AMDGPUISelLowering.h revision 286684
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_R600_AMDGPUISELLOWERING_H 17#define LLVM_LIB_TARGET_R600_AMDGPUISELLOWERING_H 18 19#include "llvm/Target/TargetLowering.h" 20 21namespace llvm { 22 23class AMDGPUMachineFunction; 24class AMDGPUSubtarget; 25class MachineRegisterInfo; 26 27class AMDGPUTargetLowering : public TargetLowering { 28protected: 29 const AMDGPUSubtarget *Subtarget; 30 31private: 32 SDValue LowerConstantInitializer(const Constant* Init, const GlobalValue *GV, 33 const SDValue &InitPtr, 34 SDValue Chain, 35 SelectionDAG &DAG) const; 36 SDValue LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const; 37 SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const; 38 SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const; 39 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const; 40 /// \brief Lower vector stores by merging the vector elements into an integer 41 /// of the same bitwidth. 42 SDValue MergeVectorStore(const SDValue &Op, SelectionDAG &DAG) const; 43 /// \brief Split a vector store into multiple scalar stores. 44 /// \returns The resulting chain. 45 46 SDValue LowerFREM(SDValue Op, SelectionDAG &DAG) const; 47 SDValue LowerFCEIL(SDValue Op, SelectionDAG &DAG) const; 48 SDValue LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const; 49 SDValue LowerFRINT(SDValue Op, SelectionDAG &DAG) const; 50 SDValue LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const; 51 52 SDValue LowerFROUND32(SDValue Op, SelectionDAG &DAG) const; 53 SDValue LowerFROUND64(SDValue Op, SelectionDAG &DAG) const; 54 SDValue LowerFROUND(SDValue Op, SelectionDAG &DAG) const; 55 SDValue LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const; 56 57 SDValue LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG, bool Signed) const; 58 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 59 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 60 61 SDValue LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG, bool Signed) const; 62 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const; 63 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const; 64 65 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const; 66 67 SDValue performStoreCombine(SDNode *N, DAGCombinerInfo &DCI) const; 68 SDValue performShlCombine(SDNode *N, DAGCombinerInfo &DCI) const; 69 SDValue performMulCombine(SDNode *N, DAGCombinerInfo &DCI) const; 70 71protected: 72 static EVT getEquivalentMemType(LLVMContext &Context, EVT VT); 73 static EVT getEquivalentLoadRegType(LLVMContext &Context, EVT VT); 74 75 virtual SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op, 76 SelectionDAG &DAG) const; 77 78 /// \brief Split a vector load into a scalar load of each component. 79 SDValue ScalarizeVectorLoad(SDValue Op, SelectionDAG &DAG) const; 80 81 /// \brief Split a vector load into 2 loads of half the vector. 82 SDValue SplitVectorLoad(SDValue Op, SelectionDAG &DAG) const; 83 84 /// \brief Split a vector store into a scalar store of each component. 85 SDValue ScalarizeVectorStore(SDValue Op, SelectionDAG &DAG) const; 86 87 /// \brief Split a vector store into 2 stores of half the vector. 88 SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const; 89 90 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const; 91 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const; 92 SDValue LowerSDIVREM(SDValue Op, SelectionDAG &DAG) const; 93 SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const; 94 SDValue LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const; 95 void LowerUDIVREM64(SDValue Op, SelectionDAG &DAG, 96 SmallVectorImpl<SDValue> &Results) const; 97 bool isHWTrueValue(SDValue Op) const; 98 bool isHWFalseValue(SDValue Op) const; 99 100 /// The SelectionDAGBuilder will automatically promote function arguments 101 /// with illegal types. However, this does not work for the AMDGPU targets 102 /// since the function arguments are stored in memory as these illegal types. 103 /// In order to handle this properly we need to get the origianl types sizes 104 /// from the LLVM IR Function and fixup the ISD:InputArg values before 105 /// passing them to AnalyzeFormalArguments() 106 void getOriginalFunctionArgs(SelectionDAG &DAG, 107 const Function *F, 108 const SmallVectorImpl<ISD::InputArg> &Ins, 109 SmallVectorImpl<ISD::InputArg> &OrigIns) const; 110 void AnalyzeFormalArguments(CCState &State, 111 const SmallVectorImpl<ISD::InputArg> &Ins) const; 112 113public: 114 AMDGPUTargetLowering(TargetMachine &TM, const AMDGPUSubtarget &STI); 115 116 bool isFAbsFree(EVT VT) const override; 117 bool isFNegFree(EVT VT) const override; 118 bool isTruncateFree(EVT Src, EVT Dest) const override; 119 bool isTruncateFree(Type *Src, Type *Dest) const override; 120 121 bool isZExtFree(Type *Src, Type *Dest) const override; 122 bool isZExtFree(EVT Src, EVT Dest) const override; 123 bool isZExtFree(SDValue Val, EVT VT2) const override; 124 125 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override; 126 127 MVT getVectorIdxTy(const DataLayout &) const override; 128 bool isSelectSupported(SelectSupportKind) const override; 129 130 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override; 131 bool ShouldShrinkFPConstant(EVT VT) const override; 132 bool shouldReduceLoadWidth(SDNode *Load, 133 ISD::LoadExtType ExtType, 134 EVT ExtVT) const override; 135 136 bool isLoadBitCastBeneficial(EVT, EVT) const override; 137 138 bool storeOfVectorConstantIsCheap(EVT MemVT, 139 unsigned NumElem, 140 unsigned AS) const override; 141 bool isCheapToSpeculateCttz() const override; 142 bool isCheapToSpeculateCtlz() const override; 143 144 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, 145 bool isVarArg, 146 const SmallVectorImpl<ISD::OutputArg> &Outs, 147 const SmallVectorImpl<SDValue> &OutVals, 148 SDLoc DL, SelectionDAG &DAG) const override; 149 SDValue LowerCall(CallLoweringInfo &CLI, 150 SmallVectorImpl<SDValue> &InVals) const override; 151 152 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; 153 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; 154 void ReplaceNodeResults(SDNode * N, 155 SmallVectorImpl<SDValue> &Results, 156 SelectionDAG &DAG) const override; 157 158 SDValue LowerIntrinsicIABS(SDValue Op, SelectionDAG &DAG) const; 159 SDValue LowerIntrinsicLRP(SDValue Op, SelectionDAG &DAG) const; 160 SDValue CombineFMinMaxLegacy(SDLoc DL, 161 EVT VT, 162 SDValue LHS, 163 SDValue RHS, 164 SDValue True, 165 SDValue False, 166 SDValue CC, 167 DAGCombinerInfo &DCI) const; 168 SDValue CombineIMinMax(SDLoc DL, 169 EVT VT, 170 SDValue LHS, 171 SDValue RHS, 172 SDValue True, 173 SDValue False, 174 SDValue CC, 175 SelectionDAG &DAG) const; 176 177 const char* getTargetNodeName(unsigned Opcode) const override; 178 179 SDValue getRsqrtEstimate(SDValue Operand, 180 DAGCombinerInfo &DCI, 181 unsigned &RefinementSteps, 182 bool &UseOneConstNR) const override; 183 SDValue getRecipEstimate(SDValue Operand, 184 DAGCombinerInfo &DCI, 185 unsigned &RefinementSteps) const override; 186 187 virtual SDNode *PostISelFolding(MachineSDNode *N, 188 SelectionDAG &DAG) const { 189 return N; 190 } 191 192 /// \brief Determine which of the bits specified in \p Mask are known to be 193 /// either zero or one and return them in the \p KnownZero and \p KnownOne 194 /// bitsets. 195 void computeKnownBitsForTargetNode(const SDValue Op, 196 APInt &KnownZero, 197 APInt &KnownOne, 198 const SelectionDAG &DAG, 199 unsigned Depth = 0) const override; 200 201 unsigned ComputeNumSignBitsForTargetNode(SDValue Op, const SelectionDAG &DAG, 202 unsigned Depth = 0) const override; 203 204 /// \brief Helper function that adds Reg to the LiveIn list of the DAG's 205 /// MachineFunction. 206 /// 207 /// \returns a RegisterSDNode representing Reg. 208 virtual SDValue CreateLiveInRegister(SelectionDAG &DAG, 209 const TargetRegisterClass *RC, 210 unsigned Reg, EVT VT) const; 211 212 enum ImplicitParameter { 213 GRID_DIM, 214 GRID_OFFSET 215 }; 216 217 /// \brief Helper function that returns the byte offset of the given 218 /// type of implicit parameter. 219 unsigned getImplicitParameterOffset(const AMDGPUMachineFunction *MFI, 220 const ImplicitParameter Param) const; 221}; 222 223namespace AMDGPUISD { 224 225enum NodeType : unsigned { 226 // AMDIL ISD Opcodes 227 FIRST_NUMBER = ISD::BUILTIN_OP_END, 228 CALL, // Function call based on a single integer 229 UMUL, // 32bit unsigned multiplication 230 RET_FLAG, 231 BRANCH_COND, 232 // End AMDIL ISD Opcodes 233 DWORDADDR, 234 FRACT, 235 CLAMP, 236 237 // SIN_HW, COS_HW - f32 for SI, 1 ULP max error, valid from -100 pi to 100 pi. 238 // Denormals handled on some parts. 239 COS_HW, 240 SIN_HW, 241 FMAX_LEGACY, 242 FMIN_LEGACY, 243 FMAX3, 244 SMAX3, 245 UMAX3, 246 FMIN3, 247 SMIN3, 248 UMIN3, 249 URECIP, 250 DIV_SCALE, 251 DIV_FMAS, 252 DIV_FIXUP, 253 TRIG_PREOP, // 1 ULP max error for f64 254 255 // RCP, RSQ - For f32, 1 ULP max error, no denormal handling. 256 // For f64, max error 2^29 ULP, handles denormals. 257 RCP, 258 RSQ, 259 RSQ_LEGACY, 260 RSQ_CLAMPED, 261 LDEXP, 262 FP_CLASS, 263 DOT4, 264 CARRY, 265 BORROW, 266 BFE_U32, // Extract range of bits with zero extension to 32-bits. 267 BFE_I32, // Extract range of bits with sign extension to 32-bits. 268 BFI, // (src0 & src1) | (~src0 & src2) 269 BFM, // Insert a range of bits into a 32-bit word. 270 BREV, // Reverse bits. 271 MUL_U24, 272 MUL_I24, 273 MAD_U24, 274 MAD_I24, 275 TEXTURE_FETCH, 276 EXPORT, 277 CONST_ADDRESS, 278 REGISTER_LOAD, 279 REGISTER_STORE, 280 LOAD_INPUT, 281 SAMPLE, 282 SAMPLEB, 283 SAMPLED, 284 SAMPLEL, 285 286 // These cvt_f32_ubyte* nodes need to remain consecutive and in order. 287 CVT_F32_UBYTE0, 288 CVT_F32_UBYTE1, 289 CVT_F32_UBYTE2, 290 CVT_F32_UBYTE3, 291 /// This node is for VLIW targets and it is used to represent a vector 292 /// that is stored in consecutive registers with the same channel. 293 /// For example: 294 /// |X |Y|Z|W| 295 /// T0|v.x| | | | 296 /// T1|v.y| | | | 297 /// T2|v.z| | | | 298 /// T3|v.w| | | | 299 BUILD_VERTICAL_VECTOR, 300 /// Pointer to the start of the shader's constant data. 301 CONST_DATA_PTR, 302 SENDMSG, 303 INTERP_MOV, 304 INTERP_P1, 305 INTERP_P2, 306 FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE, 307 STORE_MSKOR, 308 LOAD_CONSTANT, 309 TBUFFER_STORE_FORMAT, 310 LAST_AMDGPU_ISD_NUMBER 311}; 312 313 314} // End namespace AMDGPUISD 315 316} // End namespace llvm 317 318#endif 319