1 //===-- RISCVISelLowering.h - RISCV 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 RISCV uses to lower LLVM code into a 10 // selection DAG. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_LIB_TARGET_RISCV_RISCVISELLOWERING_H 15 #define LLVM_LIB_TARGET_RISCV_RISCVISELLOWERING_H 16 17 #include "RISCV.h" 18 #include "llvm/CodeGen/CallingConvLower.h" 19 #include "llvm/CodeGen/SelectionDAG.h" 20 #include "llvm/CodeGen/TargetLowering.h" 21 22 namespace llvm { 23 class RISCVSubtarget; 24 struct RISCVRegisterInfo; 25 namespace RISCVISD { 26 enum NodeType : unsigned { 27 FIRST_NUMBER = ISD::BUILTIN_OP_END, 28 RET_FLAG, 29 URET_FLAG, 30 SRET_FLAG, 31 MRET_FLAG, 32 CALL, 33 /// Select with condition operator - This selects between a true value and 34 /// a false value (ops #3 and #4) based on the boolean result of comparing 35 /// the lhs and rhs (ops #0 and #1) of a conditional expression with the 36 /// condition code in op #2, a XLenVT constant from the ISD::CondCode enum. 37 /// The lhs and rhs are XLenVT integers. The true and false values can be 38 /// integer or floating point. 39 SELECT_CC, 40 BR_CC, 41 BuildPairF64, 42 SplitF64, 43 TAIL, 44 // Multiply high for signedxunsigned. 45 MULHSU, 46 // RV64I shifts, directly matching the semantics of the named RISC-V 47 // instructions. 48 SLLW, 49 SRAW, 50 SRLW, 51 // 32-bit operations from RV64M that can't be simply matched with a pattern 52 // at instruction selection time. These have undefined behavior for division 53 // by 0 or overflow (divw) like their target independent counterparts. 54 DIVW, 55 DIVUW, 56 REMUW, 57 // RV64IB rotates, directly matching the semantics of the named RISC-V 58 // instructions. 59 ROLW, 60 RORW, 61 // RV64IZbb bit counting instructions directly matching the semantics of the 62 // named RISC-V instructions. 63 CLZW, 64 CTZW, 65 // RV64IB/RV32IB funnel shifts, with the semantics of the named RISC-V 66 // instructions. Operand order is rs1, rs3, rs2/shamt. 67 FSR, 68 FSL, 69 // RV64IB funnel shifts, with the semantics of the named RISC-V instructions. 70 // Operand order is rs1, rs3, rs2/shamt. 71 FSRW, 72 FSLW, 73 // FPR<->GPR transfer operations when the FPR is smaller than XLEN, needed as 74 // XLEN is the only legal integer width. 75 // 76 // FMV_H_X matches the semantics of the FMV.H.X. 77 // FMV_X_ANYEXTH is similar to FMV.X.H but has an any-extended result. 78 // FMV_X_SIGNEXTH is similar to FMV.X.H and has a sign-extended result. 79 // FMV_W_X_RV64 matches the semantics of the FMV.W.X. 80 // FMV_X_ANYEXTW_RV64 is similar to FMV.X.W but has an any-extended result. 81 // 82 // This is a more convenient semantic for producing dagcombines that remove 83 // unnecessary GPR->FPR->GPR moves. 84 FMV_H_X, 85 FMV_X_ANYEXTH, 86 FMV_X_SIGNEXTH, 87 FMV_W_X_RV64, 88 FMV_X_ANYEXTW_RV64, 89 // FP to XLen int conversions. Corresponds to fcvt.l(u).s/d/h on RV64 and 90 // fcvt.w(u).s/d/h on RV32. Unlike FP_TO_S/UINT these saturate out of 91 // range inputs. These are used for FP_TO_S/UINT_SAT lowering. Rounding mode 92 // is passed as a TargetConstant operand using the RISCVFPRndMode enum. 93 FCVT_X, 94 FCVT_XU, 95 // FP to 32 bit int conversions for RV64. These are used to keep track of the 96 // result being sign extended to 64 bit. These saturate out of range inputs. 97 // Used for FP_TO_S/UINT and FP_TO_S/UINT_SAT lowering. Rounding mode 98 // is passed as a TargetConstant operand using the RISCVFPRndMode enum. 99 FCVT_W_RV64, 100 FCVT_WU_RV64, 101 // READ_CYCLE_WIDE - A read of the 64-bit cycle CSR on a 32-bit target 102 // (returns (Lo, Hi)). It takes a chain operand. 103 READ_CYCLE_WIDE, 104 // Generalized Reverse and Generalized Or-Combine - directly matching the 105 // semantics of the named RISC-V instructions. Lowered as custom nodes as 106 // TableGen chokes when faced with commutative permutations in deeply-nested 107 // DAGs. Each node takes an input operand and a control operand and outputs a 108 // bit-manipulated version of input. All operands are i32 or XLenVT. 109 GREV, 110 GREVW, 111 GORC, 112 GORCW, 113 SHFL, 114 SHFLW, 115 UNSHFL, 116 UNSHFLW, 117 // Bit Compress/Decompress implement the generic bit extract and bit deposit 118 // functions. This operation is also referred to as bit gather/scatter, bit 119 // pack/unpack, parallel extract/deposit, compress/expand, or right 120 // compress/right expand. 121 BCOMPRESS, 122 BCOMPRESSW, 123 BDECOMPRESS, 124 BDECOMPRESSW, 125 // The bit field place (bfp) instruction places up to XLEN/2 LSB bits from rs2 126 // into the value in rs1. The upper bits of rs2 control the length of the bit 127 // field and target position. The layout of rs2 is chosen in a way that makes 128 // it possible to construct rs2 easily using pack[h] instructions and/or 129 // andi/lui. 130 BFP, 131 BFPW, 132 // Vector Extension 133 // VMV_V_X_VL matches the semantics of vmv.v.x but includes an extra operand 134 // for the VL value to be used for the operation. The first operand is 135 // passthru operand. 136 VMV_V_X_VL, 137 // VFMV_V_F_VL matches the semantics of vfmv.v.f but includes an extra operand 138 // for the VL value to be used for the operation. The first operand is 139 // passthru operand. 140 VFMV_V_F_VL, 141 // VMV_X_S matches the semantics of vmv.x.s. The result is always XLenVT sign 142 // extended from the vector element size. 143 VMV_X_S, 144 // VMV_S_X_VL matches the semantics of vmv.s.x. It carries a VL operand. 145 VMV_S_X_VL, 146 // VFMV_S_F_VL matches the semantics of vfmv.s.f. It carries a VL operand. 147 VFMV_S_F_VL, 148 // Splats an 64-bit value that has been split into two i32 parts. This is 149 // expanded late to two scalar stores and a stride 0 vector load. 150 // The first operand is passthru operand. 151 SPLAT_VECTOR_SPLIT_I64_VL, 152 // Read VLENB CSR 153 READ_VLENB, 154 // Truncates a RVV integer vector by one power-of-two. Carries both an extra 155 // mask and VL operand. 156 TRUNCATE_VECTOR_VL, 157 // Matches the semantics of vslideup/vslidedown. The first operand is the 158 // pass-thru operand, the second is the source vector, the third is the 159 // XLenVT index (either constant or non-constant), the fourth is the mask 160 // and the fifth the VL. 161 VSLIDEUP_VL, 162 VSLIDEDOWN_VL, 163 // Matches the semantics of vslide1up/slide1down. The first operand is 164 // passthru operand, the second is source vector, third is the XLenVT scalar 165 // value. The fourth and fifth operands are the mask and VL operands. 166 VSLIDE1UP_VL, 167 VSLIDE1DOWN_VL, 168 // Matches the semantics of the vid.v instruction, with a mask and VL 169 // operand. 170 VID_VL, 171 // Matches the semantics of the vfcnvt.rod function (Convert double-width 172 // float to single-width float, rounding towards odd). Takes a double-width 173 // float vector and produces a single-width float vector. Also has a mask and 174 // VL operand. 175 VFNCVT_ROD_VL, 176 // These nodes match the semantics of the corresponding RVV vector reduction 177 // instructions. They produce a vector result which is the reduction 178 // performed over the second vector operand plus the first element of the 179 // third vector operand. The first operand is the pass-thru operand. The 180 // second operand is an unconstrained vector type, and the result, first, and 181 // third operand's types are expected to be the corresponding full-width 182 // LMUL=1 type for the second operand: 183 // nxv8i8 = vecreduce_add nxv8i8, nxv32i8, nxv8i8 184 // nxv2i32 = vecreduce_add nxv2i32, nxv8i32, nxv2i32 185 // The different in types does introduce extra vsetvli instructions but 186 // similarly it reduces the number of registers consumed per reduction. 187 // Also has a mask and VL operand. 188 VECREDUCE_ADD_VL, 189 VECREDUCE_UMAX_VL, 190 VECREDUCE_SMAX_VL, 191 VECREDUCE_UMIN_VL, 192 VECREDUCE_SMIN_VL, 193 VECREDUCE_AND_VL, 194 VECREDUCE_OR_VL, 195 VECREDUCE_XOR_VL, 196 VECREDUCE_FADD_VL, 197 VECREDUCE_SEQ_FADD_VL, 198 VECREDUCE_FMIN_VL, 199 VECREDUCE_FMAX_VL, 200 201 // Vector binary and unary ops with a mask as a third operand, and VL as a 202 // fourth operand. 203 // FIXME: Can we replace these with ISD::VP_*? 204 ADD_VL, 205 AND_VL, 206 MUL_VL, 207 OR_VL, 208 SDIV_VL, 209 SHL_VL, 210 SREM_VL, 211 SRA_VL, 212 SRL_VL, 213 SUB_VL, 214 UDIV_VL, 215 UREM_VL, 216 XOR_VL, 217 218 SADDSAT_VL, 219 UADDSAT_VL, 220 SSUBSAT_VL, 221 USUBSAT_VL, 222 223 FADD_VL, 224 FSUB_VL, 225 FMUL_VL, 226 FDIV_VL, 227 FNEG_VL, 228 FABS_VL, 229 FSQRT_VL, 230 FMA_VL, 231 FCOPYSIGN_VL, 232 SMIN_VL, 233 SMAX_VL, 234 UMIN_VL, 235 UMAX_VL, 236 FMINNUM_VL, 237 FMAXNUM_VL, 238 MULHS_VL, 239 MULHU_VL, 240 FP_TO_SINT_VL, 241 FP_TO_UINT_VL, 242 SINT_TO_FP_VL, 243 UINT_TO_FP_VL, 244 FP_ROUND_VL, 245 FP_EXTEND_VL, 246 247 // Widening instructions 248 VWMUL_VL, 249 VWMULU_VL, 250 VWMULSU_VL, 251 VWADD_VL, 252 VWADDU_VL, 253 VWSUB_VL, 254 VWSUBU_VL, 255 VWADD_W_VL, 256 VWADDU_W_VL, 257 VWSUB_W_VL, 258 VWSUBU_W_VL, 259 260 // Vector compare producing a mask. Fourth operand is input mask. Fifth 261 // operand is VL. 262 SETCC_VL, 263 264 // Vector select with an additional VL operand. This operation is unmasked. 265 VSELECT_VL, 266 // Vector select with operand #2 (the value when the condition is false) tied 267 // to the destination and an additional VL operand. This operation is 268 // unmasked. 269 VP_MERGE_VL, 270 271 // Mask binary operators. 272 VMAND_VL, 273 VMOR_VL, 274 VMXOR_VL, 275 276 // Set mask vector to all zeros or ones. 277 VMCLR_VL, 278 VMSET_VL, 279 280 // Matches the semantics of vrgather.vx and vrgather.vv with an extra operand 281 // for VL. 282 VRGATHER_VX_VL, 283 VRGATHER_VV_VL, 284 VRGATHEREI16_VV_VL, 285 286 // Vector sign/zero extend with additional mask & VL operands. 287 VSEXT_VL, 288 VZEXT_VL, 289 290 // vcpop.m with additional mask and VL operands. 291 VCPOP_VL, 292 293 // Reads value of CSR. 294 // The first operand is a chain pointer. The second specifies address of the 295 // required CSR. Two results are produced, the read value and the new chain 296 // pointer. 297 READ_CSR, 298 // Write value to CSR. 299 // The first operand is a chain pointer, the second specifies address of the 300 // required CSR and the third is the value to write. The result is the new 301 // chain pointer. 302 WRITE_CSR, 303 // Read and write value of CSR. 304 // The first operand is a chain pointer, the second specifies address of the 305 // required CSR and the third is the value to write. Two results are produced, 306 // the value read before the modification and the new chain pointer. 307 SWAP_CSR, 308 309 // FP to 32 bit int conversions for RV64. These are used to keep track of the 310 // result being sign extended to 64 bit. These saturate out of range inputs. 311 STRICT_FCVT_W_RV64 = ISD::FIRST_TARGET_STRICTFP_OPCODE, 312 STRICT_FCVT_WU_RV64, 313 314 // WARNING: Do not add anything in the end unless you want the node to 315 // have memop! In fact, starting from FIRST_TARGET_MEMORY_OPCODE all 316 // opcodes will be thought as target memory ops! 317 }; 318 } // namespace RISCVISD 319 320 namespace RISCV { 321 // We use 64 bits as the known part in the scalable vector types. 322 static constexpr unsigned RVVBitsPerBlock = 64; 323 } // namespace RISCV 324 325 class RISCVTargetLowering : public TargetLowering { 326 const RISCVSubtarget &Subtarget; 327 328 public: 329 explicit RISCVTargetLowering(const TargetMachine &TM, 330 const RISCVSubtarget &STI); 331 332 const RISCVSubtarget &getSubtarget() const { return Subtarget; } 333 334 bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I, 335 MachineFunction &MF, 336 unsigned Intrinsic) const override; 337 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty, 338 unsigned AS, 339 Instruction *I = nullptr) const override; 340 bool isLegalICmpImmediate(int64_t Imm) const override; 341 bool isLegalAddImmediate(int64_t Imm) const override; 342 bool isTruncateFree(Type *SrcTy, Type *DstTy) const override; 343 bool isTruncateFree(EVT SrcVT, EVT DstVT) const override; 344 bool isZExtFree(SDValue Val, EVT VT2) const override; 345 bool isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const override; 346 bool signExtendConstant(const ConstantInt *CI) const override; 347 bool isCheapToSpeculateCttz() const override; 348 bool isCheapToSpeculateCtlz() const override; 349 bool hasAndNotCompare(SDValue Y) const override; 350 bool hasBitTest(SDValue X, SDValue Y) const override; 351 bool shouldSinkOperands(Instruction *I, 352 SmallVectorImpl<Use *> &Ops) const override; 353 bool isFPImmLegal(const APFloat &Imm, EVT VT, 354 bool ForCodeSize) const override; 355 356 bool softPromoteHalfType() const override { return true; } 357 358 /// Return the register type for a given MVT, ensuring vectors are treated 359 /// as a series of gpr sized integers. 360 MVT getRegisterTypeForCallingConv(LLVMContext &Context, CallingConv::ID CC, 361 EVT VT) const override; 362 363 /// Return the number of registers for a given MVT, ensuring vectors are 364 /// treated as a series of gpr sized integers. 365 unsigned getNumRegistersForCallingConv(LLVMContext &Context, 366 CallingConv::ID CC, 367 EVT VT) const override; 368 369 /// Return true if the given shuffle mask can be codegen'd directly, or if it 370 /// should be stack expanded. 371 bool isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const override; 372 373 bool hasBitPreservingFPLogic(EVT VT) const override; 374 bool 375 shouldExpandBuildVectorWithShuffles(EVT VT, 376 unsigned DefinedValues) const override; 377 378 // Provide custom lowering hooks for some operations. 379 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; 380 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results, 381 SelectionDAG &DAG) const override; 382 383 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; 384 385 bool targetShrinkDemandedConstant(SDValue Op, const APInt &DemandedBits, 386 const APInt &DemandedElts, 387 TargetLoweringOpt &TLO) const override; 388 389 void computeKnownBitsForTargetNode(const SDValue Op, 390 KnownBits &Known, 391 const APInt &DemandedElts, 392 const SelectionDAG &DAG, 393 unsigned Depth) const override; 394 unsigned ComputeNumSignBitsForTargetNode(SDValue Op, 395 const APInt &DemandedElts, 396 const SelectionDAG &DAG, 397 unsigned Depth) const override; 398 399 // This method returns the name of a target specific DAG node. 400 const char *getTargetNodeName(unsigned Opcode) const override; 401 402 ConstraintType getConstraintType(StringRef Constraint) const override; 403 404 unsigned getInlineAsmMemConstraint(StringRef ConstraintCode) const override; 405 406 std::pair<unsigned, const TargetRegisterClass *> 407 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, 408 StringRef Constraint, MVT VT) const override; 409 410 void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint, 411 std::vector<SDValue> &Ops, 412 SelectionDAG &DAG) const override; 413 414 MachineBasicBlock * 415 EmitInstrWithCustomInserter(MachineInstr &MI, 416 MachineBasicBlock *BB) const override; 417 418 void AdjustInstrPostInstrSelection(MachineInstr &MI, 419 SDNode *Node) const override; 420 421 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context, 422 EVT VT) const override; 423 424 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override { 425 return VT.isScalarInteger(); 426 } 427 bool convertSelectOfConstantsToMath(EVT VT) const override { return true; } 428 429 bool shouldInsertFencesForAtomic(const Instruction *I) const override { 430 return isa<LoadInst>(I) || isa<StoreInst>(I); 431 } 432 Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst, 433 AtomicOrdering Ord) const override; 434 Instruction *emitTrailingFence(IRBuilderBase &Builder, Instruction *Inst, 435 AtomicOrdering Ord) const override; 436 437 bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF, 438 EVT VT) const override; 439 440 ISD::NodeType getExtendForAtomicOps() const override { 441 return ISD::SIGN_EXTEND; 442 } 443 444 ISD::NodeType getExtendForAtomicCmpSwapArg() const override { 445 return ISD::SIGN_EXTEND; 446 } 447 448 bool shouldExpandShift(SelectionDAG &DAG, SDNode *N) const override { 449 if (DAG.getMachineFunction().getFunction().hasMinSize()) 450 return false; 451 return true; 452 } 453 bool isDesirableToCommuteWithShift(const SDNode *N, 454 CombineLevel Level) const override; 455 456 /// If a physical register, this returns the register that receives the 457 /// exception address on entry to an EH pad. 458 Register 459 getExceptionPointerRegister(const Constant *PersonalityFn) const override; 460 461 /// If a physical register, this returns the register that receives the 462 /// exception typeid on entry to a landing pad. 463 Register 464 getExceptionSelectorRegister(const Constant *PersonalityFn) const override; 465 466 bool shouldExtendTypeInLibCall(EVT Type) const override; 467 bool shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const override; 468 469 /// Returns the register with the specified architectural or ABI name. This 470 /// method is necessary to lower the llvm.read_register.* and 471 /// llvm.write_register.* intrinsics. Allocatable registers must be reserved 472 /// with the clang -ffixed-xX flag for access to be allowed. 473 Register getRegisterByName(const char *RegName, LLT VT, 474 const MachineFunction &MF) const override; 475 476 // Lower incoming arguments, copy physregs into vregs 477 SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, 478 bool IsVarArg, 479 const SmallVectorImpl<ISD::InputArg> &Ins, 480 const SDLoc &DL, SelectionDAG &DAG, 481 SmallVectorImpl<SDValue> &InVals) const override; 482 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF, 483 bool IsVarArg, 484 const SmallVectorImpl<ISD::OutputArg> &Outs, 485 LLVMContext &Context) const override; 486 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, 487 const SmallVectorImpl<ISD::OutputArg> &Outs, 488 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL, 489 SelectionDAG &DAG) const override; 490 SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI, 491 SmallVectorImpl<SDValue> &InVals) const override; 492 template <class NodeTy> 493 SDValue getAddr(NodeTy *N, SelectionDAG &DAG, bool IsLocal = true) const; 494 495 bool shouldConvertConstantLoadToIntImm(const APInt &Imm, 496 Type *Ty) const override { 497 return true; 498 } 499 bool mayBeEmittedAsTailCall(const CallInst *CI) const override; 500 bool shouldConsiderGEPOffsetSplit() const override { return true; } 501 502 bool decomposeMulByConstant(LLVMContext &Context, EVT VT, 503 SDValue C) const override; 504 505 bool isMulAddWithConstProfitable(SDValue AddNode, 506 SDValue ConstNode) const override; 507 508 TargetLowering::AtomicExpansionKind 509 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override; 510 Value *emitMaskedAtomicRMWIntrinsic(IRBuilderBase &Builder, AtomicRMWInst *AI, 511 Value *AlignedAddr, Value *Incr, 512 Value *Mask, Value *ShiftAmt, 513 AtomicOrdering Ord) const override; 514 TargetLowering::AtomicExpansionKind 515 shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *CI) const override; 516 Value *emitMaskedAtomicCmpXchgIntrinsic(IRBuilderBase &Builder, 517 AtomicCmpXchgInst *CI, 518 Value *AlignedAddr, Value *CmpVal, 519 Value *NewVal, Value *Mask, 520 AtomicOrdering Ord) const override; 521 522 /// Returns true if the target allows unaligned memory accesses of the 523 /// specified type. 524 bool allowsMisalignedMemoryAccesses( 525 EVT VT, unsigned AddrSpace = 0, Align Alignment = Align(1), 526 MachineMemOperand::Flags Flags = MachineMemOperand::MONone, 527 bool *Fast = nullptr) const override; 528 529 bool splitValueIntoRegisterParts(SelectionDAG &DAG, const SDLoc &DL, 530 SDValue Val, SDValue *Parts, 531 unsigned NumParts, MVT PartVT, 532 Optional<CallingConv::ID> CC) const override; 533 534 SDValue 535 joinRegisterPartsIntoValue(SelectionDAG &DAG, const SDLoc &DL, 536 const SDValue *Parts, unsigned NumParts, 537 MVT PartVT, EVT ValueVT, 538 Optional<CallingConv::ID> CC) const override; 539 540 static RISCVII::VLMUL getLMUL(MVT VT); 541 inline static unsigned computeVLMAX(unsigned VectorBits, unsigned EltSize, 542 unsigned MinSize) { 543 // Original equation: 544 // VLMAX = (VectorBits / EltSize) * LMUL 545 // where LMUL = MinSize / RISCV::RVVBitsPerBlock 546 // The following equations have been reordered to prevent loss of precision 547 // when calculating fractional LMUL. 548 return ((VectorBits / EltSize) * MinSize) / RISCV::RVVBitsPerBlock; 549 }; 550 static unsigned getRegClassIDForLMUL(RISCVII::VLMUL LMul); 551 static unsigned getSubregIndexByMVT(MVT VT, unsigned Index); 552 static unsigned getRegClassIDForVecVT(MVT VT); 553 static std::pair<unsigned, unsigned> 554 decomposeSubvectorInsertExtractToSubRegs(MVT VecVT, MVT SubVecVT, 555 unsigned InsertExtractIdx, 556 const RISCVRegisterInfo *TRI); 557 MVT getContainerForFixedLengthVector(MVT VT) const; 558 559 bool shouldRemoveExtendFromGSIndex(EVT VT) const override; 560 561 bool isLegalElementTypeForRVV(Type *ScalarTy) const; 562 563 bool shouldConvertFpToSat(unsigned Op, EVT FPVT, EVT VT) const override; 564 565 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, 566 SmallVectorImpl<SDNode *> &Created) const override; 567 568 unsigned getJumpTableEncoding() const override; 569 570 const MCExpr *LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI, 571 const MachineBasicBlock *MBB, 572 unsigned uid, 573 MCContext &Ctx) const override; 574 575 private: 576 /// RISCVCCAssignFn - This target-specific function extends the default 577 /// CCValAssign with additional information used to lower RISC-V calling 578 /// conventions. 579 typedef bool RISCVCCAssignFn(const DataLayout &DL, RISCVABI::ABI, 580 unsigned ValNo, MVT ValVT, MVT LocVT, 581 CCValAssign::LocInfo LocInfo, 582 ISD::ArgFlagsTy ArgFlags, CCState &State, 583 bool IsFixed, bool IsRet, Type *OrigTy, 584 const RISCVTargetLowering &TLI, 585 Optional<unsigned> FirstMaskArgument); 586 587 void analyzeInputArgs(MachineFunction &MF, CCState &CCInfo, 588 const SmallVectorImpl<ISD::InputArg> &Ins, bool IsRet, 589 RISCVCCAssignFn Fn) const; 590 void analyzeOutputArgs(MachineFunction &MF, CCState &CCInfo, 591 const SmallVectorImpl<ISD::OutputArg> &Outs, 592 bool IsRet, CallLoweringInfo *CLI, 593 RISCVCCAssignFn Fn) const; 594 595 SDValue getStaticTLSAddr(GlobalAddressSDNode *N, SelectionDAG &DAG, 596 bool UseGOT) const; 597 SDValue getDynamicTLSAddr(GlobalAddressSDNode *N, SelectionDAG &DAG) const; 598 599 SDValue lowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const; 600 SDValue lowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; 601 SDValue lowerConstantPool(SDValue Op, SelectionDAG &DAG) const; 602 SDValue lowerJumpTable(SDValue Op, SelectionDAG &DAG) const; 603 SDValue lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const; 604 SDValue lowerSELECT(SDValue Op, SelectionDAG &DAG) const; 605 SDValue lowerBRCOND(SDValue Op, SelectionDAG &DAG) const; 606 SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const; 607 SDValue lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; 608 SDValue lowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; 609 SDValue lowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const; 610 SDValue lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, bool IsSRA) const; 611 SDValue lowerSPLAT_VECTOR_PARTS(SDValue Op, SelectionDAG &DAG) const; 612 SDValue lowerVectorMaskSplat(SDValue Op, SelectionDAG &DAG) const; 613 SDValue lowerVectorMaskExt(SDValue Op, SelectionDAG &DAG, 614 int64_t ExtTrueVal) const; 615 SDValue lowerVectorMaskTruncLike(SDValue Op, SelectionDAG &DAG) const; 616 SDValue lowerVectorTruncLike(SDValue Op, SelectionDAG &DAG) const; 617 SDValue lowerVectorFPRoundLike(SDValue Op, SelectionDAG &DAG) const; 618 SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; 619 SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; 620 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const; 621 SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const; 622 SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const; 623 SDValue lowerVPREDUCE(SDValue Op, SelectionDAG &DAG) const; 624 SDValue lowerVECREDUCE(SDValue Op, SelectionDAG &DAG) const; 625 SDValue lowerVectorMaskVecReduction(SDValue Op, SelectionDAG &DAG, 626 bool IsVP) const; 627 SDValue lowerFPVECREDUCE(SDValue Op, SelectionDAG &DAG) const; 628 SDValue lowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const; 629 SDValue lowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const; 630 SDValue lowerSTEP_VECTOR(SDValue Op, SelectionDAG &DAG) const; 631 SDValue lowerVECTOR_REVERSE(SDValue Op, SelectionDAG &DAG) const; 632 SDValue lowerVECTOR_SPLICE(SDValue Op, SelectionDAG &DAG) const; 633 SDValue lowerABS(SDValue Op, SelectionDAG &DAG) const; 634 SDValue lowerMaskedLoad(SDValue Op, SelectionDAG &DAG) const; 635 SDValue lowerMaskedStore(SDValue Op, SelectionDAG &DAG) const; 636 SDValue lowerFixedLengthVectorFCOPYSIGNToRVV(SDValue Op, 637 SelectionDAG &DAG) const; 638 SDValue lowerMaskedGather(SDValue Op, SelectionDAG &DAG) const; 639 SDValue lowerMaskedScatter(SDValue Op, SelectionDAG &DAG) const; 640 SDValue lowerFixedLengthVectorLoadToRVV(SDValue Op, SelectionDAG &DAG) const; 641 SDValue lowerFixedLengthVectorStoreToRVV(SDValue Op, SelectionDAG &DAG) const; 642 SDValue lowerFixedLengthVectorSetccToRVV(SDValue Op, SelectionDAG &DAG) const; 643 SDValue lowerFixedLengthVectorLogicOpToRVV(SDValue Op, SelectionDAG &DAG, 644 unsigned MaskOpc, 645 unsigned VecOpc) const; 646 SDValue lowerFixedLengthVectorShiftToRVV(SDValue Op, SelectionDAG &DAG) const; 647 SDValue lowerFixedLengthVectorSelectToRVV(SDValue Op, 648 SelectionDAG &DAG) const; 649 SDValue lowerToScalableOp(SDValue Op, SelectionDAG &DAG, unsigned NewOpc, 650 bool HasMask = true) const; 651 SDValue lowerVPOp(SDValue Op, SelectionDAG &DAG, unsigned RISCVISDOpc) const; 652 SDValue lowerLogicVPOp(SDValue Op, SelectionDAG &DAG, unsigned MaskOpc, 653 unsigned VecOpc) const; 654 SDValue lowerVPExtMaskOp(SDValue Op, SelectionDAG &DAG) const; 655 SDValue lowerVPSetCCMaskOp(SDValue Op, SelectionDAG &DAG) const; 656 SDValue lowerVPFPIntConvOp(SDValue Op, SelectionDAG &DAG, 657 unsigned RISCVISDOpc) const; 658 SDValue lowerFixedLengthVectorExtendToRVV(SDValue Op, SelectionDAG &DAG, 659 unsigned ExtendOpc) const; 660 SDValue lowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const; 661 SDValue lowerSET_ROUNDING(SDValue Op, SelectionDAG &DAG) const; 662 663 SDValue expandUnalignedRVVLoad(SDValue Op, SelectionDAG &DAG) const; 664 SDValue expandUnalignedRVVStore(SDValue Op, SelectionDAG &DAG) const; 665 666 bool isEligibleForTailCallOptimization( 667 CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF, 668 const SmallVector<CCValAssign, 16> &ArgLocs) const; 669 670 /// Generate error diagnostics if any register used by CC has been marked 671 /// reserved. 672 void validateCCReservedRegs( 673 const SmallVectorImpl<std::pair<llvm::Register, llvm::SDValue>> &Regs, 674 MachineFunction &MF) const; 675 676 bool useRVVForFixedLengthVectorVT(MVT VT) const; 677 678 MVT getVPExplicitVectorLengthTy() const override; 679 680 /// RVV code generation for fixed length vectors does not lower all 681 /// BUILD_VECTORs. This makes BUILD_VECTOR legalisation a source of stores to 682 /// merge. However, merging them creates a BUILD_VECTOR that is just as 683 /// illegal as the original, thus leading to an infinite legalisation loop. 684 /// NOTE: Once BUILD_VECTOR can be custom lowered for all legal vector types, 685 /// this override can be removed. 686 bool mergeStoresAfterLegalization(EVT VT) const override; 687 688 /// Disable normalizing 689 /// select(N0&N1, X, Y) => select(N0, select(N1, X, Y), Y) and 690 /// select(N0|N1, X, Y) => select(N0, select(N1, X, Y, Y)) 691 /// RISCV doesn't have flags so it's better to perform the and/or in a GPR. 692 bool shouldNormalizeToSelectSequence(LLVMContext &, EVT) const override { 693 return false; 694 }; 695 }; 696 namespace RISCVVIntrinsicsTable { 697 698 struct RISCVVIntrinsicInfo { 699 unsigned IntrinsicID; 700 uint8_t ScalarOperand; 701 uint8_t VLOperand; 702 bool hasScalarOperand() const { 703 // 0xF is not valid. See NoScalarOperand in IntrinsicsRISCV.td. 704 return ScalarOperand != 0xF; 705 } 706 bool hasVLOperand() const { 707 // 0x1F is not valid. See NoVLOperand in IntrinsicsRISCV.td. 708 return VLOperand != 0x1F; 709 } 710 }; 711 712 using namespace RISCV; 713 714 #define GET_RISCVVIntrinsicsTable_DECL 715 #include "RISCVGenSearchableTables.inc" 716 717 } // end namespace RISCVVIntrinsicsTable 718 719 } // end namespace llvm 720 721 #endif 722