1 //===-- AMDGPUISelLowering.h - AMDGPU Lowering Interface --------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 /// \file 11 /// \brief Interface definition of the TargetLowering class that is common 12 /// to all AMD GPUs. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUISELLOWERING_H 17 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUISELLOWERING_H 18 19 #include "AMDGPU.h" 20 #include "llvm/CodeGen/CallingConvLower.h" 21 #include "llvm/Target/TargetLowering.h" 22 23 namespace llvm { 24 25 class AMDGPUMachineFunction; 26 class AMDGPUSubtarget; 27 struct ArgDescriptor; 28 29 class AMDGPUTargetLowering : public TargetLowering { 30 private: 31 /// \returns AMDGPUISD::FFBH_U32 node if the incoming \p Op may have been 32 /// legalized from a smaller type VT. Need to match pre-legalized type because 33 /// the generic legalization inserts the add/sub between the select and 34 /// compare. 35 SDValue getFFBX_U32(SelectionDAG &DAG, SDValue Op, const SDLoc &DL, unsigned Opc) const; 36 37 public: 38 static bool isOrEquivalentToAdd(SelectionDAG &DAG, SDValue Op); 39 40 protected: 41 const AMDGPUSubtarget *Subtarget; 42 AMDGPUAS AMDGPUASI; 43 44 SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const; 45 SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const; 46 /// \brief Split a vector store into multiple scalar stores. 47 /// \returns The resulting chain. 48 49 SDValue LowerFREM(SDValue Op, SelectionDAG &DAG) const; 50 SDValue LowerFCEIL(SDValue Op, SelectionDAG &DAG) const; 51 SDValue LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const; 52 SDValue LowerFRINT(SDValue Op, SelectionDAG &DAG) const; 53 SDValue LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const; 54 55 SDValue LowerFROUND32_16(SDValue Op, SelectionDAG &DAG) const; 56 SDValue LowerFROUND64(SDValue Op, SelectionDAG &DAG) const; 57 SDValue LowerFROUND(SDValue Op, SelectionDAG &DAG) const; 58 SDValue LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const; 59 60 SDValue LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) const; 61 62 SDValue LowerINT_TO_FP32(SDValue Op, SelectionDAG &DAG, bool Signed) const; 63 SDValue LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG, bool Signed) const; 64 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 65 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 66 67 SDValue LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG, bool Signed) const; 68 SDValue LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) const; 69 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const; 70 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const; 71 72 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const; 73 74 protected: 75 bool shouldCombineMemoryType(EVT VT) const; 76 SDValue performLoadCombine(SDNode *N, DAGCombinerInfo &DCI) const; 77 SDValue performStoreCombine(SDNode *N, DAGCombinerInfo &DCI) const; 78 SDValue performClampCombine(SDNode *N, DAGCombinerInfo &DCI) const; 79 SDValue performAssertSZExtCombine(SDNode *N, DAGCombinerInfo &DCI) const; 80 81 SDValue splitBinaryBitConstantOpImpl(DAGCombinerInfo &DCI, const SDLoc &SL, 82 unsigned Opc, SDValue LHS, 83 uint32_t ValLo, uint32_t ValHi) const; 84 SDValue performShlCombine(SDNode *N, DAGCombinerInfo &DCI) const; 85 SDValue performSraCombine(SDNode *N, DAGCombinerInfo &DCI) const; 86 SDValue performSrlCombine(SDNode *N, DAGCombinerInfo &DCI) const; 87 SDValue performMulCombine(SDNode *N, DAGCombinerInfo &DCI) const; 88 SDValue performMulhsCombine(SDNode *N, DAGCombinerInfo &DCI) const; 89 SDValue performMulhuCombine(SDNode *N, DAGCombinerInfo &DCI) const; 90 SDValue performMulLoHi24Combine(SDNode *N, DAGCombinerInfo &DCI) const; 91 SDValue performCtlz_CttzCombine(const SDLoc &SL, SDValue Cond, SDValue LHS, 92 SDValue RHS, DAGCombinerInfo &DCI) const; 93 SDValue performSelectCombine(SDNode *N, DAGCombinerInfo &DCI) const; 94 SDValue performFNegCombine(SDNode *N, DAGCombinerInfo &DCI) const; 95 SDValue performFAbsCombine(SDNode *N, DAGCombinerInfo &DCI) const; 96 97 static EVT getEquivalentMemType(LLVMContext &Context, EVT VT); 98 99 virtual SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op, 100 SelectionDAG &DAG) const; 101 102 /// Return 64-bit value Op as two 32-bit integers. 103 std::pair<SDValue, SDValue> split64BitValue(SDValue Op, 104 SelectionDAG &DAG) const; 105 SDValue getLoHalf64(SDValue Op, SelectionDAG &DAG) const; 106 SDValue getHiHalf64(SDValue Op, SelectionDAG &DAG) const; 107 108 /// \brief Split a vector load into 2 loads of half the vector. 109 SDValue SplitVectorLoad(SDValue Op, SelectionDAG &DAG) const; 110 111 /// \brief Split a vector store into 2 stores of half the vector. 112 SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const; 113 114 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const; 115 SDValue LowerSDIVREM(SDValue Op, SelectionDAG &DAG) const; 116 SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const; 117 SDValue LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const; 118 void LowerUDIVREM64(SDValue Op, SelectionDAG &DAG, 119 SmallVectorImpl<SDValue> &Results) const; 120 void analyzeFormalArgumentsCompute(CCState &State, 121 const SmallVectorImpl<ISD::InputArg> &Ins) const; 122 public: 123 AMDGPUTargetLowering(const TargetMachine &TM, const AMDGPUSubtarget &STI); 124 125 bool mayIgnoreSignedZero(SDValue Op) const { 126 if (getTargetMachine().Options.NoSignedZerosFPMath) 127 return true; 128 129 const auto Flags = Op.getNode()->getFlags(); 130 if (Flags.isDefined()) 131 return Flags.hasNoSignedZeros(); 132 133 return false; 134 } 135 136 static bool allUsesHaveSourceMods(const SDNode *N, 137 unsigned CostThreshold = 4); 138 bool isFAbsFree(EVT VT) const override; 139 bool isFNegFree(EVT VT) const override; 140 bool isTruncateFree(EVT Src, EVT Dest) const override; 141 bool isTruncateFree(Type *Src, Type *Dest) const override; 142 143 bool isZExtFree(Type *Src, Type *Dest) const override; 144 bool isZExtFree(EVT Src, EVT Dest) const override; 145 bool isZExtFree(SDValue Val, EVT VT2) const override; 146 bool isFPExtFoldable(unsigned Opcode, EVT DestVT, EVT SrcVT) const override; 147 148 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override; 149 150 MVT getVectorIdxTy(const DataLayout &) const override; 151 bool isSelectSupported(SelectSupportKind) const override; 152 153 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override; 154 bool ShouldShrinkFPConstant(EVT VT) const override; 155 bool shouldReduceLoadWidth(SDNode *Load, 156 ISD::LoadExtType ExtType, 157 EVT ExtVT) const override; 158 159 bool isLoadBitCastBeneficial(EVT, EVT) const final; 160 161 bool storeOfVectorConstantIsCheap(EVT MemVT, 162 unsigned NumElem, 163 unsigned AS) const override; 164 bool aggressivelyPreferBuildVectorSources(EVT VecVT) const override; 165 bool isCheapToSpeculateCttz() const override; 166 bool isCheapToSpeculateCtlz() const override; 167 168 static CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool IsVarArg); 169 static CCAssignFn *CCAssignFnForReturn(CallingConv::ID CC, bool IsVarArg); 170 171 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, 172 const SmallVectorImpl<ISD::OutputArg> &Outs, 173 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL, 174 SelectionDAG &DAG) const override; 175 176 SDValue addTokenForArgument(SDValue Chain, 177 SelectionDAG &DAG, 178 MachineFrameInfo &MFI, 179 int ClobberedFI) const; 180 181 SDValue lowerUnhandledCall(CallLoweringInfo &CLI, 182 SmallVectorImpl<SDValue> &InVals, 183 StringRef Reason) const; 184 SDValue LowerCall(CallLoweringInfo &CLI, 185 SmallVectorImpl<SDValue> &InVals) const override; 186 187 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, 188 SelectionDAG &DAG) const; 189 190 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; 191 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; 192 void ReplaceNodeResults(SDNode * N, 193 SmallVectorImpl<SDValue> &Results, 194 SelectionDAG &DAG) const override; 195 196 SDValue combineFMinMaxLegacy(const SDLoc &DL, EVT VT, SDValue LHS, 197 SDValue RHS, SDValue True, SDValue False, 198 SDValue CC, DAGCombinerInfo &DCI) const; 199 200 const char* getTargetNodeName(unsigned Opcode) const override; 201 202 bool isFsqrtCheap(SDValue Operand, SelectionDAG &DAG) const override { 203 return true; 204 } 205 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled, 206 int &RefinementSteps, bool &UseOneConstNR, 207 bool Reciprocal) const override; 208 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled, 209 int &RefinementSteps) const override; 210 211 virtual SDNode *PostISelFolding(MachineSDNode *N, 212 SelectionDAG &DAG) const = 0; 213 214 /// \brief Determine which of the bits specified in \p Mask are known to be 215 /// either zero or one and return them in the \p KnownZero and \p KnownOne 216 /// bitsets. 217 void computeKnownBitsForTargetNode(const SDValue Op, 218 KnownBits &Known, 219 const APInt &DemandedElts, 220 const SelectionDAG &DAG, 221 unsigned Depth = 0) const override; 222 223 unsigned ComputeNumSignBitsForTargetNode(SDValue Op, const APInt &DemandedElts, 224 const SelectionDAG &DAG, 225 unsigned Depth = 0) const override; 226 227 /// \brief Helper function that adds Reg to the LiveIn list of the DAG's 228 /// MachineFunction. 229 /// 230 /// \returns a RegisterSDNode representing Reg if \p RawReg is true, otherwise 231 /// a copy from the register. 232 SDValue CreateLiveInRegister(SelectionDAG &DAG, 233 const TargetRegisterClass *RC, 234 unsigned Reg, EVT VT, 235 const SDLoc &SL, 236 bool RawReg = false) const; 237 SDValue CreateLiveInRegister(SelectionDAG &DAG, 238 const TargetRegisterClass *RC, 239 unsigned Reg, EVT VT) const { 240 return CreateLiveInRegister(DAG, RC, Reg, VT, SDLoc(DAG.getEntryNode())); 241 } 242 243 // Returns the raw live in register rather than a copy from it. 244 SDValue CreateLiveInRegisterRaw(SelectionDAG &DAG, 245 const TargetRegisterClass *RC, 246 unsigned Reg, EVT VT) const { 247 return CreateLiveInRegister(DAG, RC, Reg, VT, SDLoc(DAG.getEntryNode()), true); 248 } 249 250 /// Similar to CreateLiveInRegister, except value maybe loaded from a stack 251 /// slot rather than passed in a register. 252 SDValue loadStackInputValue(SelectionDAG &DAG, 253 EVT VT, 254 const SDLoc &SL, 255 int64_t Offset) const; 256 257 SDValue storeStackInputValue(SelectionDAG &DAG, 258 const SDLoc &SL, 259 SDValue Chain, 260 SDValue StackPtr, 261 SDValue ArgVal, 262 int64_t Offset) const; 263 264 SDValue loadInputValue(SelectionDAG &DAG, 265 const TargetRegisterClass *RC, 266 EVT VT, const SDLoc &SL, 267 const ArgDescriptor &Arg) const; 268 269 enum ImplicitParameter { 270 FIRST_IMPLICIT, 271 GRID_DIM = FIRST_IMPLICIT, 272 GRID_OFFSET, 273 }; 274 275 /// \brief Helper function that returns the byte offset of the given 276 /// type of implicit parameter. 277 uint32_t getImplicitParameterOffset(const AMDGPUMachineFunction *MFI, 278 const ImplicitParameter Param) const; 279 280 AMDGPUAS getAMDGPUAS() const { 281 return AMDGPUASI; 282 } 283 284 MVT getFenceOperandTy(const DataLayout &DL) const override { 285 return MVT::i32; 286 } 287 }; 288 289 namespace AMDGPUISD { 290 291 enum NodeType : unsigned { 292 // AMDIL ISD Opcodes 293 FIRST_NUMBER = ISD::BUILTIN_OP_END, 294 UMUL, // 32bit unsigned multiplication 295 BRANCH_COND, 296 // End AMDIL ISD Opcodes 297 298 // Function call. 299 CALL, 300 TC_RETURN, 301 TRAP, 302 303 // Masked control flow nodes. 304 IF, 305 ELSE, 306 LOOP, 307 308 // A uniform kernel return that terminates the wavefront. 309 ENDPGM, 310 311 // Return to a shader part's epilog code. 312 RETURN_TO_EPILOG, 313 314 // Return with values from a non-entry function. 315 RET_FLAG, 316 317 DWORDADDR, 318 FRACT, 319 320 /// CLAMP value between 0.0 and 1.0. NaN clamped to 0, following clamp output 321 /// modifier behavior with dx10_enable. 322 CLAMP, 323 324 // This is SETCC with the full mask result which is used for a compare with a 325 // result bit per item in the wavefront. 326 SETCC, 327 SETREG, 328 // FP ops with input and output chain. 329 FMA_W_CHAIN, 330 FMUL_W_CHAIN, 331 332 // SIN_HW, COS_HW - f32 for SI, 1 ULP max error, valid from -100 pi to 100 pi. 333 // Denormals handled on some parts. 334 COS_HW, 335 SIN_HW, 336 FMAX_LEGACY, 337 FMIN_LEGACY, 338 FMAX3, 339 SMAX3, 340 UMAX3, 341 FMIN3, 342 SMIN3, 343 UMIN3, 344 FMED3, 345 SMED3, 346 UMED3, 347 URECIP, 348 DIV_SCALE, 349 DIV_FMAS, 350 DIV_FIXUP, 351 // For emitting ISD::FMAD when f32 denormals are enabled because mac/mad is 352 // treated as an illegal operation. 353 FMAD_FTZ, 354 TRIG_PREOP, // 1 ULP max error for f64 355 356 // RCP, RSQ - For f32, 1 ULP max error, no denormal handling. 357 // For f64, max error 2^29 ULP, handles denormals. 358 RCP, 359 RSQ, 360 RCP_LEGACY, 361 RSQ_LEGACY, 362 FMUL_LEGACY, 363 RSQ_CLAMP, 364 LDEXP, 365 FP_CLASS, 366 DOT4, 367 CARRY, 368 BORROW, 369 BFE_U32, // Extract range of bits with zero extension to 32-bits. 370 BFE_I32, // Extract range of bits with sign extension to 32-bits. 371 BFI, // (src0 & src1) | (~src0 & src2) 372 BFM, // Insert a range of bits into a 32-bit word. 373 FFBH_U32, // ctlz with -1 if input is zero. 374 FFBH_I32, 375 FFBL_B32, // cttz with -1 if input is zero. 376 MUL_U24, 377 MUL_I24, 378 MULHI_U24, 379 MULHI_I24, 380 MAD_U24, 381 MAD_I24, 382 MUL_LOHI_I24, 383 MUL_LOHI_U24, 384 TEXTURE_FETCH, 385 EXPORT, // exp on SI+ 386 EXPORT_DONE, // exp on SI+ with done bit set 387 R600_EXPORT, 388 CONST_ADDRESS, 389 REGISTER_LOAD, 390 REGISTER_STORE, 391 SAMPLE, 392 SAMPLEB, 393 SAMPLED, 394 SAMPLEL, 395 396 // These cvt_f32_ubyte* nodes need to remain consecutive and in order. 397 CVT_F32_UBYTE0, 398 CVT_F32_UBYTE1, 399 CVT_F32_UBYTE2, 400 CVT_F32_UBYTE3, 401 402 // Convert two float 32 numbers into a single register holding two packed f16 403 // with round to zero. 404 CVT_PKRTZ_F16_F32, 405 406 // Same as the standard node, except the high bits of the resulting integer 407 // are known 0. 408 FP_TO_FP16, 409 410 // Wrapper around fp16 results that are known to zero the high bits. 411 FP16_ZEXT, 412 413 /// This node is for VLIW targets and it is used to represent a vector 414 /// that is stored in consecutive registers with the same channel. 415 /// For example: 416 /// |X |Y|Z|W| 417 /// T0|v.x| | | | 418 /// T1|v.y| | | | 419 /// T2|v.z| | | | 420 /// T3|v.w| | | | 421 BUILD_VERTICAL_VECTOR, 422 /// Pointer to the start of the shader's constant data. 423 CONST_DATA_PTR, 424 INIT_EXEC, 425 INIT_EXEC_FROM_INPUT, 426 SENDMSG, 427 SENDMSGHALT, 428 INTERP_MOV, 429 INTERP_P1, 430 INTERP_P2, 431 PC_ADD_REL_OFFSET, 432 KILL, 433 DUMMY_CHAIN, 434 FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE, 435 STORE_MSKOR, 436 LOAD_CONSTANT, 437 TBUFFER_STORE_FORMAT, 438 TBUFFER_STORE_FORMAT_X3, 439 TBUFFER_LOAD_FORMAT, 440 ATOMIC_CMP_SWAP, 441 ATOMIC_INC, 442 ATOMIC_DEC, 443 BUFFER_LOAD, 444 BUFFER_LOAD_FORMAT, 445 LAST_AMDGPU_ISD_NUMBER 446 }; 447 448 449 } // End namespace AMDGPUISD 450 451 } // End namespace llvm 452 453 #endif 454