1 //===-- SIISelLowering.h - SI 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 /// \file 10 /// SI DAG Lowering interface definition 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_LIB_TARGET_AMDGPU_SIISELLOWERING_H 15 #define LLVM_LIB_TARGET_AMDGPU_SIISELLOWERING_H 16 17 #include "AMDGPUISelLowering.h" 18 #include "AMDGPUArgumentUsageInfo.h" 19 #include "SIInstrInfo.h" 20 21 namespace llvm { 22 23 class SITargetLowering final : public AMDGPUTargetLowering { 24 private: 25 const GCNSubtarget *Subtarget; 26 27 public: 28 MVT getRegisterTypeForCallingConv(LLVMContext &Context, 29 CallingConv::ID CC, 30 EVT VT) const override; 31 unsigned getNumRegistersForCallingConv(LLVMContext &Context, 32 CallingConv::ID CC, 33 EVT VT) const override; 34 35 unsigned getVectorTypeBreakdownForCallingConv( 36 LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT, 37 unsigned &NumIntermediates, MVT &RegisterVT) const override; 38 39 private: 40 SDValue lowerKernArgParameterPtr(SelectionDAG &DAG, const SDLoc &SL, 41 SDValue Chain, uint64_t Offset) const; 42 SDValue getImplicitArgPtr(SelectionDAG &DAG, const SDLoc &SL) const; 43 SDValue lowerKernargMemParameter(SelectionDAG &DAG, EVT VT, EVT MemVT, 44 const SDLoc &SL, SDValue Chain, 45 uint64_t Offset, unsigned Align, bool Signed, 46 const ISD::InputArg *Arg = nullptr) const; 47 48 SDValue lowerStackParameter(SelectionDAG &DAG, CCValAssign &VA, 49 const SDLoc &SL, SDValue Chain, 50 const ISD::InputArg &Arg) const; 51 SDValue getPreloadedValue(SelectionDAG &DAG, 52 const SIMachineFunctionInfo &MFI, 53 EVT VT, 54 AMDGPUFunctionArgInfo::PreloadedValue) const; 55 56 SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op, 57 SelectionDAG &DAG) const override; 58 SDValue lowerImplicitZextParam(SelectionDAG &DAG, SDValue Op, 59 MVT VT, unsigned Offset) const; 60 SDValue lowerImage(SDValue Op, const AMDGPU::ImageDimIntrinsicInfo *Intr, 61 SelectionDAG &DAG) const; 62 SDValue lowerSBuffer(EVT VT, SDLoc DL, SDValue Rsrc, SDValue Offset, 63 SDValue GLC, SDValue DLC, SelectionDAG &DAG) const; 64 65 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const; 66 SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const; 67 SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const; 68 69 // The raw.tbuffer and struct.tbuffer intrinsics have two offset args: offset 70 // (the offset that is included in bounds checking and swizzling, to be split 71 // between the instruction's voffset and immoffset fields) and soffset (the 72 // offset that is excluded from bounds checking and swizzling, to go in the 73 // instruction's soffset field). This function takes the first kind of 74 // offset and figures out how to split it between voffset and immoffset. 75 std::pair<SDValue, SDValue> splitBufferOffsets(SDValue Offset, 76 SelectionDAG &DAG) const; 77 78 SDValue widenLoad(LoadSDNode *Ld, DAGCombinerInfo &DCI) const; 79 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const; 80 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const; 81 SDValue lowerFastUnsafeFDIV(SDValue Op, SelectionDAG &DAG) const; 82 SDValue lowerFDIV_FAST(SDValue Op, SelectionDAG &DAG) const; 83 SDValue LowerFDIV16(SDValue Op, SelectionDAG &DAG) const; 84 SDValue LowerFDIV32(SDValue Op, SelectionDAG &DAG) const; 85 SDValue LowerFDIV64(SDValue Op, SelectionDAG &DAG) const; 86 SDValue LowerFDIV(SDValue Op, SelectionDAG &DAG) const; 87 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG, bool Signed) const; 88 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const; 89 SDValue LowerTrig(SDValue Op, SelectionDAG &DAG) const; 90 SDValue LowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) const; 91 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const; 92 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; 93 SDValue adjustLoadValueType(unsigned Opcode, MemSDNode *M, 94 SelectionDAG &DAG, ArrayRef<SDValue> Ops, 95 bool IsIntrinsic = false) const; 96 97 SDValue lowerIntrinsicLoad(MemSDNode *M, bool IsFormat, SelectionDAG &DAG, 98 ArrayRef<SDValue> Ops) const; 99 100 // Call DAG.getMemIntrinsicNode for a load, but first widen a dwordx3 type to 101 // dwordx4 if on SI. 102 SDValue getMemIntrinsicNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList, 103 ArrayRef<SDValue> Ops, EVT MemVT, 104 MachineMemOperand *MMO, SelectionDAG &DAG) const; 105 106 SDValue handleD16VData(SDValue VData, SelectionDAG &DAG) const; 107 108 /// Converts \p Op, which must be of floating point type, to the 109 /// floating point type \p VT, by either extending or truncating it. 110 SDValue getFPExtOrFPTrunc(SelectionDAG &DAG, 111 SDValue Op, 112 const SDLoc &DL, 113 EVT VT) const; 114 115 SDValue convertArgType( 116 SelectionDAG &DAG, EVT VT, EVT MemVT, const SDLoc &SL, SDValue Val, 117 bool Signed, const ISD::InputArg *Arg = nullptr) const; 118 119 /// Custom lowering for ISD::FP_ROUND for MVT::f16. 120 SDValue lowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const; 121 SDValue lowerFMINNUM_FMAXNUM(SDValue Op, SelectionDAG &DAG) const; 122 123 SDValue getSegmentAperture(unsigned AS, const SDLoc &DL, 124 SelectionDAG &DAG) const; 125 126 SDValue lowerADDRSPACECAST(SDValue Op, SelectionDAG &DAG) const; 127 SDValue lowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const; 128 SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; 129 SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; 130 SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const; 131 SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const; 132 SDValue lowerTRAP(SDValue Op, SelectionDAG &DAG) const; 133 SDValue lowerDEBUGTRAP(SDValue Op, SelectionDAG &DAG) const; 134 135 SDNode *adjustWritemask(MachineSDNode *&N, SelectionDAG &DAG) const; 136 137 SDValue performUCharToFloatCombine(SDNode *N, 138 DAGCombinerInfo &DCI) const; 139 SDValue performSHLPtrCombine(SDNode *N, 140 unsigned AS, 141 EVT MemVT, 142 DAGCombinerInfo &DCI) const; 143 144 SDValue performMemSDNodeCombine(MemSDNode *N, DAGCombinerInfo &DCI) const; 145 146 SDValue splitBinaryBitConstantOp(DAGCombinerInfo &DCI, const SDLoc &SL, 147 unsigned Opc, SDValue LHS, 148 const ConstantSDNode *CRHS) const; 149 150 SDValue performAndCombine(SDNode *N, DAGCombinerInfo &DCI) const; 151 SDValue performOrCombine(SDNode *N, DAGCombinerInfo &DCI) const; 152 SDValue performXorCombine(SDNode *N, DAGCombinerInfo &DCI) const; 153 SDValue performZeroExtendCombine(SDNode *N, DAGCombinerInfo &DCI) const; 154 SDValue performSignExtendInRegCombine(SDNode *N, DAGCombinerInfo &DCI) const; 155 SDValue performClassCombine(SDNode *N, DAGCombinerInfo &DCI) const; 156 SDValue getCanonicalConstantFP(SelectionDAG &DAG, const SDLoc &SL, EVT VT, 157 const APFloat &C) const; 158 SDValue performFCanonicalizeCombine(SDNode *N, DAGCombinerInfo &DCI) const; 159 160 SDValue performFPMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL, 161 SDValue Op0, SDValue Op1) const; 162 SDValue performIntMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL, 163 SDValue Op0, SDValue Op1, bool Signed) const; 164 SDValue performMinMaxCombine(SDNode *N, DAGCombinerInfo &DCI) const; 165 SDValue performFMed3Combine(SDNode *N, DAGCombinerInfo &DCI) const; 166 SDValue performCvtPkRTZCombine(SDNode *N, DAGCombinerInfo &DCI) const; 167 SDValue performExtractVectorEltCombine(SDNode *N, DAGCombinerInfo &DCI) const; 168 SDValue performInsertVectorEltCombine(SDNode *N, DAGCombinerInfo &DCI) const; 169 170 SDValue reassociateScalarOps(SDNode *N, SelectionDAG &DAG) const; 171 unsigned getFusedOpcode(const SelectionDAG &DAG, 172 const SDNode *N0, const SDNode *N1) const; 173 SDValue performAddCombine(SDNode *N, DAGCombinerInfo &DCI) const; 174 SDValue performAddCarrySubCarryCombine(SDNode *N, DAGCombinerInfo &DCI) const; 175 SDValue performSubCombine(SDNode *N, DAGCombinerInfo &DCI) const; 176 SDValue performFAddCombine(SDNode *N, DAGCombinerInfo &DCI) const; 177 SDValue performFSubCombine(SDNode *N, DAGCombinerInfo &DCI) const; 178 SDValue performFMACombine(SDNode *N, DAGCombinerInfo &DCI) const; 179 SDValue performSetCCCombine(SDNode *N, DAGCombinerInfo &DCI) const; 180 SDValue performCvtF32UByteNCombine(SDNode *N, DAGCombinerInfo &DCI) const; 181 SDValue performClampCombine(SDNode *N, DAGCombinerInfo &DCI) const; 182 SDValue performRcpCombine(SDNode *N, DAGCombinerInfo &DCI) const; 183 184 bool isLegalFlatAddressingMode(const AddrMode &AM) const; 185 bool isLegalMUBUFAddressingMode(const AddrMode &AM) const; 186 187 unsigned isCFIntrinsic(const SDNode *Intr) const; 188 189 public: 190 /// \returns True if fixup needs to be emitted for given global value \p GV, 191 /// false otherwise. 192 bool shouldEmitFixup(const GlobalValue *GV) const; 193 194 /// \returns True if GOT relocation needs to be emitted for given global value 195 /// \p GV, false otherwise. 196 bool shouldEmitGOTReloc(const GlobalValue *GV) const; 197 198 /// \returns True if PC-relative relocation needs to be emitted for given 199 /// global value \p GV, false otherwise. 200 bool shouldEmitPCReloc(const GlobalValue *GV) const; 201 202 private: 203 // Analyze a combined offset from an amdgcn_buffer_ intrinsic and store the 204 // three offsets (voffset, soffset and instoffset) into the SDValue[3] array 205 // pointed to by Offsets. 206 /// \returns 0 If there is a non-constant offset or if the offset is 0. 207 /// Otherwise returns the constant offset. 208 unsigned setBufferOffsets(SDValue CombinedOffset, SelectionDAG &DAG, 209 SDValue *Offsets, unsigned Align = 4) const; 210 211 // Handle 8 bit and 16 bit buffer loads 212 SDValue handleByteShortBufferLoads(SelectionDAG &DAG, EVT LoadVT, SDLoc DL, 213 ArrayRef<SDValue> Ops, MemSDNode *M) const; 214 215 // Handle 8 bit and 16 bit buffer stores 216 SDValue handleByteShortBufferStores(SelectionDAG &DAG, EVT VDataType, 217 SDLoc DL, SDValue Ops[], 218 MemSDNode *M) const; 219 220 public: 221 SITargetLowering(const TargetMachine &tm, const GCNSubtarget &STI); 222 223 const GCNSubtarget *getSubtarget() const; 224 225 bool isFPExtFoldable(const SelectionDAG &DAG, unsigned Opcode, EVT DestVT, 226 EVT SrcVT) const override; 227 228 bool isShuffleMaskLegal(ArrayRef<int> /*Mask*/, EVT /*VT*/) const override; 229 230 bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &, 231 MachineFunction &MF, 232 unsigned IntrinsicID) const override; 233 234 bool getAddrModeArguments(IntrinsicInst * /*I*/, 235 SmallVectorImpl<Value*> &/*Ops*/, 236 Type *&/*AccessTy*/) const override; 237 238 bool isLegalGlobalAddressingMode(const AddrMode &AM) const; 239 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty, 240 unsigned AS, 241 Instruction *I = nullptr) const override; 242 243 bool canMergeStoresTo(unsigned AS, EVT MemVT, 244 const SelectionDAG &DAG) const override; 245 246 bool allowsMisalignedMemoryAccessesImpl( 247 unsigned Size, unsigned AS, unsigned Align, 248 MachineMemOperand::Flags Flags = MachineMemOperand::MONone, 249 bool *IsFast = nullptr) const; 250 251 bool allowsMisalignedMemoryAccesses( 252 EVT VT, unsigned AS, unsigned Align, 253 MachineMemOperand::Flags Flags = MachineMemOperand::MONone, 254 bool *IsFast = nullptr) const override; 255 256 EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign, 257 unsigned SrcAlign, bool IsMemset, 258 bool ZeroMemset, 259 bool MemcpyStrSrc, 260 const AttributeList &FuncAttributes) const override; 261 262 bool isMemOpUniform(const SDNode *N) const; 263 bool isMemOpHasNoClobberedMemOperand(const SDNode *N) const; 264 265 static bool isFlatGlobalAddrSpace(unsigned AS) { 266 return AS == AMDGPUAS::GLOBAL_ADDRESS || 267 AS == AMDGPUAS::FLAT_ADDRESS || 268 AS == AMDGPUAS::CONSTANT_ADDRESS || 269 AS > AMDGPUAS::MAX_AMDGPU_ADDRESS; 270 } 271 272 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override; 273 bool isFreeAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override; 274 275 TargetLoweringBase::LegalizeTypeAction 276 getPreferredVectorAction(MVT VT) const override; 277 278 bool shouldConvertConstantLoadToIntImm(const APInt &Imm, 279 Type *Ty) const override; 280 281 bool isTypeDesirableForOp(unsigned Op, EVT VT) const override; 282 283 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override; 284 285 bool supportSplitCSR(MachineFunction *MF) const override; 286 void initializeSplitCSR(MachineBasicBlock *Entry) const override; 287 void insertCopiesSplitCSR( 288 MachineBasicBlock *Entry, 289 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override; 290 291 SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, 292 bool isVarArg, 293 const SmallVectorImpl<ISD::InputArg> &Ins, 294 const SDLoc &DL, SelectionDAG &DAG, 295 SmallVectorImpl<SDValue> &InVals) const override; 296 297 bool CanLowerReturn(CallingConv::ID CallConv, 298 MachineFunction &MF, bool isVarArg, 299 const SmallVectorImpl<ISD::OutputArg> &Outs, 300 LLVMContext &Context) const override; 301 302 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, 303 const SmallVectorImpl<ISD::OutputArg> &Outs, 304 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL, 305 SelectionDAG &DAG) const override; 306 307 void passSpecialInputs( 308 CallLoweringInfo &CLI, 309 CCState &CCInfo, 310 const SIMachineFunctionInfo &Info, 311 SmallVectorImpl<std::pair<unsigned, SDValue>> &RegsToPass, 312 SmallVectorImpl<SDValue> &MemOpChains, 313 SDValue Chain) const; 314 315 SDValue LowerCallResult(SDValue Chain, SDValue InFlag, 316 CallingConv::ID CallConv, bool isVarArg, 317 const SmallVectorImpl<ISD::InputArg> &Ins, 318 const SDLoc &DL, SelectionDAG &DAG, 319 SmallVectorImpl<SDValue> &InVals, bool isThisReturn, 320 SDValue ThisVal) const; 321 322 bool mayBeEmittedAsTailCall(const CallInst *) const override; 323 324 bool isEligibleForTailCallOptimization( 325 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg, 326 const SmallVectorImpl<ISD::OutputArg> &Outs, 327 const SmallVectorImpl<SDValue> &OutVals, 328 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const; 329 330 SDValue LowerCall(CallLoweringInfo &CLI, 331 SmallVectorImpl<SDValue> &InVals) const override; 332 333 Register getRegisterByName(const char* RegName, LLT VT, 334 const MachineFunction &MF) const override; 335 336 MachineBasicBlock *splitKillBlock(MachineInstr &MI, 337 MachineBasicBlock *BB) const; 338 339 void bundleInstWithWaitcnt(MachineInstr &MI) const; 340 MachineBasicBlock *emitGWSMemViolTestLoop(MachineInstr &MI, 341 MachineBasicBlock *BB) const; 342 343 MachineBasicBlock * 344 EmitInstrWithCustomInserter(MachineInstr &MI, 345 MachineBasicBlock *BB) const override; 346 347 bool hasBitPreservingFPLogic(EVT VT) const override; 348 bool enableAggressiveFMAFusion(EVT VT) const override; 349 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context, 350 EVT VT) const override; 351 MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override; 352 bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF, 353 EVT VT) const override; 354 bool isFMADLegalForFAddFSub(const SelectionDAG &DAG, 355 const SDNode *N) const override; 356 357 SDValue splitUnaryVectorOp(SDValue Op, SelectionDAG &DAG) const; 358 SDValue splitBinaryVectorOp(SDValue Op, SelectionDAG &DAG) const; 359 SDValue splitTernaryVectorOp(SDValue Op, SelectionDAG &DAG) const; 360 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; 361 362 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results, 363 SelectionDAG &DAG) const override; 364 365 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; 366 SDNode *PostISelFolding(MachineSDNode *N, SelectionDAG &DAG) const override; 367 void AdjustInstrPostInstrSelection(MachineInstr &MI, 368 SDNode *Node) const override; 369 370 SDNode *legalizeTargetIndependentNode(SDNode *Node, SelectionDAG &DAG) const; 371 372 MachineSDNode *wrapAddr64Rsrc(SelectionDAG &DAG, const SDLoc &DL, 373 SDValue Ptr) const; 374 MachineSDNode *buildRSRC(SelectionDAG &DAG, const SDLoc &DL, SDValue Ptr, 375 uint32_t RsrcDword1, uint64_t RsrcDword2And3) const; 376 std::pair<unsigned, const TargetRegisterClass *> 377 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, 378 StringRef Constraint, MVT VT) const override; 379 ConstraintType getConstraintType(StringRef Constraint) const override; 380 SDValue copyToM0(SelectionDAG &DAG, SDValue Chain, const SDLoc &DL, 381 SDValue V) const; 382 383 void finalizeLowering(MachineFunction &MF) const override; 384 385 void computeKnownBitsForFrameIndex(const SDValue Op, 386 KnownBits &Known, 387 const APInt &DemandedElts, 388 const SelectionDAG &DAG, 389 unsigned Depth = 0) const override; 390 391 bool isSDNodeSourceOfDivergence(const SDNode *N, 392 FunctionLoweringInfo *FLI, LegacyDivergenceAnalysis *DA) const override; 393 394 bool isCanonicalized(SelectionDAG &DAG, SDValue Op, 395 unsigned MaxDepth = 5) const; 396 bool denormalsEnabledForType(const SelectionDAG &DAG, EVT VT) const; 397 398 bool isKnownNeverNaNForTargetNode(SDValue Op, 399 const SelectionDAG &DAG, 400 bool SNaN = false, 401 unsigned Depth = 0) const override; 402 AtomicExpansionKind shouldExpandAtomicRMWInIR(AtomicRMWInst *) const override; 403 404 virtual const TargetRegisterClass * 405 getRegClassFor(MVT VT, bool isDivergent) const override; 406 virtual bool requiresUniformRegister(MachineFunction &MF, 407 const Value *V) const override; 408 Align getPrefLoopAlignment(MachineLoop *ML) const override; 409 410 void allocateHSAUserSGPRs(CCState &CCInfo, 411 MachineFunction &MF, 412 const SIRegisterInfo &TRI, 413 SIMachineFunctionInfo &Info) const; 414 415 void allocateSystemSGPRs(CCState &CCInfo, 416 MachineFunction &MF, 417 SIMachineFunctionInfo &Info, 418 CallingConv::ID CallConv, 419 bool IsShader) const; 420 421 void allocateSpecialEntryInputVGPRs(CCState &CCInfo, 422 MachineFunction &MF, 423 const SIRegisterInfo &TRI, 424 SIMachineFunctionInfo &Info) const; 425 void allocateSpecialInputSGPRs( 426 CCState &CCInfo, 427 MachineFunction &MF, 428 const SIRegisterInfo &TRI, 429 SIMachineFunctionInfo &Info) const; 430 431 void allocateSpecialInputVGPRs(CCState &CCInfo, 432 MachineFunction &MF, 433 const SIRegisterInfo &TRI, 434 SIMachineFunctionInfo &Info) const; 435 }; 436 437 } // End namespace llvm 438 439 #endif 440