1 //===-- MipsISelLowering.cpp - Mips DAG Lowering Implementation -----------===// 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 // This file defines the interfaces that Mips uses to lower LLVM code into a 11 // selection DAG. 12 // 13 //===----------------------------------------------------------------------===// 14 #include "MipsISelLowering.h" 15 #include "InstPrinter/MipsInstPrinter.h" 16 #include "MCTargetDesc/MipsBaseInfo.h" 17 #include "MipsCCState.h" 18 #include "MipsMachineFunction.h" 19 #include "MipsSubtarget.h" 20 #include "MipsTargetMachine.h" 21 #include "MipsTargetObjectFile.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/ADT/StringSwitch.h" 24 #include "llvm/CodeGen/CallingConvLower.h" 25 #include "llvm/CodeGen/FunctionLoweringInfo.h" 26 #include "llvm/CodeGen/MachineFrameInfo.h" 27 #include "llvm/CodeGen/MachineFunction.h" 28 #include "llvm/CodeGen/MachineInstrBuilder.h" 29 #include "llvm/CodeGen/MachineJumpTableInfo.h" 30 #include "llvm/CodeGen/MachineRegisterInfo.h" 31 #include "llvm/CodeGen/SelectionDAGISel.h" 32 #include "llvm/CodeGen/ValueTypes.h" 33 #include "llvm/IR/CallingConv.h" 34 #include "llvm/IR/DerivedTypes.h" 35 #include "llvm/IR/GlobalVariable.h" 36 #include "llvm/Support/CommandLine.h" 37 #include "llvm/Support/Debug.h" 38 #include "llvm/Support/ErrorHandling.h" 39 #include "llvm/Support/raw_ostream.h" 40 #include <cctype> 41 42 using namespace llvm; 43 44 #define DEBUG_TYPE "mips-lower" 45 46 STATISTIC(NumTailCalls, "Number of tail calls"); 47 48 static cl::opt<bool> 49 LargeGOT("mxgot", cl::Hidden, 50 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false)); 51 52 static cl::opt<bool> 53 NoZeroDivCheck("mno-check-zero-division", cl::Hidden, 54 cl::desc("MIPS: Don't trap on integer division by zero."), 55 cl::init(false)); 56 57 static const MCPhysReg Mips64DPRegs[8] = { 58 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64, 59 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64 60 }; 61 62 // If I is a shifted mask, set the size (Size) and the first bit of the 63 // mask (Pos), and return true. 64 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11). 65 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) { 66 if (!isShiftedMask_64(I)) 67 return false; 68 69 Size = countPopulation(I); 70 Pos = countTrailingZeros(I); 71 return true; 72 } 73 74 // The MIPS MSA ABI passes vector arguments in the integer register set. 75 // The number of integer registers used is dependant on the ABI used. 76 MVT MipsTargetLowering::getRegisterTypeForCallingConv(MVT VT) const { 77 if (VT.isVector() && Subtarget.hasMSA()) 78 return Subtarget.isABI_O32() ? MVT::i32 : MVT::i64; 79 return MipsTargetLowering::getRegisterType(VT); 80 } 81 82 MVT MipsTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context, 83 EVT VT) const { 84 if (VT.isVector()) { 85 if (Subtarget.isABI_O32()) { 86 return MVT::i32; 87 } else { 88 return (VT.getSizeInBits() == 32) ? MVT::i32 : MVT::i64; 89 } 90 } 91 return MipsTargetLowering::getRegisterType(Context, VT); 92 } 93 94 unsigned MipsTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context, 95 EVT VT) const { 96 if (VT.isVector()) 97 return std::max((VT.getSizeInBits() / (Subtarget.isABI_O32() ? 32 : 64)), 98 1U); 99 return MipsTargetLowering::getNumRegisters(Context, VT); 100 } 101 102 unsigned MipsTargetLowering::getVectorTypeBreakdownForCallingConv( 103 LLVMContext &Context, EVT VT, EVT &IntermediateVT, 104 unsigned &NumIntermediates, MVT &RegisterVT) const { 105 106 // Break down vector types to either 2 i64s or 4 i32s. 107 RegisterVT = getRegisterTypeForCallingConv(Context, VT) ; 108 IntermediateVT = RegisterVT; 109 NumIntermediates = VT.getSizeInBits() < RegisterVT.getSizeInBits() 110 ? VT.getVectorNumElements() 111 : VT.getSizeInBits() / RegisterVT.getSizeInBits(); 112 113 return NumIntermediates; 114 } 115 116 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const { 117 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>(); 118 return DAG.getRegister(FI->getGlobalBaseReg(), Ty); 119 } 120 121 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty, 122 SelectionDAG &DAG, 123 unsigned Flag) const { 124 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag); 125 } 126 127 SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty, 128 SelectionDAG &DAG, 129 unsigned Flag) const { 130 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag); 131 } 132 133 SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty, 134 SelectionDAG &DAG, 135 unsigned Flag) const { 136 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag); 137 } 138 139 SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty, 140 SelectionDAG &DAG, 141 unsigned Flag) const { 142 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag); 143 } 144 145 SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty, 146 SelectionDAG &DAG, 147 unsigned Flag) const { 148 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(), 149 N->getOffset(), Flag); 150 } 151 152 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const { 153 switch ((MipsISD::NodeType)Opcode) { 154 case MipsISD::FIRST_NUMBER: break; 155 case MipsISD::JmpLink: return "MipsISD::JmpLink"; 156 case MipsISD::TailCall: return "MipsISD::TailCall"; 157 case MipsISD::Highest: return "MipsISD::Highest"; 158 case MipsISD::Higher: return "MipsISD::Higher"; 159 case MipsISD::Hi: return "MipsISD::Hi"; 160 case MipsISD::Lo: return "MipsISD::Lo"; 161 case MipsISD::GotHi: return "MipsISD::GotHi"; 162 case MipsISD::GPRel: return "MipsISD::GPRel"; 163 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer"; 164 case MipsISD::Ret: return "MipsISD::Ret"; 165 case MipsISD::ERet: return "MipsISD::ERet"; 166 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN"; 167 case MipsISD::FPBrcond: return "MipsISD::FPBrcond"; 168 case MipsISD::FPCmp: return "MipsISD::FPCmp"; 169 case MipsISD::FSELECT: return "MipsISD::FSELECT"; 170 case MipsISD::MTC1_D64: return "MipsISD::MTC1_D64"; 171 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T"; 172 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F"; 173 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP"; 174 case MipsISD::MFHI: return "MipsISD::MFHI"; 175 case MipsISD::MFLO: return "MipsISD::MFLO"; 176 case MipsISD::MTLOHI: return "MipsISD::MTLOHI"; 177 case MipsISD::Mult: return "MipsISD::Mult"; 178 case MipsISD::Multu: return "MipsISD::Multu"; 179 case MipsISD::MAdd: return "MipsISD::MAdd"; 180 case MipsISD::MAddu: return "MipsISD::MAddu"; 181 case MipsISD::MSub: return "MipsISD::MSub"; 182 case MipsISD::MSubu: return "MipsISD::MSubu"; 183 case MipsISD::DivRem: return "MipsISD::DivRem"; 184 case MipsISD::DivRemU: return "MipsISD::DivRemU"; 185 case MipsISD::DivRem16: return "MipsISD::DivRem16"; 186 case MipsISD::DivRemU16: return "MipsISD::DivRemU16"; 187 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64"; 188 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64"; 189 case MipsISD::Wrapper: return "MipsISD::Wrapper"; 190 case MipsISD::DynAlloc: return "MipsISD::DynAlloc"; 191 case MipsISD::Sync: return "MipsISD::Sync"; 192 case MipsISD::Ext: return "MipsISD::Ext"; 193 case MipsISD::Ins: return "MipsISD::Ins"; 194 case MipsISD::CIns: return "MipsISD::CIns"; 195 case MipsISD::LWL: return "MipsISD::LWL"; 196 case MipsISD::LWR: return "MipsISD::LWR"; 197 case MipsISD::SWL: return "MipsISD::SWL"; 198 case MipsISD::SWR: return "MipsISD::SWR"; 199 case MipsISD::LDL: return "MipsISD::LDL"; 200 case MipsISD::LDR: return "MipsISD::LDR"; 201 case MipsISD::SDL: return "MipsISD::SDL"; 202 case MipsISD::SDR: return "MipsISD::SDR"; 203 case MipsISD::EXTP: return "MipsISD::EXTP"; 204 case MipsISD::EXTPDP: return "MipsISD::EXTPDP"; 205 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H"; 206 case MipsISD::EXTR_W: return "MipsISD::EXTR_W"; 207 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W"; 208 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W"; 209 case MipsISD::SHILO: return "MipsISD::SHILO"; 210 case MipsISD::MTHLIP: return "MipsISD::MTHLIP"; 211 case MipsISD::MULSAQ_S_W_PH: return "MipsISD::MULSAQ_S_W_PH"; 212 case MipsISD::MAQ_S_W_PHL: return "MipsISD::MAQ_S_W_PHL"; 213 case MipsISD::MAQ_S_W_PHR: return "MipsISD::MAQ_S_W_PHR"; 214 case MipsISD::MAQ_SA_W_PHL: return "MipsISD::MAQ_SA_W_PHL"; 215 case MipsISD::MAQ_SA_W_PHR: return "MipsISD::MAQ_SA_W_PHR"; 216 case MipsISD::DPAU_H_QBL: return "MipsISD::DPAU_H_QBL"; 217 case MipsISD::DPAU_H_QBR: return "MipsISD::DPAU_H_QBR"; 218 case MipsISD::DPSU_H_QBL: return "MipsISD::DPSU_H_QBL"; 219 case MipsISD::DPSU_H_QBR: return "MipsISD::DPSU_H_QBR"; 220 case MipsISD::DPAQ_S_W_PH: return "MipsISD::DPAQ_S_W_PH"; 221 case MipsISD::DPSQ_S_W_PH: return "MipsISD::DPSQ_S_W_PH"; 222 case MipsISD::DPAQ_SA_L_W: return "MipsISD::DPAQ_SA_L_W"; 223 case MipsISD::DPSQ_SA_L_W: return "MipsISD::DPSQ_SA_L_W"; 224 case MipsISD::DPA_W_PH: return "MipsISD::DPA_W_PH"; 225 case MipsISD::DPS_W_PH: return "MipsISD::DPS_W_PH"; 226 case MipsISD::DPAQX_S_W_PH: return "MipsISD::DPAQX_S_W_PH"; 227 case MipsISD::DPAQX_SA_W_PH: return "MipsISD::DPAQX_SA_W_PH"; 228 case MipsISD::DPAX_W_PH: return "MipsISD::DPAX_W_PH"; 229 case MipsISD::DPSX_W_PH: return "MipsISD::DPSX_W_PH"; 230 case MipsISD::DPSQX_S_W_PH: return "MipsISD::DPSQX_S_W_PH"; 231 case MipsISD::DPSQX_SA_W_PH: return "MipsISD::DPSQX_SA_W_PH"; 232 case MipsISD::MULSA_W_PH: return "MipsISD::MULSA_W_PH"; 233 case MipsISD::MULT: return "MipsISD::MULT"; 234 case MipsISD::MULTU: return "MipsISD::MULTU"; 235 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP"; 236 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP"; 237 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP"; 238 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP"; 239 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP"; 240 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP"; 241 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP"; 242 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP"; 243 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP"; 244 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO"; 245 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO"; 246 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO"; 247 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO"; 248 case MipsISD::VCEQ: return "MipsISD::VCEQ"; 249 case MipsISD::VCLE_S: return "MipsISD::VCLE_S"; 250 case MipsISD::VCLE_U: return "MipsISD::VCLE_U"; 251 case MipsISD::VCLT_S: return "MipsISD::VCLT_S"; 252 case MipsISD::VCLT_U: return "MipsISD::VCLT_U"; 253 case MipsISD::VSMAX: return "MipsISD::VSMAX"; 254 case MipsISD::VSMIN: return "MipsISD::VSMIN"; 255 case MipsISD::VUMAX: return "MipsISD::VUMAX"; 256 case MipsISD::VUMIN: return "MipsISD::VUMIN"; 257 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT"; 258 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT"; 259 case MipsISD::VNOR: return "MipsISD::VNOR"; 260 case MipsISD::VSHF: return "MipsISD::VSHF"; 261 case MipsISD::SHF: return "MipsISD::SHF"; 262 case MipsISD::ILVEV: return "MipsISD::ILVEV"; 263 case MipsISD::ILVOD: return "MipsISD::ILVOD"; 264 case MipsISD::ILVL: return "MipsISD::ILVL"; 265 case MipsISD::ILVR: return "MipsISD::ILVR"; 266 case MipsISD::PCKEV: return "MipsISD::PCKEV"; 267 case MipsISD::PCKOD: return "MipsISD::PCKOD"; 268 case MipsISD::INSVE: return "MipsISD::INSVE"; 269 } 270 return nullptr; 271 } 272 273 MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM, 274 const MipsSubtarget &STI) 275 : TargetLowering(TM), Subtarget(STI), ABI(TM.getABI()) { 276 // Mips does not have i1 type, so use i32 for 277 // setcc operations results (slt, sgt, ...). 278 setBooleanContents(ZeroOrOneBooleanContent); 279 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 280 // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA 281 // does. Integer booleans still use 0 and 1. 282 if (Subtarget.hasMips32r6()) 283 setBooleanContents(ZeroOrOneBooleanContent, 284 ZeroOrNegativeOneBooleanContent); 285 286 // Load extented operations for i1 types must be promoted 287 for (MVT VT : MVT::integer_valuetypes()) { 288 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote); 289 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote); 290 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); 291 } 292 293 // MIPS doesn't have extending float->double load/store. Set LoadExtAction 294 // for f32, f16 295 for (MVT VT : MVT::fp_valuetypes()) { 296 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand); 297 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand); 298 } 299 300 // Set LoadExtAction for f16 vectors to Expand 301 for (MVT VT : MVT::fp_vector_valuetypes()) { 302 MVT F16VT = MVT::getVectorVT(MVT::f16, VT.getVectorNumElements()); 303 if (F16VT.isValid()) 304 setLoadExtAction(ISD::EXTLOAD, VT, F16VT, Expand); 305 } 306 307 setTruncStoreAction(MVT::f32, MVT::f16, Expand); 308 setTruncStoreAction(MVT::f64, MVT::f16, Expand); 309 310 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 311 312 // Used by legalize types to correctly generate the setcc result. 313 // Without this, every float setcc comes with a AND/OR with the result, 314 // we don't want this, since the fpcmp result goes to a flag register, 315 // which is used implicitly by brcond and select operations. 316 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32); 317 318 // Mips Custom Operations 319 setOperationAction(ISD::BR_JT, MVT::Other, Expand); 320 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 321 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 322 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 323 setOperationAction(ISD::JumpTable, MVT::i32, Custom); 324 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 325 setOperationAction(ISD::SELECT, MVT::f32, Custom); 326 setOperationAction(ISD::SELECT, MVT::f64, Custom); 327 setOperationAction(ISD::SELECT, MVT::i32, Custom); 328 setOperationAction(ISD::SETCC, MVT::f32, Custom); 329 setOperationAction(ISD::SETCC, MVT::f64, Custom); 330 setOperationAction(ISD::BRCOND, MVT::Other, Custom); 331 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 332 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 333 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 334 335 if (Subtarget.isGP64bit()) { 336 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); 337 setOperationAction(ISD::BlockAddress, MVT::i64, Custom); 338 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); 339 setOperationAction(ISD::JumpTable, MVT::i64, Custom); 340 setOperationAction(ISD::ConstantPool, MVT::i64, Custom); 341 setOperationAction(ISD::SELECT, MVT::i64, Custom); 342 setOperationAction(ISD::LOAD, MVT::i64, Custom); 343 setOperationAction(ISD::STORE, MVT::i64, Custom); 344 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); 345 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom); 346 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom); 347 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom); 348 } 349 350 if (!Subtarget.isGP64bit()) { 351 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 352 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 353 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 354 } 355 356 setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom); 357 if (Subtarget.isGP64bit()) 358 setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom); 359 360 setOperationAction(ISD::SDIV, MVT::i32, Expand); 361 setOperationAction(ISD::SREM, MVT::i32, Expand); 362 setOperationAction(ISD::UDIV, MVT::i32, Expand); 363 setOperationAction(ISD::UREM, MVT::i32, Expand); 364 setOperationAction(ISD::SDIV, MVT::i64, Expand); 365 setOperationAction(ISD::SREM, MVT::i64, Expand); 366 setOperationAction(ISD::UDIV, MVT::i64, Expand); 367 setOperationAction(ISD::UREM, MVT::i64, Expand); 368 369 if (!(Subtarget.hasDSP() && Subtarget.hasMips32r2())) { 370 setOperationAction(ISD::ADDC, MVT::i32, Expand); 371 setOperationAction(ISD::ADDE, MVT::i32, Expand); 372 } 373 374 setOperationAction(ISD::ADDC, MVT::i64, Expand); 375 setOperationAction(ISD::ADDE, MVT::i64, Expand); 376 setOperationAction(ISD::SUBC, MVT::i32, Expand); 377 setOperationAction(ISD::SUBE, MVT::i32, Expand); 378 setOperationAction(ISD::SUBC, MVT::i64, Expand); 379 setOperationAction(ISD::SUBE, MVT::i64, Expand); 380 381 // Operations not directly supported by Mips. 382 setOperationAction(ISD::BR_CC, MVT::f32, Expand); 383 setOperationAction(ISD::BR_CC, MVT::f64, Expand); 384 setOperationAction(ISD::BR_CC, MVT::i32, Expand); 385 setOperationAction(ISD::BR_CC, MVT::i64, Expand); 386 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand); 387 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand); 388 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand); 389 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand); 390 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); 391 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); 392 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); 393 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); 394 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 395 if (Subtarget.hasCnMips()) { 396 setOperationAction(ISD::CTPOP, MVT::i32, Legal); 397 setOperationAction(ISD::CTPOP, MVT::i64, Legal); 398 } else { 399 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 400 setOperationAction(ISD::CTPOP, MVT::i64, Expand); 401 } 402 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 403 setOperationAction(ISD::CTTZ, MVT::i64, Expand); 404 setOperationAction(ISD::ROTL, MVT::i32, Expand); 405 setOperationAction(ISD::ROTL, MVT::i64, Expand); 406 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 407 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 408 409 if (!Subtarget.hasMips32r2()) 410 setOperationAction(ISD::ROTR, MVT::i32, Expand); 411 412 if (!Subtarget.hasMips64r2()) 413 setOperationAction(ISD::ROTR, MVT::i64, Expand); 414 415 setOperationAction(ISD::FSIN, MVT::f32, Expand); 416 setOperationAction(ISD::FSIN, MVT::f64, Expand); 417 setOperationAction(ISD::FCOS, MVT::f32, Expand); 418 setOperationAction(ISD::FCOS, MVT::f64, Expand); 419 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 420 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 421 setOperationAction(ISD::FPOW, MVT::f32, Expand); 422 setOperationAction(ISD::FPOW, MVT::f64, Expand); 423 setOperationAction(ISD::FLOG, MVT::f32, Expand); 424 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 425 setOperationAction(ISD::FLOG10, MVT::f32, Expand); 426 setOperationAction(ISD::FEXP, MVT::f32, Expand); 427 setOperationAction(ISD::FMA, MVT::f32, Expand); 428 setOperationAction(ISD::FMA, MVT::f64, Expand); 429 setOperationAction(ISD::FREM, MVT::f32, Expand); 430 setOperationAction(ISD::FREM, MVT::f64, Expand); 431 432 // Lower f16 conversion operations into library calls 433 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand); 434 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand); 435 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand); 436 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand); 437 438 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); 439 440 setOperationAction(ISD::VASTART, MVT::Other, Custom); 441 setOperationAction(ISD::VAARG, MVT::Other, Custom); 442 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 443 setOperationAction(ISD::VAEND, MVT::Other, Expand); 444 445 // Use the default for now 446 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 447 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 448 449 if (!Subtarget.isGP64bit()) { 450 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand); 451 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand); 452 } 453 454 455 if (!Subtarget.hasMips32r2()) { 456 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 457 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 458 } 459 460 // MIPS16 lacks MIPS32's clz and clo instructions. 461 if (!Subtarget.hasMips32() || Subtarget.inMips16Mode()) 462 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 463 if (!Subtarget.hasMips64()) 464 setOperationAction(ISD::CTLZ, MVT::i64, Expand); 465 466 if (!Subtarget.hasMips32r2()) 467 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 468 if (!Subtarget.hasMips64r2()) 469 setOperationAction(ISD::BSWAP, MVT::i64, Expand); 470 471 if (Subtarget.isGP64bit()) { 472 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, MVT::i32, Custom); 473 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, MVT::i32, Custom); 474 setLoadExtAction(ISD::EXTLOAD, MVT::i64, MVT::i32, Custom); 475 setTruncStoreAction(MVT::i64, MVT::i32, Custom); 476 } 477 478 setOperationAction(ISD::TRAP, MVT::Other, Legal); 479 480 setTargetDAGCombine(ISD::SDIVREM); 481 setTargetDAGCombine(ISD::UDIVREM); 482 setTargetDAGCombine(ISD::SELECT); 483 setTargetDAGCombine(ISD::AND); 484 setTargetDAGCombine(ISD::OR); 485 setTargetDAGCombine(ISD::ADD); 486 setTargetDAGCombine(ISD::SUB); 487 setTargetDAGCombine(ISD::AssertZext); 488 setTargetDAGCombine(ISD::SHL); 489 490 if (ABI.IsO32()) { 491 // These libcalls are not available in 32-bit. 492 setLibcallName(RTLIB::SHL_I128, nullptr); 493 setLibcallName(RTLIB::SRL_I128, nullptr); 494 setLibcallName(RTLIB::SRA_I128, nullptr); 495 } 496 497 setMinFunctionAlignment(Subtarget.isGP64bit() ? 3 : 2); 498 499 // The arguments on the stack are defined in terms of 4-byte slots on O32 500 // and 8-byte slots on N32/N64. 501 setMinStackArgumentAlignment((ABI.IsN32() || ABI.IsN64()) ? 8 : 4); 502 503 setStackPointerRegisterToSaveRestore(ABI.IsN64() ? Mips::SP_64 : Mips::SP); 504 505 MaxStoresPerMemcpy = 16; 506 507 isMicroMips = Subtarget.inMicroMipsMode(); 508 } 509 510 const MipsTargetLowering *MipsTargetLowering::create(const MipsTargetMachine &TM, 511 const MipsSubtarget &STI) { 512 if (STI.inMips16Mode()) 513 return llvm::createMips16TargetLowering(TM, STI); 514 515 return llvm::createMipsSETargetLowering(TM, STI); 516 } 517 518 // Create a fast isel object. 519 FastISel * 520 MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, 521 const TargetLibraryInfo *libInfo) const { 522 const MipsTargetMachine &TM = 523 static_cast<const MipsTargetMachine &>(funcInfo.MF->getTarget()); 524 525 // We support only the standard encoding [MIPS32,MIPS32R5] ISAs. 526 bool UseFastISel = TM.Options.EnableFastISel && Subtarget.hasMips32() && 527 !Subtarget.hasMips32r6() && !Subtarget.inMips16Mode() && 528 !Subtarget.inMicroMipsMode(); 529 530 // Disable if either of the following is true: 531 // We do not generate PIC, the ABI is not O32, LargeGOT is being used. 532 if (!TM.isPositionIndependent() || !TM.getABI().IsO32() || LargeGOT) 533 UseFastISel = false; 534 535 return UseFastISel ? Mips::createFastISel(funcInfo, libInfo) : nullptr; 536 } 537 538 EVT MipsTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &, 539 EVT VT) const { 540 if (!VT.isVector()) 541 return MVT::i32; 542 return VT.changeVectorElementTypeToInteger(); 543 } 544 545 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG, 546 TargetLowering::DAGCombinerInfo &DCI, 547 const MipsSubtarget &Subtarget) { 548 if (DCI.isBeforeLegalizeOps()) 549 return SDValue(); 550 551 EVT Ty = N->getValueType(0); 552 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64; 553 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64; 554 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 : 555 MipsISD::DivRemU16; 556 SDLoc DL(N); 557 558 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue, 559 N->getOperand(0), N->getOperand(1)); 560 SDValue InChain = DAG.getEntryNode(); 561 SDValue InGlue = DivRem; 562 563 // insert MFLO 564 if (N->hasAnyUseOfValue(0)) { 565 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty, 566 InGlue); 567 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo); 568 InChain = CopyFromLo.getValue(1); 569 InGlue = CopyFromLo.getValue(2); 570 } 571 572 // insert MFHI 573 if (N->hasAnyUseOfValue(1)) { 574 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL, 575 HI, Ty, InGlue); 576 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi); 577 } 578 579 return SDValue(); 580 } 581 582 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) { 583 switch (CC) { 584 default: llvm_unreachable("Unknown fp condition code!"); 585 case ISD::SETEQ: 586 case ISD::SETOEQ: return Mips::FCOND_OEQ; 587 case ISD::SETUNE: return Mips::FCOND_UNE; 588 case ISD::SETLT: 589 case ISD::SETOLT: return Mips::FCOND_OLT; 590 case ISD::SETGT: 591 case ISD::SETOGT: return Mips::FCOND_OGT; 592 case ISD::SETLE: 593 case ISD::SETOLE: return Mips::FCOND_OLE; 594 case ISD::SETGE: 595 case ISD::SETOGE: return Mips::FCOND_OGE; 596 case ISD::SETULT: return Mips::FCOND_ULT; 597 case ISD::SETULE: return Mips::FCOND_ULE; 598 case ISD::SETUGT: return Mips::FCOND_UGT; 599 case ISD::SETUGE: return Mips::FCOND_UGE; 600 case ISD::SETUO: return Mips::FCOND_UN; 601 case ISD::SETO: return Mips::FCOND_OR; 602 case ISD::SETNE: 603 case ISD::SETONE: return Mips::FCOND_ONE; 604 case ISD::SETUEQ: return Mips::FCOND_UEQ; 605 } 606 } 607 608 609 /// This function returns true if the floating point conditional branches and 610 /// conditional moves which use condition code CC should be inverted. 611 static bool invertFPCondCodeUser(Mips::CondCode CC) { 612 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT) 613 return false; 614 615 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) && 616 "Illegal Condition Code"); 617 618 return true; 619 } 620 621 // Creates and returns an FPCmp node from a setcc node. 622 // Returns Op if setcc is not a floating point comparison. 623 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) { 624 // must be a SETCC node 625 if (Op.getOpcode() != ISD::SETCC) 626 return Op; 627 628 SDValue LHS = Op.getOperand(0); 629 630 if (!LHS.getValueType().isFloatingPoint()) 631 return Op; 632 633 SDValue RHS = Op.getOperand(1); 634 SDLoc DL(Op); 635 636 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of 637 // node if necessary. 638 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); 639 640 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS, 641 DAG.getConstant(condCodeToFCC(CC), DL, MVT::i32)); 642 } 643 644 // Creates and returns a CMovFPT/F node. 645 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True, 646 SDValue False, const SDLoc &DL) { 647 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2)); 648 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue()); 649 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 650 651 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL, 652 True.getValueType(), True, FCC0, False, Cond); 653 } 654 655 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG, 656 TargetLowering::DAGCombinerInfo &DCI, 657 const MipsSubtarget &Subtarget) { 658 if (DCI.isBeforeLegalizeOps()) 659 return SDValue(); 660 661 SDValue SetCC = N->getOperand(0); 662 663 if ((SetCC.getOpcode() != ISD::SETCC) || 664 !SetCC.getOperand(0).getValueType().isInteger()) 665 return SDValue(); 666 667 SDValue False = N->getOperand(2); 668 EVT FalseTy = False.getValueType(); 669 670 if (!FalseTy.isInteger()) 671 return SDValue(); 672 673 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(False); 674 675 // If the RHS (False) is 0, we swap the order of the operands 676 // of ISD::SELECT (obviously also inverting the condition) so that we can 677 // take advantage of conditional moves using the $0 register. 678 // Example: 679 // return (a != 0) ? x : 0; 680 // load $reg, x 681 // movz $reg, $0, a 682 if (!FalseC) 683 return SDValue(); 684 685 const SDLoc DL(N); 686 687 if (!FalseC->getZExtValue()) { 688 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 689 SDValue True = N->getOperand(1); 690 691 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 692 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 693 694 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True); 695 } 696 697 // If both operands are integer constants there's a possibility that we 698 // can do some interesting optimizations. 699 SDValue True = N->getOperand(1); 700 ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(True); 701 702 if (!TrueC || !True.getValueType().isInteger()) 703 return SDValue(); 704 705 // We'll also ignore MVT::i64 operands as this optimizations proves 706 // to be ineffective because of the required sign extensions as the result 707 // of a SETCC operator is always MVT::i32 for non-vector types. 708 if (True.getValueType() == MVT::i64) 709 return SDValue(); 710 711 int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue(); 712 713 // 1) (a < x) ? y : y-1 714 // slti $reg1, a, x 715 // addiu $reg2, $reg1, y-1 716 if (Diff == 1) 717 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, False); 718 719 // 2) (a < x) ? y-1 : y 720 // slti $reg1, a, x 721 // xor $reg1, $reg1, 1 722 // addiu $reg2, $reg1, y-1 723 if (Diff == -1) { 724 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 725 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 726 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 727 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, True); 728 } 729 730 // Couldn't optimize. 731 return SDValue(); 732 } 733 734 static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG, 735 TargetLowering::DAGCombinerInfo &DCI, 736 const MipsSubtarget &Subtarget) { 737 if (DCI.isBeforeLegalizeOps()) 738 return SDValue(); 739 740 SDValue ValueIfTrue = N->getOperand(0), ValueIfFalse = N->getOperand(2); 741 742 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(ValueIfFalse); 743 if (!FalseC || FalseC->getZExtValue()) 744 return SDValue(); 745 746 // Since RHS (False) is 0, we swap the order of the True/False operands 747 // (obviously also inverting the condition) so that we can 748 // take advantage of conditional moves using the $0 register. 749 // Example: 750 // return (a != 0) ? x : 0; 751 // load $reg, x 752 // movz $reg, $0, a 753 unsigned Opc = (N->getOpcode() == MipsISD::CMovFP_T) ? MipsISD::CMovFP_F : 754 MipsISD::CMovFP_T; 755 756 SDValue FCC = N->getOperand(1), Glue = N->getOperand(3); 757 return DAG.getNode(Opc, SDLoc(N), ValueIfFalse.getValueType(), 758 ValueIfFalse, FCC, ValueIfTrue, Glue); 759 } 760 761 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, 762 TargetLowering::DAGCombinerInfo &DCI, 763 const MipsSubtarget &Subtarget) { 764 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) 765 return SDValue(); 766 767 SDValue FirstOperand = N->getOperand(0); 768 unsigned FirstOperandOpc = FirstOperand.getOpcode(); 769 SDValue Mask = N->getOperand(1); 770 EVT ValTy = N->getValueType(0); 771 SDLoc DL(N); 772 773 uint64_t Pos = 0, SMPos, SMSize; 774 ConstantSDNode *CN; 775 SDValue NewOperand; 776 unsigned Opc; 777 778 // Op's second operand must be a shifted mask. 779 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) || 780 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize)) 781 return SDValue(); 782 783 if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL) { 784 // Pattern match EXT. 785 // $dst = and ((sra or srl) $src , pos), (2**size - 1) 786 // => ext $dst, $src, pos, size 787 788 // The second operand of the shift must be an immediate. 789 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1)))) 790 return SDValue(); 791 792 Pos = CN->getZExtValue(); 793 794 // Return if the shifted mask does not start at bit 0 or the sum of its size 795 // and Pos exceeds the word's size. 796 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits()) 797 return SDValue(); 798 799 Opc = MipsISD::Ext; 800 NewOperand = FirstOperand.getOperand(0); 801 } else if (FirstOperandOpc == ISD::SHL && Subtarget.hasCnMips()) { 802 // Pattern match CINS. 803 // $dst = and (shl $src , pos), mask 804 // => cins $dst, $src, pos, size 805 // mask is a shifted mask with consecutive 1's, pos = shift amount, 806 // size = population count. 807 808 // The second operand of the shift must be an immediate. 809 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1)))) 810 return SDValue(); 811 812 Pos = CN->getZExtValue(); 813 814 if (SMPos != Pos || Pos >= ValTy.getSizeInBits() || SMSize >= 32 || 815 Pos + SMSize > ValTy.getSizeInBits()) 816 return SDValue(); 817 818 NewOperand = FirstOperand.getOperand(0); 819 // SMSize is 'location' (position) in this case, not size. 820 SMSize--; 821 Opc = MipsISD::CIns; 822 } else { 823 // Pattern match EXT. 824 // $dst = and $src, (2**size - 1) , if size > 16 825 // => ext $dst, $src, pos, size , pos = 0 826 827 // If the mask is <= 0xffff, andi can be used instead. 828 if (CN->getZExtValue() <= 0xffff) 829 return SDValue(); 830 831 // Return if the mask doesn't start at position 0. 832 if (SMPos) 833 return SDValue(); 834 835 Opc = MipsISD::Ext; 836 NewOperand = FirstOperand; 837 } 838 return DAG.getNode(Opc, DL, ValTy, NewOperand, 839 DAG.getConstant(Pos, DL, MVT::i32), 840 DAG.getConstant(SMSize, DL, MVT::i32)); 841 } 842 843 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, 844 TargetLowering::DAGCombinerInfo &DCI, 845 const MipsSubtarget &Subtarget) { 846 // Pattern match INS. 847 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1), 848 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1 849 // => ins $dst, $src, size, pos, $src1 850 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) 851 return SDValue(); 852 853 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1); 854 uint64_t SMPos0, SMSize0, SMPos1, SMSize1; 855 ConstantSDNode *CN, *CN1; 856 857 // See if Op's first operand matches (and $src1 , mask0). 858 if (And0.getOpcode() != ISD::AND) 859 return SDValue(); 860 861 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) || 862 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0)) 863 return SDValue(); 864 865 // See if Op's second operand matches (and (shl $src, pos), mask1). 866 if (And1.getOpcode() == ISD::AND && 867 And1.getOperand(0).getOpcode() == ISD::SHL) { 868 869 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) || 870 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1)) 871 return SDValue(); 872 873 // The shift masks must have the same position and size. 874 if (SMPos0 != SMPos1 || SMSize0 != SMSize1) 875 return SDValue(); 876 877 SDValue Shl = And1.getOperand(0); 878 879 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1)))) 880 return SDValue(); 881 882 unsigned Shamt = CN->getZExtValue(); 883 884 // Return if the shift amount and the first bit position of mask are not the 885 // same. 886 EVT ValTy = N->getValueType(0); 887 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits())) 888 return SDValue(); 889 890 SDLoc DL(N); 891 return DAG.getNode(MipsISD::Ins, DL, ValTy, Shl.getOperand(0), 892 DAG.getConstant(SMPos0, DL, MVT::i32), 893 DAG.getConstant(SMSize0, DL, MVT::i32), 894 And0.getOperand(0)); 895 } else { 896 // Pattern match DINS. 897 // $dst = or (and $src, mask0), mask1 898 // where mask0 = ((1 << SMSize0) -1) << SMPos0 899 // => dins $dst, $src, pos, size 900 if (~CN->getSExtValue() == ((((int64_t)1 << SMSize0) - 1) << SMPos0) && 901 ((SMSize0 + SMPos0 <= 64 && Subtarget.hasMips64r2()) || 902 (SMSize0 + SMPos0 <= 32))) { 903 // Check if AND instruction has constant as argument 904 bool isConstCase = And1.getOpcode() != ISD::AND; 905 if (And1.getOpcode() == ISD::AND) { 906 if (!(CN1 = dyn_cast<ConstantSDNode>(And1->getOperand(1)))) 907 return SDValue(); 908 } else { 909 if (!(CN1 = dyn_cast<ConstantSDNode>(N->getOperand(1)))) 910 return SDValue(); 911 } 912 // Don't generate INS if constant OR operand doesn't fit into bits 913 // cleared by constant AND operand. 914 if (CN->getSExtValue() & CN1->getSExtValue()) 915 return SDValue(); 916 917 SDLoc DL(N); 918 EVT ValTy = N->getOperand(0)->getValueType(0); 919 SDValue Const1; 920 SDValue SrlX; 921 if (!isConstCase) { 922 Const1 = DAG.getConstant(SMPos0, DL, MVT::i32); 923 SrlX = DAG.getNode(ISD::SRL, DL, And1->getValueType(0), And1, Const1); 924 } 925 return DAG.getNode( 926 MipsISD::Ins, DL, N->getValueType(0), 927 isConstCase 928 ? DAG.getConstant(CN1->getSExtValue() >> SMPos0, DL, ValTy) 929 : SrlX, 930 DAG.getConstant(SMPos0, DL, MVT::i32), 931 DAG.getConstant(ValTy.getSizeInBits() / 8 < 8 ? SMSize0 & 31 932 : SMSize0, 933 DL, MVT::i32), 934 And0->getOperand(0)); 935 936 } 937 return SDValue(); 938 } 939 } 940 941 static SDValue performMADD_MSUBCombine(SDNode *ROOTNode, SelectionDAG &CurDAG, 942 const MipsSubtarget &Subtarget) { 943 // ROOTNode must have a multiplication as an operand for the match to be 944 // successful. 945 if (ROOTNode->getOperand(0).getOpcode() != ISD::MUL && 946 ROOTNode->getOperand(1).getOpcode() != ISD::MUL) 947 return SDValue(); 948 949 // We don't handle vector types here. 950 if (ROOTNode->getValueType(0).isVector()) 951 return SDValue(); 952 953 // For MIPS64, madd / msub instructions are inefficent to use with 64 bit 954 // arithmetic. E.g. 955 // (add (mul a b) c) => 956 // let res = (madd (mthi (drotr c 32))x(mtlo c) a b) in 957 // MIPS64: (or (dsll (mfhi res) 32) (dsrl (dsll (mflo res) 32) 32) 958 // or 959 // MIPS64R2: (dins (mflo res) (mfhi res) 32 32) 960 // 961 // The overhead of setting up the Hi/Lo registers and reassembling the 962 // result makes this a dubious optimzation for MIPS64. The core of the 963 // problem is that Hi/Lo contain the upper and lower 32 bits of the 964 // operand and result. 965 // 966 // It requires a chain of 4 add/mul for MIPS64R2 to get better code 967 // density than doing it naively, 5 for MIPS64. Additionally, using 968 // madd/msub on MIPS64 requires the operands actually be 32 bit sign 969 // extended operands, not true 64 bit values. 970 // 971 // FIXME: For the moment, disable this completely for MIPS64. 972 if (Subtarget.hasMips64()) 973 return SDValue(); 974 975 SDValue Mult = ROOTNode->getOperand(0).getOpcode() == ISD::MUL 976 ? ROOTNode->getOperand(0) 977 : ROOTNode->getOperand(1); 978 979 SDValue AddOperand = ROOTNode->getOperand(0).getOpcode() == ISD::MUL 980 ? ROOTNode->getOperand(1) 981 : ROOTNode->getOperand(0); 982 983 // Transform this to a MADD only if the user of this node is the add. 984 // If there are other users of the mul, this function returns here. 985 if (!Mult.hasOneUse()) 986 return SDValue(); 987 988 // maddu and madd are unusual instructions in that on MIPS64 bits 63..31 989 // must be in canonical form, i.e. sign extended. For MIPS32, the operands 990 // of the multiply must have 32 or more sign bits, otherwise we cannot 991 // perform this optimization. We have to check this here as we're performing 992 // this optimization pre-legalization. 993 SDValue MultLHS = Mult->getOperand(0); 994 SDValue MultRHS = Mult->getOperand(1); 995 996 bool IsSigned = MultLHS->getOpcode() == ISD::SIGN_EXTEND && 997 MultRHS->getOpcode() == ISD::SIGN_EXTEND; 998 bool IsUnsigned = MultLHS->getOpcode() == ISD::ZERO_EXTEND && 999 MultRHS->getOpcode() == ISD::ZERO_EXTEND; 1000 1001 if (!IsSigned && !IsUnsigned) 1002 return SDValue(); 1003 1004 // Initialize accumulator. 1005 SDLoc DL(ROOTNode); 1006 SDValue TopHalf; 1007 SDValue BottomHalf; 1008 BottomHalf = CurDAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, AddOperand, 1009 CurDAG.getIntPtrConstant(0, DL)); 1010 1011 TopHalf = CurDAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, AddOperand, 1012 CurDAG.getIntPtrConstant(1, DL)); 1013 SDValue ACCIn = CurDAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped, 1014 BottomHalf, 1015 TopHalf); 1016 1017 // Create MipsMAdd(u) / MipsMSub(u) node. 1018 bool IsAdd = ROOTNode->getOpcode() == ISD::ADD; 1019 unsigned Opcode = IsAdd ? (IsUnsigned ? MipsISD::MAddu : MipsISD::MAdd) 1020 : (IsUnsigned ? MipsISD::MSubu : MipsISD::MSub); 1021 SDValue MAddOps[3] = { 1022 CurDAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mult->getOperand(0)), 1023 CurDAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mult->getOperand(1)), ACCIn}; 1024 EVT VTs[2] = {MVT::i32, MVT::i32}; 1025 SDValue MAdd = CurDAG.getNode(Opcode, DL, VTs, MAddOps); 1026 1027 SDValue ResLo = CurDAG.getNode(MipsISD::MFLO, DL, MVT::i32, MAdd); 1028 SDValue ResHi = CurDAG.getNode(MipsISD::MFHI, DL, MVT::i32, MAdd); 1029 SDValue Combined = 1030 CurDAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResLo, ResHi); 1031 return Combined; 1032 } 1033 1034 static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG, 1035 TargetLowering::DAGCombinerInfo &DCI, 1036 const MipsSubtarget &Subtarget) { 1037 // (sub v0 (mul v1, v2)) => (msub v1, v2, v0) 1038 if (DCI.isBeforeLegalizeOps()) { 1039 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() && 1040 !Subtarget.inMips16Mode() && N->getValueType(0) == MVT::i64) 1041 return performMADD_MSUBCombine(N, DAG, Subtarget); 1042 1043 return SDValue(); 1044 } 1045 1046 return SDValue(); 1047 } 1048 1049 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG, 1050 TargetLowering::DAGCombinerInfo &DCI, 1051 const MipsSubtarget &Subtarget) { 1052 // (add v0 (mul v1, v2)) => (madd v1, v2, v0) 1053 if (DCI.isBeforeLegalizeOps()) { 1054 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() && 1055 !Subtarget.inMips16Mode() && N->getValueType(0) == MVT::i64) 1056 return performMADD_MSUBCombine(N, DAG, Subtarget); 1057 1058 return SDValue(); 1059 } 1060 1061 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt)) 1062 SDValue Add = N->getOperand(1); 1063 1064 if (Add.getOpcode() != ISD::ADD) 1065 return SDValue(); 1066 1067 SDValue Lo = Add.getOperand(1); 1068 1069 if ((Lo.getOpcode() != MipsISD::Lo) || 1070 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable)) 1071 return SDValue(); 1072 1073 EVT ValTy = N->getValueType(0); 1074 SDLoc DL(N); 1075 1076 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0), 1077 Add.getOperand(0)); 1078 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo); 1079 } 1080 1081 static SDValue performAssertZextCombine(SDNode *N, SelectionDAG &DAG, 1082 TargetLowering::DAGCombinerInfo &DCI, 1083 const MipsSubtarget &Subtarget) { 1084 SDValue N0 = N->getOperand(0); 1085 EVT NarrowerVT = cast<VTSDNode>(N->getOperand(1))->getVT(); 1086 1087 if (N0.getOpcode() != ISD::TRUNCATE) 1088 return SDValue(); 1089 1090 if (N0.getOperand(0).getOpcode() != ISD::AssertZext) 1091 return SDValue(); 1092 1093 // fold (AssertZext (trunc (AssertZext x))) -> (trunc (AssertZext x)) 1094 // if the type of the extension of the innermost AssertZext node is 1095 // smaller from that of the outermost node, eg: 1096 // (AssertZext:i32 (trunc:i32 (AssertZext:i64 X, i32)), i8) 1097 // -> (trunc:i32 (AssertZext X, i8)) 1098 SDValue WiderAssertZext = N0.getOperand(0); 1099 EVT WiderVT = cast<VTSDNode>(WiderAssertZext->getOperand(1))->getVT(); 1100 1101 if (NarrowerVT.bitsLT(WiderVT)) { 1102 SDValue NewAssertZext = DAG.getNode( 1103 ISD::AssertZext, SDLoc(N), WiderAssertZext.getValueType(), 1104 WiderAssertZext.getOperand(0), DAG.getValueType(NarrowerVT)); 1105 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), 1106 NewAssertZext); 1107 } 1108 1109 return SDValue(); 1110 } 1111 1112 1113 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG, 1114 TargetLowering::DAGCombinerInfo &DCI, 1115 const MipsSubtarget &Subtarget) { 1116 // Pattern match CINS. 1117 // $dst = shl (and $src , imm), pos 1118 // => cins $dst, $src, pos, size 1119 1120 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasCnMips()) 1121 return SDValue(); 1122 1123 SDValue FirstOperand = N->getOperand(0); 1124 unsigned FirstOperandOpc = FirstOperand.getOpcode(); 1125 SDValue SecondOperand = N->getOperand(1); 1126 EVT ValTy = N->getValueType(0); 1127 SDLoc DL(N); 1128 1129 uint64_t Pos = 0, SMPos, SMSize; 1130 ConstantSDNode *CN; 1131 SDValue NewOperand; 1132 1133 // The second operand of the shift must be an immediate. 1134 if (!(CN = dyn_cast<ConstantSDNode>(SecondOperand))) 1135 return SDValue(); 1136 1137 Pos = CN->getZExtValue(); 1138 1139 if (Pos >= ValTy.getSizeInBits()) 1140 return SDValue(); 1141 1142 if (FirstOperandOpc != ISD::AND) 1143 return SDValue(); 1144 1145 // AND's second operand must be a shifted mask. 1146 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))) || 1147 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize)) 1148 return SDValue(); 1149 1150 // Return if the shifted mask does not start at bit 0 or the sum of its size 1151 // and Pos exceeds the word's size. 1152 if (SMPos != 0 || SMSize > 32 || Pos + SMSize > ValTy.getSizeInBits()) 1153 return SDValue(); 1154 1155 NewOperand = FirstOperand.getOperand(0); 1156 // SMSize is 'location' (position) in this case, not size. 1157 SMSize--; 1158 1159 return DAG.getNode(MipsISD::CIns, DL, ValTy, NewOperand, 1160 DAG.getConstant(Pos, DL, MVT::i32), 1161 DAG.getConstant(SMSize, DL, MVT::i32)); 1162 } 1163 1164 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) 1165 const { 1166 SelectionDAG &DAG = DCI.DAG; 1167 unsigned Opc = N->getOpcode(); 1168 1169 switch (Opc) { 1170 default: break; 1171 case ISD::SDIVREM: 1172 case ISD::UDIVREM: 1173 return performDivRemCombine(N, DAG, DCI, Subtarget); 1174 case ISD::SELECT: 1175 return performSELECTCombine(N, DAG, DCI, Subtarget); 1176 case MipsISD::CMovFP_F: 1177 case MipsISD::CMovFP_T: 1178 return performCMovFPCombine(N, DAG, DCI, Subtarget); 1179 case ISD::AND: 1180 return performANDCombine(N, DAG, DCI, Subtarget); 1181 case ISD::OR: 1182 return performORCombine(N, DAG, DCI, Subtarget); 1183 case ISD::ADD: 1184 return performADDCombine(N, DAG, DCI, Subtarget); 1185 case ISD::AssertZext: 1186 return performAssertZextCombine(N, DAG, DCI, Subtarget); 1187 case ISD::SHL: 1188 return performSHLCombine(N, DAG, DCI, Subtarget); 1189 case ISD::SUB: 1190 return performSUBCombine(N, DAG, DCI, Subtarget); 1191 } 1192 1193 return SDValue(); 1194 } 1195 1196 bool MipsTargetLowering::isCheapToSpeculateCttz() const { 1197 return Subtarget.hasMips32(); 1198 } 1199 1200 bool MipsTargetLowering::isCheapToSpeculateCtlz() const { 1201 return Subtarget.hasMips32(); 1202 } 1203 1204 void 1205 MipsTargetLowering::LowerOperationWrapper(SDNode *N, 1206 SmallVectorImpl<SDValue> &Results, 1207 SelectionDAG &DAG) const { 1208 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 1209 1210 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I) 1211 Results.push_back(Res.getValue(I)); 1212 } 1213 1214 void 1215 MipsTargetLowering::ReplaceNodeResults(SDNode *N, 1216 SmallVectorImpl<SDValue> &Results, 1217 SelectionDAG &DAG) const { 1218 return LowerOperationWrapper(N, Results, DAG); 1219 } 1220 1221 SDValue MipsTargetLowering:: 1222 LowerOperation(SDValue Op, SelectionDAG &DAG) const 1223 { 1224 switch (Op.getOpcode()) 1225 { 1226 case ISD::BRCOND: return lowerBRCOND(Op, DAG); 1227 case ISD::ConstantPool: return lowerConstantPool(Op, DAG); 1228 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG); 1229 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG); 1230 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG); 1231 case ISD::JumpTable: return lowerJumpTable(Op, DAG); 1232 case ISD::SELECT: return lowerSELECT(Op, DAG); 1233 case ISD::SETCC: return lowerSETCC(Op, DAG); 1234 case ISD::VASTART: return lowerVASTART(Op, DAG); 1235 case ISD::VAARG: return lowerVAARG(Op, DAG); 1236 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG); 1237 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG); 1238 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG); 1239 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG); 1240 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG); 1241 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG); 1242 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true); 1243 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false); 1244 case ISD::LOAD: return lowerLOAD(Op, DAG); 1245 case ISD::STORE: return lowerSTORE(Op, DAG); 1246 case ISD::EH_DWARF_CFA: return lowerEH_DWARF_CFA(Op, DAG); 1247 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG); 1248 } 1249 return SDValue(); 1250 } 1251 1252 //===----------------------------------------------------------------------===// 1253 // Lower helper functions 1254 //===----------------------------------------------------------------------===// 1255 1256 // addLiveIn - This helper function adds the specified physical register to the 1257 // MachineFunction as a live in value. It also creates a corresponding 1258 // virtual register for it. 1259 static unsigned 1260 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC) 1261 { 1262 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC); 1263 MF.getRegInfo().addLiveIn(PReg, VReg); 1264 return VReg; 1265 } 1266 1267 static MachineBasicBlock *insertDivByZeroTrap(MachineInstr &MI, 1268 MachineBasicBlock &MBB, 1269 const TargetInstrInfo &TII, 1270 bool Is64Bit, bool IsMicroMips) { 1271 if (NoZeroDivCheck) 1272 return &MBB; 1273 1274 // Insert instruction "teq $divisor_reg, $zero, 7". 1275 MachineBasicBlock::iterator I(MI); 1276 MachineInstrBuilder MIB; 1277 MachineOperand &Divisor = MI.getOperand(2); 1278 MIB = BuildMI(MBB, std::next(I), MI.getDebugLoc(), 1279 TII.get(IsMicroMips ? Mips::TEQ_MM : Mips::TEQ)) 1280 .addReg(Divisor.getReg(), getKillRegState(Divisor.isKill())) 1281 .addReg(Mips::ZERO) 1282 .addImm(7); 1283 1284 // Use the 32-bit sub-register if this is a 64-bit division. 1285 if (Is64Bit) 1286 MIB->getOperand(0).setSubReg(Mips::sub_32); 1287 1288 // Clear Divisor's kill flag. 1289 Divisor.setIsKill(false); 1290 1291 // We would normally delete the original instruction here but in this case 1292 // we only needed to inject an additional instruction rather than replace it. 1293 1294 return &MBB; 1295 } 1296 1297 MachineBasicBlock * 1298 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, 1299 MachineBasicBlock *BB) const { 1300 switch (MI.getOpcode()) { 1301 default: 1302 llvm_unreachable("Unexpected instr type to insert"); 1303 case Mips::ATOMIC_LOAD_ADD_I8: 1304 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu); 1305 case Mips::ATOMIC_LOAD_ADD_I16: 1306 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu); 1307 case Mips::ATOMIC_LOAD_ADD_I32: 1308 return emitAtomicBinary(MI, BB, 4, Mips::ADDu); 1309 case Mips::ATOMIC_LOAD_ADD_I64: 1310 return emitAtomicBinary(MI, BB, 8, Mips::DADDu); 1311 1312 case Mips::ATOMIC_LOAD_AND_I8: 1313 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND); 1314 case Mips::ATOMIC_LOAD_AND_I16: 1315 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND); 1316 case Mips::ATOMIC_LOAD_AND_I32: 1317 return emitAtomicBinary(MI, BB, 4, Mips::AND); 1318 case Mips::ATOMIC_LOAD_AND_I64: 1319 return emitAtomicBinary(MI, BB, 8, Mips::AND64); 1320 1321 case Mips::ATOMIC_LOAD_OR_I8: 1322 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR); 1323 case Mips::ATOMIC_LOAD_OR_I16: 1324 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR); 1325 case Mips::ATOMIC_LOAD_OR_I32: 1326 return emitAtomicBinary(MI, BB, 4, Mips::OR); 1327 case Mips::ATOMIC_LOAD_OR_I64: 1328 return emitAtomicBinary(MI, BB, 8, Mips::OR64); 1329 1330 case Mips::ATOMIC_LOAD_XOR_I8: 1331 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR); 1332 case Mips::ATOMIC_LOAD_XOR_I16: 1333 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR); 1334 case Mips::ATOMIC_LOAD_XOR_I32: 1335 return emitAtomicBinary(MI, BB, 4, Mips::XOR); 1336 case Mips::ATOMIC_LOAD_XOR_I64: 1337 return emitAtomicBinary(MI, BB, 8, Mips::XOR64); 1338 1339 case Mips::ATOMIC_LOAD_NAND_I8: 1340 return emitAtomicBinaryPartword(MI, BB, 1, 0, true); 1341 case Mips::ATOMIC_LOAD_NAND_I16: 1342 return emitAtomicBinaryPartword(MI, BB, 2, 0, true); 1343 case Mips::ATOMIC_LOAD_NAND_I32: 1344 return emitAtomicBinary(MI, BB, 4, 0, true); 1345 case Mips::ATOMIC_LOAD_NAND_I64: 1346 return emitAtomicBinary(MI, BB, 8, 0, true); 1347 1348 case Mips::ATOMIC_LOAD_SUB_I8: 1349 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu); 1350 case Mips::ATOMIC_LOAD_SUB_I16: 1351 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu); 1352 case Mips::ATOMIC_LOAD_SUB_I32: 1353 return emitAtomicBinary(MI, BB, 4, Mips::SUBu); 1354 case Mips::ATOMIC_LOAD_SUB_I64: 1355 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu); 1356 1357 case Mips::ATOMIC_SWAP_I8: 1358 return emitAtomicBinaryPartword(MI, BB, 1, 0); 1359 case Mips::ATOMIC_SWAP_I16: 1360 return emitAtomicBinaryPartword(MI, BB, 2, 0); 1361 case Mips::ATOMIC_SWAP_I32: 1362 return emitAtomicBinary(MI, BB, 4, 0); 1363 case Mips::ATOMIC_SWAP_I64: 1364 return emitAtomicBinary(MI, BB, 8, 0); 1365 1366 case Mips::ATOMIC_CMP_SWAP_I8: 1367 return emitAtomicCmpSwapPartword(MI, BB, 1); 1368 case Mips::ATOMIC_CMP_SWAP_I16: 1369 return emitAtomicCmpSwapPartword(MI, BB, 2); 1370 case Mips::ATOMIC_CMP_SWAP_I32: 1371 return emitAtomicCmpSwap(MI, BB, 4); 1372 case Mips::ATOMIC_CMP_SWAP_I64: 1373 return emitAtomicCmpSwap(MI, BB, 8); 1374 case Mips::PseudoSDIV: 1375 case Mips::PseudoUDIV: 1376 case Mips::DIV: 1377 case Mips::DIVU: 1378 case Mips::MOD: 1379 case Mips::MODU: 1380 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false, 1381 false); 1382 case Mips::SDIV_MM_Pseudo: 1383 case Mips::UDIV_MM_Pseudo: 1384 case Mips::SDIV_MM: 1385 case Mips::UDIV_MM: 1386 case Mips::DIV_MMR6: 1387 case Mips::DIVU_MMR6: 1388 case Mips::MOD_MMR6: 1389 case Mips::MODU_MMR6: 1390 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false, true); 1391 case Mips::PseudoDSDIV: 1392 case Mips::PseudoDUDIV: 1393 case Mips::DDIV: 1394 case Mips::DDIVU: 1395 case Mips::DMOD: 1396 case Mips::DMODU: 1397 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), true, false); 1398 case Mips::DDIV_MM64R6: 1399 case Mips::DDIVU_MM64R6: 1400 case Mips::DMOD_MM64R6: 1401 case Mips::DMODU_MM64R6: 1402 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), true, true); 1403 1404 case Mips::PseudoSELECT_I: 1405 case Mips::PseudoSELECT_I64: 1406 case Mips::PseudoSELECT_S: 1407 case Mips::PseudoSELECT_D32: 1408 case Mips::PseudoSELECT_D64: 1409 return emitPseudoSELECT(MI, BB, false, Mips::BNE); 1410 case Mips::PseudoSELECTFP_F_I: 1411 case Mips::PseudoSELECTFP_F_I64: 1412 case Mips::PseudoSELECTFP_F_S: 1413 case Mips::PseudoSELECTFP_F_D32: 1414 case Mips::PseudoSELECTFP_F_D64: 1415 return emitPseudoSELECT(MI, BB, true, Mips::BC1F); 1416 case Mips::PseudoSELECTFP_T_I: 1417 case Mips::PseudoSELECTFP_T_I64: 1418 case Mips::PseudoSELECTFP_T_S: 1419 case Mips::PseudoSELECTFP_T_D32: 1420 case Mips::PseudoSELECTFP_T_D64: 1421 return emitPseudoSELECT(MI, BB, true, Mips::BC1T); 1422 } 1423 } 1424 1425 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and 1426 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true) 1427 MachineBasicBlock *MipsTargetLowering::emitAtomicBinary(MachineInstr &MI, 1428 MachineBasicBlock *BB, 1429 unsigned Size, 1430 unsigned BinOpcode, 1431 bool Nand) const { 1432 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary."); 1433 1434 MachineFunction *MF = BB->getParent(); 1435 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1436 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 1437 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1438 const bool ArePtrs64bit = ABI.ArePtrs64bit(); 1439 DebugLoc DL = MI.getDebugLoc(); 1440 unsigned LL, SC, AND, NOR, ZERO, BEQ; 1441 1442 if (Size == 4) { 1443 if (isMicroMips) { 1444 LL = Mips::LL_MM; 1445 SC = Mips::SC_MM; 1446 } else { 1447 LL = Subtarget.hasMips32r6() 1448 ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6) 1449 : (ArePtrs64bit ? Mips::LL64 : Mips::LL); 1450 SC = Subtarget.hasMips32r6() 1451 ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6) 1452 : (ArePtrs64bit ? Mips::SC64 : Mips::SC); 1453 } 1454 1455 AND = Mips::AND; 1456 NOR = Mips::NOR; 1457 ZERO = Mips::ZERO; 1458 BEQ = Mips::BEQ; 1459 } else { 1460 LL = Subtarget.hasMips64r6() ? Mips::LLD_R6 : Mips::LLD; 1461 SC = Subtarget.hasMips64r6() ? Mips::SCD_R6 : Mips::SCD; 1462 AND = Mips::AND64; 1463 NOR = Mips::NOR64; 1464 ZERO = Mips::ZERO_64; 1465 BEQ = Mips::BEQ64; 1466 } 1467 1468 unsigned OldVal = MI.getOperand(0).getReg(); 1469 unsigned Ptr = MI.getOperand(1).getReg(); 1470 unsigned Incr = MI.getOperand(2).getReg(); 1471 1472 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1473 unsigned AndRes = RegInfo.createVirtualRegister(RC); 1474 unsigned Success = RegInfo.createVirtualRegister(RC); 1475 1476 // insert new blocks after the current block 1477 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1478 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1479 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1480 MachineFunction::iterator It = ++BB->getIterator(); 1481 MF->insert(It, loopMBB); 1482 MF->insert(It, exitMBB); 1483 1484 // Transfer the remainder of BB and its successor edges to exitMBB. 1485 exitMBB->splice(exitMBB->begin(), BB, 1486 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1487 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1488 1489 // thisMBB: 1490 // ... 1491 // fallthrough --> loopMBB 1492 BB->addSuccessor(loopMBB); 1493 loopMBB->addSuccessor(loopMBB); 1494 loopMBB->addSuccessor(exitMBB); 1495 1496 // loopMBB: 1497 // ll oldval, 0(ptr) 1498 // <binop> storeval, oldval, incr 1499 // sc success, storeval, 0(ptr) 1500 // beq success, $0, loopMBB 1501 BB = loopMBB; 1502 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0); 1503 if (Nand) { 1504 // and andres, oldval, incr 1505 // nor storeval, $0, andres 1506 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr); 1507 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes); 1508 } else if (BinOpcode) { 1509 // <binop> storeval, oldval, incr 1510 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr); 1511 } else { 1512 StoreVal = Incr; 1513 } 1514 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0); 1515 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB); 1516 1517 MI.eraseFromParent(); // The instruction is gone now. 1518 1519 return exitMBB; 1520 } 1521 1522 MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg( 1523 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg, 1524 unsigned SrcReg) const { 1525 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1526 const DebugLoc &DL = MI.getDebugLoc(); 1527 1528 if (Subtarget.hasMips32r2() && Size == 1) { 1529 BuildMI(BB, DL, TII->get(Mips::SEB), DstReg).addReg(SrcReg); 1530 return BB; 1531 } 1532 1533 if (Subtarget.hasMips32r2() && Size == 2) { 1534 BuildMI(BB, DL, TII->get(Mips::SEH), DstReg).addReg(SrcReg); 1535 return BB; 1536 } 1537 1538 MachineFunction *MF = BB->getParent(); 1539 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1540 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1541 unsigned ScrReg = RegInfo.createVirtualRegister(RC); 1542 1543 assert(Size < 32); 1544 int64_t ShiftImm = 32 - (Size * 8); 1545 1546 BuildMI(BB, DL, TII->get(Mips::SLL), ScrReg).addReg(SrcReg).addImm(ShiftImm); 1547 BuildMI(BB, DL, TII->get(Mips::SRA), DstReg).addReg(ScrReg).addImm(ShiftImm); 1548 1549 return BB; 1550 } 1551 1552 MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword( 1553 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned BinOpcode, 1554 bool Nand) const { 1555 assert((Size == 1 || Size == 2) && 1556 "Unsupported size for EmitAtomicBinaryPartial."); 1557 1558 MachineFunction *MF = BB->getParent(); 1559 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1560 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1561 const bool ArePtrs64bit = ABI.ArePtrs64bit(); 1562 const TargetRegisterClass *RCp = 1563 getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32); 1564 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1565 DebugLoc DL = MI.getDebugLoc(); 1566 1567 unsigned Dest = MI.getOperand(0).getReg(); 1568 unsigned Ptr = MI.getOperand(1).getReg(); 1569 unsigned Incr = MI.getOperand(2).getReg(); 1570 1571 unsigned AlignedAddr = RegInfo.createVirtualRegister(RCp); 1572 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1573 unsigned Mask = RegInfo.createVirtualRegister(RC); 1574 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1575 unsigned NewVal = RegInfo.createVirtualRegister(RC); 1576 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1577 unsigned Incr2 = RegInfo.createVirtualRegister(RC); 1578 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RCp); 1579 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1580 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1581 unsigned AndRes = RegInfo.createVirtualRegister(RC); 1582 unsigned BinOpRes = RegInfo.createVirtualRegister(RC); 1583 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1584 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1585 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1586 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1587 unsigned Success = RegInfo.createVirtualRegister(RC); 1588 1589 unsigned LL, SC; 1590 if (isMicroMips) { 1591 LL = Mips::LL_MM; 1592 SC = Mips::SC_MM; 1593 } else { 1594 LL = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6) 1595 : (ArePtrs64bit ? Mips::LL64 : Mips::LL); 1596 SC = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6) 1597 : (ArePtrs64bit ? Mips::SC64 : Mips::SC); 1598 } 1599 1600 // insert new blocks after the current block 1601 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1602 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1603 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1604 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1605 MachineFunction::iterator It = ++BB->getIterator(); 1606 MF->insert(It, loopMBB); 1607 MF->insert(It, sinkMBB); 1608 MF->insert(It, exitMBB); 1609 1610 // Transfer the remainder of BB and its successor edges to exitMBB. 1611 exitMBB->splice(exitMBB->begin(), BB, 1612 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1613 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1614 1615 BB->addSuccessor(loopMBB); 1616 loopMBB->addSuccessor(loopMBB); 1617 loopMBB->addSuccessor(sinkMBB); 1618 sinkMBB->addSuccessor(exitMBB); 1619 1620 // thisMBB: 1621 // addiu masklsb2,$0,-4 # 0xfffffffc 1622 // and alignedaddr,ptr,masklsb2 1623 // andi ptrlsb2,ptr,3 1624 // sll shiftamt,ptrlsb2,3 1625 // ori maskupper,$0,255 # 0xff 1626 // sll mask,maskupper,shiftamt 1627 // nor mask2,$0,mask 1628 // sll incr2,incr,shiftamt 1629 1630 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1631 BuildMI(BB, DL, TII->get(ABI.GetPtrAddiuOp()), MaskLSB2) 1632 .addReg(ABI.GetNullPtr()).addImm(-4); 1633 BuildMI(BB, DL, TII->get(ABI.GetPtrAndOp()), AlignedAddr) 1634 .addReg(Ptr).addReg(MaskLSB2); 1635 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2) 1636 .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3); 1637 if (Subtarget.isLittle()) { 1638 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1639 } else { 1640 unsigned Off = RegInfo.createVirtualRegister(RC); 1641 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1642 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1643 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1644 } 1645 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1646 .addReg(Mips::ZERO).addImm(MaskImm); 1647 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1648 .addReg(MaskUpper).addReg(ShiftAmt); 1649 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1650 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt); 1651 1652 // atomic.load.binop 1653 // loopMBB: 1654 // ll oldval,0(alignedaddr) 1655 // binop binopres,oldval,incr2 1656 // and newval,binopres,mask 1657 // and maskedoldval0,oldval,mask2 1658 // or storeval,maskedoldval0,newval 1659 // sc success,storeval,0(alignedaddr) 1660 // beq success,$0,loopMBB 1661 1662 // atomic.swap 1663 // loopMBB: 1664 // ll oldval,0(alignedaddr) 1665 // and newval,incr2,mask 1666 // and maskedoldval0,oldval,mask2 1667 // or storeval,maskedoldval0,newval 1668 // sc success,storeval,0(alignedaddr) 1669 // beq success,$0,loopMBB 1670 1671 BB = loopMBB; 1672 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0); 1673 if (Nand) { 1674 // and andres, oldval, incr2 1675 // nor binopres, $0, andres 1676 // and newval, binopres, mask 1677 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2); 1678 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes) 1679 .addReg(Mips::ZERO).addReg(AndRes); 1680 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1681 } else if (BinOpcode) { 1682 // <binop> binopres, oldval, incr2 1683 // and newval, binopres, mask 1684 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2); 1685 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1686 } else { // atomic.swap 1687 // and newval, incr2, mask 1688 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask); 1689 } 1690 1691 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1692 .addReg(OldVal).addReg(Mask2); 1693 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1694 .addReg(MaskedOldVal0).addReg(NewVal); 1695 BuildMI(BB, DL, TII->get(SC), Success) 1696 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1697 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1698 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB); 1699 1700 // sinkMBB: 1701 // and maskedoldval1,oldval,mask 1702 // srl srlres,maskedoldval1,shiftamt 1703 // sign_extend dest,srlres 1704 BB = sinkMBB; 1705 1706 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1707 .addReg(OldVal).addReg(Mask); 1708 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1709 .addReg(MaskedOldVal1).addReg(ShiftAmt); 1710 BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes); 1711 1712 MI.eraseFromParent(); // The instruction is gone now. 1713 1714 return exitMBB; 1715 } 1716 1717 MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwap(MachineInstr &MI, 1718 MachineBasicBlock *BB, 1719 unsigned Size) const { 1720 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap."); 1721 1722 MachineFunction *MF = BB->getParent(); 1723 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1724 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 1725 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1726 const bool ArePtrs64bit = ABI.ArePtrs64bit(); 1727 DebugLoc DL = MI.getDebugLoc(); 1728 unsigned LL, SC, ZERO, BNE, BEQ; 1729 1730 if (Size == 4) { 1731 if (isMicroMips) { 1732 LL = Mips::LL_MM; 1733 SC = Mips::SC_MM; 1734 } else { 1735 LL = Subtarget.hasMips32r6() 1736 ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6) 1737 : (ArePtrs64bit ? Mips::LL64 : Mips::LL); 1738 SC = Subtarget.hasMips32r6() 1739 ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6) 1740 : (ArePtrs64bit ? Mips::SC64 : Mips::SC); 1741 } 1742 1743 ZERO = Mips::ZERO; 1744 BNE = Mips::BNE; 1745 BEQ = Mips::BEQ; 1746 } else { 1747 LL = Subtarget.hasMips64r6() ? Mips::LLD_R6 : Mips::LLD; 1748 SC = Subtarget.hasMips64r6() ? Mips::SCD_R6 : Mips::SCD; 1749 ZERO = Mips::ZERO_64; 1750 BNE = Mips::BNE64; 1751 BEQ = Mips::BEQ64; 1752 } 1753 1754 unsigned Dest = MI.getOperand(0).getReg(); 1755 unsigned Ptr = MI.getOperand(1).getReg(); 1756 unsigned OldVal = MI.getOperand(2).getReg(); 1757 unsigned NewVal = MI.getOperand(3).getReg(); 1758 1759 unsigned Success = RegInfo.createVirtualRegister(RC); 1760 1761 // insert new blocks after the current block 1762 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1763 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1764 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1765 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1766 MachineFunction::iterator It = ++BB->getIterator(); 1767 MF->insert(It, loop1MBB); 1768 MF->insert(It, loop2MBB); 1769 MF->insert(It, exitMBB); 1770 1771 // Transfer the remainder of BB and its successor edges to exitMBB. 1772 exitMBB->splice(exitMBB->begin(), BB, 1773 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1774 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1775 1776 // thisMBB: 1777 // ... 1778 // fallthrough --> loop1MBB 1779 BB->addSuccessor(loop1MBB); 1780 loop1MBB->addSuccessor(exitMBB); 1781 loop1MBB->addSuccessor(loop2MBB); 1782 loop2MBB->addSuccessor(loop1MBB); 1783 loop2MBB->addSuccessor(exitMBB); 1784 1785 // loop1MBB: 1786 // ll dest, 0(ptr) 1787 // bne dest, oldval, exitMBB 1788 BB = loop1MBB; 1789 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0); 1790 BuildMI(BB, DL, TII->get(BNE)) 1791 .addReg(Dest).addReg(OldVal).addMBB(exitMBB); 1792 1793 // loop2MBB: 1794 // sc success, newval, 0(ptr) 1795 // beq success, $0, loop1MBB 1796 BB = loop2MBB; 1797 BuildMI(BB, DL, TII->get(SC), Success) 1798 .addReg(NewVal).addReg(Ptr).addImm(0); 1799 BuildMI(BB, DL, TII->get(BEQ)) 1800 .addReg(Success).addReg(ZERO).addMBB(loop1MBB); 1801 1802 MI.eraseFromParent(); // The instruction is gone now. 1803 1804 return exitMBB; 1805 } 1806 1807 MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwapPartword( 1808 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const { 1809 assert((Size == 1 || Size == 2) && 1810 "Unsupported size for EmitAtomicCmpSwapPartial."); 1811 1812 MachineFunction *MF = BB->getParent(); 1813 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1814 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1815 const bool ArePtrs64bit = ABI.ArePtrs64bit(); 1816 const TargetRegisterClass *RCp = 1817 getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32); 1818 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1819 DebugLoc DL = MI.getDebugLoc(); 1820 1821 unsigned Dest = MI.getOperand(0).getReg(); 1822 unsigned Ptr = MI.getOperand(1).getReg(); 1823 unsigned CmpVal = MI.getOperand(2).getReg(); 1824 unsigned NewVal = MI.getOperand(3).getReg(); 1825 1826 unsigned AlignedAddr = RegInfo.createVirtualRegister(RCp); 1827 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1828 unsigned Mask = RegInfo.createVirtualRegister(RC); 1829 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1830 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC); 1831 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1832 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1833 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC); 1834 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RCp); 1835 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1836 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1837 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC); 1838 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC); 1839 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1840 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1841 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1842 unsigned Success = RegInfo.createVirtualRegister(RC); 1843 unsigned LL, SC; 1844 1845 if (isMicroMips) { 1846 LL = Mips::LL_MM; 1847 SC = Mips::SC_MM; 1848 } else { 1849 LL = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6) 1850 : (ArePtrs64bit ? Mips::LL64 : Mips::LL); 1851 SC = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6) 1852 : (ArePtrs64bit ? Mips::SC64 : Mips::SC); 1853 } 1854 1855 // insert new blocks after the current block 1856 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1857 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1858 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1859 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1860 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1861 MachineFunction::iterator It = ++BB->getIterator(); 1862 MF->insert(It, loop1MBB); 1863 MF->insert(It, loop2MBB); 1864 MF->insert(It, sinkMBB); 1865 MF->insert(It, exitMBB); 1866 1867 // Transfer the remainder of BB and its successor edges to exitMBB. 1868 exitMBB->splice(exitMBB->begin(), BB, 1869 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1870 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1871 1872 BB->addSuccessor(loop1MBB); 1873 loop1MBB->addSuccessor(sinkMBB); 1874 loop1MBB->addSuccessor(loop2MBB); 1875 loop2MBB->addSuccessor(loop1MBB); 1876 loop2MBB->addSuccessor(sinkMBB); 1877 sinkMBB->addSuccessor(exitMBB); 1878 1879 // FIXME: computation of newval2 can be moved to loop2MBB. 1880 // thisMBB: 1881 // addiu masklsb2,$0,-4 # 0xfffffffc 1882 // and alignedaddr,ptr,masklsb2 1883 // andi ptrlsb2,ptr,3 1884 // xori ptrlsb2,ptrlsb2,3 # Only for BE 1885 // sll shiftamt,ptrlsb2,3 1886 // ori maskupper,$0,255 # 0xff 1887 // sll mask,maskupper,shiftamt 1888 // nor mask2,$0,mask 1889 // andi maskedcmpval,cmpval,255 1890 // sll shiftedcmpval,maskedcmpval,shiftamt 1891 // andi maskednewval,newval,255 1892 // sll shiftednewval,maskednewval,shiftamt 1893 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1894 BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::DADDiu : Mips::ADDiu), MaskLSB2) 1895 .addReg(ABI.GetNullPtr()).addImm(-4); 1896 BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::AND64 : Mips::AND), AlignedAddr) 1897 .addReg(Ptr).addReg(MaskLSB2); 1898 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2) 1899 .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3); 1900 if (Subtarget.isLittle()) { 1901 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1902 } else { 1903 unsigned Off = RegInfo.createVirtualRegister(RC); 1904 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1905 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1906 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1907 } 1908 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1909 .addReg(Mips::ZERO).addImm(MaskImm); 1910 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1911 .addReg(MaskUpper).addReg(ShiftAmt); 1912 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1913 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal) 1914 .addReg(CmpVal).addImm(MaskImm); 1915 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal) 1916 .addReg(MaskedCmpVal).addReg(ShiftAmt); 1917 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal) 1918 .addReg(NewVal).addImm(MaskImm); 1919 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal) 1920 .addReg(MaskedNewVal).addReg(ShiftAmt); 1921 1922 // loop1MBB: 1923 // ll oldval,0(alginedaddr) 1924 // and maskedoldval0,oldval,mask 1925 // bne maskedoldval0,shiftedcmpval,sinkMBB 1926 BB = loop1MBB; 1927 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0); 1928 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1929 .addReg(OldVal).addReg(Mask); 1930 BuildMI(BB, DL, TII->get(Mips::BNE)) 1931 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB); 1932 1933 // loop2MBB: 1934 // and maskedoldval1,oldval,mask2 1935 // or storeval,maskedoldval1,shiftednewval 1936 // sc success,storeval,0(alignedaddr) 1937 // beq success,$0,loop1MBB 1938 BB = loop2MBB; 1939 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1940 .addReg(OldVal).addReg(Mask2); 1941 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1942 .addReg(MaskedOldVal1).addReg(ShiftedNewVal); 1943 BuildMI(BB, DL, TII->get(SC), Success) 1944 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1945 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1946 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB); 1947 1948 // sinkMBB: 1949 // srl srlres,maskedoldval0,shiftamt 1950 // sign_extend dest,srlres 1951 BB = sinkMBB; 1952 1953 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1954 .addReg(MaskedOldVal0).addReg(ShiftAmt); 1955 BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes); 1956 1957 MI.eraseFromParent(); // The instruction is gone now. 1958 1959 return exitMBB; 1960 } 1961 1962 SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const { 1963 // The first operand is the chain, the second is the condition, the third is 1964 // the block to branch to if the condition is true. 1965 SDValue Chain = Op.getOperand(0); 1966 SDValue Dest = Op.getOperand(2); 1967 SDLoc DL(Op); 1968 1969 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 1970 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1)); 1971 1972 // Return if flag is not set by a floating point comparison. 1973 if (CondRes.getOpcode() != MipsISD::FPCmp) 1974 return Op; 1975 1976 SDValue CCNode = CondRes.getOperand(2); 1977 Mips::CondCode CC = 1978 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue(); 1979 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T; 1980 SDValue BrCode = DAG.getConstant(Opc, DL, MVT::i32); 1981 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 1982 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode, 1983 FCC0, Dest, CondRes); 1984 } 1985 1986 SDValue MipsTargetLowering:: 1987 lowerSELECT(SDValue Op, SelectionDAG &DAG) const 1988 { 1989 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 1990 SDValue Cond = createFPCmp(DAG, Op.getOperand(0)); 1991 1992 // Return if flag is not set by a floating point comparison. 1993 if (Cond.getOpcode() != MipsISD::FPCmp) 1994 return Op; 1995 1996 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2), 1997 SDLoc(Op)); 1998 } 1999 2000 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const { 2001 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 2002 SDValue Cond = createFPCmp(DAG, Op); 2003 2004 assert(Cond.getOpcode() == MipsISD::FPCmp && 2005 "Floating point operand expected."); 2006 2007 SDLoc DL(Op); 2008 SDValue True = DAG.getConstant(1, DL, MVT::i32); 2009 SDValue False = DAG.getConstant(0, DL, MVT::i32); 2010 2011 return createCMovFP(DAG, Cond, True, False, DL); 2012 } 2013 2014 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op, 2015 SelectionDAG &DAG) const { 2016 EVT Ty = Op.getValueType(); 2017 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op); 2018 const GlobalValue *GV = N->getGlobal(); 2019 2020 if (!isPositionIndependent()) { 2021 const MipsTargetObjectFile *TLOF = 2022 static_cast<const MipsTargetObjectFile *>( 2023 getTargetMachine().getObjFileLowering()); 2024 const GlobalObject *GO = GV->getBaseObject(); 2025 if (GO && TLOF->IsGlobalInSmallSection(GO, getTargetMachine())) 2026 // %gp_rel relocation 2027 return getAddrGPRel(N, SDLoc(N), Ty, DAG); 2028 2029 // %hi/%lo relocation 2030 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2031 // %highest/%higher/%hi/%lo relocation 2032 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2033 } 2034 2035 // Every other architecture would use shouldAssumeDSOLocal in here, but 2036 // mips is special. 2037 // * In PIC code mips requires got loads even for local statics! 2038 // * To save on got entries, for local statics the got entry contains the 2039 // page and an additional add instruction takes care of the low bits. 2040 // * It is legal to access a hidden symbol with a non hidden undefined, 2041 // so one cannot guarantee that all access to a hidden symbol will know 2042 // it is hidden. 2043 // * Mips linkers don't support creating a page and a full got entry for 2044 // the same symbol. 2045 // * Given all that, we have to use a full got entry for hidden symbols :-( 2046 if (GV->hasLocalLinkage()) 2047 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2048 2049 if (LargeGOT) 2050 return getAddrGlobalLargeGOT( 2051 N, SDLoc(N), Ty, DAG, MipsII::MO_GOT_HI16, MipsII::MO_GOT_LO16, 2052 DAG.getEntryNode(), 2053 MachinePointerInfo::getGOT(DAG.getMachineFunction())); 2054 2055 return getAddrGlobal( 2056 N, SDLoc(N), Ty, DAG, 2057 (ABI.IsN32() || ABI.IsN64()) ? MipsII::MO_GOT_DISP : MipsII::MO_GOT, 2058 DAG.getEntryNode(), MachinePointerInfo::getGOT(DAG.getMachineFunction())); 2059 } 2060 2061 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op, 2062 SelectionDAG &DAG) const { 2063 BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op); 2064 EVT Ty = Op.getValueType(); 2065 2066 if (!isPositionIndependent()) 2067 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2068 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2069 2070 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2071 } 2072 2073 SDValue MipsTargetLowering:: 2074 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const 2075 { 2076 // If the relocation model is PIC, use the General Dynamic TLS Model or 2077 // Local Dynamic TLS model, otherwise use the Initial Exec or 2078 // Local Exec TLS Model. 2079 2080 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 2081 if (DAG.getTarget().Options.EmulatedTLS) 2082 return LowerToTLSEmulatedModel(GA, DAG); 2083 2084 SDLoc DL(GA); 2085 const GlobalValue *GV = GA->getGlobal(); 2086 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 2087 2088 TLSModel::Model model = getTargetMachine().getTLSModel(GV); 2089 2090 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { 2091 // General Dynamic and Local Dynamic TLS Model. 2092 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM 2093 : MipsII::MO_TLSGD; 2094 2095 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag); 2096 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, 2097 getGlobalReg(DAG, PtrVT), TGA); 2098 unsigned PtrSize = PtrVT.getSizeInBits(); 2099 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize); 2100 2101 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT); 2102 2103 ArgListTy Args; 2104 ArgListEntry Entry; 2105 Entry.Node = Argument; 2106 Entry.Ty = PtrTy; 2107 Args.push_back(Entry); 2108 2109 TargetLowering::CallLoweringInfo CLI(DAG); 2110 CLI.setDebugLoc(DL) 2111 .setChain(DAG.getEntryNode()) 2112 .setLibCallee(CallingConv::C, PtrTy, TlsGetAddr, std::move(Args)); 2113 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 2114 2115 SDValue Ret = CallResult.first; 2116 2117 if (model != TLSModel::LocalDynamic) 2118 return Ret; 2119 2120 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2121 MipsII::MO_DTPREL_HI); 2122 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 2123 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2124 MipsII::MO_DTPREL_LO); 2125 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 2126 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret); 2127 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo); 2128 } 2129 2130 SDValue Offset; 2131 if (model == TLSModel::InitialExec) { 2132 // Initial Exec TLS Model 2133 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2134 MipsII::MO_GOTTPREL); 2135 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT), 2136 TGA); 2137 Offset = 2138 DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), TGA, MachinePointerInfo()); 2139 } else { 2140 // Local Exec TLS Model 2141 assert(model == TLSModel::LocalExec); 2142 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2143 MipsII::MO_TPREL_HI); 2144 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2145 MipsII::MO_TPREL_LO); 2146 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 2147 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 2148 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo); 2149 } 2150 2151 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT); 2152 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset); 2153 } 2154 2155 SDValue MipsTargetLowering:: 2156 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const 2157 { 2158 JumpTableSDNode *N = cast<JumpTableSDNode>(Op); 2159 EVT Ty = Op.getValueType(); 2160 2161 if (!isPositionIndependent()) 2162 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2163 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2164 2165 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2166 } 2167 2168 SDValue MipsTargetLowering:: 2169 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const 2170 { 2171 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op); 2172 EVT Ty = Op.getValueType(); 2173 2174 if (!isPositionIndependent()) { 2175 const MipsTargetObjectFile *TLOF = 2176 static_cast<const MipsTargetObjectFile *>( 2177 getTargetMachine().getObjFileLowering()); 2178 2179 if (TLOF->IsConstantInSmallSection(DAG.getDataLayout(), N->getConstVal(), 2180 getTargetMachine())) 2181 // %gp_rel relocation 2182 return getAddrGPRel(N, SDLoc(N), Ty, DAG); 2183 2184 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2185 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2186 } 2187 2188 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2189 } 2190 2191 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const { 2192 MachineFunction &MF = DAG.getMachineFunction(); 2193 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 2194 2195 SDLoc DL(Op); 2196 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), 2197 getPointerTy(MF.getDataLayout())); 2198 2199 // vastart just stores the address of the VarArgsFrameIndex slot into the 2200 // memory location argument. 2201 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 2202 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1), 2203 MachinePointerInfo(SV)); 2204 } 2205 2206 SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const { 2207 SDNode *Node = Op.getNode(); 2208 EVT VT = Node->getValueType(0); 2209 SDValue Chain = Node->getOperand(0); 2210 SDValue VAListPtr = Node->getOperand(1); 2211 unsigned Align = Node->getConstantOperandVal(3); 2212 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue(); 2213 SDLoc DL(Node); 2214 unsigned ArgSlotSizeInBytes = (ABI.IsN32() || ABI.IsN64()) ? 8 : 4; 2215 2216 SDValue VAListLoad = DAG.getLoad(getPointerTy(DAG.getDataLayout()), DL, Chain, 2217 VAListPtr, MachinePointerInfo(SV)); 2218 SDValue VAList = VAListLoad; 2219 2220 // Re-align the pointer if necessary. 2221 // It should only ever be necessary for 64-bit types on O32 since the minimum 2222 // argument alignment is the same as the maximum type alignment for N32/N64. 2223 // 2224 // FIXME: We currently align too often. The code generator doesn't notice 2225 // when the pointer is still aligned from the last va_arg (or pair of 2226 // va_args for the i64 on O32 case). 2227 if (Align > getMinStackArgumentAlignment()) { 2228 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2"); 2229 2230 VAList = DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList, 2231 DAG.getConstant(Align - 1, DL, VAList.getValueType())); 2232 2233 VAList = DAG.getNode(ISD::AND, DL, VAList.getValueType(), VAList, 2234 DAG.getConstant(-(int64_t)Align, DL, 2235 VAList.getValueType())); 2236 } 2237 2238 // Increment the pointer, VAList, to the next vaarg. 2239 auto &TD = DAG.getDataLayout(); 2240 unsigned ArgSizeInBytes = 2241 TD.getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())); 2242 SDValue Tmp3 = 2243 DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList, 2244 DAG.getConstant(alignTo(ArgSizeInBytes, ArgSlotSizeInBytes), 2245 DL, VAList.getValueType())); 2246 // Store the incremented VAList to the legalized pointer 2247 Chain = DAG.getStore(VAListLoad.getValue(1), DL, Tmp3, VAListPtr, 2248 MachinePointerInfo(SV)); 2249 2250 // In big-endian mode we must adjust the pointer when the load size is smaller 2251 // than the argument slot size. We must also reduce the known alignment to 2252 // match. For example in the N64 ABI, we must add 4 bytes to the offset to get 2253 // the correct half of the slot, and reduce the alignment from 8 (slot 2254 // alignment) down to 4 (type alignment). 2255 if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) { 2256 unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes; 2257 VAList = DAG.getNode(ISD::ADD, DL, VAListPtr.getValueType(), VAList, 2258 DAG.getIntPtrConstant(Adjustment, DL)); 2259 } 2260 // Load the actual argument out of the pointer VAList 2261 return DAG.getLoad(VT, DL, Chain, VAList, MachinePointerInfo()); 2262 } 2263 2264 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, 2265 bool HasExtractInsert) { 2266 EVT TyX = Op.getOperand(0).getValueType(); 2267 EVT TyY = Op.getOperand(1).getValueType(); 2268 SDLoc DL(Op); 2269 SDValue Const1 = DAG.getConstant(1, DL, MVT::i32); 2270 SDValue Const31 = DAG.getConstant(31, DL, MVT::i32); 2271 SDValue Res; 2272 2273 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 2274 // to i32. 2275 SDValue X = (TyX == MVT::f32) ? 2276 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) : 2277 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 2278 Const1); 2279 SDValue Y = (TyY == MVT::f32) ? 2280 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) : 2281 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1), 2282 Const1); 2283 2284 if (HasExtractInsert) { 2285 // ext E, Y, 31, 1 ; extract bit31 of Y 2286 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X 2287 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1); 2288 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X); 2289 } else { 2290 // sll SllX, X, 1 2291 // srl SrlX, SllX, 1 2292 // srl SrlY, Y, 31 2293 // sll SllY, SrlX, 31 2294 // or Or, SrlX, SllY 2295 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 2296 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 2297 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31); 2298 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31); 2299 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY); 2300 } 2301 2302 if (TyX == MVT::f32) 2303 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res); 2304 2305 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2306 Op.getOperand(0), 2307 DAG.getConstant(0, DL, MVT::i32)); 2308 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 2309 } 2310 2311 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, 2312 bool HasExtractInsert) { 2313 unsigned WidthX = Op.getOperand(0).getValueSizeInBits(); 2314 unsigned WidthY = Op.getOperand(1).getValueSizeInBits(); 2315 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY); 2316 SDLoc DL(Op); 2317 SDValue Const1 = DAG.getConstant(1, DL, MVT::i32); 2318 2319 // Bitcast to integer nodes. 2320 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0)); 2321 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1)); 2322 2323 if (HasExtractInsert) { 2324 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y 2325 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X 2326 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y, 2327 DAG.getConstant(WidthY - 1, DL, MVT::i32), Const1); 2328 2329 if (WidthX > WidthY) 2330 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E); 2331 else if (WidthY > WidthX) 2332 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E); 2333 2334 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E, 2335 DAG.getConstant(WidthX - 1, DL, MVT::i32), Const1, 2336 X); 2337 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I); 2338 } 2339 2340 // (d)sll SllX, X, 1 2341 // (d)srl SrlX, SllX, 1 2342 // (d)srl SrlY, Y, width(Y)-1 2343 // (d)sll SllY, SrlX, width(Y)-1 2344 // or Or, SrlX, SllY 2345 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1); 2346 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1); 2347 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y, 2348 DAG.getConstant(WidthY - 1, DL, MVT::i32)); 2349 2350 if (WidthX > WidthY) 2351 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY); 2352 else if (WidthY > WidthX) 2353 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY); 2354 2355 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY, 2356 DAG.getConstant(WidthX - 1, DL, MVT::i32)); 2357 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY); 2358 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or); 2359 } 2360 2361 SDValue 2362 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 2363 if (Subtarget.isGP64bit()) 2364 return lowerFCOPYSIGN64(Op, DAG, Subtarget.hasExtractInsert()); 2365 2366 return lowerFCOPYSIGN32(Op, DAG, Subtarget.hasExtractInsert()); 2367 } 2368 2369 SDValue MipsTargetLowering:: 2370 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 2371 // check the depth 2372 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 2373 "Frame address can only be determined for current frame."); 2374 2375 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 2376 MFI.setFrameAddressIsTaken(true); 2377 EVT VT = Op.getValueType(); 2378 SDLoc DL(Op); 2379 SDValue FrameAddr = DAG.getCopyFromReg( 2380 DAG.getEntryNode(), DL, ABI.IsN64() ? Mips::FP_64 : Mips::FP, VT); 2381 return FrameAddr; 2382 } 2383 2384 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op, 2385 SelectionDAG &DAG) const { 2386 if (verifyReturnAddressArgumentIsConstant(Op, DAG)) 2387 return SDValue(); 2388 2389 // check the depth 2390 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 2391 "Return address can be determined only for current frame."); 2392 2393 MachineFunction &MF = DAG.getMachineFunction(); 2394 MachineFrameInfo &MFI = MF.getFrameInfo(); 2395 MVT VT = Op.getSimpleValueType(); 2396 unsigned RA = ABI.IsN64() ? Mips::RA_64 : Mips::RA; 2397 MFI.setReturnAddressIsTaken(true); 2398 2399 // Return RA, which contains the return address. Mark it an implicit live-in. 2400 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT)); 2401 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT); 2402 } 2403 2404 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is 2405 // generated from __builtin_eh_return (offset, handler) 2406 // The effect of this is to adjust the stack pointer by "offset" 2407 // and then branch to "handler". 2408 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG) 2409 const { 2410 MachineFunction &MF = DAG.getMachineFunction(); 2411 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2412 2413 MipsFI->setCallsEhReturn(); 2414 SDValue Chain = Op.getOperand(0); 2415 SDValue Offset = Op.getOperand(1); 2416 SDValue Handler = Op.getOperand(2); 2417 SDLoc DL(Op); 2418 EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32; 2419 2420 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and 2421 // EH_RETURN nodes, so that instructions are emitted back-to-back. 2422 unsigned OffsetReg = ABI.IsN64() ? Mips::V1_64 : Mips::V1; 2423 unsigned AddrReg = ABI.IsN64() ? Mips::V0_64 : Mips::V0; 2424 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue()); 2425 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1)); 2426 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain, 2427 DAG.getRegister(OffsetReg, Ty), 2428 DAG.getRegister(AddrReg, getPointerTy(MF.getDataLayout())), 2429 Chain.getValue(1)); 2430 } 2431 2432 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op, 2433 SelectionDAG &DAG) const { 2434 // FIXME: Need pseudo-fence for 'singlethread' fences 2435 // FIXME: Set SType for weaker fences where supported/appropriate. 2436 unsigned SType = 0; 2437 SDLoc DL(Op); 2438 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0), 2439 DAG.getConstant(SType, DL, MVT::i32)); 2440 } 2441 2442 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op, 2443 SelectionDAG &DAG) const { 2444 SDLoc DL(Op); 2445 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; 2446 2447 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 2448 SDValue Shamt = Op.getOperand(2); 2449 // if shamt < (VT.bits): 2450 // lo = (shl lo, shamt) 2451 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt)) 2452 // else: 2453 // lo = 0 2454 // hi = (shl lo, shamt[4:0]) 2455 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 2456 DAG.getConstant(-1, DL, MVT::i32)); 2457 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, 2458 DAG.getConstant(1, DL, VT)); 2459 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, Not); 2460 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt); 2461 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo); 2462 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt); 2463 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 2464 DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32)); 2465 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, 2466 DAG.getConstant(0, DL, VT), ShiftLeftLo); 2467 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftLeftLo, Or); 2468 2469 SDValue Ops[2] = {Lo, Hi}; 2470 return DAG.getMergeValues(Ops, DL); 2471 } 2472 2473 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, 2474 bool IsSRA) const { 2475 SDLoc DL(Op); 2476 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 2477 SDValue Shamt = Op.getOperand(2); 2478 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; 2479 2480 // if shamt < (VT.bits): 2481 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt)) 2482 // if isSRA: 2483 // hi = (sra hi, shamt) 2484 // else: 2485 // hi = (srl hi, shamt) 2486 // else: 2487 // if isSRA: 2488 // lo = (sra hi, shamt[4:0]) 2489 // hi = (sra hi, 31) 2490 // else: 2491 // lo = (srl hi, shamt[4:0]) 2492 // hi = 0 2493 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 2494 DAG.getConstant(-1, DL, MVT::i32)); 2495 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, VT, Hi, 2496 DAG.getConstant(1, DL, VT)); 2497 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, ShiftLeft1Hi, Not); 2498 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt); 2499 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo); 2500 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, 2501 DL, VT, Hi, Shamt); 2502 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 2503 DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32)); 2504 SDValue Ext = DAG.getNode(ISD::SRA, DL, VT, Hi, 2505 DAG.getConstant(VT.getSizeInBits() - 1, DL, VT)); 2506 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftRightHi, Or); 2507 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, 2508 IsSRA ? Ext : DAG.getConstant(0, DL, VT), ShiftRightHi); 2509 2510 SDValue Ops[2] = {Lo, Hi}; 2511 return DAG.getMergeValues(Ops, DL); 2512 } 2513 2514 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD, 2515 SDValue Chain, SDValue Src, unsigned Offset) { 2516 SDValue Ptr = LD->getBasePtr(); 2517 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT(); 2518 EVT BasePtrVT = Ptr.getValueType(); 2519 SDLoc DL(LD); 2520 SDVTList VTList = DAG.getVTList(VT, MVT::Other); 2521 2522 if (Offset) 2523 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2524 DAG.getConstant(Offset, DL, BasePtrVT)); 2525 2526 SDValue Ops[] = { Chain, Ptr, Src }; 2527 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT, 2528 LD->getMemOperand()); 2529 } 2530 2531 // Expand an unaligned 32 or 64-bit integer load node. 2532 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const { 2533 LoadSDNode *LD = cast<LoadSDNode>(Op); 2534 EVT MemVT = LD->getMemoryVT(); 2535 2536 if (Subtarget.systemSupportsUnalignedAccess()) 2537 return Op; 2538 2539 // Return if load is aligned or if MemVT is neither i32 nor i64. 2540 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) || 2541 ((MemVT != MVT::i32) && (MemVT != MVT::i64))) 2542 return SDValue(); 2543 2544 bool IsLittle = Subtarget.isLittle(); 2545 EVT VT = Op.getValueType(); 2546 ISD::LoadExtType ExtType = LD->getExtensionType(); 2547 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT); 2548 2549 assert((VT == MVT::i32) || (VT == MVT::i64)); 2550 2551 // Expand 2552 // (set dst, (i64 (load baseptr))) 2553 // to 2554 // (set tmp, (ldl (add baseptr, 7), undef)) 2555 // (set dst, (ldr baseptr, tmp)) 2556 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) { 2557 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef, 2558 IsLittle ? 7 : 0); 2559 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL, 2560 IsLittle ? 0 : 7); 2561 } 2562 2563 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef, 2564 IsLittle ? 3 : 0); 2565 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL, 2566 IsLittle ? 0 : 3); 2567 2568 // Expand 2569 // (set dst, (i32 (load baseptr))) or 2570 // (set dst, (i64 (sextload baseptr))) or 2571 // (set dst, (i64 (extload baseptr))) 2572 // to 2573 // (set tmp, (lwl (add baseptr, 3), undef)) 2574 // (set dst, (lwr baseptr, tmp)) 2575 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) || 2576 (ExtType == ISD::EXTLOAD)) 2577 return LWR; 2578 2579 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD)); 2580 2581 // Expand 2582 // (set dst, (i64 (zextload baseptr))) 2583 // to 2584 // (set tmp0, (lwl (add baseptr, 3), undef)) 2585 // (set tmp1, (lwr baseptr, tmp0)) 2586 // (set tmp2, (shl tmp1, 32)) 2587 // (set dst, (srl tmp2, 32)) 2588 SDLoc DL(LD); 2589 SDValue Const32 = DAG.getConstant(32, DL, MVT::i32); 2590 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32); 2591 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32); 2592 SDValue Ops[] = { SRL, LWR.getValue(1) }; 2593 return DAG.getMergeValues(Ops, DL); 2594 } 2595 2596 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD, 2597 SDValue Chain, unsigned Offset) { 2598 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue(); 2599 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType(); 2600 SDLoc DL(SD); 2601 SDVTList VTList = DAG.getVTList(MVT::Other); 2602 2603 if (Offset) 2604 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2605 DAG.getConstant(Offset, DL, BasePtrVT)); 2606 2607 SDValue Ops[] = { Chain, Value, Ptr }; 2608 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT, 2609 SD->getMemOperand()); 2610 } 2611 2612 // Expand an unaligned 32 or 64-bit integer store node. 2613 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG, 2614 bool IsLittle) { 2615 SDValue Value = SD->getValue(), Chain = SD->getChain(); 2616 EVT VT = Value.getValueType(); 2617 2618 // Expand 2619 // (store val, baseptr) or 2620 // (truncstore val, baseptr) 2621 // to 2622 // (swl val, (add baseptr, 3)) 2623 // (swr val, baseptr) 2624 if ((VT == MVT::i32) || SD->isTruncatingStore()) { 2625 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain, 2626 IsLittle ? 3 : 0); 2627 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3); 2628 } 2629 2630 assert(VT == MVT::i64); 2631 2632 // Expand 2633 // (store val, baseptr) 2634 // to 2635 // (sdl val, (add baseptr, 7)) 2636 // (sdr val, baseptr) 2637 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0); 2638 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7); 2639 } 2640 2641 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr). 2642 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) { 2643 SDValue Val = SD->getValue(); 2644 2645 if (Val.getOpcode() != ISD::FP_TO_SINT) 2646 return SDValue(); 2647 2648 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits()); 2649 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy, 2650 Val.getOperand(0)); 2651 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(), 2652 SD->getPointerInfo(), SD->getAlignment(), 2653 SD->getMemOperand()->getFlags()); 2654 } 2655 2656 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const { 2657 StoreSDNode *SD = cast<StoreSDNode>(Op); 2658 EVT MemVT = SD->getMemoryVT(); 2659 2660 // Lower unaligned integer stores. 2661 if (!Subtarget.systemSupportsUnalignedAccess() && 2662 (SD->getAlignment() < MemVT.getSizeInBits() / 8) && 2663 ((MemVT == MVT::i32) || (MemVT == MVT::i64))) 2664 return lowerUnalignedIntStore(SD, DAG, Subtarget.isLittle()); 2665 2666 return lowerFP_TO_SINT_STORE(SD, DAG); 2667 } 2668 2669 SDValue MipsTargetLowering::lowerEH_DWARF_CFA(SDValue Op, 2670 SelectionDAG &DAG) const { 2671 2672 // Return a fixed StackObject with offset 0 which points to the old stack 2673 // pointer. 2674 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 2675 EVT ValTy = Op->getValueType(0); 2676 int FI = MFI.CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false); 2677 return DAG.getFrameIndex(FI, ValTy); 2678 } 2679 2680 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op, 2681 SelectionDAG &DAG) const { 2682 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits()); 2683 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy, 2684 Op.getOperand(0)); 2685 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc); 2686 } 2687 2688 //===----------------------------------------------------------------------===// 2689 // Calling Convention Implementation 2690 //===----------------------------------------------------------------------===// 2691 2692 //===----------------------------------------------------------------------===// 2693 // TODO: Implement a generic logic using tblgen that can support this. 2694 // Mips O32 ABI rules: 2695 // --- 2696 // i32 - Passed in A0, A1, A2, A3 and stack 2697 // f32 - Only passed in f32 registers if no int reg has been used yet to hold 2698 // an argument. Otherwise, passed in A1, A2, A3 and stack. 2699 // f64 - Only passed in two aliased f32 registers if no int reg has been used 2700 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is 2701 // not used, it must be shadowed. If only A3 is available, shadow it and 2702 // go to stack. 2703 // vXiX - Received as scalarized i32s, passed in A0 - A3 and the stack. 2704 // vXf32 - Passed in either a pair of registers {A0, A1}, {A2, A3} or {A0 - A3} 2705 // with the remainder spilled to the stack. 2706 // vXf64 - Passed in either {A0, A1, A2, A3} or {A2, A3} and in both cases 2707 // spilling the remainder to the stack. 2708 // 2709 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack. 2710 //===----------------------------------------------------------------------===// 2711 2712 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, 2713 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, 2714 CCState &State, ArrayRef<MCPhysReg> F64Regs) { 2715 const MipsSubtarget &Subtarget = static_cast<const MipsSubtarget &>( 2716 State.getMachineFunction().getSubtarget()); 2717 2718 static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 }; 2719 2720 const MipsCCState * MipsState = static_cast<MipsCCState *>(&State); 2721 2722 static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 }; 2723 2724 static const MCPhysReg FloatVectorIntRegs[] = { Mips::A0, Mips::A2 }; 2725 2726 // Do not process byval args here. 2727 if (ArgFlags.isByVal()) 2728 return true; 2729 2730 // Promote i8 and i16 2731 if (ArgFlags.isInReg() && !Subtarget.isLittle()) { 2732 if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) { 2733 LocVT = MVT::i32; 2734 if (ArgFlags.isSExt()) 2735 LocInfo = CCValAssign::SExtUpper; 2736 else if (ArgFlags.isZExt()) 2737 LocInfo = CCValAssign::ZExtUpper; 2738 else 2739 LocInfo = CCValAssign::AExtUpper; 2740 } 2741 } 2742 2743 // Promote i8 and i16 2744 if (LocVT == MVT::i8 || LocVT == MVT::i16) { 2745 LocVT = MVT::i32; 2746 if (ArgFlags.isSExt()) 2747 LocInfo = CCValAssign::SExt; 2748 else if (ArgFlags.isZExt()) 2749 LocInfo = CCValAssign::ZExt; 2750 else 2751 LocInfo = CCValAssign::AExt; 2752 } 2753 2754 unsigned Reg; 2755 2756 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following 2757 // is true: function is vararg, argument is 3rd or higher, there is previous 2758 // argument which is not f32 or f64. 2759 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 || 2760 State.getFirstUnallocated(F32Regs) != ValNo; 2761 unsigned OrigAlign = ArgFlags.getOrigAlign(); 2762 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8); 2763 bool isVectorFloat = MipsState->WasOriginalArgVectorFloat(ValNo); 2764 2765 // The MIPS vector ABI for floats passes them in a pair of registers 2766 if (ValVT == MVT::i32 && isVectorFloat) { 2767 // This is the start of an vector that was scalarized into an unknown number 2768 // of components. It doesn't matter how many there are. Allocate one of the 2769 // notional 8 byte aligned registers which map onto the argument stack, and 2770 // shadow the register lost to alignment requirements. 2771 if (ArgFlags.isSplit()) { 2772 Reg = State.AllocateReg(FloatVectorIntRegs); 2773 if (Reg == Mips::A2) 2774 State.AllocateReg(Mips::A1); 2775 else if (Reg == 0) 2776 State.AllocateReg(Mips::A3); 2777 } else { 2778 // If we're an intermediate component of the split, we can just attempt to 2779 // allocate a register directly. 2780 Reg = State.AllocateReg(IntRegs); 2781 } 2782 } else if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) { 2783 Reg = State.AllocateReg(IntRegs); 2784 // If this is the first part of an i64 arg, 2785 // the allocated register must be either A0 or A2. 2786 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3)) 2787 Reg = State.AllocateReg(IntRegs); 2788 LocVT = MVT::i32; 2789 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) { 2790 // Allocate int register and shadow next int register. If first 2791 // available register is Mips::A1 or Mips::A3, shadow it too. 2792 Reg = State.AllocateReg(IntRegs); 2793 if (Reg == Mips::A1 || Reg == Mips::A3) 2794 Reg = State.AllocateReg(IntRegs); 2795 State.AllocateReg(IntRegs); 2796 LocVT = MVT::i32; 2797 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) { 2798 // we are guaranteed to find an available float register 2799 if (ValVT == MVT::f32) { 2800 Reg = State.AllocateReg(F32Regs); 2801 // Shadow int register 2802 State.AllocateReg(IntRegs); 2803 } else { 2804 Reg = State.AllocateReg(F64Regs); 2805 // Shadow int registers 2806 unsigned Reg2 = State.AllocateReg(IntRegs); 2807 if (Reg2 == Mips::A1 || Reg2 == Mips::A3) 2808 State.AllocateReg(IntRegs); 2809 State.AllocateReg(IntRegs); 2810 } 2811 } else 2812 llvm_unreachable("Cannot handle this ValVT."); 2813 2814 if (!Reg) { 2815 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3, 2816 OrigAlign); 2817 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); 2818 } else 2819 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 2820 2821 return false; 2822 } 2823 2824 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT, 2825 MVT LocVT, CCValAssign::LocInfo LocInfo, 2826 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2827 static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 }; 2828 2829 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); 2830 } 2831 2832 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT, 2833 MVT LocVT, CCValAssign::LocInfo LocInfo, 2834 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2835 static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 }; 2836 2837 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); 2838 } 2839 2840 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, 2841 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, 2842 CCState &State) LLVM_ATTRIBUTE_UNUSED; 2843 2844 #include "MipsGenCallingConv.inc" 2845 2846 //===----------------------------------------------------------------------===// 2847 // Call Calling Convention Implementation 2848 //===----------------------------------------------------------------------===// 2849 2850 // Return next O32 integer argument register. 2851 static unsigned getNextIntArgReg(unsigned Reg) { 2852 assert((Reg == Mips::A0) || (Reg == Mips::A2)); 2853 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3; 2854 } 2855 2856 SDValue MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset, 2857 SDValue Chain, SDValue Arg, 2858 const SDLoc &DL, bool IsTailCall, 2859 SelectionDAG &DAG) const { 2860 if (!IsTailCall) { 2861 SDValue PtrOff = 2862 DAG.getNode(ISD::ADD, DL, getPointerTy(DAG.getDataLayout()), StackPtr, 2863 DAG.getIntPtrConstant(Offset, DL)); 2864 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo()); 2865 } 2866 2867 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 2868 int FI = MFI.CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false); 2869 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); 2870 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(), 2871 /* Alignment = */ 0, MachineMemOperand::MOVolatile); 2872 } 2873 2874 void MipsTargetLowering:: 2875 getOpndList(SmallVectorImpl<SDValue> &Ops, 2876 std::deque< std::pair<unsigned, SDValue> > &RegsToPass, 2877 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage, 2878 bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee, 2879 SDValue Chain) const { 2880 // Insert node "GP copy globalreg" before call to function. 2881 // 2882 // R_MIPS_CALL* operators (emitted when non-internal functions are called 2883 // in PIC mode) allow symbols to be resolved via lazy binding. 2884 // The lazy binding stub requires GP to point to the GOT. 2885 // Note that we don't need GP to point to the GOT for indirect calls 2886 // (when R_MIPS_CALL* is not used for the call) because Mips linker generates 2887 // lazy binding stub for a function only when R_MIPS_CALL* are the only relocs 2888 // used for the function (that is, Mips linker doesn't generate lazy binding 2889 // stub for a function whose address is taken in the program). 2890 if (IsPICCall && !InternalLinkage && IsCallReloc) { 2891 unsigned GPReg = ABI.IsN64() ? Mips::GP_64 : Mips::GP; 2892 EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32; 2893 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty))); 2894 } 2895 2896 // Build a sequence of copy-to-reg nodes chained together with token 2897 // chain and flag operands which copy the outgoing args into registers. 2898 // The InFlag in necessary since all emitted instructions must be 2899 // stuck together. 2900 SDValue InFlag; 2901 2902 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 2903 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first, 2904 RegsToPass[i].second, InFlag); 2905 InFlag = Chain.getValue(1); 2906 } 2907 2908 // Add argument registers to the end of the list so that they are 2909 // known live into the call. 2910 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 2911 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first, 2912 RegsToPass[i].second.getValueType())); 2913 2914 // Add a register mask operand representing the call-preserved registers. 2915 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); 2916 const uint32_t *Mask = 2917 TRI->getCallPreservedMask(CLI.DAG.getMachineFunction(), CLI.CallConv); 2918 assert(Mask && "Missing call preserved mask for calling convention"); 2919 if (Subtarget.inMips16HardFloat()) { 2920 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) { 2921 llvm::StringRef Sym = G->getGlobal()->getName(); 2922 Function *F = G->getGlobal()->getParent()->getFunction(Sym); 2923 if (F && F->hasFnAttribute("__Mips16RetHelper")) { 2924 Mask = MipsRegisterInfo::getMips16RetHelperMask(); 2925 } 2926 } 2927 } 2928 Ops.push_back(CLI.DAG.getRegisterMask(Mask)); 2929 2930 if (InFlag.getNode()) 2931 Ops.push_back(InFlag); 2932 } 2933 2934 /// LowerCall - functions arguments are copied from virtual regs to 2935 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. 2936 SDValue 2937 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, 2938 SmallVectorImpl<SDValue> &InVals) const { 2939 SelectionDAG &DAG = CLI.DAG; 2940 SDLoc DL = CLI.DL; 2941 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; 2942 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; 2943 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; 2944 SDValue Chain = CLI.Chain; 2945 SDValue Callee = CLI.Callee; 2946 bool &IsTailCall = CLI.IsTailCall; 2947 CallingConv::ID CallConv = CLI.CallConv; 2948 bool IsVarArg = CLI.IsVarArg; 2949 2950 MachineFunction &MF = DAG.getMachineFunction(); 2951 MachineFrameInfo &MFI = MF.getFrameInfo(); 2952 const TargetFrameLowering *TFL = Subtarget.getFrameLowering(); 2953 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 2954 bool IsPIC = isPositionIndependent(); 2955 2956 // Analyze operands of the call, assigning locations to each operand. 2957 SmallVector<CCValAssign, 16> ArgLocs; 2958 MipsCCState CCInfo( 2959 CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, *DAG.getContext(), 2960 MipsCCState::getSpecialCallingConvForCallee(Callee.getNode(), Subtarget)); 2961 2962 // Allocate the reserved argument area. It seems strange to do this from the 2963 // caller side but removing it breaks the frame size calculation. 2964 CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1); 2965 2966 const ExternalSymbolSDNode *ES = 2967 dyn_cast_or_null<const ExternalSymbolSDNode>(Callee.getNode()); 2968 CCInfo.AnalyzeCallOperands(Outs, CC_Mips, CLI.getArgs(), 2969 ES ? ES->getSymbol() : nullptr); 2970 2971 // Get a count of how many bytes are to be pushed on the stack. 2972 unsigned NextStackOffset = CCInfo.getNextStackOffset(); 2973 2974 // Check if it's really possible to do a tail call. Restrict it to functions 2975 // that are part of this compilation unit. 2976 bool InternalLinkage = false; 2977 if (IsTailCall) { 2978 IsTailCall = isEligibleForTailCallOptimization( 2979 CCInfo, NextStackOffset, *MF.getInfo<MipsFunctionInfo>()); 2980 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 2981 InternalLinkage = G->getGlobal()->hasInternalLinkage(); 2982 IsTailCall &= (InternalLinkage || G->getGlobal()->hasLocalLinkage() || 2983 G->getGlobal()->hasPrivateLinkage() || 2984 G->getGlobal()->hasHiddenVisibility() || 2985 G->getGlobal()->hasProtectedVisibility()); 2986 } 2987 } 2988 if (!IsTailCall && CLI.CS && CLI.CS.isMustTailCall()) 2989 report_fatal_error("failed to perform tail call elimination on a call " 2990 "site marked musttail"); 2991 2992 if (IsTailCall) 2993 ++NumTailCalls; 2994 2995 // Chain is the output chain of the last Load/Store or CopyToReg node. 2996 // ByValChain is the output chain of the last Memcpy node created for copying 2997 // byval arguments to the stack. 2998 unsigned StackAlignment = TFL->getStackAlignment(); 2999 NextStackOffset = alignTo(NextStackOffset, StackAlignment); 3000 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, DL, true); 3001 3002 if (!IsTailCall) 3003 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffset, 0, DL); 3004 3005 SDValue StackPtr = 3006 DAG.getCopyFromReg(Chain, DL, ABI.IsN64() ? Mips::SP_64 : Mips::SP, 3007 getPointerTy(DAG.getDataLayout())); 3008 3009 std::deque< std::pair<unsigned, SDValue> > RegsToPass; 3010 SmallVector<SDValue, 8> MemOpChains; 3011 3012 CCInfo.rewindByValRegsInfo(); 3013 3014 // Walk the register/memloc assignments, inserting copies/loads. 3015 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 3016 SDValue Arg = OutVals[i]; 3017 CCValAssign &VA = ArgLocs[i]; 3018 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT(); 3019 ISD::ArgFlagsTy Flags = Outs[i].Flags; 3020 bool UseUpperBits = false; 3021 3022 // ByVal Arg. 3023 if (Flags.isByVal()) { 3024 unsigned FirstByValReg, LastByValReg; 3025 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); 3026 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg); 3027 3028 assert(Flags.getByValSize() && 3029 "ByVal args of size 0 should have been ignored by front-end."); 3030 assert(ByValIdx < CCInfo.getInRegsParamsCount()); 3031 assert(!IsTailCall && 3032 "Do not tail-call optimize if there is a byval argument."); 3033 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg, 3034 FirstByValReg, LastByValReg, Flags, Subtarget.isLittle(), 3035 VA); 3036 CCInfo.nextInRegsParam(); 3037 continue; 3038 } 3039 3040 // Promote the value if needed. 3041 switch (VA.getLocInfo()) { 3042 default: 3043 llvm_unreachable("Unknown loc info!"); 3044 case CCValAssign::Full: 3045 if (VA.isRegLoc()) { 3046 if ((ValVT == MVT::f32 && LocVT == MVT::i32) || 3047 (ValVT == MVT::f64 && LocVT == MVT::i64) || 3048 (ValVT == MVT::i64 && LocVT == MVT::f64)) 3049 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 3050 else if (ValVT == MVT::f64 && LocVT == MVT::i32) { 3051 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 3052 Arg, DAG.getConstant(0, DL, MVT::i32)); 3053 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 3054 Arg, DAG.getConstant(1, DL, MVT::i32)); 3055 if (!Subtarget.isLittle()) 3056 std::swap(Lo, Hi); 3057 unsigned LocRegLo = VA.getLocReg(); 3058 unsigned LocRegHigh = getNextIntArgReg(LocRegLo); 3059 RegsToPass.push_back(std::make_pair(LocRegLo, Lo)); 3060 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi)); 3061 continue; 3062 } 3063 } 3064 break; 3065 case CCValAssign::BCvt: 3066 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 3067 break; 3068 case CCValAssign::SExtUpper: 3069 UseUpperBits = true; 3070 LLVM_FALLTHROUGH; 3071 case CCValAssign::SExt: 3072 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg); 3073 break; 3074 case CCValAssign::ZExtUpper: 3075 UseUpperBits = true; 3076 LLVM_FALLTHROUGH; 3077 case CCValAssign::ZExt: 3078 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg); 3079 break; 3080 case CCValAssign::AExtUpper: 3081 UseUpperBits = true; 3082 LLVM_FALLTHROUGH; 3083 case CCValAssign::AExt: 3084 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg); 3085 break; 3086 } 3087 3088 if (UseUpperBits) { 3089 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits(); 3090 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3091 Arg = DAG.getNode( 3092 ISD::SHL, DL, VA.getLocVT(), Arg, 3093 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3094 } 3095 3096 // Arguments that can be passed on register must be kept at 3097 // RegsToPass vector 3098 if (VA.isRegLoc()) { 3099 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 3100 continue; 3101 } 3102 3103 // Register can't get to this point... 3104 assert(VA.isMemLoc()); 3105 3106 // emit ISD::STORE whichs stores the 3107 // parameter value to a stack Location 3108 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(), 3109 Chain, Arg, DL, IsTailCall, DAG)); 3110 } 3111 3112 // Transform all store nodes into one single node because all store 3113 // nodes are independent of each other. 3114 if (!MemOpChains.empty()) 3115 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains); 3116 3117 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 3118 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 3119 // node so that legalize doesn't hack it. 3120 3121 SDValue CalleeLo; 3122 EVT Ty = Callee.getValueType(); 3123 bool GlobalOrExternal = false, IsCallReloc = false; 3124 3125 // The long-calls feature is ignored in case of PIC. 3126 // While we do not support -mshared / -mno-shared properly, 3127 // ignore long-calls in case of -mabicalls too. 3128 if (!Subtarget.isABICalls() && !IsPIC) { 3129 // If the function should be called using "long call", 3130 // get its address into a register to prevent using 3131 // of the `jal` instruction for the direct call. 3132 if (auto *N = dyn_cast<ExternalSymbolSDNode>(Callee)) { 3133 if (Subtarget.useLongCalls()) 3134 Callee = Subtarget.hasSym32() 3135 ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 3136 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 3137 } else if (auto *N = dyn_cast<GlobalAddressSDNode>(Callee)) { 3138 bool UseLongCalls = Subtarget.useLongCalls(); 3139 // If the function has long-call/far/near attribute 3140 // it overrides command line switch pased to the backend. 3141 if (auto *F = dyn_cast<Function>(N->getGlobal())) { 3142 if (F->hasFnAttribute("long-call")) 3143 UseLongCalls = true; 3144 else if (F->hasFnAttribute("short-call")) 3145 UseLongCalls = false; 3146 } 3147 if (UseLongCalls) 3148 Callee = Subtarget.hasSym32() 3149 ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 3150 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 3151 } 3152 } 3153 3154 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 3155 if (IsPIC) { 3156 const GlobalValue *Val = G->getGlobal(); 3157 InternalLinkage = Val->hasInternalLinkage(); 3158 3159 if (InternalLinkage) 3160 Callee = getAddrLocal(G, DL, Ty, DAG, ABI.IsN32() || ABI.IsN64()); 3161 else if (LargeGOT) { 3162 Callee = getAddrGlobalLargeGOT(G, DL, Ty, DAG, MipsII::MO_CALL_HI16, 3163 MipsII::MO_CALL_LO16, Chain, 3164 FuncInfo->callPtrInfo(Val)); 3165 IsCallReloc = true; 3166 } else { 3167 Callee = getAddrGlobal(G, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain, 3168 FuncInfo->callPtrInfo(Val)); 3169 IsCallReloc = true; 3170 } 3171 } else 3172 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, 3173 getPointerTy(DAG.getDataLayout()), 0, 3174 MipsII::MO_NO_FLAG); 3175 GlobalOrExternal = true; 3176 } 3177 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 3178 const char *Sym = S->getSymbol(); 3179 3180 if (!IsPIC) // static 3181 Callee = DAG.getTargetExternalSymbol( 3182 Sym, getPointerTy(DAG.getDataLayout()), MipsII::MO_NO_FLAG); 3183 else if (LargeGOT) { 3184 Callee = getAddrGlobalLargeGOT(S, DL, Ty, DAG, MipsII::MO_CALL_HI16, 3185 MipsII::MO_CALL_LO16, Chain, 3186 FuncInfo->callPtrInfo(Sym)); 3187 IsCallReloc = true; 3188 } else { // PIC 3189 Callee = getAddrGlobal(S, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain, 3190 FuncInfo->callPtrInfo(Sym)); 3191 IsCallReloc = true; 3192 } 3193 3194 GlobalOrExternal = true; 3195 } 3196 3197 SmallVector<SDValue, 8> Ops(1, Chain); 3198 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 3199 3200 getOpndList(Ops, RegsToPass, IsPIC, GlobalOrExternal, InternalLinkage, 3201 IsCallReloc, CLI, Callee, Chain); 3202 3203 if (IsTailCall) { 3204 MF.getFrameInfo().setHasTailCall(); 3205 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, Ops); 3206 } 3207 3208 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, Ops); 3209 SDValue InFlag = Chain.getValue(1); 3210 3211 // Create the CALLSEQ_END node. 3212 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal, 3213 DAG.getIntPtrConstant(0, DL, true), InFlag, DL); 3214 InFlag = Chain.getValue(1); 3215 3216 // Handle result values, copying them out of physregs into vregs that we 3217 // return. 3218 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG, 3219 InVals, CLI); 3220 } 3221 3222 /// LowerCallResult - Lower the result values of a call into the 3223 /// appropriate copies out of appropriate physical registers. 3224 SDValue MipsTargetLowering::LowerCallResult( 3225 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg, 3226 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, 3227 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, 3228 TargetLowering::CallLoweringInfo &CLI) const { 3229 // Assign locations to each value returned by this call. 3230 SmallVector<CCValAssign, 16> RVLocs; 3231 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, 3232 *DAG.getContext()); 3233 3234 const ExternalSymbolSDNode *ES = 3235 dyn_cast_or_null<const ExternalSymbolSDNode>(CLI.Callee.getNode()); 3236 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips, CLI.RetTy, 3237 ES ? ES->getSymbol() : nullptr); 3238 3239 // Copy all of the result registers out of their specified physreg. 3240 for (unsigned i = 0; i != RVLocs.size(); ++i) { 3241 CCValAssign &VA = RVLocs[i]; 3242 assert(VA.isRegLoc() && "Can only return in registers!"); 3243 3244 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(), 3245 RVLocs[i].getLocVT(), InFlag); 3246 Chain = Val.getValue(1); 3247 InFlag = Val.getValue(2); 3248 3249 if (VA.isUpperBitsInLoc()) { 3250 unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits(); 3251 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3252 unsigned Shift = 3253 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; 3254 Val = DAG.getNode( 3255 Shift, DL, VA.getLocVT(), Val, 3256 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3257 } 3258 3259 switch (VA.getLocInfo()) { 3260 default: 3261 llvm_unreachable("Unknown loc info!"); 3262 case CCValAssign::Full: 3263 break; 3264 case CCValAssign::BCvt: 3265 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val); 3266 break; 3267 case CCValAssign::AExt: 3268 case CCValAssign::AExtUpper: 3269 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 3270 break; 3271 case CCValAssign::ZExt: 3272 case CCValAssign::ZExtUpper: 3273 Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val, 3274 DAG.getValueType(VA.getValVT())); 3275 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 3276 break; 3277 case CCValAssign::SExt: 3278 case CCValAssign::SExtUpper: 3279 Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val, 3280 DAG.getValueType(VA.getValVT())); 3281 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 3282 break; 3283 } 3284 3285 InVals.push_back(Val); 3286 } 3287 3288 return Chain; 3289 } 3290 3291 static SDValue UnpackFromArgumentSlot(SDValue Val, const CCValAssign &VA, 3292 EVT ArgVT, const SDLoc &DL, 3293 SelectionDAG &DAG) { 3294 MVT LocVT = VA.getLocVT(); 3295 EVT ValVT = VA.getValVT(); 3296 3297 // Shift into the upper bits if necessary. 3298 switch (VA.getLocInfo()) { 3299 default: 3300 break; 3301 case CCValAssign::AExtUpper: 3302 case CCValAssign::SExtUpper: 3303 case CCValAssign::ZExtUpper: { 3304 unsigned ValSizeInBits = ArgVT.getSizeInBits(); 3305 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3306 unsigned Opcode = 3307 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; 3308 Val = DAG.getNode( 3309 Opcode, DL, VA.getLocVT(), Val, 3310 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3311 break; 3312 } 3313 } 3314 3315 // If this is an value smaller than the argument slot size (32-bit for O32, 3316 // 64-bit for N32/N64), it has been promoted in some way to the argument slot 3317 // size. Extract the value and insert any appropriate assertions regarding 3318 // sign/zero extension. 3319 switch (VA.getLocInfo()) { 3320 default: 3321 llvm_unreachable("Unknown loc info!"); 3322 case CCValAssign::Full: 3323 break; 3324 case CCValAssign::AExtUpper: 3325 case CCValAssign::AExt: 3326 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 3327 break; 3328 case CCValAssign::SExtUpper: 3329 case CCValAssign::SExt: 3330 Val = DAG.getNode(ISD::AssertSext, DL, LocVT, Val, DAG.getValueType(ValVT)); 3331 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 3332 break; 3333 case CCValAssign::ZExtUpper: 3334 case CCValAssign::ZExt: 3335 Val = DAG.getNode(ISD::AssertZext, DL, LocVT, Val, DAG.getValueType(ValVT)); 3336 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 3337 break; 3338 case CCValAssign::BCvt: 3339 Val = DAG.getNode(ISD::BITCAST, DL, ValVT, Val); 3340 break; 3341 } 3342 3343 return Val; 3344 } 3345 3346 //===----------------------------------------------------------------------===// 3347 // Formal Arguments Calling Convention Implementation 3348 //===----------------------------------------------------------------------===// 3349 /// LowerFormalArguments - transform physical registers into virtual registers 3350 /// and generate load operations for arguments places on the stack. 3351 SDValue MipsTargetLowering::LowerFormalArguments( 3352 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, 3353 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, 3354 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { 3355 MachineFunction &MF = DAG.getMachineFunction(); 3356 MachineFrameInfo &MFI = MF.getFrameInfo(); 3357 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3358 3359 MipsFI->setVarArgsFrameIndex(0); 3360 3361 // Used with vargs to acumulate store chains. 3362 std::vector<SDValue> OutChains; 3363 3364 // Assign locations to all of the incoming arguments. 3365 SmallVector<CCValAssign, 16> ArgLocs; 3366 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, 3367 *DAG.getContext()); 3368 CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1); 3369 const Function *Func = DAG.getMachineFunction().getFunction(); 3370 Function::const_arg_iterator FuncArg = Func->arg_begin(); 3371 3372 if (Func->hasFnAttribute("interrupt") && !Func->arg_empty()) 3373 report_fatal_error( 3374 "Functions with the interrupt attribute cannot have arguments!"); 3375 3376 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips_FixedArg); 3377 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(), 3378 CCInfo.getInRegsParamsCount() > 0); 3379 3380 unsigned CurArgIdx = 0; 3381 CCInfo.rewindByValRegsInfo(); 3382 3383 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 3384 CCValAssign &VA = ArgLocs[i]; 3385 if (Ins[i].isOrigArg()) { 3386 std::advance(FuncArg, Ins[i].getOrigArgIndex() - CurArgIdx); 3387 CurArgIdx = Ins[i].getOrigArgIndex(); 3388 } 3389 EVT ValVT = VA.getValVT(); 3390 ISD::ArgFlagsTy Flags = Ins[i].Flags; 3391 bool IsRegLoc = VA.isRegLoc(); 3392 3393 if (Flags.isByVal()) { 3394 assert(Ins[i].isOrigArg() && "Byval arguments cannot be implicit"); 3395 unsigned FirstByValReg, LastByValReg; 3396 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); 3397 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg); 3398 3399 assert(Flags.getByValSize() && 3400 "ByVal args of size 0 should have been ignored by front-end."); 3401 assert(ByValIdx < CCInfo.getInRegsParamsCount()); 3402 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg, 3403 FirstByValReg, LastByValReg, VA, CCInfo); 3404 CCInfo.nextInRegsParam(); 3405 continue; 3406 } 3407 3408 // Arguments stored on registers 3409 if (IsRegLoc) { 3410 MVT RegVT = VA.getLocVT(); 3411 unsigned ArgReg = VA.getLocReg(); 3412 const TargetRegisterClass *RC = getRegClassFor(RegVT); 3413 3414 // Transform the arguments stored on 3415 // physical registers into virtual ones 3416 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC); 3417 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT); 3418 3419 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG); 3420 3421 // Handle floating point arguments passed in integer registers and 3422 // long double arguments passed in floating point registers. 3423 if ((RegVT == MVT::i32 && ValVT == MVT::f32) || 3424 (RegVT == MVT::i64 && ValVT == MVT::f64) || 3425 (RegVT == MVT::f64 && ValVT == MVT::i64)) 3426 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue); 3427 else if (ABI.IsO32() && RegVT == MVT::i32 && 3428 ValVT == MVT::f64) { 3429 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(), 3430 getNextIntArgReg(ArgReg), RC); 3431 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT); 3432 if (!Subtarget.isLittle()) 3433 std::swap(ArgValue, ArgValue2); 3434 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, 3435 ArgValue, ArgValue2); 3436 } 3437 3438 InVals.push_back(ArgValue); 3439 } else { // VA.isRegLoc() 3440 MVT LocVT = VA.getLocVT(); 3441 3442 if (ABI.IsO32()) { 3443 // We ought to be able to use LocVT directly but O32 sets it to i32 3444 // when allocating floating point values to integer registers. 3445 // This shouldn't influence how we load the value into registers unless 3446 // we are targeting softfloat. 3447 if (VA.getValVT().isFloatingPoint() && !Subtarget.useSoftFloat()) 3448 LocVT = VA.getValVT(); 3449 } 3450 3451 // sanity check 3452 assert(VA.isMemLoc()); 3453 3454 // The stack pointer offset is relative to the caller stack frame. 3455 int FI = MFI.CreateFixedObject(LocVT.getSizeInBits() / 8, 3456 VA.getLocMemOffset(), true); 3457 3458 // Create load nodes to retrieve arguments from the stack 3459 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); 3460 SDValue ArgValue = DAG.getLoad( 3461 LocVT, DL, Chain, FIN, 3462 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)); 3463 OutChains.push_back(ArgValue.getValue(1)); 3464 3465 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG); 3466 3467 InVals.push_back(ArgValue); 3468 } 3469 } 3470 3471 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 3472 // The mips ABIs for returning structs by value requires that we copy 3473 // the sret argument into $v0 for the return. Save the argument into 3474 // a virtual register so that we can access it from the return points. 3475 if (Ins[i].Flags.isSRet()) { 3476 unsigned Reg = MipsFI->getSRetReturnReg(); 3477 if (!Reg) { 3478 Reg = MF.getRegInfo().createVirtualRegister( 3479 getRegClassFor(ABI.IsN64() ? MVT::i64 : MVT::i32)); 3480 MipsFI->setSRetReturnReg(Reg); 3481 } 3482 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[i]); 3483 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain); 3484 break; 3485 } 3486 } 3487 3488 if (IsVarArg) 3489 writeVarArgRegs(OutChains, Chain, DL, DAG, CCInfo); 3490 3491 // All stores are grouped in one node to allow the matching between 3492 // the size of Ins and InVals. This only happens when on varg functions 3493 if (!OutChains.empty()) { 3494 OutChains.push_back(Chain); 3495 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains); 3496 } 3497 3498 return Chain; 3499 } 3500 3501 //===----------------------------------------------------------------------===// 3502 // Return Value Calling Convention Implementation 3503 //===----------------------------------------------------------------------===// 3504 3505 bool 3506 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv, 3507 MachineFunction &MF, bool IsVarArg, 3508 const SmallVectorImpl<ISD::OutputArg> &Outs, 3509 LLVMContext &Context) const { 3510 SmallVector<CCValAssign, 16> RVLocs; 3511 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context); 3512 return CCInfo.CheckReturn(Outs, RetCC_Mips); 3513 } 3514 3515 bool 3516 MipsTargetLowering::shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const { 3517 if (Subtarget.hasMips3() && Subtarget.useSoftFloat()) { 3518 if (Type == MVT::i32) 3519 return true; 3520 } 3521 return IsSigned; 3522 } 3523 3524 SDValue 3525 MipsTargetLowering::LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps, 3526 const SDLoc &DL, 3527 SelectionDAG &DAG) const { 3528 3529 MachineFunction &MF = DAG.getMachineFunction(); 3530 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3531 3532 MipsFI->setISR(); 3533 3534 return DAG.getNode(MipsISD::ERet, DL, MVT::Other, RetOps); 3535 } 3536 3537 SDValue 3538 MipsTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, 3539 bool IsVarArg, 3540 const SmallVectorImpl<ISD::OutputArg> &Outs, 3541 const SmallVectorImpl<SDValue> &OutVals, 3542 const SDLoc &DL, SelectionDAG &DAG) const { 3543 // CCValAssign - represent the assignment of 3544 // the return value to a location 3545 SmallVector<CCValAssign, 16> RVLocs; 3546 MachineFunction &MF = DAG.getMachineFunction(); 3547 3548 // CCState - Info about the registers and stack slot. 3549 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext()); 3550 3551 // Analyze return values. 3552 CCInfo.AnalyzeReturn(Outs, RetCC_Mips); 3553 3554 SDValue Flag; 3555 SmallVector<SDValue, 4> RetOps(1, Chain); 3556 3557 // Copy the result values into the output registers. 3558 for (unsigned i = 0; i != RVLocs.size(); ++i) { 3559 SDValue Val = OutVals[i]; 3560 CCValAssign &VA = RVLocs[i]; 3561 assert(VA.isRegLoc() && "Can only return in registers!"); 3562 bool UseUpperBits = false; 3563 3564 switch (VA.getLocInfo()) { 3565 default: 3566 llvm_unreachable("Unknown loc info!"); 3567 case CCValAssign::Full: 3568 break; 3569 case CCValAssign::BCvt: 3570 Val = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Val); 3571 break; 3572 case CCValAssign::AExtUpper: 3573 UseUpperBits = true; 3574 LLVM_FALLTHROUGH; 3575 case CCValAssign::AExt: 3576 Val = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Val); 3577 break; 3578 case CCValAssign::ZExtUpper: 3579 UseUpperBits = true; 3580 LLVM_FALLTHROUGH; 3581 case CCValAssign::ZExt: 3582 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Val); 3583 break; 3584 case CCValAssign::SExtUpper: 3585 UseUpperBits = true; 3586 LLVM_FALLTHROUGH; 3587 case CCValAssign::SExt: 3588 Val = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Val); 3589 break; 3590 } 3591 3592 if (UseUpperBits) { 3593 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits(); 3594 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3595 Val = DAG.getNode( 3596 ISD::SHL, DL, VA.getLocVT(), Val, 3597 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3598 } 3599 3600 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag); 3601 3602 // Guarantee that all emitted copies are stuck together with flags. 3603 Flag = Chain.getValue(1); 3604 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 3605 } 3606 3607 // The mips ABIs for returning structs by value requires that we copy 3608 // the sret argument into $v0 for the return. We saved the argument into 3609 // a virtual register in the entry block, so now we copy the value out 3610 // and into $v0. 3611 if (MF.getFunction()->hasStructRetAttr()) { 3612 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3613 unsigned Reg = MipsFI->getSRetReturnReg(); 3614 3615 if (!Reg) 3616 llvm_unreachable("sret virtual register not created in the entry block"); 3617 SDValue Val = 3618 DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(DAG.getDataLayout())); 3619 unsigned V0 = ABI.IsN64() ? Mips::V0_64 : Mips::V0; 3620 3621 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag); 3622 Flag = Chain.getValue(1); 3623 RetOps.push_back(DAG.getRegister(V0, getPointerTy(DAG.getDataLayout()))); 3624 } 3625 3626 RetOps[0] = Chain; // Update chain. 3627 3628 // Add the flag if we have it. 3629 if (Flag.getNode()) 3630 RetOps.push_back(Flag); 3631 3632 // ISRs must use "eret". 3633 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt")) 3634 return LowerInterruptReturn(RetOps, DL, DAG); 3635 3636 // Standard return on Mips is a "jr $ra" 3637 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, RetOps); 3638 } 3639 3640 //===----------------------------------------------------------------------===// 3641 // Mips Inline Assembly Support 3642 //===----------------------------------------------------------------------===// 3643 3644 /// getConstraintType - Given a constraint letter, return the type of 3645 /// constraint it is for this target. 3646 MipsTargetLowering::ConstraintType 3647 MipsTargetLowering::getConstraintType(StringRef Constraint) const { 3648 // Mips specific constraints 3649 // GCC config/mips/constraints.md 3650 // 3651 // 'd' : An address register. Equivalent to r 3652 // unless generating MIPS16 code. 3653 // 'y' : Equivalent to r; retained for 3654 // backwards compatibility. 3655 // 'c' : A register suitable for use in an indirect 3656 // jump. This will always be $25 for -mabicalls. 3657 // 'l' : The lo register. 1 word storage. 3658 // 'x' : The hilo register pair. Double word storage. 3659 if (Constraint.size() == 1) { 3660 switch (Constraint[0]) { 3661 default : break; 3662 case 'd': 3663 case 'y': 3664 case 'f': 3665 case 'c': 3666 case 'l': 3667 case 'x': 3668 return C_RegisterClass; 3669 case 'R': 3670 return C_Memory; 3671 } 3672 } 3673 3674 if (Constraint == "ZC") 3675 return C_Memory; 3676 3677 return TargetLowering::getConstraintType(Constraint); 3678 } 3679 3680 /// Examine constraint type and operand type and determine a weight value. 3681 /// This object must already have been set up with the operand type 3682 /// and the current alternative constraint selected. 3683 TargetLowering::ConstraintWeight 3684 MipsTargetLowering::getSingleConstraintMatchWeight( 3685 AsmOperandInfo &info, const char *constraint) const { 3686 ConstraintWeight weight = CW_Invalid; 3687 Value *CallOperandVal = info.CallOperandVal; 3688 // If we don't have a value, we can't do a match, 3689 // but allow it at the lowest weight. 3690 if (!CallOperandVal) 3691 return CW_Default; 3692 Type *type = CallOperandVal->getType(); 3693 // Look at the constraint type. 3694 switch (*constraint) { 3695 default: 3696 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 3697 break; 3698 case 'd': 3699 case 'y': 3700 if (type->isIntegerTy()) 3701 weight = CW_Register; 3702 break; 3703 case 'f': // FPU or MSA register 3704 if (Subtarget.hasMSA() && type->isVectorTy() && 3705 cast<VectorType>(type)->getBitWidth() == 128) 3706 weight = CW_Register; 3707 else if (type->isFloatTy()) 3708 weight = CW_Register; 3709 break; 3710 case 'c': // $25 for indirect jumps 3711 case 'l': // lo register 3712 case 'x': // hilo register pair 3713 if (type->isIntegerTy()) 3714 weight = CW_SpecificReg; 3715 break; 3716 case 'I': // signed 16 bit immediate 3717 case 'J': // integer zero 3718 case 'K': // unsigned 16 bit immediate 3719 case 'L': // signed 32 bit immediate where lower 16 bits are 0 3720 case 'N': // immediate in the range of -65535 to -1 (inclusive) 3721 case 'O': // signed 15 bit immediate (+- 16383) 3722 case 'P': // immediate in the range of 65535 to 1 (inclusive) 3723 if (isa<ConstantInt>(CallOperandVal)) 3724 weight = CW_Constant; 3725 break; 3726 case 'R': 3727 weight = CW_Memory; 3728 break; 3729 } 3730 return weight; 3731 } 3732 3733 /// This is a helper function to parse a physical register string and split it 3734 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag 3735 /// that is returned indicates whether parsing was successful. The second flag 3736 /// is true if the numeric part exists. 3737 static std::pair<bool, bool> parsePhysicalReg(StringRef C, StringRef &Prefix, 3738 unsigned long long &Reg) { 3739 if (C.front() != '{' || C.back() != '}') 3740 return std::make_pair(false, false); 3741 3742 // Search for the first numeric character. 3743 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1; 3744 I = std::find_if(B, E, isdigit); 3745 3746 Prefix = StringRef(B, I - B); 3747 3748 // The second flag is set to false if no numeric characters were found. 3749 if (I == E) 3750 return std::make_pair(true, false); 3751 3752 // Parse the numeric characters. 3753 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg), 3754 true); 3755 } 3756 3757 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering:: 3758 parseRegForInlineAsmConstraint(StringRef C, MVT VT) const { 3759 const TargetRegisterInfo *TRI = 3760 Subtarget.getRegisterInfo(); 3761 const TargetRegisterClass *RC; 3762 StringRef Prefix; 3763 unsigned long long Reg; 3764 3765 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg); 3766 3767 if (!R.first) 3768 return std::make_pair(0U, nullptr); 3769 3770 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo. 3771 // No numeric characters follow "hi" or "lo". 3772 if (R.second) 3773 return std::make_pair(0U, nullptr); 3774 3775 RC = TRI->getRegClass(Prefix == "hi" ? 3776 Mips::HI32RegClassID : Mips::LO32RegClassID); 3777 return std::make_pair(*(RC->begin()), RC); 3778 } else if (Prefix.startswith("$msa")) { 3779 // Parse $msa(ir|csr|access|save|modify|request|map|unmap) 3780 3781 // No numeric characters follow the name. 3782 if (R.second) 3783 return std::make_pair(0U, nullptr); 3784 3785 Reg = StringSwitch<unsigned long long>(Prefix) 3786 .Case("$msair", Mips::MSAIR) 3787 .Case("$msacsr", Mips::MSACSR) 3788 .Case("$msaaccess", Mips::MSAAccess) 3789 .Case("$msasave", Mips::MSASave) 3790 .Case("$msamodify", Mips::MSAModify) 3791 .Case("$msarequest", Mips::MSARequest) 3792 .Case("$msamap", Mips::MSAMap) 3793 .Case("$msaunmap", Mips::MSAUnmap) 3794 .Default(0); 3795 3796 if (!Reg) 3797 return std::make_pair(0U, nullptr); 3798 3799 RC = TRI->getRegClass(Mips::MSACtrlRegClassID); 3800 return std::make_pair(Reg, RC); 3801 } 3802 3803 if (!R.second) 3804 return std::make_pair(0U, nullptr); 3805 3806 if (Prefix == "$f") { // Parse $f0-$f31. 3807 // If the size of FP registers is 64-bit or Reg is an even number, select 3808 // the 64-bit register class. Otherwise, select the 32-bit register class. 3809 if (VT == MVT::Other) 3810 VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32; 3811 3812 RC = getRegClassFor(VT); 3813 3814 if (RC == &Mips::AFGR64RegClass) { 3815 assert(Reg % 2 == 0); 3816 Reg >>= 1; 3817 } 3818 } else if (Prefix == "$fcc") // Parse $fcc0-$fcc7. 3819 RC = TRI->getRegClass(Mips::FCCRegClassID); 3820 else if (Prefix == "$w") { // Parse $w0-$w31. 3821 RC = getRegClassFor((VT == MVT::Other) ? MVT::v16i8 : VT); 3822 } else { // Parse $0-$31. 3823 assert(Prefix == "$"); 3824 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT); 3825 } 3826 3827 assert(Reg < RC->getNumRegs()); 3828 return std::make_pair(*(RC->begin() + Reg), RC); 3829 } 3830 3831 /// Given a register class constraint, like 'r', if this corresponds directly 3832 /// to an LLVM register class, return a register of 0 and the register class 3833 /// pointer. 3834 std::pair<unsigned, const TargetRegisterClass *> 3835 MipsTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, 3836 StringRef Constraint, 3837 MVT VT) const { 3838 if (Constraint.size() == 1) { 3839 switch (Constraint[0]) { 3840 case 'd': // Address register. Same as 'r' unless generating MIPS16 code. 3841 case 'y': // Same as 'r'. Exists for compatibility. 3842 case 'r': 3843 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) { 3844 if (Subtarget.inMips16Mode()) 3845 return std::make_pair(0U, &Mips::CPU16RegsRegClass); 3846 return std::make_pair(0U, &Mips::GPR32RegClass); 3847 } 3848 if (VT == MVT::i64 && !Subtarget.isGP64bit()) 3849 return std::make_pair(0U, &Mips::GPR32RegClass); 3850 if (VT == MVT::i64 && Subtarget.isGP64bit()) 3851 return std::make_pair(0U, &Mips::GPR64RegClass); 3852 // This will generate an error message 3853 return std::make_pair(0U, nullptr); 3854 case 'f': // FPU or MSA register 3855 if (VT == MVT::v16i8) 3856 return std::make_pair(0U, &Mips::MSA128BRegClass); 3857 else if (VT == MVT::v8i16 || VT == MVT::v8f16) 3858 return std::make_pair(0U, &Mips::MSA128HRegClass); 3859 else if (VT == MVT::v4i32 || VT == MVT::v4f32) 3860 return std::make_pair(0U, &Mips::MSA128WRegClass); 3861 else if (VT == MVT::v2i64 || VT == MVT::v2f64) 3862 return std::make_pair(0U, &Mips::MSA128DRegClass); 3863 else if (VT == MVT::f32) 3864 return std::make_pair(0U, &Mips::FGR32RegClass); 3865 else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) { 3866 if (Subtarget.isFP64bit()) 3867 return std::make_pair(0U, &Mips::FGR64RegClass); 3868 return std::make_pair(0U, &Mips::AFGR64RegClass); 3869 } 3870 break; 3871 case 'c': // register suitable for indirect jump 3872 if (VT == MVT::i32) 3873 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass); 3874 assert(VT == MVT::i64 && "Unexpected type."); 3875 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass); 3876 case 'l': // register suitable for indirect jump 3877 if (VT == MVT::i32) 3878 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass); 3879 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass); 3880 case 'x': // register suitable for indirect jump 3881 // Fixme: Not triggering the use of both hi and low 3882 // This will generate an error message 3883 return std::make_pair(0U, nullptr); 3884 } 3885 } 3886 3887 std::pair<unsigned, const TargetRegisterClass *> R; 3888 R = parseRegForInlineAsmConstraint(Constraint, VT); 3889 3890 if (R.second) 3891 return R; 3892 3893 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); 3894 } 3895 3896 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 3897 /// vector. If it is invalid, don't add anything to Ops. 3898 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 3899 std::string &Constraint, 3900 std::vector<SDValue>&Ops, 3901 SelectionDAG &DAG) const { 3902 SDLoc DL(Op); 3903 SDValue Result; 3904 3905 // Only support length 1 constraints for now. 3906 if (Constraint.length() > 1) return; 3907 3908 char ConstraintLetter = Constraint[0]; 3909 switch (ConstraintLetter) { 3910 default: break; // This will fall through to the generic implementation 3911 case 'I': // Signed 16 bit constant 3912 // If this fails, the parent routine will give an error 3913 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3914 EVT Type = Op.getValueType(); 3915 int64_t Val = C->getSExtValue(); 3916 if (isInt<16>(Val)) { 3917 Result = DAG.getTargetConstant(Val, DL, Type); 3918 break; 3919 } 3920 } 3921 return; 3922 case 'J': // integer zero 3923 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3924 EVT Type = Op.getValueType(); 3925 int64_t Val = C->getZExtValue(); 3926 if (Val == 0) { 3927 Result = DAG.getTargetConstant(0, DL, Type); 3928 break; 3929 } 3930 } 3931 return; 3932 case 'K': // unsigned 16 bit immediate 3933 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3934 EVT Type = Op.getValueType(); 3935 uint64_t Val = (uint64_t)C->getZExtValue(); 3936 if (isUInt<16>(Val)) { 3937 Result = DAG.getTargetConstant(Val, DL, Type); 3938 break; 3939 } 3940 } 3941 return; 3942 case 'L': // signed 32 bit immediate where lower 16 bits are 0 3943 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3944 EVT Type = Op.getValueType(); 3945 int64_t Val = C->getSExtValue(); 3946 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){ 3947 Result = DAG.getTargetConstant(Val, DL, Type); 3948 break; 3949 } 3950 } 3951 return; 3952 case 'N': // immediate in the range of -65535 to -1 (inclusive) 3953 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3954 EVT Type = Op.getValueType(); 3955 int64_t Val = C->getSExtValue(); 3956 if ((Val >= -65535) && (Val <= -1)) { 3957 Result = DAG.getTargetConstant(Val, DL, Type); 3958 break; 3959 } 3960 } 3961 return; 3962 case 'O': // signed 15 bit immediate 3963 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3964 EVT Type = Op.getValueType(); 3965 int64_t Val = C->getSExtValue(); 3966 if ((isInt<15>(Val))) { 3967 Result = DAG.getTargetConstant(Val, DL, Type); 3968 break; 3969 } 3970 } 3971 return; 3972 case 'P': // immediate in the range of 1 to 65535 (inclusive) 3973 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3974 EVT Type = Op.getValueType(); 3975 int64_t Val = C->getSExtValue(); 3976 if ((Val <= 65535) && (Val >= 1)) { 3977 Result = DAG.getTargetConstant(Val, DL, Type); 3978 break; 3979 } 3980 } 3981 return; 3982 } 3983 3984 if (Result.getNode()) { 3985 Ops.push_back(Result); 3986 return; 3987 } 3988 3989 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 3990 } 3991 3992 bool MipsTargetLowering::isLegalAddressingMode(const DataLayout &DL, 3993 const AddrMode &AM, Type *Ty, 3994 unsigned AS, Instruction *I) const { 3995 // No global is ever allowed as a base. 3996 if (AM.BaseGV) 3997 return false; 3998 3999 switch (AM.Scale) { 4000 case 0: // "r+i" or just "i", depending on HasBaseReg. 4001 break; 4002 case 1: 4003 if (!AM.HasBaseReg) // allow "r+i". 4004 break; 4005 return false; // disallow "r+r" or "r+r+i". 4006 default: 4007 return false; 4008 } 4009 4010 return true; 4011 } 4012 4013 bool 4014 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 4015 // The Mips target isn't yet aware of offsets. 4016 return false; 4017 } 4018 4019 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign, 4020 unsigned SrcAlign, 4021 bool IsMemset, bool ZeroMemset, 4022 bool MemcpyStrSrc, 4023 MachineFunction &MF) const { 4024 if (Subtarget.hasMips64()) 4025 return MVT::i64; 4026 4027 return MVT::i32; 4028 } 4029 4030 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { 4031 if (VT != MVT::f32 && VT != MVT::f64) 4032 return false; 4033 if (Imm.isNegZero()) 4034 return false; 4035 return Imm.isZero(); 4036 } 4037 4038 unsigned MipsTargetLowering::getJumpTableEncoding() const { 4039 4040 // FIXME: For space reasons this should be: EK_GPRel32BlockAddress. 4041 if (ABI.IsN64() && isPositionIndependent()) 4042 return MachineJumpTableInfo::EK_GPRel64BlockAddress; 4043 4044 return TargetLowering::getJumpTableEncoding(); 4045 } 4046 4047 bool MipsTargetLowering::useSoftFloat() const { 4048 return Subtarget.useSoftFloat(); 4049 } 4050 4051 void MipsTargetLowering::copyByValRegs( 4052 SDValue Chain, const SDLoc &DL, std::vector<SDValue> &OutChains, 4053 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags, 4054 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg, 4055 unsigned FirstReg, unsigned LastReg, const CCValAssign &VA, 4056 MipsCCState &State) const { 4057 MachineFunction &MF = DAG.getMachineFunction(); 4058 MachineFrameInfo &MFI = MF.getFrameInfo(); 4059 unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes(); 4060 unsigned NumRegs = LastReg - FirstReg; 4061 unsigned RegAreaSize = NumRegs * GPRSizeInBytes; 4062 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize); 4063 int FrameObjOffset; 4064 ArrayRef<MCPhysReg> ByValArgRegs = ABI.GetByValArgRegs(); 4065 4066 if (RegAreaSize) 4067 FrameObjOffset = 4068 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) - 4069 (int)((ByValArgRegs.size() - FirstReg) * GPRSizeInBytes); 4070 else 4071 FrameObjOffset = VA.getLocMemOffset(); 4072 4073 // Create frame object. 4074 EVT PtrTy = getPointerTy(DAG.getDataLayout()); 4075 int FI = MFI.CreateFixedObject(FrameObjSize, FrameObjOffset, true); 4076 SDValue FIN = DAG.getFrameIndex(FI, PtrTy); 4077 InVals.push_back(FIN); 4078 4079 if (!NumRegs) 4080 return; 4081 4082 // Copy arg registers. 4083 MVT RegTy = MVT::getIntegerVT(GPRSizeInBytes * 8); 4084 const TargetRegisterClass *RC = getRegClassFor(RegTy); 4085 4086 for (unsigned I = 0; I < NumRegs; ++I) { 4087 unsigned ArgReg = ByValArgRegs[FirstReg + I]; 4088 unsigned VReg = addLiveIn(MF, ArgReg, RC); 4089 unsigned Offset = I * GPRSizeInBytes; 4090 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN, 4091 DAG.getConstant(Offset, DL, PtrTy)); 4092 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy), 4093 StorePtr, MachinePointerInfo(FuncArg, Offset)); 4094 OutChains.push_back(Store); 4095 } 4096 } 4097 4098 // Copy byVal arg to registers and stack. 4099 void MipsTargetLowering::passByValArg( 4100 SDValue Chain, const SDLoc &DL, 4101 std::deque<std::pair<unsigned, SDValue>> &RegsToPass, 4102 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr, 4103 MachineFrameInfo &MFI, SelectionDAG &DAG, SDValue Arg, unsigned FirstReg, 4104 unsigned LastReg, const ISD::ArgFlagsTy &Flags, bool isLittle, 4105 const CCValAssign &VA) const { 4106 unsigned ByValSizeInBytes = Flags.getByValSize(); 4107 unsigned OffsetInBytes = 0; // From beginning of struct 4108 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 4109 unsigned Alignment = std::min(Flags.getByValAlign(), RegSizeInBytes); 4110 EVT PtrTy = getPointerTy(DAG.getDataLayout()), 4111 RegTy = MVT::getIntegerVT(RegSizeInBytes * 8); 4112 unsigned NumRegs = LastReg - FirstReg; 4113 4114 if (NumRegs) { 4115 ArrayRef<MCPhysReg> ArgRegs = ABI.GetByValArgRegs(); 4116 bool LeftoverBytes = (NumRegs * RegSizeInBytes > ByValSizeInBytes); 4117 unsigned I = 0; 4118 4119 // Copy words to registers. 4120 for (; I < NumRegs - LeftoverBytes; ++I, OffsetInBytes += RegSizeInBytes) { 4121 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 4122 DAG.getConstant(OffsetInBytes, DL, PtrTy)); 4123 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr, 4124 MachinePointerInfo(), Alignment); 4125 MemOpChains.push_back(LoadVal.getValue(1)); 4126 unsigned ArgReg = ArgRegs[FirstReg + I]; 4127 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal)); 4128 } 4129 4130 // Return if the struct has been fully copied. 4131 if (ByValSizeInBytes == OffsetInBytes) 4132 return; 4133 4134 // Copy the remainder of the byval argument with sub-word loads and shifts. 4135 if (LeftoverBytes) { 4136 SDValue Val; 4137 4138 for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0; 4139 OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) { 4140 unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes; 4141 4142 if (RemainingSizeInBytes < LoadSizeInBytes) 4143 continue; 4144 4145 // Load subword. 4146 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 4147 DAG.getConstant(OffsetInBytes, DL, 4148 PtrTy)); 4149 SDValue LoadVal = DAG.getExtLoad( 4150 ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, MachinePointerInfo(), 4151 MVT::getIntegerVT(LoadSizeInBytes * 8), Alignment); 4152 MemOpChains.push_back(LoadVal.getValue(1)); 4153 4154 // Shift the loaded value. 4155 unsigned Shamt; 4156 4157 if (isLittle) 4158 Shamt = TotalBytesLoaded * 8; 4159 else 4160 Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8; 4161 4162 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal, 4163 DAG.getConstant(Shamt, DL, MVT::i32)); 4164 4165 if (Val.getNode()) 4166 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift); 4167 else 4168 Val = Shift; 4169 4170 OffsetInBytes += LoadSizeInBytes; 4171 TotalBytesLoaded += LoadSizeInBytes; 4172 Alignment = std::min(Alignment, LoadSizeInBytes); 4173 } 4174 4175 unsigned ArgReg = ArgRegs[FirstReg + I]; 4176 RegsToPass.push_back(std::make_pair(ArgReg, Val)); 4177 return; 4178 } 4179 } 4180 4181 // Copy remainder of byval arg to it with memcpy. 4182 unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes; 4183 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 4184 DAG.getConstant(OffsetInBytes, DL, PtrTy)); 4185 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr, 4186 DAG.getIntPtrConstant(VA.getLocMemOffset(), DL)); 4187 Chain = DAG.getMemcpy(Chain, DL, Dst, Src, 4188 DAG.getConstant(MemCpySize, DL, PtrTy), 4189 Alignment, /*isVolatile=*/false, /*AlwaysInline=*/false, 4190 /*isTailCall=*/false, 4191 MachinePointerInfo(), MachinePointerInfo()); 4192 MemOpChains.push_back(Chain); 4193 } 4194 4195 void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains, 4196 SDValue Chain, const SDLoc &DL, 4197 SelectionDAG &DAG, 4198 CCState &State) const { 4199 ArrayRef<MCPhysReg> ArgRegs = ABI.GetVarArgRegs(); 4200 unsigned Idx = State.getFirstUnallocated(ArgRegs); 4201 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 4202 MVT RegTy = MVT::getIntegerVT(RegSizeInBytes * 8); 4203 const TargetRegisterClass *RC = getRegClassFor(RegTy); 4204 MachineFunction &MF = DAG.getMachineFunction(); 4205 MachineFrameInfo &MFI = MF.getFrameInfo(); 4206 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 4207 4208 // Offset of the first variable argument from stack pointer. 4209 int VaArgOffset; 4210 4211 if (ArgRegs.size() == Idx) 4212 VaArgOffset = alignTo(State.getNextStackOffset(), RegSizeInBytes); 4213 else { 4214 VaArgOffset = 4215 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) - 4216 (int)(RegSizeInBytes * (ArgRegs.size() - Idx)); 4217 } 4218 4219 // Record the frame index of the first variable argument 4220 // which is a value necessary to VASTART. 4221 int FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true); 4222 MipsFI->setVarArgsFrameIndex(FI); 4223 4224 // Copy the integer registers that have not been used for argument passing 4225 // to the argument register save area. For O32, the save area is allocated 4226 // in the caller's stack frame, while for N32/64, it is allocated in the 4227 // callee's stack frame. 4228 for (unsigned I = Idx; I < ArgRegs.size(); 4229 ++I, VaArgOffset += RegSizeInBytes) { 4230 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC); 4231 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy); 4232 FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true); 4233 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); 4234 SDValue Store = 4235 DAG.getStore(Chain, DL, ArgValue, PtrOff, MachinePointerInfo()); 4236 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue( 4237 (Value *)nullptr); 4238 OutChains.push_back(Store); 4239 } 4240 } 4241 4242 void MipsTargetLowering::HandleByVal(CCState *State, unsigned &Size, 4243 unsigned Align) const { 4244 const TargetFrameLowering *TFL = Subtarget.getFrameLowering(); 4245 4246 assert(Size && "Byval argument's size shouldn't be 0."); 4247 4248 Align = std::min(Align, TFL->getStackAlignment()); 4249 4250 unsigned FirstReg = 0; 4251 unsigned NumRegs = 0; 4252 4253 if (State->getCallingConv() != CallingConv::Fast) { 4254 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 4255 ArrayRef<MCPhysReg> IntArgRegs = ABI.GetByValArgRegs(); 4256 // FIXME: The O32 case actually describes no shadow registers. 4257 const MCPhysReg *ShadowRegs = 4258 ABI.IsO32() ? IntArgRegs.data() : Mips64DPRegs; 4259 4260 // We used to check the size as well but we can't do that anymore since 4261 // CCState::HandleByVal() rounds up the size after calling this function. 4262 assert(!(Align % RegSizeInBytes) && 4263 "Byval argument's alignment should be a multiple of" 4264 "RegSizeInBytes."); 4265 4266 FirstReg = State->getFirstUnallocated(IntArgRegs); 4267 4268 // If Align > RegSizeInBytes, the first arg register must be even. 4269 // FIXME: This condition happens to do the right thing but it's not the 4270 // right way to test it. We want to check that the stack frame offset 4271 // of the register is aligned. 4272 if ((Align > RegSizeInBytes) && (FirstReg % 2)) { 4273 State->AllocateReg(IntArgRegs[FirstReg], ShadowRegs[FirstReg]); 4274 ++FirstReg; 4275 } 4276 4277 // Mark the registers allocated. 4278 Size = alignTo(Size, RegSizeInBytes); 4279 for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size()); 4280 Size -= RegSizeInBytes, ++I, ++NumRegs) 4281 State->AllocateReg(IntArgRegs[I], ShadowRegs[I]); 4282 } 4283 4284 State->addInRegsParamInfo(FirstReg, FirstReg + NumRegs); 4285 } 4286 4287 MachineBasicBlock *MipsTargetLowering::emitPseudoSELECT(MachineInstr &MI, 4288 MachineBasicBlock *BB, 4289 bool isFPCmp, 4290 unsigned Opc) const { 4291 assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) && 4292 "Subtarget already supports SELECT nodes with the use of" 4293 "conditional-move instructions."); 4294 4295 const TargetInstrInfo *TII = 4296 Subtarget.getInstrInfo(); 4297 DebugLoc DL = MI.getDebugLoc(); 4298 4299 // To "insert" a SELECT instruction, we actually have to insert the 4300 // diamond control-flow pattern. The incoming instruction knows the 4301 // destination vreg to set, the condition code register to branch on, the 4302 // true/false values to select between, and a branch opcode to use. 4303 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 4304 MachineFunction::iterator It = ++BB->getIterator(); 4305 4306 // thisMBB: 4307 // ... 4308 // TrueVal = ... 4309 // setcc r1, r2, r3 4310 // bNE r1, r0, copy1MBB 4311 // fallthrough --> copy0MBB 4312 MachineBasicBlock *thisMBB = BB; 4313 MachineFunction *F = BB->getParent(); 4314 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 4315 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 4316 F->insert(It, copy0MBB); 4317 F->insert(It, sinkMBB); 4318 4319 // Transfer the remainder of BB and its successor edges to sinkMBB. 4320 sinkMBB->splice(sinkMBB->begin(), BB, 4321 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 4322 sinkMBB->transferSuccessorsAndUpdatePHIs(BB); 4323 4324 // Next, add the true and fallthrough blocks as its successors. 4325 BB->addSuccessor(copy0MBB); 4326 BB->addSuccessor(sinkMBB); 4327 4328 if (isFPCmp) { 4329 // bc1[tf] cc, sinkMBB 4330 BuildMI(BB, DL, TII->get(Opc)) 4331 .addReg(MI.getOperand(1).getReg()) 4332 .addMBB(sinkMBB); 4333 } else { 4334 // bne rs, $0, sinkMBB 4335 BuildMI(BB, DL, TII->get(Opc)) 4336 .addReg(MI.getOperand(1).getReg()) 4337 .addReg(Mips::ZERO) 4338 .addMBB(sinkMBB); 4339 } 4340 4341 // copy0MBB: 4342 // %FalseValue = ... 4343 // # fallthrough to sinkMBB 4344 BB = copy0MBB; 4345 4346 // Update machine-CFG edges 4347 BB->addSuccessor(sinkMBB); 4348 4349 // sinkMBB: 4350 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] 4351 // ... 4352 BB = sinkMBB; 4353 4354 BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(0).getReg()) 4355 .addReg(MI.getOperand(2).getReg()) 4356 .addMBB(thisMBB) 4357 .addReg(MI.getOperand(3).getReg()) 4358 .addMBB(copy0MBB); 4359 4360 MI.eraseFromParent(); // The pseudo instruction is gone now. 4361 4362 return BB; 4363 } 4364 4365 // FIXME? Maybe this could be a TableGen attribute on some registers and 4366 // this table could be generated automatically from RegInfo. 4367 unsigned MipsTargetLowering::getRegisterByName(const char* RegName, EVT VT, 4368 SelectionDAG &DAG) const { 4369 // Named registers is expected to be fairly rare. For now, just support $28 4370 // since the linux kernel uses it. 4371 if (Subtarget.isGP64bit()) { 4372 unsigned Reg = StringSwitch<unsigned>(RegName) 4373 .Case("$28", Mips::GP_64) 4374 .Default(0); 4375 if (Reg) 4376 return Reg; 4377 } else { 4378 unsigned Reg = StringSwitch<unsigned>(RegName) 4379 .Case("$28", Mips::GP) 4380 .Default(0); 4381 if (Reg) 4382 return Reg; 4383 } 4384 report_fatal_error("Invalid register name global variable"); 4385 } 4386