1//===-- ARMBaseInstrInfo.cpp - ARM Instruction Information ----------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file contains the Base ARM implementation of the TargetInstrInfo class. 11// 12//===----------------------------------------------------------------------===// 13 14#include "ARM.h" 15#include "ARMBaseInstrInfo.h" 16#include "ARMBaseRegisterInfo.h" 17#include "ARMConstantPoolValue.h" 18#include "ARMFeatures.h" 19#include "ARMHazardRecognizer.h" 20#include "ARMMachineFunctionInfo.h" 21#include "MCTargetDesc/ARMAddressingModes.h" 22#include "llvm/ADT/STLExtras.h" 23#include "llvm/CodeGen/LiveVariables.h" 24#include "llvm/CodeGen/MachineConstantPool.h" 25#include "llvm/CodeGen/MachineFrameInfo.h" 26#include "llvm/CodeGen/MachineInstrBuilder.h" 27#include "llvm/CodeGen/MachineJumpTableInfo.h" 28#include "llvm/CodeGen/MachineMemOperand.h" 29#include "llvm/CodeGen/MachineRegisterInfo.h" 30#include "llvm/CodeGen/SelectionDAGNodes.h" 31#include "llvm/IR/Constants.h" 32#include "llvm/IR/Function.h" 33#include "llvm/IR/GlobalValue.h" 34#include "llvm/MC/MCAsmInfo.h" 35#include "llvm/MC/MCExpr.h" 36#include "llvm/Support/BranchProbability.h" 37#include "llvm/Support/CommandLine.h" 38#include "llvm/Support/Debug.h" 39#include "llvm/Support/ErrorHandling.h" 40#include "llvm/Support/raw_ostream.h" 41 42using namespace llvm; 43 44#define DEBUG_TYPE "arm-instrinfo" 45 46#define GET_INSTRINFO_CTOR_DTOR 47#include "ARMGenInstrInfo.inc" 48 49static cl::opt<bool> 50EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden, 51 cl::desc("Enable ARM 2-addr to 3-addr conv")); 52 53static cl::opt<bool> 54WidenVMOVS("widen-vmovs", cl::Hidden, cl::init(true), 55 cl::desc("Widen ARM vmovs to vmovd when possible")); 56 57static cl::opt<unsigned> 58SwiftPartialUpdateClearance("swift-partial-update-clearance", 59 cl::Hidden, cl::init(12), 60 cl::desc("Clearance before partial register updates")); 61 62/// ARM_MLxEntry - Record information about MLA / MLS instructions. 63struct ARM_MLxEntry { 64 uint16_t MLxOpc; // MLA / MLS opcode 65 uint16_t MulOpc; // Expanded multiplication opcode 66 uint16_t AddSubOpc; // Expanded add / sub opcode 67 bool NegAcc; // True if the acc is negated before the add / sub. 68 bool HasLane; // True if instruction has an extra "lane" operand. 69}; 70 71static const ARM_MLxEntry ARM_MLxTable[] = { 72 // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane 73 // fp scalar ops 74 { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false }, 75 { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false }, 76 { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false }, 77 { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false }, 78 { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false }, 79 { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false }, 80 { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false }, 81 { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false }, 82 83 // fp SIMD ops 84 { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false }, 85 { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false }, 86 { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false }, 87 { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false }, 88 { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true }, 89 { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true }, 90 { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true }, 91 { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true }, 92}; 93 94ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI) 95 : ARMGenInstrInfo(ARM::ADJCALLSTACKDOWN, ARM::ADJCALLSTACKUP), 96 Subtarget(STI) { 97 for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) { 98 if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second) 99 assert(false && "Duplicated entries?"); 100 MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc); 101 MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc); 102 } 103} 104 105// Use a ScoreboardHazardRecognizer for prepass ARM scheduling. TargetInstrImpl 106// currently defaults to no prepass hazard recognizer. 107ScheduleHazardRecognizer * 108ARMBaseInstrInfo::CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, 109 const ScheduleDAG *DAG) const { 110 if (usePreRAHazardRecognizer()) { 111 const InstrItineraryData *II = 112 static_cast<const ARMSubtarget *>(STI)->getInstrItineraryData(); 113 return new ScoreboardHazardRecognizer(II, DAG, "pre-RA-sched"); 114 } 115 return TargetInstrInfo::CreateTargetHazardRecognizer(STI, DAG); 116} 117 118ScheduleHazardRecognizer *ARMBaseInstrInfo:: 119CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 120 const ScheduleDAG *DAG) const { 121 if (Subtarget.isThumb2() || Subtarget.hasVFP2()) 122 return (ScheduleHazardRecognizer *)new ARMHazardRecognizer(II, DAG); 123 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG); 124} 125 126MachineInstr * 127ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI, 128 MachineBasicBlock::iterator &MBBI, 129 LiveVariables *LV) const { 130 // FIXME: Thumb2 support. 131 132 if (!EnableARM3Addr) 133 return nullptr; 134 135 MachineInstr *MI = MBBI; 136 MachineFunction &MF = *MI->getParent()->getParent(); 137 uint64_t TSFlags = MI->getDesc().TSFlags; 138 bool isPre = false; 139 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) { 140 default: return nullptr; 141 case ARMII::IndexModePre: 142 isPre = true; 143 break; 144 case ARMII::IndexModePost: 145 break; 146 } 147 148 // Try splitting an indexed load/store to an un-indexed one plus an add/sub 149 // operation. 150 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode()); 151 if (MemOpc == 0) 152 return nullptr; 153 154 MachineInstr *UpdateMI = nullptr; 155 MachineInstr *MemMI = nullptr; 156 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask); 157 const MCInstrDesc &MCID = MI->getDesc(); 158 unsigned NumOps = MCID.getNumOperands(); 159 bool isLoad = !MI->mayStore(); 160 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0); 161 const MachineOperand &Base = MI->getOperand(2); 162 const MachineOperand &Offset = MI->getOperand(NumOps-3); 163 unsigned WBReg = WB.getReg(); 164 unsigned BaseReg = Base.getReg(); 165 unsigned OffReg = Offset.getReg(); 166 unsigned OffImm = MI->getOperand(NumOps-2).getImm(); 167 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm(); 168 switch (AddrMode) { 169 default: llvm_unreachable("Unknown indexed op!"); 170 case ARMII::AddrMode2: { 171 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub; 172 unsigned Amt = ARM_AM::getAM2Offset(OffImm); 173 if (OffReg == 0) { 174 if (ARM_AM::getSOImmVal(Amt) == -1) 175 // Can't encode it in a so_imm operand. This transformation will 176 // add more than 1 instruction. Abandon! 177 return nullptr; 178 UpdateMI = BuildMI(MF, MI->getDebugLoc(), 179 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg) 180 .addReg(BaseReg).addImm(Amt) 181 .addImm(Pred).addReg(0).addReg(0); 182 } else if (Amt != 0) { 183 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm); 184 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt); 185 UpdateMI = BuildMI(MF, MI->getDebugLoc(), 186 get(isSub ? ARM::SUBrsi : ARM::ADDrsi), WBReg) 187 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc) 188 .addImm(Pred).addReg(0).addReg(0); 189 } else 190 UpdateMI = BuildMI(MF, MI->getDebugLoc(), 191 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg) 192 .addReg(BaseReg).addReg(OffReg) 193 .addImm(Pred).addReg(0).addReg(0); 194 break; 195 } 196 case ARMII::AddrMode3 : { 197 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub; 198 unsigned Amt = ARM_AM::getAM3Offset(OffImm); 199 if (OffReg == 0) 200 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand. 201 UpdateMI = BuildMI(MF, MI->getDebugLoc(), 202 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg) 203 .addReg(BaseReg).addImm(Amt) 204 .addImm(Pred).addReg(0).addReg(0); 205 else 206 UpdateMI = BuildMI(MF, MI->getDebugLoc(), 207 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg) 208 .addReg(BaseReg).addReg(OffReg) 209 .addImm(Pred).addReg(0).addReg(0); 210 break; 211 } 212 } 213 214 std::vector<MachineInstr*> NewMIs; 215 if (isPre) { 216 if (isLoad) 217 MemMI = BuildMI(MF, MI->getDebugLoc(), 218 get(MemOpc), MI->getOperand(0).getReg()) 219 .addReg(WBReg).addImm(0).addImm(Pred); 220 else 221 MemMI = BuildMI(MF, MI->getDebugLoc(), 222 get(MemOpc)).addReg(MI->getOperand(1).getReg()) 223 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred); 224 NewMIs.push_back(MemMI); 225 NewMIs.push_back(UpdateMI); 226 } else { 227 if (isLoad) 228 MemMI = BuildMI(MF, MI->getDebugLoc(), 229 get(MemOpc), MI->getOperand(0).getReg()) 230 .addReg(BaseReg).addImm(0).addImm(Pred); 231 else 232 MemMI = BuildMI(MF, MI->getDebugLoc(), 233 get(MemOpc)).addReg(MI->getOperand(1).getReg()) 234 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred); 235 if (WB.isDead()) 236 UpdateMI->getOperand(0).setIsDead(); 237 NewMIs.push_back(UpdateMI); 238 NewMIs.push_back(MemMI); 239 } 240 241 // Transfer LiveVariables states, kill / dead info. 242 if (LV) { 243 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 244 MachineOperand &MO = MI->getOperand(i); 245 if (MO.isReg() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) { 246 unsigned Reg = MO.getReg(); 247 248 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg); 249 if (MO.isDef()) { 250 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI; 251 if (MO.isDead()) 252 LV->addVirtualRegisterDead(Reg, NewMI); 253 } 254 if (MO.isUse() && MO.isKill()) { 255 for (unsigned j = 0; j < 2; ++j) { 256 // Look at the two new MI's in reverse order. 257 MachineInstr *NewMI = NewMIs[j]; 258 if (!NewMI->readsRegister(Reg)) 259 continue; 260 LV->addVirtualRegisterKilled(Reg, NewMI); 261 if (VI.removeKill(MI)) 262 VI.Kills.push_back(NewMI); 263 break; 264 } 265 } 266 } 267 } 268 } 269 270 MFI->insert(MBBI, NewMIs[1]); 271 MFI->insert(MBBI, NewMIs[0]); 272 return NewMIs[0]; 273} 274 275// Branch analysis. 276bool 277ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB, 278 MachineBasicBlock *&FBB, 279 SmallVectorImpl<MachineOperand> &Cond, 280 bool AllowModify) const { 281 TBB = nullptr; 282 FBB = nullptr; 283 284 MachineBasicBlock::iterator I = MBB.end(); 285 if (I == MBB.begin()) 286 return false; // Empty blocks are easy. 287 --I; 288 289 // Walk backwards from the end of the basic block until the branch is 290 // analyzed or we give up. 291 while (isPredicated(I) || I->isTerminator() || I->isDebugValue()) { 292 293 // Flag to be raised on unanalyzeable instructions. This is useful in cases 294 // where we want to clean up on the end of the basic block before we bail 295 // out. 296 bool CantAnalyze = false; 297 298 // Skip over DEBUG values and predicated nonterminators. 299 while (I->isDebugValue() || !I->isTerminator()) { 300 if (I == MBB.begin()) 301 return false; 302 --I; 303 } 304 305 if (isIndirectBranchOpcode(I->getOpcode()) || 306 isJumpTableBranchOpcode(I->getOpcode())) { 307 // Indirect branches and jump tables can't be analyzed, but we still want 308 // to clean up any instructions at the tail of the basic block. 309 CantAnalyze = true; 310 } else if (isUncondBranchOpcode(I->getOpcode())) { 311 TBB = I->getOperand(0).getMBB(); 312 } else if (isCondBranchOpcode(I->getOpcode())) { 313 // Bail out if we encounter multiple conditional branches. 314 if (!Cond.empty()) 315 return true; 316 317 assert(!FBB && "FBB should have been null."); 318 FBB = TBB; 319 TBB = I->getOperand(0).getMBB(); 320 Cond.push_back(I->getOperand(1)); 321 Cond.push_back(I->getOperand(2)); 322 } else if (I->isReturn()) { 323 // Returns can't be analyzed, but we should run cleanup. 324 CantAnalyze = !isPredicated(I); 325 } else { 326 // We encountered other unrecognized terminator. Bail out immediately. 327 return true; 328 } 329 330 // Cleanup code - to be run for unpredicated unconditional branches and 331 // returns. 332 if (!isPredicated(I) && 333 (isUncondBranchOpcode(I->getOpcode()) || 334 isIndirectBranchOpcode(I->getOpcode()) || 335 isJumpTableBranchOpcode(I->getOpcode()) || 336 I->isReturn())) { 337 // Forget any previous condition branch information - it no longer applies. 338 Cond.clear(); 339 FBB = nullptr; 340 341 // If we can modify the function, delete everything below this 342 // unconditional branch. 343 if (AllowModify) { 344 MachineBasicBlock::iterator DI = std::next(I); 345 while (DI != MBB.end()) { 346 MachineInstr *InstToDelete = DI; 347 ++DI; 348 InstToDelete->eraseFromParent(); 349 } 350 } 351 } 352 353 if (CantAnalyze) 354 return true; 355 356 if (I == MBB.begin()) 357 return false; 358 359 --I; 360 } 361 362 // We made it past the terminators without bailing out - we must have 363 // analyzed this branch successfully. 364 return false; 365} 366 367 368unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { 369 MachineBasicBlock::iterator I = MBB.end(); 370 if (I == MBB.begin()) return 0; 371 --I; 372 while (I->isDebugValue()) { 373 if (I == MBB.begin()) 374 return 0; 375 --I; 376 } 377 if (!isUncondBranchOpcode(I->getOpcode()) && 378 !isCondBranchOpcode(I->getOpcode())) 379 return 0; 380 381 // Remove the branch. 382 I->eraseFromParent(); 383 384 I = MBB.end(); 385 386 if (I == MBB.begin()) return 1; 387 --I; 388 if (!isCondBranchOpcode(I->getOpcode())) 389 return 1; 390 391 // Remove the branch. 392 I->eraseFromParent(); 393 return 2; 394} 395 396unsigned 397ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, 398 MachineBasicBlock *FBB, 399 const SmallVectorImpl<MachineOperand> &Cond, 400 DebugLoc DL) const { 401 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>(); 402 int BOpc = !AFI->isThumbFunction() 403 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB); 404 int BccOpc = !AFI->isThumbFunction() 405 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc); 406 bool isThumb = AFI->isThumbFunction() || AFI->isThumb2Function(); 407 408 // Shouldn't be a fall through. 409 assert(TBB && "InsertBranch must not be told to insert a fallthrough"); 410 assert((Cond.size() == 2 || Cond.size() == 0) && 411 "ARM branch conditions have two components!"); 412 413 if (!FBB) { 414 if (Cond.empty()) { // Unconditional branch? 415 if (isThumb) 416 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB).addImm(ARMCC::AL).addReg(0); 417 else 418 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB); 419 } else 420 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB) 421 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg()); 422 return 1; 423 } 424 425 // Two-way conditional branch. 426 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB) 427 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg()); 428 if (isThumb) 429 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB).addImm(ARMCC::AL).addReg(0); 430 else 431 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB); 432 return 2; 433} 434 435bool ARMBaseInstrInfo:: 436ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 437 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm(); 438 Cond[0].setImm(ARMCC::getOppositeCondition(CC)); 439 return false; 440} 441 442bool ARMBaseInstrInfo::isPredicated(const MachineInstr *MI) const { 443 if (MI->isBundle()) { 444 MachineBasicBlock::const_instr_iterator I = MI; 445 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end(); 446 while (++I != E && I->isInsideBundle()) { 447 int PIdx = I->findFirstPredOperandIdx(); 448 if (PIdx != -1 && I->getOperand(PIdx).getImm() != ARMCC::AL) 449 return true; 450 } 451 return false; 452 } 453 454 int PIdx = MI->findFirstPredOperandIdx(); 455 return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL; 456} 457 458bool ARMBaseInstrInfo:: 459PredicateInstruction(MachineInstr *MI, 460 const SmallVectorImpl<MachineOperand> &Pred) const { 461 unsigned Opc = MI->getOpcode(); 462 if (isUncondBranchOpcode(Opc)) { 463 MI->setDesc(get(getMatchingCondBranchOpcode(Opc))); 464 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 465 .addImm(Pred[0].getImm()) 466 .addReg(Pred[1].getReg()); 467 return true; 468 } 469 470 int PIdx = MI->findFirstPredOperandIdx(); 471 if (PIdx != -1) { 472 MachineOperand &PMO = MI->getOperand(PIdx); 473 PMO.setImm(Pred[0].getImm()); 474 MI->getOperand(PIdx+1).setReg(Pred[1].getReg()); 475 return true; 476 } 477 return false; 478} 479 480bool ARMBaseInstrInfo:: 481SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1, 482 const SmallVectorImpl<MachineOperand> &Pred2) const { 483 if (Pred1.size() > 2 || Pred2.size() > 2) 484 return false; 485 486 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm(); 487 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm(); 488 if (CC1 == CC2) 489 return true; 490 491 switch (CC1) { 492 default: 493 return false; 494 case ARMCC::AL: 495 return true; 496 case ARMCC::HS: 497 return CC2 == ARMCC::HI; 498 case ARMCC::LS: 499 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ; 500 case ARMCC::GE: 501 return CC2 == ARMCC::GT; 502 case ARMCC::LE: 503 return CC2 == ARMCC::LT; 504 } 505} 506 507bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI, 508 std::vector<MachineOperand> &Pred) const { 509 bool Found = false; 510 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 511 const MachineOperand &MO = MI->getOperand(i); 512 if ((MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) || 513 (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR)) { 514 Pred.push_back(MO); 515 Found = true; 516 } 517 } 518 519 return Found; 520} 521 522static bool isCPSRDefined(const MachineInstr *MI) { 523 for (const auto &MO : MI->operands()) 524 if (MO.isReg() && MO.getReg() == ARM::CPSR && MO.isDef()) 525 return true; 526 return false; 527} 528 529static bool isEligibleForITBlock(const MachineInstr *MI) { 530 switch (MI->getOpcode()) { 531 default: return true; 532 case ARM::tADC: // ADC (register) T1 533 case ARM::tADDi3: // ADD (immediate) T1 534 case ARM::tADDi8: // ADD (immediate) T2 535 case ARM::tADDrr: // ADD (register) T1 536 case ARM::tAND: // AND (register) T1 537 case ARM::tASRri: // ASR (immediate) T1 538 case ARM::tASRrr: // ASR (register) T1 539 case ARM::tBIC: // BIC (register) T1 540 case ARM::tEOR: // EOR (register) T1 541 case ARM::tLSLri: // LSL (immediate) T1 542 case ARM::tLSLrr: // LSL (register) T1 543 case ARM::tLSRri: // LSR (immediate) T1 544 case ARM::tLSRrr: // LSR (register) T1 545 case ARM::tMUL: // MUL T1 546 case ARM::tMVN: // MVN (register) T1 547 case ARM::tORR: // ORR (register) T1 548 case ARM::tROR: // ROR (register) T1 549 case ARM::tRSB: // RSB (immediate) T1 550 case ARM::tSBC: // SBC (register) T1 551 case ARM::tSUBi3: // SUB (immediate) T1 552 case ARM::tSUBi8: // SUB (immediate) T2 553 case ARM::tSUBrr: // SUB (register) T1 554 return !isCPSRDefined(MI); 555 } 556} 557 558/// isPredicable - Return true if the specified instruction can be predicated. 559/// By default, this returns true for every instruction with a 560/// PredicateOperand. 561bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const { 562 if (!MI->isPredicable()) 563 return false; 564 565 if (!isEligibleForITBlock(MI)) 566 return false; 567 568 ARMFunctionInfo *AFI = 569 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>(); 570 571 if (AFI->isThumb2Function()) { 572 if (getSubtarget().restrictIT()) 573 return isV8EligibleForIT(MI); 574 } else { // non-Thumb 575 if ((MI->getDesc().TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) 576 return false; 577 } 578 579 return true; 580} 581 582namespace llvm { 583template <> bool IsCPSRDead<MachineInstr>(MachineInstr *MI) { 584 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 585 const MachineOperand &MO = MI->getOperand(i); 586 if (!MO.isReg() || MO.isUndef() || MO.isUse()) 587 continue; 588 if (MO.getReg() != ARM::CPSR) 589 continue; 590 if (!MO.isDead()) 591 return false; 592 } 593 // all definitions of CPSR are dead 594 return true; 595} 596} 597 598/// GetInstSize - Return the size of the specified MachineInstr. 599/// 600unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const { 601 const MachineBasicBlock &MBB = *MI->getParent(); 602 const MachineFunction *MF = MBB.getParent(); 603 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo(); 604 605 const MCInstrDesc &MCID = MI->getDesc(); 606 if (MCID.getSize()) 607 return MCID.getSize(); 608 609 // If this machine instr is an inline asm, measure it. 610 if (MI->getOpcode() == ARM::INLINEASM) 611 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI); 612 unsigned Opc = MI->getOpcode(); 613 switch (Opc) { 614 default: 615 // pseudo-instruction sizes are zero. 616 return 0; 617 case TargetOpcode::BUNDLE: 618 return getInstBundleLength(MI); 619 case ARM::MOVi16_ga_pcrel: 620 case ARM::MOVTi16_ga_pcrel: 621 case ARM::t2MOVi16_ga_pcrel: 622 case ARM::t2MOVTi16_ga_pcrel: 623 return 4; 624 case ARM::MOVi32imm: 625 case ARM::t2MOVi32imm: 626 return 8; 627 case ARM::CONSTPOOL_ENTRY: 628 // If this machine instr is a constant pool entry, its size is recorded as 629 // operand #2. 630 return MI->getOperand(2).getImm(); 631 case ARM::Int_eh_sjlj_longjmp: 632 return 16; 633 case ARM::tInt_eh_sjlj_longjmp: 634 return 10; 635 case ARM::Int_eh_sjlj_setjmp: 636 case ARM::Int_eh_sjlj_setjmp_nofp: 637 return 20; 638 case ARM::tInt_eh_sjlj_setjmp: 639 case ARM::t2Int_eh_sjlj_setjmp: 640 case ARM::t2Int_eh_sjlj_setjmp_nofp: 641 return 12; 642 case ARM::BR_JTr: 643 case ARM::BR_JTm: 644 case ARM::BR_JTadd: 645 case ARM::tBR_JTr: 646 case ARM::t2BR_JT: 647 case ARM::t2TBB_JT: 648 case ARM::t2TBH_JT: { 649 // These are jumptable branches, i.e. a branch followed by an inlined 650 // jumptable. The size is 4 + 4 * number of entries. For TBB, each 651 // entry is one byte; TBH two byte each. 652 unsigned EntrySize = (Opc == ARM::t2TBB_JT) 653 ? 1 : ((Opc == ARM::t2TBH_JT) ? 2 : 4); 654 unsigned NumOps = MCID.getNumOperands(); 655 MachineOperand JTOP = 656 MI->getOperand(NumOps - (MI->isPredicable() ? 3 : 2)); 657 unsigned JTI = JTOP.getIndex(); 658 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); 659 assert(MJTI != nullptr); 660 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables(); 661 assert(JTI < JT.size()); 662 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte 663 // 4 aligned. The assembler / linker may add 2 byte padding just before 664 // the JT entries. The size does not include this padding; the 665 // constant islands pass does separate bookkeeping for it. 666 // FIXME: If we know the size of the function is less than (1 << 16) *2 667 // bytes, we can use 16-bit entries instead. Then there won't be an 668 // alignment issue. 669 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4; 670 unsigned NumEntries = JT[JTI].MBBs.size(); 671 if (Opc == ARM::t2TBB_JT && (NumEntries & 1)) 672 // Make sure the instruction that follows TBB is 2-byte aligned. 673 // FIXME: Constant island pass should insert an "ALIGN" instruction 674 // instead. 675 ++NumEntries; 676 return NumEntries * EntrySize + InstSize; 677 } 678 case ARM::SPACE: 679 return MI->getOperand(1).getImm(); 680 } 681} 682 683unsigned ARMBaseInstrInfo::getInstBundleLength(const MachineInstr *MI) const { 684 unsigned Size = 0; 685 MachineBasicBlock::const_instr_iterator I = MI; 686 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end(); 687 while (++I != E && I->isInsideBundle()) { 688 assert(!I->isBundle() && "No nested bundle!"); 689 Size += GetInstSizeInBytes(&*I); 690 } 691 return Size; 692} 693 694void ARMBaseInstrInfo::copyFromCPSR(MachineBasicBlock &MBB, 695 MachineBasicBlock::iterator I, 696 unsigned DestReg, bool KillSrc, 697 const ARMSubtarget &Subtarget) const { 698 unsigned Opc = Subtarget.isThumb() 699 ? (Subtarget.isMClass() ? ARM::t2MRS_M : ARM::t2MRS_AR) 700 : ARM::MRS; 701 702 MachineInstrBuilder MIB = 703 BuildMI(MBB, I, I->getDebugLoc(), get(Opc), DestReg); 704 705 // There is only 1 A/R class MRS instruction, and it always refers to 706 // APSR. However, there are lots of other possibilities on M-class cores. 707 if (Subtarget.isMClass()) 708 MIB.addImm(0x800); 709 710 AddDefaultPred(MIB); 711 712 MIB.addReg(ARM::CPSR, RegState::Implicit | getKillRegState(KillSrc)); 713} 714 715void ARMBaseInstrInfo::copyToCPSR(MachineBasicBlock &MBB, 716 MachineBasicBlock::iterator I, 717 unsigned SrcReg, bool KillSrc, 718 const ARMSubtarget &Subtarget) const { 719 unsigned Opc = Subtarget.isThumb() 720 ? (Subtarget.isMClass() ? ARM::t2MSR_M : ARM::t2MSR_AR) 721 : ARM::MSR; 722 723 MachineInstrBuilder MIB = BuildMI(MBB, I, I->getDebugLoc(), get(Opc)); 724 725 if (Subtarget.isMClass()) 726 MIB.addImm(0x800); 727 else 728 MIB.addImm(8); 729 730 MIB.addReg(SrcReg, getKillRegState(KillSrc)); 731 732 AddDefaultPred(MIB); 733 734 MIB.addReg(ARM::CPSR, RegState::Implicit | RegState::Define); 735} 736 737void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 738 MachineBasicBlock::iterator I, DebugLoc DL, 739 unsigned DestReg, unsigned SrcReg, 740 bool KillSrc) const { 741 bool GPRDest = ARM::GPRRegClass.contains(DestReg); 742 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg); 743 744 if (GPRDest && GPRSrc) { 745 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg) 746 .addReg(SrcReg, getKillRegState(KillSrc)))); 747 return; 748 } 749 750 bool SPRDest = ARM::SPRRegClass.contains(DestReg); 751 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg); 752 753 unsigned Opc = 0; 754 if (SPRDest && SPRSrc) 755 Opc = ARM::VMOVS; 756 else if (GPRDest && SPRSrc) 757 Opc = ARM::VMOVRS; 758 else if (SPRDest && GPRSrc) 759 Opc = ARM::VMOVSR; 760 else if (ARM::DPRRegClass.contains(DestReg, SrcReg) && !Subtarget.isFPOnlySP()) 761 Opc = ARM::VMOVD; 762 else if (ARM::QPRRegClass.contains(DestReg, SrcReg)) 763 Opc = ARM::VORRq; 764 765 if (Opc) { 766 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg); 767 MIB.addReg(SrcReg, getKillRegState(KillSrc)); 768 if (Opc == ARM::VORRq) 769 MIB.addReg(SrcReg, getKillRegState(KillSrc)); 770 AddDefaultPred(MIB); 771 return; 772 } 773 774 // Handle register classes that require multiple instructions. 775 unsigned BeginIdx = 0; 776 unsigned SubRegs = 0; 777 int Spacing = 1; 778 779 // Use VORRq when possible. 780 if (ARM::QQPRRegClass.contains(DestReg, SrcReg)) { 781 Opc = ARM::VORRq; 782 BeginIdx = ARM::qsub_0; 783 SubRegs = 2; 784 } else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg)) { 785 Opc = ARM::VORRq; 786 BeginIdx = ARM::qsub_0; 787 SubRegs = 4; 788 // Fall back to VMOVD. 789 } else if (ARM::DPairRegClass.contains(DestReg, SrcReg)) { 790 Opc = ARM::VMOVD; 791 BeginIdx = ARM::dsub_0; 792 SubRegs = 2; 793 } else if (ARM::DTripleRegClass.contains(DestReg, SrcReg)) { 794 Opc = ARM::VMOVD; 795 BeginIdx = ARM::dsub_0; 796 SubRegs = 3; 797 } else if (ARM::DQuadRegClass.contains(DestReg, SrcReg)) { 798 Opc = ARM::VMOVD; 799 BeginIdx = ARM::dsub_0; 800 SubRegs = 4; 801 } else if (ARM::GPRPairRegClass.contains(DestReg, SrcReg)) { 802 Opc = Subtarget.isThumb2() ? ARM::tMOVr : ARM::MOVr; 803 BeginIdx = ARM::gsub_0; 804 SubRegs = 2; 805 } else if (ARM::DPairSpcRegClass.contains(DestReg, SrcReg)) { 806 Opc = ARM::VMOVD; 807 BeginIdx = ARM::dsub_0; 808 SubRegs = 2; 809 Spacing = 2; 810 } else if (ARM::DTripleSpcRegClass.contains(DestReg, SrcReg)) { 811 Opc = ARM::VMOVD; 812 BeginIdx = ARM::dsub_0; 813 SubRegs = 3; 814 Spacing = 2; 815 } else if (ARM::DQuadSpcRegClass.contains(DestReg, SrcReg)) { 816 Opc = ARM::VMOVD; 817 BeginIdx = ARM::dsub_0; 818 SubRegs = 4; 819 Spacing = 2; 820 } else if (ARM::DPRRegClass.contains(DestReg, SrcReg) && Subtarget.isFPOnlySP()) { 821 Opc = ARM::VMOVS; 822 BeginIdx = ARM::ssub_0; 823 SubRegs = 2; 824 } else if (SrcReg == ARM::CPSR) { 825 copyFromCPSR(MBB, I, DestReg, KillSrc, Subtarget); 826 return; 827 } else if (DestReg == ARM::CPSR) { 828 copyToCPSR(MBB, I, SrcReg, KillSrc, Subtarget); 829 return; 830 } 831 832 assert(Opc && "Impossible reg-to-reg copy"); 833 834 const TargetRegisterInfo *TRI = &getRegisterInfo(); 835 MachineInstrBuilder Mov; 836 837 // Copy register tuples backward when the first Dest reg overlaps with SrcReg. 838 if (TRI->regsOverlap(SrcReg, TRI->getSubReg(DestReg, BeginIdx))) { 839 BeginIdx = BeginIdx + ((SubRegs - 1) * Spacing); 840 Spacing = -Spacing; 841 } 842#ifndef NDEBUG 843 SmallSet<unsigned, 4> DstRegs; 844#endif 845 for (unsigned i = 0; i != SubRegs; ++i) { 846 unsigned Dst = TRI->getSubReg(DestReg, BeginIdx + i * Spacing); 847 unsigned Src = TRI->getSubReg(SrcReg, BeginIdx + i * Spacing); 848 assert(Dst && Src && "Bad sub-register"); 849#ifndef NDEBUG 850 assert(!DstRegs.count(Src) && "destructive vector copy"); 851 DstRegs.insert(Dst); 852#endif 853 Mov = BuildMI(MBB, I, I->getDebugLoc(), get(Opc), Dst).addReg(Src); 854 // VORR takes two source operands. 855 if (Opc == ARM::VORRq) 856 Mov.addReg(Src); 857 Mov = AddDefaultPred(Mov); 858 // MOVr can set CC. 859 if (Opc == ARM::MOVr) 860 Mov = AddDefaultCC(Mov); 861 } 862 // Add implicit super-register defs and kills to the last instruction. 863 Mov->addRegisterDefined(DestReg, TRI); 864 if (KillSrc) 865 Mov->addRegisterKilled(SrcReg, TRI); 866} 867 868const MachineInstrBuilder & 869ARMBaseInstrInfo::AddDReg(MachineInstrBuilder &MIB, unsigned Reg, 870 unsigned SubIdx, unsigned State, 871 const TargetRegisterInfo *TRI) const { 872 if (!SubIdx) 873 return MIB.addReg(Reg, State); 874 875 if (TargetRegisterInfo::isPhysicalRegister(Reg)) 876 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State); 877 return MIB.addReg(Reg, State, SubIdx); 878} 879 880void ARMBaseInstrInfo:: 881storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, 882 unsigned SrcReg, bool isKill, int FI, 883 const TargetRegisterClass *RC, 884 const TargetRegisterInfo *TRI) const { 885 DebugLoc DL; 886 if (I != MBB.end()) DL = I->getDebugLoc(); 887 MachineFunction &MF = *MBB.getParent(); 888 MachineFrameInfo &MFI = *MF.getFrameInfo(); 889 unsigned Align = MFI.getObjectAlignment(FI); 890 891 MachineMemOperand *MMO = 892 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI), 893 MachineMemOperand::MOStore, 894 MFI.getObjectSize(FI), 895 Align); 896 897 switch (RC->getSize()) { 898 case 4: 899 if (ARM::GPRRegClass.hasSubClassEq(RC)) { 900 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STRi12)) 901 .addReg(SrcReg, getKillRegState(isKill)) 902 .addFrameIndex(FI).addImm(0).addMemOperand(MMO)); 903 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) { 904 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS)) 905 .addReg(SrcReg, getKillRegState(isKill)) 906 .addFrameIndex(FI).addImm(0).addMemOperand(MMO)); 907 } else 908 llvm_unreachable("Unknown reg class!"); 909 break; 910 case 8: 911 if (ARM::DPRRegClass.hasSubClassEq(RC)) { 912 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD)) 913 .addReg(SrcReg, getKillRegState(isKill)) 914 .addFrameIndex(FI).addImm(0).addMemOperand(MMO)); 915 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) { 916 if (Subtarget.hasV5TEOps()) { 917 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::STRD)); 918 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI); 919 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI); 920 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO); 921 922 AddDefaultPred(MIB); 923 } else { 924 // Fallback to STM instruction, which has existed since the dawn of 925 // time. 926 MachineInstrBuilder MIB = 927 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STMIA)) 928 .addFrameIndex(FI).addMemOperand(MMO)); 929 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI); 930 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI); 931 } 932 } else 933 llvm_unreachable("Unknown reg class!"); 934 break; 935 case 16: 936 if (ARM::DPairRegClass.hasSubClassEq(RC)) { 937 // Use aligned spills if the stack can be realigned. 938 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) { 939 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q64)) 940 .addFrameIndex(FI).addImm(16) 941 .addReg(SrcReg, getKillRegState(isKill)) 942 .addMemOperand(MMO)); 943 } else { 944 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQIA)) 945 .addReg(SrcReg, getKillRegState(isKill)) 946 .addFrameIndex(FI) 947 .addMemOperand(MMO)); 948 } 949 } else 950 llvm_unreachable("Unknown reg class!"); 951 break; 952 case 24: 953 if (ARM::DTripleRegClass.hasSubClassEq(RC)) { 954 // Use aligned spills if the stack can be realigned. 955 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) { 956 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64TPseudo)) 957 .addFrameIndex(FI).addImm(16) 958 .addReg(SrcReg, getKillRegState(isKill)) 959 .addMemOperand(MMO)); 960 } else { 961 MachineInstrBuilder MIB = 962 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA)) 963 .addFrameIndex(FI)) 964 .addMemOperand(MMO); 965 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI); 966 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI); 967 AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI); 968 } 969 } else 970 llvm_unreachable("Unknown reg class!"); 971 break; 972 case 32: 973 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) { 974 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) { 975 // FIXME: It's possible to only store part of the QQ register if the 976 // spilled def has a sub-register index. 977 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64QPseudo)) 978 .addFrameIndex(FI).addImm(16) 979 .addReg(SrcReg, getKillRegState(isKill)) 980 .addMemOperand(MMO)); 981 } else { 982 MachineInstrBuilder MIB = 983 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA)) 984 .addFrameIndex(FI)) 985 .addMemOperand(MMO); 986 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI); 987 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI); 988 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI); 989 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI); 990 } 991 } else 992 llvm_unreachable("Unknown reg class!"); 993 break; 994 case 64: 995 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) { 996 MachineInstrBuilder MIB = 997 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA)) 998 .addFrameIndex(FI)) 999 .addMemOperand(MMO); 1000 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI); 1001 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI); 1002 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI); 1003 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI); 1004 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI); 1005 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI); 1006 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI); 1007 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI); 1008 } else 1009 llvm_unreachable("Unknown reg class!"); 1010 break; 1011 default: 1012 llvm_unreachable("Unknown reg class!"); 1013 } 1014} 1015 1016unsigned 1017ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI, 1018 int &FrameIndex) const { 1019 switch (MI->getOpcode()) { 1020 default: break; 1021 case ARM::STRrs: 1022 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame. 1023 if (MI->getOperand(1).isFI() && 1024 MI->getOperand(2).isReg() && 1025 MI->getOperand(3).isImm() && 1026 MI->getOperand(2).getReg() == 0 && 1027 MI->getOperand(3).getImm() == 0) { 1028 FrameIndex = MI->getOperand(1).getIndex(); 1029 return MI->getOperand(0).getReg(); 1030 } 1031 break; 1032 case ARM::STRi12: 1033 case ARM::t2STRi12: 1034 case ARM::tSTRspi: 1035 case ARM::VSTRD: 1036 case ARM::VSTRS: 1037 if (MI->getOperand(1).isFI() && 1038 MI->getOperand(2).isImm() && 1039 MI->getOperand(2).getImm() == 0) { 1040 FrameIndex = MI->getOperand(1).getIndex(); 1041 return MI->getOperand(0).getReg(); 1042 } 1043 break; 1044 case ARM::VST1q64: 1045 case ARM::VST1d64TPseudo: 1046 case ARM::VST1d64QPseudo: 1047 if (MI->getOperand(0).isFI() && 1048 MI->getOperand(2).getSubReg() == 0) { 1049 FrameIndex = MI->getOperand(0).getIndex(); 1050 return MI->getOperand(2).getReg(); 1051 } 1052 break; 1053 case ARM::VSTMQIA: 1054 if (MI->getOperand(1).isFI() && 1055 MI->getOperand(0).getSubReg() == 0) { 1056 FrameIndex = MI->getOperand(1).getIndex(); 1057 return MI->getOperand(0).getReg(); 1058 } 1059 break; 1060 } 1061 1062 return 0; 1063} 1064 1065unsigned ARMBaseInstrInfo::isStoreToStackSlotPostFE(const MachineInstr *MI, 1066 int &FrameIndex) const { 1067 const MachineMemOperand *Dummy; 1068 return MI->mayStore() && hasStoreToStackSlot(MI, Dummy, FrameIndex); 1069} 1070 1071void ARMBaseInstrInfo:: 1072loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, 1073 unsigned DestReg, int FI, 1074 const TargetRegisterClass *RC, 1075 const TargetRegisterInfo *TRI) const { 1076 DebugLoc DL; 1077 if (I != MBB.end()) DL = I->getDebugLoc(); 1078 MachineFunction &MF = *MBB.getParent(); 1079 MachineFrameInfo &MFI = *MF.getFrameInfo(); 1080 unsigned Align = MFI.getObjectAlignment(FI); 1081 MachineMemOperand *MMO = 1082 MF.getMachineMemOperand( 1083 MachinePointerInfo::getFixedStack(FI), 1084 MachineMemOperand::MOLoad, 1085 MFI.getObjectSize(FI), 1086 Align); 1087 1088 switch (RC->getSize()) { 1089 case 4: 1090 if (ARM::GPRRegClass.hasSubClassEq(RC)) { 1091 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDRi12), DestReg) 1092 .addFrameIndex(FI).addImm(0).addMemOperand(MMO)); 1093 1094 } else if (ARM::SPRRegClass.hasSubClassEq(RC)) { 1095 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg) 1096 .addFrameIndex(FI).addImm(0).addMemOperand(MMO)); 1097 } else 1098 llvm_unreachable("Unknown reg class!"); 1099 break; 1100 case 8: 1101 if (ARM::DPRRegClass.hasSubClassEq(RC)) { 1102 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg) 1103 .addFrameIndex(FI).addImm(0).addMemOperand(MMO)); 1104 } else if (ARM::GPRPairRegClass.hasSubClassEq(RC)) { 1105 MachineInstrBuilder MIB; 1106 1107 if (Subtarget.hasV5TEOps()) { 1108 MIB = BuildMI(MBB, I, DL, get(ARM::LDRD)); 1109 AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI); 1110 AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI); 1111 MIB.addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO); 1112 1113 AddDefaultPred(MIB); 1114 } else { 1115 // Fallback to LDM instruction, which has existed since the dawn of 1116 // time. 1117 MIB = AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDMIA)) 1118 .addFrameIndex(FI).addMemOperand(MMO)); 1119 MIB = AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI); 1120 MIB = AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI); 1121 } 1122 1123 if (TargetRegisterInfo::isPhysicalRegister(DestReg)) 1124 MIB.addReg(DestReg, RegState::ImplicitDefine); 1125 } else 1126 llvm_unreachable("Unknown reg class!"); 1127 break; 1128 case 16: 1129 if (ARM::DPairRegClass.hasSubClassEq(RC)) { 1130 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) { 1131 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q64), DestReg) 1132 .addFrameIndex(FI).addImm(16) 1133 .addMemOperand(MMO)); 1134 } else { 1135 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQIA), DestReg) 1136 .addFrameIndex(FI) 1137 .addMemOperand(MMO)); 1138 } 1139 } else 1140 llvm_unreachable("Unknown reg class!"); 1141 break; 1142 case 24: 1143 if (ARM::DTripleRegClass.hasSubClassEq(RC)) { 1144 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) { 1145 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64TPseudo), DestReg) 1146 .addFrameIndex(FI).addImm(16) 1147 .addMemOperand(MMO)); 1148 } else { 1149 MachineInstrBuilder MIB = 1150 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA)) 1151 .addFrameIndex(FI) 1152 .addMemOperand(MMO)); 1153 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI); 1154 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI); 1155 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI); 1156 if (TargetRegisterInfo::isPhysicalRegister(DestReg)) 1157 MIB.addReg(DestReg, RegState::ImplicitDefine); 1158 } 1159 } else 1160 llvm_unreachable("Unknown reg class!"); 1161 break; 1162 case 32: 1163 if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) { 1164 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) { 1165 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg) 1166 .addFrameIndex(FI).addImm(16) 1167 .addMemOperand(MMO)); 1168 } else { 1169 MachineInstrBuilder MIB = 1170 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA)) 1171 .addFrameIndex(FI)) 1172 .addMemOperand(MMO); 1173 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI); 1174 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI); 1175 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI); 1176 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI); 1177 if (TargetRegisterInfo::isPhysicalRegister(DestReg)) 1178 MIB.addReg(DestReg, RegState::ImplicitDefine); 1179 } 1180 } else 1181 llvm_unreachable("Unknown reg class!"); 1182 break; 1183 case 64: 1184 if (ARM::QQQQPRRegClass.hasSubClassEq(RC)) { 1185 MachineInstrBuilder MIB = 1186 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA)) 1187 .addFrameIndex(FI)) 1188 .addMemOperand(MMO); 1189 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI); 1190 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI); 1191 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI); 1192 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::DefineNoRead, TRI); 1193 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::DefineNoRead, TRI); 1194 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::DefineNoRead, TRI); 1195 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::DefineNoRead, TRI); 1196 MIB = AddDReg(MIB, DestReg, ARM::dsub_7, RegState::DefineNoRead, TRI); 1197 if (TargetRegisterInfo::isPhysicalRegister(DestReg)) 1198 MIB.addReg(DestReg, RegState::ImplicitDefine); 1199 } else 1200 llvm_unreachable("Unknown reg class!"); 1201 break; 1202 default: 1203 llvm_unreachable("Unknown regclass!"); 1204 } 1205} 1206 1207unsigned 1208ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI, 1209 int &FrameIndex) const { 1210 switch (MI->getOpcode()) { 1211 default: break; 1212 case ARM::LDRrs: 1213 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame. 1214 if (MI->getOperand(1).isFI() && 1215 MI->getOperand(2).isReg() && 1216 MI->getOperand(3).isImm() && 1217 MI->getOperand(2).getReg() == 0 && 1218 MI->getOperand(3).getImm() == 0) { 1219 FrameIndex = MI->getOperand(1).getIndex(); 1220 return MI->getOperand(0).getReg(); 1221 } 1222 break; 1223 case ARM::LDRi12: 1224 case ARM::t2LDRi12: 1225 case ARM::tLDRspi: 1226 case ARM::VLDRD: 1227 case ARM::VLDRS: 1228 if (MI->getOperand(1).isFI() && 1229 MI->getOperand(2).isImm() && 1230 MI->getOperand(2).getImm() == 0) { 1231 FrameIndex = MI->getOperand(1).getIndex(); 1232 return MI->getOperand(0).getReg(); 1233 } 1234 break; 1235 case ARM::VLD1q64: 1236 case ARM::VLD1d64TPseudo: 1237 case ARM::VLD1d64QPseudo: 1238 if (MI->getOperand(1).isFI() && 1239 MI->getOperand(0).getSubReg() == 0) { 1240 FrameIndex = MI->getOperand(1).getIndex(); 1241 return MI->getOperand(0).getReg(); 1242 } 1243 break; 1244 case ARM::VLDMQIA: 1245 if (MI->getOperand(1).isFI() && 1246 MI->getOperand(0).getSubReg() == 0) { 1247 FrameIndex = MI->getOperand(1).getIndex(); 1248 return MI->getOperand(0).getReg(); 1249 } 1250 break; 1251 } 1252 1253 return 0; 1254} 1255 1256unsigned ARMBaseInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI, 1257 int &FrameIndex) const { 1258 const MachineMemOperand *Dummy; 1259 return MI->mayLoad() && hasLoadFromStackSlot(MI, Dummy, FrameIndex); 1260} 1261 1262bool 1263ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const { 1264 MachineFunction &MF = *MI->getParent()->getParent(); 1265 Reloc::Model RM = MF.getTarget().getRelocationModel(); 1266 1267 if (MI->getOpcode() == TargetOpcode::LOAD_STACK_GUARD) { 1268 assert(getSubtarget().getTargetTriple().getObjectFormat() == 1269 Triple::MachO && 1270 "LOAD_STACK_GUARD currently supported only for MachO."); 1271 expandLoadStackGuard(MI, RM); 1272 MI->getParent()->erase(MI); 1273 return true; 1274 } 1275 1276 // This hook gets to expand COPY instructions before they become 1277 // copyPhysReg() calls. Look for VMOVS instructions that can legally be 1278 // widened to VMOVD. We prefer the VMOVD when possible because it may be 1279 // changed into a VORR that can go down the NEON pipeline. 1280 if (!WidenVMOVS || !MI->isCopy() || Subtarget.isCortexA15() || 1281 Subtarget.isFPOnlySP()) 1282 return false; 1283 1284 // Look for a copy between even S-registers. That is where we keep floats 1285 // when using NEON v2f32 instructions for f32 arithmetic. 1286 unsigned DstRegS = MI->getOperand(0).getReg(); 1287 unsigned SrcRegS = MI->getOperand(1).getReg(); 1288 if (!ARM::SPRRegClass.contains(DstRegS, SrcRegS)) 1289 return false; 1290 1291 const TargetRegisterInfo *TRI = &getRegisterInfo(); 1292 unsigned DstRegD = TRI->getMatchingSuperReg(DstRegS, ARM::ssub_0, 1293 &ARM::DPRRegClass); 1294 unsigned SrcRegD = TRI->getMatchingSuperReg(SrcRegS, ARM::ssub_0, 1295 &ARM::DPRRegClass); 1296 if (!DstRegD || !SrcRegD) 1297 return false; 1298 1299 // We want to widen this into a DstRegD = VMOVD SrcRegD copy. This is only 1300 // legal if the COPY already defines the full DstRegD, and it isn't a 1301 // sub-register insertion. 1302 if (!MI->definesRegister(DstRegD, TRI) || MI->readsRegister(DstRegD, TRI)) 1303 return false; 1304 1305 // A dead copy shouldn't show up here, but reject it just in case. 1306 if (MI->getOperand(0).isDead()) 1307 return false; 1308 1309 // All clear, widen the COPY. 1310 DEBUG(dbgs() << "widening: " << *MI); 1311 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI); 1312 1313 // Get rid of the old <imp-def> of DstRegD. Leave it if it defines a Q-reg 1314 // or some other super-register. 1315 int ImpDefIdx = MI->findRegisterDefOperandIdx(DstRegD); 1316 if (ImpDefIdx != -1) 1317 MI->RemoveOperand(ImpDefIdx); 1318 1319 // Change the opcode and operands. 1320 MI->setDesc(get(ARM::VMOVD)); 1321 MI->getOperand(0).setReg(DstRegD); 1322 MI->getOperand(1).setReg(SrcRegD); 1323 AddDefaultPred(MIB); 1324 1325 // We are now reading SrcRegD instead of SrcRegS. This may upset the 1326 // register scavenger and machine verifier, so we need to indicate that we 1327 // are reading an undefined value from SrcRegD, but a proper value from 1328 // SrcRegS. 1329 MI->getOperand(1).setIsUndef(); 1330 MIB.addReg(SrcRegS, RegState::Implicit); 1331 1332 // SrcRegD may actually contain an unrelated value in the ssub_1 1333 // sub-register. Don't kill it. Only kill the ssub_0 sub-register. 1334 if (MI->getOperand(1).isKill()) { 1335 MI->getOperand(1).setIsKill(false); 1336 MI->addRegisterKilled(SrcRegS, TRI, true); 1337 } 1338 1339 DEBUG(dbgs() << "replaced by: " << *MI); 1340 return true; 1341} 1342 1343/// Create a copy of a const pool value. Update CPI to the new index and return 1344/// the label UID. 1345static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) { 1346 MachineConstantPool *MCP = MF.getConstantPool(); 1347 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1348 1349 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI]; 1350 assert(MCPE.isMachineConstantPoolEntry() && 1351 "Expecting a machine constantpool entry!"); 1352 ARMConstantPoolValue *ACPV = 1353 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal); 1354 1355 unsigned PCLabelId = AFI->createPICLabelUId(); 1356 ARMConstantPoolValue *NewCPV = nullptr; 1357 1358 // FIXME: The below assumes PIC relocation model and that the function 1359 // is Thumb mode (t1 or t2). PCAdjustment would be 8 for ARM mode PIC, and 1360 // zero for non-PIC in ARM or Thumb. The callers are all of thumb LDR 1361 // instructions, so that's probably OK, but is PIC always correct when 1362 // we get here? 1363 if (ACPV->isGlobalValue()) 1364 NewCPV = ARMConstantPoolConstant:: 1365 Create(cast<ARMConstantPoolConstant>(ACPV)->getGV(), PCLabelId, 1366 ARMCP::CPValue, 4); 1367 else if (ACPV->isExtSymbol()) 1368 NewCPV = ARMConstantPoolSymbol:: 1369 Create(MF.getFunction()->getContext(), 1370 cast<ARMConstantPoolSymbol>(ACPV)->getSymbol(), PCLabelId, 4); 1371 else if (ACPV->isBlockAddress()) 1372 NewCPV = ARMConstantPoolConstant:: 1373 Create(cast<ARMConstantPoolConstant>(ACPV)->getBlockAddress(), PCLabelId, 1374 ARMCP::CPBlockAddress, 4); 1375 else if (ACPV->isLSDA()) 1376 NewCPV = ARMConstantPoolConstant::Create(MF.getFunction(), PCLabelId, 1377 ARMCP::CPLSDA, 4); 1378 else if (ACPV->isMachineBasicBlock()) 1379 NewCPV = ARMConstantPoolMBB:: 1380 Create(MF.getFunction()->getContext(), 1381 cast<ARMConstantPoolMBB>(ACPV)->getMBB(), PCLabelId, 4); 1382 else 1383 llvm_unreachable("Unexpected ARM constantpool value type!!"); 1384 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment()); 1385 return PCLabelId; 1386} 1387 1388void ARMBaseInstrInfo:: 1389reMaterialize(MachineBasicBlock &MBB, 1390 MachineBasicBlock::iterator I, 1391 unsigned DestReg, unsigned SubIdx, 1392 const MachineInstr *Orig, 1393 const TargetRegisterInfo &TRI) const { 1394 unsigned Opcode = Orig->getOpcode(); 1395 switch (Opcode) { 1396 default: { 1397 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig); 1398 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI); 1399 MBB.insert(I, MI); 1400 break; 1401 } 1402 case ARM::tLDRpci_pic: 1403 case ARM::t2LDRpci_pic: { 1404 MachineFunction &MF = *MBB.getParent(); 1405 unsigned CPI = Orig->getOperand(1).getIndex(); 1406 unsigned PCLabelId = duplicateCPV(MF, CPI); 1407 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode), 1408 DestReg) 1409 .addConstantPoolIndex(CPI).addImm(PCLabelId); 1410 MIB->setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end()); 1411 break; 1412 } 1413 } 1414} 1415 1416MachineInstr * 1417ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const { 1418 MachineInstr *MI = TargetInstrInfo::duplicate(Orig, MF); 1419 switch(Orig->getOpcode()) { 1420 case ARM::tLDRpci_pic: 1421 case ARM::t2LDRpci_pic: { 1422 unsigned CPI = Orig->getOperand(1).getIndex(); 1423 unsigned PCLabelId = duplicateCPV(MF, CPI); 1424 Orig->getOperand(1).setIndex(CPI); 1425 Orig->getOperand(2).setImm(PCLabelId); 1426 break; 1427 } 1428 } 1429 return MI; 1430} 1431 1432bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0, 1433 const MachineInstr *MI1, 1434 const MachineRegisterInfo *MRI) const { 1435 int Opcode = MI0->getOpcode(); 1436 if (Opcode == ARM::t2LDRpci || 1437 Opcode == ARM::t2LDRpci_pic || 1438 Opcode == ARM::tLDRpci || 1439 Opcode == ARM::tLDRpci_pic || 1440 Opcode == ARM::LDRLIT_ga_pcrel || 1441 Opcode == ARM::LDRLIT_ga_pcrel_ldr || 1442 Opcode == ARM::tLDRLIT_ga_pcrel || 1443 Opcode == ARM::MOV_ga_pcrel || 1444 Opcode == ARM::MOV_ga_pcrel_ldr || 1445 Opcode == ARM::t2MOV_ga_pcrel) { 1446 if (MI1->getOpcode() != Opcode) 1447 return false; 1448 if (MI0->getNumOperands() != MI1->getNumOperands()) 1449 return false; 1450 1451 const MachineOperand &MO0 = MI0->getOperand(1); 1452 const MachineOperand &MO1 = MI1->getOperand(1); 1453 if (MO0.getOffset() != MO1.getOffset()) 1454 return false; 1455 1456 if (Opcode == ARM::LDRLIT_ga_pcrel || 1457 Opcode == ARM::LDRLIT_ga_pcrel_ldr || 1458 Opcode == ARM::tLDRLIT_ga_pcrel || 1459 Opcode == ARM::MOV_ga_pcrel || 1460 Opcode == ARM::MOV_ga_pcrel_ldr || 1461 Opcode == ARM::t2MOV_ga_pcrel) 1462 // Ignore the PC labels. 1463 return MO0.getGlobal() == MO1.getGlobal(); 1464 1465 const MachineFunction *MF = MI0->getParent()->getParent(); 1466 const MachineConstantPool *MCP = MF->getConstantPool(); 1467 int CPI0 = MO0.getIndex(); 1468 int CPI1 = MO1.getIndex(); 1469 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0]; 1470 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1]; 1471 bool isARMCP0 = MCPE0.isMachineConstantPoolEntry(); 1472 bool isARMCP1 = MCPE1.isMachineConstantPoolEntry(); 1473 if (isARMCP0 && isARMCP1) { 1474 ARMConstantPoolValue *ACPV0 = 1475 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal); 1476 ARMConstantPoolValue *ACPV1 = 1477 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal); 1478 return ACPV0->hasSameValue(ACPV1); 1479 } else if (!isARMCP0 && !isARMCP1) { 1480 return MCPE0.Val.ConstVal == MCPE1.Val.ConstVal; 1481 } 1482 return false; 1483 } else if (Opcode == ARM::PICLDR) { 1484 if (MI1->getOpcode() != Opcode) 1485 return false; 1486 if (MI0->getNumOperands() != MI1->getNumOperands()) 1487 return false; 1488 1489 unsigned Addr0 = MI0->getOperand(1).getReg(); 1490 unsigned Addr1 = MI1->getOperand(1).getReg(); 1491 if (Addr0 != Addr1) { 1492 if (!MRI || 1493 !TargetRegisterInfo::isVirtualRegister(Addr0) || 1494 !TargetRegisterInfo::isVirtualRegister(Addr1)) 1495 return false; 1496 1497 // This assumes SSA form. 1498 MachineInstr *Def0 = MRI->getVRegDef(Addr0); 1499 MachineInstr *Def1 = MRI->getVRegDef(Addr1); 1500 // Check if the loaded value, e.g. a constantpool of a global address, are 1501 // the same. 1502 if (!produceSameValue(Def0, Def1, MRI)) 1503 return false; 1504 } 1505 1506 for (unsigned i = 3, e = MI0->getNumOperands(); i != e; ++i) { 1507 // %vreg12<def> = PICLDR %vreg11, 0, pred:14, pred:%noreg 1508 const MachineOperand &MO0 = MI0->getOperand(i); 1509 const MachineOperand &MO1 = MI1->getOperand(i); 1510 if (!MO0.isIdenticalTo(MO1)) 1511 return false; 1512 } 1513 return true; 1514 } 1515 1516 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs); 1517} 1518 1519/// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to 1520/// determine if two loads are loading from the same base address. It should 1521/// only return true if the base pointers are the same and the only differences 1522/// between the two addresses is the offset. It also returns the offsets by 1523/// reference. 1524/// 1525/// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched 1526/// is permanently disabled. 1527bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 1528 int64_t &Offset1, 1529 int64_t &Offset2) const { 1530 // Don't worry about Thumb: just ARM and Thumb2. 1531 if (Subtarget.isThumb1Only()) return false; 1532 1533 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode()) 1534 return false; 1535 1536 switch (Load1->getMachineOpcode()) { 1537 default: 1538 return false; 1539 case ARM::LDRi12: 1540 case ARM::LDRBi12: 1541 case ARM::LDRD: 1542 case ARM::LDRH: 1543 case ARM::LDRSB: 1544 case ARM::LDRSH: 1545 case ARM::VLDRD: 1546 case ARM::VLDRS: 1547 case ARM::t2LDRi8: 1548 case ARM::t2LDRBi8: 1549 case ARM::t2LDRDi8: 1550 case ARM::t2LDRSHi8: 1551 case ARM::t2LDRi12: 1552 case ARM::t2LDRBi12: 1553 case ARM::t2LDRSHi12: 1554 break; 1555 } 1556 1557 switch (Load2->getMachineOpcode()) { 1558 default: 1559 return false; 1560 case ARM::LDRi12: 1561 case ARM::LDRBi12: 1562 case ARM::LDRD: 1563 case ARM::LDRH: 1564 case ARM::LDRSB: 1565 case ARM::LDRSH: 1566 case ARM::VLDRD: 1567 case ARM::VLDRS: 1568 case ARM::t2LDRi8: 1569 case ARM::t2LDRBi8: 1570 case ARM::t2LDRSHi8: 1571 case ARM::t2LDRi12: 1572 case ARM::t2LDRBi12: 1573 case ARM::t2LDRSHi12: 1574 break; 1575 } 1576 1577 // Check if base addresses and chain operands match. 1578 if (Load1->getOperand(0) != Load2->getOperand(0) || 1579 Load1->getOperand(4) != Load2->getOperand(4)) 1580 return false; 1581 1582 // Index should be Reg0. 1583 if (Load1->getOperand(3) != Load2->getOperand(3)) 1584 return false; 1585 1586 // Determine the offsets. 1587 if (isa<ConstantSDNode>(Load1->getOperand(1)) && 1588 isa<ConstantSDNode>(Load2->getOperand(1))) { 1589 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue(); 1590 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue(); 1591 return true; 1592 } 1593 1594 return false; 1595} 1596 1597/// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to 1598/// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should 1599/// be scheduled togther. On some targets if two loads are loading from 1600/// addresses in the same cache line, it's better if they are scheduled 1601/// together. This function takes two integers that represent the load offsets 1602/// from the common base address. It returns true if it decides it's desirable 1603/// to schedule the two loads together. "NumLoads" is the number of loads that 1604/// have already been scheduled after Load1. 1605/// 1606/// FIXME: remove this in favor of the MachineInstr interface once pre-RA-sched 1607/// is permanently disabled. 1608bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2, 1609 int64_t Offset1, int64_t Offset2, 1610 unsigned NumLoads) const { 1611 // Don't worry about Thumb: just ARM and Thumb2. 1612 if (Subtarget.isThumb1Only()) return false; 1613 1614 assert(Offset2 > Offset1); 1615 1616 if ((Offset2 - Offset1) / 8 > 64) 1617 return false; 1618 1619 // Check if the machine opcodes are different. If they are different 1620 // then we consider them to not be of the same base address, 1621 // EXCEPT in the case of Thumb2 byte loads where one is LDRBi8 and the other LDRBi12. 1622 // In this case, they are considered to be the same because they are different 1623 // encoding forms of the same basic instruction. 1624 if ((Load1->getMachineOpcode() != Load2->getMachineOpcode()) && 1625 !((Load1->getMachineOpcode() == ARM::t2LDRBi8 && 1626 Load2->getMachineOpcode() == ARM::t2LDRBi12) || 1627 (Load1->getMachineOpcode() == ARM::t2LDRBi12 && 1628 Load2->getMachineOpcode() == ARM::t2LDRBi8))) 1629 return false; // FIXME: overly conservative? 1630 1631 // Four loads in a row should be sufficient. 1632 if (NumLoads >= 3) 1633 return false; 1634 1635 return true; 1636} 1637 1638bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI, 1639 const MachineBasicBlock *MBB, 1640 const MachineFunction &MF) const { 1641 // Debug info is never a scheduling boundary. It's necessary to be explicit 1642 // due to the special treatment of IT instructions below, otherwise a 1643 // dbg_value followed by an IT will result in the IT instruction being 1644 // considered a scheduling hazard, which is wrong. It should be the actual 1645 // instruction preceding the dbg_value instruction(s), just like it is 1646 // when debug info is not present. 1647 if (MI->isDebugValue()) 1648 return false; 1649 1650 // Terminators and labels can't be scheduled around. 1651 if (MI->isTerminator() || MI->isPosition()) 1652 return true; 1653 1654 // Treat the start of the IT block as a scheduling boundary, but schedule 1655 // t2IT along with all instructions following it. 1656 // FIXME: This is a big hammer. But the alternative is to add all potential 1657 // true and anti dependencies to IT block instructions as implicit operands 1658 // to the t2IT instruction. The added compile time and complexity does not 1659 // seem worth it. 1660 MachineBasicBlock::const_iterator I = MI; 1661 // Make sure to skip any dbg_value instructions 1662 while (++I != MBB->end() && I->isDebugValue()) 1663 ; 1664 if (I != MBB->end() && I->getOpcode() == ARM::t2IT) 1665 return true; 1666 1667 // Don't attempt to schedule around any instruction that defines 1668 // a stack-oriented pointer, as it's unlikely to be profitable. This 1669 // saves compile time, because it doesn't require every single 1670 // stack slot reference to depend on the instruction that does the 1671 // modification. 1672 // Calls don't actually change the stack pointer, even if they have imp-defs. 1673 // No ARM calling conventions change the stack pointer. (X86 calling 1674 // conventions sometimes do). 1675 if (!MI->isCall() && MI->definesRegister(ARM::SP)) 1676 return true; 1677 1678 return false; 1679} 1680 1681bool ARMBaseInstrInfo:: 1682isProfitableToIfCvt(MachineBasicBlock &MBB, 1683 unsigned NumCycles, unsigned ExtraPredCycles, 1684 const BranchProbability &Probability) const { 1685 if (!NumCycles) 1686 return false; 1687 1688 // Attempt to estimate the relative costs of predication versus branching. 1689 unsigned UnpredCost = Probability.getNumerator() * NumCycles; 1690 UnpredCost /= Probability.getDenominator(); 1691 UnpredCost += 1; // The branch itself 1692 UnpredCost += Subtarget.getMispredictionPenalty() / 10; 1693 1694 return (NumCycles + ExtraPredCycles) <= UnpredCost; 1695} 1696 1697bool ARMBaseInstrInfo:: 1698isProfitableToIfCvt(MachineBasicBlock &TMBB, 1699 unsigned TCycles, unsigned TExtra, 1700 MachineBasicBlock &FMBB, 1701 unsigned FCycles, unsigned FExtra, 1702 const BranchProbability &Probability) const { 1703 if (!TCycles || !FCycles) 1704 return false; 1705 1706 // Attempt to estimate the relative costs of predication versus branching. 1707 unsigned TUnpredCost = Probability.getNumerator() * TCycles; 1708 TUnpredCost /= Probability.getDenominator(); 1709 1710 uint32_t Comp = Probability.getDenominator() - Probability.getNumerator(); 1711 unsigned FUnpredCost = Comp * FCycles; 1712 FUnpredCost /= Probability.getDenominator(); 1713 1714 unsigned UnpredCost = TUnpredCost + FUnpredCost; 1715 UnpredCost += 1; // The branch itself 1716 UnpredCost += Subtarget.getMispredictionPenalty() / 10; 1717 1718 return (TCycles + FCycles + TExtra + FExtra) <= UnpredCost; 1719} 1720 1721bool 1722ARMBaseInstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB, 1723 MachineBasicBlock &FMBB) const { 1724 // Reduce false anti-dependencies to let Swift's out-of-order execution 1725 // engine do its thing. 1726 return Subtarget.isSwift(); 1727} 1728 1729/// getInstrPredicate - If instruction is predicated, returns its predicate 1730/// condition, otherwise returns AL. It also returns the condition code 1731/// register by reference. 1732ARMCC::CondCodes 1733llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) { 1734 int PIdx = MI->findFirstPredOperandIdx(); 1735 if (PIdx == -1) { 1736 PredReg = 0; 1737 return ARMCC::AL; 1738 } 1739 1740 PredReg = MI->getOperand(PIdx+1).getReg(); 1741 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm(); 1742} 1743 1744 1745int llvm::getMatchingCondBranchOpcode(int Opc) { 1746 if (Opc == ARM::B) 1747 return ARM::Bcc; 1748 if (Opc == ARM::tB) 1749 return ARM::tBcc; 1750 if (Opc == ARM::t2B) 1751 return ARM::t2Bcc; 1752 1753 llvm_unreachable("Unknown unconditional branch opcode!"); 1754} 1755 1756/// commuteInstruction - Handle commutable instructions. 1757MachineInstr * 1758ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const { 1759 switch (MI->getOpcode()) { 1760 case ARM::MOVCCr: 1761 case ARM::t2MOVCCr: { 1762 // MOVCC can be commuted by inverting the condition. 1763 unsigned PredReg = 0; 1764 ARMCC::CondCodes CC = getInstrPredicate(MI, PredReg); 1765 // MOVCC AL can't be inverted. Shouldn't happen. 1766 if (CC == ARMCC::AL || PredReg != ARM::CPSR) 1767 return nullptr; 1768 MI = TargetInstrInfo::commuteInstruction(MI, NewMI); 1769 if (!MI) 1770 return nullptr; 1771 // After swapping the MOVCC operands, also invert the condition. 1772 MI->getOperand(MI->findFirstPredOperandIdx()) 1773 .setImm(ARMCC::getOppositeCondition(CC)); 1774 return MI; 1775 } 1776 } 1777 return TargetInstrInfo::commuteInstruction(MI, NewMI); 1778} 1779 1780/// Identify instructions that can be folded into a MOVCC instruction, and 1781/// return the defining instruction. 1782static MachineInstr *canFoldIntoMOVCC(unsigned Reg, 1783 const MachineRegisterInfo &MRI, 1784 const TargetInstrInfo *TII) { 1785 if (!TargetRegisterInfo::isVirtualRegister(Reg)) 1786 return nullptr; 1787 if (!MRI.hasOneNonDBGUse(Reg)) 1788 return nullptr; 1789 MachineInstr *MI = MRI.getVRegDef(Reg); 1790 if (!MI) 1791 return nullptr; 1792 // MI is folded into the MOVCC by predicating it. 1793 if (!MI->isPredicable()) 1794 return nullptr; 1795 // Check if MI has any non-dead defs or physreg uses. This also detects 1796 // predicated instructions which will be reading CPSR. 1797 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) { 1798 const MachineOperand &MO = MI->getOperand(i); 1799 // Reject frame index operands, PEI can't handle the predicated pseudos. 1800 if (MO.isFI() || MO.isCPI() || MO.isJTI()) 1801 return nullptr; 1802 if (!MO.isReg()) 1803 continue; 1804 // MI can't have any tied operands, that would conflict with predication. 1805 if (MO.isTied()) 1806 return nullptr; 1807 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) 1808 return nullptr; 1809 if (MO.isDef() && !MO.isDead()) 1810 return nullptr; 1811 } 1812 bool DontMoveAcrossStores = true; 1813 if (!MI->isSafeToMove(TII, /* AliasAnalysis = */ nullptr, 1814 DontMoveAcrossStores)) 1815 return nullptr; 1816 return MI; 1817} 1818 1819bool ARMBaseInstrInfo::analyzeSelect(const MachineInstr *MI, 1820 SmallVectorImpl<MachineOperand> &Cond, 1821 unsigned &TrueOp, unsigned &FalseOp, 1822 bool &Optimizable) const { 1823 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) && 1824 "Unknown select instruction"); 1825 // MOVCC operands: 1826 // 0: Def. 1827 // 1: True use. 1828 // 2: False use. 1829 // 3: Condition code. 1830 // 4: CPSR use. 1831 TrueOp = 1; 1832 FalseOp = 2; 1833 Cond.push_back(MI->getOperand(3)); 1834 Cond.push_back(MI->getOperand(4)); 1835 // We can always fold a def. 1836 Optimizable = true; 1837 return false; 1838} 1839 1840MachineInstr * 1841ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI, 1842 SmallPtrSetImpl<MachineInstr *> &SeenMIs, 1843 bool PreferFalse) const { 1844 assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) && 1845 "Unknown select instruction"); 1846 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo(); 1847 MachineInstr *DefMI = canFoldIntoMOVCC(MI->getOperand(2).getReg(), MRI, this); 1848 bool Invert = !DefMI; 1849 if (!DefMI) 1850 DefMI = canFoldIntoMOVCC(MI->getOperand(1).getReg(), MRI, this); 1851 if (!DefMI) 1852 return nullptr; 1853 1854 // Find new register class to use. 1855 MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1); 1856 unsigned DestReg = MI->getOperand(0).getReg(); 1857 const TargetRegisterClass *PreviousClass = MRI.getRegClass(FalseReg.getReg()); 1858 if (!MRI.constrainRegClass(DestReg, PreviousClass)) 1859 return nullptr; 1860 1861 // Create a new predicated version of DefMI. 1862 // Rfalse is the first use. 1863 MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), 1864 DefMI->getDesc(), DestReg); 1865 1866 // Copy all the DefMI operands, excluding its (null) predicate. 1867 const MCInstrDesc &DefDesc = DefMI->getDesc(); 1868 for (unsigned i = 1, e = DefDesc.getNumOperands(); 1869 i != e && !DefDesc.OpInfo[i].isPredicate(); ++i) 1870 NewMI.addOperand(DefMI->getOperand(i)); 1871 1872 unsigned CondCode = MI->getOperand(3).getImm(); 1873 if (Invert) 1874 NewMI.addImm(ARMCC::getOppositeCondition(ARMCC::CondCodes(CondCode))); 1875 else 1876 NewMI.addImm(CondCode); 1877 NewMI.addOperand(MI->getOperand(4)); 1878 1879 // DefMI is not the -S version that sets CPSR, so add an optional %noreg. 1880 if (NewMI->hasOptionalDef()) 1881 AddDefaultCC(NewMI); 1882 1883 // The output register value when the predicate is false is an implicit 1884 // register operand tied to the first def. 1885 // The tie makes the register allocator ensure the FalseReg is allocated the 1886 // same register as operand 0. 1887 FalseReg.setImplicit(); 1888 NewMI.addOperand(FalseReg); 1889 NewMI->tieOperands(0, NewMI->getNumOperands() - 1); 1890 1891 // Update SeenMIs set: register newly created MI and erase removed DefMI. 1892 SeenMIs.insert(NewMI); 1893 SeenMIs.erase(DefMI); 1894 1895 // The caller will erase MI, but not DefMI. 1896 DefMI->eraseFromParent(); 1897 return NewMI; 1898} 1899 1900/// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether the 1901/// instruction is encoded with an 'S' bit is determined by the optional CPSR 1902/// def operand. 1903/// 1904/// This will go away once we can teach tblgen how to set the optional CPSR def 1905/// operand itself. 1906struct AddSubFlagsOpcodePair { 1907 uint16_t PseudoOpc; 1908 uint16_t MachineOpc; 1909}; 1910 1911static const AddSubFlagsOpcodePair AddSubFlagsOpcodeMap[] = { 1912 {ARM::ADDSri, ARM::ADDri}, 1913 {ARM::ADDSrr, ARM::ADDrr}, 1914 {ARM::ADDSrsi, ARM::ADDrsi}, 1915 {ARM::ADDSrsr, ARM::ADDrsr}, 1916 1917 {ARM::SUBSri, ARM::SUBri}, 1918 {ARM::SUBSrr, ARM::SUBrr}, 1919 {ARM::SUBSrsi, ARM::SUBrsi}, 1920 {ARM::SUBSrsr, ARM::SUBrsr}, 1921 1922 {ARM::RSBSri, ARM::RSBri}, 1923 {ARM::RSBSrsi, ARM::RSBrsi}, 1924 {ARM::RSBSrsr, ARM::RSBrsr}, 1925 1926 {ARM::t2ADDSri, ARM::t2ADDri}, 1927 {ARM::t2ADDSrr, ARM::t2ADDrr}, 1928 {ARM::t2ADDSrs, ARM::t2ADDrs}, 1929 1930 {ARM::t2SUBSri, ARM::t2SUBri}, 1931 {ARM::t2SUBSrr, ARM::t2SUBrr}, 1932 {ARM::t2SUBSrs, ARM::t2SUBrs}, 1933 1934 {ARM::t2RSBSri, ARM::t2RSBri}, 1935 {ARM::t2RSBSrs, ARM::t2RSBrs}, 1936}; 1937 1938unsigned llvm::convertAddSubFlagsOpcode(unsigned OldOpc) { 1939 for (unsigned i = 0, e = array_lengthof(AddSubFlagsOpcodeMap); i != e; ++i) 1940 if (OldOpc == AddSubFlagsOpcodeMap[i].PseudoOpc) 1941 return AddSubFlagsOpcodeMap[i].MachineOpc; 1942 return 0; 1943} 1944 1945void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB, 1946 MachineBasicBlock::iterator &MBBI, DebugLoc dl, 1947 unsigned DestReg, unsigned BaseReg, int NumBytes, 1948 ARMCC::CondCodes Pred, unsigned PredReg, 1949 const ARMBaseInstrInfo &TII, unsigned MIFlags) { 1950 if (NumBytes == 0 && DestReg != BaseReg) { 1951 BuildMI(MBB, MBBI, dl, TII.get(ARM::MOVr), DestReg) 1952 .addReg(BaseReg, RegState::Kill) 1953 .addImm((unsigned)Pred).addReg(PredReg).addReg(0) 1954 .setMIFlags(MIFlags); 1955 return; 1956 } 1957 1958 bool isSub = NumBytes < 0; 1959 if (isSub) NumBytes = -NumBytes; 1960 1961 while (NumBytes) { 1962 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes); 1963 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt); 1964 assert(ThisVal && "Didn't extract field correctly"); 1965 1966 // We will handle these bits from offset, clear them. 1967 NumBytes &= ~ThisVal; 1968 1969 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?"); 1970 1971 // Build the new ADD / SUB. 1972 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri; 1973 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg) 1974 .addReg(BaseReg, RegState::Kill).addImm(ThisVal) 1975 .addImm((unsigned)Pred).addReg(PredReg).addReg(0) 1976 .setMIFlags(MIFlags); 1977 BaseReg = DestReg; 1978 } 1979} 1980 1981static bool isAnySubRegLive(unsigned Reg, const TargetRegisterInfo *TRI, 1982 MachineInstr *MI) { 1983 for (MCSubRegIterator Subreg(Reg, TRI, /* IncludeSelf */ true); 1984 Subreg.isValid(); ++Subreg) 1985 if (MI->getParent()->computeRegisterLiveness(TRI, *Subreg, MI) != 1986 MachineBasicBlock::LQR_Dead) 1987 return true; 1988 return false; 1989} 1990bool llvm::tryFoldSPUpdateIntoPushPop(const ARMSubtarget &Subtarget, 1991 MachineFunction &MF, MachineInstr *MI, 1992 unsigned NumBytes) { 1993 // This optimisation potentially adds lots of load and store 1994 // micro-operations, it's only really a great benefit to code-size. 1995 if (!MF.getFunction()->hasFnAttribute(Attribute::MinSize)) 1996 return false; 1997 1998 // If only one register is pushed/popped, LLVM can use an LDR/STR 1999 // instead. We can't modify those so make sure we're dealing with an 2000 // instruction we understand. 2001 bool IsPop = isPopOpcode(MI->getOpcode()); 2002 bool IsPush = isPushOpcode(MI->getOpcode()); 2003 if (!IsPush && !IsPop) 2004 return false; 2005 2006 bool IsVFPPushPop = MI->getOpcode() == ARM::VSTMDDB_UPD || 2007 MI->getOpcode() == ARM::VLDMDIA_UPD; 2008 bool IsT1PushPop = MI->getOpcode() == ARM::tPUSH || 2009 MI->getOpcode() == ARM::tPOP || 2010 MI->getOpcode() == ARM::tPOP_RET; 2011 2012 assert((IsT1PushPop || (MI->getOperand(0).getReg() == ARM::SP && 2013 MI->getOperand(1).getReg() == ARM::SP)) && 2014 "trying to fold sp update into non-sp-updating push/pop"); 2015 2016 // The VFP push & pop act on D-registers, so we can only fold an adjustment 2017 // by a multiple of 8 bytes in correctly. Similarly rN is 4-bytes. Don't try 2018 // if this is violated. 2019 if (NumBytes % (IsVFPPushPop ? 8 : 4) != 0) 2020 return false; 2021 2022 // ARM and Thumb2 push/pop insts have explicit "sp, sp" operands (+ 2023 // pred) so the list starts at 4. Thumb1 starts after the predicate. 2024 int RegListIdx = IsT1PushPop ? 2 : 4; 2025 2026 // Calculate the space we'll need in terms of registers. 2027 unsigned FirstReg = MI->getOperand(RegListIdx).getReg(); 2028 unsigned RD0Reg, RegsNeeded; 2029 if (IsVFPPushPop) { 2030 RD0Reg = ARM::D0; 2031 RegsNeeded = NumBytes / 8; 2032 } else { 2033 RD0Reg = ARM::R0; 2034 RegsNeeded = NumBytes / 4; 2035 } 2036 2037 // We're going to have to strip all list operands off before 2038 // re-adding them since the order matters, so save the existing ones 2039 // for later. 2040 SmallVector<MachineOperand, 4> RegList; 2041 for (int i = MI->getNumOperands() - 1; i >= RegListIdx; --i) 2042 RegList.push_back(MI->getOperand(i)); 2043 2044 const TargetRegisterInfo *TRI = MF.getRegInfo().getTargetRegisterInfo(); 2045 const MCPhysReg *CSRegs = TRI->getCalleeSavedRegs(&MF); 2046 2047 // Now try to find enough space in the reglist to allocate NumBytes. 2048 for (unsigned CurReg = FirstReg - 1; CurReg >= RD0Reg && RegsNeeded; 2049 --CurReg) { 2050 if (!IsPop) { 2051 // Pushing any register is completely harmless, mark the 2052 // register involved as undef since we don't care about it in 2053 // the slightest. 2054 RegList.push_back(MachineOperand::CreateReg(CurReg, false, false, 2055 false, false, true)); 2056 --RegsNeeded; 2057 continue; 2058 } 2059 2060 // However, we can only pop an extra register if it's not live. For 2061 // registers live within the function we might clobber a return value 2062 // register; the other way a register can be live here is if it's 2063 // callee-saved. 2064 // TODO: Currently, computeRegisterLiveness() does not report "live" if a 2065 // sub reg is live. When computeRegisterLiveness() works for sub reg, it 2066 // can replace isAnySubRegLive(). 2067 if (isCalleeSavedRegister(CurReg, CSRegs) || 2068 isAnySubRegLive(CurReg, TRI, MI)) { 2069 // VFP pops don't allow holes in the register list, so any skip is fatal 2070 // for our transformation. GPR pops do, so we should just keep looking. 2071 if (IsVFPPushPop) 2072 return false; 2073 else 2074 continue; 2075 } 2076 2077 // Mark the unimportant registers as <def,dead> in the POP. 2078 RegList.push_back(MachineOperand::CreateReg(CurReg, true, false, false, 2079 true)); 2080 --RegsNeeded; 2081 } 2082 2083 if (RegsNeeded > 0) 2084 return false; 2085 2086 // Finally we know we can profitably perform the optimisation so go 2087 // ahead: strip all existing registers off and add them back again 2088 // in the right order. 2089 for (int i = MI->getNumOperands() - 1; i >= RegListIdx; --i) 2090 MI->RemoveOperand(i); 2091 2092 // Add the complete list back in. 2093 MachineInstrBuilder MIB(MF, &*MI); 2094 for (int i = RegList.size() - 1; i >= 0; --i) 2095 MIB.addOperand(RegList[i]); 2096 2097 return true; 2098} 2099 2100bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx, 2101 unsigned FrameReg, int &Offset, 2102 const ARMBaseInstrInfo &TII) { 2103 unsigned Opcode = MI.getOpcode(); 2104 const MCInstrDesc &Desc = MI.getDesc(); 2105 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask); 2106 bool isSub = false; 2107 2108 // Memory operands in inline assembly always use AddrMode2. 2109 if (Opcode == ARM::INLINEASM) 2110 AddrMode = ARMII::AddrMode2; 2111 2112 if (Opcode == ARM::ADDri) { 2113 Offset += MI.getOperand(FrameRegIdx+1).getImm(); 2114 if (Offset == 0) { 2115 // Turn it into a move. 2116 MI.setDesc(TII.get(ARM::MOVr)); 2117 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 2118 MI.RemoveOperand(FrameRegIdx+1); 2119 Offset = 0; 2120 return true; 2121 } else if (Offset < 0) { 2122 Offset = -Offset; 2123 isSub = true; 2124 MI.setDesc(TII.get(ARM::SUBri)); 2125 } 2126 2127 // Common case: small offset, fits into instruction. 2128 if (ARM_AM::getSOImmVal(Offset) != -1) { 2129 // Replace the FrameIndex with sp / fp 2130 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 2131 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset); 2132 Offset = 0; 2133 return true; 2134 } 2135 2136 // Otherwise, pull as much of the immedidate into this ADDri/SUBri 2137 // as possible. 2138 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset); 2139 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt); 2140 2141 // We will handle these bits from offset, clear them. 2142 Offset &= ~ThisImmVal; 2143 2144 // Get the properly encoded SOImmVal field. 2145 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 && 2146 "Bit extraction didn't work?"); 2147 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal); 2148 } else { 2149 unsigned ImmIdx = 0; 2150 int InstrOffs = 0; 2151 unsigned NumBits = 0; 2152 unsigned Scale = 1; 2153 switch (AddrMode) { 2154 case ARMII::AddrMode_i12: { 2155 ImmIdx = FrameRegIdx + 1; 2156 InstrOffs = MI.getOperand(ImmIdx).getImm(); 2157 NumBits = 12; 2158 break; 2159 } 2160 case ARMII::AddrMode2: { 2161 ImmIdx = FrameRegIdx+2; 2162 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm()); 2163 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub) 2164 InstrOffs *= -1; 2165 NumBits = 12; 2166 break; 2167 } 2168 case ARMII::AddrMode3: { 2169 ImmIdx = FrameRegIdx+2; 2170 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm()); 2171 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub) 2172 InstrOffs *= -1; 2173 NumBits = 8; 2174 break; 2175 } 2176 case ARMII::AddrMode4: 2177 case ARMII::AddrMode6: 2178 // Can't fold any offset even if it's zero. 2179 return false; 2180 case ARMII::AddrMode5: { 2181 ImmIdx = FrameRegIdx+1; 2182 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm()); 2183 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub) 2184 InstrOffs *= -1; 2185 NumBits = 8; 2186 Scale = 4; 2187 break; 2188 } 2189 default: 2190 llvm_unreachable("Unsupported addressing mode!"); 2191 } 2192 2193 Offset += InstrOffs * Scale; 2194 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!"); 2195 if (Offset < 0) { 2196 Offset = -Offset; 2197 isSub = true; 2198 } 2199 2200 // Attempt to fold address comp. if opcode has offset bits 2201 if (NumBits > 0) { 2202 // Common case: small offset, fits into instruction. 2203 MachineOperand &ImmOp = MI.getOperand(ImmIdx); 2204 int ImmedOffset = Offset / Scale; 2205 unsigned Mask = (1 << NumBits) - 1; 2206 if ((unsigned)Offset <= Mask * Scale) { 2207 // Replace the FrameIndex with sp 2208 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 2209 // FIXME: When addrmode2 goes away, this will simplify (like the 2210 // T2 version), as the LDR.i12 versions don't need the encoding 2211 // tricks for the offset value. 2212 if (isSub) { 2213 if (AddrMode == ARMII::AddrMode_i12) 2214 ImmedOffset = -ImmedOffset; 2215 else 2216 ImmedOffset |= 1 << NumBits; 2217 } 2218 ImmOp.ChangeToImmediate(ImmedOffset); 2219 Offset = 0; 2220 return true; 2221 } 2222 2223 // Otherwise, it didn't fit. Pull in what we can to simplify the immed. 2224 ImmedOffset = ImmedOffset & Mask; 2225 if (isSub) { 2226 if (AddrMode == ARMII::AddrMode_i12) 2227 ImmedOffset = -ImmedOffset; 2228 else 2229 ImmedOffset |= 1 << NumBits; 2230 } 2231 ImmOp.ChangeToImmediate(ImmedOffset); 2232 Offset &= ~(Mask*Scale); 2233 } 2234 } 2235 2236 Offset = (isSub) ? -Offset : Offset; 2237 return Offset == 0; 2238} 2239 2240/// analyzeCompare - For a comparison instruction, return the source registers 2241/// in SrcReg and SrcReg2 if having two register operands, and the value it 2242/// compares against in CmpValue. Return true if the comparison instruction 2243/// can be analyzed. 2244bool ARMBaseInstrInfo:: 2245analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, unsigned &SrcReg2, 2246 int &CmpMask, int &CmpValue) const { 2247 switch (MI->getOpcode()) { 2248 default: break; 2249 case ARM::CMPri: 2250 case ARM::t2CMPri: 2251 SrcReg = MI->getOperand(0).getReg(); 2252 SrcReg2 = 0; 2253 CmpMask = ~0; 2254 CmpValue = MI->getOperand(1).getImm(); 2255 return true; 2256 case ARM::CMPrr: 2257 case ARM::t2CMPrr: 2258 SrcReg = MI->getOperand(0).getReg(); 2259 SrcReg2 = MI->getOperand(1).getReg(); 2260 CmpMask = ~0; 2261 CmpValue = 0; 2262 return true; 2263 case ARM::TSTri: 2264 case ARM::t2TSTri: 2265 SrcReg = MI->getOperand(0).getReg(); 2266 SrcReg2 = 0; 2267 CmpMask = MI->getOperand(1).getImm(); 2268 CmpValue = 0; 2269 return true; 2270 } 2271 2272 return false; 2273} 2274 2275/// isSuitableForMask - Identify a suitable 'and' instruction that 2276/// operates on the given source register and applies the same mask 2277/// as a 'tst' instruction. Provide a limited look-through for copies. 2278/// When successful, MI will hold the found instruction. 2279static bool isSuitableForMask(MachineInstr *&MI, unsigned SrcReg, 2280 int CmpMask, bool CommonUse) { 2281 switch (MI->getOpcode()) { 2282 case ARM::ANDri: 2283 case ARM::t2ANDri: 2284 if (CmpMask != MI->getOperand(2).getImm()) 2285 return false; 2286 if (SrcReg == MI->getOperand(CommonUse ? 1 : 0).getReg()) 2287 return true; 2288 break; 2289 case ARM::COPY: { 2290 // Walk down one instruction which is potentially an 'and'. 2291 const MachineInstr &Copy = *MI; 2292 MachineBasicBlock::iterator AND( 2293 std::next(MachineBasicBlock::iterator(MI))); 2294 if (AND == MI->getParent()->end()) return false; 2295 MI = AND; 2296 return isSuitableForMask(MI, Copy.getOperand(0).getReg(), 2297 CmpMask, true); 2298 } 2299 } 2300 2301 return false; 2302} 2303 2304/// getSwappedCondition - assume the flags are set by MI(a,b), return 2305/// the condition code if we modify the instructions such that flags are 2306/// set by MI(b,a). 2307inline static ARMCC::CondCodes getSwappedCondition(ARMCC::CondCodes CC) { 2308 switch (CC) { 2309 default: return ARMCC::AL; 2310 case ARMCC::EQ: return ARMCC::EQ; 2311 case ARMCC::NE: return ARMCC::NE; 2312 case ARMCC::HS: return ARMCC::LS; 2313 case ARMCC::LO: return ARMCC::HI; 2314 case ARMCC::HI: return ARMCC::LO; 2315 case ARMCC::LS: return ARMCC::HS; 2316 case ARMCC::GE: return ARMCC::LE; 2317 case ARMCC::LT: return ARMCC::GT; 2318 case ARMCC::GT: return ARMCC::LT; 2319 case ARMCC::LE: return ARMCC::GE; 2320 } 2321} 2322 2323/// isRedundantFlagInstr - check whether the first instruction, whose only 2324/// purpose is to update flags, can be made redundant. 2325/// CMPrr can be made redundant by SUBrr if the operands are the same. 2326/// CMPri can be made redundant by SUBri if the operands are the same. 2327/// This function can be extended later on. 2328inline static bool isRedundantFlagInstr(MachineInstr *CmpI, unsigned SrcReg, 2329 unsigned SrcReg2, int ImmValue, 2330 MachineInstr *OI) { 2331 if ((CmpI->getOpcode() == ARM::CMPrr || 2332 CmpI->getOpcode() == ARM::t2CMPrr) && 2333 (OI->getOpcode() == ARM::SUBrr || 2334 OI->getOpcode() == ARM::t2SUBrr) && 2335 ((OI->getOperand(1).getReg() == SrcReg && 2336 OI->getOperand(2).getReg() == SrcReg2) || 2337 (OI->getOperand(1).getReg() == SrcReg2 && 2338 OI->getOperand(2).getReg() == SrcReg))) 2339 return true; 2340 2341 if ((CmpI->getOpcode() == ARM::CMPri || 2342 CmpI->getOpcode() == ARM::t2CMPri) && 2343 (OI->getOpcode() == ARM::SUBri || 2344 OI->getOpcode() == ARM::t2SUBri) && 2345 OI->getOperand(1).getReg() == SrcReg && 2346 OI->getOperand(2).getImm() == ImmValue) 2347 return true; 2348 return false; 2349} 2350 2351/// optimizeCompareInstr - Convert the instruction supplying the argument to the 2352/// comparison into one that sets the zero bit in the flags register; 2353/// Remove a redundant Compare instruction if an earlier instruction can set the 2354/// flags in the same way as Compare. 2355/// E.g. SUBrr(r1,r2) and CMPrr(r1,r2). We also handle the case where two 2356/// operands are swapped: SUBrr(r1,r2) and CMPrr(r2,r1), by updating the 2357/// condition code of instructions which use the flags. 2358bool ARMBaseInstrInfo:: 2359optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2, 2360 int CmpMask, int CmpValue, 2361 const MachineRegisterInfo *MRI) const { 2362 // Get the unique definition of SrcReg. 2363 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg); 2364 if (!MI) return false; 2365 2366 // Masked compares sometimes use the same register as the corresponding 'and'. 2367 if (CmpMask != ~0) { 2368 if (!isSuitableForMask(MI, SrcReg, CmpMask, false) || isPredicated(MI)) { 2369 MI = nullptr; 2370 for (MachineRegisterInfo::use_instr_iterator 2371 UI = MRI->use_instr_begin(SrcReg), UE = MRI->use_instr_end(); 2372 UI != UE; ++UI) { 2373 if (UI->getParent() != CmpInstr->getParent()) continue; 2374 MachineInstr *PotentialAND = &*UI; 2375 if (!isSuitableForMask(PotentialAND, SrcReg, CmpMask, true) || 2376 isPredicated(PotentialAND)) 2377 continue; 2378 MI = PotentialAND; 2379 break; 2380 } 2381 if (!MI) return false; 2382 } 2383 } 2384 2385 // Get ready to iterate backward from CmpInstr. 2386 MachineBasicBlock::iterator I = CmpInstr, E = MI, 2387 B = CmpInstr->getParent()->begin(); 2388 2389 // Early exit if CmpInstr is at the beginning of the BB. 2390 if (I == B) return false; 2391 2392 // There are two possible candidates which can be changed to set CPSR: 2393 // One is MI, the other is a SUB instruction. 2394 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1). 2395 // For CMPri(r1, CmpValue), we are looking for SUBri(r1, CmpValue). 2396 MachineInstr *Sub = nullptr; 2397 if (SrcReg2 != 0) 2398 // MI is not a candidate for CMPrr. 2399 MI = nullptr; 2400 else if (MI->getParent() != CmpInstr->getParent() || CmpValue != 0) { 2401 // Conservatively refuse to convert an instruction which isn't in the same 2402 // BB as the comparison. 2403 // For CMPri w/ CmpValue != 0, a Sub may still be a candidate. 2404 // Thus we cannot return here. 2405 if (CmpInstr->getOpcode() == ARM::CMPri || 2406 CmpInstr->getOpcode() == ARM::t2CMPri) 2407 MI = nullptr; 2408 else 2409 return false; 2410 } 2411 2412 // Check that CPSR isn't set between the comparison instruction and the one we 2413 // want to change. At the same time, search for Sub. 2414 const TargetRegisterInfo *TRI = &getRegisterInfo(); 2415 --I; 2416 for (; I != E; --I) { 2417 const MachineInstr &Instr = *I; 2418 2419 if (Instr.modifiesRegister(ARM::CPSR, TRI) || 2420 Instr.readsRegister(ARM::CPSR, TRI)) 2421 // This instruction modifies or uses CPSR after the one we want to 2422 // change. We can't do this transformation. 2423 return false; 2424 2425 // Check whether CmpInstr can be made redundant by the current instruction. 2426 if (isRedundantFlagInstr(CmpInstr, SrcReg, SrcReg2, CmpValue, &*I)) { 2427 Sub = &*I; 2428 break; 2429 } 2430 2431 if (I == B) 2432 // The 'and' is below the comparison instruction. 2433 return false; 2434 } 2435 2436 // Return false if no candidates exist. 2437 if (!MI && !Sub) 2438 return false; 2439 2440 // The single candidate is called MI. 2441 if (!MI) MI = Sub; 2442 2443 // We can't use a predicated instruction - it doesn't always write the flags. 2444 if (isPredicated(MI)) 2445 return false; 2446 2447 switch (MI->getOpcode()) { 2448 default: break; 2449 case ARM::RSBrr: 2450 case ARM::RSBri: 2451 case ARM::RSCrr: 2452 case ARM::RSCri: 2453 case ARM::ADDrr: 2454 case ARM::ADDri: 2455 case ARM::ADCrr: 2456 case ARM::ADCri: 2457 case ARM::SUBrr: 2458 case ARM::SUBri: 2459 case ARM::SBCrr: 2460 case ARM::SBCri: 2461 case ARM::t2RSBri: 2462 case ARM::t2ADDrr: 2463 case ARM::t2ADDri: 2464 case ARM::t2ADCrr: 2465 case ARM::t2ADCri: 2466 case ARM::t2SUBrr: 2467 case ARM::t2SUBri: 2468 case ARM::t2SBCrr: 2469 case ARM::t2SBCri: 2470 case ARM::ANDrr: 2471 case ARM::ANDri: 2472 case ARM::t2ANDrr: 2473 case ARM::t2ANDri: 2474 case ARM::ORRrr: 2475 case ARM::ORRri: 2476 case ARM::t2ORRrr: 2477 case ARM::t2ORRri: 2478 case ARM::EORrr: 2479 case ARM::EORri: 2480 case ARM::t2EORrr: 2481 case ARM::t2EORri: { 2482 // Scan forward for the use of CPSR 2483 // When checking against MI: if it's a conditional code that requires 2484 // checking of the V bit or C bit, then this is not safe to do. 2485 // It is safe to remove CmpInstr if CPSR is redefined or killed. 2486 // If we are done with the basic block, we need to check whether CPSR is 2487 // live-out. 2488 SmallVector<std::pair<MachineOperand*, ARMCC::CondCodes>, 4> 2489 OperandsToUpdate; 2490 bool isSafe = false; 2491 I = CmpInstr; 2492 E = CmpInstr->getParent()->end(); 2493 while (!isSafe && ++I != E) { 2494 const MachineInstr &Instr = *I; 2495 for (unsigned IO = 0, EO = Instr.getNumOperands(); 2496 !isSafe && IO != EO; ++IO) { 2497 const MachineOperand &MO = Instr.getOperand(IO); 2498 if (MO.isRegMask() && MO.clobbersPhysReg(ARM::CPSR)) { 2499 isSafe = true; 2500 break; 2501 } 2502 if (!MO.isReg() || MO.getReg() != ARM::CPSR) 2503 continue; 2504 if (MO.isDef()) { 2505 isSafe = true; 2506 break; 2507 } 2508 // Condition code is after the operand before CPSR except for VSELs. 2509 ARMCC::CondCodes CC; 2510 bool IsInstrVSel = true; 2511 switch (Instr.getOpcode()) { 2512 default: 2513 IsInstrVSel = false; 2514 CC = (ARMCC::CondCodes)Instr.getOperand(IO - 1).getImm(); 2515 break; 2516 case ARM::VSELEQD: 2517 case ARM::VSELEQS: 2518 CC = ARMCC::EQ; 2519 break; 2520 case ARM::VSELGTD: 2521 case ARM::VSELGTS: 2522 CC = ARMCC::GT; 2523 break; 2524 case ARM::VSELGED: 2525 case ARM::VSELGES: 2526 CC = ARMCC::GE; 2527 break; 2528 case ARM::VSELVSS: 2529 case ARM::VSELVSD: 2530 CC = ARMCC::VS; 2531 break; 2532 } 2533 2534 if (Sub) { 2535 ARMCC::CondCodes NewCC = getSwappedCondition(CC); 2536 if (NewCC == ARMCC::AL) 2537 return false; 2538 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based 2539 // on CMP needs to be updated to be based on SUB. 2540 // Push the condition code operands to OperandsToUpdate. 2541 // If it is safe to remove CmpInstr, the condition code of these 2542 // operands will be modified. 2543 if (SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 && 2544 Sub->getOperand(2).getReg() == SrcReg) { 2545 // VSel doesn't support condition code update. 2546 if (IsInstrVSel) 2547 return false; 2548 OperandsToUpdate.push_back( 2549 std::make_pair(&((*I).getOperand(IO - 1)), NewCC)); 2550 } 2551 } else { 2552 // No Sub, so this is x = <op> y, z; cmp x, 0. 2553 switch (CC) { 2554 case ARMCC::EQ: // Z 2555 case ARMCC::NE: // Z 2556 case ARMCC::MI: // N 2557 case ARMCC::PL: // N 2558 case ARMCC::AL: // none 2559 // CPSR can be used multiple times, we should continue. 2560 break; 2561 case ARMCC::HS: // C 2562 case ARMCC::LO: // C 2563 case ARMCC::VS: // V 2564 case ARMCC::VC: // V 2565 case ARMCC::HI: // C Z 2566 case ARMCC::LS: // C Z 2567 case ARMCC::GE: // N V 2568 case ARMCC::LT: // N V 2569 case ARMCC::GT: // Z N V 2570 case ARMCC::LE: // Z N V 2571 // The instruction uses the V bit or C bit which is not safe. 2572 return false; 2573 } 2574 } 2575 } 2576 } 2577 2578 // If CPSR is not killed nor re-defined, we should check whether it is 2579 // live-out. If it is live-out, do not optimize. 2580 if (!isSafe) { 2581 MachineBasicBlock *MBB = CmpInstr->getParent(); 2582 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 2583 SE = MBB->succ_end(); SI != SE; ++SI) 2584 if ((*SI)->isLiveIn(ARM::CPSR)) 2585 return false; 2586 } 2587 2588 // Toggle the optional operand to CPSR. 2589 MI->getOperand(5).setReg(ARM::CPSR); 2590 MI->getOperand(5).setIsDef(true); 2591 assert(!isPredicated(MI) && "Can't use flags from predicated instruction"); 2592 CmpInstr->eraseFromParent(); 2593 2594 // Modify the condition code of operands in OperandsToUpdate. 2595 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to 2596 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc. 2597 for (unsigned i = 0, e = OperandsToUpdate.size(); i < e; i++) 2598 OperandsToUpdate[i].first->setImm(OperandsToUpdate[i].second); 2599 return true; 2600 } 2601 } 2602 2603 return false; 2604} 2605 2606bool ARMBaseInstrInfo::FoldImmediate(MachineInstr *UseMI, 2607 MachineInstr *DefMI, unsigned Reg, 2608 MachineRegisterInfo *MRI) const { 2609 // Fold large immediates into add, sub, or, xor. 2610 unsigned DefOpc = DefMI->getOpcode(); 2611 if (DefOpc != ARM::t2MOVi32imm && DefOpc != ARM::MOVi32imm) 2612 return false; 2613 if (!DefMI->getOperand(1).isImm()) 2614 // Could be t2MOVi32imm <ga:xx> 2615 return false; 2616 2617 if (!MRI->hasOneNonDBGUse(Reg)) 2618 return false; 2619 2620 const MCInstrDesc &DefMCID = DefMI->getDesc(); 2621 if (DefMCID.hasOptionalDef()) { 2622 unsigned NumOps = DefMCID.getNumOperands(); 2623 const MachineOperand &MO = DefMI->getOperand(NumOps-1); 2624 if (MO.getReg() == ARM::CPSR && !MO.isDead()) 2625 // If DefMI defines CPSR and it is not dead, it's obviously not safe 2626 // to delete DefMI. 2627 return false; 2628 } 2629 2630 const MCInstrDesc &UseMCID = UseMI->getDesc(); 2631 if (UseMCID.hasOptionalDef()) { 2632 unsigned NumOps = UseMCID.getNumOperands(); 2633 if (UseMI->getOperand(NumOps-1).getReg() == ARM::CPSR) 2634 // If the instruction sets the flag, do not attempt this optimization 2635 // since it may change the semantics of the code. 2636 return false; 2637 } 2638 2639 unsigned UseOpc = UseMI->getOpcode(); 2640 unsigned NewUseOpc = 0; 2641 uint32_t ImmVal = (uint32_t)DefMI->getOperand(1).getImm(); 2642 uint32_t SOImmValV1 = 0, SOImmValV2 = 0; 2643 bool Commute = false; 2644 switch (UseOpc) { 2645 default: return false; 2646 case ARM::SUBrr: 2647 case ARM::ADDrr: 2648 case ARM::ORRrr: 2649 case ARM::EORrr: 2650 case ARM::t2SUBrr: 2651 case ARM::t2ADDrr: 2652 case ARM::t2ORRrr: 2653 case ARM::t2EORrr: { 2654 Commute = UseMI->getOperand(2).getReg() != Reg; 2655 switch (UseOpc) { 2656 default: break; 2657 case ARM::SUBrr: { 2658 if (Commute) 2659 return false; 2660 ImmVal = -ImmVal; 2661 NewUseOpc = ARM::SUBri; 2662 // Fallthrough 2663 } 2664 case ARM::ADDrr: 2665 case ARM::ORRrr: 2666 case ARM::EORrr: { 2667 if (!ARM_AM::isSOImmTwoPartVal(ImmVal)) 2668 return false; 2669 SOImmValV1 = (uint32_t)ARM_AM::getSOImmTwoPartFirst(ImmVal); 2670 SOImmValV2 = (uint32_t)ARM_AM::getSOImmTwoPartSecond(ImmVal); 2671 switch (UseOpc) { 2672 default: break; 2673 case ARM::ADDrr: NewUseOpc = ARM::ADDri; break; 2674 case ARM::ORRrr: NewUseOpc = ARM::ORRri; break; 2675 case ARM::EORrr: NewUseOpc = ARM::EORri; break; 2676 } 2677 break; 2678 } 2679 case ARM::t2SUBrr: { 2680 if (Commute) 2681 return false; 2682 ImmVal = -ImmVal; 2683 NewUseOpc = ARM::t2SUBri; 2684 // Fallthrough 2685 } 2686 case ARM::t2ADDrr: 2687 case ARM::t2ORRrr: 2688 case ARM::t2EORrr: { 2689 if (!ARM_AM::isT2SOImmTwoPartVal(ImmVal)) 2690 return false; 2691 SOImmValV1 = (uint32_t)ARM_AM::getT2SOImmTwoPartFirst(ImmVal); 2692 SOImmValV2 = (uint32_t)ARM_AM::getT2SOImmTwoPartSecond(ImmVal); 2693 switch (UseOpc) { 2694 default: break; 2695 case ARM::t2ADDrr: NewUseOpc = ARM::t2ADDri; break; 2696 case ARM::t2ORRrr: NewUseOpc = ARM::t2ORRri; break; 2697 case ARM::t2EORrr: NewUseOpc = ARM::t2EORri; break; 2698 } 2699 break; 2700 } 2701 } 2702 } 2703 } 2704 2705 unsigned OpIdx = Commute ? 2 : 1; 2706 unsigned Reg1 = UseMI->getOperand(OpIdx).getReg(); 2707 bool isKill = UseMI->getOperand(OpIdx).isKill(); 2708 unsigned NewReg = MRI->createVirtualRegister(MRI->getRegClass(Reg)); 2709 AddDefaultCC(AddDefaultPred(BuildMI(*UseMI->getParent(), 2710 UseMI, UseMI->getDebugLoc(), 2711 get(NewUseOpc), NewReg) 2712 .addReg(Reg1, getKillRegState(isKill)) 2713 .addImm(SOImmValV1))); 2714 UseMI->setDesc(get(NewUseOpc)); 2715 UseMI->getOperand(1).setReg(NewReg); 2716 UseMI->getOperand(1).setIsKill(); 2717 UseMI->getOperand(2).ChangeToImmediate(SOImmValV2); 2718 DefMI->eraseFromParent(); 2719 return true; 2720} 2721 2722static unsigned getNumMicroOpsSwiftLdSt(const InstrItineraryData *ItinData, 2723 const MachineInstr *MI) { 2724 switch (MI->getOpcode()) { 2725 default: { 2726 const MCInstrDesc &Desc = MI->getDesc(); 2727 int UOps = ItinData->getNumMicroOps(Desc.getSchedClass()); 2728 assert(UOps >= 0 && "bad # UOps"); 2729 return UOps; 2730 } 2731 2732 case ARM::LDRrs: 2733 case ARM::LDRBrs: 2734 case ARM::STRrs: 2735 case ARM::STRBrs: { 2736 unsigned ShOpVal = MI->getOperand(3).getImm(); 2737 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; 2738 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); 2739 if (!isSub && 2740 (ShImm == 0 || 2741 ((ShImm == 1 || ShImm == 2 || ShImm == 3) && 2742 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) 2743 return 1; 2744 return 2; 2745 } 2746 2747 case ARM::LDRH: 2748 case ARM::STRH: { 2749 if (!MI->getOperand(2).getReg()) 2750 return 1; 2751 2752 unsigned ShOpVal = MI->getOperand(3).getImm(); 2753 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; 2754 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); 2755 if (!isSub && 2756 (ShImm == 0 || 2757 ((ShImm == 1 || ShImm == 2 || ShImm == 3) && 2758 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) 2759 return 1; 2760 return 2; 2761 } 2762 2763 case ARM::LDRSB: 2764 case ARM::LDRSH: 2765 return (ARM_AM::getAM3Op(MI->getOperand(3).getImm()) == ARM_AM::sub) ? 3:2; 2766 2767 case ARM::LDRSB_POST: 2768 case ARM::LDRSH_POST: { 2769 unsigned Rt = MI->getOperand(0).getReg(); 2770 unsigned Rm = MI->getOperand(3).getReg(); 2771 return (Rt == Rm) ? 4 : 3; 2772 } 2773 2774 case ARM::LDR_PRE_REG: 2775 case ARM::LDRB_PRE_REG: { 2776 unsigned Rt = MI->getOperand(0).getReg(); 2777 unsigned Rm = MI->getOperand(3).getReg(); 2778 if (Rt == Rm) 2779 return 3; 2780 unsigned ShOpVal = MI->getOperand(4).getImm(); 2781 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; 2782 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); 2783 if (!isSub && 2784 (ShImm == 0 || 2785 ((ShImm == 1 || ShImm == 2 || ShImm == 3) && 2786 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) 2787 return 2; 2788 return 3; 2789 } 2790 2791 case ARM::STR_PRE_REG: 2792 case ARM::STRB_PRE_REG: { 2793 unsigned ShOpVal = MI->getOperand(4).getImm(); 2794 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; 2795 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); 2796 if (!isSub && 2797 (ShImm == 0 || 2798 ((ShImm == 1 || ShImm == 2 || ShImm == 3) && 2799 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) 2800 return 2; 2801 return 3; 2802 } 2803 2804 case ARM::LDRH_PRE: 2805 case ARM::STRH_PRE: { 2806 unsigned Rt = MI->getOperand(0).getReg(); 2807 unsigned Rm = MI->getOperand(3).getReg(); 2808 if (!Rm) 2809 return 2; 2810 if (Rt == Rm) 2811 return 3; 2812 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) 2813 ? 3 : 2; 2814 } 2815 2816 case ARM::LDR_POST_REG: 2817 case ARM::LDRB_POST_REG: 2818 case ARM::LDRH_POST: { 2819 unsigned Rt = MI->getOperand(0).getReg(); 2820 unsigned Rm = MI->getOperand(3).getReg(); 2821 return (Rt == Rm) ? 3 : 2; 2822 } 2823 2824 case ARM::LDR_PRE_IMM: 2825 case ARM::LDRB_PRE_IMM: 2826 case ARM::LDR_POST_IMM: 2827 case ARM::LDRB_POST_IMM: 2828 case ARM::STRB_POST_IMM: 2829 case ARM::STRB_POST_REG: 2830 case ARM::STRB_PRE_IMM: 2831 case ARM::STRH_POST: 2832 case ARM::STR_POST_IMM: 2833 case ARM::STR_POST_REG: 2834 case ARM::STR_PRE_IMM: 2835 return 2; 2836 2837 case ARM::LDRSB_PRE: 2838 case ARM::LDRSH_PRE: { 2839 unsigned Rm = MI->getOperand(3).getReg(); 2840 if (Rm == 0) 2841 return 3; 2842 unsigned Rt = MI->getOperand(0).getReg(); 2843 if (Rt == Rm) 2844 return 4; 2845 unsigned ShOpVal = MI->getOperand(4).getImm(); 2846 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; 2847 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); 2848 if (!isSub && 2849 (ShImm == 0 || 2850 ((ShImm == 1 || ShImm == 2 || ShImm == 3) && 2851 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) 2852 return 3; 2853 return 4; 2854 } 2855 2856 case ARM::LDRD: { 2857 unsigned Rt = MI->getOperand(0).getReg(); 2858 unsigned Rn = MI->getOperand(2).getReg(); 2859 unsigned Rm = MI->getOperand(3).getReg(); 2860 if (Rm) 2861 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3; 2862 return (Rt == Rn) ? 3 : 2; 2863 } 2864 2865 case ARM::STRD: { 2866 unsigned Rm = MI->getOperand(3).getReg(); 2867 if (Rm) 2868 return (ARM_AM::getAM3Op(MI->getOperand(4).getImm()) == ARM_AM::sub) ?4:3; 2869 return 2; 2870 } 2871 2872 case ARM::LDRD_POST: 2873 case ARM::t2LDRD_POST: 2874 return 3; 2875 2876 case ARM::STRD_POST: 2877 case ARM::t2STRD_POST: 2878 return 4; 2879 2880 case ARM::LDRD_PRE: { 2881 unsigned Rt = MI->getOperand(0).getReg(); 2882 unsigned Rn = MI->getOperand(3).getReg(); 2883 unsigned Rm = MI->getOperand(4).getReg(); 2884 if (Rm) 2885 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4; 2886 return (Rt == Rn) ? 4 : 3; 2887 } 2888 2889 case ARM::t2LDRD_PRE: { 2890 unsigned Rt = MI->getOperand(0).getReg(); 2891 unsigned Rn = MI->getOperand(3).getReg(); 2892 return (Rt == Rn) ? 4 : 3; 2893 } 2894 2895 case ARM::STRD_PRE: { 2896 unsigned Rm = MI->getOperand(4).getReg(); 2897 if (Rm) 2898 return (ARM_AM::getAM3Op(MI->getOperand(5).getImm()) == ARM_AM::sub) ?5:4; 2899 return 3; 2900 } 2901 2902 case ARM::t2STRD_PRE: 2903 return 3; 2904 2905 case ARM::t2LDR_POST: 2906 case ARM::t2LDRB_POST: 2907 case ARM::t2LDRB_PRE: 2908 case ARM::t2LDRSBi12: 2909 case ARM::t2LDRSBi8: 2910 case ARM::t2LDRSBpci: 2911 case ARM::t2LDRSBs: 2912 case ARM::t2LDRH_POST: 2913 case ARM::t2LDRH_PRE: 2914 case ARM::t2LDRSBT: 2915 case ARM::t2LDRSB_POST: 2916 case ARM::t2LDRSB_PRE: 2917 case ARM::t2LDRSH_POST: 2918 case ARM::t2LDRSH_PRE: 2919 case ARM::t2LDRSHi12: 2920 case ARM::t2LDRSHi8: 2921 case ARM::t2LDRSHpci: 2922 case ARM::t2LDRSHs: 2923 return 2; 2924 2925 case ARM::t2LDRDi8: { 2926 unsigned Rt = MI->getOperand(0).getReg(); 2927 unsigned Rn = MI->getOperand(2).getReg(); 2928 return (Rt == Rn) ? 3 : 2; 2929 } 2930 2931 case ARM::t2STRB_POST: 2932 case ARM::t2STRB_PRE: 2933 case ARM::t2STRBs: 2934 case ARM::t2STRDi8: 2935 case ARM::t2STRH_POST: 2936 case ARM::t2STRH_PRE: 2937 case ARM::t2STRHs: 2938 case ARM::t2STR_POST: 2939 case ARM::t2STR_PRE: 2940 case ARM::t2STRs: 2941 return 2; 2942 } 2943} 2944 2945// Return the number of 32-bit words loaded by LDM or stored by STM. If this 2946// can't be easily determined return 0 (missing MachineMemOperand). 2947// 2948// FIXME: The current MachineInstr design does not support relying on machine 2949// mem operands to determine the width of a memory access. Instead, we expect 2950// the target to provide this information based on the instruction opcode and 2951// operands. However, using MachineMemOperand is the best solution now for 2952// two reasons: 2953// 2954// 1) getNumMicroOps tries to infer LDM memory width from the total number of MI 2955// operands. This is much more dangerous than using the MachineMemOperand 2956// sizes because CodeGen passes can insert/remove optional machine operands. In 2957// fact, it's totally incorrect for preRA passes and appears to be wrong for 2958// postRA passes as well. 2959// 2960// 2) getNumLDMAddresses is only used by the scheduling machine model and any 2961// machine model that calls this should handle the unknown (zero size) case. 2962// 2963// Long term, we should require a target hook that verifies MachineMemOperand 2964// sizes during MC lowering. That target hook should be local to MC lowering 2965// because we can't ensure that it is aware of other MI forms. Doing this will 2966// ensure that MachineMemOperands are correctly propagated through all passes. 2967unsigned ARMBaseInstrInfo::getNumLDMAddresses(const MachineInstr *MI) const { 2968 unsigned Size = 0; 2969 for (MachineInstr::mmo_iterator I = MI->memoperands_begin(), 2970 E = MI->memoperands_end(); I != E; ++I) { 2971 Size += (*I)->getSize(); 2972 } 2973 return Size / 4; 2974} 2975 2976unsigned 2977ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, 2978 const MachineInstr *MI) const { 2979 if (!ItinData || ItinData->isEmpty()) 2980 return 1; 2981 2982 const MCInstrDesc &Desc = MI->getDesc(); 2983 unsigned Class = Desc.getSchedClass(); 2984 int ItinUOps = ItinData->getNumMicroOps(Class); 2985 if (ItinUOps >= 0) { 2986 if (Subtarget.isSwift() && (Desc.mayLoad() || Desc.mayStore())) 2987 return getNumMicroOpsSwiftLdSt(ItinData, MI); 2988 2989 return ItinUOps; 2990 } 2991 2992 unsigned Opc = MI->getOpcode(); 2993 switch (Opc) { 2994 default: 2995 llvm_unreachable("Unexpected multi-uops instruction!"); 2996 case ARM::VLDMQIA: 2997 case ARM::VSTMQIA: 2998 return 2; 2999 3000 // The number of uOps for load / store multiple are determined by the number 3001 // registers. 3002 // 3003 // On Cortex-A8, each pair of register loads / stores can be scheduled on the 3004 // same cycle. The scheduling for the first load / store must be done 3005 // separately by assuming the address is not 64-bit aligned. 3006 // 3007 // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address 3008 // is not 64-bit aligned, then AGU would take an extra cycle. For VFP / NEON 3009 // load / store multiple, the formula is (#reg / 2) + (#reg % 2) + 1. 3010 case ARM::VLDMDIA: 3011 case ARM::VLDMDIA_UPD: 3012 case ARM::VLDMDDB_UPD: 3013 case ARM::VLDMSIA: 3014 case ARM::VLDMSIA_UPD: 3015 case ARM::VLDMSDB_UPD: 3016 case ARM::VSTMDIA: 3017 case ARM::VSTMDIA_UPD: 3018 case ARM::VSTMDDB_UPD: 3019 case ARM::VSTMSIA: 3020 case ARM::VSTMSIA_UPD: 3021 case ARM::VSTMSDB_UPD: { 3022 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands(); 3023 return (NumRegs / 2) + (NumRegs % 2) + 1; 3024 } 3025 3026 case ARM::LDMIA_RET: 3027 case ARM::LDMIA: 3028 case ARM::LDMDA: 3029 case ARM::LDMDB: 3030 case ARM::LDMIB: 3031 case ARM::LDMIA_UPD: 3032 case ARM::LDMDA_UPD: 3033 case ARM::LDMDB_UPD: 3034 case ARM::LDMIB_UPD: 3035 case ARM::STMIA: 3036 case ARM::STMDA: 3037 case ARM::STMDB: 3038 case ARM::STMIB: 3039 case ARM::STMIA_UPD: 3040 case ARM::STMDA_UPD: 3041 case ARM::STMDB_UPD: 3042 case ARM::STMIB_UPD: 3043 case ARM::tLDMIA: 3044 case ARM::tLDMIA_UPD: 3045 case ARM::tSTMIA_UPD: 3046 case ARM::tPOP_RET: 3047 case ARM::tPOP: 3048 case ARM::tPUSH: 3049 case ARM::t2LDMIA_RET: 3050 case ARM::t2LDMIA: 3051 case ARM::t2LDMDB: 3052 case ARM::t2LDMIA_UPD: 3053 case ARM::t2LDMDB_UPD: 3054 case ARM::t2STMIA: 3055 case ARM::t2STMDB: 3056 case ARM::t2STMIA_UPD: 3057 case ARM::t2STMDB_UPD: { 3058 unsigned NumRegs = MI->getNumOperands() - Desc.getNumOperands() + 1; 3059 if (Subtarget.isSwift()) { 3060 int UOps = 1 + NumRegs; // One for address computation, one for each ld / st. 3061 switch (Opc) { 3062 default: break; 3063 case ARM::VLDMDIA_UPD: 3064 case ARM::VLDMDDB_UPD: 3065 case ARM::VLDMSIA_UPD: 3066 case ARM::VLDMSDB_UPD: 3067 case ARM::VSTMDIA_UPD: 3068 case ARM::VSTMDDB_UPD: 3069 case ARM::VSTMSIA_UPD: 3070 case ARM::VSTMSDB_UPD: 3071 case ARM::LDMIA_UPD: 3072 case ARM::LDMDA_UPD: 3073 case ARM::LDMDB_UPD: 3074 case ARM::LDMIB_UPD: 3075 case ARM::STMIA_UPD: 3076 case ARM::STMDA_UPD: 3077 case ARM::STMDB_UPD: 3078 case ARM::STMIB_UPD: 3079 case ARM::tLDMIA_UPD: 3080 case ARM::tSTMIA_UPD: 3081 case ARM::t2LDMIA_UPD: 3082 case ARM::t2LDMDB_UPD: 3083 case ARM::t2STMIA_UPD: 3084 case ARM::t2STMDB_UPD: 3085 ++UOps; // One for base register writeback. 3086 break; 3087 case ARM::LDMIA_RET: 3088 case ARM::tPOP_RET: 3089 case ARM::t2LDMIA_RET: 3090 UOps += 2; // One for base reg wb, one for write to pc. 3091 break; 3092 } 3093 return UOps; 3094 } else if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) { 3095 if (NumRegs < 4) 3096 return 2; 3097 // 4 registers would be issued: 2, 2. 3098 // 5 registers would be issued: 2, 2, 1. 3099 int A8UOps = (NumRegs / 2); 3100 if (NumRegs % 2) 3101 ++A8UOps; 3102 return A8UOps; 3103 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) { 3104 int A9UOps = (NumRegs / 2); 3105 // If there are odd number of registers or if it's not 64-bit aligned, 3106 // then it takes an extra AGU (Address Generation Unit) cycle. 3107 if ((NumRegs % 2) || 3108 !MI->hasOneMemOperand() || 3109 (*MI->memoperands_begin())->getAlignment() < 8) 3110 ++A9UOps; 3111 return A9UOps; 3112 } else { 3113 // Assume the worst. 3114 return NumRegs; 3115 } 3116 } 3117 } 3118} 3119 3120int 3121ARMBaseInstrInfo::getVLDMDefCycle(const InstrItineraryData *ItinData, 3122 const MCInstrDesc &DefMCID, 3123 unsigned DefClass, 3124 unsigned DefIdx, unsigned DefAlign) const { 3125 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1; 3126 if (RegNo <= 0) 3127 // Def is the address writeback. 3128 return ItinData->getOperandCycle(DefClass, DefIdx); 3129 3130 int DefCycle; 3131 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) { 3132 // (regno / 2) + (regno % 2) + 1 3133 DefCycle = RegNo / 2 + 1; 3134 if (RegNo % 2) 3135 ++DefCycle; 3136 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) { 3137 DefCycle = RegNo; 3138 bool isSLoad = false; 3139 3140 switch (DefMCID.getOpcode()) { 3141 default: break; 3142 case ARM::VLDMSIA: 3143 case ARM::VLDMSIA_UPD: 3144 case ARM::VLDMSDB_UPD: 3145 isSLoad = true; 3146 break; 3147 } 3148 3149 // If there are odd number of 'S' registers or if it's not 64-bit aligned, 3150 // then it takes an extra cycle. 3151 if ((isSLoad && (RegNo % 2)) || DefAlign < 8) 3152 ++DefCycle; 3153 } else { 3154 // Assume the worst. 3155 DefCycle = RegNo + 2; 3156 } 3157 3158 return DefCycle; 3159} 3160 3161int 3162ARMBaseInstrInfo::getLDMDefCycle(const InstrItineraryData *ItinData, 3163 const MCInstrDesc &DefMCID, 3164 unsigned DefClass, 3165 unsigned DefIdx, unsigned DefAlign) const { 3166 int RegNo = (int)(DefIdx+1) - DefMCID.getNumOperands() + 1; 3167 if (RegNo <= 0) 3168 // Def is the address writeback. 3169 return ItinData->getOperandCycle(DefClass, DefIdx); 3170 3171 int DefCycle; 3172 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) { 3173 // 4 registers would be issued: 1, 2, 1. 3174 // 5 registers would be issued: 1, 2, 2. 3175 DefCycle = RegNo / 2; 3176 if (DefCycle < 1) 3177 DefCycle = 1; 3178 // Result latency is issue cycle + 2: E2. 3179 DefCycle += 2; 3180 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) { 3181 DefCycle = (RegNo / 2); 3182 // If there are odd number of registers or if it's not 64-bit aligned, 3183 // then it takes an extra AGU (Address Generation Unit) cycle. 3184 if ((RegNo % 2) || DefAlign < 8) 3185 ++DefCycle; 3186 // Result latency is AGU cycles + 2. 3187 DefCycle += 2; 3188 } else { 3189 // Assume the worst. 3190 DefCycle = RegNo + 2; 3191 } 3192 3193 return DefCycle; 3194} 3195 3196int 3197ARMBaseInstrInfo::getVSTMUseCycle(const InstrItineraryData *ItinData, 3198 const MCInstrDesc &UseMCID, 3199 unsigned UseClass, 3200 unsigned UseIdx, unsigned UseAlign) const { 3201 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1; 3202 if (RegNo <= 0) 3203 return ItinData->getOperandCycle(UseClass, UseIdx); 3204 3205 int UseCycle; 3206 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) { 3207 // (regno / 2) + (regno % 2) + 1 3208 UseCycle = RegNo / 2 + 1; 3209 if (RegNo % 2) 3210 ++UseCycle; 3211 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) { 3212 UseCycle = RegNo; 3213 bool isSStore = false; 3214 3215 switch (UseMCID.getOpcode()) { 3216 default: break; 3217 case ARM::VSTMSIA: 3218 case ARM::VSTMSIA_UPD: 3219 case ARM::VSTMSDB_UPD: 3220 isSStore = true; 3221 break; 3222 } 3223 3224 // If there are odd number of 'S' registers or if it's not 64-bit aligned, 3225 // then it takes an extra cycle. 3226 if ((isSStore && (RegNo % 2)) || UseAlign < 8) 3227 ++UseCycle; 3228 } else { 3229 // Assume the worst. 3230 UseCycle = RegNo + 2; 3231 } 3232 3233 return UseCycle; 3234} 3235 3236int 3237ARMBaseInstrInfo::getSTMUseCycle(const InstrItineraryData *ItinData, 3238 const MCInstrDesc &UseMCID, 3239 unsigned UseClass, 3240 unsigned UseIdx, unsigned UseAlign) const { 3241 int RegNo = (int)(UseIdx+1) - UseMCID.getNumOperands() + 1; 3242 if (RegNo <= 0) 3243 return ItinData->getOperandCycle(UseClass, UseIdx); 3244 3245 int UseCycle; 3246 if (Subtarget.isCortexA8() || Subtarget.isCortexA7()) { 3247 UseCycle = RegNo / 2; 3248 if (UseCycle < 2) 3249 UseCycle = 2; 3250 // Read in E3. 3251 UseCycle += 2; 3252 } else if (Subtarget.isLikeA9() || Subtarget.isSwift()) { 3253 UseCycle = (RegNo / 2); 3254 // If there are odd number of registers or if it's not 64-bit aligned, 3255 // then it takes an extra AGU (Address Generation Unit) cycle. 3256 if ((RegNo % 2) || UseAlign < 8) 3257 ++UseCycle; 3258 } else { 3259 // Assume the worst. 3260 UseCycle = 1; 3261 } 3262 return UseCycle; 3263} 3264 3265int 3266ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 3267 const MCInstrDesc &DefMCID, 3268 unsigned DefIdx, unsigned DefAlign, 3269 const MCInstrDesc &UseMCID, 3270 unsigned UseIdx, unsigned UseAlign) const { 3271 unsigned DefClass = DefMCID.getSchedClass(); 3272 unsigned UseClass = UseMCID.getSchedClass(); 3273 3274 if (DefIdx < DefMCID.getNumDefs() && UseIdx < UseMCID.getNumOperands()) 3275 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 3276 3277 // This may be a def / use of a variable_ops instruction, the operand 3278 // latency might be determinable dynamically. Let the target try to 3279 // figure it out. 3280 int DefCycle = -1; 3281 bool LdmBypass = false; 3282 switch (DefMCID.getOpcode()) { 3283 default: 3284 DefCycle = ItinData->getOperandCycle(DefClass, DefIdx); 3285 break; 3286 3287 case ARM::VLDMDIA: 3288 case ARM::VLDMDIA_UPD: 3289 case ARM::VLDMDDB_UPD: 3290 case ARM::VLDMSIA: 3291 case ARM::VLDMSIA_UPD: 3292 case ARM::VLDMSDB_UPD: 3293 DefCycle = getVLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign); 3294 break; 3295 3296 case ARM::LDMIA_RET: 3297 case ARM::LDMIA: 3298 case ARM::LDMDA: 3299 case ARM::LDMDB: 3300 case ARM::LDMIB: 3301 case ARM::LDMIA_UPD: 3302 case ARM::LDMDA_UPD: 3303 case ARM::LDMDB_UPD: 3304 case ARM::LDMIB_UPD: 3305 case ARM::tLDMIA: 3306 case ARM::tLDMIA_UPD: 3307 case ARM::tPUSH: 3308 case ARM::t2LDMIA_RET: 3309 case ARM::t2LDMIA: 3310 case ARM::t2LDMDB: 3311 case ARM::t2LDMIA_UPD: 3312 case ARM::t2LDMDB_UPD: 3313 LdmBypass = 1; 3314 DefCycle = getLDMDefCycle(ItinData, DefMCID, DefClass, DefIdx, DefAlign); 3315 break; 3316 } 3317 3318 if (DefCycle == -1) 3319 // We can't seem to determine the result latency of the def, assume it's 2. 3320 DefCycle = 2; 3321 3322 int UseCycle = -1; 3323 switch (UseMCID.getOpcode()) { 3324 default: 3325 UseCycle = ItinData->getOperandCycle(UseClass, UseIdx); 3326 break; 3327 3328 case ARM::VSTMDIA: 3329 case ARM::VSTMDIA_UPD: 3330 case ARM::VSTMDDB_UPD: 3331 case ARM::VSTMSIA: 3332 case ARM::VSTMSIA_UPD: 3333 case ARM::VSTMSDB_UPD: 3334 UseCycle = getVSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign); 3335 break; 3336 3337 case ARM::STMIA: 3338 case ARM::STMDA: 3339 case ARM::STMDB: 3340 case ARM::STMIB: 3341 case ARM::STMIA_UPD: 3342 case ARM::STMDA_UPD: 3343 case ARM::STMDB_UPD: 3344 case ARM::STMIB_UPD: 3345 case ARM::tSTMIA_UPD: 3346 case ARM::tPOP_RET: 3347 case ARM::tPOP: 3348 case ARM::t2STMIA: 3349 case ARM::t2STMDB: 3350 case ARM::t2STMIA_UPD: 3351 case ARM::t2STMDB_UPD: 3352 UseCycle = getSTMUseCycle(ItinData, UseMCID, UseClass, UseIdx, UseAlign); 3353 break; 3354 } 3355 3356 if (UseCycle == -1) 3357 // Assume it's read in the first stage. 3358 UseCycle = 1; 3359 3360 UseCycle = DefCycle - UseCycle + 1; 3361 if (UseCycle > 0) { 3362 if (LdmBypass) { 3363 // It's a variable_ops instruction so we can't use DefIdx here. Just use 3364 // first def operand. 3365 if (ItinData->hasPipelineForwarding(DefClass, DefMCID.getNumOperands()-1, 3366 UseClass, UseIdx)) 3367 --UseCycle; 3368 } else if (ItinData->hasPipelineForwarding(DefClass, DefIdx, 3369 UseClass, UseIdx)) { 3370 --UseCycle; 3371 } 3372 } 3373 3374 return UseCycle; 3375} 3376 3377static const MachineInstr *getBundledDefMI(const TargetRegisterInfo *TRI, 3378 const MachineInstr *MI, unsigned Reg, 3379 unsigned &DefIdx, unsigned &Dist) { 3380 Dist = 0; 3381 3382 MachineBasicBlock::const_iterator I = MI; ++I; 3383 MachineBasicBlock::const_instr_iterator II = std::prev(I.getInstrIterator()); 3384 assert(II->isInsideBundle() && "Empty bundle?"); 3385 3386 int Idx = -1; 3387 while (II->isInsideBundle()) { 3388 Idx = II->findRegisterDefOperandIdx(Reg, false, true, TRI); 3389 if (Idx != -1) 3390 break; 3391 --II; 3392 ++Dist; 3393 } 3394 3395 assert(Idx != -1 && "Cannot find bundled definition!"); 3396 DefIdx = Idx; 3397 return II; 3398} 3399 3400static const MachineInstr *getBundledUseMI(const TargetRegisterInfo *TRI, 3401 const MachineInstr *MI, unsigned Reg, 3402 unsigned &UseIdx, unsigned &Dist) { 3403 Dist = 0; 3404 3405 MachineBasicBlock::const_instr_iterator II = MI; ++II; 3406 assert(II->isInsideBundle() && "Empty bundle?"); 3407 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end(); 3408 3409 // FIXME: This doesn't properly handle multiple uses. 3410 int Idx = -1; 3411 while (II != E && II->isInsideBundle()) { 3412 Idx = II->findRegisterUseOperandIdx(Reg, false, TRI); 3413 if (Idx != -1) 3414 break; 3415 if (II->getOpcode() != ARM::t2IT) 3416 ++Dist; 3417 ++II; 3418 } 3419 3420 if (Idx == -1) { 3421 Dist = 0; 3422 return nullptr; 3423 } 3424 3425 UseIdx = Idx; 3426 return II; 3427} 3428 3429/// Return the number of cycles to add to (or subtract from) the static 3430/// itinerary based on the def opcode and alignment. The caller will ensure that 3431/// adjusted latency is at least one cycle. 3432static int adjustDefLatency(const ARMSubtarget &Subtarget, 3433 const MachineInstr *DefMI, 3434 const MCInstrDesc *DefMCID, unsigned DefAlign) { 3435 int Adjust = 0; 3436 if (Subtarget.isCortexA8() || Subtarget.isLikeA9() || Subtarget.isCortexA7()) { 3437 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2] 3438 // variants are one cycle cheaper. 3439 switch (DefMCID->getOpcode()) { 3440 default: break; 3441 case ARM::LDRrs: 3442 case ARM::LDRBrs: { 3443 unsigned ShOpVal = DefMI->getOperand(3).getImm(); 3444 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); 3445 if (ShImm == 0 || 3446 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)) 3447 --Adjust; 3448 break; 3449 } 3450 case ARM::t2LDRs: 3451 case ARM::t2LDRBs: 3452 case ARM::t2LDRHs: 3453 case ARM::t2LDRSHs: { 3454 // Thumb2 mode: lsl only. 3455 unsigned ShAmt = DefMI->getOperand(3).getImm(); 3456 if (ShAmt == 0 || ShAmt == 2) 3457 --Adjust; 3458 break; 3459 } 3460 } 3461 } else if (Subtarget.isSwift()) { 3462 // FIXME: Properly handle all of the latency adjustments for address 3463 // writeback. 3464 switch (DefMCID->getOpcode()) { 3465 default: break; 3466 case ARM::LDRrs: 3467 case ARM::LDRBrs: { 3468 unsigned ShOpVal = DefMI->getOperand(3).getImm(); 3469 bool isSub = ARM_AM::getAM2Op(ShOpVal) == ARM_AM::sub; 3470 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); 3471 if (!isSub && 3472 (ShImm == 0 || 3473 ((ShImm == 1 || ShImm == 2 || ShImm == 3) && 3474 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl))) 3475 Adjust -= 2; 3476 else if (!isSub && 3477 ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr) 3478 --Adjust; 3479 break; 3480 } 3481 case ARM::t2LDRs: 3482 case ARM::t2LDRBs: 3483 case ARM::t2LDRHs: 3484 case ARM::t2LDRSHs: { 3485 // Thumb2 mode: lsl only. 3486 unsigned ShAmt = DefMI->getOperand(3).getImm(); 3487 if (ShAmt == 0 || ShAmt == 1 || ShAmt == 2 || ShAmt == 3) 3488 Adjust -= 2; 3489 break; 3490 } 3491 } 3492 } 3493 3494 if (DefAlign < 8 && Subtarget.isLikeA9()) { 3495 switch (DefMCID->getOpcode()) { 3496 default: break; 3497 case ARM::VLD1q8: 3498 case ARM::VLD1q16: 3499 case ARM::VLD1q32: 3500 case ARM::VLD1q64: 3501 case ARM::VLD1q8wb_fixed: 3502 case ARM::VLD1q16wb_fixed: 3503 case ARM::VLD1q32wb_fixed: 3504 case ARM::VLD1q64wb_fixed: 3505 case ARM::VLD1q8wb_register: 3506 case ARM::VLD1q16wb_register: 3507 case ARM::VLD1q32wb_register: 3508 case ARM::VLD1q64wb_register: 3509 case ARM::VLD2d8: 3510 case ARM::VLD2d16: 3511 case ARM::VLD2d32: 3512 case ARM::VLD2q8: 3513 case ARM::VLD2q16: 3514 case ARM::VLD2q32: 3515 case ARM::VLD2d8wb_fixed: 3516 case ARM::VLD2d16wb_fixed: 3517 case ARM::VLD2d32wb_fixed: 3518 case ARM::VLD2q8wb_fixed: 3519 case ARM::VLD2q16wb_fixed: 3520 case ARM::VLD2q32wb_fixed: 3521 case ARM::VLD2d8wb_register: 3522 case ARM::VLD2d16wb_register: 3523 case ARM::VLD2d32wb_register: 3524 case ARM::VLD2q8wb_register: 3525 case ARM::VLD2q16wb_register: 3526 case ARM::VLD2q32wb_register: 3527 case ARM::VLD3d8: 3528 case ARM::VLD3d16: 3529 case ARM::VLD3d32: 3530 case ARM::VLD1d64T: 3531 case ARM::VLD3d8_UPD: 3532 case ARM::VLD3d16_UPD: 3533 case ARM::VLD3d32_UPD: 3534 case ARM::VLD1d64Twb_fixed: 3535 case ARM::VLD1d64Twb_register: 3536 case ARM::VLD3q8_UPD: 3537 case ARM::VLD3q16_UPD: 3538 case ARM::VLD3q32_UPD: 3539 case ARM::VLD4d8: 3540 case ARM::VLD4d16: 3541 case ARM::VLD4d32: 3542 case ARM::VLD1d64Q: 3543 case ARM::VLD4d8_UPD: 3544 case ARM::VLD4d16_UPD: 3545 case ARM::VLD4d32_UPD: 3546 case ARM::VLD1d64Qwb_fixed: 3547 case ARM::VLD1d64Qwb_register: 3548 case ARM::VLD4q8_UPD: 3549 case ARM::VLD4q16_UPD: 3550 case ARM::VLD4q32_UPD: 3551 case ARM::VLD1DUPq8: 3552 case ARM::VLD1DUPq16: 3553 case ARM::VLD1DUPq32: 3554 case ARM::VLD1DUPq8wb_fixed: 3555 case ARM::VLD1DUPq16wb_fixed: 3556 case ARM::VLD1DUPq32wb_fixed: 3557 case ARM::VLD1DUPq8wb_register: 3558 case ARM::VLD1DUPq16wb_register: 3559 case ARM::VLD1DUPq32wb_register: 3560 case ARM::VLD2DUPd8: 3561 case ARM::VLD2DUPd16: 3562 case ARM::VLD2DUPd32: 3563 case ARM::VLD2DUPd8wb_fixed: 3564 case ARM::VLD2DUPd16wb_fixed: 3565 case ARM::VLD2DUPd32wb_fixed: 3566 case ARM::VLD2DUPd8wb_register: 3567 case ARM::VLD2DUPd16wb_register: 3568 case ARM::VLD2DUPd32wb_register: 3569 case ARM::VLD4DUPd8: 3570 case ARM::VLD4DUPd16: 3571 case ARM::VLD4DUPd32: 3572 case ARM::VLD4DUPd8_UPD: 3573 case ARM::VLD4DUPd16_UPD: 3574 case ARM::VLD4DUPd32_UPD: 3575 case ARM::VLD1LNd8: 3576 case ARM::VLD1LNd16: 3577 case ARM::VLD1LNd32: 3578 case ARM::VLD1LNd8_UPD: 3579 case ARM::VLD1LNd16_UPD: 3580 case ARM::VLD1LNd32_UPD: 3581 case ARM::VLD2LNd8: 3582 case ARM::VLD2LNd16: 3583 case ARM::VLD2LNd32: 3584 case ARM::VLD2LNq16: 3585 case ARM::VLD2LNq32: 3586 case ARM::VLD2LNd8_UPD: 3587 case ARM::VLD2LNd16_UPD: 3588 case ARM::VLD2LNd32_UPD: 3589 case ARM::VLD2LNq16_UPD: 3590 case ARM::VLD2LNq32_UPD: 3591 case ARM::VLD4LNd8: 3592 case ARM::VLD4LNd16: 3593 case ARM::VLD4LNd32: 3594 case ARM::VLD4LNq16: 3595 case ARM::VLD4LNq32: 3596 case ARM::VLD4LNd8_UPD: 3597 case ARM::VLD4LNd16_UPD: 3598 case ARM::VLD4LNd32_UPD: 3599 case ARM::VLD4LNq16_UPD: 3600 case ARM::VLD4LNq32_UPD: 3601 // If the address is not 64-bit aligned, the latencies of these 3602 // instructions increases by one. 3603 ++Adjust; 3604 break; 3605 } 3606 } 3607 return Adjust; 3608} 3609 3610 3611 3612int 3613ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 3614 const MachineInstr *DefMI, unsigned DefIdx, 3615 const MachineInstr *UseMI, 3616 unsigned UseIdx) const { 3617 // No operand latency. The caller may fall back to getInstrLatency. 3618 if (!ItinData || ItinData->isEmpty()) 3619 return -1; 3620 3621 const MachineOperand &DefMO = DefMI->getOperand(DefIdx); 3622 unsigned Reg = DefMO.getReg(); 3623 const MCInstrDesc *DefMCID = &DefMI->getDesc(); 3624 const MCInstrDesc *UseMCID = &UseMI->getDesc(); 3625 3626 unsigned DefAdj = 0; 3627 if (DefMI->isBundle()) { 3628 DefMI = getBundledDefMI(&getRegisterInfo(), DefMI, Reg, DefIdx, DefAdj); 3629 DefMCID = &DefMI->getDesc(); 3630 } 3631 if (DefMI->isCopyLike() || DefMI->isInsertSubreg() || 3632 DefMI->isRegSequence() || DefMI->isImplicitDef()) { 3633 return 1; 3634 } 3635 3636 unsigned UseAdj = 0; 3637 if (UseMI->isBundle()) { 3638 unsigned NewUseIdx; 3639 const MachineInstr *NewUseMI = getBundledUseMI(&getRegisterInfo(), UseMI, 3640 Reg, NewUseIdx, UseAdj); 3641 if (!NewUseMI) 3642 return -1; 3643 3644 UseMI = NewUseMI; 3645 UseIdx = NewUseIdx; 3646 UseMCID = &UseMI->getDesc(); 3647 } 3648 3649 if (Reg == ARM::CPSR) { 3650 if (DefMI->getOpcode() == ARM::FMSTAT) { 3651 // fpscr -> cpsr stalls over 20 cycles on A8 (and earlier?) 3652 return Subtarget.isLikeA9() ? 1 : 20; 3653 } 3654 3655 // CPSR set and branch can be paired in the same cycle. 3656 if (UseMI->isBranch()) 3657 return 0; 3658 3659 // Otherwise it takes the instruction latency (generally one). 3660 unsigned Latency = getInstrLatency(ItinData, DefMI); 3661 3662 // For Thumb2 and -Os, prefer scheduling CPSR setting instruction close to 3663 // its uses. Instructions which are otherwise scheduled between them may 3664 // incur a code size penalty (not able to use the CPSR setting 16-bit 3665 // instructions). 3666 if (Latency > 0 && Subtarget.isThumb2()) { 3667 const MachineFunction *MF = DefMI->getParent()->getParent(); 3668 if (MF->getFunction()->hasFnAttribute(Attribute::OptimizeForSize)) 3669 --Latency; 3670 } 3671 return Latency; 3672 } 3673 3674 if (DefMO.isImplicit() || UseMI->getOperand(UseIdx).isImplicit()) 3675 return -1; 3676 3677 unsigned DefAlign = DefMI->hasOneMemOperand() 3678 ? (*DefMI->memoperands_begin())->getAlignment() : 0; 3679 unsigned UseAlign = UseMI->hasOneMemOperand() 3680 ? (*UseMI->memoperands_begin())->getAlignment() : 0; 3681 3682 // Get the itinerary's latency if possible, and handle variable_ops. 3683 int Latency = getOperandLatency(ItinData, *DefMCID, DefIdx, DefAlign, 3684 *UseMCID, UseIdx, UseAlign); 3685 // Unable to find operand latency. The caller may resort to getInstrLatency. 3686 if (Latency < 0) 3687 return Latency; 3688 3689 // Adjust for IT block position. 3690 int Adj = DefAdj + UseAdj; 3691 3692 // Adjust for dynamic def-side opcode variants not captured by the itinerary. 3693 Adj += adjustDefLatency(Subtarget, DefMI, DefMCID, DefAlign); 3694 if (Adj >= 0 || (int)Latency > -Adj) { 3695 return Latency + Adj; 3696 } 3697 // Return the itinerary latency, which may be zero but not less than zero. 3698 return Latency; 3699} 3700 3701int 3702ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 3703 SDNode *DefNode, unsigned DefIdx, 3704 SDNode *UseNode, unsigned UseIdx) const { 3705 if (!DefNode->isMachineOpcode()) 3706 return 1; 3707 3708 const MCInstrDesc &DefMCID = get(DefNode->getMachineOpcode()); 3709 3710 if (isZeroCost(DefMCID.Opcode)) 3711 return 0; 3712 3713 if (!ItinData || ItinData->isEmpty()) 3714 return DefMCID.mayLoad() ? 3 : 1; 3715 3716 if (!UseNode->isMachineOpcode()) { 3717 int Latency = ItinData->getOperandCycle(DefMCID.getSchedClass(), DefIdx); 3718 if (Subtarget.isLikeA9() || Subtarget.isSwift()) 3719 return Latency <= 2 ? 1 : Latency - 1; 3720 else 3721 return Latency <= 3 ? 1 : Latency - 2; 3722 } 3723 3724 const MCInstrDesc &UseMCID = get(UseNode->getMachineOpcode()); 3725 const MachineSDNode *DefMN = dyn_cast<MachineSDNode>(DefNode); 3726 unsigned DefAlign = !DefMN->memoperands_empty() 3727 ? (*DefMN->memoperands_begin())->getAlignment() : 0; 3728 const MachineSDNode *UseMN = dyn_cast<MachineSDNode>(UseNode); 3729 unsigned UseAlign = !UseMN->memoperands_empty() 3730 ? (*UseMN->memoperands_begin())->getAlignment() : 0; 3731 int Latency = getOperandLatency(ItinData, DefMCID, DefIdx, DefAlign, 3732 UseMCID, UseIdx, UseAlign); 3733 3734 if (Latency > 1 && 3735 (Subtarget.isCortexA8() || Subtarget.isLikeA9() || 3736 Subtarget.isCortexA7())) { 3737 // FIXME: Shifter op hack: no shift (i.e. [r +/- r]) or [r + r << 2] 3738 // variants are one cycle cheaper. 3739 switch (DefMCID.getOpcode()) { 3740 default: break; 3741 case ARM::LDRrs: 3742 case ARM::LDRBrs: { 3743 unsigned ShOpVal = 3744 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue(); 3745 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); 3746 if (ShImm == 0 || 3747 (ShImm == 2 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)) 3748 --Latency; 3749 break; 3750 } 3751 case ARM::t2LDRs: 3752 case ARM::t2LDRBs: 3753 case ARM::t2LDRHs: 3754 case ARM::t2LDRSHs: { 3755 // Thumb2 mode: lsl only. 3756 unsigned ShAmt = 3757 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue(); 3758 if (ShAmt == 0 || ShAmt == 2) 3759 --Latency; 3760 break; 3761 } 3762 } 3763 } else if (DefIdx == 0 && Latency > 2 && Subtarget.isSwift()) { 3764 // FIXME: Properly handle all of the latency adjustments for address 3765 // writeback. 3766 switch (DefMCID.getOpcode()) { 3767 default: break; 3768 case ARM::LDRrs: 3769 case ARM::LDRBrs: { 3770 unsigned ShOpVal = 3771 cast<ConstantSDNode>(DefNode->getOperand(2))->getZExtValue(); 3772 unsigned ShImm = ARM_AM::getAM2Offset(ShOpVal); 3773 if (ShImm == 0 || 3774 ((ShImm == 1 || ShImm == 2 || ShImm == 3) && 3775 ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsl)) 3776 Latency -= 2; 3777 else if (ShImm == 1 && ARM_AM::getAM2ShiftOpc(ShOpVal) == ARM_AM::lsr) 3778 --Latency; 3779 break; 3780 } 3781 case ARM::t2LDRs: 3782 case ARM::t2LDRBs: 3783 case ARM::t2LDRHs: 3784 case ARM::t2LDRSHs: { 3785 // Thumb2 mode: lsl 0-3 only. 3786 Latency -= 2; 3787 break; 3788 } 3789 } 3790 } 3791 3792 if (DefAlign < 8 && Subtarget.isLikeA9()) 3793 switch (DefMCID.getOpcode()) { 3794 default: break; 3795 case ARM::VLD1q8: 3796 case ARM::VLD1q16: 3797 case ARM::VLD1q32: 3798 case ARM::VLD1q64: 3799 case ARM::VLD1q8wb_register: 3800 case ARM::VLD1q16wb_register: 3801 case ARM::VLD1q32wb_register: 3802 case ARM::VLD1q64wb_register: 3803 case ARM::VLD1q8wb_fixed: 3804 case ARM::VLD1q16wb_fixed: 3805 case ARM::VLD1q32wb_fixed: 3806 case ARM::VLD1q64wb_fixed: 3807 case ARM::VLD2d8: 3808 case ARM::VLD2d16: 3809 case ARM::VLD2d32: 3810 case ARM::VLD2q8Pseudo: 3811 case ARM::VLD2q16Pseudo: 3812 case ARM::VLD2q32Pseudo: 3813 case ARM::VLD2d8wb_fixed: 3814 case ARM::VLD2d16wb_fixed: 3815 case ARM::VLD2d32wb_fixed: 3816 case ARM::VLD2q8PseudoWB_fixed: 3817 case ARM::VLD2q16PseudoWB_fixed: 3818 case ARM::VLD2q32PseudoWB_fixed: 3819 case ARM::VLD2d8wb_register: 3820 case ARM::VLD2d16wb_register: 3821 case ARM::VLD2d32wb_register: 3822 case ARM::VLD2q8PseudoWB_register: 3823 case ARM::VLD2q16PseudoWB_register: 3824 case ARM::VLD2q32PseudoWB_register: 3825 case ARM::VLD3d8Pseudo: 3826 case ARM::VLD3d16Pseudo: 3827 case ARM::VLD3d32Pseudo: 3828 case ARM::VLD1d64TPseudo: 3829 case ARM::VLD1d64TPseudoWB_fixed: 3830 case ARM::VLD3d8Pseudo_UPD: 3831 case ARM::VLD3d16Pseudo_UPD: 3832 case ARM::VLD3d32Pseudo_UPD: 3833 case ARM::VLD3q8Pseudo_UPD: 3834 case ARM::VLD3q16Pseudo_UPD: 3835 case ARM::VLD3q32Pseudo_UPD: 3836 case ARM::VLD3q8oddPseudo: 3837 case ARM::VLD3q16oddPseudo: 3838 case ARM::VLD3q32oddPseudo: 3839 case ARM::VLD3q8oddPseudo_UPD: 3840 case ARM::VLD3q16oddPseudo_UPD: 3841 case ARM::VLD3q32oddPseudo_UPD: 3842 case ARM::VLD4d8Pseudo: 3843 case ARM::VLD4d16Pseudo: 3844 case ARM::VLD4d32Pseudo: 3845 case ARM::VLD1d64QPseudo: 3846 case ARM::VLD1d64QPseudoWB_fixed: 3847 case ARM::VLD4d8Pseudo_UPD: 3848 case ARM::VLD4d16Pseudo_UPD: 3849 case ARM::VLD4d32Pseudo_UPD: 3850 case ARM::VLD4q8Pseudo_UPD: 3851 case ARM::VLD4q16Pseudo_UPD: 3852 case ARM::VLD4q32Pseudo_UPD: 3853 case ARM::VLD4q8oddPseudo: 3854 case ARM::VLD4q16oddPseudo: 3855 case ARM::VLD4q32oddPseudo: 3856 case ARM::VLD4q8oddPseudo_UPD: 3857 case ARM::VLD4q16oddPseudo_UPD: 3858 case ARM::VLD4q32oddPseudo_UPD: 3859 case ARM::VLD1DUPq8: 3860 case ARM::VLD1DUPq16: 3861 case ARM::VLD1DUPq32: 3862 case ARM::VLD1DUPq8wb_fixed: 3863 case ARM::VLD1DUPq16wb_fixed: 3864 case ARM::VLD1DUPq32wb_fixed: 3865 case ARM::VLD1DUPq8wb_register: 3866 case ARM::VLD1DUPq16wb_register: 3867 case ARM::VLD1DUPq32wb_register: 3868 case ARM::VLD2DUPd8: 3869 case ARM::VLD2DUPd16: 3870 case ARM::VLD2DUPd32: 3871 case ARM::VLD2DUPd8wb_fixed: 3872 case ARM::VLD2DUPd16wb_fixed: 3873 case ARM::VLD2DUPd32wb_fixed: 3874 case ARM::VLD2DUPd8wb_register: 3875 case ARM::VLD2DUPd16wb_register: 3876 case ARM::VLD2DUPd32wb_register: 3877 case ARM::VLD4DUPd8Pseudo: 3878 case ARM::VLD4DUPd16Pseudo: 3879 case ARM::VLD4DUPd32Pseudo: 3880 case ARM::VLD4DUPd8Pseudo_UPD: 3881 case ARM::VLD4DUPd16Pseudo_UPD: 3882 case ARM::VLD4DUPd32Pseudo_UPD: 3883 case ARM::VLD1LNq8Pseudo: 3884 case ARM::VLD1LNq16Pseudo: 3885 case ARM::VLD1LNq32Pseudo: 3886 case ARM::VLD1LNq8Pseudo_UPD: 3887 case ARM::VLD1LNq16Pseudo_UPD: 3888 case ARM::VLD1LNq32Pseudo_UPD: 3889 case ARM::VLD2LNd8Pseudo: 3890 case ARM::VLD2LNd16Pseudo: 3891 case ARM::VLD2LNd32Pseudo: 3892 case ARM::VLD2LNq16Pseudo: 3893 case ARM::VLD2LNq32Pseudo: 3894 case ARM::VLD2LNd8Pseudo_UPD: 3895 case ARM::VLD2LNd16Pseudo_UPD: 3896 case ARM::VLD2LNd32Pseudo_UPD: 3897 case ARM::VLD2LNq16Pseudo_UPD: 3898 case ARM::VLD2LNq32Pseudo_UPD: 3899 case ARM::VLD4LNd8Pseudo: 3900 case ARM::VLD4LNd16Pseudo: 3901 case ARM::VLD4LNd32Pseudo: 3902 case ARM::VLD4LNq16Pseudo: 3903 case ARM::VLD4LNq32Pseudo: 3904 case ARM::VLD4LNd8Pseudo_UPD: 3905 case ARM::VLD4LNd16Pseudo_UPD: 3906 case ARM::VLD4LNd32Pseudo_UPD: 3907 case ARM::VLD4LNq16Pseudo_UPD: 3908 case ARM::VLD4LNq32Pseudo_UPD: 3909 // If the address is not 64-bit aligned, the latencies of these 3910 // instructions increases by one. 3911 ++Latency; 3912 break; 3913 } 3914 3915 return Latency; 3916} 3917 3918unsigned ARMBaseInstrInfo::getPredicationCost(const MachineInstr *MI) const { 3919 if (MI->isCopyLike() || MI->isInsertSubreg() || 3920 MI->isRegSequence() || MI->isImplicitDef()) 3921 return 0; 3922 3923 if (MI->isBundle()) 3924 return 0; 3925 3926 const MCInstrDesc &MCID = MI->getDesc(); 3927 3928 if (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR)) { 3929 // When predicated, CPSR is an additional source operand for CPSR updating 3930 // instructions, this apparently increases their latencies. 3931 return 1; 3932 } 3933 return 0; 3934} 3935 3936unsigned ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 3937 const MachineInstr *MI, 3938 unsigned *PredCost) const { 3939 if (MI->isCopyLike() || MI->isInsertSubreg() || 3940 MI->isRegSequence() || MI->isImplicitDef()) 3941 return 1; 3942 3943 // An instruction scheduler typically runs on unbundled instructions, however 3944 // other passes may query the latency of a bundled instruction. 3945 if (MI->isBundle()) { 3946 unsigned Latency = 0; 3947 MachineBasicBlock::const_instr_iterator I = MI; 3948 MachineBasicBlock::const_instr_iterator E = MI->getParent()->instr_end(); 3949 while (++I != E && I->isInsideBundle()) { 3950 if (I->getOpcode() != ARM::t2IT) 3951 Latency += getInstrLatency(ItinData, I, PredCost); 3952 } 3953 return Latency; 3954 } 3955 3956 const MCInstrDesc &MCID = MI->getDesc(); 3957 if (PredCost && (MCID.isCall() || MCID.hasImplicitDefOfPhysReg(ARM::CPSR))) { 3958 // When predicated, CPSR is an additional source operand for CPSR updating 3959 // instructions, this apparently increases their latencies. 3960 *PredCost = 1; 3961 } 3962 // Be sure to call getStageLatency for an empty itinerary in case it has a 3963 // valid MinLatency property. 3964 if (!ItinData) 3965 return MI->mayLoad() ? 3 : 1; 3966 3967 unsigned Class = MCID.getSchedClass(); 3968 3969 // For instructions with variable uops, use uops as latency. 3970 if (!ItinData->isEmpty() && ItinData->getNumMicroOps(Class) < 0) 3971 return getNumMicroOps(ItinData, MI); 3972 3973 // For the common case, fall back on the itinerary's latency. 3974 unsigned Latency = ItinData->getStageLatency(Class); 3975 3976 // Adjust for dynamic def-side opcode variants not captured by the itinerary. 3977 unsigned DefAlign = MI->hasOneMemOperand() 3978 ? (*MI->memoperands_begin())->getAlignment() : 0; 3979 int Adj = adjustDefLatency(Subtarget, MI, &MCID, DefAlign); 3980 if (Adj >= 0 || (int)Latency > -Adj) { 3981 return Latency + Adj; 3982 } 3983 return Latency; 3984} 3985 3986int ARMBaseInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 3987 SDNode *Node) const { 3988 if (!Node->isMachineOpcode()) 3989 return 1; 3990 3991 if (!ItinData || ItinData->isEmpty()) 3992 return 1; 3993 3994 unsigned Opcode = Node->getMachineOpcode(); 3995 switch (Opcode) { 3996 default: 3997 return ItinData->getStageLatency(get(Opcode).getSchedClass()); 3998 case ARM::VLDMQIA: 3999 case ARM::VSTMQIA: 4000 return 2; 4001 } 4002} 4003 4004bool ARMBaseInstrInfo:: 4005hasHighOperandLatency(const InstrItineraryData *ItinData, 4006 const MachineRegisterInfo *MRI, 4007 const MachineInstr *DefMI, unsigned DefIdx, 4008 const MachineInstr *UseMI, unsigned UseIdx) const { 4009 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask; 4010 unsigned UDomain = UseMI->getDesc().TSFlags & ARMII::DomainMask; 4011 if (Subtarget.isCortexA8() && 4012 (DDomain == ARMII::DomainVFP || UDomain == ARMII::DomainVFP)) 4013 // CortexA8 VFP instructions are not pipelined. 4014 return true; 4015 4016 // Hoist VFP / NEON instructions with 4 or higher latency. 4017 int Latency = computeOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx); 4018 if (Latency < 0) 4019 Latency = getInstrLatency(ItinData, DefMI); 4020 if (Latency <= 3) 4021 return false; 4022 return DDomain == ARMII::DomainVFP || DDomain == ARMII::DomainNEON || 4023 UDomain == ARMII::DomainVFP || UDomain == ARMII::DomainNEON; 4024} 4025 4026bool ARMBaseInstrInfo:: 4027hasLowDefLatency(const InstrItineraryData *ItinData, 4028 const MachineInstr *DefMI, unsigned DefIdx) const { 4029 if (!ItinData || ItinData->isEmpty()) 4030 return false; 4031 4032 unsigned DDomain = DefMI->getDesc().TSFlags & ARMII::DomainMask; 4033 if (DDomain == ARMII::DomainGeneral) { 4034 unsigned DefClass = DefMI->getDesc().getSchedClass(); 4035 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx); 4036 return (DefCycle != -1 && DefCycle <= 2); 4037 } 4038 return false; 4039} 4040 4041bool ARMBaseInstrInfo::verifyInstruction(const MachineInstr *MI, 4042 StringRef &ErrInfo) const { 4043 if (convertAddSubFlagsOpcode(MI->getOpcode())) { 4044 ErrInfo = "Pseudo flag setting opcodes only exist in Selection DAG"; 4045 return false; 4046 } 4047 return true; 4048} 4049 4050// LoadStackGuard has so far only been implemented for MachO. Different code 4051// sequence is needed for other targets. 4052void ARMBaseInstrInfo::expandLoadStackGuardBase(MachineBasicBlock::iterator MI, 4053 unsigned LoadImmOpc, 4054 unsigned LoadOpc, 4055 Reloc::Model RM) const { 4056 MachineBasicBlock &MBB = *MI->getParent(); 4057 DebugLoc DL = MI->getDebugLoc(); 4058 unsigned Reg = MI->getOperand(0).getReg(); 4059 const GlobalValue *GV = 4060 cast<GlobalValue>((*MI->memoperands_begin())->getValue()); 4061 MachineInstrBuilder MIB; 4062 4063 BuildMI(MBB, MI, DL, get(LoadImmOpc), Reg) 4064 .addGlobalAddress(GV, 0, ARMII::MO_NONLAZY); 4065 4066 if (Subtarget.GVIsIndirectSymbol(GV, RM)) { 4067 MIB = BuildMI(MBB, MI, DL, get(LoadOpc), Reg); 4068 MIB.addReg(Reg, RegState::Kill).addImm(0); 4069 unsigned Flag = MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant; 4070 MachineMemOperand *MMO = MBB.getParent()-> 4071 getMachineMemOperand(MachinePointerInfo::getGOT(), Flag, 4, 4); 4072 MIB.addMemOperand(MMO); 4073 AddDefaultPred(MIB); 4074 } 4075 4076 MIB = BuildMI(MBB, MI, DL, get(LoadOpc), Reg); 4077 MIB.addReg(Reg, RegState::Kill).addImm(0); 4078 MIB.setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 4079 AddDefaultPred(MIB); 4080} 4081 4082bool 4083ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc, 4084 unsigned &AddSubOpc, 4085 bool &NegAcc, bool &HasLane) const { 4086 DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode); 4087 if (I == MLxEntryMap.end()) 4088 return false; 4089 4090 const ARM_MLxEntry &Entry = ARM_MLxTable[I->second]; 4091 MulOpc = Entry.MulOpc; 4092 AddSubOpc = Entry.AddSubOpc; 4093 NegAcc = Entry.NegAcc; 4094 HasLane = Entry.HasLane; 4095 return true; 4096} 4097 4098//===----------------------------------------------------------------------===// 4099// Execution domains. 4100//===----------------------------------------------------------------------===// 4101// 4102// Some instructions go down the NEON pipeline, some go down the VFP pipeline, 4103// and some can go down both. The vmov instructions go down the VFP pipeline, 4104// but they can be changed to vorr equivalents that are executed by the NEON 4105// pipeline. 4106// 4107// We use the following execution domain numbering: 4108// 4109enum ARMExeDomain { 4110 ExeGeneric = 0, 4111 ExeVFP = 1, 4112 ExeNEON = 2 4113}; 4114// 4115// Also see ARMInstrFormats.td and Domain* enums in ARMBaseInfo.h 4116// 4117std::pair<uint16_t, uint16_t> 4118ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const { 4119 // If we don't have access to NEON instructions then we won't be able 4120 // to swizzle anything to the NEON domain. Check to make sure. 4121 if (Subtarget.hasNEON()) { 4122 // VMOVD, VMOVRS and VMOVSR are VFP instructions, but can be changed to NEON 4123 // if they are not predicated. 4124 if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI)) 4125 return std::make_pair(ExeVFP, (1 << ExeVFP) | (1 << ExeNEON)); 4126 4127 // CortexA9 is particularly picky about mixing the two and wants these 4128 // converted. 4129 if (Subtarget.isCortexA9() && !isPredicated(MI) && 4130 (MI->getOpcode() == ARM::VMOVRS || MI->getOpcode() == ARM::VMOVSR || 4131 MI->getOpcode() == ARM::VMOVS)) 4132 return std::make_pair(ExeVFP, (1 << ExeVFP) | (1 << ExeNEON)); 4133 } 4134 // No other instructions can be swizzled, so just determine their domain. 4135 unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask; 4136 4137 if (Domain & ARMII::DomainNEON) 4138 return std::make_pair(ExeNEON, 0); 4139 4140 // Certain instructions can go either way on Cortex-A8. 4141 // Treat them as NEON instructions. 4142 if ((Domain & ARMII::DomainNEONA8) && Subtarget.isCortexA8()) 4143 return std::make_pair(ExeNEON, 0); 4144 4145 if (Domain & ARMII::DomainVFP) 4146 return std::make_pair(ExeVFP, 0); 4147 4148 return std::make_pair(ExeGeneric, 0); 4149} 4150 4151static unsigned getCorrespondingDRegAndLane(const TargetRegisterInfo *TRI, 4152 unsigned SReg, unsigned &Lane) { 4153 unsigned DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_0, &ARM::DPRRegClass); 4154 Lane = 0; 4155 4156 if (DReg != ARM::NoRegister) 4157 return DReg; 4158 4159 Lane = 1; 4160 DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_1, &ARM::DPRRegClass); 4161 4162 assert(DReg && "S-register with no D super-register?"); 4163 return DReg; 4164} 4165 4166/// getImplicitSPRUseForDPRUse - Given a use of a DPR register and lane, 4167/// set ImplicitSReg to a register number that must be marked as implicit-use or 4168/// zero if no register needs to be defined as implicit-use. 4169/// 4170/// If the function cannot determine if an SPR should be marked implicit use or 4171/// not, it returns false. 4172/// 4173/// This function handles cases where an instruction is being modified from taking 4174/// an SPR to a DPR[Lane]. A use of the DPR is being added, which may conflict 4175/// with an earlier def of an SPR corresponding to DPR[Lane^1] (i.e. the other 4176/// lane of the DPR). 4177/// 4178/// If the other SPR is defined, an implicit-use of it should be added. Else, 4179/// (including the case where the DPR itself is defined), it should not. 4180/// 4181static bool getImplicitSPRUseForDPRUse(const TargetRegisterInfo *TRI, 4182 MachineInstr *MI, 4183 unsigned DReg, unsigned Lane, 4184 unsigned &ImplicitSReg) { 4185 // If the DPR is defined or used already, the other SPR lane will be chained 4186 // correctly, so there is nothing to be done. 4187 if (MI->definesRegister(DReg, TRI) || MI->readsRegister(DReg, TRI)) { 4188 ImplicitSReg = 0; 4189 return true; 4190 } 4191 4192 // Otherwise we need to go searching to see if the SPR is set explicitly. 4193 ImplicitSReg = TRI->getSubReg(DReg, 4194 (Lane & 1) ? ARM::ssub_0 : ARM::ssub_1); 4195 MachineBasicBlock::LivenessQueryResult LQR = 4196 MI->getParent()->computeRegisterLiveness(TRI, ImplicitSReg, MI); 4197 4198 if (LQR == MachineBasicBlock::LQR_Live) 4199 return true; 4200 else if (LQR == MachineBasicBlock::LQR_Unknown) 4201 return false; 4202 4203 // If the register is known not to be live, there is no need to add an 4204 // implicit-use. 4205 ImplicitSReg = 0; 4206 return true; 4207} 4208 4209void 4210ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const { 4211 unsigned DstReg, SrcReg, DReg; 4212 unsigned Lane; 4213 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI); 4214 const TargetRegisterInfo *TRI = &getRegisterInfo(); 4215 switch (MI->getOpcode()) { 4216 default: 4217 llvm_unreachable("cannot handle opcode!"); 4218 break; 4219 case ARM::VMOVD: 4220 if (Domain != ExeNEON) 4221 break; 4222 4223 // Zap the predicate operands. 4224 assert(!isPredicated(MI) && "Cannot predicate a VORRd"); 4225 4226 // Make sure we've got NEON instructions. 4227 assert(Subtarget.hasNEON() && "VORRd requires NEON"); 4228 4229 // Source instruction is %DDst = VMOVD %DSrc, 14, %noreg (; implicits) 4230 DstReg = MI->getOperand(0).getReg(); 4231 SrcReg = MI->getOperand(1).getReg(); 4232 4233 for (unsigned i = MI->getDesc().getNumOperands(); i; --i) 4234 MI->RemoveOperand(i-1); 4235 4236 // Change to a %DDst = VORRd %DSrc, %DSrc, 14, %noreg (; implicits) 4237 MI->setDesc(get(ARM::VORRd)); 4238 AddDefaultPred(MIB.addReg(DstReg, RegState::Define) 4239 .addReg(SrcReg) 4240 .addReg(SrcReg)); 4241 break; 4242 case ARM::VMOVRS: 4243 if (Domain != ExeNEON) 4244 break; 4245 assert(!isPredicated(MI) && "Cannot predicate a VGETLN"); 4246 4247 // Source instruction is %RDst = VMOVRS %SSrc, 14, %noreg (; implicits) 4248 DstReg = MI->getOperand(0).getReg(); 4249 SrcReg = MI->getOperand(1).getReg(); 4250 4251 for (unsigned i = MI->getDesc().getNumOperands(); i; --i) 4252 MI->RemoveOperand(i-1); 4253 4254 DReg = getCorrespondingDRegAndLane(TRI, SrcReg, Lane); 4255 4256 // Convert to %RDst = VGETLNi32 %DSrc, Lane, 14, %noreg (; imps) 4257 // Note that DSrc has been widened and the other lane may be undef, which 4258 // contaminates the entire register. 4259 MI->setDesc(get(ARM::VGETLNi32)); 4260 AddDefaultPred(MIB.addReg(DstReg, RegState::Define) 4261 .addReg(DReg, RegState::Undef) 4262 .addImm(Lane)); 4263 4264 // The old source should be an implicit use, otherwise we might think it 4265 // was dead before here. 4266 MIB.addReg(SrcReg, RegState::Implicit); 4267 break; 4268 case ARM::VMOVSR: { 4269 if (Domain != ExeNEON) 4270 break; 4271 assert(!isPredicated(MI) && "Cannot predicate a VSETLN"); 4272 4273 // Source instruction is %SDst = VMOVSR %RSrc, 14, %noreg (; implicits) 4274 DstReg = MI->getOperand(0).getReg(); 4275 SrcReg = MI->getOperand(1).getReg(); 4276 4277 DReg = getCorrespondingDRegAndLane(TRI, DstReg, Lane); 4278 4279 unsigned ImplicitSReg; 4280 if (!getImplicitSPRUseForDPRUse(TRI, MI, DReg, Lane, ImplicitSReg)) 4281 break; 4282 4283 for (unsigned i = MI->getDesc().getNumOperands(); i; --i) 4284 MI->RemoveOperand(i-1); 4285 4286 // Convert to %DDst = VSETLNi32 %DDst, %RSrc, Lane, 14, %noreg (; imps) 4287 // Again DDst may be undefined at the beginning of this instruction. 4288 MI->setDesc(get(ARM::VSETLNi32)); 4289 MIB.addReg(DReg, RegState::Define) 4290 .addReg(DReg, getUndefRegState(!MI->readsRegister(DReg, TRI))) 4291 .addReg(SrcReg) 4292 .addImm(Lane); 4293 AddDefaultPred(MIB); 4294 4295 // The narrower destination must be marked as set to keep previous chains 4296 // in place. 4297 MIB.addReg(DstReg, RegState::Define | RegState::Implicit); 4298 if (ImplicitSReg != 0) 4299 MIB.addReg(ImplicitSReg, RegState::Implicit); 4300 break; 4301 } 4302 case ARM::VMOVS: { 4303 if (Domain != ExeNEON) 4304 break; 4305 4306 // Source instruction is %SDst = VMOVS %SSrc, 14, %noreg (; implicits) 4307 DstReg = MI->getOperand(0).getReg(); 4308 SrcReg = MI->getOperand(1).getReg(); 4309 4310 unsigned DstLane = 0, SrcLane = 0, DDst, DSrc; 4311 DDst = getCorrespondingDRegAndLane(TRI, DstReg, DstLane); 4312 DSrc = getCorrespondingDRegAndLane(TRI, SrcReg, SrcLane); 4313 4314 unsigned ImplicitSReg; 4315 if (!getImplicitSPRUseForDPRUse(TRI, MI, DSrc, SrcLane, ImplicitSReg)) 4316 break; 4317 4318 for (unsigned i = MI->getDesc().getNumOperands(); i; --i) 4319 MI->RemoveOperand(i-1); 4320 4321 if (DSrc == DDst) { 4322 // Destination can be: 4323 // %DDst = VDUPLN32d %DDst, Lane, 14, %noreg (; implicits) 4324 MI->setDesc(get(ARM::VDUPLN32d)); 4325 MIB.addReg(DDst, RegState::Define) 4326 .addReg(DDst, getUndefRegState(!MI->readsRegister(DDst, TRI))) 4327 .addImm(SrcLane); 4328 AddDefaultPred(MIB); 4329 4330 // Neither the source or the destination are naturally represented any 4331 // more, so add them in manually. 4332 MIB.addReg(DstReg, RegState::Implicit | RegState::Define); 4333 MIB.addReg(SrcReg, RegState::Implicit); 4334 if (ImplicitSReg != 0) 4335 MIB.addReg(ImplicitSReg, RegState::Implicit); 4336 break; 4337 } 4338 4339 // In general there's no single instruction that can perform an S <-> S 4340 // move in NEON space, but a pair of VEXT instructions *can* do the 4341 // job. It turns out that the VEXTs needed will only use DSrc once, with 4342 // the position based purely on the combination of lane-0 and lane-1 4343 // involved. For example 4344 // vmov s0, s2 -> vext.32 d0, d0, d1, #1 vext.32 d0, d0, d0, #1 4345 // vmov s1, s3 -> vext.32 d0, d1, d0, #1 vext.32 d0, d0, d0, #1 4346 // vmov s0, s3 -> vext.32 d0, d0, d0, #1 vext.32 d0, d1, d0, #1 4347 // vmov s1, s2 -> vext.32 d0, d0, d0, #1 vext.32 d0, d0, d1, #1 4348 // 4349 // Pattern of the MachineInstrs is: 4350 // %DDst = VEXTd32 %DSrc1, %DSrc2, Lane, 14, %noreg (;implicits) 4351 MachineInstrBuilder NewMIB; 4352 NewMIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), 4353 get(ARM::VEXTd32), DDst); 4354 4355 // On the first instruction, both DSrc and DDst may be <undef> if present. 4356 // Specifically when the original instruction didn't have them as an 4357 // <imp-use>. 4358 unsigned CurReg = SrcLane == 1 && DstLane == 1 ? DSrc : DDst; 4359 bool CurUndef = !MI->readsRegister(CurReg, TRI); 4360 NewMIB.addReg(CurReg, getUndefRegState(CurUndef)); 4361 4362 CurReg = SrcLane == 0 && DstLane == 0 ? DSrc : DDst; 4363 CurUndef = !MI->readsRegister(CurReg, TRI); 4364 NewMIB.addReg(CurReg, getUndefRegState(CurUndef)); 4365 4366 NewMIB.addImm(1); 4367 AddDefaultPred(NewMIB); 4368 4369 if (SrcLane == DstLane) 4370 NewMIB.addReg(SrcReg, RegState::Implicit); 4371 4372 MI->setDesc(get(ARM::VEXTd32)); 4373 MIB.addReg(DDst, RegState::Define); 4374 4375 // On the second instruction, DDst has definitely been defined above, so 4376 // it is not <undef>. DSrc, if present, can be <undef> as above. 4377 CurReg = SrcLane == 1 && DstLane == 0 ? DSrc : DDst; 4378 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI); 4379 MIB.addReg(CurReg, getUndefRegState(CurUndef)); 4380 4381 CurReg = SrcLane == 0 && DstLane == 1 ? DSrc : DDst; 4382 CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI); 4383 MIB.addReg(CurReg, getUndefRegState(CurUndef)); 4384 4385 MIB.addImm(1); 4386 AddDefaultPred(MIB); 4387 4388 if (SrcLane != DstLane) 4389 MIB.addReg(SrcReg, RegState::Implicit); 4390 4391 // As before, the original destination is no longer represented, add it 4392 // implicitly. 4393 MIB.addReg(DstReg, RegState::Define | RegState::Implicit); 4394 if (ImplicitSReg != 0) 4395 MIB.addReg(ImplicitSReg, RegState::Implicit); 4396 break; 4397 } 4398 } 4399 4400} 4401 4402//===----------------------------------------------------------------------===// 4403// Partial register updates 4404//===----------------------------------------------------------------------===// 4405// 4406// Swift renames NEON registers with 64-bit granularity. That means any 4407// instruction writing an S-reg implicitly reads the containing D-reg. The 4408// problem is mostly avoided by translating f32 operations to v2f32 operations 4409// on D-registers, but f32 loads are still a problem. 4410// 4411// These instructions can load an f32 into a NEON register: 4412// 4413// VLDRS - Only writes S, partial D update. 4414// VLD1LNd32 - Writes all D-regs, explicit partial D update, 2 uops. 4415// VLD1DUPd32 - Writes all D-regs, no partial reg update, 2 uops. 4416// 4417// FCONSTD can be used as a dependency-breaking instruction. 4418unsigned ARMBaseInstrInfo:: 4419getPartialRegUpdateClearance(const MachineInstr *MI, 4420 unsigned OpNum, 4421 const TargetRegisterInfo *TRI) const { 4422 if (!SwiftPartialUpdateClearance || 4423 !(Subtarget.isSwift() || Subtarget.isCortexA15())) 4424 return 0; 4425 4426 assert(TRI && "Need TRI instance"); 4427 4428 const MachineOperand &MO = MI->getOperand(OpNum); 4429 if (MO.readsReg()) 4430 return 0; 4431 unsigned Reg = MO.getReg(); 4432 int UseOp = -1; 4433 4434 switch(MI->getOpcode()) { 4435 // Normal instructions writing only an S-register. 4436 case ARM::VLDRS: 4437 case ARM::FCONSTS: 4438 case ARM::VMOVSR: 4439 case ARM::VMOVv8i8: 4440 case ARM::VMOVv4i16: 4441 case ARM::VMOVv2i32: 4442 case ARM::VMOVv2f32: 4443 case ARM::VMOVv1i64: 4444 UseOp = MI->findRegisterUseOperandIdx(Reg, false, TRI); 4445 break; 4446 4447 // Explicitly reads the dependency. 4448 case ARM::VLD1LNd32: 4449 UseOp = 3; 4450 break; 4451 default: 4452 return 0; 4453 } 4454 4455 // If this instruction actually reads a value from Reg, there is no unwanted 4456 // dependency. 4457 if (UseOp != -1 && MI->getOperand(UseOp).readsReg()) 4458 return 0; 4459 4460 // We must be able to clobber the whole D-reg. 4461 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 4462 // Virtual register must be a foo:ssub_0<def,undef> operand. 4463 if (!MO.getSubReg() || MI->readsVirtualRegister(Reg)) 4464 return 0; 4465 } else if (ARM::SPRRegClass.contains(Reg)) { 4466 // Physical register: MI must define the full D-reg. 4467 unsigned DReg = TRI->getMatchingSuperReg(Reg, ARM::ssub_0, 4468 &ARM::DPRRegClass); 4469 if (!DReg || !MI->definesRegister(DReg, TRI)) 4470 return 0; 4471 } 4472 4473 // MI has an unwanted D-register dependency. 4474 // Avoid defs in the previous N instructrions. 4475 return SwiftPartialUpdateClearance; 4476} 4477 4478// Break a partial register dependency after getPartialRegUpdateClearance 4479// returned non-zero. 4480void ARMBaseInstrInfo:: 4481breakPartialRegDependency(MachineBasicBlock::iterator MI, 4482 unsigned OpNum, 4483 const TargetRegisterInfo *TRI) const { 4484 assert(MI && OpNum < MI->getDesc().getNumDefs() && "OpNum is not a def"); 4485 assert(TRI && "Need TRI instance"); 4486 4487 const MachineOperand &MO = MI->getOperand(OpNum); 4488 unsigned Reg = MO.getReg(); 4489 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && 4490 "Can't break virtual register dependencies."); 4491 unsigned DReg = Reg; 4492 4493 // If MI defines an S-reg, find the corresponding D super-register. 4494 if (ARM::SPRRegClass.contains(Reg)) { 4495 DReg = ARM::D0 + (Reg - ARM::S0) / 2; 4496 assert(TRI->isSuperRegister(Reg, DReg) && "Register enums broken"); 4497 } 4498 4499 assert(ARM::DPRRegClass.contains(DReg) && "Can only break D-reg deps"); 4500 assert(MI->definesRegister(DReg, TRI) && "MI doesn't clobber full D-reg"); 4501 4502 // FIXME: In some cases, VLDRS can be changed to a VLD1DUPd32 which defines 4503 // the full D-register by loading the same value to both lanes. The 4504 // instruction is micro-coded with 2 uops, so don't do this until we can 4505 // properly schedule micro-coded instructions. The dispatcher stalls cause 4506 // too big regressions. 4507 4508 // Insert the dependency-breaking FCONSTD before MI. 4509 // 96 is the encoding of 0.5, but the actual value doesn't matter here. 4510 AddDefaultPred(BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), 4511 get(ARM::FCONSTD), DReg).addImm(96)); 4512 MI->addRegisterKilled(DReg, TRI, true); 4513} 4514 4515bool ARMBaseInstrInfo::hasNOP() const { 4516 return (Subtarget.getFeatureBits() & ARM::HasV6KOps) != 0; 4517} 4518 4519bool ARMBaseInstrInfo::isSwiftFastImmShift(const MachineInstr *MI) const { 4520 if (MI->getNumOperands() < 4) 4521 return true; 4522 unsigned ShOpVal = MI->getOperand(3).getImm(); 4523 unsigned ShImm = ARM_AM::getSORegOffset(ShOpVal); 4524 // Swift supports faster shifts for: lsl 2, lsl 1, and lsr 1. 4525 if ((ShImm == 1 && ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsr) || 4526 ((ShImm == 1 || ShImm == 2) && 4527 ARM_AM::getSORegShOp(ShOpVal) == ARM_AM::lsl)) 4528 return true; 4529 4530 return false; 4531} 4532 4533bool ARMBaseInstrInfo::getRegSequenceLikeInputs( 4534 const MachineInstr &MI, unsigned DefIdx, 4535 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const { 4536 assert(DefIdx < MI.getDesc().getNumDefs() && "Invalid definition index"); 4537 assert(MI.isRegSequenceLike() && "Invalid kind of instruction"); 4538 4539 switch (MI.getOpcode()) { 4540 case ARM::VMOVDRR: 4541 // dX = VMOVDRR rY, rZ 4542 // is the same as: 4543 // dX = REG_SEQUENCE rY, ssub_0, rZ, ssub_1 4544 // Populate the InputRegs accordingly. 4545 // rY 4546 const MachineOperand *MOReg = &MI.getOperand(1); 4547 InputRegs.push_back( 4548 RegSubRegPairAndIdx(MOReg->getReg(), MOReg->getSubReg(), ARM::ssub_0)); 4549 // rZ 4550 MOReg = &MI.getOperand(2); 4551 InputRegs.push_back( 4552 RegSubRegPairAndIdx(MOReg->getReg(), MOReg->getSubReg(), ARM::ssub_1)); 4553 return true; 4554 } 4555 llvm_unreachable("Target dependent opcode missing"); 4556} 4557 4558bool ARMBaseInstrInfo::getExtractSubregLikeInputs( 4559 const MachineInstr &MI, unsigned DefIdx, 4560 RegSubRegPairAndIdx &InputReg) const { 4561 assert(DefIdx < MI.getDesc().getNumDefs() && "Invalid definition index"); 4562 assert(MI.isExtractSubregLike() && "Invalid kind of instruction"); 4563 4564 switch (MI.getOpcode()) { 4565 case ARM::VMOVRRD: 4566 // rX, rY = VMOVRRD dZ 4567 // is the same as: 4568 // rX = EXTRACT_SUBREG dZ, ssub_0 4569 // rY = EXTRACT_SUBREG dZ, ssub_1 4570 const MachineOperand &MOReg = MI.getOperand(2); 4571 InputReg.Reg = MOReg.getReg(); 4572 InputReg.SubReg = MOReg.getSubReg(); 4573 InputReg.SubIdx = DefIdx == 0 ? ARM::ssub_0 : ARM::ssub_1; 4574 return true; 4575 } 4576 llvm_unreachable("Target dependent opcode missing"); 4577} 4578 4579bool ARMBaseInstrInfo::getInsertSubregLikeInputs( 4580 const MachineInstr &MI, unsigned DefIdx, RegSubRegPair &BaseReg, 4581 RegSubRegPairAndIdx &InsertedReg) const { 4582 assert(DefIdx < MI.getDesc().getNumDefs() && "Invalid definition index"); 4583 assert(MI.isInsertSubregLike() && "Invalid kind of instruction"); 4584 4585 switch (MI.getOpcode()) { 4586 case ARM::VSETLNi32: 4587 // dX = VSETLNi32 dY, rZ, imm 4588 const MachineOperand &MOBaseReg = MI.getOperand(1); 4589 const MachineOperand &MOInsertedReg = MI.getOperand(2); 4590 const MachineOperand &MOIndex = MI.getOperand(3); 4591 BaseReg.Reg = MOBaseReg.getReg(); 4592 BaseReg.SubReg = MOBaseReg.getSubReg(); 4593 4594 InsertedReg.Reg = MOInsertedReg.getReg(); 4595 InsertedReg.SubReg = MOInsertedReg.getSubReg(); 4596 InsertedReg.SubIdx = MOIndex.getImm() == 0 ? ARM::ssub_0 : ARM::ssub_1; 4597 return true; 4598 } 4599 llvm_unreachable("Target dependent opcode missing"); 4600} 4601