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 15 #include "MipsISelLowering.h" 16 #include "InstPrinter/MipsInstPrinter.h" 17 #include "MCTargetDesc/MipsBaseInfo.h" 18 #include "MCTargetDesc/MipsMCTargetDesc.h" 19 #include "MipsCCState.h" 20 #include "MipsInstrInfo.h" 21 #include "MipsMachineFunction.h" 22 #include "MipsRegisterInfo.h" 23 #include "MipsSubtarget.h" 24 #include "MipsTargetMachine.h" 25 #include "MipsTargetObjectFile.h" 26 #include "llvm/ADT/APFloat.h" 27 #include "llvm/ADT/ArrayRef.h" 28 #include "llvm/ADT/SmallVector.h" 29 #include "llvm/ADT/Statistic.h" 30 #include "llvm/ADT/StringRef.h" 31 #include "llvm/ADT/StringSwitch.h" 32 #include "llvm/CodeGen/CallingConvLower.h" 33 #include "llvm/CodeGen/FunctionLoweringInfo.h" 34 #include "llvm/CodeGen/ISDOpcodes.h" 35 #include "llvm/CodeGen/MachineBasicBlock.h" 36 #include "llvm/CodeGen/MachineFrameInfo.h" 37 #include "llvm/CodeGen/MachineFunction.h" 38 #include "llvm/CodeGen/MachineInstr.h" 39 #include "llvm/CodeGen/MachineInstrBuilder.h" 40 #include "llvm/CodeGen/MachineJumpTableInfo.h" 41 #include "llvm/CodeGen/MachineMemOperand.h" 42 #include "llvm/CodeGen/MachineOperand.h" 43 #include "llvm/CodeGen/MachineRegisterInfo.h" 44 #include "llvm/CodeGen/RuntimeLibcalls.h" 45 #include "llvm/CodeGen/SelectionDAG.h" 46 #include "llvm/CodeGen/SelectionDAGNodes.h" 47 #include "llvm/CodeGen/TargetFrameLowering.h" 48 #include "llvm/CodeGen/TargetInstrInfo.h" 49 #include "llvm/CodeGen/TargetRegisterInfo.h" 50 #include "llvm/CodeGen/ValueTypes.h" 51 #include "llvm/IR/CallingConv.h" 52 #include "llvm/IR/Constants.h" 53 #include "llvm/IR/DataLayout.h" 54 #include "llvm/IR/DebugLoc.h" 55 #include "llvm/IR/DerivedTypes.h" 56 #include "llvm/IR/Function.h" 57 #include "llvm/IR/GlobalValue.h" 58 #include "llvm/IR/Type.h" 59 #include "llvm/IR/Value.h" 60 #include "llvm/MC/MCRegisterInfo.h" 61 #include "llvm/Support/Casting.h" 62 #include "llvm/Support/CodeGen.h" 63 #include "llvm/Support/CommandLine.h" 64 #include "llvm/Support/Compiler.h" 65 #include "llvm/Support/ErrorHandling.h" 66 #include "llvm/Support/MachineValueType.h" 67 #include "llvm/Support/MathExtras.h" 68 #include "llvm/Target/TargetMachine.h" 69 #include "llvm/Target/TargetOptions.h" 70 #include <algorithm> 71 #include <cassert> 72 #include <cctype> 73 #include <cstdint> 74 #include <deque> 75 #include <iterator> 76 #include <utility> 77 #include <vector> 78 79 using namespace llvm; 80 81 #define DEBUG_TYPE "mips-lower" 82 83 STATISTIC(NumTailCalls, "Number of tail calls"); 84 85 static cl::opt<bool> 86 LargeGOT("mxgot", cl::Hidden, 87 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false)); 88 89 static cl::opt<bool> 90 NoZeroDivCheck("mno-check-zero-division", cl::Hidden, 91 cl::desc("MIPS: Don't trap on integer division by zero."), 92 cl::init(false)); 93 94 static const MCPhysReg Mips64DPRegs[8] = { 95 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64, 96 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64 97 }; 98 99 // If I is a shifted mask, set the size (Size) and the first bit of the 100 // mask (Pos), and return true. 101 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11). 102 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) { 103 if (!isShiftedMask_64(I)) 104 return false; 105 106 Size = countPopulation(I); 107 Pos = countTrailingZeros(I); 108 return true; 109 } 110 111 // The MIPS MSA ABI passes vector arguments in the integer register set. 112 // The number of integer registers used is dependant on the ABI used. 113 MVT MipsTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context, 114 EVT VT) const { 115 if (VT.isVector()) { 116 if (Subtarget.isABI_O32()) { 117 return MVT::i32; 118 } else { 119 return (VT.getSizeInBits() == 32) ? MVT::i32 : MVT::i64; 120 } 121 } 122 return MipsTargetLowering::getRegisterType(Context, VT); 123 } 124 125 unsigned MipsTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context, 126 EVT VT) const { 127 if (VT.isVector()) 128 return std::max((VT.getSizeInBits() / (Subtarget.isABI_O32() ? 32 : 64)), 129 1U); 130 return MipsTargetLowering::getNumRegisters(Context, VT); 131 } 132 133 unsigned MipsTargetLowering::getVectorTypeBreakdownForCallingConv( 134 LLVMContext &Context, EVT VT, EVT &IntermediateVT, 135 unsigned &NumIntermediates, MVT &RegisterVT) const { 136 // Break down vector types to either 2 i64s or 4 i32s. 137 RegisterVT = getRegisterTypeForCallingConv(Context, VT) ; 138 IntermediateVT = RegisterVT; 139 NumIntermediates = VT.getSizeInBits() < RegisterVT.getSizeInBits() 140 ? VT.getVectorNumElements() 141 : VT.getSizeInBits() / RegisterVT.getSizeInBits(); 142 143 return NumIntermediates; 144 } 145 146 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const { 147 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>(); 148 return DAG.getRegister(FI->getGlobalBaseReg(), Ty); 149 } 150 151 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty, 152 SelectionDAG &DAG, 153 unsigned Flag) const { 154 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag); 155 } 156 157 SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty, 158 SelectionDAG &DAG, 159 unsigned Flag) const { 160 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag); 161 } 162 163 SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty, 164 SelectionDAG &DAG, 165 unsigned Flag) const { 166 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag); 167 } 168 169 SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty, 170 SelectionDAG &DAG, 171 unsigned Flag) const { 172 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag); 173 } 174 175 SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty, 176 SelectionDAG &DAG, 177 unsigned Flag) const { 178 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(), 179 N->getOffset(), Flag); 180 } 181 182 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const { 183 switch ((MipsISD::NodeType)Opcode) { 184 case MipsISD::FIRST_NUMBER: break; 185 case MipsISD::JmpLink: return "MipsISD::JmpLink"; 186 case MipsISD::TailCall: return "MipsISD::TailCall"; 187 case MipsISD::Highest: return "MipsISD::Highest"; 188 case MipsISD::Higher: return "MipsISD::Higher"; 189 case MipsISD::Hi: return "MipsISD::Hi"; 190 case MipsISD::Lo: return "MipsISD::Lo"; 191 case MipsISD::GotHi: return "MipsISD::GotHi"; 192 case MipsISD::GPRel: return "MipsISD::GPRel"; 193 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer"; 194 case MipsISD::Ret: return "MipsISD::Ret"; 195 case MipsISD::ERet: return "MipsISD::ERet"; 196 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN"; 197 case MipsISD::FMS: return "MipsISD::FMS"; 198 case MipsISD::FPBrcond: return "MipsISD::FPBrcond"; 199 case MipsISD::FPCmp: return "MipsISD::FPCmp"; 200 case MipsISD::FSELECT: return "MipsISD::FSELECT"; 201 case MipsISD::MTC1_D64: return "MipsISD::MTC1_D64"; 202 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T"; 203 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F"; 204 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP"; 205 case MipsISD::MFHI: return "MipsISD::MFHI"; 206 case MipsISD::MFLO: return "MipsISD::MFLO"; 207 case MipsISD::MTLOHI: return "MipsISD::MTLOHI"; 208 case MipsISD::Mult: return "MipsISD::Mult"; 209 case MipsISD::Multu: return "MipsISD::Multu"; 210 case MipsISD::MAdd: return "MipsISD::MAdd"; 211 case MipsISD::MAddu: return "MipsISD::MAddu"; 212 case MipsISD::MSub: return "MipsISD::MSub"; 213 case MipsISD::MSubu: return "MipsISD::MSubu"; 214 case MipsISD::DivRem: return "MipsISD::DivRem"; 215 case MipsISD::DivRemU: return "MipsISD::DivRemU"; 216 case MipsISD::DivRem16: return "MipsISD::DivRem16"; 217 case MipsISD::DivRemU16: return "MipsISD::DivRemU16"; 218 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64"; 219 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64"; 220 case MipsISD::Wrapper: return "MipsISD::Wrapper"; 221 case MipsISD::DynAlloc: return "MipsISD::DynAlloc"; 222 case MipsISD::Sync: return "MipsISD::Sync"; 223 case MipsISD::Ext: return "MipsISD::Ext"; 224 case MipsISD::Ins: return "MipsISD::Ins"; 225 case MipsISD::CIns: return "MipsISD::CIns"; 226 case MipsISD::LWL: return "MipsISD::LWL"; 227 case MipsISD::LWR: return "MipsISD::LWR"; 228 case MipsISD::SWL: return "MipsISD::SWL"; 229 case MipsISD::SWR: return "MipsISD::SWR"; 230 case MipsISD::LDL: return "MipsISD::LDL"; 231 case MipsISD::LDR: return "MipsISD::LDR"; 232 case MipsISD::SDL: return "MipsISD::SDL"; 233 case MipsISD::SDR: return "MipsISD::SDR"; 234 case MipsISD::EXTP: return "MipsISD::EXTP"; 235 case MipsISD::EXTPDP: return "MipsISD::EXTPDP"; 236 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H"; 237 case MipsISD::EXTR_W: return "MipsISD::EXTR_W"; 238 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W"; 239 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W"; 240 case MipsISD::SHILO: return "MipsISD::SHILO"; 241 case MipsISD::MTHLIP: return "MipsISD::MTHLIP"; 242 case MipsISD::MULSAQ_S_W_PH: return "MipsISD::MULSAQ_S_W_PH"; 243 case MipsISD::MAQ_S_W_PHL: return "MipsISD::MAQ_S_W_PHL"; 244 case MipsISD::MAQ_S_W_PHR: return "MipsISD::MAQ_S_W_PHR"; 245 case MipsISD::MAQ_SA_W_PHL: return "MipsISD::MAQ_SA_W_PHL"; 246 case MipsISD::MAQ_SA_W_PHR: return "MipsISD::MAQ_SA_W_PHR"; 247 case MipsISD::DPAU_H_QBL: return "MipsISD::DPAU_H_QBL"; 248 case MipsISD::DPAU_H_QBR: return "MipsISD::DPAU_H_QBR"; 249 case MipsISD::DPSU_H_QBL: return "MipsISD::DPSU_H_QBL"; 250 case MipsISD::DPSU_H_QBR: return "MipsISD::DPSU_H_QBR"; 251 case MipsISD::DPAQ_S_W_PH: return "MipsISD::DPAQ_S_W_PH"; 252 case MipsISD::DPSQ_S_W_PH: return "MipsISD::DPSQ_S_W_PH"; 253 case MipsISD::DPAQ_SA_L_W: return "MipsISD::DPAQ_SA_L_W"; 254 case MipsISD::DPSQ_SA_L_W: return "MipsISD::DPSQ_SA_L_W"; 255 case MipsISD::DPA_W_PH: return "MipsISD::DPA_W_PH"; 256 case MipsISD::DPS_W_PH: return "MipsISD::DPS_W_PH"; 257 case MipsISD::DPAQX_S_W_PH: return "MipsISD::DPAQX_S_W_PH"; 258 case MipsISD::DPAQX_SA_W_PH: return "MipsISD::DPAQX_SA_W_PH"; 259 case MipsISD::DPAX_W_PH: return "MipsISD::DPAX_W_PH"; 260 case MipsISD::DPSX_W_PH: return "MipsISD::DPSX_W_PH"; 261 case MipsISD::DPSQX_S_W_PH: return "MipsISD::DPSQX_S_W_PH"; 262 case MipsISD::DPSQX_SA_W_PH: return "MipsISD::DPSQX_SA_W_PH"; 263 case MipsISD::MULSA_W_PH: return "MipsISD::MULSA_W_PH"; 264 case MipsISD::MULT: return "MipsISD::MULT"; 265 case MipsISD::MULTU: return "MipsISD::MULTU"; 266 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP"; 267 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP"; 268 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP"; 269 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP"; 270 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP"; 271 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP"; 272 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP"; 273 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP"; 274 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP"; 275 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO"; 276 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO"; 277 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO"; 278 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO"; 279 case MipsISD::VCEQ: return "MipsISD::VCEQ"; 280 case MipsISD::VCLE_S: return "MipsISD::VCLE_S"; 281 case MipsISD::VCLE_U: return "MipsISD::VCLE_U"; 282 case MipsISD::VCLT_S: return "MipsISD::VCLT_S"; 283 case MipsISD::VCLT_U: return "MipsISD::VCLT_U"; 284 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT"; 285 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT"; 286 case MipsISD::VNOR: return "MipsISD::VNOR"; 287 case MipsISD::VSHF: return "MipsISD::VSHF"; 288 case MipsISD::SHF: return "MipsISD::SHF"; 289 case MipsISD::ILVEV: return "MipsISD::ILVEV"; 290 case MipsISD::ILVOD: return "MipsISD::ILVOD"; 291 case MipsISD::ILVL: return "MipsISD::ILVL"; 292 case MipsISD::ILVR: return "MipsISD::ILVR"; 293 case MipsISD::PCKEV: return "MipsISD::PCKEV"; 294 case MipsISD::PCKOD: return "MipsISD::PCKOD"; 295 case MipsISD::INSVE: return "MipsISD::INSVE"; 296 } 297 return nullptr; 298 } 299 300 MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM, 301 const MipsSubtarget &STI) 302 : TargetLowering(TM), Subtarget(STI), ABI(TM.getABI()) { 303 // Mips does not have i1 type, so use i32 for 304 // setcc operations results (slt, sgt, ...). 305 setBooleanContents(ZeroOrOneBooleanContent); 306 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 307 // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA 308 // does. Integer booleans still use 0 and 1. 309 if (Subtarget.hasMips32r6()) 310 setBooleanContents(ZeroOrOneBooleanContent, 311 ZeroOrNegativeOneBooleanContent); 312 313 // Load extented operations for i1 types must be promoted 314 for (MVT VT : MVT::integer_valuetypes()) { 315 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote); 316 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote); 317 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); 318 } 319 320 // MIPS doesn't have extending float->double load/store. Set LoadExtAction 321 // for f32, f16 322 for (MVT VT : MVT::fp_valuetypes()) { 323 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand); 324 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand); 325 } 326 327 // Set LoadExtAction for f16 vectors to Expand 328 for (MVT VT : MVT::fp_vector_valuetypes()) { 329 MVT F16VT = MVT::getVectorVT(MVT::f16, VT.getVectorNumElements()); 330 if (F16VT.isValid()) 331 setLoadExtAction(ISD::EXTLOAD, VT, F16VT, Expand); 332 } 333 334 setTruncStoreAction(MVT::f32, MVT::f16, Expand); 335 setTruncStoreAction(MVT::f64, MVT::f16, Expand); 336 337 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 338 339 // Used by legalize types to correctly generate the setcc result. 340 // Without this, every float setcc comes with a AND/OR with the result, 341 // we don't want this, since the fpcmp result goes to a flag register, 342 // which is used implicitly by brcond and select operations. 343 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32); 344 345 // Mips Custom Operations 346 setOperationAction(ISD::BR_JT, MVT::Other, Expand); 347 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 348 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 349 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 350 setOperationAction(ISD::JumpTable, MVT::i32, Custom); 351 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 352 setOperationAction(ISD::SELECT, MVT::f32, Custom); 353 setOperationAction(ISD::SELECT, MVT::f64, Custom); 354 setOperationAction(ISD::SELECT, MVT::i32, Custom); 355 setOperationAction(ISD::SETCC, MVT::f32, Custom); 356 setOperationAction(ISD::SETCC, MVT::f64, Custom); 357 setOperationAction(ISD::BRCOND, MVT::Other, Custom); 358 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 359 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 360 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 361 362 if (Subtarget.isGP64bit()) { 363 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); 364 setOperationAction(ISD::BlockAddress, MVT::i64, Custom); 365 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); 366 setOperationAction(ISD::JumpTable, MVT::i64, Custom); 367 setOperationAction(ISD::ConstantPool, MVT::i64, Custom); 368 setOperationAction(ISD::SELECT, MVT::i64, Custom); 369 setOperationAction(ISD::LOAD, MVT::i64, Custom); 370 setOperationAction(ISD::STORE, MVT::i64, Custom); 371 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); 372 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom); 373 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom); 374 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom); 375 } 376 377 if (!Subtarget.isGP64bit()) { 378 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 379 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 380 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 381 } 382 383 setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom); 384 if (Subtarget.isGP64bit()) 385 setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom); 386 387 setOperationAction(ISD::SDIV, MVT::i32, Expand); 388 setOperationAction(ISD::SREM, MVT::i32, Expand); 389 setOperationAction(ISD::UDIV, MVT::i32, Expand); 390 setOperationAction(ISD::UREM, MVT::i32, Expand); 391 setOperationAction(ISD::SDIV, MVT::i64, Expand); 392 setOperationAction(ISD::SREM, MVT::i64, Expand); 393 setOperationAction(ISD::UDIV, MVT::i64, Expand); 394 setOperationAction(ISD::UREM, MVT::i64, Expand); 395 396 // Operations not directly supported by Mips. 397 setOperationAction(ISD::BR_CC, MVT::f32, Expand); 398 setOperationAction(ISD::BR_CC, MVT::f64, Expand); 399 setOperationAction(ISD::BR_CC, MVT::i32, Expand); 400 setOperationAction(ISD::BR_CC, MVT::i64, Expand); 401 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand); 402 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand); 403 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand); 404 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand); 405 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); 406 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); 407 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); 408 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); 409 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 410 if (Subtarget.hasCnMips()) { 411 setOperationAction(ISD::CTPOP, MVT::i32, Legal); 412 setOperationAction(ISD::CTPOP, MVT::i64, Legal); 413 } else { 414 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 415 setOperationAction(ISD::CTPOP, MVT::i64, Expand); 416 } 417 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 418 setOperationAction(ISD::CTTZ, MVT::i64, Expand); 419 setOperationAction(ISD::ROTL, MVT::i32, Expand); 420 setOperationAction(ISD::ROTL, MVT::i64, Expand); 421 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 422 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 423 424 if (!Subtarget.hasMips32r2()) 425 setOperationAction(ISD::ROTR, MVT::i32, Expand); 426 427 if (!Subtarget.hasMips64r2()) 428 setOperationAction(ISD::ROTR, MVT::i64, Expand); 429 430 setOperationAction(ISD::FSIN, MVT::f32, Expand); 431 setOperationAction(ISD::FSIN, MVT::f64, Expand); 432 setOperationAction(ISD::FCOS, MVT::f32, Expand); 433 setOperationAction(ISD::FCOS, MVT::f64, Expand); 434 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 435 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 436 setOperationAction(ISD::FPOW, MVT::f32, Expand); 437 setOperationAction(ISD::FPOW, MVT::f64, Expand); 438 setOperationAction(ISD::FLOG, MVT::f32, Expand); 439 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 440 setOperationAction(ISD::FLOG10, MVT::f32, Expand); 441 setOperationAction(ISD::FEXP, MVT::f32, Expand); 442 setOperationAction(ISD::FMA, MVT::f32, Expand); 443 setOperationAction(ISD::FMA, MVT::f64, Expand); 444 setOperationAction(ISD::FREM, MVT::f32, Expand); 445 setOperationAction(ISD::FREM, MVT::f64, Expand); 446 447 // Lower f16 conversion operations into library calls 448 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand); 449 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand); 450 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand); 451 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand); 452 453 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); 454 455 setOperationAction(ISD::VASTART, MVT::Other, Custom); 456 setOperationAction(ISD::VAARG, MVT::Other, Custom); 457 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 458 setOperationAction(ISD::VAEND, MVT::Other, Expand); 459 460 // Use the default for now 461 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 462 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 463 464 if (!Subtarget.isGP64bit()) { 465 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand); 466 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand); 467 } 468 469 if (!Subtarget.hasMips32r2()) { 470 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 471 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 472 } 473 474 // MIPS16 lacks MIPS32's clz and clo instructions. 475 if (!Subtarget.hasMips32() || Subtarget.inMips16Mode()) 476 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 477 if (!Subtarget.hasMips64()) 478 setOperationAction(ISD::CTLZ, MVT::i64, Expand); 479 480 if (!Subtarget.hasMips32r2()) 481 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 482 if (!Subtarget.hasMips64r2()) 483 setOperationAction(ISD::BSWAP, MVT::i64, Expand); 484 485 if (Subtarget.isGP64bit()) { 486 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, MVT::i32, Custom); 487 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, MVT::i32, Custom); 488 setLoadExtAction(ISD::EXTLOAD, MVT::i64, MVT::i32, Custom); 489 setTruncStoreAction(MVT::i64, MVT::i32, Custom); 490 } 491 492 setOperationAction(ISD::TRAP, MVT::Other, Legal); 493 494 setTargetDAGCombine(ISD::SDIVREM); 495 setTargetDAGCombine(ISD::UDIVREM); 496 setTargetDAGCombine(ISD::SELECT); 497 setTargetDAGCombine(ISD::AND); 498 setTargetDAGCombine(ISD::OR); 499 setTargetDAGCombine(ISD::ADD); 500 setTargetDAGCombine(ISD::SUB); 501 setTargetDAGCombine(ISD::AssertZext); 502 setTargetDAGCombine(ISD::SHL); 503 504 if (ABI.IsO32()) { 505 // These libcalls are not available in 32-bit. 506 setLibcallName(RTLIB::SHL_I128, nullptr); 507 setLibcallName(RTLIB::SRL_I128, nullptr); 508 setLibcallName(RTLIB::SRA_I128, nullptr); 509 } 510 511 setMinFunctionAlignment(Subtarget.isGP64bit() ? 3 : 2); 512 513 // The arguments on the stack are defined in terms of 4-byte slots on O32 514 // and 8-byte slots on N32/N64. 515 setMinStackArgumentAlignment((ABI.IsN32() || ABI.IsN64()) ? 8 : 4); 516 517 setStackPointerRegisterToSaveRestore(ABI.IsN64() ? Mips::SP_64 : Mips::SP); 518 519 MaxStoresPerMemcpy = 16; 520 521 isMicroMips = Subtarget.inMicroMipsMode(); 522 } 523 524 const MipsTargetLowering *MipsTargetLowering::create(const MipsTargetMachine &TM, 525 const MipsSubtarget &STI) { 526 if (STI.inMips16Mode()) 527 return createMips16TargetLowering(TM, STI); 528 529 return createMipsSETargetLowering(TM, STI); 530 } 531 532 // Create a fast isel object. 533 FastISel * 534 MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, 535 const TargetLibraryInfo *libInfo) const { 536 const MipsTargetMachine &TM = 537 static_cast<const MipsTargetMachine &>(funcInfo.MF->getTarget()); 538 539 // We support only the standard encoding [MIPS32,MIPS32R5] ISAs. 540 bool UseFastISel = TM.Options.EnableFastISel && Subtarget.hasMips32() && 541 !Subtarget.hasMips32r6() && !Subtarget.inMips16Mode() && 542 !Subtarget.inMicroMipsMode(); 543 544 // Disable if either of the following is true: 545 // We do not generate PIC, the ABI is not O32, LargeGOT is being used. 546 if (!TM.isPositionIndependent() || !TM.getABI().IsO32() || LargeGOT) 547 UseFastISel = false; 548 549 return UseFastISel ? Mips::createFastISel(funcInfo, libInfo) : nullptr; 550 } 551 552 EVT MipsTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &, 553 EVT VT) const { 554 if (!VT.isVector()) 555 return MVT::i32; 556 return VT.changeVectorElementTypeToInteger(); 557 } 558 559 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG, 560 TargetLowering::DAGCombinerInfo &DCI, 561 const MipsSubtarget &Subtarget) { 562 if (DCI.isBeforeLegalizeOps()) 563 return SDValue(); 564 565 EVT Ty = N->getValueType(0); 566 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64; 567 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64; 568 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 : 569 MipsISD::DivRemU16; 570 SDLoc DL(N); 571 572 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue, 573 N->getOperand(0), N->getOperand(1)); 574 SDValue InChain = DAG.getEntryNode(); 575 SDValue InGlue = DivRem; 576 577 // insert MFLO 578 if (N->hasAnyUseOfValue(0)) { 579 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty, 580 InGlue); 581 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo); 582 InChain = CopyFromLo.getValue(1); 583 InGlue = CopyFromLo.getValue(2); 584 } 585 586 // insert MFHI 587 if (N->hasAnyUseOfValue(1)) { 588 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL, 589 HI, Ty, InGlue); 590 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi); 591 } 592 593 return SDValue(); 594 } 595 596 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) { 597 switch (CC) { 598 default: llvm_unreachable("Unknown fp condition code!"); 599 case ISD::SETEQ: 600 case ISD::SETOEQ: return Mips::FCOND_OEQ; 601 case ISD::SETUNE: return Mips::FCOND_UNE; 602 case ISD::SETLT: 603 case ISD::SETOLT: return Mips::FCOND_OLT; 604 case ISD::SETGT: 605 case ISD::SETOGT: return Mips::FCOND_OGT; 606 case ISD::SETLE: 607 case ISD::SETOLE: return Mips::FCOND_OLE; 608 case ISD::SETGE: 609 case ISD::SETOGE: return Mips::FCOND_OGE; 610 case ISD::SETULT: return Mips::FCOND_ULT; 611 case ISD::SETULE: return Mips::FCOND_ULE; 612 case ISD::SETUGT: return Mips::FCOND_UGT; 613 case ISD::SETUGE: return Mips::FCOND_UGE; 614 case ISD::SETUO: return Mips::FCOND_UN; 615 case ISD::SETO: return Mips::FCOND_OR; 616 case ISD::SETNE: 617 case ISD::SETONE: return Mips::FCOND_ONE; 618 case ISD::SETUEQ: return Mips::FCOND_UEQ; 619 } 620 } 621 622 /// This function returns true if the floating point conditional branches and 623 /// conditional moves which use condition code CC should be inverted. 624 static bool invertFPCondCodeUser(Mips::CondCode CC) { 625 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT) 626 return false; 627 628 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) && 629 "Illegal Condition Code"); 630 631 return true; 632 } 633 634 // Creates and returns an FPCmp node from a setcc node. 635 // Returns Op if setcc is not a floating point comparison. 636 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) { 637 // must be a SETCC node 638 if (Op.getOpcode() != ISD::SETCC) 639 return Op; 640 641 SDValue LHS = Op.getOperand(0); 642 643 if (!LHS.getValueType().isFloatingPoint()) 644 return Op; 645 646 SDValue RHS = Op.getOperand(1); 647 SDLoc DL(Op); 648 649 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of 650 // node if necessary. 651 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); 652 653 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS, 654 DAG.getConstant(condCodeToFCC(CC), DL, MVT::i32)); 655 } 656 657 // Creates and returns a CMovFPT/F node. 658 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True, 659 SDValue False, const SDLoc &DL) { 660 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2)); 661 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue()); 662 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 663 664 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL, 665 True.getValueType(), True, FCC0, False, Cond); 666 } 667 668 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG, 669 TargetLowering::DAGCombinerInfo &DCI, 670 const MipsSubtarget &Subtarget) { 671 if (DCI.isBeforeLegalizeOps()) 672 return SDValue(); 673 674 SDValue SetCC = N->getOperand(0); 675 676 if ((SetCC.getOpcode() != ISD::SETCC) || 677 !SetCC.getOperand(0).getValueType().isInteger()) 678 return SDValue(); 679 680 SDValue False = N->getOperand(2); 681 EVT FalseTy = False.getValueType(); 682 683 if (!FalseTy.isInteger()) 684 return SDValue(); 685 686 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(False); 687 688 // If the RHS (False) is 0, we swap the order of the operands 689 // of ISD::SELECT (obviously also inverting the condition) so that we can 690 // take advantage of conditional moves using the $0 register. 691 // Example: 692 // return (a != 0) ? x : 0; 693 // load $reg, x 694 // movz $reg, $0, a 695 if (!FalseC) 696 return SDValue(); 697 698 const SDLoc DL(N); 699 700 if (!FalseC->getZExtValue()) { 701 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 702 SDValue True = N->getOperand(1); 703 704 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 705 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 706 707 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True); 708 } 709 710 // If both operands are integer constants there's a possibility that we 711 // can do some interesting optimizations. 712 SDValue True = N->getOperand(1); 713 ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(True); 714 715 if (!TrueC || !True.getValueType().isInteger()) 716 return SDValue(); 717 718 // We'll also ignore MVT::i64 operands as this optimizations proves 719 // to be ineffective because of the required sign extensions as the result 720 // of a SETCC operator is always MVT::i32 for non-vector types. 721 if (True.getValueType() == MVT::i64) 722 return SDValue(); 723 724 int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue(); 725 726 // 1) (a < x) ? y : y-1 727 // slti $reg1, a, x 728 // addiu $reg2, $reg1, y-1 729 if (Diff == 1) 730 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, False); 731 732 // 2) (a < x) ? y-1 : y 733 // slti $reg1, a, x 734 // xor $reg1, $reg1, 1 735 // addiu $reg2, $reg1, y-1 736 if (Diff == -1) { 737 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 738 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 739 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 740 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, True); 741 } 742 743 // Could not optimize. 744 return SDValue(); 745 } 746 747 static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG, 748 TargetLowering::DAGCombinerInfo &DCI, 749 const MipsSubtarget &Subtarget) { 750 if (DCI.isBeforeLegalizeOps()) 751 return SDValue(); 752 753 SDValue ValueIfTrue = N->getOperand(0), ValueIfFalse = N->getOperand(2); 754 755 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(ValueIfFalse); 756 if (!FalseC || FalseC->getZExtValue()) 757 return SDValue(); 758 759 // Since RHS (False) is 0, we swap the order of the True/False operands 760 // (obviously also inverting the condition) so that we can 761 // take advantage of conditional moves using the $0 register. 762 // Example: 763 // return (a != 0) ? x : 0; 764 // load $reg, x 765 // movz $reg, $0, a 766 unsigned Opc = (N->getOpcode() == MipsISD::CMovFP_T) ? MipsISD::CMovFP_F : 767 MipsISD::CMovFP_T; 768 769 SDValue FCC = N->getOperand(1), Glue = N->getOperand(3); 770 return DAG.getNode(Opc, SDLoc(N), ValueIfFalse.getValueType(), 771 ValueIfFalse, FCC, ValueIfTrue, Glue); 772 } 773 774 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, 775 TargetLowering::DAGCombinerInfo &DCI, 776 const MipsSubtarget &Subtarget) { 777 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) 778 return SDValue(); 779 780 SDValue FirstOperand = N->getOperand(0); 781 unsigned FirstOperandOpc = FirstOperand.getOpcode(); 782 SDValue Mask = N->getOperand(1); 783 EVT ValTy = N->getValueType(0); 784 SDLoc DL(N); 785 786 uint64_t Pos = 0, SMPos, SMSize; 787 ConstantSDNode *CN; 788 SDValue NewOperand; 789 unsigned Opc; 790 791 // Op's second operand must be a shifted mask. 792 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) || 793 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize)) 794 return SDValue(); 795 796 if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL) { 797 // Pattern match EXT. 798 // $dst = and ((sra or srl) $src , pos), (2**size - 1) 799 // => ext $dst, $src, pos, size 800 801 // The second operand of the shift must be an immediate. 802 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1)))) 803 return SDValue(); 804 805 Pos = CN->getZExtValue(); 806 807 // Return if the shifted mask does not start at bit 0 or the sum of its size 808 // and Pos exceeds the word's size. 809 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits()) 810 return SDValue(); 811 812 Opc = MipsISD::Ext; 813 NewOperand = FirstOperand.getOperand(0); 814 } else if (FirstOperandOpc == ISD::SHL && Subtarget.hasCnMips()) { 815 // Pattern match CINS. 816 // $dst = and (shl $src , pos), mask 817 // => cins $dst, $src, pos, size 818 // mask is a shifted mask with consecutive 1's, pos = shift amount, 819 // size = population count. 820 821 // The second operand of the shift must be an immediate. 822 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1)))) 823 return SDValue(); 824 825 Pos = CN->getZExtValue(); 826 827 if (SMPos != Pos || Pos >= ValTy.getSizeInBits() || SMSize >= 32 || 828 Pos + SMSize > ValTy.getSizeInBits()) 829 return SDValue(); 830 831 NewOperand = FirstOperand.getOperand(0); 832 // SMSize is 'location' (position) in this case, not size. 833 SMSize--; 834 Opc = MipsISD::CIns; 835 } else { 836 // Pattern match EXT. 837 // $dst = and $src, (2**size - 1) , if size > 16 838 // => ext $dst, $src, pos, size , pos = 0 839 840 // If the mask is <= 0xffff, andi can be used instead. 841 if (CN->getZExtValue() <= 0xffff) 842 return SDValue(); 843 844 // Return if the mask doesn't start at position 0. 845 if (SMPos) 846 return SDValue(); 847 848 Opc = MipsISD::Ext; 849 NewOperand = FirstOperand; 850 } 851 return DAG.getNode(Opc, DL, ValTy, NewOperand, 852 DAG.getConstant(Pos, DL, MVT::i32), 853 DAG.getConstant(SMSize, DL, MVT::i32)); 854 } 855 856 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, 857 TargetLowering::DAGCombinerInfo &DCI, 858 const MipsSubtarget &Subtarget) { 859 // Pattern match INS. 860 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1), 861 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1 862 // => ins $dst, $src, size, pos, $src1 863 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) 864 return SDValue(); 865 866 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1); 867 uint64_t SMPos0, SMSize0, SMPos1, SMSize1; 868 ConstantSDNode *CN, *CN1; 869 870 // See if Op's first operand matches (and $src1 , mask0). 871 if (And0.getOpcode() != ISD::AND) 872 return SDValue(); 873 874 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) || 875 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0)) 876 return SDValue(); 877 878 // See if Op's second operand matches (and (shl $src, pos), mask1). 879 if (And1.getOpcode() == ISD::AND && 880 And1.getOperand(0).getOpcode() == ISD::SHL) { 881 882 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) || 883 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1)) 884 return SDValue(); 885 886 // The shift masks must have the same position and size. 887 if (SMPos0 != SMPos1 || SMSize0 != SMSize1) 888 return SDValue(); 889 890 SDValue Shl = And1.getOperand(0); 891 892 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1)))) 893 return SDValue(); 894 895 unsigned Shamt = CN->getZExtValue(); 896 897 // Return if the shift amount and the first bit position of mask are not the 898 // same. 899 EVT ValTy = N->getValueType(0); 900 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits())) 901 return SDValue(); 902 903 SDLoc DL(N); 904 return DAG.getNode(MipsISD::Ins, DL, ValTy, Shl.getOperand(0), 905 DAG.getConstant(SMPos0, DL, MVT::i32), 906 DAG.getConstant(SMSize0, DL, MVT::i32), 907 And0.getOperand(0)); 908 } else { 909 // Pattern match DINS. 910 // $dst = or (and $src, mask0), mask1 911 // where mask0 = ((1 << SMSize0) -1) << SMPos0 912 // => dins $dst, $src, pos, size 913 if (~CN->getSExtValue() == ((((int64_t)1 << SMSize0) - 1) << SMPos0) && 914 ((SMSize0 + SMPos0 <= 64 && Subtarget.hasMips64r2()) || 915 (SMSize0 + SMPos0 <= 32))) { 916 // Check if AND instruction has constant as argument 917 bool isConstCase = And1.getOpcode() != ISD::AND; 918 if (And1.getOpcode() == ISD::AND) { 919 if (!(CN1 = dyn_cast<ConstantSDNode>(And1->getOperand(1)))) 920 return SDValue(); 921 } else { 922 if (!(CN1 = dyn_cast<ConstantSDNode>(N->getOperand(1)))) 923 return SDValue(); 924 } 925 // Don't generate INS if constant OR operand doesn't fit into bits 926 // cleared by constant AND operand. 927 if (CN->getSExtValue() & CN1->getSExtValue()) 928 return SDValue(); 929 930 SDLoc DL(N); 931 EVT ValTy = N->getOperand(0)->getValueType(0); 932 SDValue Const1; 933 SDValue SrlX; 934 if (!isConstCase) { 935 Const1 = DAG.getConstant(SMPos0, DL, MVT::i32); 936 SrlX = DAG.getNode(ISD::SRL, DL, And1->getValueType(0), And1, Const1); 937 } 938 return DAG.getNode( 939 MipsISD::Ins, DL, N->getValueType(0), 940 isConstCase 941 ? DAG.getConstant(CN1->getSExtValue() >> SMPos0, DL, ValTy) 942 : SrlX, 943 DAG.getConstant(SMPos0, DL, MVT::i32), 944 DAG.getConstant(ValTy.getSizeInBits() / 8 < 8 ? SMSize0 & 31 945 : SMSize0, 946 DL, MVT::i32), 947 And0->getOperand(0)); 948 949 } 950 return SDValue(); 951 } 952 } 953 954 static SDValue performMADD_MSUBCombine(SDNode *ROOTNode, SelectionDAG &CurDAG, 955 const MipsSubtarget &Subtarget) { 956 // ROOTNode must have a multiplication as an operand for the match to be 957 // successful. 958 if (ROOTNode->getOperand(0).getOpcode() != ISD::MUL && 959 ROOTNode->getOperand(1).getOpcode() != ISD::MUL) 960 return SDValue(); 961 962 // We don't handle vector types here. 963 if (ROOTNode->getValueType(0).isVector()) 964 return SDValue(); 965 966 // For MIPS64, madd / msub instructions are inefficent to use with 64 bit 967 // arithmetic. E.g. 968 // (add (mul a b) c) => 969 // let res = (madd (mthi (drotr c 32))x(mtlo c) a b) in 970 // MIPS64: (or (dsll (mfhi res) 32) (dsrl (dsll (mflo res) 32) 32) 971 // or 972 // MIPS64R2: (dins (mflo res) (mfhi res) 32 32) 973 // 974 // The overhead of setting up the Hi/Lo registers and reassembling the 975 // result makes this a dubious optimzation for MIPS64. The core of the 976 // problem is that Hi/Lo contain the upper and lower 32 bits of the 977 // operand and result. 978 // 979 // It requires a chain of 4 add/mul for MIPS64R2 to get better code 980 // density than doing it naively, 5 for MIPS64. Additionally, using 981 // madd/msub on MIPS64 requires the operands actually be 32 bit sign 982 // extended operands, not true 64 bit values. 983 // 984 // FIXME: For the moment, disable this completely for MIPS64. 985 if (Subtarget.hasMips64()) 986 return SDValue(); 987 988 SDValue Mult = ROOTNode->getOperand(0).getOpcode() == ISD::MUL 989 ? ROOTNode->getOperand(0) 990 : ROOTNode->getOperand(1); 991 992 SDValue AddOperand = ROOTNode->getOperand(0).getOpcode() == ISD::MUL 993 ? ROOTNode->getOperand(1) 994 : ROOTNode->getOperand(0); 995 996 // Transform this to a MADD only if the user of this node is the add. 997 // If there are other users of the mul, this function returns here. 998 if (!Mult.hasOneUse()) 999 return SDValue(); 1000 1001 // maddu and madd are unusual instructions in that on MIPS64 bits 63..31 1002 // must be in canonical form, i.e. sign extended. For MIPS32, the operands 1003 // of the multiply must have 32 or more sign bits, otherwise we cannot 1004 // perform this optimization. We have to check this here as we're performing 1005 // this optimization pre-legalization. 1006 SDValue MultLHS = Mult->getOperand(0); 1007 SDValue MultRHS = Mult->getOperand(1); 1008 1009 bool IsSigned = MultLHS->getOpcode() == ISD::SIGN_EXTEND && 1010 MultRHS->getOpcode() == ISD::SIGN_EXTEND; 1011 bool IsUnsigned = MultLHS->getOpcode() == ISD::ZERO_EXTEND && 1012 MultRHS->getOpcode() == ISD::ZERO_EXTEND; 1013 1014 if (!IsSigned && !IsUnsigned) 1015 return SDValue(); 1016 1017 // Initialize accumulator. 1018 SDLoc DL(ROOTNode); 1019 SDValue TopHalf; 1020 SDValue BottomHalf; 1021 BottomHalf = CurDAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, AddOperand, 1022 CurDAG.getIntPtrConstant(0, DL)); 1023 1024 TopHalf = CurDAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, AddOperand, 1025 CurDAG.getIntPtrConstant(1, DL)); 1026 SDValue ACCIn = CurDAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped, 1027 BottomHalf, 1028 TopHalf); 1029 1030 // Create MipsMAdd(u) / MipsMSub(u) node. 1031 bool IsAdd = ROOTNode->getOpcode() == ISD::ADD; 1032 unsigned Opcode = IsAdd ? (IsUnsigned ? MipsISD::MAddu : MipsISD::MAdd) 1033 : (IsUnsigned ? MipsISD::MSubu : MipsISD::MSub); 1034 SDValue MAddOps[3] = { 1035 CurDAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mult->getOperand(0)), 1036 CurDAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mult->getOperand(1)), ACCIn}; 1037 EVT VTs[2] = {MVT::i32, MVT::i32}; 1038 SDValue MAdd = CurDAG.getNode(Opcode, DL, VTs, MAddOps); 1039 1040 SDValue ResLo = CurDAG.getNode(MipsISD::MFLO, DL, MVT::i32, MAdd); 1041 SDValue ResHi = CurDAG.getNode(MipsISD::MFHI, DL, MVT::i32, MAdd); 1042 SDValue Combined = 1043 CurDAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResLo, ResHi); 1044 return Combined; 1045 } 1046 1047 static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG, 1048 TargetLowering::DAGCombinerInfo &DCI, 1049 const MipsSubtarget &Subtarget) { 1050 // (sub v0 (mul v1, v2)) => (msub v1, v2, v0) 1051 if (DCI.isBeforeLegalizeOps()) { 1052 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() && 1053 !Subtarget.inMips16Mode() && N->getValueType(0) == MVT::i64) 1054 return performMADD_MSUBCombine(N, DAG, Subtarget); 1055 1056 return SDValue(); 1057 } 1058 1059 return SDValue(); 1060 } 1061 1062 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG, 1063 TargetLowering::DAGCombinerInfo &DCI, 1064 const MipsSubtarget &Subtarget) { 1065 // (add v0 (mul v1, v2)) => (madd v1, v2, v0) 1066 if (DCI.isBeforeLegalizeOps()) { 1067 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() && 1068 !Subtarget.inMips16Mode() && N->getValueType(0) == MVT::i64) 1069 return performMADD_MSUBCombine(N, DAG, Subtarget); 1070 1071 return SDValue(); 1072 } 1073 1074 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt)) 1075 SDValue Add = N->getOperand(1); 1076 1077 if (Add.getOpcode() != ISD::ADD) 1078 return SDValue(); 1079 1080 SDValue Lo = Add.getOperand(1); 1081 1082 if ((Lo.getOpcode() != MipsISD::Lo) || 1083 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable)) 1084 return SDValue(); 1085 1086 EVT ValTy = N->getValueType(0); 1087 SDLoc DL(N); 1088 1089 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0), 1090 Add.getOperand(0)); 1091 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo); 1092 } 1093 1094 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG, 1095 TargetLowering::DAGCombinerInfo &DCI, 1096 const MipsSubtarget &Subtarget) { 1097 // Pattern match CINS. 1098 // $dst = shl (and $src , imm), pos 1099 // => cins $dst, $src, pos, size 1100 1101 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasCnMips()) 1102 return SDValue(); 1103 1104 SDValue FirstOperand = N->getOperand(0); 1105 unsigned FirstOperandOpc = FirstOperand.getOpcode(); 1106 SDValue SecondOperand = N->getOperand(1); 1107 EVT ValTy = N->getValueType(0); 1108 SDLoc DL(N); 1109 1110 uint64_t Pos = 0, SMPos, SMSize; 1111 ConstantSDNode *CN; 1112 SDValue NewOperand; 1113 1114 // The second operand of the shift must be an immediate. 1115 if (!(CN = dyn_cast<ConstantSDNode>(SecondOperand))) 1116 return SDValue(); 1117 1118 Pos = CN->getZExtValue(); 1119 1120 if (Pos >= ValTy.getSizeInBits()) 1121 return SDValue(); 1122 1123 if (FirstOperandOpc != ISD::AND) 1124 return SDValue(); 1125 1126 // AND's second operand must be a shifted mask. 1127 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))) || 1128 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize)) 1129 return SDValue(); 1130 1131 // Return if the shifted mask does not start at bit 0 or the sum of its size 1132 // and Pos exceeds the word's size. 1133 if (SMPos != 0 || SMSize > 32 || Pos + SMSize > ValTy.getSizeInBits()) 1134 return SDValue(); 1135 1136 NewOperand = FirstOperand.getOperand(0); 1137 // SMSize is 'location' (position) in this case, not size. 1138 SMSize--; 1139 1140 return DAG.getNode(MipsISD::CIns, DL, ValTy, NewOperand, 1141 DAG.getConstant(Pos, DL, MVT::i32), 1142 DAG.getConstant(SMSize, DL, MVT::i32)); 1143 } 1144 1145 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) 1146 const { 1147 SelectionDAG &DAG = DCI.DAG; 1148 unsigned Opc = N->getOpcode(); 1149 1150 switch (Opc) { 1151 default: break; 1152 case ISD::SDIVREM: 1153 case ISD::UDIVREM: 1154 return performDivRemCombine(N, DAG, DCI, Subtarget); 1155 case ISD::SELECT: 1156 return performSELECTCombine(N, DAG, DCI, Subtarget); 1157 case MipsISD::CMovFP_F: 1158 case MipsISD::CMovFP_T: 1159 return performCMovFPCombine(N, DAG, DCI, Subtarget); 1160 case ISD::AND: 1161 return performANDCombine(N, DAG, DCI, Subtarget); 1162 case ISD::OR: 1163 return performORCombine(N, DAG, DCI, Subtarget); 1164 case ISD::ADD: 1165 return performADDCombine(N, DAG, DCI, Subtarget); 1166 case ISD::SHL: 1167 return performSHLCombine(N, DAG, DCI, Subtarget); 1168 case ISD::SUB: 1169 return performSUBCombine(N, DAG, DCI, Subtarget); 1170 } 1171 1172 return SDValue(); 1173 } 1174 1175 bool MipsTargetLowering::isCheapToSpeculateCttz() const { 1176 return Subtarget.hasMips32(); 1177 } 1178 1179 bool MipsTargetLowering::isCheapToSpeculateCtlz() const { 1180 return Subtarget.hasMips32(); 1181 } 1182 1183 void 1184 MipsTargetLowering::LowerOperationWrapper(SDNode *N, 1185 SmallVectorImpl<SDValue> &Results, 1186 SelectionDAG &DAG) const { 1187 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 1188 1189 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I) 1190 Results.push_back(Res.getValue(I)); 1191 } 1192 1193 void 1194 MipsTargetLowering::ReplaceNodeResults(SDNode *N, 1195 SmallVectorImpl<SDValue> &Results, 1196 SelectionDAG &DAG) const { 1197 return LowerOperationWrapper(N, Results, DAG); 1198 } 1199 1200 SDValue MipsTargetLowering:: 1201 LowerOperation(SDValue Op, SelectionDAG &DAG) const 1202 { 1203 switch (Op.getOpcode()) 1204 { 1205 case ISD::BRCOND: return lowerBRCOND(Op, DAG); 1206 case ISD::ConstantPool: return lowerConstantPool(Op, DAG); 1207 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG); 1208 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG); 1209 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG); 1210 case ISD::JumpTable: return lowerJumpTable(Op, DAG); 1211 case ISD::SELECT: return lowerSELECT(Op, DAG); 1212 case ISD::SETCC: return lowerSETCC(Op, DAG); 1213 case ISD::VASTART: return lowerVASTART(Op, DAG); 1214 case ISD::VAARG: return lowerVAARG(Op, DAG); 1215 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG); 1216 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG); 1217 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG); 1218 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG); 1219 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG); 1220 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG); 1221 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true); 1222 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false); 1223 case ISD::LOAD: return lowerLOAD(Op, DAG); 1224 case ISD::STORE: return lowerSTORE(Op, DAG); 1225 case ISD::EH_DWARF_CFA: return lowerEH_DWARF_CFA(Op, DAG); 1226 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG); 1227 } 1228 return SDValue(); 1229 } 1230 1231 //===----------------------------------------------------------------------===// 1232 // Lower helper functions 1233 //===----------------------------------------------------------------------===// 1234 1235 // addLiveIn - This helper function adds the specified physical register to the 1236 // MachineFunction as a live in value. It also creates a corresponding 1237 // virtual register for it. 1238 static unsigned 1239 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC) 1240 { 1241 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC); 1242 MF.getRegInfo().addLiveIn(PReg, VReg); 1243 return VReg; 1244 } 1245 1246 static MachineBasicBlock *insertDivByZeroTrap(MachineInstr &MI, 1247 MachineBasicBlock &MBB, 1248 const TargetInstrInfo &TII, 1249 bool Is64Bit, bool IsMicroMips) { 1250 if (NoZeroDivCheck) 1251 return &MBB; 1252 1253 // Insert instruction "teq $divisor_reg, $zero, 7". 1254 MachineBasicBlock::iterator I(MI); 1255 MachineInstrBuilder MIB; 1256 MachineOperand &Divisor = MI.getOperand(2); 1257 MIB = BuildMI(MBB, std::next(I), MI.getDebugLoc(), 1258 TII.get(IsMicroMips ? Mips::TEQ_MM : Mips::TEQ)) 1259 .addReg(Divisor.getReg(), getKillRegState(Divisor.isKill())) 1260 .addReg(Mips::ZERO) 1261 .addImm(7); 1262 1263 // Use the 32-bit sub-register if this is a 64-bit division. 1264 if (Is64Bit) 1265 MIB->getOperand(0).setSubReg(Mips::sub_32); 1266 1267 // Clear Divisor's kill flag. 1268 Divisor.setIsKill(false); 1269 1270 // We would normally delete the original instruction here but in this case 1271 // we only needed to inject an additional instruction rather than replace it. 1272 1273 return &MBB; 1274 } 1275 1276 MachineBasicBlock * 1277 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, 1278 MachineBasicBlock *BB) const { 1279 switch (MI.getOpcode()) { 1280 default: 1281 llvm_unreachable("Unexpected instr type to insert"); 1282 case Mips::ATOMIC_LOAD_ADD_I8: 1283 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu); 1284 case Mips::ATOMIC_LOAD_ADD_I16: 1285 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu); 1286 case Mips::ATOMIC_LOAD_ADD_I32: 1287 return emitAtomicBinary(MI, BB, 4, Mips::ADDu); 1288 case Mips::ATOMIC_LOAD_ADD_I64: 1289 return emitAtomicBinary(MI, BB, 8, Mips::DADDu); 1290 1291 case Mips::ATOMIC_LOAD_AND_I8: 1292 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND); 1293 case Mips::ATOMIC_LOAD_AND_I16: 1294 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND); 1295 case Mips::ATOMIC_LOAD_AND_I32: 1296 return emitAtomicBinary(MI, BB, 4, Mips::AND); 1297 case Mips::ATOMIC_LOAD_AND_I64: 1298 return emitAtomicBinary(MI, BB, 8, Mips::AND64); 1299 1300 case Mips::ATOMIC_LOAD_OR_I8: 1301 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR); 1302 case Mips::ATOMIC_LOAD_OR_I16: 1303 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR); 1304 case Mips::ATOMIC_LOAD_OR_I32: 1305 return emitAtomicBinary(MI, BB, 4, Mips::OR); 1306 case Mips::ATOMIC_LOAD_OR_I64: 1307 return emitAtomicBinary(MI, BB, 8, Mips::OR64); 1308 1309 case Mips::ATOMIC_LOAD_XOR_I8: 1310 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR); 1311 case Mips::ATOMIC_LOAD_XOR_I16: 1312 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR); 1313 case Mips::ATOMIC_LOAD_XOR_I32: 1314 return emitAtomicBinary(MI, BB, 4, Mips::XOR); 1315 case Mips::ATOMIC_LOAD_XOR_I64: 1316 return emitAtomicBinary(MI, BB, 8, Mips::XOR64); 1317 1318 case Mips::ATOMIC_LOAD_NAND_I8: 1319 return emitAtomicBinaryPartword(MI, BB, 1, 0, true); 1320 case Mips::ATOMIC_LOAD_NAND_I16: 1321 return emitAtomicBinaryPartword(MI, BB, 2, 0, true); 1322 case Mips::ATOMIC_LOAD_NAND_I32: 1323 return emitAtomicBinary(MI, BB, 4, 0, true); 1324 case Mips::ATOMIC_LOAD_NAND_I64: 1325 return emitAtomicBinary(MI, BB, 8, 0, true); 1326 1327 case Mips::ATOMIC_LOAD_SUB_I8: 1328 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu); 1329 case Mips::ATOMIC_LOAD_SUB_I16: 1330 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu); 1331 case Mips::ATOMIC_LOAD_SUB_I32: 1332 return emitAtomicBinary(MI, BB, 4, Mips::SUBu); 1333 case Mips::ATOMIC_LOAD_SUB_I64: 1334 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu); 1335 1336 case Mips::ATOMIC_SWAP_I8: 1337 return emitAtomicBinaryPartword(MI, BB, 1, 0); 1338 case Mips::ATOMIC_SWAP_I16: 1339 return emitAtomicBinaryPartword(MI, BB, 2, 0); 1340 case Mips::ATOMIC_SWAP_I32: 1341 return emitAtomicBinary(MI, BB, 4, 0); 1342 case Mips::ATOMIC_SWAP_I64: 1343 return emitAtomicBinary(MI, BB, 8, 0); 1344 1345 case Mips::ATOMIC_CMP_SWAP_I8: 1346 return emitAtomicCmpSwapPartword(MI, BB, 1); 1347 case Mips::ATOMIC_CMP_SWAP_I16: 1348 return emitAtomicCmpSwapPartword(MI, BB, 2); 1349 case Mips::ATOMIC_CMP_SWAP_I32: 1350 return emitAtomicCmpSwap(MI, BB, 4); 1351 case Mips::ATOMIC_CMP_SWAP_I64: 1352 return emitAtomicCmpSwap(MI, BB, 8); 1353 case Mips::PseudoSDIV: 1354 case Mips::PseudoUDIV: 1355 case Mips::DIV: 1356 case Mips::DIVU: 1357 case Mips::MOD: 1358 case Mips::MODU: 1359 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false, 1360 false); 1361 case Mips::SDIV_MM_Pseudo: 1362 case Mips::UDIV_MM_Pseudo: 1363 case Mips::SDIV_MM: 1364 case Mips::UDIV_MM: 1365 case Mips::DIV_MMR6: 1366 case Mips::DIVU_MMR6: 1367 case Mips::MOD_MMR6: 1368 case Mips::MODU_MMR6: 1369 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false, true); 1370 case Mips::PseudoDSDIV: 1371 case Mips::PseudoDUDIV: 1372 case Mips::DDIV: 1373 case Mips::DDIVU: 1374 case Mips::DMOD: 1375 case Mips::DMODU: 1376 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), true, false); 1377 1378 case Mips::PseudoSELECT_I: 1379 case Mips::PseudoSELECT_I64: 1380 case Mips::PseudoSELECT_S: 1381 case Mips::PseudoSELECT_D32: 1382 case Mips::PseudoSELECT_D64: 1383 return emitPseudoSELECT(MI, BB, false, Mips::BNE); 1384 case Mips::PseudoSELECTFP_F_I: 1385 case Mips::PseudoSELECTFP_F_I64: 1386 case Mips::PseudoSELECTFP_F_S: 1387 case Mips::PseudoSELECTFP_F_D32: 1388 case Mips::PseudoSELECTFP_F_D64: 1389 return emitPseudoSELECT(MI, BB, true, Mips::BC1F); 1390 case Mips::PseudoSELECTFP_T_I: 1391 case Mips::PseudoSELECTFP_T_I64: 1392 case Mips::PseudoSELECTFP_T_S: 1393 case Mips::PseudoSELECTFP_T_D32: 1394 case Mips::PseudoSELECTFP_T_D64: 1395 return emitPseudoSELECT(MI, BB, true, Mips::BC1T); 1396 } 1397 } 1398 1399 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and 1400 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true) 1401 MachineBasicBlock *MipsTargetLowering::emitAtomicBinary(MachineInstr &MI, 1402 MachineBasicBlock *BB, 1403 unsigned Size, 1404 unsigned BinOpcode, 1405 bool Nand) const { 1406 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary."); 1407 1408 MachineFunction *MF = BB->getParent(); 1409 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1410 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 1411 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1412 const bool ArePtrs64bit = ABI.ArePtrs64bit(); 1413 DebugLoc DL = MI.getDebugLoc(); 1414 unsigned LL, SC, AND, NOR, ZERO, BEQ; 1415 1416 if (Size == 4) { 1417 if (isMicroMips) { 1418 LL = Subtarget.hasMips32r6() ? Mips::LL_MMR6 : Mips::LL_MM; 1419 SC = Subtarget.hasMips32r6() ? Mips::SC_MMR6 : Mips::SC_MM; 1420 } else { 1421 LL = Subtarget.hasMips32r6() 1422 ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6) 1423 : (ArePtrs64bit ? Mips::LL64 : Mips::LL); 1424 SC = Subtarget.hasMips32r6() 1425 ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6) 1426 : (ArePtrs64bit ? Mips::SC64 : Mips::SC); 1427 } 1428 1429 AND = Mips::AND; 1430 NOR = Mips::NOR; 1431 ZERO = Mips::ZERO; 1432 BEQ = Mips::BEQ; 1433 } else { 1434 LL = Subtarget.hasMips64r6() ? Mips::LLD_R6 : Mips::LLD; 1435 SC = Subtarget.hasMips64r6() ? Mips::SCD_R6 : Mips::SCD; 1436 AND = Mips::AND64; 1437 NOR = Mips::NOR64; 1438 ZERO = Mips::ZERO_64; 1439 BEQ = Mips::BEQ64; 1440 } 1441 1442 unsigned OldVal = MI.getOperand(0).getReg(); 1443 unsigned Ptr = MI.getOperand(1).getReg(); 1444 unsigned Incr = MI.getOperand(2).getReg(); 1445 1446 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1447 unsigned AndRes = RegInfo.createVirtualRegister(RC); 1448 unsigned Success = RegInfo.createVirtualRegister(RC); 1449 1450 // insert new blocks after the current block 1451 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1452 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1453 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1454 MachineFunction::iterator It = ++BB->getIterator(); 1455 MF->insert(It, loopMBB); 1456 MF->insert(It, exitMBB); 1457 1458 // Transfer the remainder of BB and its successor edges to exitMBB. 1459 exitMBB->splice(exitMBB->begin(), BB, 1460 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1461 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1462 1463 // thisMBB: 1464 // ... 1465 // fallthrough --> loopMBB 1466 BB->addSuccessor(loopMBB); 1467 loopMBB->addSuccessor(loopMBB); 1468 loopMBB->addSuccessor(exitMBB); 1469 1470 // loopMBB: 1471 // ll oldval, 0(ptr) 1472 // <binop> storeval, oldval, incr 1473 // sc success, storeval, 0(ptr) 1474 // beq success, $0, loopMBB 1475 BB = loopMBB; 1476 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0); 1477 if (Nand) { 1478 // and andres, oldval, incr 1479 // nor storeval, $0, andres 1480 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr); 1481 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes); 1482 } else if (BinOpcode) { 1483 // <binop> storeval, oldval, incr 1484 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr); 1485 } else { 1486 StoreVal = Incr; 1487 } 1488 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0); 1489 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB); 1490 1491 MI.eraseFromParent(); // The instruction is gone now. 1492 1493 return exitMBB; 1494 } 1495 1496 MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg( 1497 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg, 1498 unsigned SrcReg) const { 1499 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1500 const DebugLoc &DL = MI.getDebugLoc(); 1501 1502 if (Subtarget.hasMips32r2() && Size == 1) { 1503 BuildMI(BB, DL, TII->get(Mips::SEB), DstReg).addReg(SrcReg); 1504 return BB; 1505 } 1506 1507 if (Subtarget.hasMips32r2() && Size == 2) { 1508 BuildMI(BB, DL, TII->get(Mips::SEH), DstReg).addReg(SrcReg); 1509 return BB; 1510 } 1511 1512 MachineFunction *MF = BB->getParent(); 1513 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1514 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1515 unsigned ScrReg = RegInfo.createVirtualRegister(RC); 1516 1517 assert(Size < 32); 1518 int64_t ShiftImm = 32 - (Size * 8); 1519 1520 BuildMI(BB, DL, TII->get(Mips::SLL), ScrReg).addReg(SrcReg).addImm(ShiftImm); 1521 BuildMI(BB, DL, TII->get(Mips::SRA), DstReg).addReg(ScrReg).addImm(ShiftImm); 1522 1523 return BB; 1524 } 1525 1526 MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword( 1527 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned BinOpcode, 1528 bool Nand) const { 1529 assert((Size == 1 || Size == 2) && 1530 "Unsupported size for EmitAtomicBinaryPartial."); 1531 1532 MachineFunction *MF = BB->getParent(); 1533 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1534 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1535 const bool ArePtrs64bit = ABI.ArePtrs64bit(); 1536 const TargetRegisterClass *RCp = 1537 getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32); 1538 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1539 DebugLoc DL = MI.getDebugLoc(); 1540 1541 unsigned Dest = MI.getOperand(0).getReg(); 1542 unsigned Ptr = MI.getOperand(1).getReg(); 1543 unsigned Incr = MI.getOperand(2).getReg(); 1544 1545 unsigned AlignedAddr = RegInfo.createVirtualRegister(RCp); 1546 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1547 unsigned Mask = RegInfo.createVirtualRegister(RC); 1548 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1549 unsigned NewVal = RegInfo.createVirtualRegister(RC); 1550 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1551 unsigned Incr2 = RegInfo.createVirtualRegister(RC); 1552 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RCp); 1553 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1554 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1555 unsigned AndRes = RegInfo.createVirtualRegister(RC); 1556 unsigned BinOpRes = RegInfo.createVirtualRegister(RC); 1557 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1558 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1559 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1560 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1561 unsigned Success = RegInfo.createVirtualRegister(RC); 1562 1563 unsigned LL, SC; 1564 if (isMicroMips) { 1565 LL = Subtarget.hasMips32r6() ? Mips::LL_MMR6 : Mips::LL_MM; 1566 SC = Subtarget.hasMips32r6() ? Mips::SC_MMR6 : Mips::SC_MM; 1567 } else { 1568 LL = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6) 1569 : (ArePtrs64bit ? Mips::LL64 : Mips::LL); 1570 SC = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6) 1571 : (ArePtrs64bit ? Mips::SC64 : Mips::SC); 1572 } 1573 1574 // insert new blocks after the current block 1575 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1576 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1577 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1578 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1579 MachineFunction::iterator It = ++BB->getIterator(); 1580 MF->insert(It, loopMBB); 1581 MF->insert(It, sinkMBB); 1582 MF->insert(It, exitMBB); 1583 1584 // Transfer the remainder of BB and its successor edges to exitMBB. 1585 exitMBB->splice(exitMBB->begin(), BB, 1586 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1587 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1588 1589 BB->addSuccessor(loopMBB); 1590 loopMBB->addSuccessor(loopMBB); 1591 loopMBB->addSuccessor(sinkMBB); 1592 sinkMBB->addSuccessor(exitMBB); 1593 1594 // thisMBB: 1595 // addiu masklsb2,$0,-4 # 0xfffffffc 1596 // and alignedaddr,ptr,masklsb2 1597 // andi ptrlsb2,ptr,3 1598 // sll shiftamt,ptrlsb2,3 1599 // ori maskupper,$0,255 # 0xff 1600 // sll mask,maskupper,shiftamt 1601 // nor mask2,$0,mask 1602 // sll incr2,incr,shiftamt 1603 1604 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1605 BuildMI(BB, DL, TII->get(ABI.GetPtrAddiuOp()), MaskLSB2) 1606 .addReg(ABI.GetNullPtr()).addImm(-4); 1607 BuildMI(BB, DL, TII->get(ABI.GetPtrAndOp()), AlignedAddr) 1608 .addReg(Ptr).addReg(MaskLSB2); 1609 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2) 1610 .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3); 1611 if (Subtarget.isLittle()) { 1612 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1613 } else { 1614 unsigned Off = RegInfo.createVirtualRegister(RC); 1615 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1616 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1617 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1618 } 1619 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1620 .addReg(Mips::ZERO).addImm(MaskImm); 1621 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1622 .addReg(MaskUpper).addReg(ShiftAmt); 1623 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1624 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt); 1625 1626 // atomic.load.binop 1627 // loopMBB: 1628 // ll oldval,0(alignedaddr) 1629 // binop binopres,oldval,incr2 1630 // and newval,binopres,mask 1631 // and maskedoldval0,oldval,mask2 1632 // or storeval,maskedoldval0,newval 1633 // sc success,storeval,0(alignedaddr) 1634 // beq success,$0,loopMBB 1635 1636 // atomic.swap 1637 // loopMBB: 1638 // ll oldval,0(alignedaddr) 1639 // and newval,incr2,mask 1640 // and maskedoldval0,oldval,mask2 1641 // or storeval,maskedoldval0,newval 1642 // sc success,storeval,0(alignedaddr) 1643 // beq success,$0,loopMBB 1644 1645 BB = loopMBB; 1646 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0); 1647 if (Nand) { 1648 // and andres, oldval, incr2 1649 // nor binopres, $0, andres 1650 // and newval, binopres, mask 1651 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2); 1652 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes) 1653 .addReg(Mips::ZERO).addReg(AndRes); 1654 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1655 } else if (BinOpcode) { 1656 // <binop> binopres, oldval, incr2 1657 // and newval, binopres, mask 1658 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2); 1659 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1660 } else { // atomic.swap 1661 // and newval, incr2, mask 1662 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask); 1663 } 1664 1665 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1666 .addReg(OldVal).addReg(Mask2); 1667 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1668 .addReg(MaskedOldVal0).addReg(NewVal); 1669 BuildMI(BB, DL, TII->get(SC), Success) 1670 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1671 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1672 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB); 1673 1674 // sinkMBB: 1675 // and maskedoldval1,oldval,mask 1676 // srl srlres,maskedoldval1,shiftamt 1677 // sign_extend dest,srlres 1678 BB = sinkMBB; 1679 1680 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1681 .addReg(OldVal).addReg(Mask); 1682 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1683 .addReg(MaskedOldVal1).addReg(ShiftAmt); 1684 BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes); 1685 1686 MI.eraseFromParent(); // The instruction is gone now. 1687 1688 return exitMBB; 1689 } 1690 1691 MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwap(MachineInstr &MI, 1692 MachineBasicBlock *BB, 1693 unsigned Size) const { 1694 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap."); 1695 1696 MachineFunction *MF = BB->getParent(); 1697 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1698 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 1699 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1700 const bool ArePtrs64bit = ABI.ArePtrs64bit(); 1701 DebugLoc DL = MI.getDebugLoc(); 1702 unsigned LL, SC, ZERO, BNE, BEQ; 1703 1704 if (Size == 4) { 1705 if (isMicroMips) { 1706 LL = Subtarget.hasMips32r6() ? Mips::LL_MMR6 : Mips::LL_MM; 1707 SC = Subtarget.hasMips32r6() ? Mips::SC_MMR6 : Mips::SC_MM; 1708 } else { 1709 LL = Subtarget.hasMips32r6() 1710 ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6) 1711 : (ArePtrs64bit ? Mips::LL64 : Mips::LL); 1712 SC = Subtarget.hasMips32r6() 1713 ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6) 1714 : (ArePtrs64bit ? Mips::SC64 : Mips::SC); 1715 } 1716 1717 ZERO = Mips::ZERO; 1718 BNE = Mips::BNE; 1719 BEQ = Mips::BEQ; 1720 } else { 1721 LL = Subtarget.hasMips64r6() ? Mips::LLD_R6 : Mips::LLD; 1722 SC = Subtarget.hasMips64r6() ? Mips::SCD_R6 : Mips::SCD; 1723 ZERO = Mips::ZERO_64; 1724 BNE = Mips::BNE64; 1725 BEQ = Mips::BEQ64; 1726 } 1727 1728 unsigned Dest = MI.getOperand(0).getReg(); 1729 unsigned Ptr = MI.getOperand(1).getReg(); 1730 unsigned OldVal = MI.getOperand(2).getReg(); 1731 unsigned NewVal = MI.getOperand(3).getReg(); 1732 1733 unsigned Success = RegInfo.createVirtualRegister(RC); 1734 1735 // insert new blocks after the current block 1736 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1737 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1738 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1739 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1740 MachineFunction::iterator It = ++BB->getIterator(); 1741 MF->insert(It, loop1MBB); 1742 MF->insert(It, loop2MBB); 1743 MF->insert(It, exitMBB); 1744 1745 // Transfer the remainder of BB and its successor edges to exitMBB. 1746 exitMBB->splice(exitMBB->begin(), BB, 1747 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1748 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1749 1750 // thisMBB: 1751 // ... 1752 // fallthrough --> loop1MBB 1753 BB->addSuccessor(loop1MBB); 1754 loop1MBB->addSuccessor(exitMBB); 1755 loop1MBB->addSuccessor(loop2MBB); 1756 loop2MBB->addSuccessor(loop1MBB); 1757 loop2MBB->addSuccessor(exitMBB); 1758 1759 // loop1MBB: 1760 // ll dest, 0(ptr) 1761 // bne dest, oldval, exitMBB 1762 BB = loop1MBB; 1763 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0); 1764 BuildMI(BB, DL, TII->get(BNE)) 1765 .addReg(Dest).addReg(OldVal).addMBB(exitMBB); 1766 1767 // loop2MBB: 1768 // sc success, newval, 0(ptr) 1769 // beq success, $0, loop1MBB 1770 BB = loop2MBB; 1771 BuildMI(BB, DL, TII->get(SC), Success) 1772 .addReg(NewVal).addReg(Ptr).addImm(0); 1773 BuildMI(BB, DL, TII->get(BEQ)) 1774 .addReg(Success).addReg(ZERO).addMBB(loop1MBB); 1775 1776 MI.eraseFromParent(); // The instruction is gone now. 1777 1778 return exitMBB; 1779 } 1780 1781 MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwapPartword( 1782 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const { 1783 assert((Size == 1 || Size == 2) && 1784 "Unsupported size for EmitAtomicCmpSwapPartial."); 1785 1786 MachineFunction *MF = BB->getParent(); 1787 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1788 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1789 const bool ArePtrs64bit = ABI.ArePtrs64bit(); 1790 const TargetRegisterClass *RCp = 1791 getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32); 1792 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1793 DebugLoc DL = MI.getDebugLoc(); 1794 1795 unsigned Dest = MI.getOperand(0).getReg(); 1796 unsigned Ptr = MI.getOperand(1).getReg(); 1797 unsigned CmpVal = MI.getOperand(2).getReg(); 1798 unsigned NewVal = MI.getOperand(3).getReg(); 1799 1800 unsigned AlignedAddr = RegInfo.createVirtualRegister(RCp); 1801 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1802 unsigned Mask = RegInfo.createVirtualRegister(RC); 1803 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1804 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC); 1805 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1806 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1807 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC); 1808 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RCp); 1809 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1810 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1811 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC); 1812 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC); 1813 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1814 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1815 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1816 unsigned Success = RegInfo.createVirtualRegister(RC); 1817 unsigned LL, SC; 1818 1819 if (isMicroMips) { 1820 LL = Subtarget.hasMips32r6() ? Mips::LL_MMR6 : Mips::LL_MM; 1821 SC = Subtarget.hasMips32r6() ? Mips::SC_MMR6 : Mips::SC_MM; 1822 } else { 1823 LL = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6) 1824 : (ArePtrs64bit ? Mips::LL64 : Mips::LL); 1825 SC = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6) 1826 : (ArePtrs64bit ? Mips::SC64 : Mips::SC); 1827 } 1828 1829 // insert new blocks after the current block 1830 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1831 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1832 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1833 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1834 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1835 MachineFunction::iterator It = ++BB->getIterator(); 1836 MF->insert(It, loop1MBB); 1837 MF->insert(It, loop2MBB); 1838 MF->insert(It, sinkMBB); 1839 MF->insert(It, exitMBB); 1840 1841 // Transfer the remainder of BB and its successor edges to exitMBB. 1842 exitMBB->splice(exitMBB->begin(), BB, 1843 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1844 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1845 1846 BB->addSuccessor(loop1MBB); 1847 loop1MBB->addSuccessor(sinkMBB); 1848 loop1MBB->addSuccessor(loop2MBB); 1849 loop2MBB->addSuccessor(loop1MBB); 1850 loop2MBB->addSuccessor(sinkMBB); 1851 sinkMBB->addSuccessor(exitMBB); 1852 1853 // FIXME: computation of newval2 can be moved to loop2MBB. 1854 // thisMBB: 1855 // addiu masklsb2,$0,-4 # 0xfffffffc 1856 // and alignedaddr,ptr,masklsb2 1857 // andi ptrlsb2,ptr,3 1858 // xori ptrlsb2,ptrlsb2,3 # Only for BE 1859 // sll shiftamt,ptrlsb2,3 1860 // ori maskupper,$0,255 # 0xff 1861 // sll mask,maskupper,shiftamt 1862 // nor mask2,$0,mask 1863 // andi maskedcmpval,cmpval,255 1864 // sll shiftedcmpval,maskedcmpval,shiftamt 1865 // andi maskednewval,newval,255 1866 // sll shiftednewval,maskednewval,shiftamt 1867 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1868 BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::DADDiu : Mips::ADDiu), MaskLSB2) 1869 .addReg(ABI.GetNullPtr()).addImm(-4); 1870 BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::AND64 : Mips::AND), AlignedAddr) 1871 .addReg(Ptr).addReg(MaskLSB2); 1872 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2) 1873 .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3); 1874 if (Subtarget.isLittle()) { 1875 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1876 } else { 1877 unsigned Off = RegInfo.createVirtualRegister(RC); 1878 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1879 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1880 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1881 } 1882 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1883 .addReg(Mips::ZERO).addImm(MaskImm); 1884 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1885 .addReg(MaskUpper).addReg(ShiftAmt); 1886 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1887 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal) 1888 .addReg(CmpVal).addImm(MaskImm); 1889 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal) 1890 .addReg(MaskedCmpVal).addReg(ShiftAmt); 1891 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal) 1892 .addReg(NewVal).addImm(MaskImm); 1893 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal) 1894 .addReg(MaskedNewVal).addReg(ShiftAmt); 1895 1896 // loop1MBB: 1897 // ll oldval,0(alginedaddr) 1898 // and maskedoldval0,oldval,mask 1899 // bne maskedoldval0,shiftedcmpval,sinkMBB 1900 BB = loop1MBB; 1901 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0); 1902 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1903 .addReg(OldVal).addReg(Mask); 1904 BuildMI(BB, DL, TII->get(Mips::BNE)) 1905 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB); 1906 1907 // loop2MBB: 1908 // and maskedoldval1,oldval,mask2 1909 // or storeval,maskedoldval1,shiftednewval 1910 // sc success,storeval,0(alignedaddr) 1911 // beq success,$0,loop1MBB 1912 BB = loop2MBB; 1913 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1914 .addReg(OldVal).addReg(Mask2); 1915 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1916 .addReg(MaskedOldVal1).addReg(ShiftedNewVal); 1917 BuildMI(BB, DL, TII->get(SC), Success) 1918 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1919 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1920 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB); 1921 1922 // sinkMBB: 1923 // srl srlres,maskedoldval0,shiftamt 1924 // sign_extend dest,srlres 1925 BB = sinkMBB; 1926 1927 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1928 .addReg(MaskedOldVal0).addReg(ShiftAmt); 1929 BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes); 1930 1931 MI.eraseFromParent(); // The instruction is gone now. 1932 1933 return exitMBB; 1934 } 1935 1936 SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const { 1937 // The first operand is the chain, the second is the condition, the third is 1938 // the block to branch to if the condition is true. 1939 SDValue Chain = Op.getOperand(0); 1940 SDValue Dest = Op.getOperand(2); 1941 SDLoc DL(Op); 1942 1943 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 1944 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1)); 1945 1946 // Return if flag is not set by a floating point comparison. 1947 if (CondRes.getOpcode() != MipsISD::FPCmp) 1948 return Op; 1949 1950 SDValue CCNode = CondRes.getOperand(2); 1951 Mips::CondCode CC = 1952 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue(); 1953 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T; 1954 SDValue BrCode = DAG.getConstant(Opc, DL, MVT::i32); 1955 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 1956 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode, 1957 FCC0, Dest, CondRes); 1958 } 1959 1960 SDValue MipsTargetLowering:: 1961 lowerSELECT(SDValue Op, SelectionDAG &DAG) const 1962 { 1963 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 1964 SDValue Cond = createFPCmp(DAG, Op.getOperand(0)); 1965 1966 // Return if flag is not set by a floating point comparison. 1967 if (Cond.getOpcode() != MipsISD::FPCmp) 1968 return Op; 1969 1970 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2), 1971 SDLoc(Op)); 1972 } 1973 1974 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const { 1975 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 1976 SDValue Cond = createFPCmp(DAG, Op); 1977 1978 assert(Cond.getOpcode() == MipsISD::FPCmp && 1979 "Floating point operand expected."); 1980 1981 SDLoc DL(Op); 1982 SDValue True = DAG.getConstant(1, DL, MVT::i32); 1983 SDValue False = DAG.getConstant(0, DL, MVT::i32); 1984 1985 return createCMovFP(DAG, Cond, True, False, DL); 1986 } 1987 1988 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op, 1989 SelectionDAG &DAG) const { 1990 EVT Ty = Op.getValueType(); 1991 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op); 1992 const GlobalValue *GV = N->getGlobal(); 1993 1994 if (!isPositionIndependent()) { 1995 const MipsTargetObjectFile *TLOF = 1996 static_cast<const MipsTargetObjectFile *>( 1997 getTargetMachine().getObjFileLowering()); 1998 const GlobalObject *GO = GV->getBaseObject(); 1999 if (GO && TLOF->IsGlobalInSmallSection(GO, getTargetMachine())) 2000 // %gp_rel relocation 2001 return getAddrGPRel(N, SDLoc(N), Ty, DAG, ABI.IsN64()); 2002 2003 // %hi/%lo relocation 2004 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2005 // %highest/%higher/%hi/%lo relocation 2006 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2007 } 2008 2009 // Every other architecture would use shouldAssumeDSOLocal in here, but 2010 // mips is special. 2011 // * In PIC code mips requires got loads even for local statics! 2012 // * To save on got entries, for local statics the got entry contains the 2013 // page and an additional add instruction takes care of the low bits. 2014 // * It is legal to access a hidden symbol with a non hidden undefined, 2015 // so one cannot guarantee that all access to a hidden symbol will know 2016 // it is hidden. 2017 // * Mips linkers don't support creating a page and a full got entry for 2018 // the same symbol. 2019 // * Given all that, we have to use a full got entry for hidden symbols :-( 2020 if (GV->hasLocalLinkage()) 2021 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2022 2023 if (LargeGOT) 2024 return getAddrGlobalLargeGOT( 2025 N, SDLoc(N), Ty, DAG, MipsII::MO_GOT_HI16, MipsII::MO_GOT_LO16, 2026 DAG.getEntryNode(), 2027 MachinePointerInfo::getGOT(DAG.getMachineFunction())); 2028 2029 return getAddrGlobal( 2030 N, SDLoc(N), Ty, DAG, 2031 (ABI.IsN32() || ABI.IsN64()) ? MipsII::MO_GOT_DISP : MipsII::MO_GOT, 2032 DAG.getEntryNode(), MachinePointerInfo::getGOT(DAG.getMachineFunction())); 2033 } 2034 2035 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op, 2036 SelectionDAG &DAG) const { 2037 BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op); 2038 EVT Ty = Op.getValueType(); 2039 2040 if (!isPositionIndependent()) 2041 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2042 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2043 2044 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2045 } 2046 2047 SDValue MipsTargetLowering:: 2048 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const 2049 { 2050 // If the relocation model is PIC, use the General Dynamic TLS Model or 2051 // Local Dynamic TLS model, otherwise use the Initial Exec or 2052 // Local Exec TLS Model. 2053 2054 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 2055 if (DAG.getTarget().useEmulatedTLS()) 2056 return LowerToTLSEmulatedModel(GA, DAG); 2057 2058 SDLoc DL(GA); 2059 const GlobalValue *GV = GA->getGlobal(); 2060 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 2061 2062 TLSModel::Model model = getTargetMachine().getTLSModel(GV); 2063 2064 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { 2065 // General Dynamic and Local Dynamic TLS Model. 2066 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM 2067 : MipsII::MO_TLSGD; 2068 2069 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag); 2070 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, 2071 getGlobalReg(DAG, PtrVT), TGA); 2072 unsigned PtrSize = PtrVT.getSizeInBits(); 2073 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize); 2074 2075 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT); 2076 2077 ArgListTy Args; 2078 ArgListEntry Entry; 2079 Entry.Node = Argument; 2080 Entry.Ty = PtrTy; 2081 Args.push_back(Entry); 2082 2083 TargetLowering::CallLoweringInfo CLI(DAG); 2084 CLI.setDebugLoc(DL) 2085 .setChain(DAG.getEntryNode()) 2086 .setLibCallee(CallingConv::C, PtrTy, TlsGetAddr, std::move(Args)); 2087 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 2088 2089 SDValue Ret = CallResult.first; 2090 2091 if (model != TLSModel::LocalDynamic) 2092 return Ret; 2093 2094 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2095 MipsII::MO_DTPREL_HI); 2096 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 2097 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2098 MipsII::MO_DTPREL_LO); 2099 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 2100 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret); 2101 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo); 2102 } 2103 2104 SDValue Offset; 2105 if (model == TLSModel::InitialExec) { 2106 // Initial Exec TLS Model 2107 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2108 MipsII::MO_GOTTPREL); 2109 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT), 2110 TGA); 2111 Offset = 2112 DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), TGA, MachinePointerInfo()); 2113 } else { 2114 // Local Exec TLS Model 2115 assert(model == TLSModel::LocalExec); 2116 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2117 MipsII::MO_TPREL_HI); 2118 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2119 MipsII::MO_TPREL_LO); 2120 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 2121 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 2122 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo); 2123 } 2124 2125 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT); 2126 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset); 2127 } 2128 2129 SDValue MipsTargetLowering:: 2130 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const 2131 { 2132 JumpTableSDNode *N = cast<JumpTableSDNode>(Op); 2133 EVT Ty = Op.getValueType(); 2134 2135 if (!isPositionIndependent()) 2136 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2137 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2138 2139 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2140 } 2141 2142 SDValue MipsTargetLowering:: 2143 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const 2144 { 2145 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op); 2146 EVT Ty = Op.getValueType(); 2147 2148 if (!isPositionIndependent()) { 2149 const MipsTargetObjectFile *TLOF = 2150 static_cast<const MipsTargetObjectFile *>( 2151 getTargetMachine().getObjFileLowering()); 2152 2153 if (TLOF->IsConstantInSmallSection(DAG.getDataLayout(), N->getConstVal(), 2154 getTargetMachine())) 2155 // %gp_rel relocation 2156 return getAddrGPRel(N, SDLoc(N), Ty, DAG, ABI.IsN64()); 2157 2158 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2159 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2160 } 2161 2162 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2163 } 2164 2165 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const { 2166 MachineFunction &MF = DAG.getMachineFunction(); 2167 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 2168 2169 SDLoc DL(Op); 2170 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), 2171 getPointerTy(MF.getDataLayout())); 2172 2173 // vastart just stores the address of the VarArgsFrameIndex slot into the 2174 // memory location argument. 2175 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 2176 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1), 2177 MachinePointerInfo(SV)); 2178 } 2179 2180 SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const { 2181 SDNode *Node = Op.getNode(); 2182 EVT VT = Node->getValueType(0); 2183 SDValue Chain = Node->getOperand(0); 2184 SDValue VAListPtr = Node->getOperand(1); 2185 unsigned Align = Node->getConstantOperandVal(3); 2186 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue(); 2187 SDLoc DL(Node); 2188 unsigned ArgSlotSizeInBytes = (ABI.IsN32() || ABI.IsN64()) ? 8 : 4; 2189 2190 SDValue VAListLoad = DAG.getLoad(getPointerTy(DAG.getDataLayout()), DL, Chain, 2191 VAListPtr, MachinePointerInfo(SV)); 2192 SDValue VAList = VAListLoad; 2193 2194 // Re-align the pointer if necessary. 2195 // It should only ever be necessary for 64-bit types on O32 since the minimum 2196 // argument alignment is the same as the maximum type alignment for N32/N64. 2197 // 2198 // FIXME: We currently align too often. The code generator doesn't notice 2199 // when the pointer is still aligned from the last va_arg (or pair of 2200 // va_args for the i64 on O32 case). 2201 if (Align > getMinStackArgumentAlignment()) { 2202 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2"); 2203 2204 VAList = DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList, 2205 DAG.getConstant(Align - 1, DL, VAList.getValueType())); 2206 2207 VAList = DAG.getNode(ISD::AND, DL, VAList.getValueType(), VAList, 2208 DAG.getConstant(-(int64_t)Align, DL, 2209 VAList.getValueType())); 2210 } 2211 2212 // Increment the pointer, VAList, to the next vaarg. 2213 auto &TD = DAG.getDataLayout(); 2214 unsigned ArgSizeInBytes = 2215 TD.getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())); 2216 SDValue Tmp3 = 2217 DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList, 2218 DAG.getConstant(alignTo(ArgSizeInBytes, ArgSlotSizeInBytes), 2219 DL, VAList.getValueType())); 2220 // Store the incremented VAList to the legalized pointer 2221 Chain = DAG.getStore(VAListLoad.getValue(1), DL, Tmp3, VAListPtr, 2222 MachinePointerInfo(SV)); 2223 2224 // In big-endian mode we must adjust the pointer when the load size is smaller 2225 // than the argument slot size. We must also reduce the known alignment to 2226 // match. For example in the N64 ABI, we must add 4 bytes to the offset to get 2227 // the correct half of the slot, and reduce the alignment from 8 (slot 2228 // alignment) down to 4 (type alignment). 2229 if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) { 2230 unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes; 2231 VAList = DAG.getNode(ISD::ADD, DL, VAListPtr.getValueType(), VAList, 2232 DAG.getIntPtrConstant(Adjustment, DL)); 2233 } 2234 // Load the actual argument out of the pointer VAList 2235 return DAG.getLoad(VT, DL, Chain, VAList, MachinePointerInfo()); 2236 } 2237 2238 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, 2239 bool HasExtractInsert) { 2240 EVT TyX = Op.getOperand(0).getValueType(); 2241 EVT TyY = Op.getOperand(1).getValueType(); 2242 SDLoc DL(Op); 2243 SDValue Const1 = DAG.getConstant(1, DL, MVT::i32); 2244 SDValue Const31 = DAG.getConstant(31, DL, MVT::i32); 2245 SDValue Res; 2246 2247 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 2248 // to i32. 2249 SDValue X = (TyX == MVT::f32) ? 2250 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) : 2251 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 2252 Const1); 2253 SDValue Y = (TyY == MVT::f32) ? 2254 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) : 2255 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1), 2256 Const1); 2257 2258 if (HasExtractInsert) { 2259 // ext E, Y, 31, 1 ; extract bit31 of Y 2260 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X 2261 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1); 2262 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X); 2263 } else { 2264 // sll SllX, X, 1 2265 // srl SrlX, SllX, 1 2266 // srl SrlY, Y, 31 2267 // sll SllY, SrlX, 31 2268 // or Or, SrlX, SllY 2269 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 2270 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 2271 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31); 2272 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31); 2273 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY); 2274 } 2275 2276 if (TyX == MVT::f32) 2277 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res); 2278 2279 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2280 Op.getOperand(0), 2281 DAG.getConstant(0, DL, MVT::i32)); 2282 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 2283 } 2284 2285 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, 2286 bool HasExtractInsert) { 2287 unsigned WidthX = Op.getOperand(0).getValueSizeInBits(); 2288 unsigned WidthY = Op.getOperand(1).getValueSizeInBits(); 2289 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY); 2290 SDLoc DL(Op); 2291 SDValue Const1 = DAG.getConstant(1, DL, MVT::i32); 2292 2293 // Bitcast to integer nodes. 2294 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0)); 2295 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1)); 2296 2297 if (HasExtractInsert) { 2298 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y 2299 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X 2300 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y, 2301 DAG.getConstant(WidthY - 1, DL, MVT::i32), Const1); 2302 2303 if (WidthX > WidthY) 2304 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E); 2305 else if (WidthY > WidthX) 2306 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E); 2307 2308 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E, 2309 DAG.getConstant(WidthX - 1, DL, MVT::i32), Const1, 2310 X); 2311 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I); 2312 } 2313 2314 // (d)sll SllX, X, 1 2315 // (d)srl SrlX, SllX, 1 2316 // (d)srl SrlY, Y, width(Y)-1 2317 // (d)sll SllY, SrlX, width(Y)-1 2318 // or Or, SrlX, SllY 2319 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1); 2320 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1); 2321 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y, 2322 DAG.getConstant(WidthY - 1, DL, MVT::i32)); 2323 2324 if (WidthX > WidthY) 2325 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY); 2326 else if (WidthY > WidthX) 2327 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY); 2328 2329 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY, 2330 DAG.getConstant(WidthX - 1, DL, MVT::i32)); 2331 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY); 2332 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or); 2333 } 2334 2335 SDValue 2336 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 2337 if (Subtarget.isGP64bit()) 2338 return lowerFCOPYSIGN64(Op, DAG, Subtarget.hasExtractInsert()); 2339 2340 return lowerFCOPYSIGN32(Op, DAG, Subtarget.hasExtractInsert()); 2341 } 2342 2343 SDValue MipsTargetLowering:: 2344 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 2345 // check the depth 2346 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 2347 "Frame address can only be determined for current frame."); 2348 2349 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 2350 MFI.setFrameAddressIsTaken(true); 2351 EVT VT = Op.getValueType(); 2352 SDLoc DL(Op); 2353 SDValue FrameAddr = DAG.getCopyFromReg( 2354 DAG.getEntryNode(), DL, ABI.IsN64() ? Mips::FP_64 : Mips::FP, VT); 2355 return FrameAddr; 2356 } 2357 2358 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op, 2359 SelectionDAG &DAG) const { 2360 if (verifyReturnAddressArgumentIsConstant(Op, DAG)) 2361 return SDValue(); 2362 2363 // check the depth 2364 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 2365 "Return address can be determined only for current frame."); 2366 2367 MachineFunction &MF = DAG.getMachineFunction(); 2368 MachineFrameInfo &MFI = MF.getFrameInfo(); 2369 MVT VT = Op.getSimpleValueType(); 2370 unsigned RA = ABI.IsN64() ? Mips::RA_64 : Mips::RA; 2371 MFI.setReturnAddressIsTaken(true); 2372 2373 // Return RA, which contains the return address. Mark it an implicit live-in. 2374 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT)); 2375 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT); 2376 } 2377 2378 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is 2379 // generated from __builtin_eh_return (offset, handler) 2380 // The effect of this is to adjust the stack pointer by "offset" 2381 // and then branch to "handler". 2382 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG) 2383 const { 2384 MachineFunction &MF = DAG.getMachineFunction(); 2385 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2386 2387 MipsFI->setCallsEhReturn(); 2388 SDValue Chain = Op.getOperand(0); 2389 SDValue Offset = Op.getOperand(1); 2390 SDValue Handler = Op.getOperand(2); 2391 SDLoc DL(Op); 2392 EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32; 2393 2394 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and 2395 // EH_RETURN nodes, so that instructions are emitted back-to-back. 2396 unsigned OffsetReg = ABI.IsN64() ? Mips::V1_64 : Mips::V1; 2397 unsigned AddrReg = ABI.IsN64() ? Mips::V0_64 : Mips::V0; 2398 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue()); 2399 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1)); 2400 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain, 2401 DAG.getRegister(OffsetReg, Ty), 2402 DAG.getRegister(AddrReg, getPointerTy(MF.getDataLayout())), 2403 Chain.getValue(1)); 2404 } 2405 2406 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op, 2407 SelectionDAG &DAG) const { 2408 // FIXME: Need pseudo-fence for 'singlethread' fences 2409 // FIXME: Set SType for weaker fences where supported/appropriate. 2410 unsigned SType = 0; 2411 SDLoc DL(Op); 2412 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0), 2413 DAG.getConstant(SType, DL, MVT::i32)); 2414 } 2415 2416 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op, 2417 SelectionDAG &DAG) const { 2418 SDLoc DL(Op); 2419 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; 2420 2421 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 2422 SDValue Shamt = Op.getOperand(2); 2423 // if shamt < (VT.bits): 2424 // lo = (shl lo, shamt) 2425 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt)) 2426 // else: 2427 // lo = 0 2428 // hi = (shl lo, shamt[4:0]) 2429 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 2430 DAG.getConstant(-1, DL, MVT::i32)); 2431 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, 2432 DAG.getConstant(1, DL, VT)); 2433 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, Not); 2434 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt); 2435 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo); 2436 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt); 2437 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 2438 DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32)); 2439 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, 2440 DAG.getConstant(0, DL, VT), ShiftLeftLo); 2441 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftLeftLo, Or); 2442 2443 SDValue Ops[2] = {Lo, Hi}; 2444 return DAG.getMergeValues(Ops, DL); 2445 } 2446 2447 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, 2448 bool IsSRA) const { 2449 SDLoc DL(Op); 2450 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 2451 SDValue Shamt = Op.getOperand(2); 2452 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; 2453 2454 // if shamt < (VT.bits): 2455 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt)) 2456 // if isSRA: 2457 // hi = (sra hi, shamt) 2458 // else: 2459 // hi = (srl hi, shamt) 2460 // else: 2461 // if isSRA: 2462 // lo = (sra hi, shamt[4:0]) 2463 // hi = (sra hi, 31) 2464 // else: 2465 // lo = (srl hi, shamt[4:0]) 2466 // hi = 0 2467 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 2468 DAG.getConstant(-1, DL, MVT::i32)); 2469 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, VT, Hi, 2470 DAG.getConstant(1, DL, VT)); 2471 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, ShiftLeft1Hi, Not); 2472 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt); 2473 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo); 2474 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, 2475 DL, VT, Hi, Shamt); 2476 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 2477 DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32)); 2478 SDValue Ext = DAG.getNode(ISD::SRA, DL, VT, Hi, 2479 DAG.getConstant(VT.getSizeInBits() - 1, DL, VT)); 2480 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftRightHi, Or); 2481 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, 2482 IsSRA ? Ext : DAG.getConstant(0, DL, VT), ShiftRightHi); 2483 2484 SDValue Ops[2] = {Lo, Hi}; 2485 return DAG.getMergeValues(Ops, DL); 2486 } 2487 2488 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD, 2489 SDValue Chain, SDValue Src, unsigned Offset) { 2490 SDValue Ptr = LD->getBasePtr(); 2491 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT(); 2492 EVT BasePtrVT = Ptr.getValueType(); 2493 SDLoc DL(LD); 2494 SDVTList VTList = DAG.getVTList(VT, MVT::Other); 2495 2496 if (Offset) 2497 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2498 DAG.getConstant(Offset, DL, BasePtrVT)); 2499 2500 SDValue Ops[] = { Chain, Ptr, Src }; 2501 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT, 2502 LD->getMemOperand()); 2503 } 2504 2505 // Expand an unaligned 32 or 64-bit integer load node. 2506 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const { 2507 LoadSDNode *LD = cast<LoadSDNode>(Op); 2508 EVT MemVT = LD->getMemoryVT(); 2509 2510 if (Subtarget.systemSupportsUnalignedAccess()) 2511 return Op; 2512 2513 // Return if load is aligned or if MemVT is neither i32 nor i64. 2514 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) || 2515 ((MemVT != MVT::i32) && (MemVT != MVT::i64))) 2516 return SDValue(); 2517 2518 bool IsLittle = Subtarget.isLittle(); 2519 EVT VT = Op.getValueType(); 2520 ISD::LoadExtType ExtType = LD->getExtensionType(); 2521 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT); 2522 2523 assert((VT == MVT::i32) || (VT == MVT::i64)); 2524 2525 // Expand 2526 // (set dst, (i64 (load baseptr))) 2527 // to 2528 // (set tmp, (ldl (add baseptr, 7), undef)) 2529 // (set dst, (ldr baseptr, tmp)) 2530 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) { 2531 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef, 2532 IsLittle ? 7 : 0); 2533 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL, 2534 IsLittle ? 0 : 7); 2535 } 2536 2537 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef, 2538 IsLittle ? 3 : 0); 2539 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL, 2540 IsLittle ? 0 : 3); 2541 2542 // Expand 2543 // (set dst, (i32 (load baseptr))) or 2544 // (set dst, (i64 (sextload baseptr))) or 2545 // (set dst, (i64 (extload baseptr))) 2546 // to 2547 // (set tmp, (lwl (add baseptr, 3), undef)) 2548 // (set dst, (lwr baseptr, tmp)) 2549 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) || 2550 (ExtType == ISD::EXTLOAD)) 2551 return LWR; 2552 2553 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD)); 2554 2555 // Expand 2556 // (set dst, (i64 (zextload baseptr))) 2557 // to 2558 // (set tmp0, (lwl (add baseptr, 3), undef)) 2559 // (set tmp1, (lwr baseptr, tmp0)) 2560 // (set tmp2, (shl tmp1, 32)) 2561 // (set dst, (srl tmp2, 32)) 2562 SDLoc DL(LD); 2563 SDValue Const32 = DAG.getConstant(32, DL, MVT::i32); 2564 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32); 2565 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32); 2566 SDValue Ops[] = { SRL, LWR.getValue(1) }; 2567 return DAG.getMergeValues(Ops, DL); 2568 } 2569 2570 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD, 2571 SDValue Chain, unsigned Offset) { 2572 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue(); 2573 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType(); 2574 SDLoc DL(SD); 2575 SDVTList VTList = DAG.getVTList(MVT::Other); 2576 2577 if (Offset) 2578 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2579 DAG.getConstant(Offset, DL, BasePtrVT)); 2580 2581 SDValue Ops[] = { Chain, Value, Ptr }; 2582 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT, 2583 SD->getMemOperand()); 2584 } 2585 2586 // Expand an unaligned 32 or 64-bit integer store node. 2587 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG, 2588 bool IsLittle) { 2589 SDValue Value = SD->getValue(), Chain = SD->getChain(); 2590 EVT VT = Value.getValueType(); 2591 2592 // Expand 2593 // (store val, baseptr) or 2594 // (truncstore val, baseptr) 2595 // to 2596 // (swl val, (add baseptr, 3)) 2597 // (swr val, baseptr) 2598 if ((VT == MVT::i32) || SD->isTruncatingStore()) { 2599 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain, 2600 IsLittle ? 3 : 0); 2601 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3); 2602 } 2603 2604 assert(VT == MVT::i64); 2605 2606 // Expand 2607 // (store val, baseptr) 2608 // to 2609 // (sdl val, (add baseptr, 7)) 2610 // (sdr val, baseptr) 2611 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0); 2612 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7); 2613 } 2614 2615 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr). 2616 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) { 2617 SDValue Val = SD->getValue(); 2618 2619 if (Val.getOpcode() != ISD::FP_TO_SINT) 2620 return SDValue(); 2621 2622 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits()); 2623 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy, 2624 Val.getOperand(0)); 2625 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(), 2626 SD->getPointerInfo(), SD->getAlignment(), 2627 SD->getMemOperand()->getFlags()); 2628 } 2629 2630 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const { 2631 StoreSDNode *SD = cast<StoreSDNode>(Op); 2632 EVT MemVT = SD->getMemoryVT(); 2633 2634 // Lower unaligned integer stores. 2635 if (!Subtarget.systemSupportsUnalignedAccess() && 2636 (SD->getAlignment() < MemVT.getSizeInBits() / 8) && 2637 ((MemVT == MVT::i32) || (MemVT == MVT::i64))) 2638 return lowerUnalignedIntStore(SD, DAG, Subtarget.isLittle()); 2639 2640 return lowerFP_TO_SINT_STORE(SD, DAG); 2641 } 2642 2643 SDValue MipsTargetLowering::lowerEH_DWARF_CFA(SDValue Op, 2644 SelectionDAG &DAG) const { 2645 2646 // Return a fixed StackObject with offset 0 which points to the old stack 2647 // pointer. 2648 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 2649 EVT ValTy = Op->getValueType(0); 2650 int FI = MFI.CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false); 2651 return DAG.getFrameIndex(FI, ValTy); 2652 } 2653 2654 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op, 2655 SelectionDAG &DAG) const { 2656 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits()); 2657 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy, 2658 Op.getOperand(0)); 2659 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc); 2660 } 2661 2662 //===----------------------------------------------------------------------===// 2663 // Calling Convention Implementation 2664 //===----------------------------------------------------------------------===// 2665 2666 //===----------------------------------------------------------------------===// 2667 // TODO: Implement a generic logic using tblgen that can support this. 2668 // Mips O32 ABI rules: 2669 // --- 2670 // i32 - Passed in A0, A1, A2, A3 and stack 2671 // f32 - Only passed in f32 registers if no int reg has been used yet to hold 2672 // an argument. Otherwise, passed in A1, A2, A3 and stack. 2673 // f64 - Only passed in two aliased f32 registers if no int reg has been used 2674 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is 2675 // not used, it must be shadowed. If only A3 is available, shadow it and 2676 // go to stack. 2677 // vXiX - Received as scalarized i32s, passed in A0 - A3 and the stack. 2678 // vXf32 - Passed in either a pair of registers {A0, A1}, {A2, A3} or {A0 - A3} 2679 // with the remainder spilled to the stack. 2680 // vXf64 - Passed in either {A0, A1, A2, A3} or {A2, A3} and in both cases 2681 // spilling the remainder to the stack. 2682 // 2683 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack. 2684 //===----------------------------------------------------------------------===// 2685 2686 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, 2687 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, 2688 CCState &State, ArrayRef<MCPhysReg> F64Regs) { 2689 const MipsSubtarget &Subtarget = static_cast<const MipsSubtarget &>( 2690 State.getMachineFunction().getSubtarget()); 2691 2692 static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 }; 2693 2694 const MipsCCState * MipsState = static_cast<MipsCCState *>(&State); 2695 2696 static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 }; 2697 2698 static const MCPhysReg FloatVectorIntRegs[] = { Mips::A0, Mips::A2 }; 2699 2700 // Do not process byval args here. 2701 if (ArgFlags.isByVal()) 2702 return true; 2703 2704 // Promote i8 and i16 2705 if (ArgFlags.isInReg() && !Subtarget.isLittle()) { 2706 if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) { 2707 LocVT = MVT::i32; 2708 if (ArgFlags.isSExt()) 2709 LocInfo = CCValAssign::SExtUpper; 2710 else if (ArgFlags.isZExt()) 2711 LocInfo = CCValAssign::ZExtUpper; 2712 else 2713 LocInfo = CCValAssign::AExtUpper; 2714 } 2715 } 2716 2717 // Promote i8 and i16 2718 if (LocVT == MVT::i8 || LocVT == MVT::i16) { 2719 LocVT = MVT::i32; 2720 if (ArgFlags.isSExt()) 2721 LocInfo = CCValAssign::SExt; 2722 else if (ArgFlags.isZExt()) 2723 LocInfo = CCValAssign::ZExt; 2724 else 2725 LocInfo = CCValAssign::AExt; 2726 } 2727 2728 unsigned Reg; 2729 2730 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following 2731 // is true: function is vararg, argument is 3rd or higher, there is previous 2732 // argument which is not f32 or f64. 2733 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 || 2734 State.getFirstUnallocated(F32Regs) != ValNo; 2735 unsigned OrigAlign = ArgFlags.getOrigAlign(); 2736 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8); 2737 bool isVectorFloat = MipsState->WasOriginalArgVectorFloat(ValNo); 2738 2739 // The MIPS vector ABI for floats passes them in a pair of registers 2740 if (ValVT == MVT::i32 && isVectorFloat) { 2741 // This is the start of an vector that was scalarized into an unknown number 2742 // of components. It doesn't matter how many there are. Allocate one of the 2743 // notional 8 byte aligned registers which map onto the argument stack, and 2744 // shadow the register lost to alignment requirements. 2745 if (ArgFlags.isSplit()) { 2746 Reg = State.AllocateReg(FloatVectorIntRegs); 2747 if (Reg == Mips::A2) 2748 State.AllocateReg(Mips::A1); 2749 else if (Reg == 0) 2750 State.AllocateReg(Mips::A3); 2751 } else { 2752 // If we're an intermediate component of the split, we can just attempt to 2753 // allocate a register directly. 2754 Reg = State.AllocateReg(IntRegs); 2755 } 2756 } else if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) { 2757 Reg = State.AllocateReg(IntRegs); 2758 // If this is the first part of an i64 arg, 2759 // the allocated register must be either A0 or A2. 2760 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3)) 2761 Reg = State.AllocateReg(IntRegs); 2762 LocVT = MVT::i32; 2763 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) { 2764 // Allocate int register and shadow next int register. If first 2765 // available register is Mips::A1 or Mips::A3, shadow it too. 2766 Reg = State.AllocateReg(IntRegs); 2767 if (Reg == Mips::A1 || Reg == Mips::A3) 2768 Reg = State.AllocateReg(IntRegs); 2769 State.AllocateReg(IntRegs); 2770 LocVT = MVT::i32; 2771 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) { 2772 // we are guaranteed to find an available float register 2773 if (ValVT == MVT::f32) { 2774 Reg = State.AllocateReg(F32Regs); 2775 // Shadow int register 2776 State.AllocateReg(IntRegs); 2777 } else { 2778 Reg = State.AllocateReg(F64Regs); 2779 // Shadow int registers 2780 unsigned Reg2 = State.AllocateReg(IntRegs); 2781 if (Reg2 == Mips::A1 || Reg2 == Mips::A3) 2782 State.AllocateReg(IntRegs); 2783 State.AllocateReg(IntRegs); 2784 } 2785 } else 2786 llvm_unreachable("Cannot handle this ValVT."); 2787 2788 if (!Reg) { 2789 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(), OrigAlign); 2790 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); 2791 } else 2792 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 2793 2794 return false; 2795 } 2796 2797 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT, 2798 MVT LocVT, CCValAssign::LocInfo LocInfo, 2799 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2800 static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 }; 2801 2802 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); 2803 } 2804 2805 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT, 2806 MVT LocVT, CCValAssign::LocInfo LocInfo, 2807 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2808 static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 }; 2809 2810 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); 2811 } 2812 2813 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, 2814 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, 2815 CCState &State) LLVM_ATTRIBUTE_UNUSED; 2816 2817 #include "MipsGenCallingConv.inc" 2818 2819 CCAssignFn *MipsTargetLowering::CCAssignFnForCall() const{ 2820 return CC_Mips; 2821 } 2822 2823 CCAssignFn *MipsTargetLowering::CCAssignFnForReturn() const{ 2824 return RetCC_Mips; 2825 } 2826 //===----------------------------------------------------------------------===// 2827 // Call Calling Convention Implementation 2828 //===----------------------------------------------------------------------===// 2829 2830 // Return next O32 integer argument register. 2831 static unsigned getNextIntArgReg(unsigned Reg) { 2832 assert((Reg == Mips::A0) || (Reg == Mips::A2)); 2833 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3; 2834 } 2835 2836 SDValue MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset, 2837 SDValue Chain, SDValue Arg, 2838 const SDLoc &DL, bool IsTailCall, 2839 SelectionDAG &DAG) const { 2840 if (!IsTailCall) { 2841 SDValue PtrOff = 2842 DAG.getNode(ISD::ADD, DL, getPointerTy(DAG.getDataLayout()), StackPtr, 2843 DAG.getIntPtrConstant(Offset, DL)); 2844 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo()); 2845 } 2846 2847 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 2848 int FI = MFI.CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false); 2849 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); 2850 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(), 2851 /* Alignment = */ 0, MachineMemOperand::MOVolatile); 2852 } 2853 2854 void MipsTargetLowering:: 2855 getOpndList(SmallVectorImpl<SDValue> &Ops, 2856 std::deque<std::pair<unsigned, SDValue>> &RegsToPass, 2857 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage, 2858 bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee, 2859 SDValue Chain) const { 2860 // Insert node "GP copy globalreg" before call to function. 2861 // 2862 // R_MIPS_CALL* operators (emitted when non-internal functions are called 2863 // in PIC mode) allow symbols to be resolved via lazy binding. 2864 // The lazy binding stub requires GP to point to the GOT. 2865 // Note that we don't need GP to point to the GOT for indirect calls 2866 // (when R_MIPS_CALL* is not used for the call) because Mips linker generates 2867 // lazy binding stub for a function only when R_MIPS_CALL* are the only relocs 2868 // used for the function (that is, Mips linker doesn't generate lazy binding 2869 // stub for a function whose address is taken in the program). 2870 if (IsPICCall && !InternalLinkage && IsCallReloc) { 2871 unsigned GPReg = ABI.IsN64() ? Mips::GP_64 : Mips::GP; 2872 EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32; 2873 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty))); 2874 } 2875 2876 // Build a sequence of copy-to-reg nodes chained together with token 2877 // chain and flag operands which copy the outgoing args into registers. 2878 // The InFlag in necessary since all emitted instructions must be 2879 // stuck together. 2880 SDValue InFlag; 2881 2882 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 2883 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first, 2884 RegsToPass[i].second, InFlag); 2885 InFlag = Chain.getValue(1); 2886 } 2887 2888 // Add argument registers to the end of the list so that they are 2889 // known live into the call. 2890 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 2891 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first, 2892 RegsToPass[i].second.getValueType())); 2893 2894 // Add a register mask operand representing the call-preserved registers. 2895 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); 2896 const uint32_t *Mask = 2897 TRI->getCallPreservedMask(CLI.DAG.getMachineFunction(), CLI.CallConv); 2898 assert(Mask && "Missing call preserved mask for calling convention"); 2899 if (Subtarget.inMips16HardFloat()) { 2900 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) { 2901 StringRef Sym = G->getGlobal()->getName(); 2902 Function *F = G->getGlobal()->getParent()->getFunction(Sym); 2903 if (F && F->hasFnAttribute("__Mips16RetHelper")) { 2904 Mask = MipsRegisterInfo::getMips16RetHelperMask(); 2905 } 2906 } 2907 } 2908 Ops.push_back(CLI.DAG.getRegisterMask(Mask)); 2909 2910 if (InFlag.getNode()) 2911 Ops.push_back(InFlag); 2912 } 2913 2914 /// LowerCall - functions arguments are copied from virtual regs to 2915 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. 2916 SDValue 2917 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, 2918 SmallVectorImpl<SDValue> &InVals) const { 2919 SelectionDAG &DAG = CLI.DAG; 2920 SDLoc DL = CLI.DL; 2921 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; 2922 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; 2923 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; 2924 SDValue Chain = CLI.Chain; 2925 SDValue Callee = CLI.Callee; 2926 bool &IsTailCall = CLI.IsTailCall; 2927 CallingConv::ID CallConv = CLI.CallConv; 2928 bool IsVarArg = CLI.IsVarArg; 2929 2930 MachineFunction &MF = DAG.getMachineFunction(); 2931 MachineFrameInfo &MFI = MF.getFrameInfo(); 2932 const TargetFrameLowering *TFL = Subtarget.getFrameLowering(); 2933 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 2934 bool IsPIC = isPositionIndependent(); 2935 2936 // Analyze operands of the call, assigning locations to each operand. 2937 SmallVector<CCValAssign, 16> ArgLocs; 2938 MipsCCState CCInfo( 2939 CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, *DAG.getContext(), 2940 MipsCCState::getSpecialCallingConvForCallee(Callee.getNode(), Subtarget)); 2941 2942 const ExternalSymbolSDNode *ES = 2943 dyn_cast_or_null<const ExternalSymbolSDNode>(Callee.getNode()); 2944 2945 // There is one case where CALLSEQ_START..CALLSEQ_END can be nested, which 2946 // is during the lowering of a call with a byval argument which produces 2947 // a call to memcpy. For the O32 case, this causes the caller to allocate 2948 // stack space for the reserved argument area for the callee, then recursively 2949 // again for the memcpy call. In the NEWABI case, this doesn't occur as those 2950 // ABIs mandate that the callee allocates the reserved argument area. We do 2951 // still produce nested CALLSEQ_START..CALLSEQ_END with zero space though. 2952 // 2953 // If the callee has a byval argument and memcpy is used, we are mandated 2954 // to already have produced a reserved argument area for the callee for O32. 2955 // Therefore, the reserved argument area can be reused for both calls. 2956 // 2957 // Other cases of calling memcpy cannot have a chain with a CALLSEQ_START 2958 // present, as we have yet to hook that node onto the chain. 2959 // 2960 // Hence, the CALLSEQ_START and CALLSEQ_END nodes can be eliminated in this 2961 // case. GCC does a similar trick, in that wherever possible, it calculates 2962 // the maximum out going argument area (including the reserved area), and 2963 // preallocates the stack space on entrance to the caller. 2964 // 2965 // FIXME: We should do the same for efficency and space. 2966 2967 // Note: The check on the calling convention below must match 2968 // MipsABIInfo::GetCalleeAllocdArgSizeInBytes(). 2969 bool MemcpyInByVal = ES && 2970 StringRef(ES->getSymbol()) == StringRef("memcpy") && 2971 CallConv != CallingConv::Fast && 2972 Chain.getOpcode() == ISD::CALLSEQ_START; 2973 2974 // Allocate the reserved argument area. It seems strange to do this from the 2975 // caller side but removing it breaks the frame size calculation. 2976 unsigned ReservedArgArea = 2977 MemcpyInByVal ? 0 : ABI.GetCalleeAllocdArgSizeInBytes(CallConv); 2978 CCInfo.AllocateStack(ReservedArgArea, 1); 2979 2980 CCInfo.AnalyzeCallOperands(Outs, CC_Mips, CLI.getArgs(), 2981 ES ? ES->getSymbol() : nullptr); 2982 2983 // Get a count of how many bytes are to be pushed on the stack. 2984 unsigned NextStackOffset = CCInfo.getNextStackOffset(); 2985 2986 // Check if it's really possible to do a tail call. Restrict it to functions 2987 // that are part of this compilation unit. 2988 bool InternalLinkage = false; 2989 if (IsTailCall) { 2990 IsTailCall = isEligibleForTailCallOptimization( 2991 CCInfo, NextStackOffset, *MF.getInfo<MipsFunctionInfo>()); 2992 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 2993 InternalLinkage = G->getGlobal()->hasInternalLinkage(); 2994 IsTailCall &= (InternalLinkage || G->getGlobal()->hasLocalLinkage() || 2995 G->getGlobal()->hasPrivateLinkage() || 2996 G->getGlobal()->hasHiddenVisibility() || 2997 G->getGlobal()->hasProtectedVisibility()); 2998 } 2999 } 3000 if (!IsTailCall && CLI.CS && CLI.CS.isMustTailCall()) 3001 report_fatal_error("failed to perform tail call elimination on a call " 3002 "site marked musttail"); 3003 3004 if (IsTailCall) 3005 ++NumTailCalls; 3006 3007 // Chain is the output chain of the last Load/Store or CopyToReg node. 3008 // ByValChain is the output chain of the last Memcpy node created for copying 3009 // byval arguments to the stack. 3010 unsigned StackAlignment = TFL->getStackAlignment(); 3011 NextStackOffset = alignTo(NextStackOffset, StackAlignment); 3012 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, DL, true); 3013 3014 if (!(IsTailCall || MemcpyInByVal)) 3015 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffset, 0, DL); 3016 3017 SDValue StackPtr = 3018 DAG.getCopyFromReg(Chain, DL, ABI.IsN64() ? Mips::SP_64 : Mips::SP, 3019 getPointerTy(DAG.getDataLayout())); 3020 3021 std::deque<std::pair<unsigned, SDValue>> RegsToPass; 3022 SmallVector<SDValue, 8> MemOpChains; 3023 3024 CCInfo.rewindByValRegsInfo(); 3025 3026 // Walk the register/memloc assignments, inserting copies/loads. 3027 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 3028 SDValue Arg = OutVals[i]; 3029 CCValAssign &VA = ArgLocs[i]; 3030 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT(); 3031 ISD::ArgFlagsTy Flags = Outs[i].Flags; 3032 bool UseUpperBits = false; 3033 3034 // ByVal Arg. 3035 if (Flags.isByVal()) { 3036 unsigned FirstByValReg, LastByValReg; 3037 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); 3038 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg); 3039 3040 assert(Flags.getByValSize() && 3041 "ByVal args of size 0 should have been ignored by front-end."); 3042 assert(ByValIdx < CCInfo.getInRegsParamsCount()); 3043 assert(!IsTailCall && 3044 "Do not tail-call optimize if there is a byval argument."); 3045 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg, 3046 FirstByValReg, LastByValReg, Flags, Subtarget.isLittle(), 3047 VA); 3048 CCInfo.nextInRegsParam(); 3049 continue; 3050 } 3051 3052 // Promote the value if needed. 3053 switch (VA.getLocInfo()) { 3054 default: 3055 llvm_unreachable("Unknown loc info!"); 3056 case CCValAssign::Full: 3057 if (VA.isRegLoc()) { 3058 if ((ValVT == MVT::f32 && LocVT == MVT::i32) || 3059 (ValVT == MVT::f64 && LocVT == MVT::i64) || 3060 (ValVT == MVT::i64 && LocVT == MVT::f64)) 3061 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 3062 else if (ValVT == MVT::f64 && LocVT == MVT::i32) { 3063 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 3064 Arg, DAG.getConstant(0, DL, MVT::i32)); 3065 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 3066 Arg, DAG.getConstant(1, DL, MVT::i32)); 3067 if (!Subtarget.isLittle()) 3068 std::swap(Lo, Hi); 3069 unsigned LocRegLo = VA.getLocReg(); 3070 unsigned LocRegHigh = getNextIntArgReg(LocRegLo); 3071 RegsToPass.push_back(std::make_pair(LocRegLo, Lo)); 3072 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi)); 3073 continue; 3074 } 3075 } 3076 break; 3077 case CCValAssign::BCvt: 3078 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 3079 break; 3080 case CCValAssign::SExtUpper: 3081 UseUpperBits = true; 3082 LLVM_FALLTHROUGH; 3083 case CCValAssign::SExt: 3084 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg); 3085 break; 3086 case CCValAssign::ZExtUpper: 3087 UseUpperBits = true; 3088 LLVM_FALLTHROUGH; 3089 case CCValAssign::ZExt: 3090 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg); 3091 break; 3092 case CCValAssign::AExtUpper: 3093 UseUpperBits = true; 3094 LLVM_FALLTHROUGH; 3095 case CCValAssign::AExt: 3096 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg); 3097 break; 3098 } 3099 3100 if (UseUpperBits) { 3101 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits(); 3102 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3103 Arg = DAG.getNode( 3104 ISD::SHL, DL, VA.getLocVT(), Arg, 3105 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3106 } 3107 3108 // Arguments that can be passed on register must be kept at 3109 // RegsToPass vector 3110 if (VA.isRegLoc()) { 3111 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 3112 continue; 3113 } 3114 3115 // Register can't get to this point... 3116 assert(VA.isMemLoc()); 3117 3118 // emit ISD::STORE whichs stores the 3119 // parameter value to a stack Location 3120 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(), 3121 Chain, Arg, DL, IsTailCall, DAG)); 3122 } 3123 3124 // Transform all store nodes into one single node because all store 3125 // nodes are independent of each other. 3126 if (!MemOpChains.empty()) 3127 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains); 3128 3129 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 3130 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 3131 // node so that legalize doesn't hack it. 3132 3133 EVT Ty = Callee.getValueType(); 3134 bool GlobalOrExternal = false, IsCallReloc = false; 3135 3136 // The long-calls feature is ignored in case of PIC. 3137 // While we do not support -mshared / -mno-shared properly, 3138 // ignore long-calls in case of -mabicalls too. 3139 if (!Subtarget.isABICalls() && !IsPIC) { 3140 // If the function should be called using "long call", 3141 // get its address into a register to prevent using 3142 // of the `jal` instruction for the direct call. 3143 if (auto *N = dyn_cast<ExternalSymbolSDNode>(Callee)) { 3144 if (Subtarget.useLongCalls()) 3145 Callee = Subtarget.hasSym32() 3146 ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 3147 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 3148 } else if (auto *N = dyn_cast<GlobalAddressSDNode>(Callee)) { 3149 bool UseLongCalls = Subtarget.useLongCalls(); 3150 // If the function has long-call/far/near attribute 3151 // it overrides command line switch pased to the backend. 3152 if (auto *F = dyn_cast<Function>(N->getGlobal())) { 3153 if (F->hasFnAttribute("long-call")) 3154 UseLongCalls = true; 3155 else if (F->hasFnAttribute("short-call")) 3156 UseLongCalls = false; 3157 } 3158 if (UseLongCalls) 3159 Callee = Subtarget.hasSym32() 3160 ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 3161 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 3162 } 3163 } 3164 3165 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 3166 if (IsPIC) { 3167 const GlobalValue *Val = G->getGlobal(); 3168 InternalLinkage = Val->hasInternalLinkage(); 3169 3170 if (InternalLinkage) 3171 Callee = getAddrLocal(G, DL, Ty, DAG, ABI.IsN32() || ABI.IsN64()); 3172 else if (LargeGOT) { 3173 Callee = getAddrGlobalLargeGOT(G, DL, Ty, DAG, MipsII::MO_CALL_HI16, 3174 MipsII::MO_CALL_LO16, Chain, 3175 FuncInfo->callPtrInfo(Val)); 3176 IsCallReloc = true; 3177 } else { 3178 Callee = getAddrGlobal(G, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain, 3179 FuncInfo->callPtrInfo(Val)); 3180 IsCallReloc = true; 3181 } 3182 } else 3183 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, 3184 getPointerTy(DAG.getDataLayout()), 0, 3185 MipsII::MO_NO_FLAG); 3186 GlobalOrExternal = true; 3187 } 3188 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 3189 const char *Sym = S->getSymbol(); 3190 3191 if (!IsPIC) // static 3192 Callee = DAG.getTargetExternalSymbol( 3193 Sym, getPointerTy(DAG.getDataLayout()), MipsII::MO_NO_FLAG); 3194 else if (LargeGOT) { 3195 Callee = getAddrGlobalLargeGOT(S, DL, Ty, DAG, MipsII::MO_CALL_HI16, 3196 MipsII::MO_CALL_LO16, Chain, 3197 FuncInfo->callPtrInfo(Sym)); 3198 IsCallReloc = true; 3199 } else { // PIC 3200 Callee = getAddrGlobal(S, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain, 3201 FuncInfo->callPtrInfo(Sym)); 3202 IsCallReloc = true; 3203 } 3204 3205 GlobalOrExternal = true; 3206 } 3207 3208 SmallVector<SDValue, 8> Ops(1, Chain); 3209 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 3210 3211 getOpndList(Ops, RegsToPass, IsPIC, GlobalOrExternal, InternalLinkage, 3212 IsCallReloc, CLI, Callee, Chain); 3213 3214 if (IsTailCall) { 3215 MF.getFrameInfo().setHasTailCall(); 3216 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, Ops); 3217 } 3218 3219 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, Ops); 3220 SDValue InFlag = Chain.getValue(1); 3221 3222 // Create the CALLSEQ_END node in the case of where it is not a call to 3223 // memcpy. 3224 if (!(MemcpyInByVal)) { 3225 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal, 3226 DAG.getIntPtrConstant(0, DL, true), InFlag, DL); 3227 InFlag = Chain.getValue(1); 3228 } 3229 3230 // Handle result values, copying them out of physregs into vregs that we 3231 // return. 3232 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG, 3233 InVals, CLI); 3234 } 3235 3236 /// LowerCallResult - Lower the result values of a call into the 3237 /// appropriate copies out of appropriate physical registers. 3238 SDValue MipsTargetLowering::LowerCallResult( 3239 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg, 3240 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, 3241 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, 3242 TargetLowering::CallLoweringInfo &CLI) const { 3243 // Assign locations to each value returned by this call. 3244 SmallVector<CCValAssign, 16> RVLocs; 3245 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, 3246 *DAG.getContext()); 3247 3248 const ExternalSymbolSDNode *ES = 3249 dyn_cast_or_null<const ExternalSymbolSDNode>(CLI.Callee.getNode()); 3250 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips, CLI.RetTy, 3251 ES ? ES->getSymbol() : nullptr); 3252 3253 // Copy all of the result registers out of their specified physreg. 3254 for (unsigned i = 0; i != RVLocs.size(); ++i) { 3255 CCValAssign &VA = RVLocs[i]; 3256 assert(VA.isRegLoc() && "Can only return in registers!"); 3257 3258 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(), 3259 RVLocs[i].getLocVT(), InFlag); 3260 Chain = Val.getValue(1); 3261 InFlag = Val.getValue(2); 3262 3263 if (VA.isUpperBitsInLoc()) { 3264 unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits(); 3265 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3266 unsigned Shift = 3267 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; 3268 Val = DAG.getNode( 3269 Shift, DL, VA.getLocVT(), Val, 3270 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3271 } 3272 3273 switch (VA.getLocInfo()) { 3274 default: 3275 llvm_unreachable("Unknown loc info!"); 3276 case CCValAssign::Full: 3277 break; 3278 case CCValAssign::BCvt: 3279 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val); 3280 break; 3281 case CCValAssign::AExt: 3282 case CCValAssign::AExtUpper: 3283 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 3284 break; 3285 case CCValAssign::ZExt: 3286 case CCValAssign::ZExtUpper: 3287 Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val, 3288 DAG.getValueType(VA.getValVT())); 3289 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 3290 break; 3291 case CCValAssign::SExt: 3292 case CCValAssign::SExtUpper: 3293 Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val, 3294 DAG.getValueType(VA.getValVT())); 3295 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 3296 break; 3297 } 3298 3299 InVals.push_back(Val); 3300 } 3301 3302 return Chain; 3303 } 3304 3305 static SDValue UnpackFromArgumentSlot(SDValue Val, const CCValAssign &VA, 3306 EVT ArgVT, const SDLoc &DL, 3307 SelectionDAG &DAG) { 3308 MVT LocVT = VA.getLocVT(); 3309 EVT ValVT = VA.getValVT(); 3310 3311 // Shift into the upper bits if necessary. 3312 switch (VA.getLocInfo()) { 3313 default: 3314 break; 3315 case CCValAssign::AExtUpper: 3316 case CCValAssign::SExtUpper: 3317 case CCValAssign::ZExtUpper: { 3318 unsigned ValSizeInBits = ArgVT.getSizeInBits(); 3319 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3320 unsigned Opcode = 3321 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; 3322 Val = DAG.getNode( 3323 Opcode, DL, VA.getLocVT(), Val, 3324 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3325 break; 3326 } 3327 } 3328 3329 // If this is an value smaller than the argument slot size (32-bit for O32, 3330 // 64-bit for N32/N64), it has been promoted in some way to the argument slot 3331 // size. Extract the value and insert any appropriate assertions regarding 3332 // sign/zero extension. 3333 switch (VA.getLocInfo()) { 3334 default: 3335 llvm_unreachable("Unknown loc info!"); 3336 case CCValAssign::Full: 3337 break; 3338 case CCValAssign::AExtUpper: 3339 case CCValAssign::AExt: 3340 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 3341 break; 3342 case CCValAssign::SExtUpper: 3343 case CCValAssign::SExt: 3344 Val = DAG.getNode(ISD::AssertSext, DL, LocVT, Val, DAG.getValueType(ValVT)); 3345 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 3346 break; 3347 case CCValAssign::ZExtUpper: 3348 case CCValAssign::ZExt: 3349 Val = DAG.getNode(ISD::AssertZext, DL, LocVT, Val, DAG.getValueType(ValVT)); 3350 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 3351 break; 3352 case CCValAssign::BCvt: 3353 Val = DAG.getNode(ISD::BITCAST, DL, ValVT, Val); 3354 break; 3355 } 3356 3357 return Val; 3358 } 3359 3360 //===----------------------------------------------------------------------===// 3361 // Formal Arguments Calling Convention Implementation 3362 //===----------------------------------------------------------------------===// 3363 /// LowerFormalArguments - transform physical registers into virtual registers 3364 /// and generate load operations for arguments places on the stack. 3365 SDValue MipsTargetLowering::LowerFormalArguments( 3366 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, 3367 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, 3368 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { 3369 MachineFunction &MF = DAG.getMachineFunction(); 3370 MachineFrameInfo &MFI = MF.getFrameInfo(); 3371 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3372 3373 MipsFI->setVarArgsFrameIndex(0); 3374 3375 // Used with vargs to acumulate store chains. 3376 std::vector<SDValue> OutChains; 3377 3378 // Assign locations to all of the incoming arguments. 3379 SmallVector<CCValAssign, 16> ArgLocs; 3380 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, 3381 *DAG.getContext()); 3382 CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1); 3383 const Function &Func = DAG.getMachineFunction().getFunction(); 3384 Function::const_arg_iterator FuncArg = Func.arg_begin(); 3385 3386 if (Func.hasFnAttribute("interrupt") && !Func.arg_empty()) 3387 report_fatal_error( 3388 "Functions with the interrupt attribute cannot have arguments!"); 3389 3390 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips_FixedArg); 3391 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(), 3392 CCInfo.getInRegsParamsCount() > 0); 3393 3394 unsigned CurArgIdx = 0; 3395 CCInfo.rewindByValRegsInfo(); 3396 3397 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 3398 CCValAssign &VA = ArgLocs[i]; 3399 if (Ins[i].isOrigArg()) { 3400 std::advance(FuncArg, Ins[i].getOrigArgIndex() - CurArgIdx); 3401 CurArgIdx = Ins[i].getOrigArgIndex(); 3402 } 3403 EVT ValVT = VA.getValVT(); 3404 ISD::ArgFlagsTy Flags = Ins[i].Flags; 3405 bool IsRegLoc = VA.isRegLoc(); 3406 3407 if (Flags.isByVal()) { 3408 assert(Ins[i].isOrigArg() && "Byval arguments cannot be implicit"); 3409 unsigned FirstByValReg, LastByValReg; 3410 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); 3411 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg); 3412 3413 assert(Flags.getByValSize() && 3414 "ByVal args of size 0 should have been ignored by front-end."); 3415 assert(ByValIdx < CCInfo.getInRegsParamsCount()); 3416 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg, 3417 FirstByValReg, LastByValReg, VA, CCInfo); 3418 CCInfo.nextInRegsParam(); 3419 continue; 3420 } 3421 3422 // Arguments stored on registers 3423 if (IsRegLoc) { 3424 MVT RegVT = VA.getLocVT(); 3425 unsigned ArgReg = VA.getLocReg(); 3426 const TargetRegisterClass *RC = getRegClassFor(RegVT); 3427 3428 // Transform the arguments stored on 3429 // physical registers into virtual ones 3430 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC); 3431 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT); 3432 3433 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG); 3434 3435 // Handle floating point arguments passed in integer registers and 3436 // long double arguments passed in floating point registers. 3437 if ((RegVT == MVT::i32 && ValVT == MVT::f32) || 3438 (RegVT == MVT::i64 && ValVT == MVT::f64) || 3439 (RegVT == MVT::f64 && ValVT == MVT::i64)) 3440 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue); 3441 else if (ABI.IsO32() && RegVT == MVT::i32 && 3442 ValVT == MVT::f64) { 3443 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(), 3444 getNextIntArgReg(ArgReg), RC); 3445 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT); 3446 if (!Subtarget.isLittle()) 3447 std::swap(ArgValue, ArgValue2); 3448 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, 3449 ArgValue, ArgValue2); 3450 } 3451 3452 InVals.push_back(ArgValue); 3453 } else { // VA.isRegLoc() 3454 MVT LocVT = VA.getLocVT(); 3455 3456 if (ABI.IsO32()) { 3457 // We ought to be able to use LocVT directly but O32 sets it to i32 3458 // when allocating floating point values to integer registers. 3459 // This shouldn't influence how we load the value into registers unless 3460 // we are targeting softfloat. 3461 if (VA.getValVT().isFloatingPoint() && !Subtarget.useSoftFloat()) 3462 LocVT = VA.getValVT(); 3463 } 3464 3465 // sanity check 3466 assert(VA.isMemLoc()); 3467 3468 // The stack pointer offset is relative to the caller stack frame. 3469 int FI = MFI.CreateFixedObject(LocVT.getSizeInBits() / 8, 3470 VA.getLocMemOffset(), true); 3471 3472 // Create load nodes to retrieve arguments from the stack 3473 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); 3474 SDValue ArgValue = DAG.getLoad( 3475 LocVT, DL, Chain, FIN, 3476 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)); 3477 OutChains.push_back(ArgValue.getValue(1)); 3478 3479 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG); 3480 3481 InVals.push_back(ArgValue); 3482 } 3483 } 3484 3485 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 3486 // The mips ABIs for returning structs by value requires that we copy 3487 // the sret argument into $v0 for the return. Save the argument into 3488 // a virtual register so that we can access it from the return points. 3489 if (Ins[i].Flags.isSRet()) { 3490 unsigned Reg = MipsFI->getSRetReturnReg(); 3491 if (!Reg) { 3492 Reg = MF.getRegInfo().createVirtualRegister( 3493 getRegClassFor(ABI.IsN64() ? MVT::i64 : MVT::i32)); 3494 MipsFI->setSRetReturnReg(Reg); 3495 } 3496 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[i]); 3497 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain); 3498 break; 3499 } 3500 } 3501 3502 if (IsVarArg) 3503 writeVarArgRegs(OutChains, Chain, DL, DAG, CCInfo); 3504 3505 // All stores are grouped in one node to allow the matching between 3506 // the size of Ins and InVals. This only happens when on varg functions 3507 if (!OutChains.empty()) { 3508 OutChains.push_back(Chain); 3509 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains); 3510 } 3511 3512 return Chain; 3513 } 3514 3515 //===----------------------------------------------------------------------===// 3516 // Return Value Calling Convention Implementation 3517 //===----------------------------------------------------------------------===// 3518 3519 bool 3520 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv, 3521 MachineFunction &MF, bool IsVarArg, 3522 const SmallVectorImpl<ISD::OutputArg> &Outs, 3523 LLVMContext &Context) const { 3524 SmallVector<CCValAssign, 16> RVLocs; 3525 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context); 3526 return CCInfo.CheckReturn(Outs, RetCC_Mips); 3527 } 3528 3529 bool 3530 MipsTargetLowering::shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const { 3531 if ((ABI.IsN32() || ABI.IsN64()) && Type == MVT::i32) 3532 return true; 3533 3534 return IsSigned; 3535 } 3536 3537 SDValue 3538 MipsTargetLowering::LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps, 3539 const SDLoc &DL, 3540 SelectionDAG &DAG) const { 3541 MachineFunction &MF = DAG.getMachineFunction(); 3542 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3543 3544 MipsFI->setISR(); 3545 3546 return DAG.getNode(MipsISD::ERet, DL, MVT::Other, RetOps); 3547 } 3548 3549 SDValue 3550 MipsTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, 3551 bool IsVarArg, 3552 const SmallVectorImpl<ISD::OutputArg> &Outs, 3553 const SmallVectorImpl<SDValue> &OutVals, 3554 const SDLoc &DL, SelectionDAG &DAG) const { 3555 // CCValAssign - represent the assignment of 3556 // the return value to a location 3557 SmallVector<CCValAssign, 16> RVLocs; 3558 MachineFunction &MF = DAG.getMachineFunction(); 3559 3560 // CCState - Info about the registers and stack slot. 3561 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext()); 3562 3563 // Analyze return values. 3564 CCInfo.AnalyzeReturn(Outs, RetCC_Mips); 3565 3566 SDValue Flag; 3567 SmallVector<SDValue, 4> RetOps(1, Chain); 3568 3569 // Copy the result values into the output registers. 3570 for (unsigned i = 0; i != RVLocs.size(); ++i) { 3571 SDValue Val = OutVals[i]; 3572 CCValAssign &VA = RVLocs[i]; 3573 assert(VA.isRegLoc() && "Can only return in registers!"); 3574 bool UseUpperBits = false; 3575 3576 switch (VA.getLocInfo()) { 3577 default: 3578 llvm_unreachable("Unknown loc info!"); 3579 case CCValAssign::Full: 3580 break; 3581 case CCValAssign::BCvt: 3582 Val = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Val); 3583 break; 3584 case CCValAssign::AExtUpper: 3585 UseUpperBits = true; 3586 LLVM_FALLTHROUGH; 3587 case CCValAssign::AExt: 3588 Val = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Val); 3589 break; 3590 case CCValAssign::ZExtUpper: 3591 UseUpperBits = true; 3592 LLVM_FALLTHROUGH; 3593 case CCValAssign::ZExt: 3594 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Val); 3595 break; 3596 case CCValAssign::SExtUpper: 3597 UseUpperBits = true; 3598 LLVM_FALLTHROUGH; 3599 case CCValAssign::SExt: 3600 Val = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Val); 3601 break; 3602 } 3603 3604 if (UseUpperBits) { 3605 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits(); 3606 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3607 Val = DAG.getNode( 3608 ISD::SHL, DL, VA.getLocVT(), Val, 3609 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3610 } 3611 3612 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag); 3613 3614 // Guarantee that all emitted copies are stuck together with flags. 3615 Flag = Chain.getValue(1); 3616 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 3617 } 3618 3619 // The mips ABIs for returning structs by value requires that we copy 3620 // the sret argument into $v0 for the return. We saved the argument into 3621 // a virtual register in the entry block, so now we copy the value out 3622 // and into $v0. 3623 if (MF.getFunction().hasStructRetAttr()) { 3624 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3625 unsigned Reg = MipsFI->getSRetReturnReg(); 3626 3627 if (!Reg) 3628 llvm_unreachable("sret virtual register not created in the entry block"); 3629 SDValue Val = 3630 DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(DAG.getDataLayout())); 3631 unsigned V0 = ABI.IsN64() ? Mips::V0_64 : Mips::V0; 3632 3633 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag); 3634 Flag = Chain.getValue(1); 3635 RetOps.push_back(DAG.getRegister(V0, getPointerTy(DAG.getDataLayout()))); 3636 } 3637 3638 RetOps[0] = Chain; // Update chain. 3639 3640 // Add the flag if we have it. 3641 if (Flag.getNode()) 3642 RetOps.push_back(Flag); 3643 3644 // ISRs must use "eret". 3645 if (DAG.getMachineFunction().getFunction().hasFnAttribute("interrupt")) 3646 return LowerInterruptReturn(RetOps, DL, DAG); 3647 3648 // Standard return on Mips is a "jr $ra" 3649 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, RetOps); 3650 } 3651 3652 //===----------------------------------------------------------------------===// 3653 // Mips Inline Assembly Support 3654 //===----------------------------------------------------------------------===// 3655 3656 /// getConstraintType - Given a constraint letter, return the type of 3657 /// constraint it is for this target. 3658 MipsTargetLowering::ConstraintType 3659 MipsTargetLowering::getConstraintType(StringRef Constraint) const { 3660 // Mips specific constraints 3661 // GCC config/mips/constraints.md 3662 // 3663 // 'd' : An address register. Equivalent to r 3664 // unless generating MIPS16 code. 3665 // 'y' : Equivalent to r; retained for 3666 // backwards compatibility. 3667 // 'c' : A register suitable for use in an indirect 3668 // jump. This will always be $25 for -mabicalls. 3669 // 'l' : The lo register. 1 word storage. 3670 // 'x' : The hilo register pair. Double word storage. 3671 if (Constraint.size() == 1) { 3672 switch (Constraint[0]) { 3673 default : break; 3674 case 'd': 3675 case 'y': 3676 case 'f': 3677 case 'c': 3678 case 'l': 3679 case 'x': 3680 return C_RegisterClass; 3681 case 'R': 3682 return C_Memory; 3683 } 3684 } 3685 3686 if (Constraint == "ZC") 3687 return C_Memory; 3688 3689 return TargetLowering::getConstraintType(Constraint); 3690 } 3691 3692 /// Examine constraint type and operand type and determine a weight value. 3693 /// This object must already have been set up with the operand type 3694 /// and the current alternative constraint selected. 3695 TargetLowering::ConstraintWeight 3696 MipsTargetLowering::getSingleConstraintMatchWeight( 3697 AsmOperandInfo &info, const char *constraint) const { 3698 ConstraintWeight weight = CW_Invalid; 3699 Value *CallOperandVal = info.CallOperandVal; 3700 // If we don't have a value, we can't do a match, 3701 // but allow it at the lowest weight. 3702 if (!CallOperandVal) 3703 return CW_Default; 3704 Type *type = CallOperandVal->getType(); 3705 // Look at the constraint type. 3706 switch (*constraint) { 3707 default: 3708 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 3709 break; 3710 case 'd': 3711 case 'y': 3712 if (type->isIntegerTy()) 3713 weight = CW_Register; 3714 break; 3715 case 'f': // FPU or MSA register 3716 if (Subtarget.hasMSA() && type->isVectorTy() && 3717 cast<VectorType>(type)->getBitWidth() == 128) 3718 weight = CW_Register; 3719 else if (type->isFloatTy()) 3720 weight = CW_Register; 3721 break; 3722 case 'c': // $25 for indirect jumps 3723 case 'l': // lo register 3724 case 'x': // hilo register pair 3725 if (type->isIntegerTy()) 3726 weight = CW_SpecificReg; 3727 break; 3728 case 'I': // signed 16 bit immediate 3729 case 'J': // integer zero 3730 case 'K': // unsigned 16 bit immediate 3731 case 'L': // signed 32 bit immediate where lower 16 bits are 0 3732 case 'N': // immediate in the range of -65535 to -1 (inclusive) 3733 case 'O': // signed 15 bit immediate (+- 16383) 3734 case 'P': // immediate in the range of 65535 to 1 (inclusive) 3735 if (isa<ConstantInt>(CallOperandVal)) 3736 weight = CW_Constant; 3737 break; 3738 case 'R': 3739 weight = CW_Memory; 3740 break; 3741 } 3742 return weight; 3743 } 3744 3745 /// This is a helper function to parse a physical register string and split it 3746 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag 3747 /// that is returned indicates whether parsing was successful. The second flag 3748 /// is true if the numeric part exists. 3749 static std::pair<bool, bool> parsePhysicalReg(StringRef C, StringRef &Prefix, 3750 unsigned long long &Reg) { 3751 if (C.front() != '{' || C.back() != '}') 3752 return std::make_pair(false, false); 3753 3754 // Search for the first numeric character. 3755 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1; 3756 I = std::find_if(B, E, isdigit); 3757 3758 Prefix = StringRef(B, I - B); 3759 3760 // The second flag is set to false if no numeric characters were found. 3761 if (I == E) 3762 return std::make_pair(true, false); 3763 3764 // Parse the numeric characters. 3765 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg), 3766 true); 3767 } 3768 3769 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering:: 3770 parseRegForInlineAsmConstraint(StringRef C, MVT VT) const { 3771 const TargetRegisterInfo *TRI = 3772 Subtarget.getRegisterInfo(); 3773 const TargetRegisterClass *RC; 3774 StringRef Prefix; 3775 unsigned long long Reg; 3776 3777 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg); 3778 3779 if (!R.first) 3780 return std::make_pair(0U, nullptr); 3781 3782 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo. 3783 // No numeric characters follow "hi" or "lo". 3784 if (R.second) 3785 return std::make_pair(0U, nullptr); 3786 3787 RC = TRI->getRegClass(Prefix == "hi" ? 3788 Mips::HI32RegClassID : Mips::LO32RegClassID); 3789 return std::make_pair(*(RC->begin()), RC); 3790 } else if (Prefix.startswith("$msa")) { 3791 // Parse $msa(ir|csr|access|save|modify|request|map|unmap) 3792 3793 // No numeric characters follow the name. 3794 if (R.second) 3795 return std::make_pair(0U, nullptr); 3796 3797 Reg = StringSwitch<unsigned long long>(Prefix) 3798 .Case("$msair", Mips::MSAIR) 3799 .Case("$msacsr", Mips::MSACSR) 3800 .Case("$msaaccess", Mips::MSAAccess) 3801 .Case("$msasave", Mips::MSASave) 3802 .Case("$msamodify", Mips::MSAModify) 3803 .Case("$msarequest", Mips::MSARequest) 3804 .Case("$msamap", Mips::MSAMap) 3805 .Case("$msaunmap", Mips::MSAUnmap) 3806 .Default(0); 3807 3808 if (!Reg) 3809 return std::make_pair(0U, nullptr); 3810 3811 RC = TRI->getRegClass(Mips::MSACtrlRegClassID); 3812 return std::make_pair(Reg, RC); 3813 } 3814 3815 if (!R.second) 3816 return std::make_pair(0U, nullptr); 3817 3818 if (Prefix == "$f") { // Parse $f0-$f31. 3819 // If the size of FP registers is 64-bit or Reg is an even number, select 3820 // the 64-bit register class. Otherwise, select the 32-bit register class. 3821 if (VT == MVT::Other) 3822 VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32; 3823 3824 RC = getRegClassFor(VT); 3825 3826 if (RC == &Mips::AFGR64RegClass) { 3827 assert(Reg % 2 == 0); 3828 Reg >>= 1; 3829 } 3830 } else if (Prefix == "$fcc") // Parse $fcc0-$fcc7. 3831 RC = TRI->getRegClass(Mips::FCCRegClassID); 3832 else if (Prefix == "$w") { // Parse $w0-$w31. 3833 RC = getRegClassFor((VT == MVT::Other) ? MVT::v16i8 : VT); 3834 } else { // Parse $0-$31. 3835 assert(Prefix == "$"); 3836 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT); 3837 } 3838 3839 assert(Reg < RC->getNumRegs()); 3840 return std::make_pair(*(RC->begin() + Reg), RC); 3841 } 3842 3843 /// Given a register class constraint, like 'r', if this corresponds directly 3844 /// to an LLVM register class, return a register of 0 and the register class 3845 /// pointer. 3846 std::pair<unsigned, const TargetRegisterClass *> 3847 MipsTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, 3848 StringRef Constraint, 3849 MVT VT) const { 3850 if (Constraint.size() == 1) { 3851 switch (Constraint[0]) { 3852 case 'd': // Address register. Same as 'r' unless generating MIPS16 code. 3853 case 'y': // Same as 'r'. Exists for compatibility. 3854 case 'r': 3855 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) { 3856 if (Subtarget.inMips16Mode()) 3857 return std::make_pair(0U, &Mips::CPU16RegsRegClass); 3858 return std::make_pair(0U, &Mips::GPR32RegClass); 3859 } 3860 if (VT == MVT::i64 && !Subtarget.isGP64bit()) 3861 return std::make_pair(0U, &Mips::GPR32RegClass); 3862 if (VT == MVT::i64 && Subtarget.isGP64bit()) 3863 return std::make_pair(0U, &Mips::GPR64RegClass); 3864 // This will generate an error message 3865 return std::make_pair(0U, nullptr); 3866 case 'f': // FPU or MSA register 3867 if (VT == MVT::v16i8) 3868 return std::make_pair(0U, &Mips::MSA128BRegClass); 3869 else if (VT == MVT::v8i16 || VT == MVT::v8f16) 3870 return std::make_pair(0U, &Mips::MSA128HRegClass); 3871 else if (VT == MVT::v4i32 || VT == MVT::v4f32) 3872 return std::make_pair(0U, &Mips::MSA128WRegClass); 3873 else if (VT == MVT::v2i64 || VT == MVT::v2f64) 3874 return std::make_pair(0U, &Mips::MSA128DRegClass); 3875 else if (VT == MVT::f32) 3876 return std::make_pair(0U, &Mips::FGR32RegClass); 3877 else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) { 3878 if (Subtarget.isFP64bit()) 3879 return std::make_pair(0U, &Mips::FGR64RegClass); 3880 return std::make_pair(0U, &Mips::AFGR64RegClass); 3881 } 3882 break; 3883 case 'c': // register suitable for indirect jump 3884 if (VT == MVT::i32) 3885 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass); 3886 if (VT == MVT::i64) 3887 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass); 3888 // This will generate an error message 3889 return std::make_pair(0U, nullptr); 3890 case 'l': // use the `lo` register to store values 3891 // that are no bigger than a word 3892 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) 3893 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass); 3894 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass); 3895 case 'x': // use the concatenated `hi` and `lo` registers 3896 // to store doubleword values 3897 // Fixme: Not triggering the use of both hi and low 3898 // This will generate an error message 3899 return std::make_pair(0U, nullptr); 3900 } 3901 } 3902 3903 std::pair<unsigned, const TargetRegisterClass *> R; 3904 R = parseRegForInlineAsmConstraint(Constraint, VT); 3905 3906 if (R.second) 3907 return R; 3908 3909 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); 3910 } 3911 3912 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 3913 /// vector. If it is invalid, don't add anything to Ops. 3914 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 3915 std::string &Constraint, 3916 std::vector<SDValue>&Ops, 3917 SelectionDAG &DAG) const { 3918 SDLoc DL(Op); 3919 SDValue Result; 3920 3921 // Only support length 1 constraints for now. 3922 if (Constraint.length() > 1) return; 3923 3924 char ConstraintLetter = Constraint[0]; 3925 switch (ConstraintLetter) { 3926 default: break; // This will fall through to the generic implementation 3927 case 'I': // Signed 16 bit constant 3928 // If this fails, the parent routine will give an error 3929 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3930 EVT Type = Op.getValueType(); 3931 int64_t Val = C->getSExtValue(); 3932 if (isInt<16>(Val)) { 3933 Result = DAG.getTargetConstant(Val, DL, Type); 3934 break; 3935 } 3936 } 3937 return; 3938 case 'J': // integer zero 3939 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3940 EVT Type = Op.getValueType(); 3941 int64_t Val = C->getZExtValue(); 3942 if (Val == 0) { 3943 Result = DAG.getTargetConstant(0, DL, Type); 3944 break; 3945 } 3946 } 3947 return; 3948 case 'K': // unsigned 16 bit immediate 3949 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3950 EVT Type = Op.getValueType(); 3951 uint64_t Val = (uint64_t)C->getZExtValue(); 3952 if (isUInt<16>(Val)) { 3953 Result = DAG.getTargetConstant(Val, DL, Type); 3954 break; 3955 } 3956 } 3957 return; 3958 case 'L': // signed 32 bit immediate where lower 16 bits are 0 3959 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3960 EVT Type = Op.getValueType(); 3961 int64_t Val = C->getSExtValue(); 3962 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){ 3963 Result = DAG.getTargetConstant(Val, DL, Type); 3964 break; 3965 } 3966 } 3967 return; 3968 case 'N': // immediate in the range of -65535 to -1 (inclusive) 3969 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3970 EVT Type = Op.getValueType(); 3971 int64_t Val = C->getSExtValue(); 3972 if ((Val >= -65535) && (Val <= -1)) { 3973 Result = DAG.getTargetConstant(Val, DL, Type); 3974 break; 3975 } 3976 } 3977 return; 3978 case 'O': // signed 15 bit immediate 3979 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3980 EVT Type = Op.getValueType(); 3981 int64_t Val = C->getSExtValue(); 3982 if ((isInt<15>(Val))) { 3983 Result = DAG.getTargetConstant(Val, DL, Type); 3984 break; 3985 } 3986 } 3987 return; 3988 case 'P': // immediate in the range of 1 to 65535 (inclusive) 3989 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3990 EVT Type = Op.getValueType(); 3991 int64_t Val = C->getSExtValue(); 3992 if ((Val <= 65535) && (Val >= 1)) { 3993 Result = DAG.getTargetConstant(Val, DL, Type); 3994 break; 3995 } 3996 } 3997 return; 3998 } 3999 4000 if (Result.getNode()) { 4001 Ops.push_back(Result); 4002 return; 4003 } 4004 4005 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 4006 } 4007 4008 bool MipsTargetLowering::isLegalAddressingMode(const DataLayout &DL, 4009 const AddrMode &AM, Type *Ty, 4010 unsigned AS, Instruction *I) const { 4011 // No global is ever allowed as a base. 4012 if (AM.BaseGV) 4013 return false; 4014 4015 switch (AM.Scale) { 4016 case 0: // "r+i" or just "i", depending on HasBaseReg. 4017 break; 4018 case 1: 4019 if (!AM.HasBaseReg) // allow "r+i". 4020 break; 4021 return false; // disallow "r+r" or "r+r+i". 4022 default: 4023 return false; 4024 } 4025 4026 return true; 4027 } 4028 4029 bool 4030 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 4031 // The Mips target isn't yet aware of offsets. 4032 return false; 4033 } 4034 4035 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign, 4036 unsigned SrcAlign, 4037 bool IsMemset, bool ZeroMemset, 4038 bool MemcpyStrSrc, 4039 MachineFunction &MF) const { 4040 if (Subtarget.hasMips64()) 4041 return MVT::i64; 4042 4043 return MVT::i32; 4044 } 4045 4046 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { 4047 if (VT != MVT::f32 && VT != MVT::f64) 4048 return false; 4049 if (Imm.isNegZero()) 4050 return false; 4051 return Imm.isZero(); 4052 } 4053 4054 unsigned MipsTargetLowering::getJumpTableEncoding() const { 4055 4056 // FIXME: For space reasons this should be: EK_GPRel32BlockAddress. 4057 if (ABI.IsN64() && isPositionIndependent()) 4058 return MachineJumpTableInfo::EK_GPRel64BlockAddress; 4059 4060 return TargetLowering::getJumpTableEncoding(); 4061 } 4062 4063 bool MipsTargetLowering::useSoftFloat() const { 4064 return Subtarget.useSoftFloat(); 4065 } 4066 4067 void MipsTargetLowering::copyByValRegs( 4068 SDValue Chain, const SDLoc &DL, std::vector<SDValue> &OutChains, 4069 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags, 4070 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg, 4071 unsigned FirstReg, unsigned LastReg, const CCValAssign &VA, 4072 MipsCCState &State) const { 4073 MachineFunction &MF = DAG.getMachineFunction(); 4074 MachineFrameInfo &MFI = MF.getFrameInfo(); 4075 unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes(); 4076 unsigned NumRegs = LastReg - FirstReg; 4077 unsigned RegAreaSize = NumRegs * GPRSizeInBytes; 4078 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize); 4079 int FrameObjOffset; 4080 ArrayRef<MCPhysReg> ByValArgRegs = ABI.GetByValArgRegs(); 4081 4082 if (RegAreaSize) 4083 FrameObjOffset = 4084 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) - 4085 (int)((ByValArgRegs.size() - FirstReg) * GPRSizeInBytes); 4086 else 4087 FrameObjOffset = VA.getLocMemOffset(); 4088 4089 // Create frame object. 4090 EVT PtrTy = getPointerTy(DAG.getDataLayout()); 4091 // Make the fixed object stored to mutable so that the load instructions 4092 // referencing it have their memory dependencies added. 4093 // Set the frame object as isAliased which clears the underlying objects 4094 // vector in ScheduleDAGInstrs::buildSchedGraph() resulting in addition of all 4095 // stores as dependencies for loads referencing this fixed object. 4096 int FI = MFI.CreateFixedObject(FrameObjSize, FrameObjOffset, false, true); 4097 SDValue FIN = DAG.getFrameIndex(FI, PtrTy); 4098 InVals.push_back(FIN); 4099 4100 if (!NumRegs) 4101 return; 4102 4103 // Copy arg registers. 4104 MVT RegTy = MVT::getIntegerVT(GPRSizeInBytes * 8); 4105 const TargetRegisterClass *RC = getRegClassFor(RegTy); 4106 4107 for (unsigned I = 0; I < NumRegs; ++I) { 4108 unsigned ArgReg = ByValArgRegs[FirstReg + I]; 4109 unsigned VReg = addLiveIn(MF, ArgReg, RC); 4110 unsigned Offset = I * GPRSizeInBytes; 4111 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN, 4112 DAG.getConstant(Offset, DL, PtrTy)); 4113 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy), 4114 StorePtr, MachinePointerInfo(FuncArg, Offset)); 4115 OutChains.push_back(Store); 4116 } 4117 } 4118 4119 // Copy byVal arg to registers and stack. 4120 void MipsTargetLowering::passByValArg( 4121 SDValue Chain, const SDLoc &DL, 4122 std::deque<std::pair<unsigned, SDValue>> &RegsToPass, 4123 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr, 4124 MachineFrameInfo &MFI, SelectionDAG &DAG, SDValue Arg, unsigned FirstReg, 4125 unsigned LastReg, const ISD::ArgFlagsTy &Flags, bool isLittle, 4126 const CCValAssign &VA) const { 4127 unsigned ByValSizeInBytes = Flags.getByValSize(); 4128 unsigned OffsetInBytes = 0; // From beginning of struct 4129 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 4130 unsigned Alignment = std::min(Flags.getByValAlign(), RegSizeInBytes); 4131 EVT PtrTy = getPointerTy(DAG.getDataLayout()), 4132 RegTy = MVT::getIntegerVT(RegSizeInBytes * 8); 4133 unsigned NumRegs = LastReg - FirstReg; 4134 4135 if (NumRegs) { 4136 ArrayRef<MCPhysReg> ArgRegs = ABI.GetByValArgRegs(); 4137 bool LeftoverBytes = (NumRegs * RegSizeInBytes > ByValSizeInBytes); 4138 unsigned I = 0; 4139 4140 // Copy words to registers. 4141 for (; I < NumRegs - LeftoverBytes; ++I, OffsetInBytes += RegSizeInBytes) { 4142 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 4143 DAG.getConstant(OffsetInBytes, DL, PtrTy)); 4144 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr, 4145 MachinePointerInfo(), Alignment); 4146 MemOpChains.push_back(LoadVal.getValue(1)); 4147 unsigned ArgReg = ArgRegs[FirstReg + I]; 4148 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal)); 4149 } 4150 4151 // Return if the struct has been fully copied. 4152 if (ByValSizeInBytes == OffsetInBytes) 4153 return; 4154 4155 // Copy the remainder of the byval argument with sub-word loads and shifts. 4156 if (LeftoverBytes) { 4157 SDValue Val; 4158 4159 for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0; 4160 OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) { 4161 unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes; 4162 4163 if (RemainingSizeInBytes < LoadSizeInBytes) 4164 continue; 4165 4166 // Load subword. 4167 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 4168 DAG.getConstant(OffsetInBytes, DL, 4169 PtrTy)); 4170 SDValue LoadVal = DAG.getExtLoad( 4171 ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, MachinePointerInfo(), 4172 MVT::getIntegerVT(LoadSizeInBytes * 8), Alignment); 4173 MemOpChains.push_back(LoadVal.getValue(1)); 4174 4175 // Shift the loaded value. 4176 unsigned Shamt; 4177 4178 if (isLittle) 4179 Shamt = TotalBytesLoaded * 8; 4180 else 4181 Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8; 4182 4183 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal, 4184 DAG.getConstant(Shamt, DL, MVT::i32)); 4185 4186 if (Val.getNode()) 4187 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift); 4188 else 4189 Val = Shift; 4190 4191 OffsetInBytes += LoadSizeInBytes; 4192 TotalBytesLoaded += LoadSizeInBytes; 4193 Alignment = std::min(Alignment, LoadSizeInBytes); 4194 } 4195 4196 unsigned ArgReg = ArgRegs[FirstReg + I]; 4197 RegsToPass.push_back(std::make_pair(ArgReg, Val)); 4198 return; 4199 } 4200 } 4201 4202 // Copy remainder of byval arg to it with memcpy. 4203 unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes; 4204 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 4205 DAG.getConstant(OffsetInBytes, DL, PtrTy)); 4206 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr, 4207 DAG.getIntPtrConstant(VA.getLocMemOffset(), DL)); 4208 Chain = DAG.getMemcpy(Chain, DL, Dst, Src, 4209 DAG.getConstant(MemCpySize, DL, PtrTy), 4210 Alignment, /*isVolatile=*/false, /*AlwaysInline=*/false, 4211 /*isTailCall=*/false, 4212 MachinePointerInfo(), MachinePointerInfo()); 4213 MemOpChains.push_back(Chain); 4214 } 4215 4216 void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains, 4217 SDValue Chain, const SDLoc &DL, 4218 SelectionDAG &DAG, 4219 CCState &State) const { 4220 ArrayRef<MCPhysReg> ArgRegs = ABI.GetVarArgRegs(); 4221 unsigned Idx = State.getFirstUnallocated(ArgRegs); 4222 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 4223 MVT RegTy = MVT::getIntegerVT(RegSizeInBytes * 8); 4224 const TargetRegisterClass *RC = getRegClassFor(RegTy); 4225 MachineFunction &MF = DAG.getMachineFunction(); 4226 MachineFrameInfo &MFI = MF.getFrameInfo(); 4227 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 4228 4229 // Offset of the first variable argument from stack pointer. 4230 int VaArgOffset; 4231 4232 if (ArgRegs.size() == Idx) 4233 VaArgOffset = alignTo(State.getNextStackOffset(), RegSizeInBytes); 4234 else { 4235 VaArgOffset = 4236 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) - 4237 (int)(RegSizeInBytes * (ArgRegs.size() - Idx)); 4238 } 4239 4240 // Record the frame index of the first variable argument 4241 // which is a value necessary to VASTART. 4242 int FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true); 4243 MipsFI->setVarArgsFrameIndex(FI); 4244 4245 // Copy the integer registers that have not been used for argument passing 4246 // to the argument register save area. For O32, the save area is allocated 4247 // in the caller's stack frame, while for N32/64, it is allocated in the 4248 // callee's stack frame. 4249 for (unsigned I = Idx; I < ArgRegs.size(); 4250 ++I, VaArgOffset += RegSizeInBytes) { 4251 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC); 4252 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy); 4253 FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true); 4254 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); 4255 SDValue Store = 4256 DAG.getStore(Chain, DL, ArgValue, PtrOff, MachinePointerInfo()); 4257 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue( 4258 (Value *)nullptr); 4259 OutChains.push_back(Store); 4260 } 4261 } 4262 4263 void MipsTargetLowering::HandleByVal(CCState *State, unsigned &Size, 4264 unsigned Align) const { 4265 const TargetFrameLowering *TFL = Subtarget.getFrameLowering(); 4266 4267 assert(Size && "Byval argument's size shouldn't be 0."); 4268 4269 Align = std::min(Align, TFL->getStackAlignment()); 4270 4271 unsigned FirstReg = 0; 4272 unsigned NumRegs = 0; 4273 4274 if (State->getCallingConv() != CallingConv::Fast) { 4275 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 4276 ArrayRef<MCPhysReg> IntArgRegs = ABI.GetByValArgRegs(); 4277 // FIXME: The O32 case actually describes no shadow registers. 4278 const MCPhysReg *ShadowRegs = 4279 ABI.IsO32() ? IntArgRegs.data() : Mips64DPRegs; 4280 4281 // We used to check the size as well but we can't do that anymore since 4282 // CCState::HandleByVal() rounds up the size after calling this function. 4283 assert(!(Align % RegSizeInBytes) && 4284 "Byval argument's alignment should be a multiple of" 4285 "RegSizeInBytes."); 4286 4287 FirstReg = State->getFirstUnallocated(IntArgRegs); 4288 4289 // If Align > RegSizeInBytes, the first arg register must be even. 4290 // FIXME: This condition happens to do the right thing but it's not the 4291 // right way to test it. We want to check that the stack frame offset 4292 // of the register is aligned. 4293 if ((Align > RegSizeInBytes) && (FirstReg % 2)) { 4294 State->AllocateReg(IntArgRegs[FirstReg], ShadowRegs[FirstReg]); 4295 ++FirstReg; 4296 } 4297 4298 // Mark the registers allocated. 4299 Size = alignTo(Size, RegSizeInBytes); 4300 for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size()); 4301 Size -= RegSizeInBytes, ++I, ++NumRegs) 4302 State->AllocateReg(IntArgRegs[I], ShadowRegs[I]); 4303 } 4304 4305 State->addInRegsParamInfo(FirstReg, FirstReg + NumRegs); 4306 } 4307 4308 MachineBasicBlock *MipsTargetLowering::emitPseudoSELECT(MachineInstr &MI, 4309 MachineBasicBlock *BB, 4310 bool isFPCmp, 4311 unsigned Opc) const { 4312 assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) && 4313 "Subtarget already supports SELECT nodes with the use of" 4314 "conditional-move instructions."); 4315 4316 const TargetInstrInfo *TII = 4317 Subtarget.getInstrInfo(); 4318 DebugLoc DL = MI.getDebugLoc(); 4319 4320 // To "insert" a SELECT instruction, we actually have to insert the 4321 // diamond control-flow pattern. The incoming instruction knows the 4322 // destination vreg to set, the condition code register to branch on, the 4323 // true/false values to select between, and a branch opcode to use. 4324 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 4325 MachineFunction::iterator It = ++BB->getIterator(); 4326 4327 // thisMBB: 4328 // ... 4329 // TrueVal = ... 4330 // setcc r1, r2, r3 4331 // bNE r1, r0, copy1MBB 4332 // fallthrough --> copy0MBB 4333 MachineBasicBlock *thisMBB = BB; 4334 MachineFunction *F = BB->getParent(); 4335 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 4336 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 4337 F->insert(It, copy0MBB); 4338 F->insert(It, sinkMBB); 4339 4340 // Transfer the remainder of BB and its successor edges to sinkMBB. 4341 sinkMBB->splice(sinkMBB->begin(), BB, 4342 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 4343 sinkMBB->transferSuccessorsAndUpdatePHIs(BB); 4344 4345 // Next, add the true and fallthrough blocks as its successors. 4346 BB->addSuccessor(copy0MBB); 4347 BB->addSuccessor(sinkMBB); 4348 4349 if (isFPCmp) { 4350 // bc1[tf] cc, sinkMBB 4351 BuildMI(BB, DL, TII->get(Opc)) 4352 .addReg(MI.getOperand(1).getReg()) 4353 .addMBB(sinkMBB); 4354 } else { 4355 // bne rs, $0, sinkMBB 4356 BuildMI(BB, DL, TII->get(Opc)) 4357 .addReg(MI.getOperand(1).getReg()) 4358 .addReg(Mips::ZERO) 4359 .addMBB(sinkMBB); 4360 } 4361 4362 // copy0MBB: 4363 // %FalseValue = ... 4364 // # fallthrough to sinkMBB 4365 BB = copy0MBB; 4366 4367 // Update machine-CFG edges 4368 BB->addSuccessor(sinkMBB); 4369 4370 // sinkMBB: 4371 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] 4372 // ... 4373 BB = sinkMBB; 4374 4375 BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(0).getReg()) 4376 .addReg(MI.getOperand(2).getReg()) 4377 .addMBB(thisMBB) 4378 .addReg(MI.getOperand(3).getReg()) 4379 .addMBB(copy0MBB); 4380 4381 MI.eraseFromParent(); // The pseudo instruction is gone now. 4382 4383 return BB; 4384 } 4385 4386 // FIXME? Maybe this could be a TableGen attribute on some registers and 4387 // this table could be generated automatically from RegInfo. 4388 unsigned MipsTargetLowering::getRegisterByName(const char* RegName, EVT VT, 4389 SelectionDAG &DAG) const { 4390 // Named registers is expected to be fairly rare. For now, just support $28 4391 // since the linux kernel uses it. 4392 if (Subtarget.isGP64bit()) { 4393 unsigned Reg = StringSwitch<unsigned>(RegName) 4394 .Case("$28", Mips::GP_64) 4395 .Default(0); 4396 if (Reg) 4397 return Reg; 4398 } else { 4399 unsigned Reg = StringSwitch<unsigned>(RegName) 4400 .Case("$28", Mips::GP) 4401 .Default(0); 4402 if (Reg) 4403 return Reg; 4404 } 4405 report_fatal_error("Invalid register name global variable"); 4406 } 4407