #include "llvm/IR/InlineAsm.h" |
| 27 | #include <cstdint> |
| 28 | #include <utility> |
| 29 | |
| 30 | namespace llvm { |
| 31 | |
| 32 | namespace HexagonISD { |
| 33 | |
| 34 | enum NodeType : unsigned { |
| 35 | OP_BEGIN = ISD::BUILTIN_OP_END, |
| 36 | |
| 37 | CONST32 = OP_BEGIN, |
| 38 | CONST32_GP, // For marking data present in GP. |
| 39 | ADDC, // Add with carry: (X, Y, Cin) -> (X+Y, Cout). |
| 40 | SUBC, // Sub with carry: (X, Y, Cin) -> (X+~Y+Cin, Cout). |
| 41 | ALLOCA, |
| 42 | |
| 43 | AT_GOT, // Index in GOT. |
| 44 | AT_PCREL, // Offset relative to PC. |
| 45 | |
| 46 | CALL, // Function call. |
| 47 | CALLnr, // Function call that does not return. |
| 48 | CALLR, |
| 49 | |
| 50 | RET_GLUE, // Return with a glue operand. |
| 51 | BARRIER, // Memory barrier. |
| 52 | JT, // Jump table. |
| 53 | CP, // Constant pool. |
| 54 | |
| 55 | COMBINE, |
| 56 | VASL, // Vector shifts by a scalar value |
| 57 | VASR, |
| 58 | VLSR, |
| 59 | MFSHL, // Funnel shifts with the shift amount guaranteed to be |
| 60 | MFSHR, // within the range of the bit width of the element. |
| 61 | |
| 62 | SSAT, // Signed saturate. |
| 63 | USAT, // Unsigned saturate. |
| 64 | SMUL_LOHI, // Same as ISD::SMUL_LOHI, but opaque to the combiner. |
| 65 | UMUL_LOHI, // Same as ISD::UMUL_LOHI, but opaque to the combiner. |
| 66 | // We want to legalize MULH[SU] to [SU]MUL_LOHI, but the |
| 67 | // combiner will keep rewriting it back to MULH[SU]. |
| 68 | USMUL_LOHI, // Like SMUL_LOHI, but unsigned*signed. |
| 69 | |
| 70 | TSTBIT, |
| 71 | INSERT, |
| 72 | , |
| 73 | , |
| 74 | VINSERTW0, |
| 75 | VROR, |
| 76 | TC_RETURN, |
| 77 | EH_RETURN, |
| 78 | DCFETCH, |
| 79 | READCYCLE, |
| 80 | READTIMER, |
| 81 | PTRUE, |
| 82 | PFALSE, |
| 83 | D2P, // Convert 8-byte value to 8-bit predicate register. [*] |
| 84 | P2D, // Convert 8-bit predicate register to 8-byte value. [*] |
| 85 | V2Q, // Convert HVX vector to a vector predicate reg. [*] |
| 86 | Q2V, // Convert vector predicate to an HVX vector. [*] |
| 87 | // [*] The equivalence is defined as "Q <=> (V != 0)", |
| 88 | // where the != operation compares bytes. |
| 89 | // Note: V != 0 is implemented as V >u 0. |
| 90 | QCAT, |
| 91 | QTRUE, |
| 92 | QFALSE, |
| 93 | |
| 94 | TL_EXTEND, // Wrappers for ISD::*_EXTEND and ISD::TRUNCATE to prevent DAG |
| 95 | TL_TRUNCATE, // from auto-folding operations, e.g. |
| 96 | // (i32 ext (i16 ext i8)) would be folded to (i32 ext i8). |
| 97 | // To simplify the type legalization, we want to keep these |
| 98 | // single steps separate during type legalization. |
| 99 | // TL_[EXTEND|TRUNCATE] Inp, i128 _, i32 Opc |
| 100 | // * Inp is the original input to extend/truncate, |
| 101 | // * _ is a dummy operand with an illegal type (can be undef), |
| 102 | // * Opc is the original opcode. |
| 103 | // The legalization process (in Hexagon lowering code) will |
| 104 | // first deal with the "real" types (i.e. Inp and the result), |
| 105 | // and once all of them are processed, the wrapper node will |
| 106 | // be replaced with the original ISD node. The dummy illegal |
| 107 | // operand is there to make sure that the legalization hooks |
| 108 | // are called again after everything else is legal, giving |
| 109 | // us the opportunity to undo the wrapping. |
| 110 | |
| 111 | TYPECAST, // No-op that's used to convert between different legal |
| 112 | // types in a register. |
| 113 | VALIGN, // Align two vectors (in Op0, Op1) to one that would have |
| 114 | // been loaded from address in Op2. |
| 115 | VALIGNADDR, // Align vector address: Op0 & -Op1, except when it is |
| 116 | // an address in a vector load, then it's a no-op. |
| 117 | ISEL, // Marker for nodes that were created during ISel, and |
| 118 | // which need explicit selection (would have been left |
| 119 | // unselected otherwise). |
| 120 | OP_END |
| 121 | }; |
| 122 | |
| 123 | } // end namespace HexagonISD |
| 124 | |
| 125 | class HexagonSubtarget; |
| 126 | |
| 127 | class HexagonTargetLowering : public TargetLowering { |
| 128 | int VarArgsFrameOffset; // Frame offset to start of varargs area. |
| 129 | const HexagonTargetMachine &HTM; |
| 130 | const HexagonSubtarget &Subtarget; |
| 131 | |
| 132 | public: |
| 133 | explicit HexagonTargetLowering(const TargetMachine &TM, |
| 134 | const HexagonSubtarget &ST); |
| 135 | |
| 136 | /// IsEligibleForTailCallOptimization - Check whether the call is eligible |
| 137 | /// for tail call optimization. Targets which want to do tail call |
| 138 | /// optimization should implement this function. |
| 139 | bool IsEligibleForTailCallOptimization(SDValue Callee, |
| 140 | CallingConv::ID CalleeCC, bool isVarArg, bool isCalleeStructRet, |
| 141 | bool isCallerStructRet, const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 142 | const SmallVectorImpl<SDValue> &OutVals, |
| 143 | const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG& DAG) const; |
| 144 | |
| 145 | bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I, |
| 146 | MachineFunction &MF, |
| 147 | unsigned Intrinsic) const override; |
| 148 | |
| 149 | bool isTruncateFree(Type *Ty1, Type *Ty2) const override; |
| 150 | bool isTruncateFree(EVT VT1, EVT VT2) const override; |
| 151 | |
| 152 | bool isCheapToSpeculateCttz(Type *) const override { return true; } |
| 153 | bool isCheapToSpeculateCtlz(Type *) const override { return true; } |
| 154 | bool isCtlzFast() const override { return true; } |
| 155 | |
| 156 | bool hasBitTest(SDValue X, SDValue Y) const override; |
| 157 | |
| 158 | bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override; |
| 159 | |
| 160 | /// Return true if an FMA operation is faster than a pair of mul and add |
| 161 | /// instructions. fmuladd intrinsics will be expanded to FMAs when this |
| 162 | /// method returns true (and FMAs are legal), otherwise fmuladd is |
| 163 | /// expanded to mul + add. |
| 164 | bool isFMAFasterThanFMulAndFAdd(const MachineFunction &, |
| 165 | EVT) const override; |
| 166 | |
| 167 | // Should we expand the build vector with shuffles? |
| 168 | bool shouldExpandBuildVectorWithShuffles(EVT VT, |
| 169 | unsigned DefinedValues) const override; |
| 170 | bool (EVT ResVT, EVT SrcVT, |
| 171 | unsigned Index) const override; |
| 172 | |
| 173 | bool isTargetCanonicalConstantNode(SDValue Op) const override; |
| 174 | |
| 175 | bool isShuffleMaskLegal(ArrayRef<int> Mask, EVT VT) const override; |
| 176 | LegalizeTypeAction getPreferredVectorAction(MVT VT) const override; |
| 177 | LegalizeAction getCustomOperationAction(SDNode &Op) const override; |
| 178 | |
| 179 | SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; |
| 180 | void LowerOperationWrapper(SDNode *N, SmallVectorImpl<SDValue> &Results, |
| 181 | SelectionDAG &DAG) const override; |
| 182 | void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results, |
| 183 | SelectionDAG &DAG) const override; |
| 184 | |
| 185 | const char *getTargetNodeName(unsigned Opcode) const override; |
| 186 | |
| 187 | SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const; |
| 188 | SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const; |
| 189 | SDValue (SDValue Op, SelectionDAG &DAG) const; |
| 190 | SDValue (SDValue Op, SelectionDAG &DAG) const; |
| 191 | SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; |
| 192 | SDValue LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const; |
| 193 | SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const; |
| 194 | SDValue LowerVECTOR_SHIFT(SDValue Op, SelectionDAG &DAG) const; |
| 195 | SDValue LowerROTL(SDValue Op, SelectionDAG &DAG) const; |
| 196 | SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) const; |
| 197 | SDValue LowerANY_EXTEND(SDValue Op, SelectionDAG &DAG) const; |
| 198 | SDValue LowerSIGN_EXTEND(SDValue Op, SelectionDAG &DAG) const; |
| 199 | SDValue LowerZERO_EXTEND(SDValue Op, SelectionDAG &DAG) const; |
| 200 | SDValue LowerLoad(SDValue Op, SelectionDAG &DAG) const; |
| 201 | SDValue LowerStore(SDValue Op, SelectionDAG &DAG) const; |
| 202 | SDValue LowerUnalignedLoad(SDValue Op, SelectionDAG &DAG) const; |
| 203 | SDValue LowerUAddSubO(SDValue Op, SelectionDAG &DAG) const; |
| 204 | SDValue LowerUAddSubOCarry(SDValue Op, SelectionDAG &DAG) const; |
| 205 | |
| 206 | SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const; |
| 207 | SDValue LowerINLINEASM(SDValue Op, SelectionDAG &DAG) const; |
| 208 | SDValue LowerFDIV(SDValue Op, SelectionDAG &DAG) const; |
| 209 | SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) const; |
| 210 | SDValue LowerREADCYCLECOUNTER(SDValue Op, SelectionDAG &DAG) const; |
| 211 | SDValue LowerREADSTEADYCOUNTER(SDValue Op, SelectionDAG &DAG) const; |
| 212 | SDValue LowerEH_LABEL(SDValue Op, SelectionDAG &DAG) const; |
| 213 | SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const; |
| 214 | SDValue |
| 215 | LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
| 216 | const SmallVectorImpl<ISD::InputArg> &Ins, |
| 217 | const SDLoc &dl, SelectionDAG &DAG, |
| 218 | SmallVectorImpl<SDValue> &InVals) const override; |
| 219 | SDValue LowerGLOBALADDRESS(SDValue Op, SelectionDAG &DAG) const; |
| 220 | SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; |
| 221 | SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const; |
| 222 | SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, |
| 223 | SelectionDAG &DAG) const; |
| 224 | SDValue LowerToTLSInitialExecModel(GlobalAddressSDNode *GA, |
| 225 | SelectionDAG &DAG) const; |
| 226 | SDValue LowerToTLSLocalExecModel(GlobalAddressSDNode *GA, |
| 227 | SelectionDAG &DAG) const; |
| 228 | SDValue GetDynamicTLSAddr(SelectionDAG &DAG, SDValue Chain, |
| 229 | GlobalAddressSDNode *GA, SDValue InGlue, EVT PtrVT, |
| 230 | unsigned ReturnReg, unsigned char OperandGlues) const; |
| 231 | SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const; |
| 232 | |
| 233 | SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI, |
| 234 | SmallVectorImpl<SDValue> &InVals) const override; |
| 235 | SDValue LowerCallResult(SDValue Chain, SDValue InGlue, |
| 236 | CallingConv::ID CallConv, bool isVarArg, |
| 237 | const SmallVectorImpl<ISD::InputArg> &Ins, |
| 238 | const SDLoc &dl, SelectionDAG &DAG, |
| 239 | SmallVectorImpl<SDValue> &InVals, |
| 240 | const SmallVectorImpl<SDValue> &OutVals, |
| 241 | SDValue Callee) const; |
| 242 | |
| 243 | SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const; |
| 244 | SDValue LowerVSELECT(SDValue Op, SelectionDAG &DAG) const; |
| 245 | SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; |
| 246 | SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG& DAG) const; |
| 247 | SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; |
| 248 | |
| 249 | bool CanLowerReturn(CallingConv::ID CallConv, |
| 250 | MachineFunction &MF, bool isVarArg, |
| 251 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 252 | LLVMContext &Context, const Type *RetTy) const override; |
| 253 | |
| 254 | SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
| 255 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 256 | const SmallVectorImpl<SDValue> &OutVals, |
| 257 | const SDLoc &dl, SelectionDAG &DAG) const override; |
| 258 | |
| 259 | SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; |
| 260 | |
| 261 | bool mayBeEmittedAsTailCall(const CallInst *CI) const override; |
| 262 | |
| 263 | Register getRegisterByName(const char* RegName, LLT VT, |
| 264 | const MachineFunction &MF) const override; |
| 265 | |
| 266 | /// If a physical register, this returns the register that receives the |
| 267 | /// exception address on entry to an EH pad. |
| 268 | Register |
| 269 | getExceptionPointerRegister(const Constant *PersonalityFn) const override { |
| 270 | return Hexagon::R0; |
| 271 | } |
| 272 | |
| 273 | /// If a physical register, this returns the register that receives the |
| 274 | /// exception typeid on entry to a landing pad. |
| 275 | Register |
| 276 | getExceptionSelectorRegister(const Constant *PersonalityFn) const override { |
| 277 | return Hexagon::R1; |
| 278 | } |
| 279 | |
| 280 | SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const; |
| 281 | SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const; |
| 282 | SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const; |
| 283 | SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const; |
| 284 | |
| 285 | EVT getSetCCResultType(const DataLayout &, LLVMContext &C, |
| 286 | EVT VT) const override { |
| 287 | if (!VT.isVector()) |
| 288 | return MVT::i1; |
| 289 | else |
| 290 | return EVT::getVectorVT(Context&: C, VT: MVT::i1, NumElements: VT.getVectorNumElements()); |
| 291 | } |
| 292 | |
| 293 | bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, |
| 294 | SDValue &Base, SDValue &Offset, |
| 295 | ISD::MemIndexedMode &AM, |
| 296 | SelectionDAG &DAG) const override; |
| 297 | |
| 298 | ConstraintType getConstraintType(StringRef Constraint) const override; |
| 299 | |
| 300 | std::pair<unsigned, const TargetRegisterClass *> |
| 301 | getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
| 302 | StringRef Constraint, MVT VT) const override; |
| 303 | |
| 304 | // Intrinsics |
| 305 | SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const; |
| 306 | SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const; |
| 307 | /// isLegalAddressingMode - Return true if the addressing mode represented |
| 308 | /// by AM is legal for this target, for a load/store of the specified type. |
| 309 | /// The type may be VoidTy, in which case only return true if the addressing |
| 310 | /// mode is legal for a load/store of any legal type. |
| 311 | /// TODO: Handle pre/postinc as well. |
| 312 | bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, |
| 313 | Type *Ty, unsigned AS, |
| 314 | Instruction *I = nullptr) const override; |
| 315 | /// Return true if folding a constant offset with the given GlobalAddress |
| 316 | /// is legal. It is frequently not legal in PIC relocation models. |
| 317 | bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override; |
| 318 | |
| 319 | bool isFPImmLegal(const APFloat &Imm, EVT VT, |
| 320 | bool ForCodeSize) const override; |
| 321 | |
| 322 | /// isLegalICmpImmediate - Return true if the specified immediate is legal |
| 323 | /// icmp immediate, that is the target has icmp instructions which can |
| 324 | /// compare a register against the immediate without having to materialize |
| 325 | /// the immediate into a register. |
| 326 | bool isLegalICmpImmediate(int64_t Imm) const override; |
| 327 | |
| 328 | EVT getOptimalMemOpType(const MemOp &Op, |
| 329 | const AttributeList &FuncAttributes) const override; |
| 330 | |
| 331 | bool allowsMemoryAccess(LLVMContext &Context, const DataLayout &DL, EVT VT, |
| 332 | unsigned AddrSpace, Align Alignment, |
| 333 | MachineMemOperand::Flags Flags, |
| 334 | unsigned *Fast) const override; |
| 335 | |
| 336 | bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AddrSpace, |
| 337 | Align Alignment, |
| 338 | MachineMemOperand::Flags Flags, |
| 339 | unsigned *Fast) const override; |
| 340 | |
| 341 | /// Returns relocation base for the given PIC jumptable. |
| 342 | SDValue getPICJumpTableRelocBase(SDValue Table, SelectionDAG &DAG) |
| 343 | const override; |
| 344 | |
| 345 | /// Returns true if it is beneficial to convert a load of a constant |
| 346 | /// to just the constant itself. |
| 347 | bool shouldConvertConstantLoadToIntImm(const APInt &Imm, |
| 348 | Type *Ty) const override; |
| 349 | |
| 350 | bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy, EVT NewVT, |
| 351 | std::optional<unsigned> ByteOffset) const override; |
| 352 | |
| 353 | void AdjustInstrPostInstrSelection(MachineInstr &MI, |
| 354 | SDNode *Node) const override; |
| 355 | |
| 356 | // Handling of atomic RMW instructions. |
| 357 | Value *emitLoadLinked(IRBuilderBase &Builder, Type *ValueTy, Value *Addr, |
| 358 | AtomicOrdering Ord) const override; |
| 359 | Value *emitStoreConditional(IRBuilderBase &Builder, Value *Val, Value *Addr, |
| 360 | AtomicOrdering Ord) const override; |
| 361 | AtomicExpansionKind shouldExpandAtomicLoadInIR(LoadInst *LI) const override; |
| 362 | AtomicExpansionKind shouldExpandAtomicStoreInIR(StoreInst *SI) const override; |
| 363 | AtomicExpansionKind |
| 364 | shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override; |
| 365 | |
| 366 | AtomicExpansionKind |
| 367 | shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override { |
| 368 | return AtomicExpansionKind::LLSC; |
| 369 | } |
| 370 | bool softPromoteHalfType() const override { return true; } |
| 371 | |
| 372 | private: |
| 373 | void initializeHVXLowering(); |
| 374 | unsigned getPreferredHvxVectorAction(MVT VecTy) const; |
| 375 | unsigned getCustomHvxOperationAction(SDNode &Op) const; |
| 376 | |
| 377 | bool validateConstPtrAlignment(SDValue Ptr, Align NeedAlign, const SDLoc &dl, |
| 378 | SelectionDAG &DAG) const; |
| 379 | SDValue replaceMemWithUndef(SDValue Op, SelectionDAG &DAG) const; |
| 380 | |
| 381 | std::pair<SDValue,int> getBaseAndOffset(SDValue Addr) const; |
| 382 | |
| 383 | bool getBuildVectorConstInts(ArrayRef<SDValue> Values, MVT VecTy, |
| 384 | SelectionDAG &DAG, |
| 385 | MutableArrayRef<ConstantInt*> Consts) const; |
| 386 | SDValue buildVector32(ArrayRef<SDValue> Elem, const SDLoc &dl, MVT VecTy, |
| 387 | SelectionDAG &DAG) const; |
| 388 | SDValue buildVector64(ArrayRef<SDValue> Elem, const SDLoc &dl, MVT VecTy, |
| 389 | SelectionDAG &DAG) const; |
| 390 | SDValue (SDValue VecV, SDValue IdxV, const SDLoc &dl, |
| 391 | MVT ValTy, MVT ResTy, SelectionDAG &DAG) const; |
| 392 | SDValue (SDValue VecV, SDValue IdxV, const SDLoc &dl, |
| 393 | MVT ValTy, MVT ResTy, SelectionDAG &DAG) const; |
| 394 | SDValue insertVector(SDValue VecV, SDValue ValV, SDValue IdxV, |
| 395 | const SDLoc &dl, MVT ValTy, SelectionDAG &DAG) const; |
| 396 | SDValue insertVectorPred(SDValue VecV, SDValue ValV, SDValue IdxV, |
| 397 | const SDLoc &dl, MVT ValTy, SelectionDAG &DAG) const; |
| 398 | SDValue expandPredicate(SDValue Vec32, const SDLoc &dl, |
| 399 | SelectionDAG &DAG) const; |
| 400 | SDValue contractPredicate(SDValue Vec64, const SDLoc &dl, |
| 401 | SelectionDAG &DAG) const; |
| 402 | SDValue getSplatValue(SDValue Op, SelectionDAG &DAG) const; |
| 403 | SDValue getVectorShiftByInt(SDValue Op, SelectionDAG &DAG) const; |
| 404 | SDValue appendUndef(SDValue Val, MVT ResTy, SelectionDAG &DAG) const; |
| 405 | SDValue getCombine(SDValue Hi, SDValue Lo, const SDLoc &dl, MVT ResTy, |
| 406 | SelectionDAG &DAG) const; |
| 407 | |
| 408 | bool isUndef(SDValue Op) const { |
| 409 | if (Op.isMachineOpcode()) |
| 410 | return Op.getMachineOpcode() == TargetOpcode::IMPLICIT_DEF; |
| 411 | return Op.getOpcode() == ISD::UNDEF; |
| 412 | } |
| 413 | SDValue getInstr(unsigned MachineOpc, const SDLoc &dl, MVT Ty, |
| 414 | ArrayRef<SDValue> Ops, SelectionDAG &DAG) const { |
| 415 | SDNode *N = DAG.getMachineNode(Opcode: MachineOpc, dl, VT: Ty, Ops); |
| 416 | return SDValue(N, 0); |
| 417 | } |
| 418 | SDValue getZero(const SDLoc &dl, MVT Ty, SelectionDAG &DAG) const; |
| 419 | |
| 420 | using VectorPair = std::pair<SDValue, SDValue>; |
| 421 | using TypePair = std::pair<MVT, MVT>; |
| 422 | |
| 423 | SDValue getInt(unsigned IntId, MVT ResTy, ArrayRef<SDValue> Ops, |
| 424 | const SDLoc &dl, SelectionDAG &DAG) const; |
| 425 | |
| 426 | MVT ty(SDValue Op) const { |
| 427 | return Op.getValueType().getSimpleVT(); |
| 428 | } |
| 429 | TypePair ty(const VectorPair &Ops) const { |
| 430 | return { Ops.first.getValueType().getSimpleVT(), |
| 431 | Ops.second.getValueType().getSimpleVT() }; |
| 432 | } |
| 433 | MVT tyScalar(MVT Ty) const { |
| 434 | if (!Ty.isVector()) |
| 435 | return Ty; |
| 436 | return MVT::getIntegerVT(BitWidth: Ty.getSizeInBits()); |
| 437 | } |
| 438 | MVT tyVector(MVT Ty, MVT ElemTy) const { |
| 439 | if (Ty.isVector() && Ty.getVectorElementType() == ElemTy) |
| 440 | return Ty; |
| 441 | unsigned TyWidth = Ty.getSizeInBits(); |
| 442 | unsigned ElemWidth = ElemTy.getSizeInBits(); |
| 443 | assert((TyWidth % ElemWidth) == 0); |
| 444 | return MVT::getVectorVT(VT: ElemTy, NumElements: TyWidth/ElemWidth); |
| 445 | } |
| 446 | |
| 447 | MVT typeJoin(const TypePair &Tys) const; |
| 448 | TypePair typeSplit(MVT Ty) const; |
| 449 | MVT typeExtElem(MVT VecTy, unsigned Factor) const; |
| 450 | MVT typeTruncElem(MVT VecTy, unsigned Factor) const; |
| 451 | TypePair typeExtendToWider(MVT Ty0, MVT Ty1) const; |
| 452 | TypePair typeWidenToWider(MVT Ty0, MVT Ty1) const; |
| 453 | MVT typeLegalize(MVT Ty, SelectionDAG &DAG) const; |
| 454 | MVT typeWidenToHvx(MVT Ty) const; |
| 455 | |
| 456 | SDValue opJoin(const VectorPair &Ops, const SDLoc &dl, |
| 457 | SelectionDAG &DAG) const; |
| 458 | VectorPair opSplit(SDValue Vec, const SDLoc &dl, SelectionDAG &DAG) const; |
| 459 | SDValue opCastElem(SDValue Vec, MVT ElemTy, SelectionDAG &DAG) const; |
| 460 | |
| 461 | SDValue LoHalf(SDValue V, SelectionDAG &DAG) const { |
| 462 | MVT Ty = ty(Op: V); |
| 463 | const SDLoc &dl(V); |
| 464 | if (!Ty.isVector()) { |
| 465 | assert(Ty.getSizeInBits() == 64); |
| 466 | return DAG.getTargetExtractSubreg(SRIdx: Hexagon::isub_lo, DL: dl, VT: MVT::i32, Operand: V); |
| 467 | } |
| 468 | MVT HalfTy = typeSplit(Ty).first; |
| 469 | SDValue Idx = getZero(dl, Ty: MVT::i32, DAG); |
| 470 | return DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL: dl, VT: HalfTy, N1: V, N2: Idx); |
| 471 | } |
| 472 | SDValue HiHalf(SDValue V, SelectionDAG &DAG) const { |
| 473 | MVT Ty = ty(Op: V); |
| 474 | const SDLoc &dl(V); |
| 475 | if (!Ty.isVector()) { |
| 476 | assert(Ty.getSizeInBits() == 64); |
| 477 | return DAG.getTargetExtractSubreg(SRIdx: Hexagon::isub_hi, DL: dl, VT: MVT::i32, Operand: V); |
| 478 | } |
| 479 | MVT HalfTy = typeSplit(Ty).first; |
| 480 | SDValue Idx = DAG.getConstant(Val: HalfTy.getVectorNumElements(), DL: dl, VT: MVT::i32); |
| 481 | return DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL: dl, VT: HalfTy, N1: V, N2: Idx); |
| 482 | } |
| 483 | |
| 484 | bool allowsHvxMemoryAccess(MVT VecTy, MachineMemOperand::Flags Flags, |
| 485 | unsigned *Fast) const; |
| 486 | bool allowsHvxMisalignedMemoryAccesses(MVT VecTy, |
| 487 | MachineMemOperand::Flags Flags, |
| 488 | unsigned *Fast) const; |
| 489 | void AdjustHvxInstrPostInstrSelection(MachineInstr &MI, SDNode *Node) const; |
| 490 | |
| 491 | bool isHvxSingleTy(MVT Ty) const; |
| 492 | bool isHvxPairTy(MVT Ty) const; |
| 493 | bool isHvxBoolTy(MVT Ty) const; |
| 494 | SDValue convertToByteIndex(SDValue ElemIdx, MVT ElemTy, |
| 495 | SelectionDAG &DAG) const; |
| 496 | SDValue getIndexInWord32(SDValue Idx, MVT ElemTy, SelectionDAG &DAG) const; |
| 497 | SDValue getByteShuffle(const SDLoc &dl, SDValue Op0, SDValue Op1, |
| 498 | ArrayRef<int> Mask, SelectionDAG &DAG) const; |
| 499 | |
| 500 | SDValue buildHvxVectorReg(ArrayRef<SDValue> Values, const SDLoc &dl, |
| 501 | MVT VecTy, SelectionDAG &DAG) const; |
| 502 | SDValue buildHvxVectorPred(ArrayRef<SDValue> Values, const SDLoc &dl, |
| 503 | MVT VecTy, SelectionDAG &DAG) const; |
| 504 | SDValue createHvxPrefixPred(SDValue PredV, const SDLoc &dl, |
| 505 | unsigned BitBytes, bool ZeroFill, |
| 506 | SelectionDAG &DAG) const; |
| 507 | SDValue (SDValue VecV, SDValue IdxV, const SDLoc &dl, |
| 508 | MVT ResTy, SelectionDAG &DAG) const; |
| 509 | SDValue (SDValue VecV, SDValue IdxV, const SDLoc &dl, |
| 510 | MVT ResTy, SelectionDAG &DAG) const; |
| 511 | SDValue insertHvxElementReg(SDValue VecV, SDValue IdxV, SDValue ValV, |
| 512 | const SDLoc &dl, SelectionDAG &DAG) const; |
| 513 | SDValue insertHvxElementPred(SDValue VecV, SDValue IdxV, SDValue ValV, |
| 514 | const SDLoc &dl, SelectionDAG &DAG) const; |
| 515 | SDValue (SDValue OrigOp, SDValue VecV, SDValue IdxV, |
| 516 | const SDLoc &dl, MVT ResTy, SelectionDAG &DAG) |
| 517 | const; |
| 518 | SDValue (SDValue VecV, SDValue IdxV, const SDLoc &dl, |
| 519 | MVT ResTy, SelectionDAG &DAG) const; |
| 520 | SDValue insertHvxSubvectorReg(SDValue VecV, SDValue SubV, SDValue IdxV, |
| 521 | const SDLoc &dl, SelectionDAG &DAG) const; |
| 522 | SDValue insertHvxSubvectorPred(SDValue VecV, SDValue SubV, SDValue IdxV, |
| 523 | const SDLoc &dl, SelectionDAG &DAG) const; |
| 524 | SDValue extendHvxVectorPred(SDValue VecV, const SDLoc &dl, MVT ResTy, |
| 525 | bool ZeroExt, SelectionDAG &DAG) const; |
| 526 | SDValue compressHvxPred(SDValue VecQ, const SDLoc &dl, MVT ResTy, |
| 527 | SelectionDAG &DAG) const; |
| 528 | SDValue resizeToWidth(SDValue VecV, MVT ResTy, bool Signed, const SDLoc &dl, |
| 529 | SelectionDAG &DAG) const; |
| 530 | SDValue (SDValue Vec, MVT SubTy, unsigned SubIdx, |
| 531 | SelectionDAG &DAG) const; |
| 532 | VectorPair emitHvxAddWithOverflow(SDValue A, SDValue B, const SDLoc &dl, |
| 533 | bool Signed, SelectionDAG &DAG) const; |
| 534 | VectorPair emitHvxShiftRightRnd(SDValue Val, unsigned Amt, bool Signed, |
| 535 | SelectionDAG &DAG) const; |
| 536 | SDValue emitHvxMulHsV60(SDValue A, SDValue B, const SDLoc &dl, |
| 537 | SelectionDAG &DAG) const; |
| 538 | SDValue emitHvxMulLoHiV60(SDValue A, bool SignedA, SDValue B, bool SignedB, |
| 539 | const SDLoc &dl, SelectionDAG &DAG) const; |
| 540 | SDValue emitHvxMulLoHiV62(SDValue A, bool SignedA, SDValue B, bool SignedB, |
| 541 | const SDLoc &dl, SelectionDAG &DAG) const; |
| 542 | |
| 543 | SDValue LowerHvxBuildVector(SDValue Op, SelectionDAG &DAG) const; |
| 544 | SDValue LowerHvxSplatVector(SDValue Op, SelectionDAG &DAG) const; |
| 545 | SDValue LowerHvxConcatVectors(SDValue Op, SelectionDAG &DAG) const; |
| 546 | SDValue (SDValue Op, SelectionDAG &DAG) const; |
| 547 | SDValue LowerHvxInsertElement(SDValue Op, SelectionDAG &DAG) const; |
| 548 | SDValue (SDValue Op, SelectionDAG &DAG) const; |
| 549 | SDValue LowerHvxInsertSubvector(SDValue Op, SelectionDAG &DAG) const; |
| 550 | SDValue LowerHvxBitcast(SDValue Op, SelectionDAG &DAG) const; |
| 551 | SDValue LowerHvxAnyExt(SDValue Op, SelectionDAG &DAG) const; |
| 552 | SDValue LowerHvxSignExt(SDValue Op, SelectionDAG &DAG) const; |
| 553 | SDValue LowerHvxZeroExt(SDValue Op, SelectionDAG &DAG) const; |
| 554 | SDValue LowerHvxCttz(SDValue Op, SelectionDAG &DAG) const; |
| 555 | SDValue LowerHvxMulh(SDValue Op, SelectionDAG &DAG) const; |
| 556 | SDValue LowerHvxMulLoHi(SDValue Op, SelectionDAG &DAG) const; |
| 557 | SDValue LowerHvxExtend(SDValue Op, SelectionDAG &DAG) const; |
| 558 | SDValue LowerHvxSelect(SDValue Op, SelectionDAG &DAG) const; |
| 559 | SDValue LowerHvxShift(SDValue Op, SelectionDAG &DAG) const; |
| 560 | SDValue LowerHvxFunnelShift(SDValue Op, SelectionDAG &DAG) const; |
| 561 | SDValue LowerHvxIntrinsic(SDValue Op, SelectionDAG &DAG) const; |
| 562 | SDValue LowerHvxMaskedOp(SDValue Op, SelectionDAG &DAG) const; |
| 563 | SDValue LowerHvxFpExtend(SDValue Op, SelectionDAG &DAG) const; |
| 564 | SDValue LowerHvxFpToInt(SDValue Op, SelectionDAG &DAG) const; |
| 565 | SDValue LowerHvxIntToFp(SDValue Op, SelectionDAG &DAG) const; |
| 566 | SDValue ExpandHvxFpToInt(SDValue Op, SelectionDAG &DAG) const; |
| 567 | SDValue ExpandHvxIntToFp(SDValue Op, SelectionDAG &DAG) const; |
| 568 | |
| 569 | VectorPair SplitVectorOp(SDValue Op, SelectionDAG &DAG) const; |
| 570 | |
| 571 | SDValue SplitHvxMemOp(SDValue Op, SelectionDAG &DAG) const; |
| 572 | SDValue WidenHvxLoad(SDValue Op, SelectionDAG &DAG) const; |
| 573 | SDValue WidenHvxStore(SDValue Op, SelectionDAG &DAG) const; |
| 574 | SDValue WidenHvxSetCC(SDValue Op, SelectionDAG &DAG) const; |
| 575 | SDValue LegalizeHvxResize(SDValue Op, SelectionDAG &DAG) const; |
| 576 | SDValue ExpandHvxResizeIntoSteps(SDValue Op, SelectionDAG &DAG) const; |
| 577 | SDValue EqualizeFpIntConversion(SDValue Op, SelectionDAG &DAG) const; |
| 578 | |
| 579 | SDValue CreateTLWrapper(SDValue Op, SelectionDAG &DAG) const; |
| 580 | SDValue RemoveTLWrapper(SDValue Op, SelectionDAG &DAG) const; |
| 581 | |
| 582 | std::pair<const TargetRegisterClass*, uint8_t> |
| 583 | findRepresentativeClass(const TargetRegisterInfo *TRI, MVT VT) |
| 584 | const override; |
| 585 | |
| 586 | bool shouldSplitToHvx(MVT Ty, SelectionDAG &DAG) const; |
| 587 | bool shouldWidenToHvx(MVT Ty, SelectionDAG &DAG) const; |
| 588 | bool isHvxOperation(SDNode *N, SelectionDAG &DAG) const; |
| 589 | SDValue LowerHvxOperation(SDValue Op, SelectionDAG &DAG) const; |
| 590 | void LowerHvxOperationWrapper(SDNode *N, SmallVectorImpl<SDValue> &Results, |
| 591 | SelectionDAG &DAG) const; |
| 592 | void ReplaceHvxNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results, |
| 593 | SelectionDAG &DAG) const; |
| 594 | |
| 595 | SDValue combineTruncateBeforeLegal(SDValue Op, DAGCombinerInfo &DCI) const; |
| 596 | SDValue combineConcatVectorsBeforeLegal(SDValue Op, DAGCombinerInfo & DCI) |
| 597 | const; |
| 598 | SDValue combineVectorShuffleBeforeLegal(SDValue Op, DAGCombinerInfo & DCI) |
| 599 | const; |
| 600 | |
| 601 | SDValue PerformHvxDAGCombine(SDNode * N, DAGCombinerInfo & DCI) const; |
| 602 | }; |
| 603 | |
| 604 | } // end namespace llvm |
| 605 | |
| 606 | #endif // LLVM_LIB_TARGET_HEXAGON_HEXAGONISELLOWERING_H |
| 607 | |