lib/Target/AMDGPU/SIISelLowering.h

   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   // Call DAG.getMemIntrinsicNode for a load, but first widen a dwordx3 type to
  98   // dwordx4 if on SI.
  99   SDValue getMemIntrinsicNode(unsigned Opcode, const SDLoc &DL, SDVTList VTList,
 100                               ArrayRef<SDValue> Ops, EVT MemVT,
 101                               MachineMemOperand *MMO, SelectionDAG &DAG) const;
 102
 103   SDValue handleD16VData(SDValue VData, SelectionDAG &DAG) const;
 104
 105   /// Converts \p Op, which must be of floating point type, to the
 106   /// floating point type \p VT, by either extending or truncating it.
 107   SDValue getFPExtOrFPTrunc(SelectionDAG &DAG,
 108                             SDValue Op,
 109                             const SDLoc &DL,
 110                             EVT VT) const;
 111
 112   SDValue convertArgType(
 113     SelectionDAG &DAG, EVT VT, EVT MemVT, const SDLoc &SL, SDValue Val,
 114     bool Signed, const ISD::InputArg *Arg = nullptr) const;
 115
 116   /// Custom lowering for ISD::FP_ROUND for MVT::f16.
 117   SDValue lowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
 118   SDValue lowerFMINNUM_FMAXNUM(SDValue Op, SelectionDAG &DAG) const;
 119
 120   SDValue getSegmentAperture(unsigned AS, const SDLoc &DL,
 121                              SelectionDAG &DAG) const;
 122
 123   SDValue lowerADDRSPACECAST(SDValue Op, SelectionDAG &DAG) const;
 124   SDValue lowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
 125   SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
 126   SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
 127   SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
 128   SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
 129   SDValue lowerTRAP(SDValue Op, SelectionDAG &DAG) const;
 130   SDValue lowerDEBUGTRAP(SDValue Op, SelectionDAG &DAG) const;
 131
 132   SDNode *adjustWritemask(MachineSDNode *&N, SelectionDAG &DAG) const;
 133
 134   SDValue performUCharToFloatCombine(SDNode *N,
 135                                      DAGCombinerInfo &DCI) const;
 136   SDValue performSHLPtrCombine(SDNode *N,
 137                                unsigned AS,
 138                                EVT MemVT,
 139                                DAGCombinerInfo &DCI) const;
 140
 141   SDValue performMemSDNodeCombine(MemSDNode *N, DAGCombinerInfo &DCI) const;
 142
 143   SDValue splitBinaryBitConstantOp(DAGCombinerInfo &DCI, const SDLoc &SL,
 144                                    unsigned Opc, SDValue LHS,
 145                                    const ConstantSDNode *CRHS) const;
 146
 147   SDValue performAndCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 148   SDValue performOrCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 149   SDValue performXorCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 150   SDValue performZeroExtendCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 151   SDValue performSignExtendInRegCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 152   SDValue performClassCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 153   SDValue getCanonicalConstantFP(SelectionDAG &DAG, const SDLoc &SL, EVT VT,
 154                                  const APFloat &C) const;
 155   SDValue performFCanonicalizeCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 156
 157   SDValue performFPMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL,
 158                                   SDValue Op0, SDValue Op1) const;
 159   SDValue performIntMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL,
 160                                    SDValue Op0, SDValue Op1, bool Signed) const;
 161   SDValue performMinMaxCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 162   SDValue performFMed3Combine(SDNode *N, DAGCombinerInfo &DCI) const;
 163   SDValue performCvtPkRTZCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 164   SDValue performExtractVectorEltCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 165   SDValue performInsertVectorEltCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 166
 167   SDValue reassociateScalarOps(SDNode *N, SelectionDAG &DAG) const;
 168   unsigned getFusedOpcode(const SelectionDAG &DAG,
 169                           const SDNode *N0, const SDNode *N1) const;
 170   SDValue performAddCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 171   SDValue performAddCarrySubCarryCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 172   SDValue performSubCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 173   SDValue performFAddCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 174   SDValue performFSubCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 175   SDValue performFMACombine(SDNode *N, DAGCombinerInfo &DCI) const;
 176   SDValue performSetCCCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 177   SDValue performCvtF32UByteNCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 178   SDValue performClampCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 179   SDValue performRcpCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 180
 181   bool isLegalFlatAddressingMode(const AddrMode &AM) const;
 182   bool isLegalMUBUFAddressingMode(const AddrMode &AM) const;
 183
 184   unsigned isCFIntrinsic(const SDNode *Intr) const;
 185
 186   /// \returns True if fixup needs to be emitted for given global value \p GV,
 187   /// false otherwise.
 188   bool shouldEmitFixup(const GlobalValue *GV) const;
 189
 190   /// \returns True if GOT relocation needs to be emitted for given global value
 191   /// \p GV, false otherwise.
 192   bool shouldEmitGOTReloc(const GlobalValue *GV) const;
 193
 194   /// \returns True if PC-relative relocation needs to be emitted for given
 195   /// global value \p GV, false otherwise.
 196   bool shouldEmitPCReloc(const GlobalValue *GV) const;
 197
 198   // Analyze a combined offset from an amdgcn_buffer_ intrinsic and store the
 199   // three offsets (voffset, soffset and instoffset) into the SDValue[3] array
 200   // pointed to by Offsets.
 201   void setBufferOffsets(SDValue CombinedOffset, SelectionDAG &DAG,
 202                         SDValue *Offsets, unsigned Align = 4) const;
 203
 204   // Handle 8 bit and 16 bit buffer loads
 205   SDValue handleByteShortBufferLoads(SelectionDAG &DAG, EVT LoadVT, SDLoc DL,
 206                                      ArrayRef<SDValue> Ops, MemSDNode *M) const;
 207
 208   // Handle 8 bit and 16 bit buffer stores
 209   SDValue handleByteShortBufferStores(SelectionDAG &DAG, EVT VDataType,
 210                                       SDLoc DL, SDValue Ops[],
 211                                       MemSDNode *M) const;
 212
 213 public:
 214   SITargetLowering(const TargetMachine &tm, const GCNSubtarget &STI);
 215
 216   const GCNSubtarget *getSubtarget() const;
 217
 218   bool isFPExtFoldable(unsigned Opcode, EVT DestVT, EVT SrcVT) const override;
 219
 220   bool isShuffleMaskLegal(ArrayRef<int> /*Mask*/, EVT /*VT*/) const override;
 221
 222   bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &,
 223                           MachineFunction &MF,
 224                           unsigned IntrinsicID) const override;
 225
 226   bool getAddrModeArguments(IntrinsicInst * /*I*/,
 227                             SmallVectorImpl<Value*> &/*Ops*/,
 228                             Type *&/*AccessTy*/) const override;
 229
 230   bool isLegalGlobalAddressingMode(const AddrMode &AM) const;
 231   bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
 232                              unsigned AS,
 233                              Instruction *I = nullptr) const override;
 234
 235   bool canMergeStoresTo(unsigned AS, EVT MemVT,
 236                         const SelectionDAG &DAG) const override;
 237
 238   bool allowsMisalignedMemoryAccesses(
 239       EVT VT, unsigned AS, unsigned Align,
 240       MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
 241       bool *IsFast = nullptr) const override;
 242
 243   EVT getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
 244                           unsigned SrcAlign, bool IsMemset,
 245                           bool ZeroMemset,
 246                           bool MemcpyStrSrc,
 247                           const AttributeList &FuncAttributes) const override;
 248
 249   bool isMemOpUniform(const SDNode *N) const;
 250   bool isMemOpHasNoClobberedMemOperand(const SDNode *N) const;
 251   bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
 252   bool isFreeAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
 253
 254   TargetLoweringBase::LegalizeTypeAction
 255   getPreferredVectorAction(MVT VT) const override;
 256
 257   bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
 258                                         Type *Ty) const override;
 259
 260   bool isTypeDesirableForOp(unsigned Op, EVT VT) const override;
 261
 262   bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
 263
 264   bool supportSplitCSR(MachineFunction *MF) const override;
 265   void initializeSplitCSR(MachineBasicBlock *Entry) const override;
 266   void insertCopiesSplitCSR(
 267     MachineBasicBlock *Entry,
 268     const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
 269
 270   SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
 271                                bool isVarArg,
 272                                const SmallVectorImpl<ISD::InputArg> &Ins,
 273                                const SDLoc &DL, SelectionDAG &DAG,
 274                                SmallVectorImpl<SDValue> &InVals) const override;
 275
 276   bool CanLowerReturn(CallingConv::ID CallConv,
 277                       MachineFunction &MF, bool isVarArg,
 278                       const SmallVectorImpl<ISD::OutputArg> &Outs,
 279                       LLVMContext &Context) const override;
 280
 281   SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
 282                       const SmallVectorImpl<ISD::OutputArg> &Outs,
 283                       const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
 284                       SelectionDAG &DAG) const override;
 285
 286   void passSpecialInputs(
 287     CallLoweringInfo &CLI,
 288     CCState &CCInfo,
 289     const SIMachineFunctionInfo &Info,
 290     SmallVectorImpl<std::pair<unsigned, SDValue>> &RegsToPass,
 291     SmallVectorImpl<SDValue> &MemOpChains,
 292     SDValue Chain) const;
 293
 294   SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
 295                           CallingConv::ID CallConv, bool isVarArg,
 296                           const SmallVectorImpl<ISD::InputArg> &Ins,
 297                           const SDLoc &DL, SelectionDAG &DAG,
 298                           SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
 299                           SDValue ThisVal) const;
 300
 301   bool mayBeEmittedAsTailCall(const CallInst *) const override;
 302
 303   bool isEligibleForTailCallOptimization(
 304     SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
 305     const SmallVectorImpl<ISD::OutputArg> &Outs,
 306     const SmallVectorImpl<SDValue> &OutVals,
 307     const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const;
 308
 309   SDValue LowerCall(CallLoweringInfo &CLI,
 310                     SmallVectorImpl<SDValue> &InVals) const override;
 311
 312   unsigned getRegisterByName(const char* RegName, EVT VT,
 313                              SelectionDAG &DAG) const override;
 314
 315   MachineBasicBlock *splitKillBlock(MachineInstr &MI,
 316                                     MachineBasicBlock *BB) const;
 317
 318   void bundleInstWithWaitcnt(MachineInstr &MI) const;
 319   MachineBasicBlock *emitGWSMemViolTestLoop(MachineInstr &MI,
 320                                             MachineBasicBlock *BB) const;
 321
 322   MachineBasicBlock *
 323   EmitInstrWithCustomInserter(MachineInstr &MI,
 324                               MachineBasicBlock *BB) const override;
 325
 326   bool hasBitPreservingFPLogic(EVT VT) const override;
 327   bool enableAggressiveFMAFusion(EVT VT) const override;
 328   EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
 329                          EVT VT) const override;
 330   MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override;
 331   bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
 332   SDValue splitUnaryVectorOp(SDValue Op, SelectionDAG &DAG) const;
 333   SDValue splitBinaryVectorOp(SDValue Op, SelectionDAG &DAG) const;
 334   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
 335
 336   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
 337                           SelectionDAG &DAG) const override;
 338
 339   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
 340   SDNode *PostISelFolding(MachineSDNode *N, SelectionDAG &DAG) const override;
 341   void AdjustInstrPostInstrSelection(MachineInstr &MI,
 342                                      SDNode *Node) const override;
 343
 344   SDNode *legalizeTargetIndependentNode(SDNode *Node, SelectionDAG &DAG) const;
 345
 346   MachineSDNode *wrapAddr64Rsrc(SelectionDAG &DAG, const SDLoc &DL,
 347                                 SDValue Ptr) const;
 348   MachineSDNode *buildRSRC(SelectionDAG &DAG, const SDLoc &DL, SDValue Ptr,
 349                            uint32_t RsrcDword1, uint64_t RsrcDword2And3) const;
 350   std::pair<unsigned, const TargetRegisterClass *>
 351   getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
 352                                StringRef Constraint, MVT VT) const override;
 353   ConstraintType getConstraintType(StringRef Constraint) const override;
 354   SDValue copyToM0(SelectionDAG &DAG, SDValue Chain, const SDLoc &DL,
 355                    SDValue V) const;
 356
 357   void finalizeLowering(MachineFunction &MF) const override;
 358
 359   void computeKnownBitsForFrameIndex(const SDValue Op,
 360                                      KnownBits &Known,
 361                                      const APInt &DemandedElts,
 362                                      const SelectionDAG &DAG,
 363                                      unsigned Depth = 0) const override;
 364
 365   bool isSDNodeSourceOfDivergence(const SDNode *N,
 366     FunctionLoweringInfo *FLI, LegacyDivergenceAnalysis *DA) const override;
 367
 368   bool isCanonicalized(SelectionDAG &DAG, SDValue Op,
 369                        unsigned MaxDepth = 5) const;
 370   bool denormalsEnabledForType(EVT VT) const;
 371
 372   bool isKnownNeverNaNForTargetNode(SDValue Op,
 373                                     const SelectionDAG &DAG,
 374                                     bool SNaN = false,
 375                                     unsigned Depth = 0) const override;
 376   AtomicExpansionKind shouldExpandAtomicRMWInIR(AtomicRMWInst *) const override;
 377
 378   unsigned getPrefLoopAlignment(MachineLoop *ML) const override;
 379
 380
 381   void allocateHSAUserSGPRs(CCState &CCInfo,
 382                             MachineFunction &MF,
 383                             const SIRegisterInfo &TRI,
 384                             SIMachineFunctionInfo &Info) const;
 385
 386   void allocateSystemSGPRs(CCState &CCInfo,
 387                            MachineFunction &MF,
 388                            SIMachineFunctionInfo &Info,
 389                            CallingConv::ID CallConv,
 390                            bool IsShader) const;
 391
 392   void allocateSpecialEntryInputVGPRs(CCState &CCInfo,
 393                                       MachineFunction &MF,
 394                                       const SIRegisterInfo &TRI,
 395                                       SIMachineFunctionInfo &Info) const;
 396   void allocateSpecialInputSGPRs(
 397     CCState &CCInfo,
 398     MachineFunction &MF,
 399     const SIRegisterInfo &TRI,
 400     SIMachineFunctionInfo &Info) const;
 401
 402   void allocateSpecialInputVGPRs(CCState &CCInfo,
 403                                  MachineFunction &MF,
 404                                  const SIRegisterInfo &TRI,
 405                                  SIMachineFunctionInfo &Info) const;
 406 };
 407
 408 } // End namespace llvm
 409
 410 #endif