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 AMDGPUISELLOWERING_H 17 #define AMDGPUISELLOWERING_H 18 19 #include "llvm/Target/TargetLowering.h" 20 21 namespace llvm { 22 23 class AMDGPUMachineFunction; 24 class AMDGPUSubtarget; 25 class MachineRegisterInfo; 26 27 class AMDGPUTargetLowering : public TargetLowering { 28 protected: 29 const AMDGPUSubtarget *Subtarget; 30 31 private: 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 LowerSDIV(SDValue Op, SelectionDAG &DAG) const; 47 SDValue LowerSDIV24(SDValue Op, SelectionDAG &DAG) const; 48 SDValue LowerSDIV32(SDValue Op, SelectionDAG &DAG) const; 49 SDValue LowerSDIV64(SDValue Op, SelectionDAG &DAG) const; 50 SDValue LowerSREM(SDValue Op, SelectionDAG &DAG) const; 51 SDValue LowerSREM32(SDValue Op, SelectionDAG &DAG) const; 52 SDValue LowerSREM64(SDValue Op, SelectionDAG &DAG) const; 53 SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const; 54 SDValue LowerFCEIL(SDValue Op, SelectionDAG &DAG) const; 55 SDValue LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const; 56 SDValue LowerFRINT(SDValue Op, SelectionDAG &DAG) const; 57 SDValue LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const; 58 SDValue LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const; 59 60 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 61 62 SDValue ExpandSIGN_EXTEND_INREG(SDValue Op, 63 unsigned BitsDiff, 64 SelectionDAG &DAG) const; 65 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const; 66 67 SDValue performMulCombine(SDNode *N, DAGCombinerInfo &DCI) const; 68 69 protected: 70 static EVT getEquivalentMemType(LLVMContext &Context, EVT VT); 71 static EVT getEquivalentLoadRegType(LLVMContext &Context, EVT VT); 72 73 /// \brief Helper function that adds Reg to the LiveIn list of the DAG's 74 /// MachineFunction. 75 /// 76 /// \returns a RegisterSDNode representing Reg. 77 virtual SDValue CreateLiveInRegister(SelectionDAG &DAG, 78 const TargetRegisterClass *RC, 79 unsigned Reg, EVT VT) const; 80 SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op, 81 SelectionDAG &DAG) const; 82 /// \brief Split a vector load into multiple scalar loads. 83 SDValue SplitVectorLoad(const SDValue &Op, SelectionDAG &DAG) const; 84 SDValue SplitVectorStore(SDValue Op, SelectionDAG &DAG) const; 85 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const; 86 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const; 87 SDValue LowerSDIVREM(SDValue Op, SelectionDAG &DAG) const; 88 bool isHWTrueValue(SDValue Op) const; 89 bool isHWFalseValue(SDValue Op) const; 90 91 /// The SelectionDAGBuilder will automatically promote function arguments 92 /// with illegal types. However, this does not work for the AMDGPU targets 93 /// since the function arguments are stored in memory as these illegal types. 94 /// In order to handle this properly we need to get the origianl types sizes 95 /// from the LLVM IR Function and fixup the ISD:InputArg values before 96 /// passing them to AnalyzeFormalArguments() 97 void getOriginalFunctionArgs(SelectionDAG &DAG, 98 const Function *F, 99 const SmallVectorImpl<ISD::InputArg> &Ins, 100 SmallVectorImpl<ISD::InputArg> &OrigIns) const; 101 void AnalyzeFormalArguments(CCState &State, 102 const SmallVectorImpl<ISD::InputArg> &Ins) const; 103 104 public: 105 AMDGPUTargetLowering(TargetMachine &TM); 106 107 bool isFAbsFree(EVT VT) const override; 108 bool isFNegFree(EVT VT) const override; 109 bool isTruncateFree(EVT Src, EVT Dest) const override; 110 bool isTruncateFree(Type *Src, Type *Dest) const override; 111 112 bool isZExtFree(Type *Src, Type *Dest) const override; 113 bool isZExtFree(EVT Src, EVT Dest) const override; 114 bool isZExtFree(SDValue Val, EVT VT2) const override; 115 116 bool isNarrowingProfitable(EVT VT1, EVT VT2) const override; 117 118 MVT getVectorIdxTy() const override; 119 bool isSelectSupported(SelectSupportKind) const override; 120 121 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override; 122 bool ShouldShrinkFPConstant(EVT VT) const override; 123 124 bool isLoadBitCastBeneficial(EVT, EVT) const override; 125 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, 126 bool isVarArg, 127 const SmallVectorImpl<ISD::OutputArg> &Outs, 128 const SmallVectorImpl<SDValue> &OutVals, 129 SDLoc DL, SelectionDAG &DAG) const override; 130 SDValue LowerCall(CallLoweringInfo &CLI, 131 SmallVectorImpl<SDValue> &InVals) const override; 132 133 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; 134 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; 135 void ReplaceNodeResults(SDNode * N, 136 SmallVectorImpl<SDValue> &Results, 137 SelectionDAG &DAG) const override; 138 139 SDValue LowerIntrinsicIABS(SDValue Op, SelectionDAG &DAG) const; 140 SDValue LowerIntrinsicLRP(SDValue Op, SelectionDAG &DAG) const; 141 SDValue CombineMinMax(SDNode *N, SelectionDAG &DAG) const; 142 const char* getTargetNodeName(unsigned Opcode) const override; 143 PostISelFolding(MachineSDNode * N,SelectionDAG & DAG)144 virtual SDNode *PostISelFolding(MachineSDNode *N, 145 SelectionDAG &DAG) const { 146 return N; 147 } 148 149 /// \brief Determine which of the bits specified in \p Mask are known to be 150 /// either zero or one and return them in the \p KnownZero and \p KnownOne 151 /// bitsets. 152 void computeKnownBitsForTargetNode(const SDValue Op, 153 APInt &KnownZero, 154 APInt &KnownOne, 155 const SelectionDAG &DAG, 156 unsigned Depth = 0) const override; 157 158 virtual unsigned ComputeNumSignBitsForTargetNode( 159 SDValue Op, 160 const SelectionDAG &DAG, 161 unsigned Depth = 0) const override; 162 }; 163 164 namespace AMDGPUISD { 165 166 enum { 167 // AMDIL ISD Opcodes 168 FIRST_NUMBER = ISD::BUILTIN_OP_END, 169 CALL, // Function call based on a single integer 170 UMUL, // 32bit unsigned multiplication 171 RET_FLAG, 172 BRANCH_COND, 173 // End AMDIL ISD Opcodes 174 DWORDADDR, 175 FRACT, 176 CLAMP, 177 178 // SIN_HW, COS_HW - f32 for SI, 1 ULP max error, valid from -100 pi to 100 pi. 179 // Denormals handled on some parts. 180 COS_HW, 181 SIN_HW, 182 FMAX, 183 SMAX, 184 UMAX, 185 FMIN, 186 SMIN, 187 UMIN, 188 URECIP, 189 DIV_SCALE, 190 DIV_FMAS, 191 DIV_FIXUP, 192 TRIG_PREOP, // 1 ULP max error for f64 193 194 // RCP, RSQ - For f32, 1 ULP max error, no denormal handling. 195 // For f64, max error 2^29 ULP, handles denormals. 196 RCP, 197 RSQ, 198 RSQ_LEGACY, 199 RSQ_CLAMPED, 200 DOT4, 201 BFE_U32, // Extract range of bits with zero extension to 32-bits. 202 BFE_I32, // Extract range of bits with sign extension to 32-bits. 203 BFI, // (src0 & src1) | (~src0 & src2) 204 BFM, // Insert a range of bits into a 32-bit word. 205 BREV, // Reverse bits. 206 MUL_U24, 207 MUL_I24, 208 MAD_U24, 209 MAD_I24, 210 TEXTURE_FETCH, 211 EXPORT, 212 CONST_ADDRESS, 213 REGISTER_LOAD, 214 REGISTER_STORE, 215 LOAD_INPUT, 216 SAMPLE, 217 SAMPLEB, 218 SAMPLED, 219 SAMPLEL, 220 221 // These cvt_f32_ubyte* nodes need to remain consecutive and in order. 222 CVT_F32_UBYTE0, 223 CVT_F32_UBYTE1, 224 CVT_F32_UBYTE2, 225 CVT_F32_UBYTE3, 226 /// This node is for VLIW targets and it is used to represent a vector 227 /// that is stored in consecutive registers with the same channel. 228 /// For example: 229 /// |X |Y|Z|W| 230 /// T0|v.x| | | | 231 /// T1|v.y| | | | 232 /// T2|v.z| | | | 233 /// T3|v.w| | | | 234 BUILD_VERTICAL_VECTOR, 235 FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE, 236 STORE_MSKOR, 237 LOAD_CONSTANT, 238 TBUFFER_STORE_FORMAT, 239 LAST_AMDGPU_ISD_NUMBER 240 }; 241 242 243 } // End namespace AMDGPUISD 244 245 } // End namespace llvm 246 247 #endif // AMDGPUISELLOWERING_H 248