1 //===- MipsISelLowering.cpp - Mips DAG Lowering Implementation ------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file defines the interfaces that Mips uses to lower LLVM code into a 10 // selection DAG. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "MipsISelLowering.h" 15 #include "MCTargetDesc/MipsBaseInfo.h" 16 #include "MCTargetDesc/MipsInstPrinter.h" 17 #include "MCTargetDesc/MipsMCTargetDesc.h" 18 #include "MipsCCState.h" 19 #include "MipsInstrInfo.h" 20 #include "MipsMachineFunction.h" 21 #include "MipsRegisterInfo.h" 22 #include "MipsSubtarget.h" 23 #include "MipsTargetMachine.h" 24 #include "MipsTargetObjectFile.h" 25 #include "llvm/ADT/APFloat.h" 26 #include "llvm/ADT/ArrayRef.h" 27 #include "llvm/ADT/SmallVector.h" 28 #include "llvm/ADT/Statistic.h" 29 #include "llvm/ADT/StringRef.h" 30 #include "llvm/ADT/StringSwitch.h" 31 #include "llvm/CodeGen/CallingConvLower.h" 32 #include "llvm/CodeGen/FunctionLoweringInfo.h" 33 #include "llvm/CodeGen/ISDOpcodes.h" 34 #include "llvm/CodeGen/MachineBasicBlock.h" 35 #include "llvm/CodeGen/MachineFrameInfo.h" 36 #include "llvm/CodeGen/MachineFunction.h" 37 #include "llvm/CodeGen/MachineInstr.h" 38 #include "llvm/CodeGen/MachineInstrBuilder.h" 39 #include "llvm/CodeGen/MachineJumpTableInfo.h" 40 #include "llvm/CodeGen/MachineMemOperand.h" 41 #include "llvm/CodeGen/MachineOperand.h" 42 #include "llvm/CodeGen/MachineRegisterInfo.h" 43 #include "llvm/CodeGen/RuntimeLibcalls.h" 44 #include "llvm/CodeGen/SelectionDAG.h" 45 #include "llvm/CodeGen/SelectionDAGNodes.h" 46 #include "llvm/CodeGen/TargetFrameLowering.h" 47 #include "llvm/CodeGen/TargetInstrInfo.h" 48 #include "llvm/CodeGen/TargetRegisterInfo.h" 49 #include "llvm/CodeGen/ValueTypes.h" 50 #include "llvm/IR/CallingConv.h" 51 #include "llvm/IR/Constants.h" 52 #include "llvm/IR/DataLayout.h" 53 #include "llvm/IR/DebugLoc.h" 54 #include "llvm/IR/DerivedTypes.h" 55 #include "llvm/IR/Function.h" 56 #include "llvm/IR/GlobalValue.h" 57 #include "llvm/IR/Type.h" 58 #include "llvm/IR/Value.h" 59 #include "llvm/MC/MCContext.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 NoZeroDivCheck("mno-check-zero-division", cl::Hidden, 87 cl::desc("MIPS: Don't trap on integer division by zero."), 88 cl::init(false)); 89 90 extern cl::opt<bool> EmitJalrReloc; 91 92 static const MCPhysReg Mips64DPRegs[8] = { 93 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64, 94 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64 95 }; 96 97 // If I is a shifted mask, set the size (Size) and the first bit of the 98 // mask (Pos), and return true. 99 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11). 100 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) { 101 if (!isShiftedMask_64(I)) 102 return false; 103 104 Size = countPopulation(I); 105 Pos = countTrailingZeros(I); 106 return true; 107 } 108 109 // The MIPS MSA ABI passes vector arguments in the integer register set. 110 // The number of integer registers used is dependant on the ABI used. 111 MVT MipsTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context, 112 CallingConv::ID CC, 113 EVT VT) const { 114 if (!VT.isVector()) 115 return getRegisterType(Context, VT); 116 117 return Subtarget.isABI_O32() || VT.getSizeInBits() == 32 ? MVT::i32 118 : MVT::i64; 119 } 120 121 unsigned MipsTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context, 122 CallingConv::ID CC, 123 EVT VT) const { 124 if (VT.isVector()) 125 return std::max(((unsigned)VT.getSizeInBits() / 126 (Subtarget.isABI_O32() ? 32 : 64)), 127 1U); 128 return MipsTargetLowering::getNumRegisters(Context, VT); 129 } 130 131 unsigned MipsTargetLowering::getVectorTypeBreakdownForCallingConv( 132 LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT, 133 unsigned &NumIntermediates, MVT &RegisterVT) const { 134 // Break down vector types to either 2 i64s or 4 i32s. 135 RegisterVT = getRegisterTypeForCallingConv(Context, CC, VT); 136 IntermediateVT = RegisterVT; 137 NumIntermediates = VT.getSizeInBits() < RegisterVT.getSizeInBits() 138 ? VT.getVectorNumElements() 139 : VT.getSizeInBits() / RegisterVT.getSizeInBits(); 140 141 return NumIntermediates; 142 } 143 144 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const { 145 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>(); 146 return DAG.getRegister(FI->getGlobalBaseReg(), Ty); 147 } 148 149 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty, 150 SelectionDAG &DAG, 151 unsigned Flag) const { 152 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag); 153 } 154 155 SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty, 156 SelectionDAG &DAG, 157 unsigned Flag) const { 158 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag); 159 } 160 161 SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty, 162 SelectionDAG &DAG, 163 unsigned Flag) const { 164 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag); 165 } 166 167 SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty, 168 SelectionDAG &DAG, 169 unsigned Flag) const { 170 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag); 171 } 172 173 SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty, 174 SelectionDAG &DAG, 175 unsigned Flag) const { 176 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(), 177 N->getOffset(), Flag); 178 } 179 180 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const { 181 switch ((MipsISD::NodeType)Opcode) { 182 case MipsISD::FIRST_NUMBER: break; 183 case MipsISD::JmpLink: return "MipsISD::JmpLink"; 184 case MipsISD::TailCall: return "MipsISD::TailCall"; 185 case MipsISD::Highest: return "MipsISD::Highest"; 186 case MipsISD::Higher: return "MipsISD::Higher"; 187 case MipsISD::Hi: return "MipsISD::Hi"; 188 case MipsISD::Lo: return "MipsISD::Lo"; 189 case MipsISD::GotHi: return "MipsISD::GotHi"; 190 case MipsISD::TlsHi: return "MipsISD::TlsHi"; 191 case MipsISD::GPRel: return "MipsISD::GPRel"; 192 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer"; 193 case MipsISD::Ret: return "MipsISD::Ret"; 194 case MipsISD::ERet: return "MipsISD::ERet"; 195 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN"; 196 case MipsISD::FMS: return "MipsISD::FMS"; 197 case MipsISD::FPBrcond: return "MipsISD::FPBrcond"; 198 case MipsISD::FPCmp: return "MipsISD::FPCmp"; 199 case MipsISD::FSELECT: return "MipsISD::FSELECT"; 200 case MipsISD::MTC1_D64: return "MipsISD::MTC1_D64"; 201 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T"; 202 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F"; 203 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP"; 204 case MipsISD::MFHI: return "MipsISD::MFHI"; 205 case MipsISD::MFLO: return "MipsISD::MFLO"; 206 case MipsISD::MTLOHI: return "MipsISD::MTLOHI"; 207 case MipsISD::Mult: return "MipsISD::Mult"; 208 case MipsISD::Multu: return "MipsISD::Multu"; 209 case MipsISD::MAdd: return "MipsISD::MAdd"; 210 case MipsISD::MAddu: return "MipsISD::MAddu"; 211 case MipsISD::MSub: return "MipsISD::MSub"; 212 case MipsISD::MSubu: return "MipsISD::MSubu"; 213 case MipsISD::DivRem: return "MipsISD::DivRem"; 214 case MipsISD::DivRemU: return "MipsISD::DivRemU"; 215 case MipsISD::DivRem16: return "MipsISD::DivRem16"; 216 case MipsISD::DivRemU16: return "MipsISD::DivRemU16"; 217 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64"; 218 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64"; 219 case MipsISD::Wrapper: return "MipsISD::Wrapper"; 220 case MipsISD::DynAlloc: return "MipsISD::DynAlloc"; 221 case MipsISD::Sync: return "MipsISD::Sync"; 222 case MipsISD::Ext: return "MipsISD::Ext"; 223 case MipsISD::Ins: return "MipsISD::Ins"; 224 case MipsISD::CIns: return "MipsISD::CIns"; 225 case MipsISD::LWL: return "MipsISD::LWL"; 226 case MipsISD::LWR: return "MipsISD::LWR"; 227 case MipsISD::SWL: return "MipsISD::SWL"; 228 case MipsISD::SWR: return "MipsISD::SWR"; 229 case MipsISD::LDL: return "MipsISD::LDL"; 230 case MipsISD::LDR: return "MipsISD::LDR"; 231 case MipsISD::SDL: return "MipsISD::SDL"; 232 case MipsISD::SDR: return "MipsISD::SDR"; 233 case MipsISD::EXTP: return "MipsISD::EXTP"; 234 case MipsISD::EXTPDP: return "MipsISD::EXTPDP"; 235 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H"; 236 case MipsISD::EXTR_W: return "MipsISD::EXTR_W"; 237 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W"; 238 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W"; 239 case MipsISD::SHILO: return "MipsISD::SHILO"; 240 case MipsISD::MTHLIP: return "MipsISD::MTHLIP"; 241 case MipsISD::MULSAQ_S_W_PH: return "MipsISD::MULSAQ_S_W_PH"; 242 case MipsISD::MAQ_S_W_PHL: return "MipsISD::MAQ_S_W_PHL"; 243 case MipsISD::MAQ_S_W_PHR: return "MipsISD::MAQ_S_W_PHR"; 244 case MipsISD::MAQ_SA_W_PHL: return "MipsISD::MAQ_SA_W_PHL"; 245 case MipsISD::MAQ_SA_W_PHR: return "MipsISD::MAQ_SA_W_PHR"; 246 case MipsISD::DPAU_H_QBL: return "MipsISD::DPAU_H_QBL"; 247 case MipsISD::DPAU_H_QBR: return "MipsISD::DPAU_H_QBR"; 248 case MipsISD::DPSU_H_QBL: return "MipsISD::DPSU_H_QBL"; 249 case MipsISD::DPSU_H_QBR: return "MipsISD::DPSU_H_QBR"; 250 case MipsISD::DPAQ_S_W_PH: return "MipsISD::DPAQ_S_W_PH"; 251 case MipsISD::DPSQ_S_W_PH: return "MipsISD::DPSQ_S_W_PH"; 252 case MipsISD::DPAQ_SA_L_W: return "MipsISD::DPAQ_SA_L_W"; 253 case MipsISD::DPSQ_SA_L_W: return "MipsISD::DPSQ_SA_L_W"; 254 case MipsISD::DPA_W_PH: return "MipsISD::DPA_W_PH"; 255 case MipsISD::DPS_W_PH: return "MipsISD::DPS_W_PH"; 256 case MipsISD::DPAQX_S_W_PH: return "MipsISD::DPAQX_S_W_PH"; 257 case MipsISD::DPAQX_SA_W_PH: return "MipsISD::DPAQX_SA_W_PH"; 258 case MipsISD::DPAX_W_PH: return "MipsISD::DPAX_W_PH"; 259 case MipsISD::DPSX_W_PH: return "MipsISD::DPSX_W_PH"; 260 case MipsISD::DPSQX_S_W_PH: return "MipsISD::DPSQX_S_W_PH"; 261 case MipsISD::DPSQX_SA_W_PH: return "MipsISD::DPSQX_SA_W_PH"; 262 case MipsISD::MULSA_W_PH: return "MipsISD::MULSA_W_PH"; 263 case MipsISD::MULT: return "MipsISD::MULT"; 264 case MipsISD::MULTU: return "MipsISD::MULTU"; 265 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP"; 266 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP"; 267 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP"; 268 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP"; 269 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP"; 270 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP"; 271 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP"; 272 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP"; 273 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP"; 274 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO"; 275 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO"; 276 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO"; 277 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO"; 278 case MipsISD::VCEQ: return "MipsISD::VCEQ"; 279 case MipsISD::VCLE_S: return "MipsISD::VCLE_S"; 280 case MipsISD::VCLE_U: return "MipsISD::VCLE_U"; 281 case MipsISD::VCLT_S: return "MipsISD::VCLT_S"; 282 case MipsISD::VCLT_U: return "MipsISD::VCLT_U"; 283 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT"; 284 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT"; 285 case MipsISD::VNOR: return "MipsISD::VNOR"; 286 case MipsISD::VSHF: return "MipsISD::VSHF"; 287 case MipsISD::SHF: return "MipsISD::SHF"; 288 case MipsISD::ILVEV: return "MipsISD::ILVEV"; 289 case MipsISD::ILVOD: return "MipsISD::ILVOD"; 290 case MipsISD::ILVL: return "MipsISD::ILVL"; 291 case MipsISD::ILVR: return "MipsISD::ILVR"; 292 case MipsISD::PCKEV: return "MipsISD::PCKEV"; 293 case MipsISD::PCKOD: return "MipsISD::PCKOD"; 294 case MipsISD::INSVE: return "MipsISD::INSVE"; 295 } 296 return nullptr; 297 } 298 299 MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM, 300 const MipsSubtarget &STI) 301 : TargetLowering(TM), Subtarget(STI), ABI(TM.getABI()) { 302 // Mips does not have i1 type, so use i32 for 303 // setcc operations results (slt, sgt, ...). 304 setBooleanContents(ZeroOrOneBooleanContent); 305 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 306 // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA 307 // does. Integer booleans still use 0 and 1. 308 if (Subtarget.hasMips32r6()) 309 setBooleanContents(ZeroOrOneBooleanContent, 310 ZeroOrNegativeOneBooleanContent); 311 312 // Load extented operations for i1 types must be promoted 313 for (MVT VT : MVT::integer_valuetypes()) { 314 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote); 315 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote); 316 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); 317 } 318 319 // MIPS doesn't have extending float->double load/store. Set LoadExtAction 320 // for f32, f16 321 for (MVT VT : MVT::fp_valuetypes()) { 322 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand); 323 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand); 324 } 325 326 // Set LoadExtAction for f16 vectors to Expand 327 for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) { 328 MVT F16VT = MVT::getVectorVT(MVT::f16, VT.getVectorNumElements()); 329 if (F16VT.isValid()) 330 setLoadExtAction(ISD::EXTLOAD, VT, F16VT, Expand); 331 } 332 333 setTruncStoreAction(MVT::f32, MVT::f16, Expand); 334 setTruncStoreAction(MVT::f64, MVT::f16, Expand); 335 336 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 337 338 // Used by legalize types to correctly generate the setcc result. 339 // Without this, every float setcc comes with a AND/OR with the result, 340 // we don't want this, since the fpcmp result goes to a flag register, 341 // which is used implicitly by brcond and select operations. 342 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32); 343 344 // Mips Custom Operations 345 setOperationAction(ISD::BR_JT, MVT::Other, Expand); 346 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 347 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 348 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 349 setOperationAction(ISD::JumpTable, MVT::i32, Custom); 350 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 351 setOperationAction(ISD::SELECT, MVT::f32, Custom); 352 setOperationAction(ISD::SELECT, MVT::f64, Custom); 353 setOperationAction(ISD::SELECT, MVT::i32, Custom); 354 setOperationAction(ISD::SETCC, MVT::f32, Custom); 355 setOperationAction(ISD::SETCC, MVT::f64, Custom); 356 setOperationAction(ISD::BRCOND, MVT::Other, Custom); 357 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 358 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 359 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 360 361 if (!(TM.Options.NoNaNsFPMath || Subtarget.inAbs2008Mode())) { 362 setOperationAction(ISD::FABS, MVT::f32, Custom); 363 setOperationAction(ISD::FABS, MVT::f64, Custom); 364 } 365 366 if (Subtarget.isGP64bit()) { 367 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); 368 setOperationAction(ISD::BlockAddress, MVT::i64, Custom); 369 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); 370 setOperationAction(ISD::JumpTable, MVT::i64, Custom); 371 setOperationAction(ISD::ConstantPool, MVT::i64, Custom); 372 setOperationAction(ISD::SELECT, MVT::i64, Custom); 373 setOperationAction(ISD::LOAD, MVT::i64, Custom); 374 setOperationAction(ISD::STORE, MVT::i64, Custom); 375 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); 376 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom); 377 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom); 378 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom); 379 } 380 381 if (!Subtarget.isGP64bit()) { 382 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 383 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 384 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 385 } 386 387 setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom); 388 if (Subtarget.isGP64bit()) 389 setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom); 390 391 setOperationAction(ISD::SDIV, MVT::i32, Expand); 392 setOperationAction(ISD::SREM, MVT::i32, Expand); 393 setOperationAction(ISD::UDIV, MVT::i32, Expand); 394 setOperationAction(ISD::UREM, MVT::i32, Expand); 395 setOperationAction(ISD::SDIV, MVT::i64, Expand); 396 setOperationAction(ISD::SREM, MVT::i64, Expand); 397 setOperationAction(ISD::UDIV, MVT::i64, Expand); 398 setOperationAction(ISD::UREM, MVT::i64, Expand); 399 400 // Operations not directly supported by Mips. 401 setOperationAction(ISD::BR_CC, MVT::f32, Expand); 402 setOperationAction(ISD::BR_CC, MVT::f64, Expand); 403 setOperationAction(ISD::BR_CC, MVT::i32, Expand); 404 setOperationAction(ISD::BR_CC, MVT::i64, Expand); 405 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand); 406 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand); 407 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand); 408 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand); 409 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); 410 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); 411 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); 412 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); 413 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 414 if (Subtarget.hasCnMips()) { 415 setOperationAction(ISD::CTPOP, MVT::i32, Legal); 416 setOperationAction(ISD::CTPOP, MVT::i64, Legal); 417 } else { 418 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 419 setOperationAction(ISD::CTPOP, MVT::i64, Expand); 420 } 421 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 422 setOperationAction(ISD::CTTZ, MVT::i64, Expand); 423 setOperationAction(ISD::ROTL, MVT::i32, Expand); 424 setOperationAction(ISD::ROTL, MVT::i64, Expand); 425 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 426 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 427 428 if (!Subtarget.hasMips32r2()) 429 setOperationAction(ISD::ROTR, MVT::i32, Expand); 430 431 if (!Subtarget.hasMips64r2()) 432 setOperationAction(ISD::ROTR, MVT::i64, Expand); 433 434 setOperationAction(ISD::FSIN, MVT::f32, Expand); 435 setOperationAction(ISD::FSIN, MVT::f64, Expand); 436 setOperationAction(ISD::FCOS, MVT::f32, Expand); 437 setOperationAction(ISD::FCOS, MVT::f64, Expand); 438 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 439 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 440 setOperationAction(ISD::FPOW, MVT::f32, Expand); 441 setOperationAction(ISD::FPOW, MVT::f64, Expand); 442 setOperationAction(ISD::FLOG, MVT::f32, Expand); 443 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 444 setOperationAction(ISD::FLOG10, MVT::f32, Expand); 445 setOperationAction(ISD::FEXP, MVT::f32, Expand); 446 setOperationAction(ISD::FMA, MVT::f32, Expand); 447 setOperationAction(ISD::FMA, MVT::f64, Expand); 448 setOperationAction(ISD::FREM, MVT::f32, Expand); 449 setOperationAction(ISD::FREM, MVT::f64, Expand); 450 451 // Lower f16 conversion operations into library calls 452 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand); 453 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand); 454 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand); 455 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand); 456 457 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); 458 459 setOperationAction(ISD::VASTART, MVT::Other, Custom); 460 setOperationAction(ISD::VAARG, MVT::Other, Custom); 461 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 462 setOperationAction(ISD::VAEND, MVT::Other, Expand); 463 464 // Use the default for now 465 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 466 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 467 468 if (!Subtarget.isGP64bit()) { 469 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand); 470 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand); 471 } 472 473 if (!Subtarget.hasMips32r2()) { 474 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 475 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 476 } 477 478 // MIPS16 lacks MIPS32's clz and clo instructions. 479 if (!Subtarget.hasMips32() || Subtarget.inMips16Mode()) 480 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 481 if (!Subtarget.hasMips64()) 482 setOperationAction(ISD::CTLZ, MVT::i64, Expand); 483 484 if (!Subtarget.hasMips32r2()) 485 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 486 if (!Subtarget.hasMips64r2()) 487 setOperationAction(ISD::BSWAP, MVT::i64, Expand); 488 489 if (Subtarget.isGP64bit()) { 490 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, MVT::i32, Custom); 491 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, MVT::i32, Custom); 492 setLoadExtAction(ISD::EXTLOAD, MVT::i64, MVT::i32, Custom); 493 setTruncStoreAction(MVT::i64, MVT::i32, Custom); 494 } 495 496 setOperationAction(ISD::TRAP, MVT::Other, Legal); 497 498 setTargetDAGCombine(ISD::SDIVREM); 499 setTargetDAGCombine(ISD::UDIVREM); 500 setTargetDAGCombine(ISD::SELECT); 501 setTargetDAGCombine(ISD::AND); 502 setTargetDAGCombine(ISD::OR); 503 setTargetDAGCombine(ISD::ADD); 504 setTargetDAGCombine(ISD::SUB); 505 setTargetDAGCombine(ISD::AssertZext); 506 setTargetDAGCombine(ISD::SHL); 507 508 if (ABI.IsO32()) { 509 // These libcalls are not available in 32-bit. 510 setLibcallName(RTLIB::SHL_I128, nullptr); 511 setLibcallName(RTLIB::SRL_I128, nullptr); 512 setLibcallName(RTLIB::SRA_I128, nullptr); 513 } 514 515 setMinFunctionAlignment(Subtarget.isGP64bit() ? Align(8) : Align(4)); 516 517 // The arguments on the stack are defined in terms of 4-byte slots on O32 518 // and 8-byte slots on N32/N64. 519 setMinStackArgumentAlignment((ABI.IsN32() || ABI.IsN64()) ? Align(8) 520 : Align(4)); 521 522 setStackPointerRegisterToSaveRestore(ABI.IsN64() ? Mips::SP_64 : Mips::SP); 523 524 MaxStoresPerMemcpy = 16; 525 526 isMicroMips = Subtarget.inMicroMipsMode(); 527 } 528 529 const MipsTargetLowering * 530 MipsTargetLowering::create(const MipsTargetMachine &TM, 531 const MipsSubtarget &STI) { 532 if (STI.inMips16Mode()) 533 return createMips16TargetLowering(TM, STI); 534 535 return createMipsSETargetLowering(TM, STI); 536 } 537 538 // Create a fast isel object. 539 FastISel * 540 MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, 541 const TargetLibraryInfo *libInfo) const { 542 const MipsTargetMachine &TM = 543 static_cast<const MipsTargetMachine &>(funcInfo.MF->getTarget()); 544 545 // We support only the standard encoding [MIPS32,MIPS32R5] ISAs. 546 bool UseFastISel = TM.Options.EnableFastISel && Subtarget.hasMips32() && 547 !Subtarget.hasMips32r6() && !Subtarget.inMips16Mode() && 548 !Subtarget.inMicroMipsMode(); 549 550 // Disable if either of the following is true: 551 // We do not generate PIC, the ABI is not O32, XGOT is being used. 552 if (!TM.isPositionIndependent() || !TM.getABI().IsO32() || 553 Subtarget.useXGOT()) 554 UseFastISel = false; 555 556 return UseFastISel ? Mips::createFastISel(funcInfo, libInfo) : nullptr; 557 } 558 559 EVT MipsTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &, 560 EVT VT) const { 561 if (!VT.isVector()) 562 return MVT::i32; 563 return VT.changeVectorElementTypeToInteger(); 564 } 565 566 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG, 567 TargetLowering::DAGCombinerInfo &DCI, 568 const MipsSubtarget &Subtarget) { 569 if (DCI.isBeforeLegalizeOps()) 570 return SDValue(); 571 572 EVT Ty = N->getValueType(0); 573 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64; 574 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64; 575 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 : 576 MipsISD::DivRemU16; 577 SDLoc DL(N); 578 579 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue, 580 N->getOperand(0), N->getOperand(1)); 581 SDValue InChain = DAG.getEntryNode(); 582 SDValue InGlue = DivRem; 583 584 // insert MFLO 585 if (N->hasAnyUseOfValue(0)) { 586 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty, 587 InGlue); 588 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo); 589 InChain = CopyFromLo.getValue(1); 590 InGlue = CopyFromLo.getValue(2); 591 } 592 593 // insert MFHI 594 if (N->hasAnyUseOfValue(1)) { 595 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL, 596 HI, Ty, InGlue); 597 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi); 598 } 599 600 return SDValue(); 601 } 602 603 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) { 604 switch (CC) { 605 default: llvm_unreachable("Unknown fp condition code!"); 606 case ISD::SETEQ: 607 case ISD::SETOEQ: return Mips::FCOND_OEQ; 608 case ISD::SETUNE: return Mips::FCOND_UNE; 609 case ISD::SETLT: 610 case ISD::SETOLT: return Mips::FCOND_OLT; 611 case ISD::SETGT: 612 case ISD::SETOGT: return Mips::FCOND_OGT; 613 case ISD::SETLE: 614 case ISD::SETOLE: return Mips::FCOND_OLE; 615 case ISD::SETGE: 616 case ISD::SETOGE: return Mips::FCOND_OGE; 617 case ISD::SETULT: return Mips::FCOND_ULT; 618 case ISD::SETULE: return Mips::FCOND_ULE; 619 case ISD::SETUGT: return Mips::FCOND_UGT; 620 case ISD::SETUGE: return Mips::FCOND_UGE; 621 case ISD::SETUO: return Mips::FCOND_UN; 622 case ISD::SETO: return Mips::FCOND_OR; 623 case ISD::SETNE: 624 case ISD::SETONE: return Mips::FCOND_ONE; 625 case ISD::SETUEQ: return Mips::FCOND_UEQ; 626 } 627 } 628 629 /// This function returns true if the floating point conditional branches and 630 /// conditional moves which use condition code CC should be inverted. 631 static bool invertFPCondCodeUser(Mips::CondCode CC) { 632 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT) 633 return false; 634 635 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) && 636 "Illegal Condition Code"); 637 638 return true; 639 } 640 641 // Creates and returns an FPCmp node from a setcc node. 642 // Returns Op if setcc is not a floating point comparison. 643 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) { 644 // must be a SETCC node 645 if (Op.getOpcode() != ISD::SETCC) 646 return Op; 647 648 SDValue LHS = Op.getOperand(0); 649 650 if (!LHS.getValueType().isFloatingPoint()) 651 return Op; 652 653 SDValue RHS = Op.getOperand(1); 654 SDLoc DL(Op); 655 656 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of 657 // node if necessary. 658 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); 659 660 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS, 661 DAG.getConstant(condCodeToFCC(CC), DL, MVT::i32)); 662 } 663 664 // Creates and returns a CMovFPT/F node. 665 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True, 666 SDValue False, const SDLoc &DL) { 667 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2)); 668 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue()); 669 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 670 671 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL, 672 True.getValueType(), True, FCC0, False, Cond); 673 } 674 675 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG, 676 TargetLowering::DAGCombinerInfo &DCI, 677 const MipsSubtarget &Subtarget) { 678 if (DCI.isBeforeLegalizeOps()) 679 return SDValue(); 680 681 SDValue SetCC = N->getOperand(0); 682 683 if ((SetCC.getOpcode() != ISD::SETCC) || 684 !SetCC.getOperand(0).getValueType().isInteger()) 685 return SDValue(); 686 687 SDValue False = N->getOperand(2); 688 EVT FalseTy = False.getValueType(); 689 690 if (!FalseTy.isInteger()) 691 return SDValue(); 692 693 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(False); 694 695 // If the RHS (False) is 0, we swap the order of the operands 696 // of ISD::SELECT (obviously also inverting the condition) so that we can 697 // take advantage of conditional moves using the $0 register. 698 // Example: 699 // return (a != 0) ? x : 0; 700 // load $reg, x 701 // movz $reg, $0, a 702 if (!FalseC) 703 return SDValue(); 704 705 const SDLoc DL(N); 706 707 if (!FalseC->getZExtValue()) { 708 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 709 SDValue True = N->getOperand(1); 710 711 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 712 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 713 714 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True); 715 } 716 717 // If both operands are integer constants there's a possibility that we 718 // can do some interesting optimizations. 719 SDValue True = N->getOperand(1); 720 ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(True); 721 722 if (!TrueC || !True.getValueType().isInteger()) 723 return SDValue(); 724 725 // We'll also ignore MVT::i64 operands as this optimizations proves 726 // to be ineffective because of the required sign extensions as the result 727 // of a SETCC operator is always MVT::i32 for non-vector types. 728 if (True.getValueType() == MVT::i64) 729 return SDValue(); 730 731 int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue(); 732 733 // 1) (a < x) ? y : y-1 734 // slti $reg1, a, x 735 // addiu $reg2, $reg1, y-1 736 if (Diff == 1) 737 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, False); 738 739 // 2) (a < x) ? y-1 : y 740 // slti $reg1, a, x 741 // xor $reg1, $reg1, 1 742 // addiu $reg2, $reg1, y-1 743 if (Diff == -1) { 744 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 745 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 746 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 747 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, True); 748 } 749 750 // Could not optimize. 751 return SDValue(); 752 } 753 754 static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG, 755 TargetLowering::DAGCombinerInfo &DCI, 756 const MipsSubtarget &Subtarget) { 757 if (DCI.isBeforeLegalizeOps()) 758 return SDValue(); 759 760 SDValue ValueIfTrue = N->getOperand(0), ValueIfFalse = N->getOperand(2); 761 762 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(ValueIfFalse); 763 if (!FalseC || FalseC->getZExtValue()) 764 return SDValue(); 765 766 // Since RHS (False) is 0, we swap the order of the True/False operands 767 // (obviously also inverting the condition) so that we can 768 // take advantage of conditional moves using the $0 register. 769 // Example: 770 // return (a != 0) ? x : 0; 771 // load $reg, x 772 // movz $reg, $0, a 773 unsigned Opc = (N->getOpcode() == MipsISD::CMovFP_T) ? MipsISD::CMovFP_F : 774 MipsISD::CMovFP_T; 775 776 SDValue FCC = N->getOperand(1), Glue = N->getOperand(3); 777 return DAG.getNode(Opc, SDLoc(N), ValueIfFalse.getValueType(), 778 ValueIfFalse, FCC, ValueIfTrue, Glue); 779 } 780 781 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, 782 TargetLowering::DAGCombinerInfo &DCI, 783 const MipsSubtarget &Subtarget) { 784 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) 785 return SDValue(); 786 787 SDValue FirstOperand = N->getOperand(0); 788 unsigned FirstOperandOpc = FirstOperand.getOpcode(); 789 SDValue Mask = N->getOperand(1); 790 EVT ValTy = N->getValueType(0); 791 SDLoc DL(N); 792 793 uint64_t Pos = 0, SMPos, SMSize; 794 ConstantSDNode *CN; 795 SDValue NewOperand; 796 unsigned Opc; 797 798 // Op's second operand must be a shifted mask. 799 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) || 800 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize)) 801 return SDValue(); 802 803 if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL) { 804 // Pattern match EXT. 805 // $dst = and ((sra or srl) $src , pos), (2**size - 1) 806 // => ext $dst, $src, pos, size 807 808 // The second operand of the shift must be an immediate. 809 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1)))) 810 return SDValue(); 811 812 Pos = CN->getZExtValue(); 813 814 // Return if the shifted mask does not start at bit 0 or the sum of its size 815 // and Pos exceeds the word's size. 816 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits()) 817 return SDValue(); 818 819 Opc = MipsISD::Ext; 820 NewOperand = FirstOperand.getOperand(0); 821 } else if (FirstOperandOpc == ISD::SHL && Subtarget.hasCnMips()) { 822 // Pattern match CINS. 823 // $dst = and (shl $src , pos), mask 824 // => cins $dst, $src, pos, size 825 // mask is a shifted mask with consecutive 1's, pos = shift amount, 826 // size = population count. 827 828 // The second operand of the shift must be an immediate. 829 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1)))) 830 return SDValue(); 831 832 Pos = CN->getZExtValue(); 833 834 if (SMPos != Pos || Pos >= ValTy.getSizeInBits() || SMSize >= 32 || 835 Pos + SMSize > ValTy.getSizeInBits()) 836 return SDValue(); 837 838 NewOperand = FirstOperand.getOperand(0); 839 // SMSize is 'location' (position) in this case, not size. 840 SMSize--; 841 Opc = MipsISD::CIns; 842 } else { 843 // Pattern match EXT. 844 // $dst = and $src, (2**size - 1) , if size > 16 845 // => ext $dst, $src, pos, size , pos = 0 846 847 // If the mask is <= 0xffff, andi can be used instead. 848 if (CN->getZExtValue() <= 0xffff) 849 return SDValue(); 850 851 // Return if the mask doesn't start at position 0. 852 if (SMPos) 853 return SDValue(); 854 855 Opc = MipsISD::Ext; 856 NewOperand = FirstOperand; 857 } 858 return DAG.getNode(Opc, DL, ValTy, NewOperand, 859 DAG.getConstant(Pos, DL, MVT::i32), 860 DAG.getConstant(SMSize, DL, MVT::i32)); 861 } 862 863 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, 864 TargetLowering::DAGCombinerInfo &DCI, 865 const MipsSubtarget &Subtarget) { 866 // Pattern match INS. 867 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1), 868 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1 869 // => ins $dst, $src, size, pos, $src1 870 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) 871 return SDValue(); 872 873 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1); 874 uint64_t SMPos0, SMSize0, SMPos1, SMSize1; 875 ConstantSDNode *CN, *CN1; 876 877 // See if Op's first operand matches (and $src1 , mask0). 878 if (And0.getOpcode() != ISD::AND) 879 return SDValue(); 880 881 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) || 882 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0)) 883 return SDValue(); 884 885 // See if Op's second operand matches (and (shl $src, pos), mask1). 886 if (And1.getOpcode() == ISD::AND && 887 And1.getOperand(0).getOpcode() == ISD::SHL) { 888 889 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) || 890 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1)) 891 return SDValue(); 892 893 // The shift masks must have the same position and size. 894 if (SMPos0 != SMPos1 || SMSize0 != SMSize1) 895 return SDValue(); 896 897 SDValue Shl = And1.getOperand(0); 898 899 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1)))) 900 return SDValue(); 901 902 unsigned Shamt = CN->getZExtValue(); 903 904 // Return if the shift amount and the first bit position of mask are not the 905 // same. 906 EVT ValTy = N->getValueType(0); 907 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits())) 908 return SDValue(); 909 910 SDLoc DL(N); 911 return DAG.getNode(MipsISD::Ins, DL, ValTy, Shl.getOperand(0), 912 DAG.getConstant(SMPos0, DL, MVT::i32), 913 DAG.getConstant(SMSize0, DL, MVT::i32), 914 And0.getOperand(0)); 915 } else { 916 // Pattern match DINS. 917 // $dst = or (and $src, mask0), mask1 918 // where mask0 = ((1 << SMSize0) -1) << SMPos0 919 // => dins $dst, $src, pos, size 920 if (~CN->getSExtValue() == ((((int64_t)1 << SMSize0) - 1) << SMPos0) && 921 ((SMSize0 + SMPos0 <= 64 && Subtarget.hasMips64r2()) || 922 (SMSize0 + SMPos0 <= 32))) { 923 // Check if AND instruction has constant as argument 924 bool isConstCase = And1.getOpcode() != ISD::AND; 925 if (And1.getOpcode() == ISD::AND) { 926 if (!(CN1 = dyn_cast<ConstantSDNode>(And1->getOperand(1)))) 927 return SDValue(); 928 } else { 929 if (!(CN1 = dyn_cast<ConstantSDNode>(N->getOperand(1)))) 930 return SDValue(); 931 } 932 // Don't generate INS if constant OR operand doesn't fit into bits 933 // cleared by constant AND operand. 934 if (CN->getSExtValue() & CN1->getSExtValue()) 935 return SDValue(); 936 937 SDLoc DL(N); 938 EVT ValTy = N->getOperand(0)->getValueType(0); 939 SDValue Const1; 940 SDValue SrlX; 941 if (!isConstCase) { 942 Const1 = DAG.getConstant(SMPos0, DL, MVT::i32); 943 SrlX = DAG.getNode(ISD::SRL, DL, And1->getValueType(0), And1, Const1); 944 } 945 return DAG.getNode( 946 MipsISD::Ins, DL, N->getValueType(0), 947 isConstCase 948 ? DAG.getConstant(CN1->getSExtValue() >> SMPos0, DL, ValTy) 949 : SrlX, 950 DAG.getConstant(SMPos0, DL, MVT::i32), 951 DAG.getConstant(ValTy.getSizeInBits() / 8 < 8 ? SMSize0 & 31 952 : SMSize0, 953 DL, MVT::i32), 954 And0->getOperand(0)); 955 956 } 957 return SDValue(); 958 } 959 } 960 961 static SDValue performMADD_MSUBCombine(SDNode *ROOTNode, SelectionDAG &CurDAG, 962 const MipsSubtarget &Subtarget) { 963 // ROOTNode must have a multiplication as an operand for the match to be 964 // successful. 965 if (ROOTNode->getOperand(0).getOpcode() != ISD::MUL && 966 ROOTNode->getOperand(1).getOpcode() != ISD::MUL) 967 return SDValue(); 968 969 // We don't handle vector types here. 970 if (ROOTNode->getValueType(0).isVector()) 971 return SDValue(); 972 973 // For MIPS64, madd / msub instructions are inefficent to use with 64 bit 974 // arithmetic. E.g. 975 // (add (mul a b) c) => 976 // let res = (madd (mthi (drotr c 32))x(mtlo c) a b) in 977 // MIPS64: (or (dsll (mfhi res) 32) (dsrl (dsll (mflo res) 32) 32) 978 // or 979 // MIPS64R2: (dins (mflo res) (mfhi res) 32 32) 980 // 981 // The overhead of setting up the Hi/Lo registers and reassembling the 982 // result makes this a dubious optimzation for MIPS64. The core of the 983 // problem is that Hi/Lo contain the upper and lower 32 bits of the 984 // operand and result. 985 // 986 // It requires a chain of 4 add/mul for MIPS64R2 to get better code 987 // density than doing it naively, 5 for MIPS64. Additionally, using 988 // madd/msub on MIPS64 requires the operands actually be 32 bit sign 989 // extended operands, not true 64 bit values. 990 // 991 // FIXME: For the moment, disable this completely for MIPS64. 992 if (Subtarget.hasMips64()) 993 return SDValue(); 994 995 SDValue Mult = ROOTNode->getOperand(0).getOpcode() == ISD::MUL 996 ? ROOTNode->getOperand(0) 997 : ROOTNode->getOperand(1); 998 999 SDValue AddOperand = ROOTNode->getOperand(0).getOpcode() == ISD::MUL 1000 ? ROOTNode->getOperand(1) 1001 : ROOTNode->getOperand(0); 1002 1003 // Transform this to a MADD only if the user of this node is the add. 1004 // If there are other users of the mul, this function returns here. 1005 if (!Mult.hasOneUse()) 1006 return SDValue(); 1007 1008 // maddu and madd are unusual instructions in that on MIPS64 bits 63..31 1009 // must be in canonical form, i.e. sign extended. For MIPS32, the operands 1010 // of the multiply must have 32 or more sign bits, otherwise we cannot 1011 // perform this optimization. We have to check this here as we're performing 1012 // this optimization pre-legalization. 1013 SDValue MultLHS = Mult->getOperand(0); 1014 SDValue MultRHS = Mult->getOperand(1); 1015 1016 bool IsSigned = MultLHS->getOpcode() == ISD::SIGN_EXTEND && 1017 MultRHS->getOpcode() == ISD::SIGN_EXTEND; 1018 bool IsUnsigned = MultLHS->getOpcode() == ISD::ZERO_EXTEND && 1019 MultRHS->getOpcode() == ISD::ZERO_EXTEND; 1020 1021 if (!IsSigned && !IsUnsigned) 1022 return SDValue(); 1023 1024 // Initialize accumulator. 1025 SDLoc DL(ROOTNode); 1026 SDValue TopHalf; 1027 SDValue BottomHalf; 1028 BottomHalf = CurDAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, AddOperand, 1029 CurDAG.getIntPtrConstant(0, DL)); 1030 1031 TopHalf = CurDAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, AddOperand, 1032 CurDAG.getIntPtrConstant(1, DL)); 1033 SDValue ACCIn = CurDAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped, 1034 BottomHalf, 1035 TopHalf); 1036 1037 // Create MipsMAdd(u) / MipsMSub(u) node. 1038 bool IsAdd = ROOTNode->getOpcode() == ISD::ADD; 1039 unsigned Opcode = IsAdd ? (IsUnsigned ? MipsISD::MAddu : MipsISD::MAdd) 1040 : (IsUnsigned ? MipsISD::MSubu : MipsISD::MSub); 1041 SDValue MAddOps[3] = { 1042 CurDAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mult->getOperand(0)), 1043 CurDAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mult->getOperand(1)), ACCIn}; 1044 EVT VTs[2] = {MVT::i32, MVT::i32}; 1045 SDValue MAdd = CurDAG.getNode(Opcode, DL, VTs, MAddOps); 1046 1047 SDValue ResLo = CurDAG.getNode(MipsISD::MFLO, DL, MVT::i32, MAdd); 1048 SDValue ResHi = CurDAG.getNode(MipsISD::MFHI, DL, MVT::i32, MAdd); 1049 SDValue Combined = 1050 CurDAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResLo, ResHi); 1051 return Combined; 1052 } 1053 1054 static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG, 1055 TargetLowering::DAGCombinerInfo &DCI, 1056 const MipsSubtarget &Subtarget) { 1057 // (sub v0 (mul v1, v2)) => (msub v1, v2, v0) 1058 if (DCI.isBeforeLegalizeOps()) { 1059 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() && 1060 !Subtarget.inMips16Mode() && N->getValueType(0) == MVT::i64) 1061 return performMADD_MSUBCombine(N, DAG, Subtarget); 1062 1063 return SDValue(); 1064 } 1065 1066 return SDValue(); 1067 } 1068 1069 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG, 1070 TargetLowering::DAGCombinerInfo &DCI, 1071 const MipsSubtarget &Subtarget) { 1072 // (add v0 (mul v1, v2)) => (madd v1, v2, v0) 1073 if (DCI.isBeforeLegalizeOps()) { 1074 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() && 1075 !Subtarget.inMips16Mode() && N->getValueType(0) == MVT::i64) 1076 return performMADD_MSUBCombine(N, DAG, Subtarget); 1077 1078 return SDValue(); 1079 } 1080 1081 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt)) 1082 SDValue Add = N->getOperand(1); 1083 1084 if (Add.getOpcode() != ISD::ADD) 1085 return SDValue(); 1086 1087 SDValue Lo = Add.getOperand(1); 1088 1089 if ((Lo.getOpcode() != MipsISD::Lo) || 1090 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable)) 1091 return SDValue(); 1092 1093 EVT ValTy = N->getValueType(0); 1094 SDLoc DL(N); 1095 1096 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0), 1097 Add.getOperand(0)); 1098 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo); 1099 } 1100 1101 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG, 1102 TargetLowering::DAGCombinerInfo &DCI, 1103 const MipsSubtarget &Subtarget) { 1104 // Pattern match CINS. 1105 // $dst = shl (and $src , imm), pos 1106 // => cins $dst, $src, pos, size 1107 1108 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasCnMips()) 1109 return SDValue(); 1110 1111 SDValue FirstOperand = N->getOperand(0); 1112 unsigned FirstOperandOpc = FirstOperand.getOpcode(); 1113 SDValue SecondOperand = N->getOperand(1); 1114 EVT ValTy = N->getValueType(0); 1115 SDLoc DL(N); 1116 1117 uint64_t Pos = 0, SMPos, SMSize; 1118 ConstantSDNode *CN; 1119 SDValue NewOperand; 1120 1121 // The second operand of the shift must be an immediate. 1122 if (!(CN = dyn_cast<ConstantSDNode>(SecondOperand))) 1123 return SDValue(); 1124 1125 Pos = CN->getZExtValue(); 1126 1127 if (Pos >= ValTy.getSizeInBits()) 1128 return SDValue(); 1129 1130 if (FirstOperandOpc != ISD::AND) 1131 return SDValue(); 1132 1133 // AND's second operand must be a shifted mask. 1134 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))) || 1135 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize)) 1136 return SDValue(); 1137 1138 // Return if the shifted mask does not start at bit 0 or the sum of its size 1139 // and Pos exceeds the word's size. 1140 if (SMPos != 0 || SMSize > 32 || Pos + SMSize > ValTy.getSizeInBits()) 1141 return SDValue(); 1142 1143 NewOperand = FirstOperand.getOperand(0); 1144 // SMSize is 'location' (position) in this case, not size. 1145 SMSize--; 1146 1147 return DAG.getNode(MipsISD::CIns, DL, ValTy, NewOperand, 1148 DAG.getConstant(Pos, DL, MVT::i32), 1149 DAG.getConstant(SMSize, DL, MVT::i32)); 1150 } 1151 1152 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) 1153 const { 1154 SelectionDAG &DAG = DCI.DAG; 1155 unsigned Opc = N->getOpcode(); 1156 1157 switch (Opc) { 1158 default: break; 1159 case ISD::SDIVREM: 1160 case ISD::UDIVREM: 1161 return performDivRemCombine(N, DAG, DCI, Subtarget); 1162 case ISD::SELECT: 1163 return performSELECTCombine(N, DAG, DCI, Subtarget); 1164 case MipsISD::CMovFP_F: 1165 case MipsISD::CMovFP_T: 1166 return performCMovFPCombine(N, DAG, DCI, Subtarget); 1167 case ISD::AND: 1168 return performANDCombine(N, DAG, DCI, Subtarget); 1169 case ISD::OR: 1170 return performORCombine(N, DAG, DCI, Subtarget); 1171 case ISD::ADD: 1172 return performADDCombine(N, DAG, DCI, Subtarget); 1173 case ISD::SHL: 1174 return performSHLCombine(N, DAG, DCI, Subtarget); 1175 case ISD::SUB: 1176 return performSUBCombine(N, DAG, DCI, Subtarget); 1177 } 1178 1179 return SDValue(); 1180 } 1181 1182 bool MipsTargetLowering::isCheapToSpeculateCttz() const { 1183 return Subtarget.hasMips32(); 1184 } 1185 1186 bool MipsTargetLowering::isCheapToSpeculateCtlz() const { 1187 return Subtarget.hasMips32(); 1188 } 1189 1190 bool MipsTargetLowering::shouldFoldConstantShiftPairToMask( 1191 const SDNode *N, CombineLevel Level) const { 1192 if (N->getOperand(0).getValueType().isVector()) 1193 return false; 1194 return true; 1195 } 1196 1197 void 1198 MipsTargetLowering::LowerOperationWrapper(SDNode *N, 1199 SmallVectorImpl<SDValue> &Results, 1200 SelectionDAG &DAG) const { 1201 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 1202 1203 if (Res) 1204 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I) 1205 Results.push_back(Res.getValue(I)); 1206 } 1207 1208 void 1209 MipsTargetLowering::ReplaceNodeResults(SDNode *N, 1210 SmallVectorImpl<SDValue> &Results, 1211 SelectionDAG &DAG) const { 1212 return LowerOperationWrapper(N, Results, DAG); 1213 } 1214 1215 SDValue MipsTargetLowering:: 1216 LowerOperation(SDValue Op, SelectionDAG &DAG) const 1217 { 1218 switch (Op.getOpcode()) 1219 { 1220 case ISD::BRCOND: return lowerBRCOND(Op, DAG); 1221 case ISD::ConstantPool: return lowerConstantPool(Op, DAG); 1222 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG); 1223 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG); 1224 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG); 1225 case ISD::JumpTable: return lowerJumpTable(Op, DAG); 1226 case ISD::SELECT: return lowerSELECT(Op, DAG); 1227 case ISD::SETCC: return lowerSETCC(Op, DAG); 1228 case ISD::VASTART: return lowerVASTART(Op, DAG); 1229 case ISD::VAARG: return lowerVAARG(Op, DAG); 1230 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG); 1231 case ISD::FABS: return lowerFABS(Op, DAG); 1232 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG); 1233 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG); 1234 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG); 1235 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG); 1236 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG); 1237 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true); 1238 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false); 1239 case ISD::LOAD: return lowerLOAD(Op, DAG); 1240 case ISD::STORE: return lowerSTORE(Op, DAG); 1241 case ISD::EH_DWARF_CFA: return lowerEH_DWARF_CFA(Op, DAG); 1242 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG); 1243 } 1244 return SDValue(); 1245 } 1246 1247 //===----------------------------------------------------------------------===// 1248 // Lower helper functions 1249 //===----------------------------------------------------------------------===// 1250 1251 // addLiveIn - This helper function adds the specified physical register to the 1252 // MachineFunction as a live in value. It also creates a corresponding 1253 // virtual register for it. 1254 static unsigned 1255 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC) 1256 { 1257 Register VReg = MF.getRegInfo().createVirtualRegister(RC); 1258 MF.getRegInfo().addLiveIn(PReg, VReg); 1259 return VReg; 1260 } 1261 1262 static MachineBasicBlock *insertDivByZeroTrap(MachineInstr &MI, 1263 MachineBasicBlock &MBB, 1264 const TargetInstrInfo &TII, 1265 bool Is64Bit, bool IsMicroMips) { 1266 if (NoZeroDivCheck) 1267 return &MBB; 1268 1269 // Insert instruction "teq $divisor_reg, $zero, 7". 1270 MachineBasicBlock::iterator I(MI); 1271 MachineInstrBuilder MIB; 1272 MachineOperand &Divisor = MI.getOperand(2); 1273 MIB = BuildMI(MBB, std::next(I), MI.getDebugLoc(), 1274 TII.get(IsMicroMips ? Mips::TEQ_MM : Mips::TEQ)) 1275 .addReg(Divisor.getReg(), getKillRegState(Divisor.isKill())) 1276 .addReg(Mips::ZERO) 1277 .addImm(7); 1278 1279 // Use the 32-bit sub-register if this is a 64-bit division. 1280 if (Is64Bit) 1281 MIB->getOperand(0).setSubReg(Mips::sub_32); 1282 1283 // Clear Divisor's kill flag. 1284 Divisor.setIsKill(false); 1285 1286 // We would normally delete the original instruction here but in this case 1287 // we only needed to inject an additional instruction rather than replace it. 1288 1289 return &MBB; 1290 } 1291 1292 MachineBasicBlock * 1293 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, 1294 MachineBasicBlock *BB) const { 1295 switch (MI.getOpcode()) { 1296 default: 1297 llvm_unreachable("Unexpected instr type to insert"); 1298 case Mips::ATOMIC_LOAD_ADD_I8: 1299 return emitAtomicBinaryPartword(MI, BB, 1); 1300 case Mips::ATOMIC_LOAD_ADD_I16: 1301 return emitAtomicBinaryPartword(MI, BB, 2); 1302 case Mips::ATOMIC_LOAD_ADD_I32: 1303 return emitAtomicBinary(MI, BB); 1304 case Mips::ATOMIC_LOAD_ADD_I64: 1305 return emitAtomicBinary(MI, BB); 1306 1307 case Mips::ATOMIC_LOAD_AND_I8: 1308 return emitAtomicBinaryPartword(MI, BB, 1); 1309 case Mips::ATOMIC_LOAD_AND_I16: 1310 return emitAtomicBinaryPartword(MI, BB, 2); 1311 case Mips::ATOMIC_LOAD_AND_I32: 1312 return emitAtomicBinary(MI, BB); 1313 case Mips::ATOMIC_LOAD_AND_I64: 1314 return emitAtomicBinary(MI, BB); 1315 1316 case Mips::ATOMIC_LOAD_OR_I8: 1317 return emitAtomicBinaryPartword(MI, BB, 1); 1318 case Mips::ATOMIC_LOAD_OR_I16: 1319 return emitAtomicBinaryPartword(MI, BB, 2); 1320 case Mips::ATOMIC_LOAD_OR_I32: 1321 return emitAtomicBinary(MI, BB); 1322 case Mips::ATOMIC_LOAD_OR_I64: 1323 return emitAtomicBinary(MI, BB); 1324 1325 case Mips::ATOMIC_LOAD_XOR_I8: 1326 return emitAtomicBinaryPartword(MI, BB, 1); 1327 case Mips::ATOMIC_LOAD_XOR_I16: 1328 return emitAtomicBinaryPartword(MI, BB, 2); 1329 case Mips::ATOMIC_LOAD_XOR_I32: 1330 return emitAtomicBinary(MI, BB); 1331 case Mips::ATOMIC_LOAD_XOR_I64: 1332 return emitAtomicBinary(MI, BB); 1333 1334 case Mips::ATOMIC_LOAD_NAND_I8: 1335 return emitAtomicBinaryPartword(MI, BB, 1); 1336 case Mips::ATOMIC_LOAD_NAND_I16: 1337 return emitAtomicBinaryPartword(MI, BB, 2); 1338 case Mips::ATOMIC_LOAD_NAND_I32: 1339 return emitAtomicBinary(MI, BB); 1340 case Mips::ATOMIC_LOAD_NAND_I64: 1341 return emitAtomicBinary(MI, BB); 1342 1343 case Mips::ATOMIC_LOAD_SUB_I8: 1344 return emitAtomicBinaryPartword(MI, BB, 1); 1345 case Mips::ATOMIC_LOAD_SUB_I16: 1346 return emitAtomicBinaryPartword(MI, BB, 2); 1347 case Mips::ATOMIC_LOAD_SUB_I32: 1348 return emitAtomicBinary(MI, BB); 1349 case Mips::ATOMIC_LOAD_SUB_I64: 1350 return emitAtomicBinary(MI, BB); 1351 1352 case Mips::ATOMIC_SWAP_I8: 1353 return emitAtomicBinaryPartword(MI, BB, 1); 1354 case Mips::ATOMIC_SWAP_I16: 1355 return emitAtomicBinaryPartword(MI, BB, 2); 1356 case Mips::ATOMIC_SWAP_I32: 1357 return emitAtomicBinary(MI, BB); 1358 case Mips::ATOMIC_SWAP_I64: 1359 return emitAtomicBinary(MI, BB); 1360 1361 case Mips::ATOMIC_CMP_SWAP_I8: 1362 return emitAtomicCmpSwapPartword(MI, BB, 1); 1363 case Mips::ATOMIC_CMP_SWAP_I16: 1364 return emitAtomicCmpSwapPartword(MI, BB, 2); 1365 case Mips::ATOMIC_CMP_SWAP_I32: 1366 return emitAtomicCmpSwap(MI, BB); 1367 case Mips::ATOMIC_CMP_SWAP_I64: 1368 return emitAtomicCmpSwap(MI, BB); 1369 1370 case Mips::ATOMIC_LOAD_MIN_I8: 1371 return emitAtomicBinaryPartword(MI, BB, 1); 1372 case Mips::ATOMIC_LOAD_MIN_I16: 1373 return emitAtomicBinaryPartword(MI, BB, 2); 1374 case Mips::ATOMIC_LOAD_MIN_I32: 1375 return emitAtomicBinary(MI, BB); 1376 case Mips::ATOMIC_LOAD_MIN_I64: 1377 return emitAtomicBinary(MI, BB); 1378 1379 case Mips::ATOMIC_LOAD_MAX_I8: 1380 return emitAtomicBinaryPartword(MI, BB, 1); 1381 case Mips::ATOMIC_LOAD_MAX_I16: 1382 return emitAtomicBinaryPartword(MI, BB, 2); 1383 case Mips::ATOMIC_LOAD_MAX_I32: 1384 return emitAtomicBinary(MI, BB); 1385 case Mips::ATOMIC_LOAD_MAX_I64: 1386 return emitAtomicBinary(MI, BB); 1387 1388 case Mips::ATOMIC_LOAD_UMIN_I8: 1389 return emitAtomicBinaryPartword(MI, BB, 1); 1390 case Mips::ATOMIC_LOAD_UMIN_I16: 1391 return emitAtomicBinaryPartword(MI, BB, 2); 1392 case Mips::ATOMIC_LOAD_UMIN_I32: 1393 return emitAtomicBinary(MI, BB); 1394 case Mips::ATOMIC_LOAD_UMIN_I64: 1395 return emitAtomicBinary(MI, BB); 1396 1397 case Mips::ATOMIC_LOAD_UMAX_I8: 1398 return emitAtomicBinaryPartword(MI, BB, 1); 1399 case Mips::ATOMIC_LOAD_UMAX_I16: 1400 return emitAtomicBinaryPartword(MI, BB, 2); 1401 case Mips::ATOMIC_LOAD_UMAX_I32: 1402 return emitAtomicBinary(MI, BB); 1403 case Mips::ATOMIC_LOAD_UMAX_I64: 1404 return emitAtomicBinary(MI, BB); 1405 1406 case Mips::PseudoSDIV: 1407 case Mips::PseudoUDIV: 1408 case Mips::DIV: 1409 case Mips::DIVU: 1410 case Mips::MOD: 1411 case Mips::MODU: 1412 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false, 1413 false); 1414 case Mips::SDIV_MM_Pseudo: 1415 case Mips::UDIV_MM_Pseudo: 1416 case Mips::SDIV_MM: 1417 case Mips::UDIV_MM: 1418 case Mips::DIV_MMR6: 1419 case Mips::DIVU_MMR6: 1420 case Mips::MOD_MMR6: 1421 case Mips::MODU_MMR6: 1422 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false, true); 1423 case Mips::PseudoDSDIV: 1424 case Mips::PseudoDUDIV: 1425 case Mips::DDIV: 1426 case Mips::DDIVU: 1427 case Mips::DMOD: 1428 case Mips::DMODU: 1429 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), true, false); 1430 1431 case Mips::PseudoSELECT_I: 1432 case Mips::PseudoSELECT_I64: 1433 case Mips::PseudoSELECT_S: 1434 case Mips::PseudoSELECT_D32: 1435 case Mips::PseudoSELECT_D64: 1436 return emitPseudoSELECT(MI, BB, false, Mips::BNE); 1437 case Mips::PseudoSELECTFP_F_I: 1438 case Mips::PseudoSELECTFP_F_I64: 1439 case Mips::PseudoSELECTFP_F_S: 1440 case Mips::PseudoSELECTFP_F_D32: 1441 case Mips::PseudoSELECTFP_F_D64: 1442 return emitPseudoSELECT(MI, BB, true, Mips::BC1F); 1443 case Mips::PseudoSELECTFP_T_I: 1444 case Mips::PseudoSELECTFP_T_I64: 1445 case Mips::PseudoSELECTFP_T_S: 1446 case Mips::PseudoSELECTFP_T_D32: 1447 case Mips::PseudoSELECTFP_T_D64: 1448 return emitPseudoSELECT(MI, BB, true, Mips::BC1T); 1449 case Mips::PseudoD_SELECT_I: 1450 case Mips::PseudoD_SELECT_I64: 1451 return emitPseudoD_SELECT(MI, BB); 1452 } 1453 } 1454 1455 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and 1456 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true) 1457 MachineBasicBlock * 1458 MipsTargetLowering::emitAtomicBinary(MachineInstr &MI, 1459 MachineBasicBlock *BB) const { 1460 1461 MachineFunction *MF = BB->getParent(); 1462 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1463 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1464 DebugLoc DL = MI.getDebugLoc(); 1465 1466 unsigned AtomicOp; 1467 bool NeedsAdditionalReg = false; 1468 switch (MI.getOpcode()) { 1469 case Mips::ATOMIC_LOAD_ADD_I32: 1470 AtomicOp = Mips::ATOMIC_LOAD_ADD_I32_POSTRA; 1471 break; 1472 case Mips::ATOMIC_LOAD_SUB_I32: 1473 AtomicOp = Mips::ATOMIC_LOAD_SUB_I32_POSTRA; 1474 break; 1475 case Mips::ATOMIC_LOAD_AND_I32: 1476 AtomicOp = Mips::ATOMIC_LOAD_AND_I32_POSTRA; 1477 break; 1478 case Mips::ATOMIC_LOAD_OR_I32: 1479 AtomicOp = Mips::ATOMIC_LOAD_OR_I32_POSTRA; 1480 break; 1481 case Mips::ATOMIC_LOAD_XOR_I32: 1482 AtomicOp = Mips::ATOMIC_LOAD_XOR_I32_POSTRA; 1483 break; 1484 case Mips::ATOMIC_LOAD_NAND_I32: 1485 AtomicOp = Mips::ATOMIC_LOAD_NAND_I32_POSTRA; 1486 break; 1487 case Mips::ATOMIC_SWAP_I32: 1488 AtomicOp = Mips::ATOMIC_SWAP_I32_POSTRA; 1489 break; 1490 case Mips::ATOMIC_LOAD_ADD_I64: 1491 AtomicOp = Mips::ATOMIC_LOAD_ADD_I64_POSTRA; 1492 break; 1493 case Mips::ATOMIC_LOAD_SUB_I64: 1494 AtomicOp = Mips::ATOMIC_LOAD_SUB_I64_POSTRA; 1495 break; 1496 case Mips::ATOMIC_LOAD_AND_I64: 1497 AtomicOp = Mips::ATOMIC_LOAD_AND_I64_POSTRA; 1498 break; 1499 case Mips::ATOMIC_LOAD_OR_I64: 1500 AtomicOp = Mips::ATOMIC_LOAD_OR_I64_POSTRA; 1501 break; 1502 case Mips::ATOMIC_LOAD_XOR_I64: 1503 AtomicOp = Mips::ATOMIC_LOAD_XOR_I64_POSTRA; 1504 break; 1505 case Mips::ATOMIC_LOAD_NAND_I64: 1506 AtomicOp = Mips::ATOMIC_LOAD_NAND_I64_POSTRA; 1507 break; 1508 case Mips::ATOMIC_SWAP_I64: 1509 AtomicOp = Mips::ATOMIC_SWAP_I64_POSTRA; 1510 break; 1511 case Mips::ATOMIC_LOAD_MIN_I32: 1512 AtomicOp = Mips::ATOMIC_LOAD_MIN_I32_POSTRA; 1513 NeedsAdditionalReg = true; 1514 break; 1515 case Mips::ATOMIC_LOAD_MAX_I32: 1516 AtomicOp = Mips::ATOMIC_LOAD_MAX_I32_POSTRA; 1517 NeedsAdditionalReg = true; 1518 break; 1519 case Mips::ATOMIC_LOAD_UMIN_I32: 1520 AtomicOp = Mips::ATOMIC_LOAD_UMIN_I32_POSTRA; 1521 NeedsAdditionalReg = true; 1522 break; 1523 case Mips::ATOMIC_LOAD_UMAX_I32: 1524 AtomicOp = Mips::ATOMIC_LOAD_UMAX_I32_POSTRA; 1525 NeedsAdditionalReg = true; 1526 break; 1527 case Mips::ATOMIC_LOAD_MIN_I64: 1528 AtomicOp = Mips::ATOMIC_LOAD_MIN_I64_POSTRA; 1529 NeedsAdditionalReg = true; 1530 break; 1531 case Mips::ATOMIC_LOAD_MAX_I64: 1532 AtomicOp = Mips::ATOMIC_LOAD_MAX_I64_POSTRA; 1533 NeedsAdditionalReg = true; 1534 break; 1535 case Mips::ATOMIC_LOAD_UMIN_I64: 1536 AtomicOp = Mips::ATOMIC_LOAD_UMIN_I64_POSTRA; 1537 NeedsAdditionalReg = true; 1538 break; 1539 case Mips::ATOMIC_LOAD_UMAX_I64: 1540 AtomicOp = Mips::ATOMIC_LOAD_UMAX_I64_POSTRA; 1541 NeedsAdditionalReg = true; 1542 break; 1543 default: 1544 llvm_unreachable("Unknown pseudo atomic for replacement!"); 1545 } 1546 1547 Register OldVal = MI.getOperand(0).getReg(); 1548 Register Ptr = MI.getOperand(1).getReg(); 1549 Register Incr = MI.getOperand(2).getReg(); 1550 Register Scratch = RegInfo.createVirtualRegister(RegInfo.getRegClass(OldVal)); 1551 1552 MachineBasicBlock::iterator II(MI); 1553 1554 // The scratch registers here with the EarlyClobber | Define | Implicit 1555 // flags is used to persuade the register allocator and the machine 1556 // verifier to accept the usage of this register. This has to be a real 1557 // register which has an UNDEF value but is dead after the instruction which 1558 // is unique among the registers chosen for the instruction. 1559 1560 // The EarlyClobber flag has the semantic properties that the operand it is 1561 // attached to is clobbered before the rest of the inputs are read. Hence it 1562 // must be unique among the operands to the instruction. 1563 // The Define flag is needed to coerce the machine verifier that an Undef 1564 // value isn't a problem. 1565 // The Dead flag is needed as the value in scratch isn't used by any other 1566 // instruction. Kill isn't used as Dead is more precise. 1567 // The implicit flag is here due to the interaction between the other flags 1568 // and the machine verifier. 1569 1570 // For correctness purpose, a new pseudo is introduced here. We need this 1571 // new pseudo, so that FastRegisterAllocator does not see an ll/sc sequence 1572 // that is spread over >1 basic blocks. A register allocator which 1573 // introduces (or any codegen infact) a store, can violate the expectations 1574 // of the hardware. 1575 // 1576 // An atomic read-modify-write sequence starts with a linked load 1577 // instruction and ends with a store conditional instruction. The atomic 1578 // read-modify-write sequence fails if any of the following conditions 1579 // occur between the execution of ll and sc: 1580 // * A coherent store is completed by another process or coherent I/O 1581 // module into the block of synchronizable physical memory containing 1582 // the word. The size and alignment of the block is 1583 // implementation-dependent. 1584 // * A coherent store is executed between an LL and SC sequence on the 1585 // same processor to the block of synchornizable physical memory 1586 // containing the word. 1587 // 1588 1589 Register PtrCopy = RegInfo.createVirtualRegister(RegInfo.getRegClass(Ptr)); 1590 Register IncrCopy = RegInfo.createVirtualRegister(RegInfo.getRegClass(Incr)); 1591 1592 BuildMI(*BB, II, DL, TII->get(Mips::COPY), IncrCopy).addReg(Incr); 1593 BuildMI(*BB, II, DL, TII->get(Mips::COPY), PtrCopy).addReg(Ptr); 1594 1595 MachineInstrBuilder MIB = 1596 BuildMI(*BB, II, DL, TII->get(AtomicOp)) 1597 .addReg(OldVal, RegState::Define | RegState::EarlyClobber) 1598 .addReg(PtrCopy) 1599 .addReg(IncrCopy) 1600 .addReg(Scratch, RegState::Define | RegState::EarlyClobber | 1601 RegState::Implicit | RegState::Dead); 1602 if (NeedsAdditionalReg) { 1603 Register Scratch2 = 1604 RegInfo.createVirtualRegister(RegInfo.getRegClass(OldVal)); 1605 MIB.addReg(Scratch2, RegState::Define | RegState::EarlyClobber | 1606 RegState::Implicit | RegState::Dead); 1607 } 1608 1609 MI.eraseFromParent(); 1610 1611 return BB; 1612 } 1613 1614 MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg( 1615 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg, 1616 unsigned SrcReg) const { 1617 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1618 const DebugLoc &DL = MI.getDebugLoc(); 1619 1620 if (Subtarget.hasMips32r2() && Size == 1) { 1621 BuildMI(BB, DL, TII->get(Mips::SEB), DstReg).addReg(SrcReg); 1622 return BB; 1623 } 1624 1625 if (Subtarget.hasMips32r2() && Size == 2) { 1626 BuildMI(BB, DL, TII->get(Mips::SEH), DstReg).addReg(SrcReg); 1627 return BB; 1628 } 1629 1630 MachineFunction *MF = BB->getParent(); 1631 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1632 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1633 Register ScrReg = RegInfo.createVirtualRegister(RC); 1634 1635 assert(Size < 32); 1636 int64_t ShiftImm = 32 - (Size * 8); 1637 1638 BuildMI(BB, DL, TII->get(Mips::SLL), ScrReg).addReg(SrcReg).addImm(ShiftImm); 1639 BuildMI(BB, DL, TII->get(Mips::SRA), DstReg).addReg(ScrReg).addImm(ShiftImm); 1640 1641 return BB; 1642 } 1643 1644 MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword( 1645 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const { 1646 assert((Size == 1 || Size == 2) && 1647 "Unsupported size for EmitAtomicBinaryPartial."); 1648 1649 MachineFunction *MF = BB->getParent(); 1650 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1651 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1652 const bool ArePtrs64bit = ABI.ArePtrs64bit(); 1653 const TargetRegisterClass *RCp = 1654 getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32); 1655 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1656 DebugLoc DL = MI.getDebugLoc(); 1657 1658 Register Dest = MI.getOperand(0).getReg(); 1659 Register Ptr = MI.getOperand(1).getReg(); 1660 Register Incr = MI.getOperand(2).getReg(); 1661 1662 Register AlignedAddr = RegInfo.createVirtualRegister(RCp); 1663 Register ShiftAmt = RegInfo.createVirtualRegister(RC); 1664 Register Mask = RegInfo.createVirtualRegister(RC); 1665 Register Mask2 = RegInfo.createVirtualRegister(RC); 1666 Register Incr2 = RegInfo.createVirtualRegister(RC); 1667 Register MaskLSB2 = RegInfo.createVirtualRegister(RCp); 1668 Register PtrLSB2 = RegInfo.createVirtualRegister(RC); 1669 Register MaskUpper = RegInfo.createVirtualRegister(RC); 1670 Register Scratch = RegInfo.createVirtualRegister(RC); 1671 Register Scratch2 = RegInfo.createVirtualRegister(RC); 1672 Register Scratch3 = RegInfo.createVirtualRegister(RC); 1673 1674 unsigned AtomicOp = 0; 1675 bool NeedsAdditionalReg = false; 1676 switch (MI.getOpcode()) { 1677 case Mips::ATOMIC_LOAD_NAND_I8: 1678 AtomicOp = Mips::ATOMIC_LOAD_NAND_I8_POSTRA; 1679 break; 1680 case Mips::ATOMIC_LOAD_NAND_I16: 1681 AtomicOp = Mips::ATOMIC_LOAD_NAND_I16_POSTRA; 1682 break; 1683 case Mips::ATOMIC_SWAP_I8: 1684 AtomicOp = Mips::ATOMIC_SWAP_I8_POSTRA; 1685 break; 1686 case Mips::ATOMIC_SWAP_I16: 1687 AtomicOp = Mips::ATOMIC_SWAP_I16_POSTRA; 1688 break; 1689 case Mips::ATOMIC_LOAD_ADD_I8: 1690 AtomicOp = Mips::ATOMIC_LOAD_ADD_I8_POSTRA; 1691 break; 1692 case Mips::ATOMIC_LOAD_ADD_I16: 1693 AtomicOp = Mips::ATOMIC_LOAD_ADD_I16_POSTRA; 1694 break; 1695 case Mips::ATOMIC_LOAD_SUB_I8: 1696 AtomicOp = Mips::ATOMIC_LOAD_SUB_I8_POSTRA; 1697 break; 1698 case Mips::ATOMIC_LOAD_SUB_I16: 1699 AtomicOp = Mips::ATOMIC_LOAD_SUB_I16_POSTRA; 1700 break; 1701 case Mips::ATOMIC_LOAD_AND_I8: 1702 AtomicOp = Mips::ATOMIC_LOAD_AND_I8_POSTRA; 1703 break; 1704 case Mips::ATOMIC_LOAD_AND_I16: 1705 AtomicOp = Mips::ATOMIC_LOAD_AND_I16_POSTRA; 1706 break; 1707 case Mips::ATOMIC_LOAD_OR_I8: 1708 AtomicOp = Mips::ATOMIC_LOAD_OR_I8_POSTRA; 1709 break; 1710 case Mips::ATOMIC_LOAD_OR_I16: 1711 AtomicOp = Mips::ATOMIC_LOAD_OR_I16_POSTRA; 1712 break; 1713 case Mips::ATOMIC_LOAD_XOR_I8: 1714 AtomicOp = Mips::ATOMIC_LOAD_XOR_I8_POSTRA; 1715 break; 1716 case Mips::ATOMIC_LOAD_XOR_I16: 1717 AtomicOp = Mips::ATOMIC_LOAD_XOR_I16_POSTRA; 1718 break; 1719 case Mips::ATOMIC_LOAD_MIN_I8: 1720 AtomicOp = Mips::ATOMIC_LOAD_MIN_I8_POSTRA; 1721 NeedsAdditionalReg = true; 1722 break; 1723 case Mips::ATOMIC_LOAD_MIN_I16: 1724 AtomicOp = Mips::ATOMIC_LOAD_MIN_I16_POSTRA; 1725 NeedsAdditionalReg = true; 1726 break; 1727 case Mips::ATOMIC_LOAD_MAX_I8: 1728 AtomicOp = Mips::ATOMIC_LOAD_MAX_I8_POSTRA; 1729 NeedsAdditionalReg = true; 1730 break; 1731 case Mips::ATOMIC_LOAD_MAX_I16: 1732 AtomicOp = Mips::ATOMIC_LOAD_MAX_I16_POSTRA; 1733 NeedsAdditionalReg = true; 1734 break; 1735 case Mips::ATOMIC_LOAD_UMIN_I8: 1736 AtomicOp = Mips::ATOMIC_LOAD_UMIN_I8_POSTRA; 1737 NeedsAdditionalReg = true; 1738 break; 1739 case Mips::ATOMIC_LOAD_UMIN_I16: 1740 AtomicOp = Mips::ATOMIC_LOAD_UMIN_I16_POSTRA; 1741 NeedsAdditionalReg = true; 1742 break; 1743 case Mips::ATOMIC_LOAD_UMAX_I8: 1744 AtomicOp = Mips::ATOMIC_LOAD_UMAX_I8_POSTRA; 1745 NeedsAdditionalReg = true; 1746 break; 1747 case Mips::ATOMIC_LOAD_UMAX_I16: 1748 AtomicOp = Mips::ATOMIC_LOAD_UMAX_I16_POSTRA; 1749 NeedsAdditionalReg = true; 1750 break; 1751 default: 1752 llvm_unreachable("Unknown subword atomic pseudo for expansion!"); 1753 } 1754 1755 // insert new blocks after the current block 1756 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1757 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1758 MachineFunction::iterator It = ++BB->getIterator(); 1759 MF->insert(It, exitMBB); 1760 1761 // Transfer the remainder of BB and its successor edges to exitMBB. 1762 exitMBB->splice(exitMBB->begin(), BB, 1763 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1764 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1765 1766 BB->addSuccessor(exitMBB, BranchProbability::getOne()); 1767 1768 // thisMBB: 1769 // addiu masklsb2,$0,-4 # 0xfffffffc 1770 // and alignedaddr,ptr,masklsb2 1771 // andi ptrlsb2,ptr,3 1772 // sll shiftamt,ptrlsb2,3 1773 // ori maskupper,$0,255 # 0xff 1774 // sll mask,maskupper,shiftamt 1775 // nor mask2,$0,mask 1776 // sll incr2,incr,shiftamt 1777 1778 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1779 BuildMI(BB, DL, TII->get(ABI.GetPtrAddiuOp()), MaskLSB2) 1780 .addReg(ABI.GetNullPtr()).addImm(-4); 1781 BuildMI(BB, DL, TII->get(ABI.GetPtrAndOp()), AlignedAddr) 1782 .addReg(Ptr).addReg(MaskLSB2); 1783 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2) 1784 .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3); 1785 if (Subtarget.isLittle()) { 1786 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1787 } else { 1788 Register Off = RegInfo.createVirtualRegister(RC); 1789 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1790 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1791 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1792 } 1793 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1794 .addReg(Mips::ZERO).addImm(MaskImm); 1795 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1796 .addReg(MaskUpper).addReg(ShiftAmt); 1797 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1798 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt); 1799 1800 1801 // The purposes of the flags on the scratch registers is explained in 1802 // emitAtomicBinary. In summary, we need a scratch register which is going to 1803 // be undef, that is unique among registers chosen for the instruction. 1804 1805 MachineInstrBuilder MIB = 1806 BuildMI(BB, DL, TII->get(AtomicOp)) 1807 .addReg(Dest, RegState::Define | RegState::EarlyClobber) 1808 .addReg(AlignedAddr) 1809 .addReg(Incr2) 1810 .addReg(Mask) 1811 .addReg(Mask2) 1812 .addReg(ShiftAmt) 1813 .addReg(Scratch, RegState::EarlyClobber | RegState::Define | 1814 RegState::Dead | RegState::Implicit) 1815 .addReg(Scratch2, RegState::EarlyClobber | RegState::Define | 1816 RegState::Dead | RegState::Implicit) 1817 .addReg(Scratch3, RegState::EarlyClobber | RegState::Define | 1818 RegState::Dead | RegState::Implicit); 1819 if (NeedsAdditionalReg) { 1820 Register Scratch4 = RegInfo.createVirtualRegister(RC); 1821 MIB.addReg(Scratch4, RegState::EarlyClobber | RegState::Define | 1822 RegState::Dead | RegState::Implicit); 1823 } 1824 1825 MI.eraseFromParent(); // The instruction is gone now. 1826 1827 return exitMBB; 1828 } 1829 1830 // Lower atomic compare and swap to a pseudo instruction, taking care to 1831 // define a scratch register for the pseudo instruction's expansion. The 1832 // instruction is expanded after the register allocator as to prevent 1833 // the insertion of stores between the linked load and the store conditional. 1834 1835 MachineBasicBlock * 1836 MipsTargetLowering::emitAtomicCmpSwap(MachineInstr &MI, 1837 MachineBasicBlock *BB) const { 1838 1839 assert((MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 || 1840 MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I64) && 1841 "Unsupported atomic pseudo for EmitAtomicCmpSwap."); 1842 1843 const unsigned Size = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 ? 4 : 8; 1844 1845 MachineFunction *MF = BB->getParent(); 1846 MachineRegisterInfo &MRI = MF->getRegInfo(); 1847 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 1848 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1849 DebugLoc DL = MI.getDebugLoc(); 1850 1851 unsigned AtomicOp = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 1852 ? Mips::ATOMIC_CMP_SWAP_I32_POSTRA 1853 : Mips::ATOMIC_CMP_SWAP_I64_POSTRA; 1854 Register Dest = MI.getOperand(0).getReg(); 1855 Register Ptr = MI.getOperand(1).getReg(); 1856 Register OldVal = MI.getOperand(2).getReg(); 1857 Register NewVal = MI.getOperand(3).getReg(); 1858 1859 Register Scratch = MRI.createVirtualRegister(RC); 1860 MachineBasicBlock::iterator II(MI); 1861 1862 // We need to create copies of the various registers and kill them at the 1863 // atomic pseudo. If the copies are not made, when the atomic is expanded 1864 // after fast register allocation, the spills will end up outside of the 1865 // blocks that their values are defined in, causing livein errors. 1866 1867 Register PtrCopy = MRI.createVirtualRegister(MRI.getRegClass(Ptr)); 1868 Register OldValCopy = MRI.createVirtualRegister(MRI.getRegClass(OldVal)); 1869 Register NewValCopy = MRI.createVirtualRegister(MRI.getRegClass(NewVal)); 1870 1871 BuildMI(*BB, II, DL, TII->get(Mips::COPY), PtrCopy).addReg(Ptr); 1872 BuildMI(*BB, II, DL, TII->get(Mips::COPY), OldValCopy).addReg(OldVal); 1873 BuildMI(*BB, II, DL, TII->get(Mips::COPY), NewValCopy).addReg(NewVal); 1874 1875 // The purposes of the flags on the scratch registers is explained in 1876 // emitAtomicBinary. In summary, we need a scratch register which is going to 1877 // be undef, that is unique among registers chosen for the instruction. 1878 1879 BuildMI(*BB, II, DL, TII->get(AtomicOp)) 1880 .addReg(Dest, RegState::Define | RegState::EarlyClobber) 1881 .addReg(PtrCopy, RegState::Kill) 1882 .addReg(OldValCopy, RegState::Kill) 1883 .addReg(NewValCopy, RegState::Kill) 1884 .addReg(Scratch, RegState::EarlyClobber | RegState::Define | 1885 RegState::Dead | RegState::Implicit); 1886 1887 MI.eraseFromParent(); // The instruction is gone now. 1888 1889 return BB; 1890 } 1891 1892 MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwapPartword( 1893 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const { 1894 assert((Size == 1 || Size == 2) && 1895 "Unsupported size for EmitAtomicCmpSwapPartial."); 1896 1897 MachineFunction *MF = BB->getParent(); 1898 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1899 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1900 const bool ArePtrs64bit = ABI.ArePtrs64bit(); 1901 const TargetRegisterClass *RCp = 1902 getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32); 1903 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 1904 DebugLoc DL = MI.getDebugLoc(); 1905 1906 Register Dest = MI.getOperand(0).getReg(); 1907 Register Ptr = MI.getOperand(1).getReg(); 1908 Register CmpVal = MI.getOperand(2).getReg(); 1909 Register NewVal = MI.getOperand(3).getReg(); 1910 1911 Register AlignedAddr = RegInfo.createVirtualRegister(RCp); 1912 Register ShiftAmt = RegInfo.createVirtualRegister(RC); 1913 Register Mask = RegInfo.createVirtualRegister(RC); 1914 Register Mask2 = RegInfo.createVirtualRegister(RC); 1915 Register ShiftedCmpVal = RegInfo.createVirtualRegister(RC); 1916 Register ShiftedNewVal = RegInfo.createVirtualRegister(RC); 1917 Register MaskLSB2 = RegInfo.createVirtualRegister(RCp); 1918 Register PtrLSB2 = RegInfo.createVirtualRegister(RC); 1919 Register MaskUpper = RegInfo.createVirtualRegister(RC); 1920 Register MaskedCmpVal = RegInfo.createVirtualRegister(RC); 1921 Register MaskedNewVal = RegInfo.createVirtualRegister(RC); 1922 unsigned AtomicOp = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I8 1923 ? Mips::ATOMIC_CMP_SWAP_I8_POSTRA 1924 : Mips::ATOMIC_CMP_SWAP_I16_POSTRA; 1925 1926 // The scratch registers here with the EarlyClobber | Define | Dead | Implicit 1927 // flags are used to coerce the register allocator and the machine verifier to 1928 // accept the usage of these registers. 1929 // The EarlyClobber flag has the semantic properties that the operand it is 1930 // attached to is clobbered before the rest of the inputs are read. Hence it 1931 // must be unique among the operands to the instruction. 1932 // The Define flag is needed to coerce the machine verifier that an Undef 1933 // value isn't a problem. 1934 // The Dead flag is needed as the value in scratch isn't used by any other 1935 // instruction. Kill isn't used as Dead is more precise. 1936 Register Scratch = RegInfo.createVirtualRegister(RC); 1937 Register Scratch2 = RegInfo.createVirtualRegister(RC); 1938 1939 // insert new blocks after the current block 1940 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1941 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1942 MachineFunction::iterator It = ++BB->getIterator(); 1943 MF->insert(It, exitMBB); 1944 1945 // Transfer the remainder of BB and its successor edges to exitMBB. 1946 exitMBB->splice(exitMBB->begin(), BB, 1947 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1948 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1949 1950 BB->addSuccessor(exitMBB, BranchProbability::getOne()); 1951 1952 // thisMBB: 1953 // addiu masklsb2,$0,-4 # 0xfffffffc 1954 // and alignedaddr,ptr,masklsb2 1955 // andi ptrlsb2,ptr,3 1956 // xori ptrlsb2,ptrlsb2,3 # Only for BE 1957 // sll shiftamt,ptrlsb2,3 1958 // ori maskupper,$0,255 # 0xff 1959 // sll mask,maskupper,shiftamt 1960 // nor mask2,$0,mask 1961 // andi maskedcmpval,cmpval,255 1962 // sll shiftedcmpval,maskedcmpval,shiftamt 1963 // andi maskednewval,newval,255 1964 // sll shiftednewval,maskednewval,shiftamt 1965 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1966 BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::DADDiu : Mips::ADDiu), MaskLSB2) 1967 .addReg(ABI.GetNullPtr()).addImm(-4); 1968 BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::AND64 : Mips::AND), AlignedAddr) 1969 .addReg(Ptr).addReg(MaskLSB2); 1970 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2) 1971 .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3); 1972 if (Subtarget.isLittle()) { 1973 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1974 } else { 1975 Register Off = RegInfo.createVirtualRegister(RC); 1976 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1977 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1978 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1979 } 1980 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1981 .addReg(Mips::ZERO).addImm(MaskImm); 1982 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1983 .addReg(MaskUpper).addReg(ShiftAmt); 1984 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1985 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal) 1986 .addReg(CmpVal).addImm(MaskImm); 1987 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal) 1988 .addReg(MaskedCmpVal).addReg(ShiftAmt); 1989 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal) 1990 .addReg(NewVal).addImm(MaskImm); 1991 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal) 1992 .addReg(MaskedNewVal).addReg(ShiftAmt); 1993 1994 // The purposes of the flags on the scratch registers are explained in 1995 // emitAtomicBinary. In summary, we need a scratch register which is going to 1996 // be undef, that is unique among the register chosen for the instruction. 1997 1998 BuildMI(BB, DL, TII->get(AtomicOp)) 1999 .addReg(Dest, RegState::Define | RegState::EarlyClobber) 2000 .addReg(AlignedAddr) 2001 .addReg(Mask) 2002 .addReg(ShiftedCmpVal) 2003 .addReg(Mask2) 2004 .addReg(ShiftedNewVal) 2005 .addReg(ShiftAmt) 2006 .addReg(Scratch, RegState::EarlyClobber | RegState::Define | 2007 RegState::Dead | RegState::Implicit) 2008 .addReg(Scratch2, RegState::EarlyClobber | RegState::Define | 2009 RegState::Dead | RegState::Implicit); 2010 2011 MI.eraseFromParent(); // The instruction is gone now. 2012 2013 return exitMBB; 2014 } 2015 2016 SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const { 2017 // The first operand is the chain, the second is the condition, the third is 2018 // the block to branch to if the condition is true. 2019 SDValue Chain = Op.getOperand(0); 2020 SDValue Dest = Op.getOperand(2); 2021 SDLoc DL(Op); 2022 2023 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 2024 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1)); 2025 2026 // Return if flag is not set by a floating point comparison. 2027 if (CondRes.getOpcode() != MipsISD::FPCmp) 2028 return Op; 2029 2030 SDValue CCNode = CondRes.getOperand(2); 2031 Mips::CondCode CC = 2032 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue(); 2033 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T; 2034 SDValue BrCode = DAG.getConstant(Opc, DL, MVT::i32); 2035 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 2036 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode, 2037 FCC0, Dest, CondRes); 2038 } 2039 2040 SDValue MipsTargetLowering:: 2041 lowerSELECT(SDValue Op, SelectionDAG &DAG) const 2042 { 2043 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 2044 SDValue Cond = createFPCmp(DAG, Op.getOperand(0)); 2045 2046 // Return if flag is not set by a floating point comparison. 2047 if (Cond.getOpcode() != MipsISD::FPCmp) 2048 return Op; 2049 2050 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2), 2051 SDLoc(Op)); 2052 } 2053 2054 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const { 2055 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 2056 SDValue Cond = createFPCmp(DAG, Op); 2057 2058 assert(Cond.getOpcode() == MipsISD::FPCmp && 2059 "Floating point operand expected."); 2060 2061 SDLoc DL(Op); 2062 SDValue True = DAG.getConstant(1, DL, MVT::i32); 2063 SDValue False = DAG.getConstant(0, DL, MVT::i32); 2064 2065 return createCMovFP(DAG, Cond, True, False, DL); 2066 } 2067 2068 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op, 2069 SelectionDAG &DAG) const { 2070 EVT Ty = Op.getValueType(); 2071 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op); 2072 const GlobalValue *GV = N->getGlobal(); 2073 2074 if (!isPositionIndependent()) { 2075 const MipsTargetObjectFile *TLOF = 2076 static_cast<const MipsTargetObjectFile *>( 2077 getTargetMachine().getObjFileLowering()); 2078 const GlobalObject *GO = GV->getBaseObject(); 2079 if (GO && TLOF->IsGlobalInSmallSection(GO, getTargetMachine())) 2080 // %gp_rel relocation 2081 return getAddrGPRel(N, SDLoc(N), Ty, DAG, ABI.IsN64()); 2082 2083 // %hi/%lo relocation 2084 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2085 // %highest/%higher/%hi/%lo relocation 2086 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2087 } 2088 2089 // Every other architecture would use shouldAssumeDSOLocal in here, but 2090 // mips is special. 2091 // * In PIC code mips requires got loads even for local statics! 2092 // * To save on got entries, for local statics the got entry contains the 2093 // page and an additional add instruction takes care of the low bits. 2094 // * It is legal to access a hidden symbol with a non hidden undefined, 2095 // so one cannot guarantee that all access to a hidden symbol will know 2096 // it is hidden. 2097 // * Mips linkers don't support creating a page and a full got entry for 2098 // the same symbol. 2099 // * Given all that, we have to use a full got entry for hidden symbols :-( 2100 if (GV->hasLocalLinkage()) 2101 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2102 2103 if (Subtarget.useXGOT()) 2104 return getAddrGlobalLargeGOT( 2105 N, SDLoc(N), Ty, DAG, MipsII::MO_GOT_HI16, MipsII::MO_GOT_LO16, 2106 DAG.getEntryNode(), 2107 MachinePointerInfo::getGOT(DAG.getMachineFunction())); 2108 2109 return getAddrGlobal( 2110 N, SDLoc(N), Ty, DAG, 2111 (ABI.IsN32() || ABI.IsN64()) ? MipsII::MO_GOT_DISP : MipsII::MO_GOT, 2112 DAG.getEntryNode(), MachinePointerInfo::getGOT(DAG.getMachineFunction())); 2113 } 2114 2115 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op, 2116 SelectionDAG &DAG) const { 2117 BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op); 2118 EVT Ty = Op.getValueType(); 2119 2120 if (!isPositionIndependent()) 2121 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2122 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2123 2124 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2125 } 2126 2127 SDValue MipsTargetLowering:: 2128 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const 2129 { 2130 // If the relocation model is PIC, use the General Dynamic TLS Model or 2131 // Local Dynamic TLS model, otherwise use the Initial Exec or 2132 // Local Exec TLS Model. 2133 2134 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 2135 if (DAG.getTarget().useEmulatedTLS()) 2136 return LowerToTLSEmulatedModel(GA, DAG); 2137 2138 SDLoc DL(GA); 2139 const GlobalValue *GV = GA->getGlobal(); 2140 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 2141 2142 TLSModel::Model model = getTargetMachine().getTLSModel(GV); 2143 2144 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { 2145 // General Dynamic and Local Dynamic TLS Model. 2146 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM 2147 : MipsII::MO_TLSGD; 2148 2149 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag); 2150 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, 2151 getGlobalReg(DAG, PtrVT), TGA); 2152 unsigned PtrSize = PtrVT.getSizeInBits(); 2153 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize); 2154 2155 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT); 2156 2157 ArgListTy Args; 2158 ArgListEntry Entry; 2159 Entry.Node = Argument; 2160 Entry.Ty = PtrTy; 2161 Args.push_back(Entry); 2162 2163 TargetLowering::CallLoweringInfo CLI(DAG); 2164 CLI.setDebugLoc(DL) 2165 .setChain(DAG.getEntryNode()) 2166 .setLibCallee(CallingConv::C, PtrTy, TlsGetAddr, std::move(Args)); 2167 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 2168 2169 SDValue Ret = CallResult.first; 2170 2171 if (model != TLSModel::LocalDynamic) 2172 return Ret; 2173 2174 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2175 MipsII::MO_DTPREL_HI); 2176 SDValue Hi = DAG.getNode(MipsISD::TlsHi, DL, PtrVT, TGAHi); 2177 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2178 MipsII::MO_DTPREL_LO); 2179 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 2180 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret); 2181 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo); 2182 } 2183 2184 SDValue Offset; 2185 if (model == TLSModel::InitialExec) { 2186 // Initial Exec TLS Model 2187 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2188 MipsII::MO_GOTTPREL); 2189 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT), 2190 TGA); 2191 Offset = 2192 DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), TGA, MachinePointerInfo()); 2193 } else { 2194 // Local Exec TLS Model 2195 assert(model == TLSModel::LocalExec); 2196 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2197 MipsII::MO_TPREL_HI); 2198 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 2199 MipsII::MO_TPREL_LO); 2200 SDValue Hi = DAG.getNode(MipsISD::TlsHi, DL, PtrVT, TGAHi); 2201 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 2202 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo); 2203 } 2204 2205 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT); 2206 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset); 2207 } 2208 2209 SDValue MipsTargetLowering:: 2210 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const 2211 { 2212 JumpTableSDNode *N = cast<JumpTableSDNode>(Op); 2213 EVT Ty = Op.getValueType(); 2214 2215 if (!isPositionIndependent()) 2216 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2217 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2218 2219 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2220 } 2221 2222 SDValue MipsTargetLowering:: 2223 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const 2224 { 2225 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op); 2226 EVT Ty = Op.getValueType(); 2227 2228 if (!isPositionIndependent()) { 2229 const MipsTargetObjectFile *TLOF = 2230 static_cast<const MipsTargetObjectFile *>( 2231 getTargetMachine().getObjFileLowering()); 2232 2233 if (TLOF->IsConstantInSmallSection(DAG.getDataLayout(), N->getConstVal(), 2234 getTargetMachine())) 2235 // %gp_rel relocation 2236 return getAddrGPRel(N, SDLoc(N), Ty, DAG, ABI.IsN64()); 2237 2238 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 2239 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 2240 } 2241 2242 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64()); 2243 } 2244 2245 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const { 2246 MachineFunction &MF = DAG.getMachineFunction(); 2247 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 2248 2249 SDLoc DL(Op); 2250 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), 2251 getPointerTy(MF.getDataLayout())); 2252 2253 // vastart just stores the address of the VarArgsFrameIndex slot into the 2254 // memory location argument. 2255 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 2256 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1), 2257 MachinePointerInfo(SV)); 2258 } 2259 2260 SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const { 2261 SDNode *Node = Op.getNode(); 2262 EVT VT = Node->getValueType(0); 2263 SDValue Chain = Node->getOperand(0); 2264 SDValue VAListPtr = Node->getOperand(1); 2265 const Align Align = 2266 llvm::MaybeAlign(Node->getConstantOperandVal(3)).valueOrOne(); 2267 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue(); 2268 SDLoc DL(Node); 2269 unsigned ArgSlotSizeInBytes = (ABI.IsN32() || ABI.IsN64()) ? 8 : 4; 2270 2271 SDValue VAListLoad = DAG.getLoad(getPointerTy(DAG.getDataLayout()), DL, Chain, 2272 VAListPtr, MachinePointerInfo(SV)); 2273 SDValue VAList = VAListLoad; 2274 2275 // Re-align the pointer if necessary. 2276 // It should only ever be necessary for 64-bit types on O32 since the minimum 2277 // argument alignment is the same as the maximum type alignment for N32/N64. 2278 // 2279 // FIXME: We currently align too often. The code generator doesn't notice 2280 // when the pointer is still aligned from the last va_arg (or pair of 2281 // va_args for the i64 on O32 case). 2282 if (Align > getMinStackArgumentAlignment()) { 2283 VAList = DAG.getNode( 2284 ISD::ADD, DL, VAList.getValueType(), VAList, 2285 DAG.getConstant(Align.value() - 1, DL, VAList.getValueType())); 2286 2287 VAList = DAG.getNode( 2288 ISD::AND, DL, VAList.getValueType(), VAList, 2289 DAG.getConstant(-(int64_t)Align.value(), DL, VAList.getValueType())); 2290 } 2291 2292 // Increment the pointer, VAList, to the next vaarg. 2293 auto &TD = DAG.getDataLayout(); 2294 unsigned ArgSizeInBytes = 2295 TD.getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())); 2296 SDValue Tmp3 = 2297 DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList, 2298 DAG.getConstant(alignTo(ArgSizeInBytes, ArgSlotSizeInBytes), 2299 DL, VAList.getValueType())); 2300 // Store the incremented VAList to the legalized pointer 2301 Chain = DAG.getStore(VAListLoad.getValue(1), DL, Tmp3, VAListPtr, 2302 MachinePointerInfo(SV)); 2303 2304 // In big-endian mode we must adjust the pointer when the load size is smaller 2305 // than the argument slot size. We must also reduce the known alignment to 2306 // match. For example in the N64 ABI, we must add 4 bytes to the offset to get 2307 // the correct half of the slot, and reduce the alignment from 8 (slot 2308 // alignment) down to 4 (type alignment). 2309 if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) { 2310 unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes; 2311 VAList = DAG.getNode(ISD::ADD, DL, VAListPtr.getValueType(), VAList, 2312 DAG.getIntPtrConstant(Adjustment, DL)); 2313 } 2314 // Load the actual argument out of the pointer VAList 2315 return DAG.getLoad(VT, DL, Chain, VAList, MachinePointerInfo()); 2316 } 2317 2318 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, 2319 bool HasExtractInsert) { 2320 EVT TyX = Op.getOperand(0).getValueType(); 2321 EVT TyY = Op.getOperand(1).getValueType(); 2322 SDLoc DL(Op); 2323 SDValue Const1 = DAG.getConstant(1, DL, MVT::i32); 2324 SDValue Const31 = DAG.getConstant(31, DL, MVT::i32); 2325 SDValue Res; 2326 2327 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 2328 // to i32. 2329 SDValue X = (TyX == MVT::f32) ? 2330 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) : 2331 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 2332 Const1); 2333 SDValue Y = (TyY == MVT::f32) ? 2334 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) : 2335 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1), 2336 Const1); 2337 2338 if (HasExtractInsert) { 2339 // ext E, Y, 31, 1 ; extract bit31 of Y 2340 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X 2341 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1); 2342 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X); 2343 } else { 2344 // sll SllX, X, 1 2345 // srl SrlX, SllX, 1 2346 // srl SrlY, Y, 31 2347 // sll SllY, SrlX, 31 2348 // or Or, SrlX, SllY 2349 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 2350 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 2351 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31); 2352 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31); 2353 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY); 2354 } 2355 2356 if (TyX == MVT::f32) 2357 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res); 2358 2359 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2360 Op.getOperand(0), 2361 DAG.getConstant(0, DL, MVT::i32)); 2362 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 2363 } 2364 2365 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, 2366 bool HasExtractInsert) { 2367 unsigned WidthX = Op.getOperand(0).getValueSizeInBits(); 2368 unsigned WidthY = Op.getOperand(1).getValueSizeInBits(); 2369 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY); 2370 SDLoc DL(Op); 2371 SDValue Const1 = DAG.getConstant(1, DL, MVT::i32); 2372 2373 // Bitcast to integer nodes. 2374 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0)); 2375 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1)); 2376 2377 if (HasExtractInsert) { 2378 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y 2379 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X 2380 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y, 2381 DAG.getConstant(WidthY - 1, DL, MVT::i32), Const1); 2382 2383 if (WidthX > WidthY) 2384 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E); 2385 else if (WidthY > WidthX) 2386 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E); 2387 2388 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E, 2389 DAG.getConstant(WidthX - 1, DL, MVT::i32), Const1, 2390 X); 2391 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I); 2392 } 2393 2394 // (d)sll SllX, X, 1 2395 // (d)srl SrlX, SllX, 1 2396 // (d)srl SrlY, Y, width(Y)-1 2397 // (d)sll SllY, SrlX, width(Y)-1 2398 // or Or, SrlX, SllY 2399 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1); 2400 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1); 2401 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y, 2402 DAG.getConstant(WidthY - 1, DL, MVT::i32)); 2403 2404 if (WidthX > WidthY) 2405 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY); 2406 else if (WidthY > WidthX) 2407 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY); 2408 2409 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY, 2410 DAG.getConstant(WidthX - 1, DL, MVT::i32)); 2411 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY); 2412 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or); 2413 } 2414 2415 SDValue 2416 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 2417 if (Subtarget.isGP64bit()) 2418 return lowerFCOPYSIGN64(Op, DAG, Subtarget.hasExtractInsert()); 2419 2420 return lowerFCOPYSIGN32(Op, DAG, Subtarget.hasExtractInsert()); 2421 } 2422 2423 static SDValue lowerFABS32(SDValue Op, SelectionDAG &DAG, 2424 bool HasExtractInsert) { 2425 SDLoc DL(Op); 2426 SDValue Res, Const1 = DAG.getConstant(1, DL, MVT::i32); 2427 2428 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 2429 // to i32. 2430 SDValue X = (Op.getValueType() == MVT::f32) 2431 ? DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) 2432 : DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2433 Op.getOperand(0), Const1); 2434 2435 // Clear MSB. 2436 if (HasExtractInsert) 2437 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, 2438 DAG.getRegister(Mips::ZERO, MVT::i32), 2439 DAG.getConstant(31, DL, MVT::i32), Const1, X); 2440 else { 2441 // TODO: Provide DAG patterns which transform (and x, cst) 2442 // back to a (shl (srl x (clz cst)) (clz cst)) sequence. 2443 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 2444 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 2445 } 2446 2447 if (Op.getValueType() == MVT::f32) 2448 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res); 2449 2450 // FIXME: For mips32r2, the sequence of (BuildPairF64 (ins (ExtractElementF64 2451 // Op 1), $zero, 31 1) (ExtractElementF64 Op 0)) and the Op has one use, we 2452 // should be able to drop the usage of mfc1/mtc1 and rewrite the register in 2453 // place. 2454 SDValue LowX = 2455 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 2456 DAG.getConstant(0, DL, MVT::i32)); 2457 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 2458 } 2459 2460 static SDValue lowerFABS64(SDValue Op, SelectionDAG &DAG, 2461 bool HasExtractInsert) { 2462 SDLoc DL(Op); 2463 SDValue Res, Const1 = DAG.getConstant(1, DL, MVT::i32); 2464 2465 // Bitcast to integer node. 2466 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0)); 2467 2468 // Clear MSB. 2469 if (HasExtractInsert) 2470 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64, 2471 DAG.getRegister(Mips::ZERO_64, MVT::i64), 2472 DAG.getConstant(63, DL, MVT::i32), Const1, X); 2473 else { 2474 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1); 2475 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1); 2476 } 2477 2478 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res); 2479 } 2480 2481 SDValue MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const { 2482 if ((ABI.IsN32() || ABI.IsN64()) && (Op.getValueType() == MVT::f64)) 2483 return lowerFABS64(Op, DAG, Subtarget.hasExtractInsert()); 2484 2485 return lowerFABS32(Op, DAG, Subtarget.hasExtractInsert()); 2486 } 2487 2488 SDValue MipsTargetLowering:: 2489 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 2490 // check the depth 2491 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() != 0) { 2492 DAG.getContext()->emitError( 2493 "return address can be determined only for current frame"); 2494 return SDValue(); 2495 } 2496 2497 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 2498 MFI.setFrameAddressIsTaken(true); 2499 EVT VT = Op.getValueType(); 2500 SDLoc DL(Op); 2501 SDValue FrameAddr = DAG.getCopyFromReg( 2502 DAG.getEntryNode(), DL, ABI.IsN64() ? Mips::FP_64 : Mips::FP, VT); 2503 return FrameAddr; 2504 } 2505 2506 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op, 2507 SelectionDAG &DAG) const { 2508 if (verifyReturnAddressArgumentIsConstant(Op, DAG)) 2509 return SDValue(); 2510 2511 // check the depth 2512 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() != 0) { 2513 DAG.getContext()->emitError( 2514 "return address can be determined only for current frame"); 2515 return SDValue(); 2516 } 2517 2518 MachineFunction &MF = DAG.getMachineFunction(); 2519 MachineFrameInfo &MFI = MF.getFrameInfo(); 2520 MVT VT = Op.getSimpleValueType(); 2521 unsigned RA = ABI.IsN64() ? Mips::RA_64 : Mips::RA; 2522 MFI.setReturnAddressIsTaken(true); 2523 2524 // Return RA, which contains the return address. Mark it an implicit live-in. 2525 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT)); 2526 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT); 2527 } 2528 2529 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is 2530 // generated from __builtin_eh_return (offset, handler) 2531 // The effect of this is to adjust the stack pointer by "offset" 2532 // and then branch to "handler". 2533 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG) 2534 const { 2535 MachineFunction &MF = DAG.getMachineFunction(); 2536 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2537 2538 MipsFI->setCallsEhReturn(); 2539 SDValue Chain = Op.getOperand(0); 2540 SDValue Offset = Op.getOperand(1); 2541 SDValue Handler = Op.getOperand(2); 2542 SDLoc DL(Op); 2543 EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32; 2544 2545 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and 2546 // EH_RETURN nodes, so that instructions are emitted back-to-back. 2547 unsigned OffsetReg = ABI.IsN64() ? Mips::V1_64 : Mips::V1; 2548 unsigned AddrReg = ABI.IsN64() ? Mips::V0_64 : Mips::V0; 2549 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue()); 2550 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1)); 2551 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain, 2552 DAG.getRegister(OffsetReg, Ty), 2553 DAG.getRegister(AddrReg, getPointerTy(MF.getDataLayout())), 2554 Chain.getValue(1)); 2555 } 2556 2557 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op, 2558 SelectionDAG &DAG) const { 2559 // FIXME: Need pseudo-fence for 'singlethread' fences 2560 // FIXME: Set SType for weaker fences where supported/appropriate. 2561 unsigned SType = 0; 2562 SDLoc DL(Op); 2563 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0), 2564 DAG.getConstant(SType, DL, MVT::i32)); 2565 } 2566 2567 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op, 2568 SelectionDAG &DAG) const { 2569 SDLoc DL(Op); 2570 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; 2571 2572 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 2573 SDValue Shamt = Op.getOperand(2); 2574 // if shamt < (VT.bits): 2575 // lo = (shl lo, shamt) 2576 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt)) 2577 // else: 2578 // lo = 0 2579 // hi = (shl lo, shamt[4:0]) 2580 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 2581 DAG.getConstant(-1, DL, MVT::i32)); 2582 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, 2583 DAG.getConstant(1, DL, VT)); 2584 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, Not); 2585 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt); 2586 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo); 2587 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt); 2588 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 2589 DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32)); 2590 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, 2591 DAG.getConstant(0, DL, VT), ShiftLeftLo); 2592 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftLeftLo, Or); 2593 2594 SDValue Ops[2] = {Lo, Hi}; 2595 return DAG.getMergeValues(Ops, DL); 2596 } 2597 2598 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, 2599 bool IsSRA) const { 2600 SDLoc DL(Op); 2601 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 2602 SDValue Shamt = Op.getOperand(2); 2603 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; 2604 2605 // if shamt < (VT.bits): 2606 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt)) 2607 // if isSRA: 2608 // hi = (sra hi, shamt) 2609 // else: 2610 // hi = (srl hi, shamt) 2611 // else: 2612 // if isSRA: 2613 // lo = (sra hi, shamt[4:0]) 2614 // hi = (sra hi, 31) 2615 // else: 2616 // lo = (srl hi, shamt[4:0]) 2617 // hi = 0 2618 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 2619 DAG.getConstant(-1, DL, MVT::i32)); 2620 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, VT, Hi, 2621 DAG.getConstant(1, DL, VT)); 2622 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, ShiftLeft1Hi, Not); 2623 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt); 2624 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo); 2625 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, 2626 DL, VT, Hi, Shamt); 2627 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 2628 DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32)); 2629 SDValue Ext = DAG.getNode(ISD::SRA, DL, VT, Hi, 2630 DAG.getConstant(VT.getSizeInBits() - 1, DL, VT)); 2631 2632 if (!(Subtarget.hasMips4() || Subtarget.hasMips32())) { 2633 SDVTList VTList = DAG.getVTList(VT, VT); 2634 return DAG.getNode(Subtarget.isGP64bit() ? Mips::PseudoD_SELECT_I64 2635 : Mips::PseudoD_SELECT_I, 2636 DL, VTList, Cond, ShiftRightHi, 2637 IsSRA ? Ext : DAG.getConstant(0, DL, VT), Or, 2638 ShiftRightHi); 2639 } 2640 2641 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftRightHi, Or); 2642 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, 2643 IsSRA ? Ext : DAG.getConstant(0, DL, VT), ShiftRightHi); 2644 2645 SDValue Ops[2] = {Lo, Hi}; 2646 return DAG.getMergeValues(Ops, DL); 2647 } 2648 2649 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD, 2650 SDValue Chain, SDValue Src, unsigned Offset) { 2651 SDValue Ptr = LD->getBasePtr(); 2652 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT(); 2653 EVT BasePtrVT = Ptr.getValueType(); 2654 SDLoc DL(LD); 2655 SDVTList VTList = DAG.getVTList(VT, MVT::Other); 2656 2657 if (Offset) 2658 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2659 DAG.getConstant(Offset, DL, BasePtrVT)); 2660 2661 SDValue Ops[] = { Chain, Ptr, Src }; 2662 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT, 2663 LD->getMemOperand()); 2664 } 2665 2666 // Expand an unaligned 32 or 64-bit integer load node. 2667 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const { 2668 LoadSDNode *LD = cast<LoadSDNode>(Op); 2669 EVT MemVT = LD->getMemoryVT(); 2670 2671 if (Subtarget.systemSupportsUnalignedAccess()) 2672 return Op; 2673 2674 // Return if load is aligned or if MemVT is neither i32 nor i64. 2675 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) || 2676 ((MemVT != MVT::i32) && (MemVT != MVT::i64))) 2677 return SDValue(); 2678 2679 bool IsLittle = Subtarget.isLittle(); 2680 EVT VT = Op.getValueType(); 2681 ISD::LoadExtType ExtType = LD->getExtensionType(); 2682 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT); 2683 2684 assert((VT == MVT::i32) || (VT == MVT::i64)); 2685 2686 // Expand 2687 // (set dst, (i64 (load baseptr))) 2688 // to 2689 // (set tmp, (ldl (add baseptr, 7), undef)) 2690 // (set dst, (ldr baseptr, tmp)) 2691 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) { 2692 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef, 2693 IsLittle ? 7 : 0); 2694 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL, 2695 IsLittle ? 0 : 7); 2696 } 2697 2698 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef, 2699 IsLittle ? 3 : 0); 2700 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL, 2701 IsLittle ? 0 : 3); 2702 2703 // Expand 2704 // (set dst, (i32 (load baseptr))) or 2705 // (set dst, (i64 (sextload baseptr))) or 2706 // (set dst, (i64 (extload baseptr))) 2707 // to 2708 // (set tmp, (lwl (add baseptr, 3), undef)) 2709 // (set dst, (lwr baseptr, tmp)) 2710 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) || 2711 (ExtType == ISD::EXTLOAD)) 2712 return LWR; 2713 2714 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD)); 2715 2716 // Expand 2717 // (set dst, (i64 (zextload baseptr))) 2718 // to 2719 // (set tmp0, (lwl (add baseptr, 3), undef)) 2720 // (set tmp1, (lwr baseptr, tmp0)) 2721 // (set tmp2, (shl tmp1, 32)) 2722 // (set dst, (srl tmp2, 32)) 2723 SDLoc DL(LD); 2724 SDValue Const32 = DAG.getConstant(32, DL, MVT::i32); 2725 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32); 2726 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32); 2727 SDValue Ops[] = { SRL, LWR.getValue(1) }; 2728 return DAG.getMergeValues(Ops, DL); 2729 } 2730 2731 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD, 2732 SDValue Chain, unsigned Offset) { 2733 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue(); 2734 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType(); 2735 SDLoc DL(SD); 2736 SDVTList VTList = DAG.getVTList(MVT::Other); 2737 2738 if (Offset) 2739 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2740 DAG.getConstant(Offset, DL, BasePtrVT)); 2741 2742 SDValue Ops[] = { Chain, Value, Ptr }; 2743 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT, 2744 SD->getMemOperand()); 2745 } 2746 2747 // Expand an unaligned 32 or 64-bit integer store node. 2748 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG, 2749 bool IsLittle) { 2750 SDValue Value = SD->getValue(), Chain = SD->getChain(); 2751 EVT VT = Value.getValueType(); 2752 2753 // Expand 2754 // (store val, baseptr) or 2755 // (truncstore val, baseptr) 2756 // to 2757 // (swl val, (add baseptr, 3)) 2758 // (swr val, baseptr) 2759 if ((VT == MVT::i32) || SD->isTruncatingStore()) { 2760 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain, 2761 IsLittle ? 3 : 0); 2762 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3); 2763 } 2764 2765 assert(VT == MVT::i64); 2766 2767 // Expand 2768 // (store val, baseptr) 2769 // to 2770 // (sdl val, (add baseptr, 7)) 2771 // (sdr val, baseptr) 2772 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0); 2773 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7); 2774 } 2775 2776 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr). 2777 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG, 2778 bool SingleFloat) { 2779 SDValue Val = SD->getValue(); 2780 2781 if (Val.getOpcode() != ISD::FP_TO_SINT || 2782 (Val.getValueSizeInBits() > 32 && SingleFloat)) 2783 return SDValue(); 2784 2785 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits()); 2786 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy, 2787 Val.getOperand(0)); 2788 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(), 2789 SD->getPointerInfo(), SD->getAlignment(), 2790 SD->getMemOperand()->getFlags()); 2791 } 2792 2793 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const { 2794 StoreSDNode *SD = cast<StoreSDNode>(Op); 2795 EVT MemVT = SD->getMemoryVT(); 2796 2797 // Lower unaligned integer stores. 2798 if (!Subtarget.systemSupportsUnalignedAccess() && 2799 (SD->getAlignment() < MemVT.getSizeInBits() / 8) && 2800 ((MemVT == MVT::i32) || (MemVT == MVT::i64))) 2801 return lowerUnalignedIntStore(SD, DAG, Subtarget.isLittle()); 2802 2803 return lowerFP_TO_SINT_STORE(SD, DAG, Subtarget.isSingleFloat()); 2804 } 2805 2806 SDValue MipsTargetLowering::lowerEH_DWARF_CFA(SDValue Op, 2807 SelectionDAG &DAG) const { 2808 2809 // Return a fixed StackObject with offset 0 which points to the old stack 2810 // pointer. 2811 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 2812 EVT ValTy = Op->getValueType(0); 2813 int FI = MFI.CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false); 2814 return DAG.getFrameIndex(FI, ValTy); 2815 } 2816 2817 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op, 2818 SelectionDAG &DAG) const { 2819 if (Op.getValueSizeInBits() > 32 && Subtarget.isSingleFloat()) 2820 return SDValue(); 2821 2822 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits()); 2823 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy, 2824 Op.getOperand(0)); 2825 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc); 2826 } 2827 2828 //===----------------------------------------------------------------------===// 2829 // Calling Convention Implementation 2830 //===----------------------------------------------------------------------===// 2831 2832 //===----------------------------------------------------------------------===// 2833 // TODO: Implement a generic logic using tblgen that can support this. 2834 // Mips O32 ABI rules: 2835 // --- 2836 // i32 - Passed in A0, A1, A2, A3 and stack 2837 // f32 - Only passed in f32 registers if no int reg has been used yet to hold 2838 // an argument. Otherwise, passed in A1, A2, A3 and stack. 2839 // f64 - Only passed in two aliased f32 registers if no int reg has been used 2840 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is 2841 // not used, it must be shadowed. If only A3 is available, shadow it and 2842 // go to stack. 2843 // vXiX - Received as scalarized i32s, passed in A0 - A3 and the stack. 2844 // vXf32 - Passed in either a pair of registers {A0, A1}, {A2, A3} or {A0 - A3} 2845 // with the remainder spilled to the stack. 2846 // vXf64 - Passed in either {A0, A1, A2, A3} or {A2, A3} and in both cases 2847 // spilling the remainder to the stack. 2848 // 2849 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack. 2850 //===----------------------------------------------------------------------===// 2851 2852 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, 2853 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, 2854 CCState &State, ArrayRef<MCPhysReg> F64Regs) { 2855 const MipsSubtarget &Subtarget = static_cast<const MipsSubtarget &>( 2856 State.getMachineFunction().getSubtarget()); 2857 2858 static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 }; 2859 2860 const MipsCCState * MipsState = static_cast<MipsCCState *>(&State); 2861 2862 static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 }; 2863 2864 static const MCPhysReg FloatVectorIntRegs[] = { Mips::A0, Mips::A2 }; 2865 2866 // Do not process byval args here. 2867 if (ArgFlags.isByVal()) 2868 return true; 2869 2870 // Promote i8 and i16 2871 if (ArgFlags.isInReg() && !Subtarget.isLittle()) { 2872 if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) { 2873 LocVT = MVT::i32; 2874 if (ArgFlags.isSExt()) 2875 LocInfo = CCValAssign::SExtUpper; 2876 else if (ArgFlags.isZExt()) 2877 LocInfo = CCValAssign::ZExtUpper; 2878 else 2879 LocInfo = CCValAssign::AExtUpper; 2880 } 2881 } 2882 2883 // Promote i8 and i16 2884 if (LocVT == MVT::i8 || LocVT == MVT::i16) { 2885 LocVT = MVT::i32; 2886 if (ArgFlags.isSExt()) 2887 LocInfo = CCValAssign::SExt; 2888 else if (ArgFlags.isZExt()) 2889 LocInfo = CCValAssign::ZExt; 2890 else 2891 LocInfo = CCValAssign::AExt; 2892 } 2893 2894 unsigned Reg; 2895 2896 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following 2897 // is true: function is vararg, argument is 3rd or higher, there is previous 2898 // argument which is not f32 or f64. 2899 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 || 2900 State.getFirstUnallocated(F32Regs) != ValNo; 2901 unsigned OrigAlign = ArgFlags.getOrigAlign(); 2902 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8); 2903 bool isVectorFloat = MipsState->WasOriginalArgVectorFloat(ValNo); 2904 2905 // The MIPS vector ABI for floats passes them in a pair of registers 2906 if (ValVT == MVT::i32 && isVectorFloat) { 2907 // This is the start of an vector that was scalarized into an unknown number 2908 // of components. It doesn't matter how many there are. Allocate one of the 2909 // notional 8 byte aligned registers which map onto the argument stack, and 2910 // shadow the register lost to alignment requirements. 2911 if (ArgFlags.isSplit()) { 2912 Reg = State.AllocateReg(FloatVectorIntRegs); 2913 if (Reg == Mips::A2) 2914 State.AllocateReg(Mips::A1); 2915 else if (Reg == 0) 2916 State.AllocateReg(Mips::A3); 2917 } else { 2918 // If we're an intermediate component of the split, we can just attempt to 2919 // allocate a register directly. 2920 Reg = State.AllocateReg(IntRegs); 2921 } 2922 } else if (ValVT == MVT::i32 || 2923 (ValVT == MVT::f32 && AllocateFloatsInIntReg)) { 2924 Reg = State.AllocateReg(IntRegs); 2925 // If this is the first part of an i64 arg, 2926 // the allocated register must be either A0 or A2. 2927 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3)) 2928 Reg = State.AllocateReg(IntRegs); 2929 LocVT = MVT::i32; 2930 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) { 2931 // Allocate int register and shadow next int register. If first 2932 // available register is Mips::A1 or Mips::A3, shadow it too. 2933 Reg = State.AllocateReg(IntRegs); 2934 if (Reg == Mips::A1 || Reg == Mips::A3) 2935 Reg = State.AllocateReg(IntRegs); 2936 State.AllocateReg(IntRegs); 2937 LocVT = MVT::i32; 2938 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) { 2939 // we are guaranteed to find an available float register 2940 if (ValVT == MVT::f32) { 2941 Reg = State.AllocateReg(F32Regs); 2942 // Shadow int register 2943 State.AllocateReg(IntRegs); 2944 } else { 2945 Reg = State.AllocateReg(F64Regs); 2946 // Shadow int registers 2947 unsigned Reg2 = State.AllocateReg(IntRegs); 2948 if (Reg2 == Mips::A1 || Reg2 == Mips::A3) 2949 State.AllocateReg(IntRegs); 2950 State.AllocateReg(IntRegs); 2951 } 2952 } else 2953 llvm_unreachable("Cannot handle this ValVT."); 2954 2955 if (!Reg) { 2956 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(), OrigAlign); 2957 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); 2958 } else 2959 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 2960 2961 return false; 2962 } 2963 2964 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT, 2965 MVT LocVT, CCValAssign::LocInfo LocInfo, 2966 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2967 static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 }; 2968 2969 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); 2970 } 2971 2972 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT, 2973 MVT LocVT, CCValAssign::LocInfo LocInfo, 2974 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2975 static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 }; 2976 2977 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); 2978 } 2979 2980 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, 2981 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, 2982 CCState &State) LLVM_ATTRIBUTE_UNUSED; 2983 2984 #include "MipsGenCallingConv.inc" 2985 2986 CCAssignFn *MipsTargetLowering::CCAssignFnForCall() const{ 2987 return CC_Mips_FixedArg; 2988 } 2989 2990 CCAssignFn *MipsTargetLowering::CCAssignFnForReturn() const{ 2991 return RetCC_Mips; 2992 } 2993 //===----------------------------------------------------------------------===// 2994 // Call Calling Convention Implementation 2995 //===----------------------------------------------------------------------===// 2996 2997 // Return next O32 integer argument register. 2998 static unsigned getNextIntArgReg(unsigned Reg) { 2999 assert((Reg == Mips::A0) || (Reg == Mips::A2)); 3000 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3; 3001 } 3002 3003 SDValue MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset, 3004 SDValue Chain, SDValue Arg, 3005 const SDLoc &DL, bool IsTailCall, 3006 SelectionDAG &DAG) const { 3007 if (!IsTailCall) { 3008 SDValue PtrOff = 3009 DAG.getNode(ISD::ADD, DL, getPointerTy(DAG.getDataLayout()), StackPtr, 3010 DAG.getIntPtrConstant(Offset, DL)); 3011 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo()); 3012 } 3013 3014 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 3015 int FI = MFI.CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false); 3016 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); 3017 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(), 3018 /* Alignment = */ 0, MachineMemOperand::MOVolatile); 3019 } 3020 3021 void MipsTargetLowering:: 3022 getOpndList(SmallVectorImpl<SDValue> &Ops, 3023 std::deque<std::pair<unsigned, SDValue>> &RegsToPass, 3024 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage, 3025 bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee, 3026 SDValue Chain) const { 3027 // Insert node "GP copy globalreg" before call to function. 3028 // 3029 // R_MIPS_CALL* operators (emitted when non-internal functions are called 3030 // in PIC mode) allow symbols to be resolved via lazy binding. 3031 // The lazy binding stub requires GP to point to the GOT. 3032 // Note that we don't need GP to point to the GOT for indirect calls 3033 // (when R_MIPS_CALL* is not used for the call) because Mips linker generates 3034 // lazy binding stub for a function only when R_MIPS_CALL* are the only relocs 3035 // used for the function (that is, Mips linker doesn't generate lazy binding 3036 // stub for a function whose address is taken in the program). 3037 if (IsPICCall && !InternalLinkage && IsCallReloc) { 3038 unsigned GPReg = ABI.IsN64() ? Mips::GP_64 : Mips::GP; 3039 EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32; 3040 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty))); 3041 } 3042 3043 // Build a sequence of copy-to-reg nodes chained together with token 3044 // chain and flag operands which copy the outgoing args into registers. 3045 // The InFlag in necessary since all emitted instructions must be 3046 // stuck together. 3047 SDValue InFlag; 3048 3049 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 3050 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first, 3051 RegsToPass[i].second, InFlag); 3052 InFlag = Chain.getValue(1); 3053 } 3054 3055 // Add argument registers to the end of the list so that they are 3056 // known live into the call. 3057 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 3058 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first, 3059 RegsToPass[i].second.getValueType())); 3060 3061 // Add a register mask operand representing the call-preserved registers. 3062 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); 3063 const uint32_t *Mask = 3064 TRI->getCallPreservedMask(CLI.DAG.getMachineFunction(), CLI.CallConv); 3065 assert(Mask && "Missing call preserved mask for calling convention"); 3066 if (Subtarget.inMips16HardFloat()) { 3067 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) { 3068 StringRef Sym = G->getGlobal()->getName(); 3069 Function *F = G->getGlobal()->getParent()->getFunction(Sym); 3070 if (F && F->hasFnAttribute("__Mips16RetHelper")) { 3071 Mask = MipsRegisterInfo::getMips16RetHelperMask(); 3072 } 3073 } 3074 } 3075 Ops.push_back(CLI.DAG.getRegisterMask(Mask)); 3076 3077 if (InFlag.getNode()) 3078 Ops.push_back(InFlag); 3079 } 3080 3081 void MipsTargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI, 3082 SDNode *Node) const { 3083 switch (MI.getOpcode()) { 3084 default: 3085 return; 3086 case Mips::JALR: 3087 case Mips::JALRPseudo: 3088 case Mips::JALR64: 3089 case Mips::JALR64Pseudo: 3090 case Mips::JALR16_MM: 3091 case Mips::JALRC16_MMR6: 3092 case Mips::TAILCALLREG: 3093 case Mips::TAILCALLREG64: 3094 case Mips::TAILCALLR6REG: 3095 case Mips::TAILCALL64R6REG: 3096 case Mips::TAILCALLREG_MM: 3097 case Mips::TAILCALLREG_MMR6: { 3098 if (!EmitJalrReloc || 3099 Subtarget.inMips16Mode() || 3100 !isPositionIndependent() || 3101 Node->getNumOperands() < 1 || 3102 Node->getOperand(0).getNumOperands() < 2) { 3103 return; 3104 } 3105 // We are after the callee address, set by LowerCall(). 3106 // If added to MI, asm printer will emit .reloc R_MIPS_JALR for the 3107 // symbol. 3108 const SDValue TargetAddr = Node->getOperand(0).getOperand(1); 3109 StringRef Sym; 3110 if (const GlobalAddressSDNode *G = 3111 dyn_cast_or_null<const GlobalAddressSDNode>(TargetAddr)) { 3112 Sym = G->getGlobal()->getName(); 3113 } 3114 else if (const ExternalSymbolSDNode *ES = 3115 dyn_cast_or_null<const ExternalSymbolSDNode>(TargetAddr)) { 3116 Sym = ES->getSymbol(); 3117 } 3118 3119 if (Sym.empty()) 3120 return; 3121 3122 MachineFunction *MF = MI.getParent()->getParent(); 3123 MCSymbol *S = MF->getContext().getOrCreateSymbol(Sym); 3124 MI.addOperand(MachineOperand::CreateMCSymbol(S, MipsII::MO_JALR)); 3125 } 3126 } 3127 } 3128 3129 /// LowerCall - functions arguments are copied from virtual regs to 3130 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. 3131 SDValue 3132 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, 3133 SmallVectorImpl<SDValue> &InVals) const { 3134 SelectionDAG &DAG = CLI.DAG; 3135 SDLoc DL = CLI.DL; 3136 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; 3137 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; 3138 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; 3139 SDValue Chain = CLI.Chain; 3140 SDValue Callee = CLI.Callee; 3141 bool &IsTailCall = CLI.IsTailCall; 3142 CallingConv::ID CallConv = CLI.CallConv; 3143 bool IsVarArg = CLI.IsVarArg; 3144 3145 MachineFunction &MF = DAG.getMachineFunction(); 3146 MachineFrameInfo &MFI = MF.getFrameInfo(); 3147 const TargetFrameLowering *TFL = Subtarget.getFrameLowering(); 3148 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 3149 bool IsPIC = isPositionIndependent(); 3150 3151 // Analyze operands of the call, assigning locations to each operand. 3152 SmallVector<CCValAssign, 16> ArgLocs; 3153 MipsCCState CCInfo( 3154 CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, *DAG.getContext(), 3155 MipsCCState::getSpecialCallingConvForCallee(Callee.getNode(), Subtarget)); 3156 3157 const ExternalSymbolSDNode *ES = 3158 dyn_cast_or_null<const ExternalSymbolSDNode>(Callee.getNode()); 3159 3160 // There is one case where CALLSEQ_START..CALLSEQ_END can be nested, which 3161 // is during the lowering of a call with a byval argument which produces 3162 // a call to memcpy. For the O32 case, this causes the caller to allocate 3163 // stack space for the reserved argument area for the callee, then recursively 3164 // again for the memcpy call. In the NEWABI case, this doesn't occur as those 3165 // ABIs mandate that the callee allocates the reserved argument area. We do 3166 // still produce nested CALLSEQ_START..CALLSEQ_END with zero space though. 3167 // 3168 // If the callee has a byval argument and memcpy is used, we are mandated 3169 // to already have produced a reserved argument area for the callee for O32. 3170 // Therefore, the reserved argument area can be reused for both calls. 3171 // 3172 // Other cases of calling memcpy cannot have a chain with a CALLSEQ_START 3173 // present, as we have yet to hook that node onto the chain. 3174 // 3175 // Hence, the CALLSEQ_START and CALLSEQ_END nodes can be eliminated in this 3176 // case. GCC does a similar trick, in that wherever possible, it calculates 3177 // the maximum out going argument area (including the reserved area), and 3178 // preallocates the stack space on entrance to the caller. 3179 // 3180 // FIXME: We should do the same for efficiency and space. 3181 3182 // Note: The check on the calling convention below must match 3183 // MipsABIInfo::GetCalleeAllocdArgSizeInBytes(). 3184 bool MemcpyInByVal = ES && 3185 StringRef(ES->getSymbol()) == StringRef("memcpy") && 3186 CallConv != CallingConv::Fast && 3187 Chain.getOpcode() == ISD::CALLSEQ_START; 3188 3189 // Allocate the reserved argument area. It seems strange to do this from the 3190 // caller side but removing it breaks the frame size calculation. 3191 unsigned ReservedArgArea = 3192 MemcpyInByVal ? 0 : ABI.GetCalleeAllocdArgSizeInBytes(CallConv); 3193 CCInfo.AllocateStack(ReservedArgArea, 1); 3194 3195 CCInfo.AnalyzeCallOperands(Outs, CC_Mips, CLI.getArgs(), 3196 ES ? ES->getSymbol() : nullptr); 3197 3198 // Get a count of how many bytes are to be pushed on the stack. 3199 unsigned NextStackOffset = CCInfo.getNextStackOffset(); 3200 3201 // Check if it's really possible to do a tail call. Restrict it to functions 3202 // that are part of this compilation unit. 3203 bool InternalLinkage = false; 3204 if (IsTailCall) { 3205 IsTailCall = isEligibleForTailCallOptimization( 3206 CCInfo, NextStackOffset, *MF.getInfo<MipsFunctionInfo>()); 3207 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 3208 InternalLinkage = G->getGlobal()->hasInternalLinkage(); 3209 IsTailCall &= (InternalLinkage || G->getGlobal()->hasLocalLinkage() || 3210 G->getGlobal()->hasPrivateLinkage() || 3211 G->getGlobal()->hasHiddenVisibility() || 3212 G->getGlobal()->hasProtectedVisibility()); 3213 } 3214 } 3215 if (!IsTailCall && CLI.CS && CLI.CS.isMustTailCall()) 3216 report_fatal_error("failed to perform tail call elimination on a call " 3217 "site marked musttail"); 3218 3219 if (IsTailCall) 3220 ++NumTailCalls; 3221 3222 // Chain is the output chain of the last Load/Store or CopyToReg node. 3223 // ByValChain is the output chain of the last Memcpy node created for copying 3224 // byval arguments to the stack. 3225 unsigned StackAlignment = TFL->getStackAlignment(); 3226 NextStackOffset = alignTo(NextStackOffset, StackAlignment); 3227 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, DL, true); 3228 3229 if (!(IsTailCall || MemcpyInByVal)) 3230 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffset, 0, DL); 3231 3232 SDValue StackPtr = 3233 DAG.getCopyFromReg(Chain, DL, ABI.IsN64() ? Mips::SP_64 : Mips::SP, 3234 getPointerTy(DAG.getDataLayout())); 3235 3236 std::deque<std::pair<unsigned, SDValue>> RegsToPass; 3237 SmallVector<SDValue, 8> MemOpChains; 3238 3239 CCInfo.rewindByValRegsInfo(); 3240 3241 // Walk the register/memloc assignments, inserting copies/loads. 3242 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 3243 SDValue Arg = OutVals[i]; 3244 CCValAssign &VA = ArgLocs[i]; 3245 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT(); 3246 ISD::ArgFlagsTy Flags = Outs[i].Flags; 3247 bool UseUpperBits = false; 3248 3249 // ByVal Arg. 3250 if (Flags.isByVal()) { 3251 unsigned FirstByValReg, LastByValReg; 3252 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); 3253 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg); 3254 3255 assert(Flags.getByValSize() && 3256 "ByVal args of size 0 should have been ignored by front-end."); 3257 assert(ByValIdx < CCInfo.getInRegsParamsCount()); 3258 assert(!IsTailCall && 3259 "Do not tail-call optimize if there is a byval argument."); 3260 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg, 3261 FirstByValReg, LastByValReg, Flags, Subtarget.isLittle(), 3262 VA); 3263 CCInfo.nextInRegsParam(); 3264 continue; 3265 } 3266 3267 // Promote the value if needed. 3268 switch (VA.getLocInfo()) { 3269 default: 3270 llvm_unreachable("Unknown loc info!"); 3271 case CCValAssign::Full: 3272 if (VA.isRegLoc()) { 3273 if ((ValVT == MVT::f32 && LocVT == MVT::i32) || 3274 (ValVT == MVT::f64 && LocVT == MVT::i64) || 3275 (ValVT == MVT::i64 && LocVT == MVT::f64)) 3276 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 3277 else if (ValVT == MVT::f64 && LocVT == MVT::i32) { 3278 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 3279 Arg, DAG.getConstant(0, DL, MVT::i32)); 3280 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 3281 Arg, DAG.getConstant(1, DL, MVT::i32)); 3282 if (!Subtarget.isLittle()) 3283 std::swap(Lo, Hi); 3284 Register LocRegLo = VA.getLocReg(); 3285 unsigned LocRegHigh = getNextIntArgReg(LocRegLo); 3286 RegsToPass.push_back(std::make_pair(LocRegLo, Lo)); 3287 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi)); 3288 continue; 3289 } 3290 } 3291 break; 3292 case CCValAssign::BCvt: 3293 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 3294 break; 3295 case CCValAssign::SExtUpper: 3296 UseUpperBits = true; 3297 LLVM_FALLTHROUGH; 3298 case CCValAssign::SExt: 3299 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg); 3300 break; 3301 case CCValAssign::ZExtUpper: 3302 UseUpperBits = true; 3303 LLVM_FALLTHROUGH; 3304 case CCValAssign::ZExt: 3305 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg); 3306 break; 3307 case CCValAssign::AExtUpper: 3308 UseUpperBits = true; 3309 LLVM_FALLTHROUGH; 3310 case CCValAssign::AExt: 3311 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg); 3312 break; 3313 } 3314 3315 if (UseUpperBits) { 3316 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits(); 3317 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3318 Arg = DAG.getNode( 3319 ISD::SHL, DL, VA.getLocVT(), Arg, 3320 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3321 } 3322 3323 // Arguments that can be passed on register must be kept at 3324 // RegsToPass vector 3325 if (VA.isRegLoc()) { 3326 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 3327 continue; 3328 } 3329 3330 // Register can't get to this point... 3331 assert(VA.isMemLoc()); 3332 3333 // emit ISD::STORE whichs stores the 3334 // parameter value to a stack Location 3335 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(), 3336 Chain, Arg, DL, IsTailCall, DAG)); 3337 } 3338 3339 // Transform all store nodes into one single node because all store 3340 // nodes are independent of each other. 3341 if (!MemOpChains.empty()) 3342 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains); 3343 3344 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 3345 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 3346 // node so that legalize doesn't hack it. 3347 3348 EVT Ty = Callee.getValueType(); 3349 bool GlobalOrExternal = false, IsCallReloc = false; 3350 3351 // The long-calls feature is ignored in case of PIC. 3352 // While we do not support -mshared / -mno-shared properly, 3353 // ignore long-calls in case of -mabicalls too. 3354 if (!Subtarget.isABICalls() && !IsPIC) { 3355 // If the function should be called using "long call", 3356 // get its address into a register to prevent using 3357 // of the `jal` instruction for the direct call. 3358 if (auto *N = dyn_cast<ExternalSymbolSDNode>(Callee)) { 3359 if (Subtarget.useLongCalls()) 3360 Callee = Subtarget.hasSym32() 3361 ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 3362 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 3363 } else if (auto *N = dyn_cast<GlobalAddressSDNode>(Callee)) { 3364 bool UseLongCalls = Subtarget.useLongCalls(); 3365 // If the function has long-call/far/near attribute 3366 // it overrides command line switch pased to the backend. 3367 if (auto *F = dyn_cast<Function>(N->getGlobal())) { 3368 if (F->hasFnAttribute("long-call")) 3369 UseLongCalls = true; 3370 else if (F->hasFnAttribute("short-call")) 3371 UseLongCalls = false; 3372 } 3373 if (UseLongCalls) 3374 Callee = Subtarget.hasSym32() 3375 ? getAddrNonPIC(N, SDLoc(N), Ty, DAG) 3376 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG); 3377 } 3378 } 3379 3380 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 3381 if (IsPIC) { 3382 const GlobalValue *Val = G->getGlobal(); 3383 InternalLinkage = Val->hasInternalLinkage(); 3384 3385 if (InternalLinkage) 3386 Callee = getAddrLocal(G, DL, Ty, DAG, ABI.IsN32() || ABI.IsN64()); 3387 else if (Subtarget.useXGOT()) { 3388 Callee = getAddrGlobalLargeGOT(G, DL, Ty, DAG, MipsII::MO_CALL_HI16, 3389 MipsII::MO_CALL_LO16, Chain, 3390 FuncInfo->callPtrInfo(Val)); 3391 IsCallReloc = true; 3392 } else { 3393 Callee = getAddrGlobal(G, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain, 3394 FuncInfo->callPtrInfo(Val)); 3395 IsCallReloc = true; 3396 } 3397 } else 3398 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, 3399 getPointerTy(DAG.getDataLayout()), 0, 3400 MipsII::MO_NO_FLAG); 3401 GlobalOrExternal = true; 3402 } 3403 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 3404 const char *Sym = S->getSymbol(); 3405 3406 if (!IsPIC) // static 3407 Callee = DAG.getTargetExternalSymbol( 3408 Sym, getPointerTy(DAG.getDataLayout()), MipsII::MO_NO_FLAG); 3409 else if (Subtarget.useXGOT()) { 3410 Callee = getAddrGlobalLargeGOT(S, DL, Ty, DAG, MipsII::MO_CALL_HI16, 3411 MipsII::MO_CALL_LO16, Chain, 3412 FuncInfo->callPtrInfo(Sym)); 3413 IsCallReloc = true; 3414 } else { // PIC 3415 Callee = getAddrGlobal(S, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain, 3416 FuncInfo->callPtrInfo(Sym)); 3417 IsCallReloc = true; 3418 } 3419 3420 GlobalOrExternal = true; 3421 } 3422 3423 SmallVector<SDValue, 8> Ops(1, Chain); 3424 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 3425 3426 getOpndList(Ops, RegsToPass, IsPIC, GlobalOrExternal, InternalLinkage, 3427 IsCallReloc, CLI, Callee, Chain); 3428 3429 if (IsTailCall) { 3430 MF.getFrameInfo().setHasTailCall(); 3431 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, Ops); 3432 } 3433 3434 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, Ops); 3435 SDValue InFlag = Chain.getValue(1); 3436 3437 // Create the CALLSEQ_END node in the case of where it is not a call to 3438 // memcpy. 3439 if (!(MemcpyInByVal)) { 3440 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal, 3441 DAG.getIntPtrConstant(0, DL, true), InFlag, DL); 3442 InFlag = Chain.getValue(1); 3443 } 3444 3445 // Handle result values, copying them out of physregs into vregs that we 3446 // return. 3447 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG, 3448 InVals, CLI); 3449 } 3450 3451 /// LowerCallResult - Lower the result values of a call into the 3452 /// appropriate copies out of appropriate physical registers. 3453 SDValue MipsTargetLowering::LowerCallResult( 3454 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg, 3455 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, 3456 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, 3457 TargetLowering::CallLoweringInfo &CLI) const { 3458 // Assign locations to each value returned by this call. 3459 SmallVector<CCValAssign, 16> RVLocs; 3460 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, 3461 *DAG.getContext()); 3462 3463 const ExternalSymbolSDNode *ES = 3464 dyn_cast_or_null<const ExternalSymbolSDNode>(CLI.Callee.getNode()); 3465 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips, CLI.RetTy, 3466 ES ? ES->getSymbol() : nullptr); 3467 3468 // Copy all of the result registers out of their specified physreg. 3469 for (unsigned i = 0; i != RVLocs.size(); ++i) { 3470 CCValAssign &VA = RVLocs[i]; 3471 assert(VA.isRegLoc() && "Can only return in registers!"); 3472 3473 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(), 3474 RVLocs[i].getLocVT(), InFlag); 3475 Chain = Val.getValue(1); 3476 InFlag = Val.getValue(2); 3477 3478 if (VA.isUpperBitsInLoc()) { 3479 unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits(); 3480 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3481 unsigned Shift = 3482 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; 3483 Val = DAG.getNode( 3484 Shift, DL, VA.getLocVT(), Val, 3485 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3486 } 3487 3488 switch (VA.getLocInfo()) { 3489 default: 3490 llvm_unreachable("Unknown loc info!"); 3491 case CCValAssign::Full: 3492 break; 3493 case CCValAssign::BCvt: 3494 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val); 3495 break; 3496 case CCValAssign::AExt: 3497 case CCValAssign::AExtUpper: 3498 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 3499 break; 3500 case CCValAssign::ZExt: 3501 case CCValAssign::ZExtUpper: 3502 Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val, 3503 DAG.getValueType(VA.getValVT())); 3504 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 3505 break; 3506 case CCValAssign::SExt: 3507 case CCValAssign::SExtUpper: 3508 Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val, 3509 DAG.getValueType(VA.getValVT())); 3510 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 3511 break; 3512 } 3513 3514 InVals.push_back(Val); 3515 } 3516 3517 return Chain; 3518 } 3519 3520 static SDValue UnpackFromArgumentSlot(SDValue Val, const CCValAssign &VA, 3521 EVT ArgVT, const SDLoc &DL, 3522 SelectionDAG &DAG) { 3523 MVT LocVT = VA.getLocVT(); 3524 EVT ValVT = VA.getValVT(); 3525 3526 // Shift into the upper bits if necessary. 3527 switch (VA.getLocInfo()) { 3528 default: 3529 break; 3530 case CCValAssign::AExtUpper: 3531 case CCValAssign::SExtUpper: 3532 case CCValAssign::ZExtUpper: { 3533 unsigned ValSizeInBits = ArgVT.getSizeInBits(); 3534 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3535 unsigned Opcode = 3536 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; 3537 Val = DAG.getNode( 3538 Opcode, DL, VA.getLocVT(), Val, 3539 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3540 break; 3541 } 3542 } 3543 3544 // If this is an value smaller than the argument slot size (32-bit for O32, 3545 // 64-bit for N32/N64), it has been promoted in some way to the argument slot 3546 // size. Extract the value and insert any appropriate assertions regarding 3547 // sign/zero extension. 3548 switch (VA.getLocInfo()) { 3549 default: 3550 llvm_unreachable("Unknown loc info!"); 3551 case CCValAssign::Full: 3552 break; 3553 case CCValAssign::AExtUpper: 3554 case CCValAssign::AExt: 3555 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 3556 break; 3557 case CCValAssign::SExtUpper: 3558 case CCValAssign::SExt: 3559 Val = DAG.getNode(ISD::AssertSext, DL, LocVT, Val, DAG.getValueType(ValVT)); 3560 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 3561 break; 3562 case CCValAssign::ZExtUpper: 3563 case CCValAssign::ZExt: 3564 Val = DAG.getNode(ISD::AssertZext, DL, LocVT, Val, DAG.getValueType(ValVT)); 3565 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 3566 break; 3567 case CCValAssign::BCvt: 3568 Val = DAG.getNode(ISD::BITCAST, DL, ValVT, Val); 3569 break; 3570 } 3571 3572 return Val; 3573 } 3574 3575 //===----------------------------------------------------------------------===// 3576 // Formal Arguments Calling Convention Implementation 3577 //===----------------------------------------------------------------------===// 3578 /// LowerFormalArguments - transform physical registers into virtual registers 3579 /// and generate load operations for arguments places on the stack. 3580 SDValue MipsTargetLowering::LowerFormalArguments( 3581 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, 3582 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, 3583 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { 3584 MachineFunction &MF = DAG.getMachineFunction(); 3585 MachineFrameInfo &MFI = MF.getFrameInfo(); 3586 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3587 3588 MipsFI->setVarArgsFrameIndex(0); 3589 3590 // Used with vargs to acumulate store chains. 3591 std::vector<SDValue> OutChains; 3592 3593 // Assign locations to all of the incoming arguments. 3594 SmallVector<CCValAssign, 16> ArgLocs; 3595 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, 3596 *DAG.getContext()); 3597 CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1); 3598 const Function &Func = DAG.getMachineFunction().getFunction(); 3599 Function::const_arg_iterator FuncArg = Func.arg_begin(); 3600 3601 if (Func.hasFnAttribute("interrupt") && !Func.arg_empty()) 3602 report_fatal_error( 3603 "Functions with the interrupt attribute cannot have arguments!"); 3604 3605 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips_FixedArg); 3606 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(), 3607 CCInfo.getInRegsParamsCount() > 0); 3608 3609 unsigned CurArgIdx = 0; 3610 CCInfo.rewindByValRegsInfo(); 3611 3612 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 3613 CCValAssign &VA = ArgLocs[i]; 3614 if (Ins[i].isOrigArg()) { 3615 std::advance(FuncArg, Ins[i].getOrigArgIndex() - CurArgIdx); 3616 CurArgIdx = Ins[i].getOrigArgIndex(); 3617 } 3618 EVT ValVT = VA.getValVT(); 3619 ISD::ArgFlagsTy Flags = Ins[i].Flags; 3620 bool IsRegLoc = VA.isRegLoc(); 3621 3622 if (Flags.isByVal()) { 3623 assert(Ins[i].isOrigArg() && "Byval arguments cannot be implicit"); 3624 unsigned FirstByValReg, LastByValReg; 3625 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); 3626 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg); 3627 3628 assert(Flags.getByValSize() && 3629 "ByVal args of size 0 should have been ignored by front-end."); 3630 assert(ByValIdx < CCInfo.getInRegsParamsCount()); 3631 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg, 3632 FirstByValReg, LastByValReg, VA, CCInfo); 3633 CCInfo.nextInRegsParam(); 3634 continue; 3635 } 3636 3637 // Arguments stored on registers 3638 if (IsRegLoc) { 3639 MVT RegVT = VA.getLocVT(); 3640 Register ArgReg = VA.getLocReg(); 3641 const TargetRegisterClass *RC = getRegClassFor(RegVT); 3642 3643 // Transform the arguments stored on 3644 // physical registers into virtual ones 3645 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC); 3646 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT); 3647 3648 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG); 3649 3650 // Handle floating point arguments passed in integer registers and 3651 // long double arguments passed in floating point registers. 3652 if ((RegVT == MVT::i32 && ValVT == MVT::f32) || 3653 (RegVT == MVT::i64 && ValVT == MVT::f64) || 3654 (RegVT == MVT::f64 && ValVT == MVT::i64)) 3655 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue); 3656 else if (ABI.IsO32() && RegVT == MVT::i32 && 3657 ValVT == MVT::f64) { 3658 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(), 3659 getNextIntArgReg(ArgReg), RC); 3660 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT); 3661 if (!Subtarget.isLittle()) 3662 std::swap(ArgValue, ArgValue2); 3663 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, 3664 ArgValue, ArgValue2); 3665 } 3666 3667 InVals.push_back(ArgValue); 3668 } else { // VA.isRegLoc() 3669 MVT LocVT = VA.getLocVT(); 3670 3671 if (ABI.IsO32()) { 3672 // We ought to be able to use LocVT directly but O32 sets it to i32 3673 // when allocating floating point values to integer registers. 3674 // This shouldn't influence how we load the value into registers unless 3675 // we are targeting softfloat. 3676 if (VA.getValVT().isFloatingPoint() && !Subtarget.useSoftFloat()) 3677 LocVT = VA.getValVT(); 3678 } 3679 3680 // sanity check 3681 assert(VA.isMemLoc()); 3682 3683 // The stack pointer offset is relative to the caller stack frame. 3684 int FI = MFI.CreateFixedObject(LocVT.getSizeInBits() / 8, 3685 VA.getLocMemOffset(), true); 3686 3687 // Create load nodes to retrieve arguments from the stack 3688 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); 3689 SDValue ArgValue = DAG.getLoad( 3690 LocVT, DL, Chain, FIN, 3691 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)); 3692 OutChains.push_back(ArgValue.getValue(1)); 3693 3694 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG); 3695 3696 InVals.push_back(ArgValue); 3697 } 3698 } 3699 3700 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 3701 // The mips ABIs for returning structs by value requires that we copy 3702 // the sret argument into $v0 for the return. Save the argument into 3703 // a virtual register so that we can access it from the return points. 3704 if (Ins[i].Flags.isSRet()) { 3705 unsigned Reg = MipsFI->getSRetReturnReg(); 3706 if (!Reg) { 3707 Reg = MF.getRegInfo().createVirtualRegister( 3708 getRegClassFor(ABI.IsN64() ? MVT::i64 : MVT::i32)); 3709 MipsFI->setSRetReturnReg(Reg); 3710 } 3711 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[i]); 3712 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain); 3713 break; 3714 } 3715 } 3716 3717 if (IsVarArg) 3718 writeVarArgRegs(OutChains, Chain, DL, DAG, CCInfo); 3719 3720 // All stores are grouped in one node to allow the matching between 3721 // the size of Ins and InVals. This only happens when on varg functions 3722 if (!OutChains.empty()) { 3723 OutChains.push_back(Chain); 3724 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains); 3725 } 3726 3727 return Chain; 3728 } 3729 3730 //===----------------------------------------------------------------------===// 3731 // Return Value Calling Convention Implementation 3732 //===----------------------------------------------------------------------===// 3733 3734 bool 3735 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv, 3736 MachineFunction &MF, bool IsVarArg, 3737 const SmallVectorImpl<ISD::OutputArg> &Outs, 3738 LLVMContext &Context) const { 3739 SmallVector<CCValAssign, 16> RVLocs; 3740 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context); 3741 return CCInfo.CheckReturn(Outs, RetCC_Mips); 3742 } 3743 3744 bool MipsTargetLowering::shouldSignExtendTypeInLibCall(EVT Type, 3745 bool IsSigned) const { 3746 if ((ABI.IsN32() || ABI.IsN64()) && Type == MVT::i32) 3747 return true; 3748 3749 return IsSigned; 3750 } 3751 3752 SDValue 3753 MipsTargetLowering::LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps, 3754 const SDLoc &DL, 3755 SelectionDAG &DAG) const { 3756 MachineFunction &MF = DAG.getMachineFunction(); 3757 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3758 3759 MipsFI->setISR(); 3760 3761 return DAG.getNode(MipsISD::ERet, DL, MVT::Other, RetOps); 3762 } 3763 3764 SDValue 3765 MipsTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, 3766 bool IsVarArg, 3767 const SmallVectorImpl<ISD::OutputArg> &Outs, 3768 const SmallVectorImpl<SDValue> &OutVals, 3769 const SDLoc &DL, SelectionDAG &DAG) const { 3770 // CCValAssign - represent the assignment of 3771 // the return value to a location 3772 SmallVector<CCValAssign, 16> RVLocs; 3773 MachineFunction &MF = DAG.getMachineFunction(); 3774 3775 // CCState - Info about the registers and stack slot. 3776 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext()); 3777 3778 // Analyze return values. 3779 CCInfo.AnalyzeReturn(Outs, RetCC_Mips); 3780 3781 SDValue Flag; 3782 SmallVector<SDValue, 4> RetOps(1, Chain); 3783 3784 // Copy the result values into the output registers. 3785 for (unsigned i = 0; i != RVLocs.size(); ++i) { 3786 SDValue Val = OutVals[i]; 3787 CCValAssign &VA = RVLocs[i]; 3788 assert(VA.isRegLoc() && "Can only return in registers!"); 3789 bool UseUpperBits = false; 3790 3791 switch (VA.getLocInfo()) { 3792 default: 3793 llvm_unreachable("Unknown loc info!"); 3794 case CCValAssign::Full: 3795 break; 3796 case CCValAssign::BCvt: 3797 Val = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Val); 3798 break; 3799 case CCValAssign::AExtUpper: 3800 UseUpperBits = true; 3801 LLVM_FALLTHROUGH; 3802 case CCValAssign::AExt: 3803 Val = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Val); 3804 break; 3805 case CCValAssign::ZExtUpper: 3806 UseUpperBits = true; 3807 LLVM_FALLTHROUGH; 3808 case CCValAssign::ZExt: 3809 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Val); 3810 break; 3811 case CCValAssign::SExtUpper: 3812 UseUpperBits = true; 3813 LLVM_FALLTHROUGH; 3814 case CCValAssign::SExt: 3815 Val = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Val); 3816 break; 3817 } 3818 3819 if (UseUpperBits) { 3820 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits(); 3821 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3822 Val = DAG.getNode( 3823 ISD::SHL, DL, VA.getLocVT(), Val, 3824 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT())); 3825 } 3826 3827 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag); 3828 3829 // Guarantee that all emitted copies are stuck together with flags. 3830 Flag = Chain.getValue(1); 3831 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 3832 } 3833 3834 // The mips ABIs for returning structs by value requires that we copy 3835 // the sret argument into $v0 for the return. We saved the argument into 3836 // a virtual register in the entry block, so now we copy the value out 3837 // and into $v0. 3838 if (MF.getFunction().hasStructRetAttr()) { 3839 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3840 unsigned Reg = MipsFI->getSRetReturnReg(); 3841 3842 if (!Reg) 3843 llvm_unreachable("sret virtual register not created in the entry block"); 3844 SDValue Val = 3845 DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(DAG.getDataLayout())); 3846 unsigned V0 = ABI.IsN64() ? Mips::V0_64 : Mips::V0; 3847 3848 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag); 3849 Flag = Chain.getValue(1); 3850 RetOps.push_back(DAG.getRegister(V0, getPointerTy(DAG.getDataLayout()))); 3851 } 3852 3853 RetOps[0] = Chain; // Update chain. 3854 3855 // Add the flag if we have it. 3856 if (Flag.getNode()) 3857 RetOps.push_back(Flag); 3858 3859 // ISRs must use "eret". 3860 if (DAG.getMachineFunction().getFunction().hasFnAttribute("interrupt")) 3861 return LowerInterruptReturn(RetOps, DL, DAG); 3862 3863 // Standard return on Mips is a "jr $ra" 3864 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, RetOps); 3865 } 3866 3867 //===----------------------------------------------------------------------===// 3868 // Mips Inline Assembly Support 3869 //===----------------------------------------------------------------------===// 3870 3871 /// getConstraintType - Given a constraint letter, return the type of 3872 /// constraint it is for this target. 3873 MipsTargetLowering::ConstraintType 3874 MipsTargetLowering::getConstraintType(StringRef Constraint) const { 3875 // Mips specific constraints 3876 // GCC config/mips/constraints.md 3877 // 3878 // 'd' : An address register. Equivalent to r 3879 // unless generating MIPS16 code. 3880 // 'y' : Equivalent to r; retained for 3881 // backwards compatibility. 3882 // 'c' : A register suitable for use in an indirect 3883 // jump. This will always be $25 for -mabicalls. 3884 // 'l' : The lo register. 1 word storage. 3885 // 'x' : The hilo register pair. Double word storage. 3886 if (Constraint.size() == 1) { 3887 switch (Constraint[0]) { 3888 default : break; 3889 case 'd': 3890 case 'y': 3891 case 'f': 3892 case 'c': 3893 case 'l': 3894 case 'x': 3895 return C_RegisterClass; 3896 case 'R': 3897 return C_Memory; 3898 } 3899 } 3900 3901 if (Constraint == "ZC") 3902 return C_Memory; 3903 3904 return TargetLowering::getConstraintType(Constraint); 3905 } 3906 3907 /// Examine constraint type and operand type and determine a weight value. 3908 /// This object must already have been set up with the operand type 3909 /// and the current alternative constraint selected. 3910 TargetLowering::ConstraintWeight 3911 MipsTargetLowering::getSingleConstraintMatchWeight( 3912 AsmOperandInfo &info, const char *constraint) const { 3913 ConstraintWeight weight = CW_Invalid; 3914 Value *CallOperandVal = info.CallOperandVal; 3915 // If we don't have a value, we can't do a match, 3916 // but allow it at the lowest weight. 3917 if (!CallOperandVal) 3918 return CW_Default; 3919 Type *type = CallOperandVal->getType(); 3920 // Look at the constraint type. 3921 switch (*constraint) { 3922 default: 3923 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 3924 break; 3925 case 'd': 3926 case 'y': 3927 if (type->isIntegerTy()) 3928 weight = CW_Register; 3929 break; 3930 case 'f': // FPU or MSA register 3931 if (Subtarget.hasMSA() && type->isVectorTy() && 3932 cast<VectorType>(type)->getBitWidth() == 128) 3933 weight = CW_Register; 3934 else if (type->isFloatTy()) 3935 weight = CW_Register; 3936 break; 3937 case 'c': // $25 for indirect jumps 3938 case 'l': // lo register 3939 case 'x': // hilo register pair 3940 if (type->isIntegerTy()) 3941 weight = CW_SpecificReg; 3942 break; 3943 case 'I': // signed 16 bit immediate 3944 case 'J': // integer zero 3945 case 'K': // unsigned 16 bit immediate 3946 case 'L': // signed 32 bit immediate where lower 16 bits are 0 3947 case 'N': // immediate in the range of -65535 to -1 (inclusive) 3948 case 'O': // signed 15 bit immediate (+- 16383) 3949 case 'P': // immediate in the range of 65535 to 1 (inclusive) 3950 if (isa<ConstantInt>(CallOperandVal)) 3951 weight = CW_Constant; 3952 break; 3953 case 'R': 3954 weight = CW_Memory; 3955 break; 3956 } 3957 return weight; 3958 } 3959 3960 /// This is a helper function to parse a physical register string and split it 3961 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag 3962 /// that is returned indicates whether parsing was successful. The second flag 3963 /// is true if the numeric part exists. 3964 static std::pair<bool, bool> parsePhysicalReg(StringRef C, StringRef &Prefix, 3965 unsigned long long &Reg) { 3966 if (C.front() != '{' || C.back() != '}') 3967 return std::make_pair(false, false); 3968 3969 // Search for the first numeric character. 3970 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1; 3971 I = std::find_if(B, E, isdigit); 3972 3973 Prefix = StringRef(B, I - B); 3974 3975 // The second flag is set to false if no numeric characters were found. 3976 if (I == E) 3977 return std::make_pair(true, false); 3978 3979 // Parse the numeric characters. 3980 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg), 3981 true); 3982 } 3983 3984 EVT MipsTargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT, 3985 ISD::NodeType) const { 3986 bool Cond = !Subtarget.isABI_O32() && VT.getSizeInBits() == 32; 3987 EVT MinVT = getRegisterType(Context, Cond ? MVT::i64 : MVT::i32); 3988 return VT.bitsLT(MinVT) ? MinVT : VT; 3989 } 3990 3991 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering:: 3992 parseRegForInlineAsmConstraint(StringRef C, MVT VT) const { 3993 const TargetRegisterInfo *TRI = 3994 Subtarget.getRegisterInfo(); 3995 const TargetRegisterClass *RC; 3996 StringRef Prefix; 3997 unsigned long long Reg; 3998 3999 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg); 4000 4001 if (!R.first) 4002 return std::make_pair(0U, nullptr); 4003 4004 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo. 4005 // No numeric characters follow "hi" or "lo". 4006 if (R.second) 4007 return std::make_pair(0U, nullptr); 4008 4009 RC = TRI->getRegClass(Prefix == "hi" ? 4010 Mips::HI32RegClassID : Mips::LO32RegClassID); 4011 return std::make_pair(*(RC->begin()), RC); 4012 } else if (Prefix.startswith("$msa")) { 4013 // Parse $msa(ir|csr|access|save|modify|request|map|unmap) 4014 4015 // No numeric characters follow the name. 4016 if (R.second) 4017 return std::make_pair(0U, nullptr); 4018 4019 Reg = StringSwitch<unsigned long long>(Prefix) 4020 .Case("$msair", Mips::MSAIR) 4021 .Case("$msacsr", Mips::MSACSR) 4022 .Case("$msaaccess", Mips::MSAAccess) 4023 .Case("$msasave", Mips::MSASave) 4024 .Case("$msamodify", Mips::MSAModify) 4025 .Case("$msarequest", Mips::MSARequest) 4026 .Case("$msamap", Mips::MSAMap) 4027 .Case("$msaunmap", Mips::MSAUnmap) 4028 .Default(0); 4029 4030 if (!Reg) 4031 return std::make_pair(0U, nullptr); 4032 4033 RC = TRI->getRegClass(Mips::MSACtrlRegClassID); 4034 return std::make_pair(Reg, RC); 4035 } 4036 4037 if (!R.second) 4038 return std::make_pair(0U, nullptr); 4039 4040 if (Prefix == "$f") { // Parse $f0-$f31. 4041 // If the size of FP registers is 64-bit or Reg is an even number, select 4042 // the 64-bit register class. Otherwise, select the 32-bit register class. 4043 if (VT == MVT::Other) 4044 VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32; 4045 4046 RC = getRegClassFor(VT); 4047 4048 if (RC == &Mips::AFGR64RegClass) { 4049 assert(Reg % 2 == 0); 4050 Reg >>= 1; 4051 } 4052 } else if (Prefix == "$fcc") // Parse $fcc0-$fcc7. 4053 RC = TRI->getRegClass(Mips::FCCRegClassID); 4054 else if (Prefix == "$w") { // Parse $w0-$w31. 4055 RC = getRegClassFor((VT == MVT::Other) ? MVT::v16i8 : VT); 4056 } else { // Parse $0-$31. 4057 assert(Prefix == "$"); 4058 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT); 4059 } 4060 4061 assert(Reg < RC->getNumRegs()); 4062 return std::make_pair(*(RC->begin() + Reg), RC); 4063 } 4064 4065 /// Given a register class constraint, like 'r', if this corresponds directly 4066 /// to an LLVM register class, return a register of 0 and the register class 4067 /// pointer. 4068 std::pair<unsigned, const TargetRegisterClass *> 4069 MipsTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, 4070 StringRef Constraint, 4071 MVT VT) const { 4072 if (Constraint.size() == 1) { 4073 switch (Constraint[0]) { 4074 case 'd': // Address register. Same as 'r' unless generating MIPS16 code. 4075 case 'y': // Same as 'r'. Exists for compatibility. 4076 case 'r': 4077 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) { 4078 if (Subtarget.inMips16Mode()) 4079 return std::make_pair(0U, &Mips::CPU16RegsRegClass); 4080 return std::make_pair(0U, &Mips::GPR32RegClass); 4081 } 4082 if (VT == MVT::i64 && !Subtarget.isGP64bit()) 4083 return std::make_pair(0U, &Mips::GPR32RegClass); 4084 if (VT == MVT::i64 && Subtarget.isGP64bit()) 4085 return std::make_pair(0U, &Mips::GPR64RegClass); 4086 // This will generate an error message 4087 return std::make_pair(0U, nullptr); 4088 case 'f': // FPU or MSA register 4089 if (VT == MVT::v16i8) 4090 return std::make_pair(0U, &Mips::MSA128BRegClass); 4091 else if (VT == MVT::v8i16 || VT == MVT::v8f16) 4092 return std::make_pair(0U, &Mips::MSA128HRegClass); 4093 else if (VT == MVT::v4i32 || VT == MVT::v4f32) 4094 return std::make_pair(0U, &Mips::MSA128WRegClass); 4095 else if (VT == MVT::v2i64 || VT == MVT::v2f64) 4096 return std::make_pair(0U, &Mips::MSA128DRegClass); 4097 else if (VT == MVT::f32) 4098 return std::make_pair(0U, &Mips::FGR32RegClass); 4099 else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) { 4100 if (Subtarget.isFP64bit()) 4101 return std::make_pair(0U, &Mips::FGR64RegClass); 4102 return std::make_pair(0U, &Mips::AFGR64RegClass); 4103 } 4104 break; 4105 case 'c': // register suitable for indirect jump 4106 if (VT == MVT::i32) 4107 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass); 4108 if (VT == MVT::i64) 4109 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass); 4110 // This will generate an error message 4111 return std::make_pair(0U, nullptr); 4112 case 'l': // use the `lo` register to store values 4113 // that are no bigger than a word 4114 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) 4115 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass); 4116 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass); 4117 case 'x': // use the concatenated `hi` and `lo` registers 4118 // to store doubleword values 4119 // Fixme: Not triggering the use of both hi and low 4120 // This will generate an error message 4121 return std::make_pair(0U, nullptr); 4122 } 4123 } 4124 4125 if (!Constraint.empty()) { 4126 std::pair<unsigned, const TargetRegisterClass *> R; 4127 R = parseRegForInlineAsmConstraint(Constraint, VT); 4128 4129 if (R.second) 4130 return R; 4131 } 4132 4133 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); 4134 } 4135 4136 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 4137 /// vector. If it is invalid, don't add anything to Ops. 4138 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 4139 std::string &Constraint, 4140 std::vector<SDValue>&Ops, 4141 SelectionDAG &DAG) const { 4142 SDLoc DL(Op); 4143 SDValue Result; 4144 4145 // Only support length 1 constraints for now. 4146 if (Constraint.length() > 1) return; 4147 4148 char ConstraintLetter = Constraint[0]; 4149 switch (ConstraintLetter) { 4150 default: break; // This will fall through to the generic implementation 4151 case 'I': // Signed 16 bit constant 4152 // If this fails, the parent routine will give an error 4153 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 4154 EVT Type = Op.getValueType(); 4155 int64_t Val = C->getSExtValue(); 4156 if (isInt<16>(Val)) { 4157 Result = DAG.getTargetConstant(Val, DL, Type); 4158 break; 4159 } 4160 } 4161 return; 4162 case 'J': // integer zero 4163 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 4164 EVT Type = Op.getValueType(); 4165 int64_t Val = C->getZExtValue(); 4166 if (Val == 0) { 4167 Result = DAG.getTargetConstant(0, DL, Type); 4168 break; 4169 } 4170 } 4171 return; 4172 case 'K': // unsigned 16 bit immediate 4173 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 4174 EVT Type = Op.getValueType(); 4175 uint64_t Val = (uint64_t)C->getZExtValue(); 4176 if (isUInt<16>(Val)) { 4177 Result = DAG.getTargetConstant(Val, DL, Type); 4178 break; 4179 } 4180 } 4181 return; 4182 case 'L': // signed 32 bit immediate where lower 16 bits are 0 4183 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 4184 EVT Type = Op.getValueType(); 4185 int64_t Val = C->getSExtValue(); 4186 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){ 4187 Result = DAG.getTargetConstant(Val, DL, Type); 4188 break; 4189 } 4190 } 4191 return; 4192 case 'N': // immediate in the range of -65535 to -1 (inclusive) 4193 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 4194 EVT Type = Op.getValueType(); 4195 int64_t Val = C->getSExtValue(); 4196 if ((Val >= -65535) && (Val <= -1)) { 4197 Result = DAG.getTargetConstant(Val, DL, Type); 4198 break; 4199 } 4200 } 4201 return; 4202 case 'O': // signed 15 bit immediate 4203 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 4204 EVT Type = Op.getValueType(); 4205 int64_t Val = C->getSExtValue(); 4206 if ((isInt<15>(Val))) { 4207 Result = DAG.getTargetConstant(Val, DL, Type); 4208 break; 4209 } 4210 } 4211 return; 4212 case 'P': // immediate in the range of 1 to 65535 (inclusive) 4213 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 4214 EVT Type = Op.getValueType(); 4215 int64_t Val = C->getSExtValue(); 4216 if ((Val <= 65535) && (Val >= 1)) { 4217 Result = DAG.getTargetConstant(Val, DL, Type); 4218 break; 4219 } 4220 } 4221 return; 4222 } 4223 4224 if (Result.getNode()) { 4225 Ops.push_back(Result); 4226 return; 4227 } 4228 4229 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 4230 } 4231 4232 bool MipsTargetLowering::isLegalAddressingMode(const DataLayout &DL, 4233 const AddrMode &AM, Type *Ty, 4234 unsigned AS, 4235 Instruction *I) const { 4236 // No global is ever allowed as a base. 4237 if (AM.BaseGV) 4238 return false; 4239 4240 switch (AM.Scale) { 4241 case 0: // "r+i" or just "i", depending on HasBaseReg. 4242 break; 4243 case 1: 4244 if (!AM.HasBaseReg) // allow "r+i". 4245 break; 4246 return false; // disallow "r+r" or "r+r+i". 4247 default: 4248 return false; 4249 } 4250 4251 return true; 4252 } 4253 4254 bool 4255 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 4256 // The Mips target isn't yet aware of offsets. 4257 return false; 4258 } 4259 4260 EVT MipsTargetLowering::getOptimalMemOpType( 4261 uint64_t Size, unsigned DstAlign, unsigned SrcAlign, bool IsMemset, 4262 bool ZeroMemset, bool MemcpyStrSrc, 4263 const AttributeList &FuncAttributes) const { 4264 if (Subtarget.hasMips64()) 4265 return MVT::i64; 4266 4267 return MVT::i32; 4268 } 4269 4270 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, 4271 bool ForCodeSize) const { 4272 if (VT != MVT::f32 && VT != MVT::f64) 4273 return false; 4274 if (Imm.isNegZero()) 4275 return false; 4276 return Imm.isZero(); 4277 } 4278 4279 unsigned MipsTargetLowering::getJumpTableEncoding() const { 4280 4281 // FIXME: For space reasons this should be: EK_GPRel32BlockAddress. 4282 if (ABI.IsN64() && isPositionIndependent()) 4283 return MachineJumpTableInfo::EK_GPRel64BlockAddress; 4284 4285 return TargetLowering::getJumpTableEncoding(); 4286 } 4287 4288 bool MipsTargetLowering::useSoftFloat() const { 4289 return Subtarget.useSoftFloat(); 4290 } 4291 4292 void MipsTargetLowering::copyByValRegs( 4293 SDValue Chain, const SDLoc &DL, std::vector<SDValue> &OutChains, 4294 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags, 4295 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg, 4296 unsigned FirstReg, unsigned LastReg, const CCValAssign &VA, 4297 MipsCCState &State) const { 4298 MachineFunction &MF = DAG.getMachineFunction(); 4299 MachineFrameInfo &MFI = MF.getFrameInfo(); 4300 unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes(); 4301 unsigned NumRegs = LastReg - FirstReg; 4302 unsigned RegAreaSize = NumRegs * GPRSizeInBytes; 4303 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize); 4304 int FrameObjOffset; 4305 ArrayRef<MCPhysReg> ByValArgRegs = ABI.GetByValArgRegs(); 4306 4307 if (RegAreaSize) 4308 FrameObjOffset = 4309 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) - 4310 (int)((ByValArgRegs.size() - FirstReg) * GPRSizeInBytes); 4311 else 4312 FrameObjOffset = VA.getLocMemOffset(); 4313 4314 // Create frame object. 4315 EVT PtrTy = getPointerTy(DAG.getDataLayout()); 4316 // Make the fixed object stored to mutable so that the load instructions 4317 // referencing it have their memory dependencies added. 4318 // Set the frame object as isAliased which clears the underlying objects 4319 // vector in ScheduleDAGInstrs::buildSchedGraph() resulting in addition of all 4320 // stores as dependencies for loads referencing this fixed object. 4321 int FI = MFI.CreateFixedObject(FrameObjSize, FrameObjOffset, false, true); 4322 SDValue FIN = DAG.getFrameIndex(FI, PtrTy); 4323 InVals.push_back(FIN); 4324 4325 if (!NumRegs) 4326 return; 4327 4328 // Copy arg registers. 4329 MVT RegTy = MVT::getIntegerVT(GPRSizeInBytes * 8); 4330 const TargetRegisterClass *RC = getRegClassFor(RegTy); 4331 4332 for (unsigned I = 0; I < NumRegs; ++I) { 4333 unsigned ArgReg = ByValArgRegs[FirstReg + I]; 4334 unsigned VReg = addLiveIn(MF, ArgReg, RC); 4335 unsigned Offset = I * GPRSizeInBytes; 4336 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN, 4337 DAG.getConstant(Offset, DL, PtrTy)); 4338 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy), 4339 StorePtr, MachinePointerInfo(FuncArg, Offset)); 4340 OutChains.push_back(Store); 4341 } 4342 } 4343 4344 // Copy byVal arg to registers and stack. 4345 void MipsTargetLowering::passByValArg( 4346 SDValue Chain, const SDLoc &DL, 4347 std::deque<std::pair<unsigned, SDValue>> &RegsToPass, 4348 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr, 4349 MachineFrameInfo &MFI, SelectionDAG &DAG, SDValue Arg, unsigned FirstReg, 4350 unsigned LastReg, const ISD::ArgFlagsTy &Flags, bool isLittle, 4351 const CCValAssign &VA) const { 4352 unsigned ByValSizeInBytes = Flags.getByValSize(); 4353 unsigned OffsetInBytes = 0; // From beginning of struct 4354 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 4355 unsigned Alignment = std::min(Flags.getByValAlign(), RegSizeInBytes); 4356 EVT PtrTy = getPointerTy(DAG.getDataLayout()), 4357 RegTy = MVT::getIntegerVT(RegSizeInBytes * 8); 4358 unsigned NumRegs = LastReg - FirstReg; 4359 4360 if (NumRegs) { 4361 ArrayRef<MCPhysReg> ArgRegs = ABI.GetByValArgRegs(); 4362 bool LeftoverBytes = (NumRegs * RegSizeInBytes > ByValSizeInBytes); 4363 unsigned I = 0; 4364 4365 // Copy words to registers. 4366 for (; I < NumRegs - LeftoverBytes; ++I, OffsetInBytes += RegSizeInBytes) { 4367 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 4368 DAG.getConstant(OffsetInBytes, DL, PtrTy)); 4369 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr, 4370 MachinePointerInfo(), Alignment); 4371 MemOpChains.push_back(LoadVal.getValue(1)); 4372 unsigned ArgReg = ArgRegs[FirstReg + I]; 4373 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal)); 4374 } 4375 4376 // Return if the struct has been fully copied. 4377 if (ByValSizeInBytes == OffsetInBytes) 4378 return; 4379 4380 // Copy the remainder of the byval argument with sub-word loads and shifts. 4381 if (LeftoverBytes) { 4382 SDValue Val; 4383 4384 for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0; 4385 OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) { 4386 unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes; 4387 4388 if (RemainingSizeInBytes < LoadSizeInBytes) 4389 continue; 4390 4391 // Load subword. 4392 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 4393 DAG.getConstant(OffsetInBytes, DL, 4394 PtrTy)); 4395 SDValue LoadVal = DAG.getExtLoad( 4396 ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, MachinePointerInfo(), 4397 MVT::getIntegerVT(LoadSizeInBytes * 8), Alignment); 4398 MemOpChains.push_back(LoadVal.getValue(1)); 4399 4400 // Shift the loaded value. 4401 unsigned Shamt; 4402 4403 if (isLittle) 4404 Shamt = TotalBytesLoaded * 8; 4405 else 4406 Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8; 4407 4408 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal, 4409 DAG.getConstant(Shamt, DL, MVT::i32)); 4410 4411 if (Val.getNode()) 4412 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift); 4413 else 4414 Val = Shift; 4415 4416 OffsetInBytes += LoadSizeInBytes; 4417 TotalBytesLoaded += LoadSizeInBytes; 4418 Alignment = std::min(Alignment, LoadSizeInBytes); 4419 } 4420 4421 unsigned ArgReg = ArgRegs[FirstReg + I]; 4422 RegsToPass.push_back(std::make_pair(ArgReg, Val)); 4423 return; 4424 } 4425 } 4426 4427 // Copy remainder of byval arg to it with memcpy. 4428 unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes; 4429 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 4430 DAG.getConstant(OffsetInBytes, DL, PtrTy)); 4431 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr, 4432 DAG.getIntPtrConstant(VA.getLocMemOffset(), DL)); 4433 Chain = DAG.getMemcpy(Chain, DL, Dst, Src, 4434 DAG.getConstant(MemCpySize, DL, PtrTy), 4435 Alignment, /*isVolatile=*/false, /*AlwaysInline=*/false, 4436 /*isTailCall=*/false, 4437 MachinePointerInfo(), MachinePointerInfo()); 4438 MemOpChains.push_back(Chain); 4439 } 4440 4441 void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains, 4442 SDValue Chain, const SDLoc &DL, 4443 SelectionDAG &DAG, 4444 CCState &State) const { 4445 ArrayRef<MCPhysReg> ArgRegs = ABI.GetVarArgRegs(); 4446 unsigned Idx = State.getFirstUnallocated(ArgRegs); 4447 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 4448 MVT RegTy = MVT::getIntegerVT(RegSizeInBytes * 8); 4449 const TargetRegisterClass *RC = getRegClassFor(RegTy); 4450 MachineFunction &MF = DAG.getMachineFunction(); 4451 MachineFrameInfo &MFI = MF.getFrameInfo(); 4452 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 4453 4454 // Offset of the first variable argument from stack pointer. 4455 int VaArgOffset; 4456 4457 if (ArgRegs.size() == Idx) 4458 VaArgOffset = alignTo(State.getNextStackOffset(), RegSizeInBytes); 4459 else { 4460 VaArgOffset = 4461 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) - 4462 (int)(RegSizeInBytes * (ArgRegs.size() - Idx)); 4463 } 4464 4465 // Record the frame index of the first variable argument 4466 // which is a value necessary to VASTART. 4467 int FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true); 4468 MipsFI->setVarArgsFrameIndex(FI); 4469 4470 // Copy the integer registers that have not been used for argument passing 4471 // to the argument register save area. For O32, the save area is allocated 4472 // in the caller's stack frame, while for N32/64, it is allocated in the 4473 // callee's stack frame. 4474 for (unsigned I = Idx; I < ArgRegs.size(); 4475 ++I, VaArgOffset += RegSizeInBytes) { 4476 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC); 4477 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy); 4478 FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true); 4479 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); 4480 SDValue Store = 4481 DAG.getStore(Chain, DL, ArgValue, PtrOff, MachinePointerInfo()); 4482 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue( 4483 (Value *)nullptr); 4484 OutChains.push_back(Store); 4485 } 4486 } 4487 4488 void MipsTargetLowering::HandleByVal(CCState *State, unsigned &Size, 4489 unsigned Align) const { 4490 const TargetFrameLowering *TFL = Subtarget.getFrameLowering(); 4491 4492 assert(Size && "Byval argument's size shouldn't be 0."); 4493 4494 Align = std::min(Align, TFL->getStackAlignment()); 4495 4496 unsigned FirstReg = 0; 4497 unsigned NumRegs = 0; 4498 4499 if (State->getCallingConv() != CallingConv::Fast) { 4500 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 4501 ArrayRef<MCPhysReg> IntArgRegs = ABI.GetByValArgRegs(); 4502 // FIXME: The O32 case actually describes no shadow registers. 4503 const MCPhysReg *ShadowRegs = 4504 ABI.IsO32() ? IntArgRegs.data() : Mips64DPRegs; 4505 4506 // We used to check the size as well but we can't do that anymore since 4507 // CCState::HandleByVal() rounds up the size after calling this function. 4508 assert(!(Align % RegSizeInBytes) && 4509 "Byval argument's alignment should be a multiple of" 4510 "RegSizeInBytes."); 4511 4512 FirstReg = State->getFirstUnallocated(IntArgRegs); 4513 4514 // If Align > RegSizeInBytes, the first arg register must be even. 4515 // FIXME: This condition happens to do the right thing but it's not the 4516 // right way to test it. We want to check that the stack frame offset 4517 // of the register is aligned. 4518 if ((Align > RegSizeInBytes) && (FirstReg % 2)) { 4519 State->AllocateReg(IntArgRegs[FirstReg], ShadowRegs[FirstReg]); 4520 ++FirstReg; 4521 } 4522 4523 // Mark the registers allocated. 4524 Size = alignTo(Size, RegSizeInBytes); 4525 for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size()); 4526 Size -= RegSizeInBytes, ++I, ++NumRegs) 4527 State->AllocateReg(IntArgRegs[I], ShadowRegs[I]); 4528 } 4529 4530 State->addInRegsParamInfo(FirstReg, FirstReg + NumRegs); 4531 } 4532 4533 MachineBasicBlock *MipsTargetLowering::emitPseudoSELECT(MachineInstr &MI, 4534 MachineBasicBlock *BB, 4535 bool isFPCmp, 4536 unsigned Opc) const { 4537 assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) && 4538 "Subtarget already supports SELECT nodes with the use of" 4539 "conditional-move instructions."); 4540 4541 const TargetInstrInfo *TII = 4542 Subtarget.getInstrInfo(); 4543 DebugLoc DL = MI.getDebugLoc(); 4544 4545 // To "insert" a SELECT instruction, we actually have to insert the 4546 // diamond control-flow pattern. The incoming instruction knows the 4547 // destination vreg to set, the condition code register to branch on, the 4548 // true/false values to select between, and a branch opcode to use. 4549 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 4550 MachineFunction::iterator It = ++BB->getIterator(); 4551 4552 // thisMBB: 4553 // ... 4554 // TrueVal = ... 4555 // setcc r1, r2, r3 4556 // bNE r1, r0, copy1MBB 4557 // fallthrough --> copy0MBB 4558 MachineBasicBlock *thisMBB = BB; 4559 MachineFunction *F = BB->getParent(); 4560 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 4561 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 4562 F->insert(It, copy0MBB); 4563 F->insert(It, sinkMBB); 4564 4565 // Transfer the remainder of BB and its successor edges to sinkMBB. 4566 sinkMBB->splice(sinkMBB->begin(), BB, 4567 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 4568 sinkMBB->transferSuccessorsAndUpdatePHIs(BB); 4569 4570 // Next, add the true and fallthrough blocks as its successors. 4571 BB->addSuccessor(copy0MBB); 4572 BB->addSuccessor(sinkMBB); 4573 4574 if (isFPCmp) { 4575 // bc1[tf] cc, sinkMBB 4576 BuildMI(BB, DL, TII->get(Opc)) 4577 .addReg(MI.getOperand(1).getReg()) 4578 .addMBB(sinkMBB); 4579 } else { 4580 // bne rs, $0, sinkMBB 4581 BuildMI(BB, DL, TII->get(Opc)) 4582 .addReg(MI.getOperand(1).getReg()) 4583 .addReg(Mips::ZERO) 4584 .addMBB(sinkMBB); 4585 } 4586 4587 // copy0MBB: 4588 // %FalseValue = ... 4589 // # fallthrough to sinkMBB 4590 BB = copy0MBB; 4591 4592 // Update machine-CFG edges 4593 BB->addSuccessor(sinkMBB); 4594 4595 // sinkMBB: 4596 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] 4597 // ... 4598 BB = sinkMBB; 4599 4600 BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(0).getReg()) 4601 .addReg(MI.getOperand(2).getReg()) 4602 .addMBB(thisMBB) 4603 .addReg(MI.getOperand(3).getReg()) 4604 .addMBB(copy0MBB); 4605 4606 MI.eraseFromParent(); // The pseudo instruction is gone now. 4607 4608 return BB; 4609 } 4610 4611 MachineBasicBlock * 4612 MipsTargetLowering::emitPseudoD_SELECT(MachineInstr &MI, 4613 MachineBasicBlock *BB) const { 4614 assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) && 4615 "Subtarget already supports SELECT nodes with the use of" 4616 "conditional-move instructions."); 4617 4618 const TargetInstrInfo *TII = Subtarget.getInstrInfo(); 4619 DebugLoc DL = MI.getDebugLoc(); 4620 4621 // D_SELECT substitutes two SELECT nodes that goes one after another and 4622 // have the same condition operand. On machines which don't have 4623 // conditional-move instruction, it reduces unnecessary branch instructions 4624 // which are result of using two diamond patterns that are result of two 4625 // SELECT pseudo instructions. 4626 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 4627 MachineFunction::iterator It = ++BB->getIterator(); 4628 4629 // thisMBB: 4630 // ... 4631 // TrueVal = ... 4632 // setcc r1, r2, r3 4633 // bNE r1, r0, copy1MBB 4634 // fallthrough --> copy0MBB 4635 MachineBasicBlock *thisMBB = BB; 4636 MachineFunction *F = BB->getParent(); 4637 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 4638 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 4639 F->insert(It, copy0MBB); 4640 F->insert(It, sinkMBB); 4641 4642 // Transfer the remainder of BB and its successor edges to sinkMBB. 4643 sinkMBB->splice(sinkMBB->begin(), BB, 4644 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 4645 sinkMBB->transferSuccessorsAndUpdatePHIs(BB); 4646 4647 // Next, add the true and fallthrough blocks as its successors. 4648 BB->addSuccessor(copy0MBB); 4649 BB->addSuccessor(sinkMBB); 4650 4651 // bne rs, $0, sinkMBB 4652 BuildMI(BB, DL, TII->get(Mips::BNE)) 4653 .addReg(MI.getOperand(2).getReg()) 4654 .addReg(Mips::ZERO) 4655 .addMBB(sinkMBB); 4656 4657 // copy0MBB: 4658 // %FalseValue = ... 4659 // # fallthrough to sinkMBB 4660 BB = copy0MBB; 4661 4662 // Update machine-CFG edges 4663 BB->addSuccessor(sinkMBB); 4664 4665 // sinkMBB: 4666 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] 4667 // ... 4668 BB = sinkMBB; 4669 4670 // Use two PHI nodes to select two reults 4671 BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(0).getReg()) 4672 .addReg(MI.getOperand(3).getReg()) 4673 .addMBB(thisMBB) 4674 .addReg(MI.getOperand(5).getReg()) 4675 .addMBB(copy0MBB); 4676 BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(1).getReg()) 4677 .addReg(MI.getOperand(4).getReg()) 4678 .addMBB(thisMBB) 4679 .addReg(MI.getOperand(6).getReg()) 4680 .addMBB(copy0MBB); 4681 4682 MI.eraseFromParent(); // The pseudo instruction is gone now. 4683 4684 return BB; 4685 } 4686 4687 // FIXME? Maybe this could be a TableGen attribute on some registers and 4688 // this table could be generated automatically from RegInfo. 4689 Register 4690 MipsTargetLowering::getRegisterByName(const char *RegName, EVT VT, 4691 const MachineFunction &MF) const { 4692 // Named registers is expected to be fairly rare. For now, just support $28 4693 // since the linux kernel uses it. 4694 if (Subtarget.isGP64bit()) { 4695 Register Reg = StringSwitch<Register>(RegName) 4696 .Case("$28", Mips::GP_64) 4697 .Default(Register()); 4698 if (Reg) 4699 return Reg; 4700 } else { 4701 Register Reg = StringSwitch<Register>(RegName) 4702 .Case("$28", Mips::GP) 4703 .Default(Register()); 4704 if (Reg) 4705 return Reg; 4706 } 4707 report_fatal_error("Invalid register name global variable"); 4708 } 4709