1 //===-- TargetInstrInfo.cpp - Target 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 implements the TargetInstrInfo class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Target/TargetInstrInfo.h" 15 #include "llvm/CodeGen/MachineFrameInfo.h" 16 #include "llvm/CodeGen/MachineInstrBuilder.h" 17 #include "llvm/CodeGen/MachineMemOperand.h" 18 #include "llvm/CodeGen/MachineRegisterInfo.h" 19 #include "llvm/CodeGen/PseudoSourceValue.h" 20 #include "llvm/CodeGen/ScoreboardHazardRecognizer.h" 21 #include "llvm/CodeGen/StackMaps.h" 22 #include "llvm/CodeGen/TargetSchedule.h" 23 #include "llvm/IR/DataLayout.h" 24 #include "llvm/MC/MCAsmInfo.h" 25 #include "llvm/MC/MCInstrItineraries.h" 26 #include "llvm/Support/CommandLine.h" 27 #include "llvm/Support/ErrorHandling.h" 28 #include "llvm/Support/raw_ostream.h" 29 #include "llvm/Target/TargetFrameLowering.h" 30 #include "llvm/Target/TargetLowering.h" 31 #include "llvm/Target/TargetMachine.h" 32 #include "llvm/Target/TargetRegisterInfo.h" 33 #include <cctype> 34 using namespace llvm; 35 36 static cl::opt<bool> DisableHazardRecognizer( 37 "disable-sched-hazard", cl::Hidden, cl::init(false), 38 cl::desc("Disable hazard detection during preRA scheduling")); 39 40 TargetInstrInfo::~TargetInstrInfo() { 41 } 42 43 const TargetRegisterClass* 44 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum, 45 const TargetRegisterInfo *TRI, 46 const MachineFunction &MF) const { 47 if (OpNum >= MCID.getNumOperands()) 48 return nullptr; 49 50 short RegClass = MCID.OpInfo[OpNum].RegClass; 51 if (MCID.OpInfo[OpNum].isLookupPtrRegClass()) 52 return TRI->getPointerRegClass(MF, RegClass); 53 54 // Instructions like INSERT_SUBREG do not have fixed register classes. 55 if (RegClass < 0) 56 return nullptr; 57 58 // Otherwise just look it up normally. 59 return TRI->getRegClass(RegClass); 60 } 61 62 /// insertNoop - Insert a noop into the instruction stream at the specified 63 /// point. 64 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB, 65 MachineBasicBlock::iterator MI) const { 66 llvm_unreachable("Target didn't implement insertNoop!"); 67 } 68 69 /// Measure the specified inline asm to determine an approximation of its 70 /// length. 71 /// Comments (which run till the next SeparatorString or newline) do not 72 /// count as an instruction. 73 /// Any other non-whitespace text is considered an instruction, with 74 /// multiple instructions separated by SeparatorString or newlines. 75 /// Variable-length instructions are not handled here; this function 76 /// may be overloaded in the target code to do that. 77 unsigned TargetInstrInfo::getInlineAsmLength(const char *Str, 78 const MCAsmInfo &MAI) const { 79 80 81 // Count the number of instructions in the asm. 82 bool atInsnStart = true; 83 unsigned Length = 0; 84 for (; *Str; ++Str) { 85 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(), 86 strlen(MAI.getSeparatorString())) == 0) 87 atInsnStart = true; 88 if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) { 89 Length += MAI.getMaxInstLength(); 90 atInsnStart = false; 91 } 92 if (atInsnStart && strncmp(Str, MAI.getCommentString(), 93 strlen(MAI.getCommentString())) == 0) 94 atInsnStart = false; 95 } 96 97 return Length; 98 } 99 100 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 101 /// after it, replacing it with an unconditional branch to NewDest. 102 void 103 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail, 104 MachineBasicBlock *NewDest) const { 105 MachineBasicBlock *MBB = Tail->getParent(); 106 107 // Remove all the old successors of MBB from the CFG. 108 while (!MBB->succ_empty()) 109 MBB->removeSuccessor(MBB->succ_begin()); 110 111 // Remove all the dead instructions from the end of MBB. 112 MBB->erase(Tail, MBB->end()); 113 114 // If MBB isn't immediately before MBB, insert a branch to it. 115 if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest)) 116 InsertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(), 117 Tail->getDebugLoc()); 118 MBB->addSuccessor(NewDest); 119 } 120 121 // commuteInstruction - The default implementation of this method just exchanges 122 // the two operands returned by findCommutedOpIndices. 123 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr *MI, 124 bool NewMI) const { 125 const MCInstrDesc &MCID = MI->getDesc(); 126 bool HasDef = MCID.getNumDefs(); 127 if (HasDef && !MI->getOperand(0).isReg()) 128 // No idea how to commute this instruction. Target should implement its own. 129 return nullptr; 130 unsigned Idx1, Idx2; 131 if (!findCommutedOpIndices(MI, Idx1, Idx2)) { 132 assert(MI->isCommutable() && "Precondition violation: MI must be commutable."); 133 return nullptr; 134 } 135 136 assert(MI->getOperand(Idx1).isReg() && MI->getOperand(Idx2).isReg() && 137 "This only knows how to commute register operands so far"); 138 unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0; 139 unsigned Reg1 = MI->getOperand(Idx1).getReg(); 140 unsigned Reg2 = MI->getOperand(Idx2).getReg(); 141 unsigned SubReg0 = HasDef ? MI->getOperand(0).getSubReg() : 0; 142 unsigned SubReg1 = MI->getOperand(Idx1).getSubReg(); 143 unsigned SubReg2 = MI->getOperand(Idx2).getSubReg(); 144 bool Reg1IsKill = MI->getOperand(Idx1).isKill(); 145 bool Reg2IsKill = MI->getOperand(Idx2).isKill(); 146 bool Reg1IsUndef = MI->getOperand(Idx1).isUndef(); 147 bool Reg2IsUndef = MI->getOperand(Idx2).isUndef(); 148 bool Reg1IsInternal = MI->getOperand(Idx1).isInternalRead(); 149 bool Reg2IsInternal = MI->getOperand(Idx2).isInternalRead(); 150 // If destination is tied to either of the commuted source register, then 151 // it must be updated. 152 if (HasDef && Reg0 == Reg1 && 153 MI->getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) { 154 Reg2IsKill = false; 155 Reg0 = Reg2; 156 SubReg0 = SubReg2; 157 } else if (HasDef && Reg0 == Reg2 && 158 MI->getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) { 159 Reg1IsKill = false; 160 Reg0 = Reg1; 161 SubReg0 = SubReg1; 162 } 163 164 if (NewMI) { 165 // Create a new instruction. 166 MachineFunction &MF = *MI->getParent()->getParent(); 167 MI = MF.CloneMachineInstr(MI); 168 } 169 170 if (HasDef) { 171 MI->getOperand(0).setReg(Reg0); 172 MI->getOperand(0).setSubReg(SubReg0); 173 } 174 MI->getOperand(Idx2).setReg(Reg1); 175 MI->getOperand(Idx1).setReg(Reg2); 176 MI->getOperand(Idx2).setSubReg(SubReg1); 177 MI->getOperand(Idx1).setSubReg(SubReg2); 178 MI->getOperand(Idx2).setIsKill(Reg1IsKill); 179 MI->getOperand(Idx1).setIsKill(Reg2IsKill); 180 MI->getOperand(Idx2).setIsUndef(Reg1IsUndef); 181 MI->getOperand(Idx1).setIsUndef(Reg2IsUndef); 182 MI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal); 183 MI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal); 184 return MI; 185 } 186 187 /// findCommutedOpIndices - If specified MI is commutable, return the two 188 /// operand indices that would swap value. Return true if the instruction 189 /// is not in a form which this routine understands. 190 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr *MI, 191 unsigned &SrcOpIdx1, 192 unsigned &SrcOpIdx2) const { 193 assert(!MI->isBundle() && 194 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles"); 195 196 const MCInstrDesc &MCID = MI->getDesc(); 197 if (!MCID.isCommutable()) 198 return false; 199 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this 200 // is not true, then the target must implement this. 201 SrcOpIdx1 = MCID.getNumDefs(); 202 SrcOpIdx2 = SrcOpIdx1 + 1; 203 if (!MI->getOperand(SrcOpIdx1).isReg() || 204 !MI->getOperand(SrcOpIdx2).isReg()) 205 // No idea. 206 return false; 207 return true; 208 } 209 210 211 bool 212 TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const { 213 if (!MI->isTerminator()) return false; 214 215 // Conditional branch is a special case. 216 if (MI->isBranch() && !MI->isBarrier()) 217 return true; 218 if (!MI->isPredicable()) 219 return true; 220 return !isPredicated(MI); 221 } 222 223 bool TargetInstrInfo::PredicateInstruction( 224 MachineInstr *MI, ArrayRef<MachineOperand> Pred) const { 225 bool MadeChange = false; 226 227 assert(!MI->isBundle() && 228 "TargetInstrInfo::PredicateInstruction() can't handle bundles"); 229 230 const MCInstrDesc &MCID = MI->getDesc(); 231 if (!MI->isPredicable()) 232 return false; 233 234 for (unsigned j = 0, i = 0, e = MI->getNumOperands(); i != e; ++i) { 235 if (MCID.OpInfo[i].isPredicate()) { 236 MachineOperand &MO = MI->getOperand(i); 237 if (MO.isReg()) { 238 MO.setReg(Pred[j].getReg()); 239 MadeChange = true; 240 } else if (MO.isImm()) { 241 MO.setImm(Pred[j].getImm()); 242 MadeChange = true; 243 } else if (MO.isMBB()) { 244 MO.setMBB(Pred[j].getMBB()); 245 MadeChange = true; 246 } 247 ++j; 248 } 249 } 250 return MadeChange; 251 } 252 253 bool TargetInstrInfo::hasLoadFromStackSlot(const MachineInstr *MI, 254 const MachineMemOperand *&MMO, 255 int &FrameIndex) const { 256 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(), 257 oe = MI->memoperands_end(); 258 o != oe; 259 ++o) { 260 if ((*o)->isLoad()) { 261 if (const FixedStackPseudoSourceValue *Value = 262 dyn_cast_or_null<FixedStackPseudoSourceValue>( 263 (*o)->getPseudoValue())) { 264 FrameIndex = Value->getFrameIndex(); 265 MMO = *o; 266 return true; 267 } 268 } 269 } 270 return false; 271 } 272 273 bool TargetInstrInfo::hasStoreToStackSlot(const MachineInstr *MI, 274 const MachineMemOperand *&MMO, 275 int &FrameIndex) const { 276 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(), 277 oe = MI->memoperands_end(); 278 o != oe; 279 ++o) { 280 if ((*o)->isStore()) { 281 if (const FixedStackPseudoSourceValue *Value = 282 dyn_cast_or_null<FixedStackPseudoSourceValue>( 283 (*o)->getPseudoValue())) { 284 FrameIndex = Value->getFrameIndex(); 285 MMO = *o; 286 return true; 287 } 288 } 289 } 290 return false; 291 } 292 293 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC, 294 unsigned SubIdx, unsigned &Size, 295 unsigned &Offset, 296 const MachineFunction &MF) const { 297 if (!SubIdx) { 298 Size = RC->getSize(); 299 Offset = 0; 300 return true; 301 } 302 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 303 unsigned BitSize = TRI->getSubRegIdxSize(SubIdx); 304 // Convert bit size to byte size to be consistent with 305 // MCRegisterClass::getSize(). 306 if (BitSize % 8) 307 return false; 308 309 int BitOffset = TRI->getSubRegIdxOffset(SubIdx); 310 if (BitOffset < 0 || BitOffset % 8) 311 return false; 312 313 Size = BitSize /= 8; 314 Offset = (unsigned)BitOffset / 8; 315 316 assert(RC->getSize() >= (Offset + Size) && "bad subregister range"); 317 318 if (!MF.getDataLayout().isLittleEndian()) { 319 Offset = RC->getSize() - (Offset + Size); 320 } 321 return true; 322 } 323 324 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB, 325 MachineBasicBlock::iterator I, 326 unsigned DestReg, 327 unsigned SubIdx, 328 const MachineInstr *Orig, 329 const TargetRegisterInfo &TRI) const { 330 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig); 331 MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI); 332 MBB.insert(I, MI); 333 } 334 335 bool 336 TargetInstrInfo::produceSameValue(const MachineInstr *MI0, 337 const MachineInstr *MI1, 338 const MachineRegisterInfo *MRI) const { 339 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs); 340 } 341 342 MachineInstr *TargetInstrInfo::duplicate(MachineInstr *Orig, 343 MachineFunction &MF) const { 344 assert(!Orig->isNotDuplicable() && 345 "Instruction cannot be duplicated"); 346 return MF.CloneMachineInstr(Orig); 347 } 348 349 // If the COPY instruction in MI can be folded to a stack operation, return 350 // the register class to use. 351 static const TargetRegisterClass *canFoldCopy(const MachineInstr *MI, 352 unsigned FoldIdx) { 353 assert(MI->isCopy() && "MI must be a COPY instruction"); 354 if (MI->getNumOperands() != 2) 355 return nullptr; 356 assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand"); 357 358 const MachineOperand &FoldOp = MI->getOperand(FoldIdx); 359 const MachineOperand &LiveOp = MI->getOperand(1-FoldIdx); 360 361 if (FoldOp.getSubReg() || LiveOp.getSubReg()) 362 return nullptr; 363 364 unsigned FoldReg = FoldOp.getReg(); 365 unsigned LiveReg = LiveOp.getReg(); 366 367 assert(TargetRegisterInfo::isVirtualRegister(FoldReg) && 368 "Cannot fold physregs"); 369 370 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo(); 371 const TargetRegisterClass *RC = MRI.getRegClass(FoldReg); 372 373 if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg())) 374 return RC->contains(LiveOp.getReg()) ? RC : nullptr; 375 376 if (RC->hasSubClassEq(MRI.getRegClass(LiveReg))) 377 return RC; 378 379 // FIXME: Allow folding when register classes are memory compatible. 380 return nullptr; 381 } 382 383 void TargetInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const { 384 llvm_unreachable("Not a MachO target"); 385 } 386 387 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr *MI, 388 ArrayRef<unsigned> Ops, int FrameIndex, 389 const TargetInstrInfo &TII) { 390 unsigned StartIdx = 0; 391 switch (MI->getOpcode()) { 392 case TargetOpcode::STACKMAP: 393 StartIdx = 2; // Skip ID, nShadowBytes. 394 break; 395 case TargetOpcode::PATCHPOINT: { 396 // For PatchPoint, the call args are not foldable. 397 PatchPointOpers opers(MI); 398 StartIdx = opers.getVarIdx(); 399 break; 400 } 401 default: 402 llvm_unreachable("unexpected stackmap opcode"); 403 } 404 405 // Return false if any operands requested for folding are not foldable (not 406 // part of the stackmap's live values). 407 for (unsigned Op : Ops) { 408 if (Op < StartIdx) 409 return nullptr; 410 } 411 412 MachineInstr *NewMI = 413 MF.CreateMachineInstr(TII.get(MI->getOpcode()), MI->getDebugLoc(), true); 414 MachineInstrBuilder MIB(MF, NewMI); 415 416 // No need to fold return, the meta data, and function arguments 417 for (unsigned i = 0; i < StartIdx; ++i) 418 MIB.addOperand(MI->getOperand(i)); 419 420 for (unsigned i = StartIdx; i < MI->getNumOperands(); ++i) { 421 MachineOperand &MO = MI->getOperand(i); 422 if (std::find(Ops.begin(), Ops.end(), i) != Ops.end()) { 423 unsigned SpillSize; 424 unsigned SpillOffset; 425 // Compute the spill slot size and offset. 426 const TargetRegisterClass *RC = 427 MF.getRegInfo().getRegClass(MO.getReg()); 428 bool Valid = 429 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF); 430 if (!Valid) 431 report_fatal_error("cannot spill patchpoint subregister operand"); 432 MIB.addImm(StackMaps::IndirectMemRefOp); 433 MIB.addImm(SpillSize); 434 MIB.addFrameIndex(FrameIndex); 435 MIB.addImm(SpillOffset); 436 } 437 else 438 MIB.addOperand(MO); 439 } 440 return NewMI; 441 } 442 443 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack 444 /// slot into the specified machine instruction for the specified operand(s). 445 /// If this is possible, a new instruction is returned with the specified 446 /// operand folded, otherwise NULL is returned. The client is responsible for 447 /// removing the old instruction and adding the new one in the instruction 448 /// stream. 449 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI, 450 ArrayRef<unsigned> Ops, 451 int FI) const { 452 unsigned Flags = 0; 453 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 454 if (MI->getOperand(Ops[i]).isDef()) 455 Flags |= MachineMemOperand::MOStore; 456 else 457 Flags |= MachineMemOperand::MOLoad; 458 459 MachineBasicBlock *MBB = MI->getParent(); 460 assert(MBB && "foldMemoryOperand needs an inserted instruction"); 461 MachineFunction &MF = *MBB->getParent(); 462 463 MachineInstr *NewMI = nullptr; 464 465 if (MI->getOpcode() == TargetOpcode::STACKMAP || 466 MI->getOpcode() == TargetOpcode::PATCHPOINT) { 467 // Fold stackmap/patchpoint. 468 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this); 469 if (NewMI) 470 MBB->insert(MI, NewMI); 471 } else { 472 // Ask the target to do the actual folding. 473 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI); 474 } 475 476 if (NewMI) { 477 NewMI->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 478 // Add a memory operand, foldMemoryOperandImpl doesn't do that. 479 assert((!(Flags & MachineMemOperand::MOStore) || 480 NewMI->mayStore()) && 481 "Folded a def to a non-store!"); 482 assert((!(Flags & MachineMemOperand::MOLoad) || 483 NewMI->mayLoad()) && 484 "Folded a use to a non-load!"); 485 const MachineFrameInfo &MFI = *MF.getFrameInfo(); 486 assert(MFI.getObjectOffset(FI) != -1); 487 MachineMemOperand *MMO = MF.getMachineMemOperand( 488 MachinePointerInfo::getFixedStack(MF, FI), Flags, MFI.getObjectSize(FI), 489 MFI.getObjectAlignment(FI)); 490 NewMI->addMemOperand(MF, MMO); 491 492 return NewMI; 493 } 494 495 // Straight COPY may fold as load/store. 496 if (!MI->isCopy() || Ops.size() != 1) 497 return nullptr; 498 499 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]); 500 if (!RC) 501 return nullptr; 502 503 const MachineOperand &MO = MI->getOperand(1-Ops[0]); 504 MachineBasicBlock::iterator Pos = MI; 505 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 506 507 if (Flags == MachineMemOperand::MOStore) 508 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI); 509 else 510 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI); 511 return --Pos; 512 } 513 514 /// foldMemoryOperand - Same as the previous version except it allows folding 515 /// of any load and store from / to any address, not just from a specific 516 /// stack slot. 517 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI, 518 ArrayRef<unsigned> Ops, 519 MachineInstr *LoadMI) const { 520 assert(LoadMI->canFoldAsLoad() && "LoadMI isn't foldable!"); 521 #ifndef NDEBUG 522 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 523 assert(MI->getOperand(Ops[i]).isUse() && "Folding load into def!"); 524 #endif 525 MachineBasicBlock &MBB = *MI->getParent(); 526 MachineFunction &MF = *MBB.getParent(); 527 528 // Ask the target to do the actual folding. 529 MachineInstr *NewMI = nullptr; 530 int FrameIndex = 0; 531 532 if ((MI->getOpcode() == TargetOpcode::STACKMAP || 533 MI->getOpcode() == TargetOpcode::PATCHPOINT) && 534 isLoadFromStackSlot(LoadMI, FrameIndex)) { 535 // Fold stackmap/patchpoint. 536 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this); 537 if (NewMI) 538 NewMI = MBB.insert(MI, NewMI); 539 } else { 540 // Ask the target to do the actual folding. 541 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI); 542 } 543 544 if (!NewMI) return nullptr; 545 546 // Copy the memoperands from the load to the folded instruction. 547 if (MI->memoperands_empty()) { 548 NewMI->setMemRefs(LoadMI->memoperands_begin(), 549 LoadMI->memoperands_end()); 550 } 551 else { 552 // Handle the rare case of folding multiple loads. 553 NewMI->setMemRefs(MI->memoperands_begin(), 554 MI->memoperands_end()); 555 for (MachineInstr::mmo_iterator I = LoadMI->memoperands_begin(), 556 E = LoadMI->memoperands_end(); I != E; ++I) { 557 NewMI->addMemOperand(MF, *I); 558 } 559 } 560 return NewMI; 561 } 562 563 bool TargetInstrInfo:: 564 isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI, 565 AliasAnalysis *AA) const { 566 const MachineFunction &MF = *MI->getParent()->getParent(); 567 const MachineRegisterInfo &MRI = MF.getRegInfo(); 568 569 // Remat clients assume operand 0 is the defined register. 570 if (!MI->getNumOperands() || !MI->getOperand(0).isReg()) 571 return false; 572 unsigned DefReg = MI->getOperand(0).getReg(); 573 574 // A sub-register definition can only be rematerialized if the instruction 575 // doesn't read the other parts of the register. Otherwise it is really a 576 // read-modify-write operation on the full virtual register which cannot be 577 // moved safely. 578 if (TargetRegisterInfo::isVirtualRegister(DefReg) && 579 MI->getOperand(0).getSubReg() && MI->readsVirtualRegister(DefReg)) 580 return false; 581 582 // A load from a fixed stack slot can be rematerialized. This may be 583 // redundant with subsequent checks, but it's target-independent, 584 // simple, and a common case. 585 int FrameIdx = 0; 586 if (isLoadFromStackSlot(MI, FrameIdx) && 587 MF.getFrameInfo()->isImmutableObjectIndex(FrameIdx)) 588 return true; 589 590 // Avoid instructions obviously unsafe for remat. 591 if (MI->isNotDuplicable() || MI->mayStore() || 592 MI->hasUnmodeledSideEffects()) 593 return false; 594 595 // Don't remat inline asm. We have no idea how expensive it is 596 // even if it's side effect free. 597 if (MI->isInlineAsm()) 598 return false; 599 600 // Avoid instructions which load from potentially varying memory. 601 if (MI->mayLoad() && !MI->isInvariantLoad(AA)) 602 return false; 603 604 // If any of the registers accessed are non-constant, conservatively assume 605 // the instruction is not rematerializable. 606 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 607 const MachineOperand &MO = MI->getOperand(i); 608 if (!MO.isReg()) continue; 609 unsigned Reg = MO.getReg(); 610 if (Reg == 0) 611 continue; 612 613 // Check for a well-behaved physical register. 614 if (TargetRegisterInfo::isPhysicalRegister(Reg)) { 615 if (MO.isUse()) { 616 // If the physreg has no defs anywhere, it's just an ambient register 617 // and we can freely move its uses. Alternatively, if it's allocatable, 618 // it could get allocated to something with a def during allocation. 619 if (!MRI.isConstantPhysReg(Reg, MF)) 620 return false; 621 } else { 622 // A physreg def. We can't remat it. 623 return false; 624 } 625 continue; 626 } 627 628 // Only allow one virtual-register def. There may be multiple defs of the 629 // same virtual register, though. 630 if (MO.isDef() && Reg != DefReg) 631 return false; 632 633 // Don't allow any virtual-register uses. Rematting an instruction with 634 // virtual register uses would length the live ranges of the uses, which 635 // is not necessarily a good idea, certainly not "trivial". 636 if (MO.isUse()) 637 return false; 638 } 639 640 // Everything checked out. 641 return true; 642 } 643 644 int TargetInstrInfo::getSPAdjust(const MachineInstr *MI) const { 645 const MachineFunction *MF = MI->getParent()->getParent(); 646 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering(); 647 bool StackGrowsDown = 648 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 649 650 unsigned FrameSetupOpcode = getCallFrameSetupOpcode(); 651 unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode(); 652 653 if (MI->getOpcode() != FrameSetupOpcode && 654 MI->getOpcode() != FrameDestroyOpcode) 655 return 0; 656 657 int SPAdj = MI->getOperand(0).getImm(); 658 659 if ((!StackGrowsDown && MI->getOpcode() == FrameSetupOpcode) || 660 (StackGrowsDown && MI->getOpcode() == FrameDestroyOpcode)) 661 SPAdj = -SPAdj; 662 663 return SPAdj; 664 } 665 666 /// isSchedulingBoundary - Test if the given instruction should be 667 /// considered a scheduling boundary. This primarily includes labels 668 /// and terminators. 669 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr *MI, 670 const MachineBasicBlock *MBB, 671 const MachineFunction &MF) const { 672 // Terminators and labels can't be scheduled around. 673 if (MI->isTerminator() || MI->isPosition()) 674 return true; 675 676 // Don't attempt to schedule around any instruction that defines 677 // a stack-oriented pointer, as it's unlikely to be profitable. This 678 // saves compile time, because it doesn't require every single 679 // stack slot reference to depend on the instruction that does the 680 // modification. 681 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering(); 682 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 683 if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI)) 684 return true; 685 686 return false; 687 } 688 689 // Provide a global flag for disabling the PreRA hazard recognizer that targets 690 // may choose to honor. 691 bool TargetInstrInfo::usePreRAHazardRecognizer() const { 692 return !DisableHazardRecognizer; 693 } 694 695 // Default implementation of CreateTargetRAHazardRecognizer. 696 ScheduleHazardRecognizer *TargetInstrInfo:: 697 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, 698 const ScheduleDAG *DAG) const { 699 // Dummy hazard recognizer allows all instructions to issue. 700 return new ScheduleHazardRecognizer(); 701 } 702 703 // Default implementation of CreateTargetMIHazardRecognizer. 704 ScheduleHazardRecognizer *TargetInstrInfo:: 705 CreateTargetMIHazardRecognizer(const InstrItineraryData *II, 706 const ScheduleDAG *DAG) const { 707 return (ScheduleHazardRecognizer *) 708 new ScoreboardHazardRecognizer(II, DAG, "misched"); 709 } 710 711 // Default implementation of CreateTargetPostRAHazardRecognizer. 712 ScheduleHazardRecognizer *TargetInstrInfo:: 713 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 714 const ScheduleDAG *DAG) const { 715 return (ScheduleHazardRecognizer *) 716 new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched"); 717 } 718 719 //===----------------------------------------------------------------------===// 720 // SelectionDAG latency interface. 721 //===----------------------------------------------------------------------===// 722 723 int 724 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 725 SDNode *DefNode, unsigned DefIdx, 726 SDNode *UseNode, unsigned UseIdx) const { 727 if (!ItinData || ItinData->isEmpty()) 728 return -1; 729 730 if (!DefNode->isMachineOpcode()) 731 return -1; 732 733 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass(); 734 if (!UseNode->isMachineOpcode()) 735 return ItinData->getOperandCycle(DefClass, DefIdx); 736 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass(); 737 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 738 } 739 740 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 741 SDNode *N) const { 742 if (!ItinData || ItinData->isEmpty()) 743 return 1; 744 745 if (!N->isMachineOpcode()) 746 return 1; 747 748 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass()); 749 } 750 751 //===----------------------------------------------------------------------===// 752 // MachineInstr latency interface. 753 //===----------------------------------------------------------------------===// 754 755 unsigned 756 TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, 757 const MachineInstr *MI) const { 758 if (!ItinData || ItinData->isEmpty()) 759 return 1; 760 761 unsigned Class = MI->getDesc().getSchedClass(); 762 int UOps = ItinData->Itineraries[Class].NumMicroOps; 763 if (UOps >= 0) 764 return UOps; 765 766 // The # of u-ops is dynamically determined. The specific target should 767 // override this function to return the right number. 768 return 1; 769 } 770 771 /// Return the default expected latency for a def based on it's opcode. 772 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel, 773 const MachineInstr *DefMI) const { 774 if (DefMI->isTransient()) 775 return 0; 776 if (DefMI->mayLoad()) 777 return SchedModel.LoadLatency; 778 if (isHighLatencyDef(DefMI->getOpcode())) 779 return SchedModel.HighLatency; 780 return 1; 781 } 782 783 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr *) const { 784 return 0; 785 } 786 787 unsigned TargetInstrInfo:: 788 getInstrLatency(const InstrItineraryData *ItinData, 789 const MachineInstr *MI, 790 unsigned *PredCost) const { 791 // Default to one cycle for no itinerary. However, an "empty" itinerary may 792 // still have a MinLatency property, which getStageLatency checks. 793 if (!ItinData) 794 return MI->mayLoad() ? 2 : 1; 795 796 return ItinData->getStageLatency(MI->getDesc().getSchedClass()); 797 } 798 799 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel, 800 const MachineInstr *DefMI, 801 unsigned DefIdx) const { 802 const InstrItineraryData *ItinData = SchedModel.getInstrItineraries(); 803 if (!ItinData || ItinData->isEmpty()) 804 return false; 805 806 unsigned DefClass = DefMI->getDesc().getSchedClass(); 807 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx); 808 return (DefCycle != -1 && DefCycle <= 1); 809 } 810 811 /// Both DefMI and UseMI must be valid. By default, call directly to the 812 /// itinerary. This may be overriden by the target. 813 int TargetInstrInfo:: 814 getOperandLatency(const InstrItineraryData *ItinData, 815 const MachineInstr *DefMI, unsigned DefIdx, 816 const MachineInstr *UseMI, unsigned UseIdx) const { 817 unsigned DefClass = DefMI->getDesc().getSchedClass(); 818 unsigned UseClass = UseMI->getDesc().getSchedClass(); 819 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 820 } 821 822 /// If we can determine the operand latency from the def only, without itinerary 823 /// lookup, do so. Otherwise return -1. 824 int TargetInstrInfo::computeDefOperandLatency( 825 const InstrItineraryData *ItinData, 826 const MachineInstr *DefMI) const { 827 828 // Let the target hook getInstrLatency handle missing itineraries. 829 if (!ItinData) 830 return getInstrLatency(ItinData, DefMI); 831 832 if(ItinData->isEmpty()) 833 return defaultDefLatency(ItinData->SchedModel, DefMI); 834 835 // ...operand lookup required 836 return -1; 837 } 838 839 /// computeOperandLatency - Compute and return the latency of the given data 840 /// dependent def and use when the operand indices are already known. UseMI may 841 /// be NULL for an unknown use. 842 /// 843 /// FindMin may be set to get the minimum vs. expected latency. Minimum 844 /// latency is used for scheduling groups, while expected latency is for 845 /// instruction cost and critical path. 846 /// 847 /// Depending on the subtarget's itinerary properties, this may or may not need 848 /// to call getOperandLatency(). For most subtargets, we don't need DefIdx or 849 /// UseIdx to compute min latency. 850 unsigned TargetInstrInfo:: 851 computeOperandLatency(const InstrItineraryData *ItinData, 852 const MachineInstr *DefMI, unsigned DefIdx, 853 const MachineInstr *UseMI, unsigned UseIdx) const { 854 855 int DefLatency = computeDefOperandLatency(ItinData, DefMI); 856 if (DefLatency >= 0) 857 return DefLatency; 858 859 assert(ItinData && !ItinData->isEmpty() && "computeDefOperandLatency fail"); 860 861 int OperLatency = 0; 862 if (UseMI) 863 OperLatency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx); 864 else { 865 unsigned DefClass = DefMI->getDesc().getSchedClass(); 866 OperLatency = ItinData->getOperandCycle(DefClass, DefIdx); 867 } 868 if (OperLatency >= 0) 869 return OperLatency; 870 871 // No operand latency was found. 872 unsigned InstrLatency = getInstrLatency(ItinData, DefMI); 873 874 // Expected latency is the max of the stage latency and itinerary props. 875 InstrLatency = std::max(InstrLatency, 876 defaultDefLatency(ItinData->SchedModel, DefMI)); 877 return InstrLatency; 878 } 879 880 bool TargetInstrInfo::getRegSequenceInputs( 881 const MachineInstr &MI, unsigned DefIdx, 882 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const { 883 assert((MI.isRegSequence() || 884 MI.isRegSequenceLike()) && "Instruction do not have the proper type"); 885 886 if (!MI.isRegSequence()) 887 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs); 888 889 // We are looking at: 890 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ... 891 assert(DefIdx == 0 && "REG_SEQUENCE only has one def"); 892 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx; 893 OpIdx += 2) { 894 const MachineOperand &MOReg = MI.getOperand(OpIdx); 895 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1); 896 assert(MOSubIdx.isImm() && 897 "One of the subindex of the reg_sequence is not an immediate"); 898 // Record Reg:SubReg, SubIdx. 899 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(), 900 (unsigned)MOSubIdx.getImm())); 901 } 902 return true; 903 } 904 905 bool TargetInstrInfo::getExtractSubregInputs( 906 const MachineInstr &MI, unsigned DefIdx, 907 RegSubRegPairAndIdx &InputReg) const { 908 assert((MI.isExtractSubreg() || 909 MI.isExtractSubregLike()) && "Instruction do not have the proper type"); 910 911 if (!MI.isExtractSubreg()) 912 return getExtractSubregLikeInputs(MI, DefIdx, InputReg); 913 914 // We are looking at: 915 // Def = EXTRACT_SUBREG v0.sub1, sub0. 916 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def"); 917 const MachineOperand &MOReg = MI.getOperand(1); 918 const MachineOperand &MOSubIdx = MI.getOperand(2); 919 assert(MOSubIdx.isImm() && 920 "The subindex of the extract_subreg is not an immediate"); 921 922 InputReg.Reg = MOReg.getReg(); 923 InputReg.SubReg = MOReg.getSubReg(); 924 InputReg.SubIdx = (unsigned)MOSubIdx.getImm(); 925 return true; 926 } 927 928 bool TargetInstrInfo::getInsertSubregInputs( 929 const MachineInstr &MI, unsigned DefIdx, 930 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const { 931 assert((MI.isInsertSubreg() || 932 MI.isInsertSubregLike()) && "Instruction do not have the proper type"); 933 934 if (!MI.isInsertSubreg()) 935 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg); 936 937 // We are looking at: 938 // Def = INSERT_SEQUENCE v0, v1, sub0. 939 assert(DefIdx == 0 && "INSERT_SUBREG only has one def"); 940 const MachineOperand &MOBaseReg = MI.getOperand(1); 941 const MachineOperand &MOInsertedReg = MI.getOperand(2); 942 const MachineOperand &MOSubIdx = MI.getOperand(3); 943 assert(MOSubIdx.isImm() && 944 "One of the subindex of the reg_sequence is not an immediate"); 945 BaseReg.Reg = MOBaseReg.getReg(); 946 BaseReg.SubReg = MOBaseReg.getSubReg(); 947 948 InsertedReg.Reg = MOInsertedReg.getReg(); 949 InsertedReg.SubReg = MOInsertedReg.getSubReg(); 950 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm(); 951 return true; 952 } 953