1 //===- X86InstructionSelector.cpp -----------------------------------------===// 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 /// \file 9 /// This file implements the targeting of the InstructionSelector class for 10 /// X86. 11 /// \todo This should be generated by TableGen. 12 //===----------------------------------------------------------------------===// 13 14 #include "MCTargetDesc/X86BaseInfo.h" 15 #include "X86InstrBuilder.h" 16 #include "X86InstrInfo.h" 17 #include "X86RegisterBankInfo.h" 18 #include "X86RegisterInfo.h" 19 #include "X86Subtarget.h" 20 #include "X86TargetMachine.h" 21 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h" 22 #include "llvm/CodeGen/GlobalISel/InstructionSelectorImpl.h" 23 #include "llvm/CodeGen/GlobalISel/RegisterBank.h" 24 #include "llvm/CodeGen/GlobalISel/Utils.h" 25 #include "llvm/CodeGen/MachineBasicBlock.h" 26 #include "llvm/CodeGen/MachineConstantPool.h" 27 #include "llvm/CodeGen/MachineFunction.h" 28 #include "llvm/CodeGen/MachineInstr.h" 29 #include "llvm/CodeGen/MachineInstrBuilder.h" 30 #include "llvm/CodeGen/MachineMemOperand.h" 31 #include "llvm/CodeGen/MachineOperand.h" 32 #include "llvm/CodeGen/MachineRegisterInfo.h" 33 #include "llvm/CodeGen/TargetOpcodes.h" 34 #include "llvm/CodeGen/TargetRegisterInfo.h" 35 #include "llvm/IR/DataLayout.h" 36 #include "llvm/IR/InstrTypes.h" 37 #include "llvm/Support/AtomicOrdering.h" 38 #include "llvm/Support/CodeGen.h" 39 #include "llvm/Support/Debug.h" 40 #include "llvm/Support/ErrorHandling.h" 41 #include "llvm/Support/LowLevelTypeImpl.h" 42 #include "llvm/Support/MathExtras.h" 43 #include "llvm/Support/raw_ostream.h" 44 #include <cassert> 45 #include <cstdint> 46 #include <tuple> 47 48 #define DEBUG_TYPE "X86-isel" 49 50 using namespace llvm; 51 52 namespace { 53 54 #define GET_GLOBALISEL_PREDICATE_BITSET 55 #include "X86GenGlobalISel.inc" 56 #undef GET_GLOBALISEL_PREDICATE_BITSET 57 58 class X86InstructionSelector : public InstructionSelector { 59 public: 60 X86InstructionSelector(const X86TargetMachine &TM, const X86Subtarget &STI, 61 const X86RegisterBankInfo &RBI); 62 63 bool select(MachineInstr &I) override; 64 static const char *getName() { return DEBUG_TYPE; } 65 66 private: 67 /// tblgen-erated 'select' implementation, used as the initial selector for 68 /// the patterns that don't require complex C++. 69 bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const; 70 71 // TODO: remove after supported by Tablegen-erated instruction selection. 72 unsigned getLoadStoreOp(const LLT &Ty, const RegisterBank &RB, unsigned Opc, 73 uint64_t Alignment) const; 74 75 bool selectLoadStoreOp(MachineInstr &I, MachineRegisterInfo &MRI, 76 MachineFunction &MF) const; 77 bool selectFrameIndexOrGep(MachineInstr &I, MachineRegisterInfo &MRI, 78 MachineFunction &MF) const; 79 bool selectGlobalValue(MachineInstr &I, MachineRegisterInfo &MRI, 80 MachineFunction &MF) const; 81 bool selectConstant(MachineInstr &I, MachineRegisterInfo &MRI, 82 MachineFunction &MF) const; 83 bool selectTruncOrPtrToInt(MachineInstr &I, MachineRegisterInfo &MRI, 84 MachineFunction &MF) const; 85 bool selectZext(MachineInstr &I, MachineRegisterInfo &MRI, 86 MachineFunction &MF) const; 87 bool selectAnyext(MachineInstr &I, MachineRegisterInfo &MRI, 88 MachineFunction &MF) const; 89 bool selectCmp(MachineInstr &I, MachineRegisterInfo &MRI, 90 MachineFunction &MF) const; 91 bool selectFCmp(MachineInstr &I, MachineRegisterInfo &MRI, 92 MachineFunction &MF) const; 93 bool selectUadde(MachineInstr &I, MachineRegisterInfo &MRI, 94 MachineFunction &MF) const; 95 bool selectCopy(MachineInstr &I, MachineRegisterInfo &MRI) const; 96 bool selectUnmergeValues(MachineInstr &I, MachineRegisterInfo &MRI, 97 MachineFunction &MF); 98 bool selectMergeValues(MachineInstr &I, MachineRegisterInfo &MRI, 99 MachineFunction &MF); 100 bool selectInsert(MachineInstr &I, MachineRegisterInfo &MRI, 101 MachineFunction &MF) const; 102 bool selectExtract(MachineInstr &I, MachineRegisterInfo &MRI, 103 MachineFunction &MF) const; 104 bool selectCondBranch(MachineInstr &I, MachineRegisterInfo &MRI, 105 MachineFunction &MF) const; 106 bool selectTurnIntoCOPY(MachineInstr &I, MachineRegisterInfo &MRI, 107 const unsigned DstReg, 108 const TargetRegisterClass *DstRC, 109 const unsigned SrcReg, 110 const TargetRegisterClass *SrcRC) const; 111 bool materializeFP(MachineInstr &I, MachineRegisterInfo &MRI, 112 MachineFunction &MF) const; 113 bool selectImplicitDefOrPHI(MachineInstr &I, MachineRegisterInfo &MRI) const; 114 bool selectShift(MachineInstr &I, MachineRegisterInfo &MRI, 115 MachineFunction &MF) const; 116 bool selectDivRem(MachineInstr &I, MachineRegisterInfo &MRI, 117 MachineFunction &MF) const; 118 bool selectIntrinsicWSideEffects(MachineInstr &I, MachineRegisterInfo &MRI, 119 MachineFunction &MF) const; 120 121 // emit insert subreg instruction and insert it before MachineInstr &I 122 bool emitInsertSubreg(unsigned DstReg, unsigned SrcReg, MachineInstr &I, 123 MachineRegisterInfo &MRI, MachineFunction &MF) const; 124 // emit extract subreg instruction and insert it before MachineInstr &I 125 bool emitExtractSubreg(unsigned DstReg, unsigned SrcReg, MachineInstr &I, 126 MachineRegisterInfo &MRI, MachineFunction &MF) const; 127 128 const TargetRegisterClass *getRegClass(LLT Ty, const RegisterBank &RB) const; 129 const TargetRegisterClass *getRegClass(LLT Ty, unsigned Reg, 130 MachineRegisterInfo &MRI) const; 131 132 const X86TargetMachine &TM; 133 const X86Subtarget &STI; 134 const X86InstrInfo &TII; 135 const X86RegisterInfo &TRI; 136 const X86RegisterBankInfo &RBI; 137 138 #define GET_GLOBALISEL_PREDICATES_DECL 139 #include "X86GenGlobalISel.inc" 140 #undef GET_GLOBALISEL_PREDICATES_DECL 141 142 #define GET_GLOBALISEL_TEMPORARIES_DECL 143 #include "X86GenGlobalISel.inc" 144 #undef GET_GLOBALISEL_TEMPORARIES_DECL 145 }; 146 147 } // end anonymous namespace 148 149 #define GET_GLOBALISEL_IMPL 150 #include "X86GenGlobalISel.inc" 151 #undef GET_GLOBALISEL_IMPL 152 153 X86InstructionSelector::X86InstructionSelector(const X86TargetMachine &TM, 154 const X86Subtarget &STI, 155 const X86RegisterBankInfo &RBI) 156 : InstructionSelector(), TM(TM), STI(STI), TII(*STI.getInstrInfo()), 157 TRI(*STI.getRegisterInfo()), RBI(RBI), 158 #define GET_GLOBALISEL_PREDICATES_INIT 159 #include "X86GenGlobalISel.inc" 160 #undef GET_GLOBALISEL_PREDICATES_INIT 161 #define GET_GLOBALISEL_TEMPORARIES_INIT 162 #include "X86GenGlobalISel.inc" 163 #undef GET_GLOBALISEL_TEMPORARIES_INIT 164 { 165 } 166 167 // FIXME: This should be target-independent, inferred from the types declared 168 // for each class in the bank. 169 const TargetRegisterClass * 170 X86InstructionSelector::getRegClass(LLT Ty, const RegisterBank &RB) const { 171 if (RB.getID() == X86::GPRRegBankID) { 172 if (Ty.getSizeInBits() <= 8) 173 return &X86::GR8RegClass; 174 if (Ty.getSizeInBits() == 16) 175 return &X86::GR16RegClass; 176 if (Ty.getSizeInBits() == 32) 177 return &X86::GR32RegClass; 178 if (Ty.getSizeInBits() == 64) 179 return &X86::GR64RegClass; 180 } 181 if (RB.getID() == X86::VECRRegBankID) { 182 if (Ty.getSizeInBits() == 32) 183 return STI.hasAVX512() ? &X86::FR32XRegClass : &X86::FR32RegClass; 184 if (Ty.getSizeInBits() == 64) 185 return STI.hasAVX512() ? &X86::FR64XRegClass : &X86::FR64RegClass; 186 if (Ty.getSizeInBits() == 128) 187 return STI.hasAVX512() ? &X86::VR128XRegClass : &X86::VR128RegClass; 188 if (Ty.getSizeInBits() == 256) 189 return STI.hasAVX512() ? &X86::VR256XRegClass : &X86::VR256RegClass; 190 if (Ty.getSizeInBits() == 512) 191 return &X86::VR512RegClass; 192 } 193 194 llvm_unreachable("Unknown RegBank!"); 195 } 196 197 const TargetRegisterClass * 198 X86InstructionSelector::getRegClass(LLT Ty, unsigned Reg, 199 MachineRegisterInfo &MRI) const { 200 const RegisterBank &RegBank = *RBI.getRegBank(Reg, MRI, TRI); 201 return getRegClass(Ty, RegBank); 202 } 203 204 static unsigned getSubRegIndex(const TargetRegisterClass *RC) { 205 unsigned SubIdx = X86::NoSubRegister; 206 if (RC == &X86::GR32RegClass) { 207 SubIdx = X86::sub_32bit; 208 } else if (RC == &X86::GR16RegClass) { 209 SubIdx = X86::sub_16bit; 210 } else if (RC == &X86::GR8RegClass) { 211 SubIdx = X86::sub_8bit; 212 } 213 214 return SubIdx; 215 } 216 217 static const TargetRegisterClass *getRegClassFromGRPhysReg(unsigned Reg) { 218 assert(Register::isPhysicalRegister(Reg)); 219 if (X86::GR64RegClass.contains(Reg)) 220 return &X86::GR64RegClass; 221 if (X86::GR32RegClass.contains(Reg)) 222 return &X86::GR32RegClass; 223 if (X86::GR16RegClass.contains(Reg)) 224 return &X86::GR16RegClass; 225 if (X86::GR8RegClass.contains(Reg)) 226 return &X86::GR8RegClass; 227 228 llvm_unreachable("Unknown RegClass for PhysReg!"); 229 } 230 231 // Set X86 Opcode and constrain DestReg. 232 bool X86InstructionSelector::selectCopy(MachineInstr &I, 233 MachineRegisterInfo &MRI) const { 234 Register DstReg = I.getOperand(0).getReg(); 235 const unsigned DstSize = RBI.getSizeInBits(DstReg, MRI, TRI); 236 const RegisterBank &DstRegBank = *RBI.getRegBank(DstReg, MRI, TRI); 237 238 Register SrcReg = I.getOperand(1).getReg(); 239 const unsigned SrcSize = RBI.getSizeInBits(SrcReg, MRI, TRI); 240 const RegisterBank &SrcRegBank = *RBI.getRegBank(SrcReg, MRI, TRI); 241 242 if (Register::isPhysicalRegister(DstReg)) { 243 assert(I.isCopy() && "Generic operators do not allow physical registers"); 244 245 if (DstSize > SrcSize && SrcRegBank.getID() == X86::GPRRegBankID && 246 DstRegBank.getID() == X86::GPRRegBankID) { 247 248 const TargetRegisterClass *SrcRC = 249 getRegClass(MRI.getType(SrcReg), SrcRegBank); 250 const TargetRegisterClass *DstRC = getRegClassFromGRPhysReg(DstReg); 251 252 if (SrcRC != DstRC) { 253 // This case can be generated by ABI lowering, performe anyext 254 Register ExtSrc = MRI.createVirtualRegister(DstRC); 255 BuildMI(*I.getParent(), I, I.getDebugLoc(), 256 TII.get(TargetOpcode::SUBREG_TO_REG)) 257 .addDef(ExtSrc) 258 .addImm(0) 259 .addReg(SrcReg) 260 .addImm(getSubRegIndex(SrcRC)); 261 262 I.getOperand(1).setReg(ExtSrc); 263 } 264 } 265 266 return true; 267 } 268 269 assert((!Register::isPhysicalRegister(SrcReg) || I.isCopy()) && 270 "No phys reg on generic operators"); 271 assert((DstSize == SrcSize || 272 // Copies are a mean to setup initial types, the number of 273 // bits may not exactly match. 274 (Register::isPhysicalRegister(SrcReg) && 275 DstSize <= RBI.getSizeInBits(SrcReg, MRI, TRI))) && 276 "Copy with different width?!"); 277 278 const TargetRegisterClass *DstRC = 279 getRegClass(MRI.getType(DstReg), DstRegBank); 280 281 if (SrcRegBank.getID() == X86::GPRRegBankID && 282 DstRegBank.getID() == X86::GPRRegBankID && SrcSize > DstSize && 283 Register::isPhysicalRegister(SrcReg)) { 284 // Change the physical register to performe truncate. 285 286 const TargetRegisterClass *SrcRC = getRegClassFromGRPhysReg(SrcReg); 287 288 if (DstRC != SrcRC) { 289 I.getOperand(1).setSubReg(getSubRegIndex(DstRC)); 290 I.getOperand(1).substPhysReg(SrcReg, TRI); 291 } 292 } 293 294 // No need to constrain SrcReg. It will get constrained when 295 // we hit another of its use or its defs. 296 // Copies do not have constraints. 297 const TargetRegisterClass *OldRC = MRI.getRegClassOrNull(DstReg); 298 if (!OldRC || !DstRC->hasSubClassEq(OldRC)) { 299 if (!RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 300 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 301 << " operand\n"); 302 return false; 303 } 304 } 305 I.setDesc(TII.get(X86::COPY)); 306 return true; 307 } 308 309 bool X86InstructionSelector::select(MachineInstr &I) { 310 assert(I.getParent() && "Instruction should be in a basic block!"); 311 assert(I.getParent()->getParent() && "Instruction should be in a function!"); 312 313 MachineBasicBlock &MBB = *I.getParent(); 314 MachineFunction &MF = *MBB.getParent(); 315 MachineRegisterInfo &MRI = MF.getRegInfo(); 316 317 unsigned Opcode = I.getOpcode(); 318 if (!isPreISelGenericOpcode(Opcode)) { 319 // Certain non-generic instructions also need some special handling. 320 321 if (Opcode == TargetOpcode::LOAD_STACK_GUARD) 322 return false; 323 324 if (I.isCopy()) 325 return selectCopy(I, MRI); 326 327 return true; 328 } 329 330 assert(I.getNumOperands() == I.getNumExplicitOperands() && 331 "Generic instruction has unexpected implicit operands\n"); 332 333 if (selectImpl(I, *CoverageInfo)) 334 return true; 335 336 LLVM_DEBUG(dbgs() << " C++ instruction selection: "; I.print(dbgs())); 337 338 // TODO: This should be implemented by tblgen. 339 switch (I.getOpcode()) { 340 default: 341 return false; 342 case TargetOpcode::G_STORE: 343 case TargetOpcode::G_LOAD: 344 return selectLoadStoreOp(I, MRI, MF); 345 case TargetOpcode::G_GEP: 346 case TargetOpcode::G_FRAME_INDEX: 347 return selectFrameIndexOrGep(I, MRI, MF); 348 case TargetOpcode::G_GLOBAL_VALUE: 349 return selectGlobalValue(I, MRI, MF); 350 case TargetOpcode::G_CONSTANT: 351 return selectConstant(I, MRI, MF); 352 case TargetOpcode::G_FCONSTANT: 353 return materializeFP(I, MRI, MF); 354 case TargetOpcode::G_PTRTOINT: 355 case TargetOpcode::G_TRUNC: 356 return selectTruncOrPtrToInt(I, MRI, MF); 357 case TargetOpcode::G_INTTOPTR: 358 return selectCopy(I, MRI); 359 case TargetOpcode::G_ZEXT: 360 return selectZext(I, MRI, MF); 361 case TargetOpcode::G_ANYEXT: 362 return selectAnyext(I, MRI, MF); 363 case TargetOpcode::G_ICMP: 364 return selectCmp(I, MRI, MF); 365 case TargetOpcode::G_FCMP: 366 return selectFCmp(I, MRI, MF); 367 case TargetOpcode::G_UADDE: 368 return selectUadde(I, MRI, MF); 369 case TargetOpcode::G_UNMERGE_VALUES: 370 return selectUnmergeValues(I, MRI, MF); 371 case TargetOpcode::G_MERGE_VALUES: 372 case TargetOpcode::G_CONCAT_VECTORS: 373 return selectMergeValues(I, MRI, MF); 374 case TargetOpcode::G_EXTRACT: 375 return selectExtract(I, MRI, MF); 376 case TargetOpcode::G_INSERT: 377 return selectInsert(I, MRI, MF); 378 case TargetOpcode::G_BRCOND: 379 return selectCondBranch(I, MRI, MF); 380 case TargetOpcode::G_IMPLICIT_DEF: 381 case TargetOpcode::G_PHI: 382 return selectImplicitDefOrPHI(I, MRI); 383 case TargetOpcode::G_SHL: 384 case TargetOpcode::G_ASHR: 385 case TargetOpcode::G_LSHR: 386 return selectShift(I, MRI, MF); 387 case TargetOpcode::G_SDIV: 388 case TargetOpcode::G_UDIV: 389 case TargetOpcode::G_SREM: 390 case TargetOpcode::G_UREM: 391 return selectDivRem(I, MRI, MF); 392 case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS: 393 return selectIntrinsicWSideEffects(I, MRI, MF); 394 } 395 396 return false; 397 } 398 399 unsigned X86InstructionSelector::getLoadStoreOp(const LLT &Ty, 400 const RegisterBank &RB, 401 unsigned Opc, 402 uint64_t Alignment) const { 403 bool Isload = (Opc == TargetOpcode::G_LOAD); 404 bool HasAVX = STI.hasAVX(); 405 bool HasAVX512 = STI.hasAVX512(); 406 bool HasVLX = STI.hasVLX(); 407 408 if (Ty == LLT::scalar(8)) { 409 if (X86::GPRRegBankID == RB.getID()) 410 return Isload ? X86::MOV8rm : X86::MOV8mr; 411 } else if (Ty == LLT::scalar(16)) { 412 if (X86::GPRRegBankID == RB.getID()) 413 return Isload ? X86::MOV16rm : X86::MOV16mr; 414 } else if (Ty == LLT::scalar(32) || Ty == LLT::pointer(0, 32)) { 415 if (X86::GPRRegBankID == RB.getID()) 416 return Isload ? X86::MOV32rm : X86::MOV32mr; 417 if (X86::VECRRegBankID == RB.getID()) 418 return Isload ? (HasAVX512 ? X86::VMOVSSZrm_alt : 419 HasAVX ? X86::VMOVSSrm_alt : 420 X86::MOVSSrm_alt) 421 : (HasAVX512 ? X86::VMOVSSZmr : 422 HasAVX ? X86::VMOVSSmr : 423 X86::MOVSSmr); 424 } else if (Ty == LLT::scalar(64) || Ty == LLT::pointer(0, 64)) { 425 if (X86::GPRRegBankID == RB.getID()) 426 return Isload ? X86::MOV64rm : X86::MOV64mr; 427 if (X86::VECRRegBankID == RB.getID()) 428 return Isload ? (HasAVX512 ? X86::VMOVSDZrm_alt : 429 HasAVX ? X86::VMOVSDrm_alt : 430 X86::MOVSDrm_alt) 431 : (HasAVX512 ? X86::VMOVSDZmr : 432 HasAVX ? X86::VMOVSDmr : 433 X86::MOVSDmr); 434 } else if (Ty.isVector() && Ty.getSizeInBits() == 128) { 435 if (Alignment >= 16) 436 return Isload ? (HasVLX ? X86::VMOVAPSZ128rm 437 : HasAVX512 438 ? X86::VMOVAPSZ128rm_NOVLX 439 : HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm) 440 : (HasVLX ? X86::VMOVAPSZ128mr 441 : HasAVX512 442 ? X86::VMOVAPSZ128mr_NOVLX 443 : HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr); 444 else 445 return Isload ? (HasVLX ? X86::VMOVUPSZ128rm 446 : HasAVX512 447 ? X86::VMOVUPSZ128rm_NOVLX 448 : HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm) 449 : (HasVLX ? X86::VMOVUPSZ128mr 450 : HasAVX512 451 ? X86::VMOVUPSZ128mr_NOVLX 452 : HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr); 453 } else if (Ty.isVector() && Ty.getSizeInBits() == 256) { 454 if (Alignment >= 32) 455 return Isload ? (HasVLX ? X86::VMOVAPSZ256rm 456 : HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX 457 : X86::VMOVAPSYrm) 458 : (HasVLX ? X86::VMOVAPSZ256mr 459 : HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX 460 : X86::VMOVAPSYmr); 461 else 462 return Isload ? (HasVLX ? X86::VMOVUPSZ256rm 463 : HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX 464 : X86::VMOVUPSYrm) 465 : (HasVLX ? X86::VMOVUPSZ256mr 466 : HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX 467 : X86::VMOVUPSYmr); 468 } else if (Ty.isVector() && Ty.getSizeInBits() == 512) { 469 if (Alignment >= 64) 470 return Isload ? X86::VMOVAPSZrm : X86::VMOVAPSZmr; 471 else 472 return Isload ? X86::VMOVUPSZrm : X86::VMOVUPSZmr; 473 } 474 return Opc; 475 } 476 477 // Fill in an address from the given instruction. 478 static void X86SelectAddress(const MachineInstr &I, 479 const MachineRegisterInfo &MRI, 480 X86AddressMode &AM) { 481 assert(I.getOperand(0).isReg() && "unsupported opperand."); 482 assert(MRI.getType(I.getOperand(0).getReg()).isPointer() && 483 "unsupported type."); 484 485 if (I.getOpcode() == TargetOpcode::G_GEP) { 486 if (auto COff = getConstantVRegVal(I.getOperand(2).getReg(), MRI)) { 487 int64_t Imm = *COff; 488 if (isInt<32>(Imm)) { // Check for displacement overflow. 489 AM.Disp = static_cast<int32_t>(Imm); 490 AM.Base.Reg = I.getOperand(1).getReg(); 491 return; 492 } 493 } 494 } else if (I.getOpcode() == TargetOpcode::G_FRAME_INDEX) { 495 AM.Base.FrameIndex = I.getOperand(1).getIndex(); 496 AM.BaseType = X86AddressMode::FrameIndexBase; 497 return; 498 } 499 500 // Default behavior. 501 AM.Base.Reg = I.getOperand(0).getReg(); 502 } 503 504 bool X86InstructionSelector::selectLoadStoreOp(MachineInstr &I, 505 MachineRegisterInfo &MRI, 506 MachineFunction &MF) const { 507 unsigned Opc = I.getOpcode(); 508 509 assert((Opc == TargetOpcode::G_STORE || Opc == TargetOpcode::G_LOAD) && 510 "unexpected instruction"); 511 512 const Register DefReg = I.getOperand(0).getReg(); 513 LLT Ty = MRI.getType(DefReg); 514 const RegisterBank &RB = *RBI.getRegBank(DefReg, MRI, TRI); 515 516 assert(I.hasOneMemOperand()); 517 auto &MemOp = **I.memoperands_begin(); 518 if (MemOp.isAtomic()) { 519 // Note: for unordered operations, we rely on the fact the appropriate MMO 520 // is already on the instruction we're mutating, and thus we don't need to 521 // make any changes. So long as we select an opcode which is capable of 522 // loading or storing the appropriate size atomically, the rest of the 523 // backend is required to respect the MMO state. 524 if (!MemOp.isUnordered()) { 525 LLVM_DEBUG(dbgs() << "Atomic ordering not supported yet\n"); 526 return false; 527 } 528 if (MemOp.getAlignment() < Ty.getSizeInBits()/8) { 529 LLVM_DEBUG(dbgs() << "Unaligned atomics not supported yet\n"); 530 return false; 531 } 532 } 533 534 unsigned NewOpc = getLoadStoreOp(Ty, RB, Opc, MemOp.getAlignment()); 535 if (NewOpc == Opc) 536 return false; 537 538 X86AddressMode AM; 539 X86SelectAddress(*MRI.getVRegDef(I.getOperand(1).getReg()), MRI, AM); 540 541 I.setDesc(TII.get(NewOpc)); 542 MachineInstrBuilder MIB(MF, I); 543 if (Opc == TargetOpcode::G_LOAD) { 544 I.RemoveOperand(1); 545 addFullAddress(MIB, AM); 546 } else { 547 // G_STORE (VAL, Addr), X86Store instruction (Addr, VAL) 548 I.RemoveOperand(1); 549 I.RemoveOperand(0); 550 addFullAddress(MIB, AM).addUse(DefReg); 551 } 552 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 553 } 554 555 static unsigned getLeaOP(LLT Ty, const X86Subtarget &STI) { 556 if (Ty == LLT::pointer(0, 64)) 557 return X86::LEA64r; 558 else if (Ty == LLT::pointer(0, 32)) 559 return STI.isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r; 560 else 561 llvm_unreachable("Can't get LEA opcode. Unsupported type."); 562 } 563 564 bool X86InstructionSelector::selectFrameIndexOrGep(MachineInstr &I, 565 MachineRegisterInfo &MRI, 566 MachineFunction &MF) const { 567 unsigned Opc = I.getOpcode(); 568 569 assert((Opc == TargetOpcode::G_FRAME_INDEX || Opc == TargetOpcode::G_GEP) && 570 "unexpected instruction"); 571 572 const Register DefReg = I.getOperand(0).getReg(); 573 LLT Ty = MRI.getType(DefReg); 574 575 // Use LEA to calculate frame index and GEP 576 unsigned NewOpc = getLeaOP(Ty, STI); 577 I.setDesc(TII.get(NewOpc)); 578 MachineInstrBuilder MIB(MF, I); 579 580 if (Opc == TargetOpcode::G_FRAME_INDEX) { 581 addOffset(MIB, 0); 582 } else { 583 MachineOperand &InxOp = I.getOperand(2); 584 I.addOperand(InxOp); // set IndexReg 585 InxOp.ChangeToImmediate(1); // set Scale 586 MIB.addImm(0).addReg(0); 587 } 588 589 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 590 } 591 592 bool X86InstructionSelector::selectGlobalValue(MachineInstr &I, 593 MachineRegisterInfo &MRI, 594 MachineFunction &MF) const { 595 assert((I.getOpcode() == TargetOpcode::G_GLOBAL_VALUE) && 596 "unexpected instruction"); 597 598 auto GV = I.getOperand(1).getGlobal(); 599 if (GV->isThreadLocal()) { 600 return false; // TODO: we don't support TLS yet. 601 } 602 603 // Can't handle alternate code models yet. 604 if (TM.getCodeModel() != CodeModel::Small) 605 return false; 606 607 X86AddressMode AM; 608 AM.GV = GV; 609 AM.GVOpFlags = STI.classifyGlobalReference(GV); 610 611 // TODO: The ABI requires an extra load. not supported yet. 612 if (isGlobalStubReference(AM.GVOpFlags)) 613 return false; 614 615 // TODO: This reference is relative to the pic base. not supported yet. 616 if (isGlobalRelativeToPICBase(AM.GVOpFlags)) 617 return false; 618 619 if (STI.isPICStyleRIPRel()) { 620 // Use rip-relative addressing. 621 assert(AM.Base.Reg == 0 && AM.IndexReg == 0); 622 AM.Base.Reg = X86::RIP; 623 } 624 625 const Register DefReg = I.getOperand(0).getReg(); 626 LLT Ty = MRI.getType(DefReg); 627 unsigned NewOpc = getLeaOP(Ty, STI); 628 629 I.setDesc(TII.get(NewOpc)); 630 MachineInstrBuilder MIB(MF, I); 631 632 I.RemoveOperand(1); 633 addFullAddress(MIB, AM); 634 635 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 636 } 637 638 bool X86InstructionSelector::selectConstant(MachineInstr &I, 639 MachineRegisterInfo &MRI, 640 MachineFunction &MF) const { 641 assert((I.getOpcode() == TargetOpcode::G_CONSTANT) && 642 "unexpected instruction"); 643 644 const Register DefReg = I.getOperand(0).getReg(); 645 LLT Ty = MRI.getType(DefReg); 646 647 if (RBI.getRegBank(DefReg, MRI, TRI)->getID() != X86::GPRRegBankID) 648 return false; 649 650 uint64_t Val = 0; 651 if (I.getOperand(1).isCImm()) { 652 Val = I.getOperand(1).getCImm()->getZExtValue(); 653 I.getOperand(1).ChangeToImmediate(Val); 654 } else if (I.getOperand(1).isImm()) { 655 Val = I.getOperand(1).getImm(); 656 } else 657 llvm_unreachable("Unsupported operand type."); 658 659 unsigned NewOpc; 660 switch (Ty.getSizeInBits()) { 661 case 8: 662 NewOpc = X86::MOV8ri; 663 break; 664 case 16: 665 NewOpc = X86::MOV16ri; 666 break; 667 case 32: 668 NewOpc = X86::MOV32ri; 669 break; 670 case 64: 671 // TODO: in case isUInt<32>(Val), X86::MOV32ri can be used 672 if (isInt<32>(Val)) 673 NewOpc = X86::MOV64ri32; 674 else 675 NewOpc = X86::MOV64ri; 676 break; 677 default: 678 llvm_unreachable("Can't select G_CONSTANT, unsupported type."); 679 } 680 681 I.setDesc(TII.get(NewOpc)); 682 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 683 } 684 685 // Helper function for selectTruncOrPtrToInt and selectAnyext. 686 // Returns true if DstRC lives on a floating register class and 687 // SrcRC lives on a 128-bit vector class. 688 static bool canTurnIntoCOPY(const TargetRegisterClass *DstRC, 689 const TargetRegisterClass *SrcRC) { 690 return (DstRC == &X86::FR32RegClass || DstRC == &X86::FR32XRegClass || 691 DstRC == &X86::FR64RegClass || DstRC == &X86::FR64XRegClass) && 692 (SrcRC == &X86::VR128RegClass || SrcRC == &X86::VR128XRegClass); 693 } 694 695 bool X86InstructionSelector::selectTurnIntoCOPY( 696 MachineInstr &I, MachineRegisterInfo &MRI, const unsigned DstReg, 697 const TargetRegisterClass *DstRC, const unsigned SrcReg, 698 const TargetRegisterClass *SrcRC) const { 699 700 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 701 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 702 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 703 << " operand\n"); 704 return false; 705 } 706 I.setDesc(TII.get(X86::COPY)); 707 return true; 708 } 709 710 bool X86InstructionSelector::selectTruncOrPtrToInt(MachineInstr &I, 711 MachineRegisterInfo &MRI, 712 MachineFunction &MF) const { 713 assert((I.getOpcode() == TargetOpcode::G_TRUNC || 714 I.getOpcode() == TargetOpcode::G_PTRTOINT) && 715 "unexpected instruction"); 716 717 const Register DstReg = I.getOperand(0).getReg(); 718 const Register SrcReg = I.getOperand(1).getReg(); 719 720 const LLT DstTy = MRI.getType(DstReg); 721 const LLT SrcTy = MRI.getType(SrcReg); 722 723 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI); 724 const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI); 725 726 if (DstRB.getID() != SrcRB.getID()) { 727 LLVM_DEBUG(dbgs() << TII.getName(I.getOpcode()) 728 << " input/output on different banks\n"); 729 return false; 730 } 731 732 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB); 733 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB); 734 735 if (!DstRC || !SrcRC) 736 return false; 737 738 // If that's truncation of the value that lives on the vector class and goes 739 // into the floating class, just replace it with copy, as we are able to 740 // select it as a regular move. 741 if (canTurnIntoCOPY(DstRC, SrcRC)) 742 return selectTurnIntoCOPY(I, MRI, DstReg, DstRC, SrcReg, SrcRC); 743 744 if (DstRB.getID() != X86::GPRRegBankID) 745 return false; 746 747 unsigned SubIdx; 748 if (DstRC == SrcRC) { 749 // Nothing to be done 750 SubIdx = X86::NoSubRegister; 751 } else if (DstRC == &X86::GR32RegClass) { 752 SubIdx = X86::sub_32bit; 753 } else if (DstRC == &X86::GR16RegClass) { 754 SubIdx = X86::sub_16bit; 755 } else if (DstRC == &X86::GR8RegClass) { 756 SubIdx = X86::sub_8bit; 757 } else { 758 return false; 759 } 760 761 SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubIdx); 762 763 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 764 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 765 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 766 << "\n"); 767 return false; 768 } 769 770 I.getOperand(1).setSubReg(SubIdx); 771 772 I.setDesc(TII.get(X86::COPY)); 773 return true; 774 } 775 776 bool X86InstructionSelector::selectZext(MachineInstr &I, 777 MachineRegisterInfo &MRI, 778 MachineFunction &MF) const { 779 assert((I.getOpcode() == TargetOpcode::G_ZEXT) && "unexpected instruction"); 780 781 const Register DstReg = I.getOperand(0).getReg(); 782 const Register SrcReg = I.getOperand(1).getReg(); 783 784 const LLT DstTy = MRI.getType(DstReg); 785 const LLT SrcTy = MRI.getType(SrcReg); 786 787 assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(32)) && 788 "8=>32 Zext is handled by tablegen"); 789 assert(!(SrcTy == LLT::scalar(16) && DstTy == LLT::scalar(32)) && 790 "16=>32 Zext is handled by tablegen"); 791 792 const static struct ZextEntry { 793 LLT SrcTy; 794 LLT DstTy; 795 unsigned MovOp; 796 bool NeedSubregToReg; 797 } OpTable[] = { 798 {LLT::scalar(8), LLT::scalar(16), X86::MOVZX16rr8, false}, // i8 => i16 799 {LLT::scalar(8), LLT::scalar(64), X86::MOVZX32rr8, true}, // i8 => i64 800 {LLT::scalar(16), LLT::scalar(64), X86::MOVZX32rr16, true}, // i16 => i64 801 {LLT::scalar(32), LLT::scalar(64), 0, true} // i32 => i64 802 }; 803 804 auto ZextEntryIt = 805 std::find_if(std::begin(OpTable), std::end(OpTable), 806 [SrcTy, DstTy](const ZextEntry &El) { 807 return El.DstTy == DstTy && El.SrcTy == SrcTy; 808 }); 809 810 // Here we try to select Zext into a MOVZ and/or SUBREG_TO_REG instruction. 811 if (ZextEntryIt != std::end(OpTable)) { 812 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI); 813 const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI); 814 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB); 815 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB); 816 817 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 818 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 819 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 820 << " operand\n"); 821 return false; 822 } 823 824 unsigned TransitRegTo = DstReg; 825 unsigned TransitRegFrom = SrcReg; 826 if (ZextEntryIt->MovOp) { 827 // If we select Zext into MOVZ + SUBREG_TO_REG, we need to have 828 // a transit register in between: create it here. 829 if (ZextEntryIt->NeedSubregToReg) { 830 TransitRegFrom = MRI.createVirtualRegister( 831 getRegClass(LLT::scalar(32), DstReg, MRI)); 832 TransitRegTo = TransitRegFrom; 833 } 834 835 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(ZextEntryIt->MovOp)) 836 .addDef(TransitRegTo) 837 .addReg(SrcReg); 838 } 839 if (ZextEntryIt->NeedSubregToReg) { 840 BuildMI(*I.getParent(), I, I.getDebugLoc(), 841 TII.get(TargetOpcode::SUBREG_TO_REG)) 842 .addDef(DstReg) 843 .addImm(0) 844 .addReg(TransitRegFrom) 845 .addImm(X86::sub_32bit); 846 } 847 I.eraseFromParent(); 848 return true; 849 } 850 851 if (SrcTy != LLT::scalar(1)) 852 return false; 853 854 unsigned AndOpc; 855 if (DstTy == LLT::scalar(8)) 856 AndOpc = X86::AND8ri; 857 else if (DstTy == LLT::scalar(16)) 858 AndOpc = X86::AND16ri8; 859 else if (DstTy == LLT::scalar(32)) 860 AndOpc = X86::AND32ri8; 861 else if (DstTy == LLT::scalar(64)) 862 AndOpc = X86::AND64ri8; 863 else 864 return false; 865 866 unsigned DefReg = SrcReg; 867 if (DstTy != LLT::scalar(8)) { 868 DefReg = MRI.createVirtualRegister(getRegClass(DstTy, DstReg, MRI)); 869 BuildMI(*I.getParent(), I, I.getDebugLoc(), 870 TII.get(TargetOpcode::SUBREG_TO_REG), DefReg) 871 .addImm(0) 872 .addReg(SrcReg) 873 .addImm(X86::sub_8bit); 874 } 875 876 MachineInstr &AndInst = 877 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(AndOpc), DstReg) 878 .addReg(DefReg) 879 .addImm(1); 880 881 constrainSelectedInstRegOperands(AndInst, TII, TRI, RBI); 882 883 I.eraseFromParent(); 884 return true; 885 } 886 887 bool X86InstructionSelector::selectAnyext(MachineInstr &I, 888 MachineRegisterInfo &MRI, 889 MachineFunction &MF) const { 890 assert((I.getOpcode() == TargetOpcode::G_ANYEXT) && "unexpected instruction"); 891 892 const Register DstReg = I.getOperand(0).getReg(); 893 const Register SrcReg = I.getOperand(1).getReg(); 894 895 const LLT DstTy = MRI.getType(DstReg); 896 const LLT SrcTy = MRI.getType(SrcReg); 897 898 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI); 899 const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI); 900 901 assert(DstRB.getID() == SrcRB.getID() && 902 "G_ANYEXT input/output on different banks\n"); 903 904 assert(DstTy.getSizeInBits() > SrcTy.getSizeInBits() && 905 "G_ANYEXT incorrect operand size"); 906 907 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB); 908 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB); 909 910 // If that's ANY_EXT of the value that lives on the floating class and goes 911 // into the vector class, just replace it with copy, as we are able to select 912 // it as a regular move. 913 if (canTurnIntoCOPY(SrcRC, DstRC)) 914 return selectTurnIntoCOPY(I, MRI, SrcReg, SrcRC, DstReg, DstRC); 915 916 if (DstRB.getID() != X86::GPRRegBankID) 917 return false; 918 919 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 920 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 921 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 922 << " operand\n"); 923 return false; 924 } 925 926 if (SrcRC == DstRC) { 927 I.setDesc(TII.get(X86::COPY)); 928 return true; 929 } 930 931 BuildMI(*I.getParent(), I, I.getDebugLoc(), 932 TII.get(TargetOpcode::SUBREG_TO_REG)) 933 .addDef(DstReg) 934 .addImm(0) 935 .addReg(SrcReg) 936 .addImm(getSubRegIndex(SrcRC)); 937 938 I.eraseFromParent(); 939 return true; 940 } 941 942 bool X86InstructionSelector::selectCmp(MachineInstr &I, 943 MachineRegisterInfo &MRI, 944 MachineFunction &MF) const { 945 assert((I.getOpcode() == TargetOpcode::G_ICMP) && "unexpected instruction"); 946 947 X86::CondCode CC; 948 bool SwapArgs; 949 std::tie(CC, SwapArgs) = X86::getX86ConditionCode( 950 (CmpInst::Predicate)I.getOperand(1).getPredicate()); 951 952 Register LHS = I.getOperand(2).getReg(); 953 Register RHS = I.getOperand(3).getReg(); 954 955 if (SwapArgs) 956 std::swap(LHS, RHS); 957 958 unsigned OpCmp; 959 LLT Ty = MRI.getType(LHS); 960 961 switch (Ty.getSizeInBits()) { 962 default: 963 return false; 964 case 8: 965 OpCmp = X86::CMP8rr; 966 break; 967 case 16: 968 OpCmp = X86::CMP16rr; 969 break; 970 case 32: 971 OpCmp = X86::CMP32rr; 972 break; 973 case 64: 974 OpCmp = X86::CMP64rr; 975 break; 976 } 977 978 MachineInstr &CmpInst = 979 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp)) 980 .addReg(LHS) 981 .addReg(RHS); 982 983 MachineInstr &SetInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 984 TII.get(X86::SETCCr), I.getOperand(0).getReg()).addImm(CC); 985 986 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI); 987 constrainSelectedInstRegOperands(SetInst, TII, TRI, RBI); 988 989 I.eraseFromParent(); 990 return true; 991 } 992 993 bool X86InstructionSelector::selectFCmp(MachineInstr &I, 994 MachineRegisterInfo &MRI, 995 MachineFunction &MF) const { 996 assert((I.getOpcode() == TargetOpcode::G_FCMP) && "unexpected instruction"); 997 998 Register LhsReg = I.getOperand(2).getReg(); 999 Register RhsReg = I.getOperand(3).getReg(); 1000 CmpInst::Predicate Predicate = 1001 (CmpInst::Predicate)I.getOperand(1).getPredicate(); 1002 1003 // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction. 1004 static const uint16_t SETFOpcTable[2][3] = { 1005 {X86::COND_E, X86::COND_NP, X86::AND8rr}, 1006 {X86::COND_NE, X86::COND_P, X86::OR8rr}}; 1007 const uint16_t *SETFOpc = nullptr; 1008 switch (Predicate) { 1009 default: 1010 break; 1011 case CmpInst::FCMP_OEQ: 1012 SETFOpc = &SETFOpcTable[0][0]; 1013 break; 1014 case CmpInst::FCMP_UNE: 1015 SETFOpc = &SETFOpcTable[1][0]; 1016 break; 1017 } 1018 1019 // Compute the opcode for the CMP instruction. 1020 unsigned OpCmp; 1021 LLT Ty = MRI.getType(LhsReg); 1022 switch (Ty.getSizeInBits()) { 1023 default: 1024 return false; 1025 case 32: 1026 OpCmp = X86::UCOMISSrr; 1027 break; 1028 case 64: 1029 OpCmp = X86::UCOMISDrr; 1030 break; 1031 } 1032 1033 Register ResultReg = I.getOperand(0).getReg(); 1034 RBI.constrainGenericRegister( 1035 ResultReg, 1036 *getRegClass(LLT::scalar(8), *RBI.getRegBank(ResultReg, MRI, TRI)), MRI); 1037 if (SETFOpc) { 1038 MachineInstr &CmpInst = 1039 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp)) 1040 .addReg(LhsReg) 1041 .addReg(RhsReg); 1042 1043 Register FlagReg1 = MRI.createVirtualRegister(&X86::GR8RegClass); 1044 Register FlagReg2 = MRI.createVirtualRegister(&X86::GR8RegClass); 1045 MachineInstr &Set1 = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1046 TII.get(X86::SETCCr), FlagReg1).addImm(SETFOpc[0]); 1047 MachineInstr &Set2 = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1048 TII.get(X86::SETCCr), FlagReg2).addImm(SETFOpc[1]); 1049 MachineInstr &Set3 = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1050 TII.get(SETFOpc[2]), ResultReg) 1051 .addReg(FlagReg1) 1052 .addReg(FlagReg2); 1053 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI); 1054 constrainSelectedInstRegOperands(Set1, TII, TRI, RBI); 1055 constrainSelectedInstRegOperands(Set2, TII, TRI, RBI); 1056 constrainSelectedInstRegOperands(Set3, TII, TRI, RBI); 1057 1058 I.eraseFromParent(); 1059 return true; 1060 } 1061 1062 X86::CondCode CC; 1063 bool SwapArgs; 1064 std::tie(CC, SwapArgs) = X86::getX86ConditionCode(Predicate); 1065 assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code."); 1066 1067 if (SwapArgs) 1068 std::swap(LhsReg, RhsReg); 1069 1070 // Emit a compare of LHS/RHS. 1071 MachineInstr &CmpInst = 1072 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp)) 1073 .addReg(LhsReg) 1074 .addReg(RhsReg); 1075 1076 MachineInstr &Set = 1077 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::SETCCr), ResultReg).addImm(CC); 1078 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI); 1079 constrainSelectedInstRegOperands(Set, TII, TRI, RBI); 1080 I.eraseFromParent(); 1081 return true; 1082 } 1083 1084 bool X86InstructionSelector::selectUadde(MachineInstr &I, 1085 MachineRegisterInfo &MRI, 1086 MachineFunction &MF) const { 1087 assert((I.getOpcode() == TargetOpcode::G_UADDE) && "unexpected instruction"); 1088 1089 const Register DstReg = I.getOperand(0).getReg(); 1090 const Register CarryOutReg = I.getOperand(1).getReg(); 1091 const Register Op0Reg = I.getOperand(2).getReg(); 1092 const Register Op1Reg = I.getOperand(3).getReg(); 1093 Register CarryInReg = I.getOperand(4).getReg(); 1094 1095 const LLT DstTy = MRI.getType(DstReg); 1096 1097 if (DstTy != LLT::scalar(32)) 1098 return false; 1099 1100 // find CarryIn def instruction. 1101 MachineInstr *Def = MRI.getVRegDef(CarryInReg); 1102 while (Def->getOpcode() == TargetOpcode::G_TRUNC) { 1103 CarryInReg = Def->getOperand(1).getReg(); 1104 Def = MRI.getVRegDef(CarryInReg); 1105 } 1106 1107 unsigned Opcode; 1108 if (Def->getOpcode() == TargetOpcode::G_UADDE) { 1109 // carry set by prev ADD. 1110 1111 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), X86::EFLAGS) 1112 .addReg(CarryInReg); 1113 1114 if (!RBI.constrainGenericRegister(CarryInReg, X86::GR32RegClass, MRI)) 1115 return false; 1116 1117 Opcode = X86::ADC32rr; 1118 } else if (auto val = getConstantVRegVal(CarryInReg, MRI)) { 1119 // carry is constant, support only 0. 1120 if (*val != 0) 1121 return false; 1122 1123 Opcode = X86::ADD32rr; 1124 } else 1125 return false; 1126 1127 MachineInstr &AddInst = 1128 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode), DstReg) 1129 .addReg(Op0Reg) 1130 .addReg(Op1Reg); 1131 1132 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), CarryOutReg) 1133 .addReg(X86::EFLAGS); 1134 1135 if (!constrainSelectedInstRegOperands(AddInst, TII, TRI, RBI) || 1136 !RBI.constrainGenericRegister(CarryOutReg, X86::GR32RegClass, MRI)) 1137 return false; 1138 1139 I.eraseFromParent(); 1140 return true; 1141 } 1142 1143 bool X86InstructionSelector::selectExtract(MachineInstr &I, 1144 MachineRegisterInfo &MRI, 1145 MachineFunction &MF) const { 1146 assert((I.getOpcode() == TargetOpcode::G_EXTRACT) && 1147 "unexpected instruction"); 1148 1149 const Register DstReg = I.getOperand(0).getReg(); 1150 const Register SrcReg = I.getOperand(1).getReg(); 1151 int64_t Index = I.getOperand(2).getImm(); 1152 1153 const LLT DstTy = MRI.getType(DstReg); 1154 const LLT SrcTy = MRI.getType(SrcReg); 1155 1156 // Meanwile handle vector type only. 1157 if (!DstTy.isVector()) 1158 return false; 1159 1160 if (Index % DstTy.getSizeInBits() != 0) 1161 return false; // Not extract subvector. 1162 1163 if (Index == 0) { 1164 // Replace by extract subreg copy. 1165 if (!emitExtractSubreg(DstReg, SrcReg, I, MRI, MF)) 1166 return false; 1167 1168 I.eraseFromParent(); 1169 return true; 1170 } 1171 1172 bool HasAVX = STI.hasAVX(); 1173 bool HasAVX512 = STI.hasAVX512(); 1174 bool HasVLX = STI.hasVLX(); 1175 1176 if (SrcTy.getSizeInBits() == 256 && DstTy.getSizeInBits() == 128) { 1177 if (HasVLX) 1178 I.setDesc(TII.get(X86::VEXTRACTF32x4Z256rr)); 1179 else if (HasAVX) 1180 I.setDesc(TII.get(X86::VEXTRACTF128rr)); 1181 else 1182 return false; 1183 } else if (SrcTy.getSizeInBits() == 512 && HasAVX512) { 1184 if (DstTy.getSizeInBits() == 128) 1185 I.setDesc(TII.get(X86::VEXTRACTF32x4Zrr)); 1186 else if (DstTy.getSizeInBits() == 256) 1187 I.setDesc(TII.get(X86::VEXTRACTF64x4Zrr)); 1188 else 1189 return false; 1190 } else 1191 return false; 1192 1193 // Convert to X86 VEXTRACT immediate. 1194 Index = Index / DstTy.getSizeInBits(); 1195 I.getOperand(2).setImm(Index); 1196 1197 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 1198 } 1199 1200 bool X86InstructionSelector::emitExtractSubreg(unsigned DstReg, unsigned SrcReg, 1201 MachineInstr &I, 1202 MachineRegisterInfo &MRI, 1203 MachineFunction &MF) const { 1204 const LLT DstTy = MRI.getType(DstReg); 1205 const LLT SrcTy = MRI.getType(SrcReg); 1206 unsigned SubIdx = X86::NoSubRegister; 1207 1208 if (!DstTy.isVector() || !SrcTy.isVector()) 1209 return false; 1210 1211 assert(SrcTy.getSizeInBits() > DstTy.getSizeInBits() && 1212 "Incorrect Src/Dst register size"); 1213 1214 if (DstTy.getSizeInBits() == 128) 1215 SubIdx = X86::sub_xmm; 1216 else if (DstTy.getSizeInBits() == 256) 1217 SubIdx = X86::sub_ymm; 1218 else 1219 return false; 1220 1221 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstReg, MRI); 1222 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcReg, MRI); 1223 1224 SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubIdx); 1225 1226 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 1227 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 1228 LLVM_DEBUG(dbgs() << "Failed to constrain G_TRUNC\n"); 1229 return false; 1230 } 1231 1232 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), DstReg) 1233 .addReg(SrcReg, 0, SubIdx); 1234 1235 return true; 1236 } 1237 1238 bool X86InstructionSelector::emitInsertSubreg(unsigned DstReg, unsigned SrcReg, 1239 MachineInstr &I, 1240 MachineRegisterInfo &MRI, 1241 MachineFunction &MF) const { 1242 const LLT DstTy = MRI.getType(DstReg); 1243 const LLT SrcTy = MRI.getType(SrcReg); 1244 unsigned SubIdx = X86::NoSubRegister; 1245 1246 // TODO: support scalar types 1247 if (!DstTy.isVector() || !SrcTy.isVector()) 1248 return false; 1249 1250 assert(SrcTy.getSizeInBits() < DstTy.getSizeInBits() && 1251 "Incorrect Src/Dst register size"); 1252 1253 if (SrcTy.getSizeInBits() == 128) 1254 SubIdx = X86::sub_xmm; 1255 else if (SrcTy.getSizeInBits() == 256) 1256 SubIdx = X86::sub_ymm; 1257 else 1258 return false; 1259 1260 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcReg, MRI); 1261 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstReg, MRI); 1262 1263 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) || 1264 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) { 1265 LLVM_DEBUG(dbgs() << "Failed to constrain INSERT_SUBREG\n"); 1266 return false; 1267 } 1268 1269 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY)) 1270 .addReg(DstReg, RegState::DefineNoRead, SubIdx) 1271 .addReg(SrcReg); 1272 1273 return true; 1274 } 1275 1276 bool X86InstructionSelector::selectInsert(MachineInstr &I, 1277 MachineRegisterInfo &MRI, 1278 MachineFunction &MF) const { 1279 assert((I.getOpcode() == TargetOpcode::G_INSERT) && "unexpected instruction"); 1280 1281 const Register DstReg = I.getOperand(0).getReg(); 1282 const Register SrcReg = I.getOperand(1).getReg(); 1283 const Register InsertReg = I.getOperand(2).getReg(); 1284 int64_t Index = I.getOperand(3).getImm(); 1285 1286 const LLT DstTy = MRI.getType(DstReg); 1287 const LLT InsertRegTy = MRI.getType(InsertReg); 1288 1289 // Meanwile handle vector type only. 1290 if (!DstTy.isVector()) 1291 return false; 1292 1293 if (Index % InsertRegTy.getSizeInBits() != 0) 1294 return false; // Not insert subvector. 1295 1296 if (Index == 0 && MRI.getVRegDef(SrcReg)->isImplicitDef()) { 1297 // Replace by subreg copy. 1298 if (!emitInsertSubreg(DstReg, InsertReg, I, MRI, MF)) 1299 return false; 1300 1301 I.eraseFromParent(); 1302 return true; 1303 } 1304 1305 bool HasAVX = STI.hasAVX(); 1306 bool HasAVX512 = STI.hasAVX512(); 1307 bool HasVLX = STI.hasVLX(); 1308 1309 if (DstTy.getSizeInBits() == 256 && InsertRegTy.getSizeInBits() == 128) { 1310 if (HasVLX) 1311 I.setDesc(TII.get(X86::VINSERTF32x4Z256rr)); 1312 else if (HasAVX) 1313 I.setDesc(TII.get(X86::VINSERTF128rr)); 1314 else 1315 return false; 1316 } else if (DstTy.getSizeInBits() == 512 && HasAVX512) { 1317 if (InsertRegTy.getSizeInBits() == 128) 1318 I.setDesc(TII.get(X86::VINSERTF32x4Zrr)); 1319 else if (InsertRegTy.getSizeInBits() == 256) 1320 I.setDesc(TII.get(X86::VINSERTF64x4Zrr)); 1321 else 1322 return false; 1323 } else 1324 return false; 1325 1326 // Convert to X86 VINSERT immediate. 1327 Index = Index / InsertRegTy.getSizeInBits(); 1328 1329 I.getOperand(3).setImm(Index); 1330 1331 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 1332 } 1333 1334 bool X86InstructionSelector::selectUnmergeValues( 1335 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) { 1336 assert((I.getOpcode() == TargetOpcode::G_UNMERGE_VALUES) && 1337 "unexpected instruction"); 1338 1339 // Split to extracts. 1340 unsigned NumDefs = I.getNumOperands() - 1; 1341 Register SrcReg = I.getOperand(NumDefs).getReg(); 1342 unsigned DefSize = MRI.getType(I.getOperand(0).getReg()).getSizeInBits(); 1343 1344 for (unsigned Idx = 0; Idx < NumDefs; ++Idx) { 1345 MachineInstr &ExtrInst = 1346 *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1347 TII.get(TargetOpcode::G_EXTRACT), I.getOperand(Idx).getReg()) 1348 .addReg(SrcReg) 1349 .addImm(Idx * DefSize); 1350 1351 if (!select(ExtrInst)) 1352 return false; 1353 } 1354 1355 I.eraseFromParent(); 1356 return true; 1357 } 1358 1359 bool X86InstructionSelector::selectMergeValues( 1360 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) { 1361 assert((I.getOpcode() == TargetOpcode::G_MERGE_VALUES || 1362 I.getOpcode() == TargetOpcode::G_CONCAT_VECTORS) && 1363 "unexpected instruction"); 1364 1365 // Split to inserts. 1366 Register DstReg = I.getOperand(0).getReg(); 1367 Register SrcReg0 = I.getOperand(1).getReg(); 1368 1369 const LLT DstTy = MRI.getType(DstReg); 1370 const LLT SrcTy = MRI.getType(SrcReg0); 1371 unsigned SrcSize = SrcTy.getSizeInBits(); 1372 1373 const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI); 1374 1375 // For the first src use insertSubReg. 1376 Register DefReg = MRI.createGenericVirtualRegister(DstTy); 1377 MRI.setRegBank(DefReg, RegBank); 1378 if (!emitInsertSubreg(DefReg, I.getOperand(1).getReg(), I, MRI, MF)) 1379 return false; 1380 1381 for (unsigned Idx = 2; Idx < I.getNumOperands(); ++Idx) { 1382 Register Tmp = MRI.createGenericVirtualRegister(DstTy); 1383 MRI.setRegBank(Tmp, RegBank); 1384 1385 MachineInstr &InsertInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1386 TII.get(TargetOpcode::G_INSERT), Tmp) 1387 .addReg(DefReg) 1388 .addReg(I.getOperand(Idx).getReg()) 1389 .addImm((Idx - 1) * SrcSize); 1390 1391 DefReg = Tmp; 1392 1393 if (!select(InsertInst)) 1394 return false; 1395 } 1396 1397 MachineInstr &CopyInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(), 1398 TII.get(TargetOpcode::COPY), DstReg) 1399 .addReg(DefReg); 1400 1401 if (!select(CopyInst)) 1402 return false; 1403 1404 I.eraseFromParent(); 1405 return true; 1406 } 1407 1408 bool X86InstructionSelector::selectCondBranch(MachineInstr &I, 1409 MachineRegisterInfo &MRI, 1410 MachineFunction &MF) const { 1411 assert((I.getOpcode() == TargetOpcode::G_BRCOND) && "unexpected instruction"); 1412 1413 const Register CondReg = I.getOperand(0).getReg(); 1414 MachineBasicBlock *DestMBB = I.getOperand(1).getMBB(); 1415 1416 MachineInstr &TestInst = 1417 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::TEST8ri)) 1418 .addReg(CondReg) 1419 .addImm(1); 1420 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::JCC_1)) 1421 .addMBB(DestMBB).addImm(X86::COND_NE); 1422 1423 constrainSelectedInstRegOperands(TestInst, TII, TRI, RBI); 1424 1425 I.eraseFromParent(); 1426 return true; 1427 } 1428 1429 bool X86InstructionSelector::materializeFP(MachineInstr &I, 1430 MachineRegisterInfo &MRI, 1431 MachineFunction &MF) const { 1432 assert((I.getOpcode() == TargetOpcode::G_FCONSTANT) && 1433 "unexpected instruction"); 1434 1435 // Can't handle alternate code models yet. 1436 CodeModel::Model CM = TM.getCodeModel(); 1437 if (CM != CodeModel::Small && CM != CodeModel::Large) 1438 return false; 1439 1440 const Register DstReg = I.getOperand(0).getReg(); 1441 const LLT DstTy = MRI.getType(DstReg); 1442 const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI); 1443 unsigned Align = DstTy.getSizeInBits(); 1444 const DebugLoc &DbgLoc = I.getDebugLoc(); 1445 1446 unsigned Opc = getLoadStoreOp(DstTy, RegBank, TargetOpcode::G_LOAD, Align); 1447 1448 // Create the load from the constant pool. 1449 const ConstantFP *CFP = I.getOperand(1).getFPImm(); 1450 unsigned CPI = MF.getConstantPool()->getConstantPoolIndex(CFP, Align); 1451 MachineInstr *LoadInst = nullptr; 1452 unsigned char OpFlag = STI.classifyLocalReference(nullptr); 1453 1454 if (CM == CodeModel::Large && STI.is64Bit()) { 1455 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so 1456 // they cannot be folded into immediate fields. 1457 1458 Register AddrReg = MRI.createVirtualRegister(&X86::GR64RegClass); 1459 BuildMI(*I.getParent(), I, DbgLoc, TII.get(X86::MOV64ri), AddrReg) 1460 .addConstantPoolIndex(CPI, 0, OpFlag); 1461 1462 MachineMemOperand *MMO = MF.getMachineMemOperand( 1463 MachinePointerInfo::getConstantPool(MF), MachineMemOperand::MOLoad, 1464 MF.getDataLayout().getPointerSize(), Align); 1465 1466 LoadInst = 1467 addDirectMem(BuildMI(*I.getParent(), I, DbgLoc, TII.get(Opc), DstReg), 1468 AddrReg) 1469 .addMemOperand(MMO); 1470 1471 } else if (CM == CodeModel::Small || !STI.is64Bit()) { 1472 // Handle the case when globals fit in our immediate field. 1473 // This is true for X86-32 always and X86-64 when in -mcmodel=small mode. 1474 1475 // x86-32 PIC requires a PIC base register for constant pools. 1476 unsigned PICBase = 0; 1477 if (OpFlag == X86II::MO_PIC_BASE_OFFSET || OpFlag == X86II::MO_GOTOFF) { 1478 // PICBase can be allocated by TII.getGlobalBaseReg(&MF). 1479 // In DAGISEL the code that initialize it generated by the CGBR pass. 1480 return false; // TODO support the mode. 1481 } else if (STI.is64Bit() && TM.getCodeModel() == CodeModel::Small) 1482 PICBase = X86::RIP; 1483 1484 LoadInst = addConstantPoolReference( 1485 BuildMI(*I.getParent(), I, DbgLoc, TII.get(Opc), DstReg), CPI, PICBase, 1486 OpFlag); 1487 } else 1488 return false; 1489 1490 constrainSelectedInstRegOperands(*LoadInst, TII, TRI, RBI); 1491 I.eraseFromParent(); 1492 return true; 1493 } 1494 1495 bool X86InstructionSelector::selectImplicitDefOrPHI( 1496 MachineInstr &I, MachineRegisterInfo &MRI) const { 1497 assert((I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF || 1498 I.getOpcode() == TargetOpcode::G_PHI) && 1499 "unexpected instruction"); 1500 1501 Register DstReg = I.getOperand(0).getReg(); 1502 1503 if (!MRI.getRegClassOrNull(DstReg)) { 1504 const LLT DstTy = MRI.getType(DstReg); 1505 const TargetRegisterClass *RC = getRegClass(DstTy, DstReg, MRI); 1506 1507 if (!RBI.constrainGenericRegister(DstReg, *RC, MRI)) { 1508 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 1509 << " operand\n"); 1510 return false; 1511 } 1512 } 1513 1514 if (I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF) 1515 I.setDesc(TII.get(X86::IMPLICIT_DEF)); 1516 else 1517 I.setDesc(TII.get(X86::PHI)); 1518 1519 return true; 1520 } 1521 1522 // Currently GlobalIsel TableGen generates patterns for shift imm and shift 1, 1523 // but with shiftCount i8. In G_LSHR/G_ASHR/G_SHL like LLVM-IR both arguments 1524 // has the same type, so for now only shift i8 can use auto generated 1525 // TableGen patterns. 1526 bool X86InstructionSelector::selectShift(MachineInstr &I, 1527 MachineRegisterInfo &MRI, 1528 MachineFunction &MF) const { 1529 1530 assert((I.getOpcode() == TargetOpcode::G_SHL || 1531 I.getOpcode() == TargetOpcode::G_ASHR || 1532 I.getOpcode() == TargetOpcode::G_LSHR) && 1533 "unexpected instruction"); 1534 1535 Register DstReg = I.getOperand(0).getReg(); 1536 const LLT DstTy = MRI.getType(DstReg); 1537 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI); 1538 1539 const static struct ShiftEntry { 1540 unsigned SizeInBits; 1541 unsigned OpLSHR; 1542 unsigned OpASHR; 1543 unsigned OpSHL; 1544 } OpTable[] = { 1545 {8, X86::SHR8rCL, X86::SAR8rCL, X86::SHL8rCL}, // i8 1546 {16, X86::SHR16rCL, X86::SAR16rCL, X86::SHL16rCL}, // i16 1547 {32, X86::SHR32rCL, X86::SAR32rCL, X86::SHL32rCL}, // i32 1548 {64, X86::SHR64rCL, X86::SAR64rCL, X86::SHL64rCL} // i64 1549 }; 1550 1551 if (DstRB.getID() != X86::GPRRegBankID) 1552 return false; 1553 1554 auto ShiftEntryIt = std::find_if( 1555 std::begin(OpTable), std::end(OpTable), [DstTy](const ShiftEntry &El) { 1556 return El.SizeInBits == DstTy.getSizeInBits(); 1557 }); 1558 if (ShiftEntryIt == std::end(OpTable)) 1559 return false; 1560 1561 unsigned Opcode = 0; 1562 switch (I.getOpcode()) { 1563 case TargetOpcode::G_SHL: 1564 Opcode = ShiftEntryIt->OpSHL; 1565 break; 1566 case TargetOpcode::G_ASHR: 1567 Opcode = ShiftEntryIt->OpASHR; 1568 break; 1569 case TargetOpcode::G_LSHR: 1570 Opcode = ShiftEntryIt->OpLSHR; 1571 break; 1572 default: 1573 return false; 1574 } 1575 1576 Register Op0Reg = I.getOperand(1).getReg(); 1577 Register Op1Reg = I.getOperand(2).getReg(); 1578 1579 assert(MRI.getType(Op1Reg).getSizeInBits() == 8); 1580 1581 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::COPY), 1582 X86::CL) 1583 .addReg(Op1Reg); 1584 1585 MachineInstr &ShiftInst = 1586 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode), DstReg) 1587 .addReg(Op0Reg); 1588 1589 constrainSelectedInstRegOperands(ShiftInst, TII, TRI, RBI); 1590 I.eraseFromParent(); 1591 return true; 1592 } 1593 1594 bool X86InstructionSelector::selectDivRem(MachineInstr &I, 1595 MachineRegisterInfo &MRI, 1596 MachineFunction &MF) const { 1597 // The implementation of this function is taken from X86FastISel. 1598 assert((I.getOpcode() == TargetOpcode::G_SDIV || 1599 I.getOpcode() == TargetOpcode::G_SREM || 1600 I.getOpcode() == TargetOpcode::G_UDIV || 1601 I.getOpcode() == TargetOpcode::G_UREM) && 1602 "unexpected instruction"); 1603 1604 const Register DstReg = I.getOperand(0).getReg(); 1605 const Register Op1Reg = I.getOperand(1).getReg(); 1606 const Register Op2Reg = I.getOperand(2).getReg(); 1607 1608 const LLT RegTy = MRI.getType(DstReg); 1609 assert(RegTy == MRI.getType(Op1Reg) && RegTy == MRI.getType(Op2Reg) && 1610 "Arguments and return value types must match"); 1611 1612 const RegisterBank *RegRB = RBI.getRegBank(DstReg, MRI, TRI); 1613 if (!RegRB || RegRB->getID() != X86::GPRRegBankID) 1614 return false; 1615 1616 const static unsigned NumTypes = 4; // i8, i16, i32, i64 1617 const static unsigned NumOps = 4; // SDiv, SRem, UDiv, URem 1618 const static bool S = true; // IsSigned 1619 const static bool U = false; // !IsSigned 1620 const static unsigned Copy = TargetOpcode::COPY; 1621 // For the X86 IDIV instruction, in most cases the dividend 1622 // (numerator) must be in a specific register pair highreg:lowreg, 1623 // producing the quotient in lowreg and the remainder in highreg. 1624 // For most data types, to set up the instruction, the dividend is 1625 // copied into lowreg, and lowreg is sign-extended into highreg. The 1626 // exception is i8, where the dividend is defined as a single register rather 1627 // than a register pair, and we therefore directly sign-extend the dividend 1628 // into lowreg, instead of copying, and ignore the highreg. 1629 const static struct DivRemEntry { 1630 // The following portion depends only on the data type. 1631 unsigned SizeInBits; 1632 unsigned LowInReg; // low part of the register pair 1633 unsigned HighInReg; // high part of the register pair 1634 // The following portion depends on both the data type and the operation. 1635 struct DivRemResult { 1636 unsigned OpDivRem; // The specific DIV/IDIV opcode to use. 1637 unsigned OpSignExtend; // Opcode for sign-extending lowreg into 1638 // highreg, or copying a zero into highreg. 1639 unsigned OpCopy; // Opcode for copying dividend into lowreg, or 1640 // zero/sign-extending into lowreg for i8. 1641 unsigned DivRemResultReg; // Register containing the desired result. 1642 bool IsOpSigned; // Whether to use signed or unsigned form. 1643 } ResultTable[NumOps]; 1644 } OpTable[NumTypes] = { 1645 {8, 1646 X86::AX, 1647 0, 1648 { 1649 {X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AL, S}, // SDiv 1650 {X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AH, S}, // SRem 1651 {X86::DIV8r, 0, X86::MOVZX16rr8, X86::AL, U}, // UDiv 1652 {X86::DIV8r, 0, X86::MOVZX16rr8, X86::AH, U}, // URem 1653 }}, // i8 1654 {16, 1655 X86::AX, 1656 X86::DX, 1657 { 1658 {X86::IDIV16r, X86::CWD, Copy, X86::AX, S}, // SDiv 1659 {X86::IDIV16r, X86::CWD, Copy, X86::DX, S}, // SRem 1660 {X86::DIV16r, X86::MOV32r0, Copy, X86::AX, U}, // UDiv 1661 {X86::DIV16r, X86::MOV32r0, Copy, X86::DX, U}, // URem 1662 }}, // i16 1663 {32, 1664 X86::EAX, 1665 X86::EDX, 1666 { 1667 {X86::IDIV32r, X86::CDQ, Copy, X86::EAX, S}, // SDiv 1668 {X86::IDIV32r, X86::CDQ, Copy, X86::EDX, S}, // SRem 1669 {X86::DIV32r, X86::MOV32r0, Copy, X86::EAX, U}, // UDiv 1670 {X86::DIV32r, X86::MOV32r0, Copy, X86::EDX, U}, // URem 1671 }}, // i32 1672 {64, 1673 X86::RAX, 1674 X86::RDX, 1675 { 1676 {X86::IDIV64r, X86::CQO, Copy, X86::RAX, S}, // SDiv 1677 {X86::IDIV64r, X86::CQO, Copy, X86::RDX, S}, // SRem 1678 {X86::DIV64r, X86::MOV32r0, Copy, X86::RAX, U}, // UDiv 1679 {X86::DIV64r, X86::MOV32r0, Copy, X86::RDX, U}, // URem 1680 }}, // i64 1681 }; 1682 1683 auto OpEntryIt = std::find_if(std::begin(OpTable), std::end(OpTable), 1684 [RegTy](const DivRemEntry &El) { 1685 return El.SizeInBits == RegTy.getSizeInBits(); 1686 }); 1687 if (OpEntryIt == std::end(OpTable)) 1688 return false; 1689 1690 unsigned OpIndex; 1691 switch (I.getOpcode()) { 1692 default: 1693 llvm_unreachable("Unexpected div/rem opcode"); 1694 case TargetOpcode::G_SDIV: 1695 OpIndex = 0; 1696 break; 1697 case TargetOpcode::G_SREM: 1698 OpIndex = 1; 1699 break; 1700 case TargetOpcode::G_UDIV: 1701 OpIndex = 2; 1702 break; 1703 case TargetOpcode::G_UREM: 1704 OpIndex = 3; 1705 break; 1706 } 1707 1708 const DivRemEntry &TypeEntry = *OpEntryIt; 1709 const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex]; 1710 1711 const TargetRegisterClass *RegRC = getRegClass(RegTy, *RegRB); 1712 if (!RBI.constrainGenericRegister(Op1Reg, *RegRC, MRI) || 1713 !RBI.constrainGenericRegister(Op2Reg, *RegRC, MRI) || 1714 !RBI.constrainGenericRegister(DstReg, *RegRC, MRI)) { 1715 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode()) 1716 << " operand\n"); 1717 return false; 1718 } 1719 1720 // Move op1 into low-order input register. 1721 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpEntry.OpCopy), 1722 TypeEntry.LowInReg) 1723 .addReg(Op1Reg); 1724 // Zero-extend or sign-extend into high-order input register. 1725 if (OpEntry.OpSignExtend) { 1726 if (OpEntry.IsOpSigned) 1727 BuildMI(*I.getParent(), I, I.getDebugLoc(), 1728 TII.get(OpEntry.OpSignExtend)); 1729 else { 1730 Register Zero32 = MRI.createVirtualRegister(&X86::GR32RegClass); 1731 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::MOV32r0), 1732 Zero32); 1733 1734 // Copy the zero into the appropriate sub/super/identical physical 1735 // register. Unfortunately the operations needed are not uniform enough 1736 // to fit neatly into the table above. 1737 if (RegTy.getSizeInBits() == 16) { 1738 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy), 1739 TypeEntry.HighInReg) 1740 .addReg(Zero32, 0, X86::sub_16bit); 1741 } else if (RegTy.getSizeInBits() == 32) { 1742 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy), 1743 TypeEntry.HighInReg) 1744 .addReg(Zero32); 1745 } else if (RegTy.getSizeInBits() == 64) { 1746 BuildMI(*I.getParent(), I, I.getDebugLoc(), 1747 TII.get(TargetOpcode::SUBREG_TO_REG), TypeEntry.HighInReg) 1748 .addImm(0) 1749 .addReg(Zero32) 1750 .addImm(X86::sub_32bit); 1751 } 1752 } 1753 } 1754 // Generate the DIV/IDIV instruction. 1755 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpEntry.OpDivRem)) 1756 .addReg(Op2Reg); 1757 // For i8 remainder, we can't reference ah directly, as we'll end 1758 // up with bogus copies like %r9b = COPY %ah. Reference ax 1759 // instead to prevent ah references in a rex instruction. 1760 // 1761 // The current assumption of the fast register allocator is that isel 1762 // won't generate explicit references to the GR8_NOREX registers. If 1763 // the allocator and/or the backend get enhanced to be more robust in 1764 // that regard, this can be, and should be, removed. 1765 if ((I.getOpcode() == Instruction::SRem || 1766 I.getOpcode() == Instruction::URem) && 1767 OpEntry.DivRemResultReg == X86::AH && STI.is64Bit()) { 1768 Register SourceSuperReg = MRI.createVirtualRegister(&X86::GR16RegClass); 1769 Register ResultSuperReg = MRI.createVirtualRegister(&X86::GR16RegClass); 1770 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy), SourceSuperReg) 1771 .addReg(X86::AX); 1772 1773 // Shift AX right by 8 bits instead of using AH. 1774 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::SHR16ri), 1775 ResultSuperReg) 1776 .addReg(SourceSuperReg) 1777 .addImm(8); 1778 1779 // Now reference the 8-bit subreg of the result. 1780 BuildMI(*I.getParent(), I, I.getDebugLoc(), 1781 TII.get(TargetOpcode::SUBREG_TO_REG)) 1782 .addDef(DstReg) 1783 .addImm(0) 1784 .addReg(ResultSuperReg) 1785 .addImm(X86::sub_8bit); 1786 } else { 1787 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::COPY), 1788 DstReg) 1789 .addReg(OpEntry.DivRemResultReg); 1790 } 1791 I.eraseFromParent(); 1792 return true; 1793 } 1794 1795 bool X86InstructionSelector::selectIntrinsicWSideEffects( 1796 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) const { 1797 1798 assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS && 1799 "unexpected instruction"); 1800 1801 if (I.getOperand(0).getIntrinsicID() != Intrinsic::trap) 1802 return false; 1803 1804 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::TRAP)); 1805 1806 I.eraseFromParent(); 1807 return true; 1808 } 1809 1810 InstructionSelector * 1811 llvm::createX86InstructionSelector(const X86TargetMachine &TM, 1812 X86Subtarget &Subtarget, 1813 X86RegisterBankInfo &RBI) { 1814 return new X86InstructionSelector(TM, Subtarget, RBI); 1815 } 1816