1 //===- X86InstructionSelector.cpp -----------------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 /// \file 10 /// This file implements the targeting of the InstructionSelector class for 11 /// X86. 12 /// \todo This should be generated by TableGen. 13 //===----------------------------------------------------------------------===// 14 15 #include "MCTargetDesc/X86BaseInfo.h" 16 #include "X86InstrBuilder.h" 17 #include "X86InstrInfo.h" 18 #include "X86RegisterBankInfo.h" 19 #include "X86RegisterInfo.h" 20 #include "X86Subtarget.h" 21 #include "X86TargetMachine.h" 22 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h" 23 #include "llvm/CodeGen/GlobalISel/InstructionSelectorImpl.h" 24 #include "llvm/CodeGen/GlobalISel/RegisterBank.h" 25 #include "llvm/CodeGen/GlobalISel/Utils.h" 26 #include "llvm/CodeGen/MachineBasicBlock.h" 27 #include "llvm/CodeGen/MachineConstantPool.h" 28 #include "llvm/CodeGen/MachineFunction.h" 29 #include "llvm/CodeGen/MachineInstr.h" 30 #include "llvm/CodeGen/MachineInstrBuilder.h" 31 #include "llvm/CodeGen/MachineMemOperand.h" 32 #include "llvm/CodeGen/MachineOperand.h" 33 #include "llvm/CodeGen/MachineRegisterInfo.h" 34 #include "llvm/CodeGen/TargetOpcodes.h" 35 #include "llvm/CodeGen/TargetRegisterInfo.h" 36 #include "llvm/IR/DataLayout.h" 37 #include "llvm/IR/InstrTypes.h" 38 #include "llvm/Support/AtomicOrdering.h" 39 #include "llvm/Support/CodeGen.h" 40 #include "llvm/Support/Debug.h" 41 #include "llvm/Support/ErrorHandling.h" 42 #include "llvm/Support/LowLevelTypeImpl.h" 43 #include "llvm/Support/MathExtras.h" 44 #include "llvm/Support/raw_ostream.h" 45 #include <cassert> 46 #include <cstdint> 47 #include <tuple> 48 49 #define DEBUG_TYPE "X86-isel" 50 51 using namespace llvm; 52 53 namespace { 54 55 #define GET_GLOBALISEL_PREDICATE_BITSET 56 #include "X86GenGlobalISel.inc" 57 #undef GET_GLOBALISEL_PREDICATE_BITSET 58 59 class X86InstructionSelector : public InstructionSelector { 60 public: 61 X86InstructionSelector(const X86TargetMachine &TM, const X86Subtarget &STI, 62 const X86RegisterBankInfo &RBI); 63 64 bool select(MachineInstr &I, CodeGenCoverage &CoverageInfo) const override; 65 static const char *getName() { return DEBUG_TYPE; } 66 67 private: 68 /// tblgen-erated 'select' implementation, used as the initial selector for 69 /// the patterns that don't require complex C++. 70 bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const; 71 72 // TODO: remove after supported by Tablegen-erated instruction selection. 73 unsigned getLoadStoreOp(const LLT &Ty, const RegisterBank &RB, unsigned Opc, 74 uint64_t Alignment) const; 75 76 bool selectLoadStoreOp(MachineInstr &I, MachineRegisterInfo &MRI, 77 MachineFunction &MF) const; 78 bool selectFrameIndexOrGep(MachineInstr &I, MachineRegisterInfo &MRI, 79 MachineFunction &MF) const; 80 bool selectGlobalValue(MachineInstr &I, MachineRegisterInfo &MRI, 81 MachineFunction &MF) const; 82 bool selectConstant(MachineInstr &I, MachineRegisterInfo &MRI, 83 MachineFunction &MF) const; 84 bool selectTruncOrPtrToInt(MachineInstr &I, MachineRegisterInfo &MRI, 85 MachineFunction &MF) const; 86 bool selectZext(MachineInstr &I, MachineRegisterInfo &MRI, 87 MachineFunction &MF) const; 88 bool selectAnyext(MachineInstr &I, MachineRegisterInfo &MRI, 89 MachineFunction &MF) const; 90 bool selectCmp(MachineInstr &I, MachineRegisterInfo &MRI, 91 MachineFunction &MF) const; 92 bool selectFCmp(MachineInstr &I, MachineRegisterInfo &MRI, 93 MachineFunction &MF) const; 94 bool selectUadde(MachineInstr &I, MachineRegisterInfo &MRI, 95 MachineFunction &MF) const; 96 bool selectCopy(MachineInstr &I, MachineRegisterInfo &MRI) const; 97 bool selectUnmergeValues(MachineInstr &I, MachineRegisterInfo &MRI, 98 MachineFunction &MF, 99 CodeGenCoverage &CoverageInfo) const; 100 bool selectMergeValues(MachineInstr &I, MachineRegisterInfo &MRI, 101 MachineFunction &MF, 102 CodeGenCoverage &CoverageInfo) const; 103 bool selectInsert(MachineInstr &I, MachineRegisterInfo &MRI, 104 MachineFunction &MF) const; 105 bool selectExtract(MachineInstr &I, MachineRegisterInfo &MRI, 106 MachineFunction &MF) const; 107 bool selectCondBranch(MachineInstr &I, MachineRegisterInfo &MRI, 108 MachineFunction &MF) const; 109 bool selectTurnIntoCOPY(MachineInstr &I, MachineRegisterInfo &MRI, 110 const unsigned DstReg, 111 const TargetRegisterClass *DstRC, 112 const unsigned SrcReg, 113 const TargetRegisterClass *SrcRC) const; 114 bool materializeFP(MachineInstr &I, MachineRegisterInfo &MRI, 115 MachineFunction &MF) const; 116 bool selectImplicitDefOrPHI(MachineInstr &I, MachineRegisterInfo &MRI) const; 117 bool selectShift(MachineInstr &I, MachineRegisterInfo &MRI, 118 MachineFunction &MF) const; 119 bool selectDivRem(MachineInstr &I, MachineRegisterInfo &MRI, 120 MachineFunction &MF) const; 121 bool selectIntrinsicWSideEffects(MachineInstr &I, MachineRegisterInfo &MRI, 122 MachineFunction &MF) const; 123 124 // emit insert subreg instruction and insert it before MachineInstr &I 125 bool emitInsertSubreg(unsigned DstReg, unsigned SrcReg, MachineInstr &I, 126 MachineRegisterInfo &MRI, MachineFunction &MF) const; 127 // emit extract subreg instruction and insert it before MachineInstr &I 128 bool emitExtractSubreg(unsigned DstReg, unsigned SrcReg, MachineInstr &I, 129 MachineRegisterInfo &MRI, MachineFunction &MF) const; 130 131 const TargetRegisterClass *getRegClass(LLT Ty, const RegisterBank &RB) const; 132 const TargetRegisterClass *getRegClass(LLT Ty, unsigned Reg, 133 MachineRegisterInfo &MRI) const; 134 135 const X86TargetMachine &TM; 136 const X86Subtarget &STI; 137 const X86InstrInfo &TII; 138 const X86RegisterInfo &TRI; 139 const X86RegisterBankInfo &RBI; 140 141 #define GET_GLOBALISEL_PREDICATES_DECL 142 #include "X86GenGlobalISel.inc" 143 #undef GET_GLOBALISEL_PREDICATES_DECL 144 145 #define GET_GLOBALISEL_TEMPORARIES_DECL 146 #include "X86GenGlobalISel.inc" 147 #undef GET_GLOBALISEL_TEMPORARIES_DECL 148 }; 149 150 } // end anonymous namespace 151 152 #define GET_GLOBALISEL_IMPL 153 #include "X86GenGlobalISel.inc" 154 #undef GET_GLOBALISEL_IMPL 155 156 X86InstructionSelector::X86InstructionSelector(const X86TargetMachine &TM, 157 const X86Subtarget &STI, 158 const X86RegisterBankInfo &RBI) 159 : InstructionSelector(), TM(TM), STI(STI), TII(*STI.getInstrInfo()), 160 TRI(*STI.getRegisterInfo()), RBI(RBI), 161 #define GET_GLOBALISEL_PREDICATES_INIT 162 #include "X86GenGlobalISel.inc" 163 #undef GET_GLOBALISEL_PREDICATES_INIT 164 #define GET_GLOBALISEL_TEMPORARIES_INIT 165 #include "X86GenGlobalISel.inc" 166 #undef GET_GLOBALISEL_TEMPORARIES_INIT 167 { 168 } 169 170 // FIXME: This should be target-independent, inferred from the types declared 171 // for each class in the bank. 172 const TargetRegisterClass * 173 X86InstructionSelector::getRegClass(LLT Ty, const RegisterBank &RB) const { 174 if (RB.getID() == X86::GPRRegBankID) { 175 if (Ty.getSizeInBits() <= 8) 176 return &X86::GR8RegClass; 177 if (Ty.getSizeInBits() == 16) 178 return &X86::GR16RegClass; 179 if (Ty.getSizeInBits() == 32) 180 return &X86::GR32RegClass; 181 if (Ty.getSizeInBits() == 64) 182 return &X86::GR64RegClass; 183 } 184 if (RB.getID() == X86::VECRRegBankID) { 185 if (Ty.getSizeInBits() == 32) 186 return STI.hasAVX512() ? &X86::FR32XRegClass : &X86::FR32RegClass; 187 if (Ty.getSizeInBits() == 64) 188 return STI.hasAVX512() ? &X86::FR64XRegClass : &X86::FR64RegClass; 189 if (Ty.getSizeInBits() == 128) 190 return STI.hasAVX512() ? &X86::VR128XRegClass : &X86::VR128RegClass; 191 if (Ty.getSizeInBits() == 256) 192 return STI.hasAVX512() ? &X86::VR256XRegClass : &X86::VR256RegClass; 193 if (Ty.getSizeInBits() == 512) 194 return &X86::VR512RegClass; 195 } 196 197 llvm_unreachable("Unknown RegBank!"); 198 } 199 200 const TargetRegisterClass * 201 X86InstructionSelector::getRegClass(LLT Ty, unsigned Reg, 202 MachineRegisterInfo &MRI) const { 203 const RegisterBank &RegBank = *RBI.getRegBank(Reg, MRI, TRI); 204 return getRegClass(Ty, RegBank); 205 } 206 207 static unsigned getSubRegIndex(const TargetRegisterClass *RC) { 208 unsigned SubIdx = X86::NoSubRegister; 209 if (RC == &X86::GR32RegClass) { 210 SubIdx = X86::sub_32bit; 211 } else if (RC == &X86::GR16RegClass) { 212 SubIdx = X86::sub_16bit; 213 } else if (RC == &X86::GR8RegClass) { 214 SubIdx = X86::sub_8bit; 215 } 216 217 return SubIdx; 218 } 219 220 static const TargetRegisterClass *getRegClassFromGRPhysReg(unsigned Reg) { 221 assert(TargetRegisterInfo::isPhysicalRegister(Reg)); 222 if (X86::GR64RegClass.contains(Reg)) 223 return &X86::GR64RegClass; 224 if (X86::GR32RegClass.contains(Reg)) 225 return &X86::GR32RegClass; 226 if (X86::GR16RegClass.contains(Reg)) 227 return &X86::GR16RegClass; 228 if (X86::GR8RegClass.contains(Reg)) 229 return &X86::GR8RegClass; 230 231 llvm_unreachable("Unknown RegClass for PhysReg!"); 232 } 233 234 // Set X86 Opcode and constrain DestReg. 235 bool X86InstructionSelector::selectCopy(MachineInstr &I, 236 MachineRegisterInfo &MRI) const { 237 unsigned DstReg = I.getOperand(0).getReg(); 238 const unsigned DstSize = RBI.getSizeInBits(DstReg, MRI, TRI); 239 const RegisterBank &DstRegBank = *RBI.getRegBank(DstReg, MRI, TRI); 240 241 unsigned SrcReg = I.getOperand(1).getReg(); 242 const unsigned SrcSize = RBI.getSizeInBits(SrcReg, MRI, TRI); 243 const RegisterBank &SrcRegBank = *RBI.getRegBank(SrcReg, MRI, TRI); 244 245 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) { 246 assert(I.isCopy() && "Generic operators do not allow physical registers"); 247 248 if (DstSize > SrcSize && SrcRegBank.getID() == X86::GPRRegBankID && 249 DstRegBank.getID() == X86::GPRRegBankID) { 250 251 const TargetRegisterClass *SrcRC = 252 getRegClass(MRI.getType(SrcReg), SrcRegBank); 253 const TargetRegisterClass *DstRC = getRegClassFromGRPhysReg(DstReg); 254 255 if (SrcRC != DstRC) { 256 // This case can be generated by ABI lowering, performe anyext 257 unsigned ExtSrc = MRI.createVirtualRegister(DstRC); 258 BuildMI(*I.getParent(), I, I.getDebugLoc(), 259 TII.get(TargetOpcode::SUBREG_TO_REG)) 260 .addDef(ExtSrc) 261 .addImm(0) 262 .addReg(SrcReg) 263 .addImm(getSubRegIndex(SrcRC)); 264 265 I.getOperand(1).setReg(ExtSrc); 266 } 267 } 268 269 return true; 270 } 271 272 assert((!TargetRegisterInfo::isPhysicalRegister(SrcReg) || I.isCopy()) && 273 "No phys reg on generic operators"); 274 assert((DstSize == SrcSize || 275 // Copies are a mean to setup initial types, the number of 276 // bits may not exactly match. 277 (TargetRegisterInfo::isPhysicalRegister(SrcReg) && 278 DstSize <= RBI.getSizeInBits(SrcReg, MRI, TRI))) && 279 "Copy with different width?!"); 280 281 const TargetRegisterClass *DstRC = 282 getRegClass(MRI.getType(DstReg), DstRegBank); 283 284 if (SrcRegBank.getID() == X86::GPRRegBankID && 285 DstRegBank.getID() == X86::GPRRegBankID && SrcSize > DstSize && 286 TargetRegisterInfo::isPhysicalRegister(SrcReg)) { 287 // Change the physical register to performe truncate. 288 289 const TargetRegisterClass *SrcRC = getRegClassFromGRPhysReg(SrcReg); 290 291 if (DstRC != SrcRC) { 292 I.getOperand(1).setSubReg(getSubRegIndex(DstRC)); 293 I.getOperand(1).substPhysReg(SrcReg, TRI); 294 } 295 } 296 297 // No need to constrain SrcReg. It will get constrained when 298 // we hit another of its use or its defs. 299 // Copies do not have constraints. 300 const TargetRegisterClass *OldRC = MRI.getRegClassOrNull(DstReg); 301 if (!OldRC || !DstRC->hasSubClassEq(OldRC)) { 302 if (!RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 303 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 304 << " operand\n"); 305 return false; 306 } 307 } 308 I.setDesc(TII.get(X86::COPY)); 309 return true; 310 } 311 312 bool X86InstructionSelector::select(MachineInstr &I, 313 CodeGenCoverage &CoverageInfo) const { 314 assert(I.getParent() && "Instruction should be in a basic block!"); 315 assert(I.getParent()->getParent() && "Instruction should be in a function!"); 316 317 MachineBasicBlock &MBB = *I.getParent(); 318 MachineFunction &MF = *MBB.getParent(); 319 MachineRegisterInfo &MRI = MF.getRegInfo(); 320 321 unsigned Opcode = I.getOpcode(); 322 if (!isPreISelGenericOpcode(Opcode)) { 323 // Certain non-generic instructions also need some special handling. 324 325 if (Opcode == TargetOpcode::LOAD_STACK_GUARD) 326 return false; 327 328 if (I.isCopy()) 329 return selectCopy(I, MRI); 330 331 return true; 332 } 333 334 assert(I.getNumOperands() == I.getNumExplicitOperands() && 335 "Generic instruction has unexpected implicit operands\n"); 336 337 if (selectImpl(I, CoverageInfo)) 338 return true; 339 340 LLVM_DEBUG(dbgs() << " C++ instruction selection: "; I.print(dbgs())); 341 342 // TODO: This should be implemented by tblgen. 343 switch (I.getOpcode()) { 344 default: 345 return false; 346 case TargetOpcode::G_STORE: 347 case TargetOpcode::G_LOAD: 348 return selectLoadStoreOp(I, MRI, MF); 349 case TargetOpcode::G_GEP: 350 case TargetOpcode::G_FRAME_INDEX: 351 return selectFrameIndexOrGep(I, MRI, MF); 352 case TargetOpcode::G_GLOBAL_VALUE: 353 return selectGlobalValue(I, MRI, MF); 354 case TargetOpcode::G_CONSTANT: 355 return selectConstant(I, MRI, MF); 356 case TargetOpcode::G_FCONSTANT: 357 return materializeFP(I, MRI, MF); 358 case TargetOpcode::G_PTRTOINT: 359 case TargetOpcode::G_TRUNC: 360 return selectTruncOrPtrToInt(I, MRI, MF); 361 case TargetOpcode::G_INTTOPTR: 362 return selectCopy(I, MRI); 363 case TargetOpcode::G_ZEXT: 364 return selectZext(I, MRI, MF); 365 case TargetOpcode::G_ANYEXT: 366 return selectAnyext(I, MRI, MF); 367 case TargetOpcode::G_ICMP: 368 return selectCmp(I, MRI, MF); 369 case TargetOpcode::G_FCMP: 370 return selectFCmp(I, MRI, MF); 371 case TargetOpcode::G_UADDE: 372 return selectUadde(I, MRI, MF); 373 case TargetOpcode::G_UNMERGE_VALUES: 374 return selectUnmergeValues(I, MRI, MF, CoverageInfo); 375 case TargetOpcode::G_MERGE_VALUES: 376 return selectMergeValues(I, MRI, MF, CoverageInfo); 377 case TargetOpcode::G_EXTRACT: 378 return selectExtract(I, MRI, MF); 379 case TargetOpcode::G_INSERT: 380 return selectInsert(I, MRI, MF); 381 case TargetOpcode::G_BRCOND: 382 return selectCondBranch(I, MRI, MF); 383 case TargetOpcode::G_IMPLICIT_DEF: 384 case TargetOpcode::G_PHI: 385 return selectImplicitDefOrPHI(I, MRI); 386 case TargetOpcode::G_SHL: 387 case TargetOpcode::G_ASHR: 388 case TargetOpcode::G_LSHR: 389 return selectShift(I, MRI, MF); 390 case TargetOpcode::G_SDIV: 391 case TargetOpcode::G_UDIV: 392 case TargetOpcode::G_SREM: 393 case TargetOpcode::G_UREM: 394 return selectDivRem(I, MRI, MF); 395 case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS: 396 return selectIntrinsicWSideEffects(I, MRI, MF); 397 } 398 399 return false; 400 } 401 402 unsigned X86InstructionSelector::getLoadStoreOp(const LLT &Ty, 403 const RegisterBank &RB, 404 unsigned Opc, 405 uint64_t Alignment) const { 406 bool Isload = (Opc == TargetOpcode::G_LOAD); 407 bool HasAVX = STI.hasAVX(); 408 bool HasAVX512 = STI.hasAVX512(); 409 bool HasVLX = STI.hasVLX(); 410 411 if (Ty == LLT::scalar(8)) { 412 if (X86::GPRRegBankID == RB.getID()) 413 return Isload ? X86::MOV8rm : X86::MOV8mr; 414 } else if (Ty == LLT::scalar(16)) { 415 if (X86::GPRRegBankID == RB.getID()) 416 return Isload ? X86::MOV16rm : X86::MOV16mr; 417 } else if (Ty == LLT::scalar(32) || Ty == LLT::pointer(0, 32)) { 418 if (X86::GPRRegBankID == RB.getID()) 419 return Isload ? X86::MOV32rm : X86::MOV32mr; 420 if (X86::VECRRegBankID == RB.getID()) 421 return Isload ? (HasAVX512 ? X86::VMOVSSZrm 422 : HasAVX ? X86::VMOVSSrm : X86::MOVSSrm) 423 : (HasAVX512 ? X86::VMOVSSZmr 424 : HasAVX ? X86::VMOVSSmr : X86::MOVSSmr); 425 } else if (Ty == LLT::scalar(64) || Ty == LLT::pointer(0, 64)) { 426 if (X86::GPRRegBankID == RB.getID()) 427 return Isload ? X86::MOV64rm : X86::MOV64mr; 428 if (X86::VECRRegBankID == RB.getID()) 429 return Isload ? (HasAVX512 ? X86::VMOVSDZrm 430 : HasAVX ? X86::VMOVSDrm : X86::MOVSDrm) 431 : (HasAVX512 ? X86::VMOVSDZmr 432 : HasAVX ? X86::VMOVSDmr : X86::MOVSDmr); 433 } else if (Ty.isVector() && Ty.getSizeInBits() == 128) { 434 if (Alignment >= 16) 435 return Isload ? (HasVLX ? X86::VMOVAPSZ128rm 436 : HasAVX512 437 ? X86::VMOVAPSZ128rm_NOVLX 438 : HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) 439 : (HasVLX ? X86::VMOVAPSZ128mr 440 : HasAVX512 441 ? X86::VMOVAPSZ128mr_NOVLX 442 : HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr); 443 else 444 return Isload ? (HasVLX ? X86::VMOVUPSZ128rm 445 : HasAVX512 446 ? X86::VMOVUPSZ128rm_NOVLX 447 : HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) 448 : (HasVLX ? X86::VMOVUPSZ128mr 449 : HasAVX512 450 ? X86::VMOVUPSZ128mr_NOVLX 451 : HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr); 452 } else if (Ty.isVector() && Ty.getSizeInBits() == 256) { 453 if (Alignment >= 32) 454 return Isload ? (HasVLX ? X86::VMOVAPSZ256rm 455 : HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX 456 : X86::VMOVAPSYrm) 457 : (HasVLX ? X86::VMOVAPSZ256mr 458 : HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX 459 : X86::VMOVAPSYmr); 460 else 461 return Isload ? (HasVLX ? X86::VMOVUPSZ256rm 462 : HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX 463 : X86::VMOVUPSYrm) 464 : (HasVLX ? X86::VMOVUPSZ256mr 465 : HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX 466 : X86::VMOVUPSYmr); 467 } else if (Ty.isVector() && Ty.getSizeInBits() == 512) { 468 if (Alignment >= 64) 469 return Isload ? X86::VMOVAPSZrm : X86::VMOVAPSZmr; 470 else 471 return Isload ? X86::VMOVUPSZrm : X86::VMOVUPSZmr; 472 } 473 return Opc; 474 } 475 476 // Fill in an address from the given instruction. 477 static void X86SelectAddress(const MachineInstr &I, 478 const MachineRegisterInfo &MRI, 479 X86AddressMode &AM) { 480 assert(I.getOperand(0).isReg() && "unsupported opperand."); 481 assert(MRI.getType(I.getOperand(0).getReg()).isPointer() && 482 "unsupported type."); 483 484 if (I.getOpcode() == TargetOpcode::G_GEP) { 485 if (auto COff = getConstantVRegVal(I.getOperand(2).getReg(), MRI)) { 486 int64_t Imm = *COff; 487 if (isInt<32>(Imm)) { // Check for displacement overflow. 488 AM.Disp = static_cast<int32_t>(Imm); 489 AM.Base.Reg = I.getOperand(1).getReg(); 490 return; 491 } 492 } 493 } else if (I.getOpcode() == TargetOpcode::G_FRAME_INDEX) { 494 AM.Base.FrameIndex = I.getOperand(1).getIndex(); 495 AM.BaseType = X86AddressMode::FrameIndexBase; 496 return; 497 } 498 499 // Default behavior. 500 AM.Base.Reg = I.getOperand(0).getReg(); 501 } 502 503 bool X86InstructionSelector::selectLoadStoreOp(MachineInstr &I, 504 MachineRegisterInfo &MRI, 505 MachineFunction &MF) const { 506 unsigned Opc = I.getOpcode(); 507 508 assert((Opc == TargetOpcode::G_STORE || Opc == TargetOpcode::G_LOAD) && 509 "unexpected instruction"); 510 511 const unsigned DefReg = I.getOperand(0).getReg(); 512 LLT Ty = MRI.getType(DefReg); 513 const RegisterBank &RB = *RBI.getRegBank(DefReg, MRI, TRI); 514 515 auto &MemOp = **I.memoperands_begin(); 516 if (MemOp.getOrdering() != AtomicOrdering::NotAtomic) { 517 LLVM_DEBUG(dbgs() << "Atomic load/store not supported yet\n"); 518 return false; 519 } 520 521 unsigned NewOpc = getLoadStoreOp(Ty, RB, Opc, MemOp.getAlignment()); 522 if (NewOpc == Opc) 523 return false; 524 525 X86AddressMode AM; 526 X86SelectAddress(*MRI.getVRegDef(I.getOperand(1).getReg()), MRI, AM); 527 528 I.setDesc(TII.get(NewOpc)); 529 MachineInstrBuilder MIB(MF, I); 530 if (Opc == TargetOpcode::G_LOAD) { 531 I.RemoveOperand(1); 532 addFullAddress(MIB, AM); 533 } else { 534 // G_STORE (VAL, Addr), X86Store instruction (Addr, VAL) 535 I.RemoveOperand(1); 536 I.RemoveOperand(0); 537 addFullAddress(MIB, AM).addUse(DefReg); 538 } 539 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 540 } 541 542 static unsigned getLeaOP(LLT Ty, const X86Subtarget &STI) { 543 if (Ty == LLT::pointer(0, 64)) 544 return X86::LEA64r; 545 else if (Ty == LLT::pointer(0, 32)) 546 return STI.isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r; 547 else 548 llvm_unreachable("Can't get LEA opcode. Unsupported type."); 549 } 550 551 bool X86InstructionSelector::selectFrameIndexOrGep(MachineInstr &I, 552 MachineRegisterInfo &MRI, 553 MachineFunction &MF) const { 554 unsigned Opc = I.getOpcode(); 555 556 assert((Opc == TargetOpcode::G_FRAME_INDEX || Opc == TargetOpcode::G_GEP) && 557 "unexpected instruction"); 558 559 const unsigned DefReg = I.getOperand(0).getReg(); 560 LLT Ty = MRI.getType(DefReg); 561 562 // Use LEA to calculate frame index and GEP 563 unsigned NewOpc = getLeaOP(Ty, STI); 564 I.setDesc(TII.get(NewOpc)); 565 MachineInstrBuilder MIB(MF, I); 566 567 if (Opc == TargetOpcode::G_FRAME_INDEX) { 568 addOffset(MIB, 0); 569 } else { 570 MachineOperand &InxOp = I.getOperand(2); 571 I.addOperand(InxOp); // set IndexReg 572 InxOp.ChangeToImmediate(1); // set Scale 573 MIB.addImm(0).addReg(0); 574 } 575 576 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 577 } 578 579 bool X86InstructionSelector::selectGlobalValue(MachineInstr &I, 580 MachineRegisterInfo &MRI, 581 MachineFunction &MF) const { 582 assert((I.getOpcode() == TargetOpcode::G_GLOBAL_VALUE) && 583 "unexpected instruction"); 584 585 auto GV = I.getOperand(1).getGlobal(); 586 if (GV->isThreadLocal()) { 587 return false; // TODO: we don't support TLS yet. 588 } 589 590 // Can't handle alternate code models yet. 591 if (TM.getCodeModel() != CodeModel::Small) 592 return false; 593 594 X86AddressMode AM; 595 AM.GV = GV; 596 AM.GVOpFlags = STI.classifyGlobalReference(GV); 597 598 // TODO: The ABI requires an extra load. not supported yet. 599 if (isGlobalStubReference(AM.GVOpFlags)) 600 return false; 601 602 // TODO: This reference is relative to the pic base. not supported yet. 603 if (isGlobalRelativeToPICBase(AM.GVOpFlags)) 604 return false; 605 606 if (STI.isPICStyleRIPRel()) { 607 // Use rip-relative addressing. 608 assert(AM.Base.Reg == 0 && AM.IndexReg == 0); 609 AM.Base.Reg = X86::RIP; 610 } 611 612 const unsigned DefReg = I.getOperand(0).getReg(); 613 LLT Ty = MRI.getType(DefReg); 614 unsigned NewOpc = getLeaOP(Ty, STI); 615 616 I.setDesc(TII.get(NewOpc)); 617 MachineInstrBuilder MIB(MF, I); 618 619 I.RemoveOperand(1); 620 addFullAddress(MIB, AM); 621 622 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 623 } 624 625 bool X86InstructionSelector::selectConstant(MachineInstr &I, 626 MachineRegisterInfo &MRI, 627 MachineFunction &MF) const { 628 assert((I.getOpcode() == TargetOpcode::G_CONSTANT) && 629 "unexpected instruction"); 630 631 const unsigned DefReg = I.getOperand(0).getReg(); 632 LLT Ty = MRI.getType(DefReg); 633 634 if (RBI.getRegBank(DefReg, MRI, TRI)->getID() != X86::GPRRegBankID) 635 return false; 636 637 uint64_t Val = 0; 638 if (I.getOperand(1).isCImm()) { 639 Val = I.getOperand(1).getCImm()->getZExtValue(); 640 I.getOperand(1).ChangeToImmediate(Val); 641 } else if (I.getOperand(1).isImm()) { 642 Val = I.getOperand(1).getImm(); 643 } else 644 llvm_unreachable("Unsupported operand type."); 645 646 unsigned NewOpc; 647 switch (Ty.getSizeInBits()) { 648 case 8: 649 NewOpc = X86::MOV8ri; 650 break; 651 case 16: 652 NewOpc = X86::MOV16ri; 653 break; 654 case 32: 655 NewOpc = X86::MOV32ri; 656 break; 657 case 64: 658 // TODO: in case isUInt<32>(Val), X86::MOV32ri can be used 659 if (isInt<32>(Val)) 660 NewOpc = X86::MOV64ri32; 661 else 662 NewOpc = X86::MOV64ri; 663 break; 664 default: 665 llvm_unreachable("Can't select G_CONSTANT, unsupported type."); 666 } 667 668 I.setDesc(TII.get(NewOpc)); 669 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 670 } 671 672 // Helper function for selectTruncOrPtrToInt and selectAnyext. 673 // Returns true if DstRC lives on a floating register class and 674 // SrcRC lives on a 128-bit vector class. 675 static bool canTurnIntoCOPY(const TargetRegisterClass *DstRC, 676 const TargetRegisterClass *SrcRC) { 677 return (DstRC == &X86::FR32RegClass || DstRC == &X86::FR32XRegClass || 678 DstRC == &X86::FR64RegClass || DstRC == &X86::FR64XRegClass) && 679 (SrcRC == &X86::VR128RegClass || SrcRC == &X86::VR128XRegClass); 680 } 681 682 bool X86InstructionSelector::selectTurnIntoCOPY( 683 MachineInstr &I, MachineRegisterInfo &MRI, const unsigned DstReg, 684 const TargetRegisterClass *DstRC, const unsigned SrcReg, 685 const TargetRegisterClass *SrcRC) const { 686 687 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 688 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 689 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 690 << " operand\n"); 691 return false; 692 } 693 I.setDesc(TII.get(X86::COPY)); 694 return true; 695 } 696 697 bool X86InstructionSelector::selectTruncOrPtrToInt(MachineInstr &I, 698 MachineRegisterInfo &MRI, 699 MachineFunction &MF) const { 700 assert((I.getOpcode() == TargetOpcode::G_TRUNC || 701 I.getOpcode() == TargetOpcode::G_PTRTOINT) && 702 "unexpected instruction"); 703 704 const unsigned DstReg = I.getOperand(0).getReg(); 705 const unsigned SrcReg = I.getOperand(1).getReg(); 706 707 const LLT DstTy = MRI.getType(DstReg); 708 const LLT SrcTy = MRI.getType(SrcReg); 709 710 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI); 711 const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI); 712 713 if (DstRB.getID() != SrcRB.getID()) { 714 LLVM_DEBUG(dbgs() << TII.getName(I.getOpcode()) 715 << " input/output on different banks\n"); 716 return false; 717 } 718 719 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB); 720 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB); 721 722 if (!DstRC || !SrcRC) 723 return false; 724 725 // If that's truncation of the value that lives on the vector class and goes 726 // into the floating class, just replace it with copy, as we are able to 727 // select it as a regular move. 728 if (canTurnIntoCOPY(DstRC, SrcRC)) 729 return selectTurnIntoCOPY(I, MRI, DstReg, DstRC, SrcReg, SrcRC); 730 731 if (DstRB.getID() != X86::GPRRegBankID) 732 return false; 733 734 unsigned SubIdx; 735 if (DstRC == SrcRC) { 736 // Nothing to be done 737 SubIdx = X86::NoSubRegister; 738 } else if (DstRC == &X86::GR32RegClass) { 739 SubIdx = X86::sub_32bit; 740 } else if (DstRC == &X86::GR16RegClass) { 741 SubIdx = X86::sub_16bit; 742 } else if (DstRC == &X86::GR8RegClass) { 743 SubIdx = X86::sub_8bit; 744 } else { 745 return false; 746 } 747 748 SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubIdx); 749 750 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 751 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 752 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 753 << "\n"); 754 return false; 755 } 756 757 I.getOperand(1).setSubReg(SubIdx); 758 759 I.setDesc(TII.get(X86::COPY)); 760 return true; 761 } 762 763 bool X86InstructionSelector::selectZext(MachineInstr &I, 764 MachineRegisterInfo &MRI, 765 MachineFunction &MF) const { 766 assert((I.getOpcode() == TargetOpcode::G_ZEXT) && "unexpected instruction"); 767 768 const unsigned DstReg = I.getOperand(0).getReg(); 769 const unsigned SrcReg = I.getOperand(1).getReg(); 770 771 const LLT DstTy = MRI.getType(DstReg); 772 const LLT SrcTy = MRI.getType(SrcReg); 773 774 assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(32)) && 775 "8=>32 Zext is handled by tablegen"); 776 assert(!(SrcTy == LLT::scalar(16) && DstTy == LLT::scalar(32)) && 777 "16=>32 Zext is handled by tablegen"); 778 779 const static struct ZextEntry { 780 LLT SrcTy; 781 LLT DstTy; 782 unsigned MovOp; 783 bool NeedSubregToReg; 784 } OpTable[] = { 785 {LLT::scalar(8), LLT::scalar(16), X86::MOVZX16rr8, false}, // i8 => i16 786 {LLT::scalar(8), LLT::scalar(64), X86::MOVZX32rr8, true}, // i8 => i64 787 {LLT::scalar(16), LLT::scalar(64), X86::MOVZX32rr16, true}, // i16 => i64 788 {LLT::scalar(32), LLT::scalar(64), 0, true} // i32 => i64 789 }; 790 791 auto ZextEntryIt = 792 std::find_if(std::begin(OpTable), std::end(OpTable), 793 [SrcTy, DstTy](const ZextEntry &El) { 794 return El.DstTy == DstTy && El.SrcTy == SrcTy; 795 }); 796 797 // Here we try to select Zext into a MOVZ and/or SUBREG_TO_REG instruction. 798 if (ZextEntryIt != std::end(OpTable)) { 799 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI); 800 const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI); 801 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB); 802 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB); 803 804 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 805 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 806 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 807 << " operand\n"); 808 return false; 809 } 810 811 unsigned TransitRegTo = DstReg; 812 unsigned TransitRegFrom = SrcReg; 813 if (ZextEntryIt->MovOp) { 814 // If we select Zext into MOVZ + SUBREG_TO_REG, we need to have 815 // a transit register in between: create it here. 816 if (ZextEntryIt->NeedSubregToReg) { 817 TransitRegFrom = MRI.createVirtualRegister( 818 getRegClass(LLT::scalar(32), DstReg, MRI)); 819 TransitRegTo = TransitRegFrom; 820 } 821 822 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(ZextEntryIt->MovOp)) 823 .addDef(TransitRegTo) 824 .addReg(SrcReg); 825 } 826 if (ZextEntryIt->NeedSubregToReg) { 827 BuildMI(*I.getParent(), I, I.getDebugLoc(), 828 TII.get(TargetOpcode::SUBREG_TO_REG)) 829 .addDef(DstReg) 830 .addImm(0) 831 .addReg(TransitRegFrom) 832 .addImm(X86::sub_32bit); 833 } 834 I.eraseFromParent(); 835 return true; 836 } 837 838 if (SrcTy != LLT::scalar(1)) 839 return false; 840 841 unsigned AndOpc; 842 if (DstTy == LLT::scalar(8)) 843 AndOpc = X86::AND8ri; 844 else if (DstTy == LLT::scalar(16)) 845 AndOpc = X86::AND16ri8; 846 else if (DstTy == LLT::scalar(32)) 847 AndOpc = X86::AND32ri8; 848 else if (DstTy == LLT::scalar(64)) 849 AndOpc = X86::AND64ri8; 850 else 851 return false; 852 853 unsigned DefReg = SrcReg; 854 if (DstTy != LLT::scalar(8)) { 855 DefReg = MRI.createVirtualRegister(getRegClass(DstTy, DstReg, MRI)); 856 BuildMI(*I.getParent(), I, I.getDebugLoc(), 857 TII.get(TargetOpcode::SUBREG_TO_REG), DefReg) 858 .addImm(0) 859 .addReg(SrcReg) 860 .addImm(X86::sub_8bit); 861 } 862 863 MachineInstr &AndInst = 864 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(AndOpc), DstReg) 865 .addReg(DefReg) 866 .addImm(1); 867 868 constrainSelectedInstRegOperands(AndInst, TII, TRI, RBI); 869 870 I.eraseFromParent(); 871 return true; 872 } 873 874 bool X86InstructionSelector::selectAnyext(MachineInstr &I, 875 MachineRegisterInfo &MRI, 876 MachineFunction &MF) const { 877 assert((I.getOpcode() == TargetOpcode::G_ANYEXT) && "unexpected instruction"); 878 879 const unsigned DstReg = I.getOperand(0).getReg(); 880 const unsigned SrcReg = I.getOperand(1).getReg(); 881 882 const LLT DstTy = MRI.getType(DstReg); 883 const LLT SrcTy = MRI.getType(SrcReg); 884 885 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI); 886 const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI); 887 888 assert(DstRB.getID() == SrcRB.getID() && 889 "G_ANYEXT input/output on different banks\n"); 890 891 assert(DstTy.getSizeInBits() > SrcTy.getSizeInBits() && 892 "G_ANYEXT incorrect operand size"); 893 894 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB); 895 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB); 896 897 // If that's ANY_EXT of the value that lives on the floating class and goes 898 // into the vector class, just replace it with copy, as we are able to select 899 // it as a regular move. 900 if (canTurnIntoCOPY(SrcRC, DstRC)) 901 return selectTurnIntoCOPY(I, MRI, SrcReg, SrcRC, DstReg, DstRC); 902 903 if (DstRB.getID() != X86::GPRRegBankID) 904 return false; 905 906 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 907 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 908 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 909 << " operand\n"); 910 return false; 911 } 912 913 if (SrcRC == DstRC) { 914 I.setDesc(TII.get(X86::COPY)); 915 return true; 916 } 917 918 BuildMI(*I.getParent(), I, I.getDebugLoc(), 919 TII.get(TargetOpcode::SUBREG_TO_REG)) 920 .addDef(DstReg) 921 .addImm(0) 922 .addReg(SrcReg) 923 .addImm(getSubRegIndex(SrcRC)); 924 925 I.eraseFromParent(); 926 return true; 927 } 928 929 bool X86InstructionSelector::selectCmp(MachineInstr &I, 930 MachineRegisterInfo &MRI, 931 MachineFunction &MF) const { 932 assert((I.getOpcode() == TargetOpcode::G_ICMP) && "unexpected instruction"); 933 934 X86::CondCode CC; 935 bool SwapArgs; 936 std::tie(CC, SwapArgs) = X86::getX86ConditionCode( 937 (CmpInst::Predicate)I.getOperand(1).getPredicate()); 938 unsigned OpSet = X86::getSETFromCond(CC); 939 940 unsigned LHS = I.getOperand(2).getReg(); 941 unsigned RHS = I.getOperand(3).getReg(); 942 943 if (SwapArgs) 944 std::swap(LHS, RHS); 945 946 unsigned OpCmp; 947 LLT Ty = MRI.getType(LHS); 948 949 switch (Ty.getSizeInBits()) { 950 default: 951 return false; 952 case 8: 953 OpCmp = X86::CMP8rr; 954 break; 955 case 16: 956 OpCmp = X86::CMP16rr; 957 break; 958 case 32: 959 OpCmp = X86::CMP32rr; 960 break; 961 case 64: 962 OpCmp = X86::CMP64rr; 963 break; 964 } 965 966 MachineInstr &CmpInst = 967 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp)) 968 .addReg(LHS) 969 .addReg(RHS); 970 971 MachineInstr &SetInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 972 TII.get(OpSet), I.getOperand(0).getReg()); 973 974 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI); 975 constrainSelectedInstRegOperands(SetInst, TII, TRI, RBI); 976 977 I.eraseFromParent(); 978 return true; 979 } 980 981 bool X86InstructionSelector::selectFCmp(MachineInstr &I, 982 MachineRegisterInfo &MRI, 983 MachineFunction &MF) const { 984 assert((I.getOpcode() == TargetOpcode::G_FCMP) && "unexpected instruction"); 985 986 unsigned LhsReg = I.getOperand(2).getReg(); 987 unsigned RhsReg = I.getOperand(3).getReg(); 988 CmpInst::Predicate Predicate = 989 (CmpInst::Predicate)I.getOperand(1).getPredicate(); 990 991 // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction. 992 static const uint16_t SETFOpcTable[2][3] = { 993 {X86::SETEr, X86::SETNPr, X86::AND8rr}, 994 {X86::SETNEr, X86::SETPr, X86::OR8rr}}; 995 const uint16_t *SETFOpc = nullptr; 996 switch (Predicate) { 997 default: 998 break; 999 case CmpInst::FCMP_OEQ: 1000 SETFOpc = &SETFOpcTable[0][0]; 1001 break; 1002 case CmpInst::FCMP_UNE: 1003 SETFOpc = &SETFOpcTable[1][0]; 1004 break; 1005 } 1006 1007 // Compute the opcode for the CMP instruction. 1008 unsigned OpCmp; 1009 LLT Ty = MRI.getType(LhsReg); 1010 switch (Ty.getSizeInBits()) { 1011 default: 1012 return false; 1013 case 32: 1014 OpCmp = X86::UCOMISSrr; 1015 break; 1016 case 64: 1017 OpCmp = X86::UCOMISDrr; 1018 break; 1019 } 1020 1021 unsigned ResultReg = I.getOperand(0).getReg(); 1022 RBI.constrainGenericRegister( 1023 ResultReg, 1024 *getRegClass(LLT::scalar(8), *RBI.getRegBank(ResultReg, MRI, TRI)), MRI); 1025 if (SETFOpc) { 1026 MachineInstr &CmpInst = 1027 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp)) 1028 .addReg(LhsReg) 1029 .addReg(RhsReg); 1030 1031 unsigned FlagReg1 = MRI.createVirtualRegister(&X86::GR8RegClass); 1032 unsigned FlagReg2 = MRI.createVirtualRegister(&X86::GR8RegClass); 1033 MachineInstr &Set1 = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1034 TII.get(SETFOpc[0]), FlagReg1); 1035 MachineInstr &Set2 = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1036 TII.get(SETFOpc[1]), FlagReg2); 1037 MachineInstr &Set3 = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1038 TII.get(SETFOpc[2]), ResultReg) 1039 .addReg(FlagReg1) 1040 .addReg(FlagReg2); 1041 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI); 1042 constrainSelectedInstRegOperands(Set1, TII, TRI, RBI); 1043 constrainSelectedInstRegOperands(Set2, TII, TRI, RBI); 1044 constrainSelectedInstRegOperands(Set3, TII, TRI, RBI); 1045 1046 I.eraseFromParent(); 1047 return true; 1048 } 1049 1050 X86::CondCode CC; 1051 bool SwapArgs; 1052 std::tie(CC, SwapArgs) = X86::getX86ConditionCode(Predicate); 1053 assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code."); 1054 unsigned Opc = X86::getSETFromCond(CC); 1055 1056 if (SwapArgs) 1057 std::swap(LhsReg, RhsReg); 1058 1059 // Emit a compare of LHS/RHS. 1060 MachineInstr &CmpInst = 1061 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp)) 1062 .addReg(LhsReg) 1063 .addReg(RhsReg); 1064 1065 MachineInstr &Set = 1066 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opc), ResultReg); 1067 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI); 1068 constrainSelectedInstRegOperands(Set, TII, TRI, RBI); 1069 I.eraseFromParent(); 1070 return true; 1071 } 1072 1073 bool X86InstructionSelector::selectUadde(MachineInstr &I, 1074 MachineRegisterInfo &MRI, 1075 MachineFunction &MF) const { 1076 assert((I.getOpcode() == TargetOpcode::G_UADDE) && "unexpected instruction"); 1077 1078 const unsigned DstReg = I.getOperand(0).getReg(); 1079 const unsigned CarryOutReg = I.getOperand(1).getReg(); 1080 const unsigned Op0Reg = I.getOperand(2).getReg(); 1081 const unsigned Op1Reg = I.getOperand(3).getReg(); 1082 unsigned CarryInReg = I.getOperand(4).getReg(); 1083 1084 const LLT DstTy = MRI.getType(DstReg); 1085 1086 if (DstTy != LLT::scalar(32)) 1087 return false; 1088 1089 // find CarryIn def instruction. 1090 MachineInstr *Def = MRI.getVRegDef(CarryInReg); 1091 while (Def->getOpcode() == TargetOpcode::G_TRUNC) { 1092 CarryInReg = Def->getOperand(1).getReg(); 1093 Def = MRI.getVRegDef(CarryInReg); 1094 } 1095 1096 unsigned Opcode; 1097 if (Def->getOpcode() == TargetOpcode::G_UADDE) { 1098 // carry set by prev ADD. 1099 1100 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), X86::EFLAGS) 1101 .addReg(CarryInReg); 1102 1103 if (!RBI.constrainGenericRegister(CarryInReg, X86::GR32RegClass, MRI)) 1104 return false; 1105 1106 Opcode = X86::ADC32rr; 1107 } else if (auto val = getConstantVRegVal(CarryInReg, MRI)) { 1108 // carry is constant, support only 0. 1109 if (*val != 0) 1110 return false; 1111 1112 Opcode = X86::ADD32rr; 1113 } else 1114 return false; 1115 1116 MachineInstr &AddInst = 1117 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode), DstReg) 1118 .addReg(Op0Reg) 1119 .addReg(Op1Reg); 1120 1121 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), CarryOutReg) 1122 .addReg(X86::EFLAGS); 1123 1124 if (!constrainSelectedInstRegOperands(AddInst, TII, TRI, RBI) || 1125 !RBI.constrainGenericRegister(CarryOutReg, X86::GR32RegClass, MRI)) 1126 return false; 1127 1128 I.eraseFromParent(); 1129 return true; 1130 } 1131 1132 bool X86InstructionSelector::selectExtract(MachineInstr &I, 1133 MachineRegisterInfo &MRI, 1134 MachineFunction &MF) const { 1135 assert((I.getOpcode() == TargetOpcode::G_EXTRACT) && 1136 "unexpected instruction"); 1137 1138 const unsigned DstReg = I.getOperand(0).getReg(); 1139 const unsigned SrcReg = I.getOperand(1).getReg(); 1140 int64_t Index = I.getOperand(2).getImm(); 1141 1142 const LLT DstTy = MRI.getType(DstReg); 1143 const LLT SrcTy = MRI.getType(SrcReg); 1144 1145 // Meanwile handle vector type only. 1146 if (!DstTy.isVector()) 1147 return false; 1148 1149 if (Index % DstTy.getSizeInBits() != 0) 1150 return false; // Not extract subvector. 1151 1152 if (Index == 0) { 1153 // Replace by extract subreg copy. 1154 if (!emitExtractSubreg(DstReg, SrcReg, I, MRI, MF)) 1155 return false; 1156 1157 I.eraseFromParent(); 1158 return true; 1159 } 1160 1161 bool HasAVX = STI.hasAVX(); 1162 bool HasAVX512 = STI.hasAVX512(); 1163 bool HasVLX = STI.hasVLX(); 1164 1165 if (SrcTy.getSizeInBits() == 256 && DstTy.getSizeInBits() == 128) { 1166 if (HasVLX) 1167 I.setDesc(TII.get(X86::VEXTRACTF32x4Z256rr)); 1168 else if (HasAVX) 1169 I.setDesc(TII.get(X86::VEXTRACTF128rr)); 1170 else 1171 return false; 1172 } else if (SrcTy.getSizeInBits() == 512 && HasAVX512) { 1173 if (DstTy.getSizeInBits() == 128) 1174 I.setDesc(TII.get(X86::VEXTRACTF32x4Zrr)); 1175 else if (DstTy.getSizeInBits() == 256) 1176 I.setDesc(TII.get(X86::VEXTRACTF64x4Zrr)); 1177 else 1178 return false; 1179 } else 1180 return false; 1181 1182 // Convert to X86 VEXTRACT immediate. 1183 Index = Index / DstTy.getSizeInBits(); 1184 I.getOperand(2).setImm(Index); 1185 1186 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 1187 } 1188 1189 bool X86InstructionSelector::emitExtractSubreg(unsigned DstReg, unsigned SrcReg, 1190 MachineInstr &I, 1191 MachineRegisterInfo &MRI, 1192 MachineFunction &MF) const { 1193 const LLT DstTy = MRI.getType(DstReg); 1194 const LLT SrcTy = MRI.getType(SrcReg); 1195 unsigned SubIdx = X86::NoSubRegister; 1196 1197 if (!DstTy.isVector() || !SrcTy.isVector()) 1198 return false; 1199 1200 assert(SrcTy.getSizeInBits() > DstTy.getSizeInBits() && 1201 "Incorrect Src/Dst register size"); 1202 1203 if (DstTy.getSizeInBits() == 128) 1204 SubIdx = X86::sub_xmm; 1205 else if (DstTy.getSizeInBits() == 256) 1206 SubIdx = X86::sub_ymm; 1207 else 1208 return false; 1209 1210 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstReg, MRI); 1211 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcReg, MRI); 1212 1213 SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubIdx); 1214 1215 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 1216 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 1217 LLVM_DEBUG(dbgs() << "Failed to constrain G_TRUNC\n"); 1218 return false; 1219 } 1220 1221 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), DstReg) 1222 .addReg(SrcReg, 0, SubIdx); 1223 1224 return true; 1225 } 1226 1227 bool X86InstructionSelector::emitInsertSubreg(unsigned DstReg, unsigned SrcReg, 1228 MachineInstr &I, 1229 MachineRegisterInfo &MRI, 1230 MachineFunction &MF) const { 1231 const LLT DstTy = MRI.getType(DstReg); 1232 const LLT SrcTy = MRI.getType(SrcReg); 1233 unsigned SubIdx = X86::NoSubRegister; 1234 1235 // TODO: support scalar types 1236 if (!DstTy.isVector() || !SrcTy.isVector()) 1237 return false; 1238 1239 assert(SrcTy.getSizeInBits() < DstTy.getSizeInBits() && 1240 "Incorrect Src/Dst register size"); 1241 1242 if (SrcTy.getSizeInBits() == 128) 1243 SubIdx = X86::sub_xmm; 1244 else if (SrcTy.getSizeInBits() == 256) 1245 SubIdx = X86::sub_ymm; 1246 else 1247 return false; 1248 1249 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcReg, MRI); 1250 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstReg, MRI); 1251 1252 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 1253 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 1254 LLVM_DEBUG(dbgs() << "Failed to constrain INSERT_SUBREG\n"); 1255 return false; 1256 } 1257 1258 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY)) 1259 .addReg(DstReg, RegState::DefineNoRead, SubIdx) 1260 .addReg(SrcReg); 1261 1262 return true; 1263 } 1264 1265 bool X86InstructionSelector::selectInsert(MachineInstr &I, 1266 MachineRegisterInfo &MRI, 1267 MachineFunction &MF) const { 1268 assert((I.getOpcode() == TargetOpcode::G_INSERT) && "unexpected instruction"); 1269 1270 const unsigned DstReg = I.getOperand(0).getReg(); 1271 const unsigned SrcReg = I.getOperand(1).getReg(); 1272 const unsigned InsertReg = I.getOperand(2).getReg(); 1273 int64_t Index = I.getOperand(3).getImm(); 1274 1275 const LLT DstTy = MRI.getType(DstReg); 1276 const LLT InsertRegTy = MRI.getType(InsertReg); 1277 1278 // Meanwile handle vector type only. 1279 if (!DstTy.isVector()) 1280 return false; 1281 1282 if (Index % InsertRegTy.getSizeInBits() != 0) 1283 return false; // Not insert subvector. 1284 1285 if (Index == 0 && MRI.getVRegDef(SrcReg)->isImplicitDef()) { 1286 // Replace by subreg copy. 1287 if (!emitInsertSubreg(DstReg, InsertReg, I, MRI, MF)) 1288 return false; 1289 1290 I.eraseFromParent(); 1291 return true; 1292 } 1293 1294 bool HasAVX = STI.hasAVX(); 1295 bool HasAVX512 = STI.hasAVX512(); 1296 bool HasVLX = STI.hasVLX(); 1297 1298 if (DstTy.getSizeInBits() == 256 && InsertRegTy.getSizeInBits() == 128) { 1299 if (HasVLX) 1300 I.setDesc(TII.get(X86::VINSERTF32x4Z256rr)); 1301 else if (HasAVX) 1302 I.setDesc(TII.get(X86::VINSERTF128rr)); 1303 else 1304 return false; 1305 } else if (DstTy.getSizeInBits() == 512 && HasAVX512) { 1306 if (InsertRegTy.getSizeInBits() == 128) 1307 I.setDesc(TII.get(X86::VINSERTF32x4Zrr)); 1308 else if (InsertRegTy.getSizeInBits() == 256) 1309 I.setDesc(TII.get(X86::VINSERTF64x4Zrr)); 1310 else 1311 return false; 1312 } else 1313 return false; 1314 1315 // Convert to X86 VINSERT immediate. 1316 Index = Index / InsertRegTy.getSizeInBits(); 1317 1318 I.getOperand(3).setImm(Index); 1319 1320 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 1321 } 1322 1323 bool X86InstructionSelector::selectUnmergeValues( 1324 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF, 1325 CodeGenCoverage &CoverageInfo) const { 1326 assert((I.getOpcode() == TargetOpcode::G_UNMERGE_VALUES) && 1327 "unexpected instruction"); 1328 1329 // Split to extracts. 1330 unsigned NumDefs = I.getNumOperands() - 1; 1331 unsigned SrcReg = I.getOperand(NumDefs).getReg(); 1332 unsigned DefSize = MRI.getType(I.getOperand(0).getReg()).getSizeInBits(); 1333 1334 for (unsigned Idx = 0; Idx < NumDefs; ++Idx) { 1335 MachineInstr &ExtrInst = 1336 *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1337 TII.get(TargetOpcode::G_EXTRACT), I.getOperand(Idx).getReg()) 1338 .addReg(SrcReg) 1339 .addImm(Idx * DefSize); 1340 1341 if (!select(ExtrInst, CoverageInfo)) 1342 return false; 1343 } 1344 1345 I.eraseFromParent(); 1346 return true; 1347 } 1348 1349 bool X86InstructionSelector::selectMergeValues( 1350 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF, 1351 CodeGenCoverage &CoverageInfo) const { 1352 assert((I.getOpcode() == TargetOpcode::G_MERGE_VALUES) && 1353 "unexpected instruction"); 1354 1355 // Split to inserts. 1356 unsigned DstReg = I.getOperand(0).getReg(); 1357 unsigned SrcReg0 = I.getOperand(1).getReg(); 1358 1359 const LLT DstTy = MRI.getType(DstReg); 1360 const LLT SrcTy = MRI.getType(SrcReg0); 1361 unsigned SrcSize = SrcTy.getSizeInBits(); 1362 1363 const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI); 1364 1365 // For the first src use insertSubReg. 1366 unsigned DefReg = MRI.createGenericVirtualRegister(DstTy); 1367 MRI.setRegBank(DefReg, RegBank); 1368 if (!emitInsertSubreg(DefReg, I.getOperand(1).getReg(), I, MRI, MF)) 1369 return false; 1370 1371 for (unsigned Idx = 2; Idx < I.getNumOperands(); ++Idx) { 1372 unsigned Tmp = MRI.createGenericVirtualRegister(DstTy); 1373 MRI.setRegBank(Tmp, RegBank); 1374 1375 MachineInstr &InsertInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1376 TII.get(TargetOpcode::G_INSERT), Tmp) 1377 .addReg(DefReg) 1378 .addReg(I.getOperand(Idx).getReg()) 1379 .addImm((Idx - 1) * SrcSize); 1380 1381 DefReg = Tmp; 1382 1383 if (!select(InsertInst, CoverageInfo)) 1384 return false; 1385 } 1386 1387 MachineInstr &CopyInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1388 TII.get(TargetOpcode::COPY), DstReg) 1389 .addReg(DefReg); 1390 1391 if (!select(CopyInst, CoverageInfo)) 1392 return false; 1393 1394 I.eraseFromParent(); 1395 return true; 1396 } 1397 1398 bool X86InstructionSelector::selectCondBranch(MachineInstr &I, 1399 MachineRegisterInfo &MRI, 1400 MachineFunction &MF) const { 1401 assert((I.getOpcode() == TargetOpcode::G_BRCOND) && "unexpected instruction"); 1402 1403 const unsigned CondReg = I.getOperand(0).getReg(); 1404 MachineBasicBlock *DestMBB = I.getOperand(1).getMBB(); 1405 1406 MachineInstr &TestInst = 1407 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::TEST8ri)) 1408 .addReg(CondReg) 1409 .addImm(1); 1410 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::JNE_1)) 1411 .addMBB(DestMBB); 1412 1413 constrainSelectedInstRegOperands(TestInst, TII, TRI, RBI); 1414 1415 I.eraseFromParent(); 1416 return true; 1417 } 1418 1419 bool X86InstructionSelector::materializeFP(MachineInstr &I, 1420 MachineRegisterInfo &MRI, 1421 MachineFunction &MF) const { 1422 assert((I.getOpcode() == TargetOpcode::G_FCONSTANT) && 1423 "unexpected instruction"); 1424 1425 // Can't handle alternate code models yet. 1426 CodeModel::Model CM = TM.getCodeModel(); 1427 if (CM != CodeModel::Small && CM != CodeModel::Large) 1428 return false; 1429 1430 const unsigned DstReg = I.getOperand(0).getReg(); 1431 const LLT DstTy = MRI.getType(DstReg); 1432 const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI); 1433 unsigned Align = DstTy.getSizeInBits(); 1434 const DebugLoc &DbgLoc = I.getDebugLoc(); 1435 1436 unsigned Opc = getLoadStoreOp(DstTy, RegBank, TargetOpcode::G_LOAD, Align); 1437 1438 // Create the load from the constant pool. 1439 const ConstantFP *CFP = I.getOperand(1).getFPImm(); 1440 unsigned CPI = MF.getConstantPool()->getConstantPoolIndex(CFP, Align); 1441 MachineInstr *LoadInst = nullptr; 1442 unsigned char OpFlag = STI.classifyLocalReference(nullptr); 1443 1444 if (CM == CodeModel::Large && STI.is64Bit()) { 1445 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so 1446 // they cannot be folded into immediate fields. 1447 1448 unsigned AddrReg = MRI.createVirtualRegister(&X86::GR64RegClass); 1449 BuildMI(*I.getParent(), I, DbgLoc, TII.get(X86::MOV64ri), AddrReg) 1450 .addConstantPoolIndex(CPI, 0, OpFlag); 1451 1452 MachineMemOperand *MMO = MF.getMachineMemOperand( 1453 MachinePointerInfo::getConstantPool(MF), MachineMemOperand::MOLoad, 1454 MF.getDataLayout().getPointerSize(), Align); 1455 1456 LoadInst = 1457 addDirectMem(BuildMI(*I.getParent(), I, DbgLoc, TII.get(Opc), DstReg), 1458 AddrReg) 1459 .addMemOperand(MMO); 1460 1461 } else if (CM == CodeModel::Small || !STI.is64Bit()) { 1462 // Handle the case when globals fit in our immediate field. 1463 // This is true for X86-32 always and X86-64 when in -mcmodel=small mode. 1464 1465 // x86-32 PIC requires a PIC base register for constant pools. 1466 unsigned PICBase = 0; 1467 if (OpFlag == X86II::MO_PIC_BASE_OFFSET || OpFlag == X86II::MO_GOTOFF) { 1468 // PICBase can be allocated by TII.getGlobalBaseReg(&MF). 1469 // In DAGISEL the code that initialize it generated by the CGBR pass. 1470 return false; // TODO support the mode. 1471 } else if (STI.is64Bit() && TM.getCodeModel() == CodeModel::Small) 1472 PICBase = X86::RIP; 1473 1474 LoadInst = addConstantPoolReference( 1475 BuildMI(*I.getParent(), I, DbgLoc, TII.get(Opc), DstReg), CPI, PICBase, 1476 OpFlag); 1477 } else 1478 return false; 1479 1480 constrainSelectedInstRegOperands(*LoadInst, TII, TRI, RBI); 1481 I.eraseFromParent(); 1482 return true; 1483 } 1484 1485 bool X86InstructionSelector::selectImplicitDefOrPHI( 1486 MachineInstr &I, MachineRegisterInfo &MRI) const { 1487 assert((I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF || 1488 I.getOpcode() == TargetOpcode::G_PHI) && 1489 "unexpected instruction"); 1490 1491 unsigned DstReg = I.getOperand(0).getReg(); 1492 1493 if (!MRI.getRegClassOrNull(DstReg)) { 1494 const LLT DstTy = MRI.getType(DstReg); 1495 const TargetRegisterClass *RC = getRegClass(DstTy, DstReg, MRI); 1496 1497 if (!RBI.constrainGenericRegister(DstReg, *RC, MRI)) { 1498 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 1499 << " operand\n"); 1500 return false; 1501 } 1502 } 1503 1504 if (I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF) 1505 I.setDesc(TII.get(X86::IMPLICIT_DEF)); 1506 else 1507 I.setDesc(TII.get(X86::PHI)); 1508 1509 return true; 1510 } 1511 1512 // Currently GlobalIsel TableGen generates patterns for shift imm and shift 1, 1513 // but with shiftCount i8. In G_LSHR/G_ASHR/G_SHL like LLVM-IR both arguments 1514 // has the same type, so for now only shift i8 can use auto generated 1515 // TableGen patterns. 1516 bool X86InstructionSelector::selectShift(MachineInstr &I, 1517 MachineRegisterInfo &MRI, 1518 MachineFunction &MF) const { 1519 1520 assert((I.getOpcode() == TargetOpcode::G_SHL || 1521 I.getOpcode() == TargetOpcode::G_ASHR || 1522 I.getOpcode() == TargetOpcode::G_LSHR) && 1523 "unexpected instruction"); 1524 1525 unsigned DstReg = I.getOperand(0).getReg(); 1526 const LLT DstTy = MRI.getType(DstReg); 1527 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI); 1528 1529 const static struct ShiftEntry { 1530 unsigned SizeInBits; 1531 unsigned CReg; 1532 unsigned OpLSHR; 1533 unsigned OpASHR; 1534 unsigned OpSHL; 1535 } OpTable[] = { 1536 {8, X86::CL, X86::SHR8rCL, X86::SAR8rCL, X86::SHL8rCL}, // i8 1537 {16, X86::CX, X86::SHR16rCL, X86::SAR16rCL, X86::SHL16rCL}, // i16 1538 {32, X86::ECX, X86::SHR32rCL, X86::SAR32rCL, X86::SHL32rCL}, // i32 1539 {64, X86::RCX, X86::SHR64rCL, X86::SAR64rCL, X86::SHL64rCL} // i64 1540 }; 1541 1542 if (DstRB.getID() != X86::GPRRegBankID) 1543 return false; 1544 1545 auto ShiftEntryIt = std::find_if( 1546 std::begin(OpTable), std::end(OpTable), [DstTy](const ShiftEntry &El) { 1547 return El.SizeInBits == DstTy.getSizeInBits(); 1548 }); 1549 if (ShiftEntryIt == std::end(OpTable)) 1550 return false; 1551 1552 unsigned CReg = ShiftEntryIt->CReg; 1553 unsigned Opcode = 0; 1554 switch (I.getOpcode()) { 1555 case TargetOpcode::G_SHL: 1556 Opcode = ShiftEntryIt->OpSHL; 1557 break; 1558 case TargetOpcode::G_ASHR: 1559 Opcode = ShiftEntryIt->OpASHR; 1560 break; 1561 case TargetOpcode::G_LSHR: 1562 Opcode = ShiftEntryIt->OpLSHR; 1563 break; 1564 default: 1565 return false; 1566 } 1567 1568 unsigned Op0Reg = I.getOperand(1).getReg(); 1569 unsigned Op1Reg = I.getOperand(2).getReg(); 1570 1571 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::COPY), 1572 ShiftEntryIt->CReg) 1573 .addReg(Op1Reg); 1574 1575 // The shift instruction uses X86::CL. If we defined a super-register 1576 // of X86::CL, emit a subreg KILL to precisely describe what we're doing here. 1577 if (CReg != X86::CL) 1578 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::KILL), 1579 X86::CL) 1580 .addReg(CReg, RegState::Kill); 1581 1582 MachineInstr &ShiftInst = 1583 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode), DstReg) 1584 .addReg(Op0Reg); 1585 1586 constrainSelectedInstRegOperands(ShiftInst, TII, TRI, RBI); 1587 I.eraseFromParent(); 1588 return true; 1589 } 1590 1591 bool X86InstructionSelector::selectDivRem(MachineInstr &I, 1592 MachineRegisterInfo &MRI, 1593 MachineFunction &MF) const { 1594 // The implementation of this function is taken from X86FastISel. 1595 assert((I.getOpcode() == TargetOpcode::G_SDIV || 1596 I.getOpcode() == TargetOpcode::G_SREM || 1597 I.getOpcode() == TargetOpcode::G_UDIV || 1598 I.getOpcode() == TargetOpcode::G_UREM) && 1599 "unexpected instruction"); 1600 1601 const unsigned DstReg = I.getOperand(0).getReg(); 1602 const unsigned Op1Reg = I.getOperand(1).getReg(); 1603 const unsigned Op2Reg = I.getOperand(2).getReg(); 1604 1605 const LLT RegTy = MRI.getType(DstReg); 1606 assert(RegTy == MRI.getType(Op1Reg) && RegTy == MRI.getType(Op2Reg) && 1607 "Arguments and return value types must match"); 1608 1609 const RegisterBank &RegRB = *RBI.getRegBank(DstReg, MRI, TRI); 1610 if (RegRB.getID() != X86::GPRRegBankID) 1611 return false; 1612 1613 const static unsigned NumTypes = 4; // i8, i16, i32, i64 1614 const static unsigned NumOps = 4; // SDiv, SRem, UDiv, URem 1615 const static bool S = true; // IsSigned 1616 const static bool U = false; // !IsSigned 1617 const static unsigned Copy = TargetOpcode::COPY; 1618 // For the X86 IDIV instruction, in most cases the dividend 1619 // (numerator) must be in a specific register pair highreg:lowreg, 1620 // producing the quotient in lowreg and the remainder in highreg. 1621 // For most data types, to set up the instruction, the dividend is 1622 // copied into lowreg, and lowreg is sign-extended into highreg. The 1623 // exception is i8, where the dividend is defined as a single register rather 1624 // than a register pair, and we therefore directly sign-extend the dividend 1625 // into lowreg, instead of copying, and ignore the highreg. 1626 const static struct DivRemEntry { 1627 // The following portion depends only on the data type. 1628 unsigned SizeInBits; 1629 unsigned LowInReg; // low part of the register pair 1630 unsigned HighInReg; // high part of the register pair 1631 // The following portion depends on both the data type and the operation. 1632 struct DivRemResult { 1633 unsigned OpDivRem; // The specific DIV/IDIV opcode to use. 1634 unsigned OpSignExtend; // Opcode for sign-extending lowreg into 1635 // highreg, or copying a zero into highreg. 1636 unsigned OpCopy; // Opcode for copying dividend into lowreg, or 1637 // zero/sign-extending into lowreg for i8. 1638 unsigned DivRemResultReg; // Register containing the desired result. 1639 bool IsOpSigned; // Whether to use signed or unsigned form. 1640 } ResultTable[NumOps]; 1641 } OpTable[NumTypes] = { 1642 {8, 1643 X86::AX, 1644 0, 1645 { 1646 {X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AL, S}, // SDiv 1647 {X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AH, S}, // SRem 1648 {X86::DIV8r, 0, X86::MOVZX16rr8, X86::AL, U}, // UDiv 1649 {X86::DIV8r, 0, X86::MOVZX16rr8, X86::AH, U}, // URem 1650 }}, // i8 1651 {16, 1652 X86::AX, 1653 X86::DX, 1654 { 1655 {X86::IDIV16r, X86::CWD, Copy, X86::AX, S}, // SDiv 1656 {X86::IDIV16r, X86::CWD, Copy, X86::DX, S}, // SRem 1657 {X86::DIV16r, X86::MOV32r0, Copy, X86::AX, U}, // UDiv 1658 {X86::DIV16r, X86::MOV32r0, Copy, X86::DX, U}, // URem 1659 }}, // i16 1660 {32, 1661 X86::EAX, 1662 X86::EDX, 1663 { 1664 {X86::IDIV32r, X86::CDQ, Copy, X86::EAX, S}, // SDiv 1665 {X86::IDIV32r, X86::CDQ, Copy, X86::EDX, S}, // SRem 1666 {X86::DIV32r, X86::MOV32r0, Copy, X86::EAX, U}, // UDiv 1667 {X86::DIV32r, X86::MOV32r0, Copy, X86::EDX, U}, // URem 1668 }}, // i32 1669 {64, 1670 X86::RAX, 1671 X86::RDX, 1672 { 1673 {X86::IDIV64r, X86::CQO, Copy, X86::RAX, S}, // SDiv 1674 {X86::IDIV64r, X86::CQO, Copy, X86::RDX, S}, // SRem 1675 {X86::DIV64r, X86::MOV32r0, Copy, X86::RAX, U}, // UDiv 1676 {X86::DIV64r, X86::MOV32r0, Copy, X86::RDX, U}, // URem 1677 }}, // i64 1678 }; 1679 1680 auto OpEntryIt = std::find_if(std::begin(OpTable), std::end(OpTable), 1681 [RegTy](const DivRemEntry &El) { 1682 return El.SizeInBits == RegTy.getSizeInBits(); 1683 }); 1684 if (OpEntryIt == std::end(OpTable)) 1685 return false; 1686 1687 unsigned OpIndex; 1688 switch (I.getOpcode()) { 1689 default: 1690 llvm_unreachable("Unexpected div/rem opcode"); 1691 case TargetOpcode::G_SDIV: 1692 OpIndex = 0; 1693 break; 1694 case TargetOpcode::G_SREM: 1695 OpIndex = 1; 1696 break; 1697 case TargetOpcode::G_UDIV: 1698 OpIndex = 2; 1699 break; 1700 case TargetOpcode::G_UREM: 1701 OpIndex = 3; 1702 break; 1703 } 1704 1705 const DivRemEntry &TypeEntry = *OpEntryIt; 1706 const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex]; 1707 1708 const TargetRegisterClass *RegRC = getRegClass(RegTy, RegRB); 1709 if (!RBI.constrainGenericRegister(Op1Reg, *RegRC, MRI) || 1710 !RBI.constrainGenericRegister(Op2Reg, *RegRC, MRI) || 1711 !RBI.constrainGenericRegister(DstReg, *RegRC, MRI)) { 1712 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 1713 << " operand\n"); 1714 return false; 1715 } 1716 1717 // Move op1 into low-order input register. 1718 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpEntry.OpCopy), 1719 TypeEntry.LowInReg) 1720 .addReg(Op1Reg); 1721 // Zero-extend or sign-extend into high-order input register. 1722 if (OpEntry.OpSignExtend) { 1723 if (OpEntry.IsOpSigned) 1724 BuildMI(*I.getParent(), I, I.getDebugLoc(), 1725 TII.get(OpEntry.OpSignExtend)); 1726 else { 1727 unsigned Zero32 = MRI.createVirtualRegister(&X86::GR32RegClass); 1728 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::MOV32r0), 1729 Zero32); 1730 1731 // Copy the zero into the appropriate sub/super/identical physical 1732 // register. Unfortunately the operations needed are not uniform enough 1733 // to fit neatly into the table above. 1734 if (RegTy.getSizeInBits() == 16) { 1735 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy), 1736 TypeEntry.HighInReg) 1737 .addReg(Zero32, 0, X86::sub_16bit); 1738 } else if (RegTy.getSizeInBits() == 32) { 1739 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy), 1740 TypeEntry.HighInReg) 1741 .addReg(Zero32); 1742 } else if (RegTy.getSizeInBits() == 64) { 1743 BuildMI(*I.getParent(), I, I.getDebugLoc(), 1744 TII.get(TargetOpcode::SUBREG_TO_REG), TypeEntry.HighInReg) 1745 .addImm(0) 1746 .addReg(Zero32) 1747 .addImm(X86::sub_32bit); 1748 } 1749 } 1750 } 1751 // Generate the DIV/IDIV instruction. 1752 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpEntry.OpDivRem)) 1753 .addReg(Op2Reg); 1754 // For i8 remainder, we can't reference ah directly, as we'll end 1755 // up with bogus copies like %r9b = COPY %ah. Reference ax 1756 // instead to prevent ah references in a rex instruction. 1757 // 1758 // The current assumption of the fast register allocator is that isel 1759 // won't generate explicit references to the GR8_NOREX registers. If 1760 // the allocator and/or the backend get enhanced to be more robust in 1761 // that regard, this can be, and should be, removed. 1762 if ((I.getOpcode() == Instruction::SRem || 1763 I.getOpcode() == Instruction::URem) && 1764 OpEntry.DivRemResultReg == X86::AH && STI.is64Bit()) { 1765 unsigned SourceSuperReg = MRI.createVirtualRegister(&X86::GR16RegClass); 1766 unsigned ResultSuperReg = MRI.createVirtualRegister(&X86::GR16RegClass); 1767 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy), SourceSuperReg) 1768 .addReg(X86::AX); 1769 1770 // Shift AX right by 8 bits instead of using AH. 1771 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::SHR16ri), 1772 ResultSuperReg) 1773 .addReg(SourceSuperReg) 1774 .addImm(8); 1775 1776 // Now reference the 8-bit subreg of the result. 1777 BuildMI(*I.getParent(), I, I.getDebugLoc(), 1778 TII.get(TargetOpcode::SUBREG_TO_REG)) 1779 .addDef(DstReg) 1780 .addImm(0) 1781 .addReg(ResultSuperReg) 1782 .addImm(X86::sub_8bit); 1783 } else { 1784 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::COPY), 1785 DstReg) 1786 .addReg(OpEntry.DivRemResultReg); 1787 } 1788 I.eraseFromParent(); 1789 return true; 1790 } 1791 1792 bool X86InstructionSelector::selectIntrinsicWSideEffects( 1793 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) const { 1794 1795 assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS && 1796 "unexpected instruction"); 1797 1798 if (I.getOperand(0).getIntrinsicID() != Intrinsic::trap) 1799 return false; 1800 1801 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::TRAP)); 1802 1803 I.eraseFromParent(); 1804 return true; 1805 } 1806 1807 InstructionSelector * 1808 llvm::createX86InstructionSelector(const X86TargetMachine &TM, 1809 X86Subtarget &Subtarget, 1810 X86RegisterBankInfo &RBI) { 1811 return new X86InstructionSelector(TM, Subtarget, RBI); 1812 } 1813