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/IR/DataLayout.h" 23 #include "llvm/MC/MCAsmInfo.h" 24 #include "llvm/MC/MCInstrItineraries.h" 25 #include "llvm/Support/CommandLine.h" 26 #include "llvm/Support/ErrorHandling.h" 27 #include "llvm/Support/raw_ostream.h" 28 #include "llvm/Target/TargetFrameLowering.h" 29 #include "llvm/Target/TargetLowering.h" 30 #include "llvm/Target/TargetMachine.h" 31 #include "llvm/Target/TargetRegisterInfo.h" 32 #include <cctype> 33 using namespace llvm; 34 35 static cl::opt<bool> DisableHazardRecognizer( 36 "disable-sched-hazard", cl::Hidden, cl::init(false), 37 cl::desc("Disable hazard detection during preRA scheduling")); 38 39 TargetInstrInfo::~TargetInstrInfo() { 40 } 41 42 const TargetRegisterClass* 43 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum, 44 const TargetRegisterInfo *TRI, 45 const MachineFunction &MF) const { 46 if (OpNum >= MCID.getNumOperands()) 47 return nullptr; 48 49 short RegClass = MCID.OpInfo[OpNum].RegClass; 50 if (MCID.OpInfo[OpNum].isLookupPtrRegClass()) 51 return TRI->getPointerRegClass(MF, RegClass); 52 53 // Instructions like INSERT_SUBREG do not have fixed register classes. 54 if (RegClass < 0) 55 return nullptr; 56 57 // Otherwise just look it up normally. 58 return TRI->getRegClass(RegClass); 59 } 60 61 /// insertNoop - Insert a noop into the instruction stream at the specified 62 /// point. 63 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB, 64 MachineBasicBlock::iterator MI) const { 65 llvm_unreachable("Target didn't implement insertNoop!"); 66 } 67 68 /// Measure the specified inline asm to determine an approximation of its 69 /// length. 70 /// Comments (which run till the next SeparatorString or newline) do not 71 /// count as an instruction. 72 /// Any other non-whitespace text is considered an instruction, with 73 /// multiple instructions separated by SeparatorString or newlines. 74 /// Variable-length instructions are not handled here; this function 75 /// may be overloaded in the target code to do that. 76 unsigned TargetInstrInfo::getInlineAsmLength(const char *Str, 77 const MCAsmInfo &MAI) const { 78 79 80 // Count the number of instructions in the asm. 81 bool atInsnStart = true; 82 unsigned Length = 0; 83 for (; *Str; ++Str) { 84 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(), 85 strlen(MAI.getSeparatorString())) == 0) 86 atInsnStart = true; 87 if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) { 88 Length += MAI.getMaxInstLength(); 89 atInsnStart = false; 90 } 91 if (atInsnStart && strncmp(Str, MAI.getCommentString(), 92 strlen(MAI.getCommentString())) == 0) 93 atInsnStart = false; 94 } 95 96 return Length; 97 } 98 99 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 100 /// after it, replacing it with an unconditional branch to NewDest. 101 void 102 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail, 103 MachineBasicBlock *NewDest) const { 104 MachineBasicBlock *MBB = Tail->getParent(); 105 106 // Remove all the old successors of MBB from the CFG. 107 while (!MBB->succ_empty()) 108 MBB->removeSuccessor(MBB->succ_begin()); 109 110 // Remove all the dead instructions from the end of MBB. 111 MBB->erase(Tail, MBB->end()); 112 113 // If MBB isn't immediately before MBB, insert a branch to it. 114 if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest)) 115 InsertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(), 116 Tail->getDebugLoc()); 117 MBB->addSuccessor(NewDest); 118 } 119 120 // commuteInstruction - The default implementation of this method just exchanges 121 // the two operands returned by findCommutedOpIndices. 122 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr *MI, 123 bool NewMI) const { 124 const MCInstrDesc &MCID = MI->getDesc(); 125 bool HasDef = MCID.getNumDefs(); 126 if (HasDef && !MI->getOperand(0).isReg()) 127 // No idea how to commute this instruction. Target should implement its own. 128 return nullptr; 129 unsigned Idx1, Idx2; 130 if (!findCommutedOpIndices(MI, Idx1, Idx2)) { 131 assert(MI->isCommutable() && "Precondition violation: MI must be commutable."); 132 return nullptr; 133 } 134 135 assert(MI->getOperand(Idx1).isReg() && MI->getOperand(Idx2).isReg() && 136 "This only knows how to commute register operands so far"); 137 unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0; 138 unsigned Reg1 = MI->getOperand(Idx1).getReg(); 139 unsigned Reg2 = MI->getOperand(Idx2).getReg(); 140 unsigned SubReg0 = HasDef ? MI->getOperand(0).getSubReg() : 0; 141 unsigned SubReg1 = MI->getOperand(Idx1).getSubReg(); 142 unsigned SubReg2 = MI->getOperand(Idx2).getSubReg(); 143 bool Reg1IsKill = MI->getOperand(Idx1).isKill(); 144 bool Reg2IsKill = MI->getOperand(Idx2).isKill(); 145 bool Reg1IsUndef = MI->getOperand(Idx1).isUndef(); 146 bool Reg2IsUndef = MI->getOperand(Idx2).isUndef(); 147 bool Reg1IsInternal = MI->getOperand(Idx1).isInternalRead(); 148 bool Reg2IsInternal = MI->getOperand(Idx2).isInternalRead(); 149 // If destination is tied to either of the commuted source register, then 150 // it must be updated. 151 if (HasDef && Reg0 == Reg1 && 152 MI->getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) { 153 Reg2IsKill = false; 154 Reg0 = Reg2; 155 SubReg0 = SubReg2; 156 } else if (HasDef && Reg0 == Reg2 && 157 MI->getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) { 158 Reg1IsKill = false; 159 Reg0 = Reg1; 160 SubReg0 = SubReg1; 161 } 162 163 if (NewMI) { 164 // Create a new instruction. 165 MachineFunction &MF = *MI->getParent()->getParent(); 166 MI = MF.CloneMachineInstr(MI); 167 } 168 169 if (HasDef) { 170 MI->getOperand(0).setReg(Reg0); 171 MI->getOperand(0).setSubReg(SubReg0); 172 } 173 MI->getOperand(Idx2).setReg(Reg1); 174 MI->getOperand(Idx1).setReg(Reg2); 175 MI->getOperand(Idx2).setSubReg(SubReg1); 176 MI->getOperand(Idx1).setSubReg(SubReg2); 177 MI->getOperand(Idx2).setIsKill(Reg1IsKill); 178 MI->getOperand(Idx1).setIsKill(Reg2IsKill); 179 MI->getOperand(Idx2).setIsUndef(Reg1IsUndef); 180 MI->getOperand(Idx1).setIsUndef(Reg2IsUndef); 181 MI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal); 182 MI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal); 183 return MI; 184 } 185 186 /// findCommutedOpIndices - If specified MI is commutable, return the two 187 /// operand indices that would swap value. Return true if the instruction 188 /// is not in a form which this routine understands. 189 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr *MI, 190 unsigned &SrcOpIdx1, 191 unsigned &SrcOpIdx2) const { 192 assert(!MI->isBundle() && 193 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles"); 194 195 const MCInstrDesc &MCID = MI->getDesc(); 196 if (!MCID.isCommutable()) 197 return false; 198 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this 199 // is not true, then the target must implement this. 200 SrcOpIdx1 = MCID.getNumDefs(); 201 SrcOpIdx2 = SrcOpIdx1 + 1; 202 if (!MI->getOperand(SrcOpIdx1).isReg() || 203 !MI->getOperand(SrcOpIdx2).isReg()) 204 // No idea. 205 return false; 206 return true; 207 } 208 209 210 bool 211 TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const { 212 if (!MI->isTerminator()) return false; 213 214 // Conditional branch is a special case. 215 if (MI->isBranch() && !MI->isBarrier()) 216 return true; 217 if (!MI->isPredicable()) 218 return true; 219 return !isPredicated(MI); 220 } 221 222 223 bool TargetInstrInfo::PredicateInstruction(MachineInstr *MI, 224 const SmallVectorImpl<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.getTarget().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 bool TargetInstrInfo::canFoldMemoryOperand(const MachineInstr *MI, 388 ArrayRef<unsigned> Ops) const { 389 return MI->isCopy() && Ops.size() == 1 && canFoldCopy(MI, Ops[0]); 390 } 391 392 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr *MI, 393 ArrayRef<unsigned> Ops, int FrameIndex, 394 const TargetInstrInfo &TII) { 395 unsigned StartIdx = 0; 396 switch (MI->getOpcode()) { 397 case TargetOpcode::STACKMAP: 398 StartIdx = 2; // Skip ID, nShadowBytes. 399 break; 400 case TargetOpcode::PATCHPOINT: { 401 // For PatchPoint, the call args are not foldable. 402 PatchPointOpers opers(MI); 403 StartIdx = opers.getVarIdx(); 404 break; 405 } 406 default: 407 llvm_unreachable("unexpected stackmap opcode"); 408 } 409 410 // Return false if any operands requested for folding are not foldable (not 411 // part of the stackmap's live values). 412 for (unsigned Op : Ops) { 413 if (Op < StartIdx) 414 return nullptr; 415 } 416 417 MachineInstr *NewMI = 418 MF.CreateMachineInstr(TII.get(MI->getOpcode()), MI->getDebugLoc(), true); 419 MachineInstrBuilder MIB(MF, NewMI); 420 421 // No need to fold return, the meta data, and function arguments 422 for (unsigned i = 0; i < StartIdx; ++i) 423 MIB.addOperand(MI->getOperand(i)); 424 425 for (unsigned i = StartIdx; i < MI->getNumOperands(); ++i) { 426 MachineOperand &MO = MI->getOperand(i); 427 if (std::find(Ops.begin(), Ops.end(), i) != Ops.end()) { 428 unsigned SpillSize; 429 unsigned SpillOffset; 430 // Compute the spill slot size and offset. 431 const TargetRegisterClass *RC = 432 MF.getRegInfo().getRegClass(MO.getReg()); 433 bool Valid = 434 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF); 435 if (!Valid) 436 report_fatal_error("cannot spill patchpoint subregister operand"); 437 MIB.addImm(StackMaps::IndirectMemRefOp); 438 MIB.addImm(SpillSize); 439 MIB.addFrameIndex(FrameIndex); 440 MIB.addImm(SpillOffset); 441 } 442 else 443 MIB.addOperand(MO); 444 } 445 return NewMI; 446 } 447 448 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack 449 /// slot into the specified machine instruction for the specified operand(s). 450 /// If this is possible, a new instruction is returned with the specified 451 /// operand folded, otherwise NULL is returned. The client is responsible for 452 /// removing the old instruction and adding the new one in the instruction 453 /// stream. 454 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI, 455 ArrayRef<unsigned> Ops, 456 int FI) const { 457 unsigned Flags = 0; 458 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 459 if (MI->getOperand(Ops[i]).isDef()) 460 Flags |= MachineMemOperand::MOStore; 461 else 462 Flags |= MachineMemOperand::MOLoad; 463 464 MachineBasicBlock *MBB = MI->getParent(); 465 assert(MBB && "foldMemoryOperand needs an inserted instruction"); 466 MachineFunction &MF = *MBB->getParent(); 467 468 MachineInstr *NewMI = nullptr; 469 470 if (MI->getOpcode() == TargetOpcode::STACKMAP || 471 MI->getOpcode() == TargetOpcode::PATCHPOINT) { 472 // Fold stackmap/patchpoint. 473 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this); 474 if (NewMI) 475 MBB->insert(MI, NewMI); 476 } else { 477 // Ask the target to do the actual folding. 478 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI); 479 } 480 481 if (NewMI) { 482 NewMI->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 483 // Add a memory operand, foldMemoryOperandImpl doesn't do that. 484 assert((!(Flags & MachineMemOperand::MOStore) || 485 NewMI->mayStore()) && 486 "Folded a def to a non-store!"); 487 assert((!(Flags & MachineMemOperand::MOLoad) || 488 NewMI->mayLoad()) && 489 "Folded a use to a non-load!"); 490 const MachineFrameInfo &MFI = *MF.getFrameInfo(); 491 assert(MFI.getObjectOffset(FI) != -1); 492 MachineMemOperand *MMO = 493 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI), 494 Flags, MFI.getObjectSize(FI), 495 MFI.getObjectAlignment(FI)); 496 NewMI->addMemOperand(MF, MMO); 497 498 return NewMI; 499 } 500 501 // Straight COPY may fold as load/store. 502 if (!MI->isCopy() || Ops.size() != 1) 503 return nullptr; 504 505 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]); 506 if (!RC) 507 return nullptr; 508 509 const MachineOperand &MO = MI->getOperand(1-Ops[0]); 510 MachineBasicBlock::iterator Pos = MI; 511 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 512 513 if (Flags == MachineMemOperand::MOStore) 514 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI); 515 else 516 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI); 517 return --Pos; 518 } 519 520 /// foldMemoryOperand - Same as the previous version except it allows folding 521 /// of any load and store from / to any address, not just from a specific 522 /// stack slot. 523 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI, 524 ArrayRef<unsigned> Ops, 525 MachineInstr *LoadMI) const { 526 assert(LoadMI->canFoldAsLoad() && "LoadMI isn't foldable!"); 527 #ifndef NDEBUG 528 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 529 assert(MI->getOperand(Ops[i]).isUse() && "Folding load into def!"); 530 #endif 531 MachineBasicBlock &MBB = *MI->getParent(); 532 MachineFunction &MF = *MBB.getParent(); 533 534 // Ask the target to do the actual folding. 535 MachineInstr *NewMI = nullptr; 536 int FrameIndex = 0; 537 538 if ((MI->getOpcode() == TargetOpcode::STACKMAP || 539 MI->getOpcode() == TargetOpcode::PATCHPOINT) && 540 isLoadFromStackSlot(LoadMI, FrameIndex)) { 541 // Fold stackmap/patchpoint. 542 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this); 543 if (NewMI) 544 NewMI = MBB.insert(MI, NewMI); 545 } else { 546 // Ask the target to do the actual folding. 547 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI); 548 } 549 550 if (!NewMI) return nullptr; 551 552 // Copy the memoperands from the load to the folded instruction. 553 if (MI->memoperands_empty()) { 554 NewMI->setMemRefs(LoadMI->memoperands_begin(), 555 LoadMI->memoperands_end()); 556 } 557 else { 558 // Handle the rare case of folding multiple loads. 559 NewMI->setMemRefs(MI->memoperands_begin(), 560 MI->memoperands_end()); 561 for (MachineInstr::mmo_iterator I = LoadMI->memoperands_begin(), 562 E = LoadMI->memoperands_end(); I != E; ++I) { 563 NewMI->addMemOperand(MF, *I); 564 } 565 } 566 return NewMI; 567 } 568 569 bool TargetInstrInfo:: 570 isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI, 571 AliasAnalysis *AA) const { 572 const MachineFunction &MF = *MI->getParent()->getParent(); 573 const MachineRegisterInfo &MRI = MF.getRegInfo(); 574 575 // Remat clients assume operand 0 is the defined register. 576 if (!MI->getNumOperands() || !MI->getOperand(0).isReg()) 577 return false; 578 unsigned DefReg = MI->getOperand(0).getReg(); 579 580 // A sub-register definition can only be rematerialized if the instruction 581 // doesn't read the other parts of the register. Otherwise it is really a 582 // read-modify-write operation on the full virtual register which cannot be 583 // moved safely. 584 if (TargetRegisterInfo::isVirtualRegister(DefReg) && 585 MI->getOperand(0).getSubReg() && MI->readsVirtualRegister(DefReg)) 586 return false; 587 588 // A load from a fixed stack slot can be rematerialized. This may be 589 // redundant with subsequent checks, but it's target-independent, 590 // simple, and a common case. 591 int FrameIdx = 0; 592 if (isLoadFromStackSlot(MI, FrameIdx) && 593 MF.getFrameInfo()->isImmutableObjectIndex(FrameIdx)) 594 return true; 595 596 // Avoid instructions obviously unsafe for remat. 597 if (MI->isNotDuplicable() || MI->mayStore() || 598 MI->hasUnmodeledSideEffects()) 599 return false; 600 601 // Don't remat inline asm. We have no idea how expensive it is 602 // even if it's side effect free. 603 if (MI->isInlineAsm()) 604 return false; 605 606 // Avoid instructions which load from potentially varying memory. 607 if (MI->mayLoad() && !MI->isInvariantLoad(AA)) 608 return false; 609 610 // If any of the registers accessed are non-constant, conservatively assume 611 // the instruction is not rematerializable. 612 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 613 const MachineOperand &MO = MI->getOperand(i); 614 if (!MO.isReg()) continue; 615 unsigned Reg = MO.getReg(); 616 if (Reg == 0) 617 continue; 618 619 // Check for a well-behaved physical register. 620 if (TargetRegisterInfo::isPhysicalRegister(Reg)) { 621 if (MO.isUse()) { 622 // If the physreg has no defs anywhere, it's just an ambient register 623 // and we can freely move its uses. Alternatively, if it's allocatable, 624 // it could get allocated to something with a def during allocation. 625 if (!MRI.isConstantPhysReg(Reg, MF)) 626 return false; 627 } else { 628 // A physreg def. We can't remat it. 629 return false; 630 } 631 continue; 632 } 633 634 // Only allow one virtual-register def. There may be multiple defs of the 635 // same virtual register, though. 636 if (MO.isDef() && Reg != DefReg) 637 return false; 638 639 // Don't allow any virtual-register uses. Rematting an instruction with 640 // virtual register uses would length the live ranges of the uses, which 641 // is not necessarily a good idea, certainly not "trivial". 642 if (MO.isUse()) 643 return false; 644 } 645 646 // Everything checked out. 647 return true; 648 } 649 650 int TargetInstrInfo::getSPAdjust(const MachineInstr *MI) const { 651 const MachineFunction *MF = MI->getParent()->getParent(); 652 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering(); 653 bool StackGrowsDown = 654 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 655 656 unsigned FrameSetupOpcode = getCallFrameSetupOpcode(); 657 unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode(); 658 659 if (MI->getOpcode() != FrameSetupOpcode && 660 MI->getOpcode() != FrameDestroyOpcode) 661 return 0; 662 663 int SPAdj = MI->getOperand(0).getImm(); 664 665 if ((!StackGrowsDown && MI->getOpcode() == FrameSetupOpcode) || 666 (StackGrowsDown && MI->getOpcode() == FrameDestroyOpcode)) 667 SPAdj = -SPAdj; 668 669 return SPAdj; 670 } 671 672 /// isSchedulingBoundary - Test if the given instruction should be 673 /// considered a scheduling boundary. This primarily includes labels 674 /// and terminators. 675 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr *MI, 676 const MachineBasicBlock *MBB, 677 const MachineFunction &MF) const { 678 // Terminators and labels can't be scheduled around. 679 if (MI->isTerminator() || MI->isPosition()) 680 return true; 681 682 // Don't attempt to schedule around any instruction that defines 683 // a stack-oriented pointer, as it's unlikely to be profitable. This 684 // saves compile time, because it doesn't require every single 685 // stack slot reference to depend on the instruction that does the 686 // modification. 687 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering(); 688 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 689 if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI)) 690 return true; 691 692 return false; 693 } 694 695 // Provide a global flag for disabling the PreRA hazard recognizer that targets 696 // may choose to honor. 697 bool TargetInstrInfo::usePreRAHazardRecognizer() const { 698 return !DisableHazardRecognizer; 699 } 700 701 // Default implementation of CreateTargetRAHazardRecognizer. 702 ScheduleHazardRecognizer *TargetInstrInfo:: 703 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, 704 const ScheduleDAG *DAG) const { 705 // Dummy hazard recognizer allows all instructions to issue. 706 return new ScheduleHazardRecognizer(); 707 } 708 709 // Default implementation of CreateTargetMIHazardRecognizer. 710 ScheduleHazardRecognizer *TargetInstrInfo:: 711 CreateTargetMIHazardRecognizer(const InstrItineraryData *II, 712 const ScheduleDAG *DAG) const { 713 return (ScheduleHazardRecognizer *) 714 new ScoreboardHazardRecognizer(II, DAG, "misched"); 715 } 716 717 // Default implementation of CreateTargetPostRAHazardRecognizer. 718 ScheduleHazardRecognizer *TargetInstrInfo:: 719 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 720 const ScheduleDAG *DAG) const { 721 return (ScheduleHazardRecognizer *) 722 new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched"); 723 } 724 725 //===----------------------------------------------------------------------===// 726 // SelectionDAG latency interface. 727 //===----------------------------------------------------------------------===// 728 729 int 730 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 731 SDNode *DefNode, unsigned DefIdx, 732 SDNode *UseNode, unsigned UseIdx) const { 733 if (!ItinData || ItinData->isEmpty()) 734 return -1; 735 736 if (!DefNode->isMachineOpcode()) 737 return -1; 738 739 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass(); 740 if (!UseNode->isMachineOpcode()) 741 return ItinData->getOperandCycle(DefClass, DefIdx); 742 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass(); 743 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 744 } 745 746 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 747 SDNode *N) const { 748 if (!ItinData || ItinData->isEmpty()) 749 return 1; 750 751 if (!N->isMachineOpcode()) 752 return 1; 753 754 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass()); 755 } 756 757 //===----------------------------------------------------------------------===// 758 // MachineInstr latency interface. 759 //===----------------------------------------------------------------------===// 760 761 unsigned 762 TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, 763 const MachineInstr *MI) const { 764 if (!ItinData || ItinData->isEmpty()) 765 return 1; 766 767 unsigned Class = MI->getDesc().getSchedClass(); 768 int UOps = ItinData->Itineraries[Class].NumMicroOps; 769 if (UOps >= 0) 770 return UOps; 771 772 // The # of u-ops is dynamically determined. The specific target should 773 // override this function to return the right number. 774 return 1; 775 } 776 777 /// Return the default expected latency for a def based on it's opcode. 778 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel, 779 const MachineInstr *DefMI) const { 780 if (DefMI->isTransient()) 781 return 0; 782 if (DefMI->mayLoad()) 783 return SchedModel.LoadLatency; 784 if (isHighLatencyDef(DefMI->getOpcode())) 785 return SchedModel.HighLatency; 786 return 1; 787 } 788 789 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr *) const { 790 return 0; 791 } 792 793 unsigned TargetInstrInfo:: 794 getInstrLatency(const InstrItineraryData *ItinData, 795 const MachineInstr *MI, 796 unsigned *PredCost) const { 797 // Default to one cycle for no itinerary. However, an "empty" itinerary may 798 // still have a MinLatency property, which getStageLatency checks. 799 if (!ItinData) 800 return MI->mayLoad() ? 2 : 1; 801 802 return ItinData->getStageLatency(MI->getDesc().getSchedClass()); 803 } 804 805 bool TargetInstrInfo::hasLowDefLatency(const InstrItineraryData *ItinData, 806 const MachineInstr *DefMI, 807 unsigned DefIdx) const { 808 if (!ItinData || ItinData->isEmpty()) 809 return false; 810 811 unsigned DefClass = DefMI->getDesc().getSchedClass(); 812 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx); 813 return (DefCycle != -1 && DefCycle <= 1); 814 } 815 816 /// Both DefMI and UseMI must be valid. By default, call directly to the 817 /// itinerary. This may be overriden by the target. 818 int TargetInstrInfo:: 819 getOperandLatency(const InstrItineraryData *ItinData, 820 const MachineInstr *DefMI, unsigned DefIdx, 821 const MachineInstr *UseMI, unsigned UseIdx) const { 822 unsigned DefClass = DefMI->getDesc().getSchedClass(); 823 unsigned UseClass = UseMI->getDesc().getSchedClass(); 824 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 825 } 826 827 /// If we can determine the operand latency from the def only, without itinerary 828 /// lookup, do so. Otherwise return -1. 829 int TargetInstrInfo::computeDefOperandLatency( 830 const InstrItineraryData *ItinData, 831 const MachineInstr *DefMI) const { 832 833 // Let the target hook getInstrLatency handle missing itineraries. 834 if (!ItinData) 835 return getInstrLatency(ItinData, DefMI); 836 837 if(ItinData->isEmpty()) 838 return defaultDefLatency(ItinData->SchedModel, DefMI); 839 840 // ...operand lookup required 841 return -1; 842 } 843 844 /// computeOperandLatency - Compute and return the latency of the given data 845 /// dependent def and use when the operand indices are already known. UseMI may 846 /// be NULL for an unknown use. 847 /// 848 /// FindMin may be set to get the minimum vs. expected latency. Minimum 849 /// latency is used for scheduling groups, while expected latency is for 850 /// instruction cost and critical path. 851 /// 852 /// Depending on the subtarget's itinerary properties, this may or may not need 853 /// to call getOperandLatency(). For most subtargets, we don't need DefIdx or 854 /// UseIdx to compute min latency. 855 unsigned TargetInstrInfo:: 856 computeOperandLatency(const InstrItineraryData *ItinData, 857 const MachineInstr *DefMI, unsigned DefIdx, 858 const MachineInstr *UseMI, unsigned UseIdx) const { 859 860 int DefLatency = computeDefOperandLatency(ItinData, DefMI); 861 if (DefLatency >= 0) 862 return DefLatency; 863 864 assert(ItinData && !ItinData->isEmpty() && "computeDefOperandLatency fail"); 865 866 int OperLatency = 0; 867 if (UseMI) 868 OperLatency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx); 869 else { 870 unsigned DefClass = DefMI->getDesc().getSchedClass(); 871 OperLatency = ItinData->getOperandCycle(DefClass, DefIdx); 872 } 873 if (OperLatency >= 0) 874 return OperLatency; 875 876 // No operand latency was found. 877 unsigned InstrLatency = getInstrLatency(ItinData, DefMI); 878 879 // Expected latency is the max of the stage latency and itinerary props. 880 InstrLatency = std::max(InstrLatency, 881 defaultDefLatency(ItinData->SchedModel, DefMI)); 882 return InstrLatency; 883 } 884 885 bool TargetInstrInfo::getRegSequenceInputs( 886 const MachineInstr &MI, unsigned DefIdx, 887 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const { 888 assert((MI.isRegSequence() || 889 MI.isRegSequenceLike()) && "Instruction do not have the proper type"); 890 891 if (!MI.isRegSequence()) 892 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs); 893 894 // We are looking at: 895 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ... 896 assert(DefIdx == 0 && "REG_SEQUENCE only has one def"); 897 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx; 898 OpIdx += 2) { 899 const MachineOperand &MOReg = MI.getOperand(OpIdx); 900 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1); 901 assert(MOSubIdx.isImm() && 902 "One of the subindex of the reg_sequence is not an immediate"); 903 // Record Reg:SubReg, SubIdx. 904 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(), 905 (unsigned)MOSubIdx.getImm())); 906 } 907 return true; 908 } 909 910 bool TargetInstrInfo::getExtractSubregInputs( 911 const MachineInstr &MI, unsigned DefIdx, 912 RegSubRegPairAndIdx &InputReg) const { 913 assert((MI.isExtractSubreg() || 914 MI.isExtractSubregLike()) && "Instruction do not have the proper type"); 915 916 if (!MI.isExtractSubreg()) 917 return getExtractSubregLikeInputs(MI, DefIdx, InputReg); 918 919 // We are looking at: 920 // Def = EXTRACT_SUBREG v0.sub1, sub0. 921 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def"); 922 const MachineOperand &MOReg = MI.getOperand(1); 923 const MachineOperand &MOSubIdx = MI.getOperand(2); 924 assert(MOSubIdx.isImm() && 925 "The subindex of the extract_subreg is not an immediate"); 926 927 InputReg.Reg = MOReg.getReg(); 928 InputReg.SubReg = MOReg.getSubReg(); 929 InputReg.SubIdx = (unsigned)MOSubIdx.getImm(); 930 return true; 931 } 932 933 bool TargetInstrInfo::getInsertSubregInputs( 934 const MachineInstr &MI, unsigned DefIdx, 935 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const { 936 assert((MI.isInsertSubreg() || 937 MI.isInsertSubregLike()) && "Instruction do not have the proper type"); 938 939 if (!MI.isInsertSubreg()) 940 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg); 941 942 // We are looking at: 943 // Def = INSERT_SEQUENCE v0, v1, sub0. 944 assert(DefIdx == 0 && "INSERT_SUBREG only has one def"); 945 const MachineOperand &MOBaseReg = MI.getOperand(1); 946 const MachineOperand &MOInsertedReg = MI.getOperand(2); 947 const MachineOperand &MOSubIdx = MI.getOperand(3); 948 assert(MOSubIdx.isImm() && 949 "One of the subindex of the reg_sequence is not an immediate"); 950 BaseReg.Reg = MOBaseReg.getReg(); 951 BaseReg.SubReg = MOBaseReg.getSubReg(); 952 953 InsertedReg.Reg = MOInsertedReg.getReg(); 954 InsertedReg.SubReg = MOInsertedReg.getSubReg(); 955 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm(); 956 return true; 957 } 958