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 35 using namespace llvm; 36 37 static cl::opt<bool> DisableHazardRecognizer( 38 "disable-sched-hazard", cl::Hidden, cl::init(false), 39 cl::desc("Disable hazard detection during preRA scheduling")); 40 41 TargetInstrInfo::~TargetInstrInfo() { 42 } 43 44 const TargetRegisterClass* 45 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum, 46 const TargetRegisterInfo *TRI, 47 const MachineFunction &MF) const { 48 if (OpNum >= MCID.getNumOperands()) 49 return nullptr; 50 51 short RegClass = MCID.OpInfo[OpNum].RegClass; 52 if (MCID.OpInfo[OpNum].isLookupPtrRegClass()) 53 return TRI->getPointerRegClass(MF, RegClass); 54 55 // Instructions like INSERT_SUBREG do not have fixed register classes. 56 if (RegClass < 0) 57 return nullptr; 58 59 // Otherwise just look it up normally. 60 return TRI->getRegClass(RegClass); 61 } 62 63 /// insertNoop - Insert a noop into the instruction stream at the specified 64 /// point. 65 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB, 66 MachineBasicBlock::iterator MI) const { 67 llvm_unreachable("Target didn't implement insertNoop!"); 68 } 69 70 /// Measure the specified inline asm to determine an approximation of its 71 /// length. 72 /// Comments (which run till the next SeparatorString or newline) do not 73 /// count as an instruction. 74 /// Any other non-whitespace text is considered an instruction, with 75 /// multiple instructions separated by SeparatorString or newlines. 76 /// Variable-length instructions are not handled here; this function 77 /// may be overloaded in the target code to do that. 78 unsigned TargetInstrInfo::getInlineAsmLength(const char *Str, 79 const MCAsmInfo &MAI) const { 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 // Save off the debug loc before erasing the instruction. 111 DebugLoc DL = Tail->getDebugLoc(); 112 113 // Remove all the dead instructions from the end of MBB. 114 MBB->erase(Tail, MBB->end()); 115 116 // If MBB isn't immediately before MBB, insert a branch to it. 117 if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest)) 118 InsertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(), DL); 119 MBB->addSuccessor(NewDest); 120 } 121 122 MachineInstr *TargetInstrInfo::commuteInstructionImpl(MachineInstr *MI, 123 bool NewMI, 124 unsigned Idx1, 125 unsigned Idx2) const { 126 const MCInstrDesc &MCID = MI->getDesc(); 127 bool HasDef = MCID.getNumDefs(); 128 if (HasDef && !MI->getOperand(0).isReg()) 129 // No idea how to commute this instruction. Target should implement its own. 130 return nullptr; 131 132 unsigned CommutableOpIdx1 = Idx1; (void)CommutableOpIdx1; 133 unsigned CommutableOpIdx2 = Idx2; (void)CommutableOpIdx2; 134 assert(findCommutedOpIndices(MI, CommutableOpIdx1, CommutableOpIdx2) && 135 CommutableOpIdx1 == Idx1 && CommutableOpIdx2 == Idx2 && 136 "TargetInstrInfo::CommuteInstructionImpl(): not commutable operands."); 137 assert(MI->getOperand(Idx1).isReg() && MI->getOperand(Idx2).isReg() && 138 "This only knows how to commute register operands so far"); 139 140 unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0; 141 unsigned Reg1 = MI->getOperand(Idx1).getReg(); 142 unsigned Reg2 = MI->getOperand(Idx2).getReg(); 143 unsigned SubReg0 = HasDef ? MI->getOperand(0).getSubReg() : 0; 144 unsigned SubReg1 = MI->getOperand(Idx1).getSubReg(); 145 unsigned SubReg2 = MI->getOperand(Idx2).getSubReg(); 146 bool Reg1IsKill = MI->getOperand(Idx1).isKill(); 147 bool Reg2IsKill = MI->getOperand(Idx2).isKill(); 148 bool Reg1IsUndef = MI->getOperand(Idx1).isUndef(); 149 bool Reg2IsUndef = MI->getOperand(Idx2).isUndef(); 150 bool Reg1IsInternal = MI->getOperand(Idx1).isInternalRead(); 151 bool Reg2IsInternal = MI->getOperand(Idx2).isInternalRead(); 152 // If destination is tied to either of the commuted source register, then 153 // it must be updated. 154 if (HasDef && Reg0 == Reg1 && 155 MI->getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) { 156 Reg2IsKill = false; 157 Reg0 = Reg2; 158 SubReg0 = SubReg2; 159 } else if (HasDef && Reg0 == Reg2 && 160 MI->getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) { 161 Reg1IsKill = false; 162 Reg0 = Reg1; 163 SubReg0 = SubReg1; 164 } 165 166 if (NewMI) { 167 // Create a new instruction. 168 MachineFunction &MF = *MI->getParent()->getParent(); 169 MI = MF.CloneMachineInstr(MI); 170 } 171 172 if (HasDef) { 173 MI->getOperand(0).setReg(Reg0); 174 MI->getOperand(0).setSubReg(SubReg0); 175 } 176 MI->getOperand(Idx2).setReg(Reg1); 177 MI->getOperand(Idx1).setReg(Reg2); 178 MI->getOperand(Idx2).setSubReg(SubReg1); 179 MI->getOperand(Idx1).setSubReg(SubReg2); 180 MI->getOperand(Idx2).setIsKill(Reg1IsKill); 181 MI->getOperand(Idx1).setIsKill(Reg2IsKill); 182 MI->getOperand(Idx2).setIsUndef(Reg1IsUndef); 183 MI->getOperand(Idx1).setIsUndef(Reg2IsUndef); 184 MI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal); 185 MI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal); 186 return MI; 187 } 188 189 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr *MI, 190 bool NewMI, 191 unsigned OpIdx1, 192 unsigned OpIdx2) const { 193 // If OpIdx1 or OpIdx2 is not specified, then this method is free to choose 194 // any commutable operand, which is done in findCommutedOpIndices() method 195 // called below. 196 if ((OpIdx1 == CommuteAnyOperandIndex || OpIdx2 == CommuteAnyOperandIndex) && 197 !findCommutedOpIndices(MI, OpIdx1, OpIdx2)) { 198 assert(MI->isCommutable() && 199 "Precondition violation: MI must be commutable."); 200 return nullptr; 201 } 202 return commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 203 } 204 205 bool TargetInstrInfo::fixCommutedOpIndices(unsigned &ResultIdx1, 206 unsigned &ResultIdx2, 207 unsigned CommutableOpIdx1, 208 unsigned CommutableOpIdx2) { 209 if (ResultIdx1 == CommuteAnyOperandIndex && 210 ResultIdx2 == CommuteAnyOperandIndex) { 211 ResultIdx1 = CommutableOpIdx1; 212 ResultIdx2 = CommutableOpIdx2; 213 } else if (ResultIdx1 == CommuteAnyOperandIndex) { 214 if (ResultIdx2 == CommutableOpIdx1) 215 ResultIdx1 = CommutableOpIdx2; 216 else if (ResultIdx2 == CommutableOpIdx2) 217 ResultIdx1 = CommutableOpIdx1; 218 else 219 return false; 220 } else if (ResultIdx2 == CommuteAnyOperandIndex) { 221 if (ResultIdx1 == CommutableOpIdx1) 222 ResultIdx2 = CommutableOpIdx2; 223 else if (ResultIdx1 == CommutableOpIdx2) 224 ResultIdx2 = CommutableOpIdx1; 225 else 226 return false; 227 } else 228 // Check that the result operand indices match the given commutable 229 // operand indices. 230 return (ResultIdx1 == CommutableOpIdx1 && ResultIdx2 == CommutableOpIdx2) || 231 (ResultIdx1 == CommutableOpIdx2 && ResultIdx2 == CommutableOpIdx1); 232 233 return true; 234 } 235 236 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr *MI, 237 unsigned &SrcOpIdx1, 238 unsigned &SrcOpIdx2) const { 239 assert(!MI->isBundle() && 240 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles"); 241 242 const MCInstrDesc &MCID = MI->getDesc(); 243 if (!MCID.isCommutable()) 244 return false; 245 246 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this 247 // is not true, then the target must implement this. 248 unsigned CommutableOpIdx1 = MCID.getNumDefs(); 249 unsigned CommutableOpIdx2 = CommutableOpIdx1 + 1; 250 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 251 CommutableOpIdx1, CommutableOpIdx2)) 252 return false; 253 254 if (!MI->getOperand(SrcOpIdx1).isReg() || 255 !MI->getOperand(SrcOpIdx2).isReg()) 256 // No idea. 257 return false; 258 return true; 259 } 260 261 bool TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const { 262 if (!MI.isTerminator()) return false; 263 264 // Conditional branch is a special case. 265 if (MI.isBranch() && !MI.isBarrier()) 266 return true; 267 if (!MI.isPredicable()) 268 return true; 269 return !isPredicated(MI); 270 } 271 272 bool TargetInstrInfo::PredicateInstruction( 273 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { 274 bool MadeChange = false; 275 276 assert(!MI.isBundle() && 277 "TargetInstrInfo::PredicateInstruction() can't handle bundles"); 278 279 const MCInstrDesc &MCID = MI.getDesc(); 280 if (!MI.isPredicable()) 281 return false; 282 283 for (unsigned j = 0, i = 0, e = MI.getNumOperands(); i != e; ++i) { 284 if (MCID.OpInfo[i].isPredicate()) { 285 MachineOperand &MO = MI.getOperand(i); 286 if (MO.isReg()) { 287 MO.setReg(Pred[j].getReg()); 288 MadeChange = true; 289 } else if (MO.isImm()) { 290 MO.setImm(Pred[j].getImm()); 291 MadeChange = true; 292 } else if (MO.isMBB()) { 293 MO.setMBB(Pred[j].getMBB()); 294 MadeChange = true; 295 } 296 ++j; 297 } 298 } 299 return MadeChange; 300 } 301 302 bool TargetInstrInfo::hasLoadFromStackSlot(const MachineInstr *MI, 303 const MachineMemOperand *&MMO, 304 int &FrameIndex) const { 305 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(), 306 oe = MI->memoperands_end(); 307 o != oe; 308 ++o) { 309 if ((*o)->isLoad()) { 310 if (const FixedStackPseudoSourceValue *Value = 311 dyn_cast_or_null<FixedStackPseudoSourceValue>( 312 (*o)->getPseudoValue())) { 313 FrameIndex = Value->getFrameIndex(); 314 MMO = *o; 315 return true; 316 } 317 } 318 } 319 return false; 320 } 321 322 bool TargetInstrInfo::hasStoreToStackSlot(const MachineInstr *MI, 323 const MachineMemOperand *&MMO, 324 int &FrameIndex) const { 325 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(), 326 oe = MI->memoperands_end(); 327 o != oe; 328 ++o) { 329 if ((*o)->isStore()) { 330 if (const FixedStackPseudoSourceValue *Value = 331 dyn_cast_or_null<FixedStackPseudoSourceValue>( 332 (*o)->getPseudoValue())) { 333 FrameIndex = Value->getFrameIndex(); 334 MMO = *o; 335 return true; 336 } 337 } 338 } 339 return false; 340 } 341 342 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC, 343 unsigned SubIdx, unsigned &Size, 344 unsigned &Offset, 345 const MachineFunction &MF) const { 346 if (!SubIdx) { 347 Size = RC->getSize(); 348 Offset = 0; 349 return true; 350 } 351 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 352 unsigned BitSize = TRI->getSubRegIdxSize(SubIdx); 353 // Convert bit size to byte size to be consistent with 354 // MCRegisterClass::getSize(). 355 if (BitSize % 8) 356 return false; 357 358 int BitOffset = TRI->getSubRegIdxOffset(SubIdx); 359 if (BitOffset < 0 || BitOffset % 8) 360 return false; 361 362 Size = BitSize /= 8; 363 Offset = (unsigned)BitOffset / 8; 364 365 assert(RC->getSize() >= (Offset + Size) && "bad subregister range"); 366 367 if (!MF.getDataLayout().isLittleEndian()) { 368 Offset = RC->getSize() - (Offset + Size); 369 } 370 return true; 371 } 372 373 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB, 374 MachineBasicBlock::iterator I, 375 unsigned DestReg, 376 unsigned SubIdx, 377 const MachineInstr *Orig, 378 const TargetRegisterInfo &TRI) const { 379 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig); 380 MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI); 381 MBB.insert(I, MI); 382 } 383 384 bool 385 TargetInstrInfo::produceSameValue(const MachineInstr *MI0, 386 const MachineInstr *MI1, 387 const MachineRegisterInfo *MRI) const { 388 return MI0->isIdenticalTo(*MI1, MachineInstr::IgnoreVRegDefs); 389 } 390 391 MachineInstr *TargetInstrInfo::duplicate(MachineInstr *Orig, 392 MachineFunction &MF) const { 393 assert(!Orig->isNotDuplicable() && 394 "Instruction cannot be duplicated"); 395 return MF.CloneMachineInstr(Orig); 396 } 397 398 // If the COPY instruction in MI can be folded to a stack operation, return 399 // the register class to use. 400 static const TargetRegisterClass *canFoldCopy(const MachineInstr *MI, 401 unsigned FoldIdx) { 402 assert(MI->isCopy() && "MI must be a COPY instruction"); 403 if (MI->getNumOperands() != 2) 404 return nullptr; 405 assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand"); 406 407 const MachineOperand &FoldOp = MI->getOperand(FoldIdx); 408 const MachineOperand &LiveOp = MI->getOperand(1-FoldIdx); 409 410 if (FoldOp.getSubReg() || LiveOp.getSubReg()) 411 return nullptr; 412 413 unsigned FoldReg = FoldOp.getReg(); 414 unsigned LiveReg = LiveOp.getReg(); 415 416 assert(TargetRegisterInfo::isVirtualRegister(FoldReg) && 417 "Cannot fold physregs"); 418 419 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo(); 420 const TargetRegisterClass *RC = MRI.getRegClass(FoldReg); 421 422 if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg())) 423 return RC->contains(LiveOp.getReg()) ? RC : nullptr; 424 425 if (RC->hasSubClassEq(MRI.getRegClass(LiveReg))) 426 return RC; 427 428 // FIXME: Allow folding when register classes are memory compatible. 429 return nullptr; 430 } 431 432 void TargetInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const { 433 llvm_unreachable("Not a MachO target"); 434 } 435 436 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr *MI, 437 ArrayRef<unsigned> Ops, int FrameIndex, 438 const TargetInstrInfo &TII) { 439 unsigned StartIdx = 0; 440 switch (MI->getOpcode()) { 441 case TargetOpcode::STACKMAP: 442 StartIdx = 2; // Skip ID, nShadowBytes. 443 break; 444 case TargetOpcode::PATCHPOINT: { 445 // For PatchPoint, the call args are not foldable. 446 PatchPointOpers opers(MI); 447 StartIdx = opers.getVarIdx(); 448 break; 449 } 450 default: 451 llvm_unreachable("unexpected stackmap opcode"); 452 } 453 454 // Return false if any operands requested for folding are not foldable (not 455 // part of the stackmap's live values). 456 for (unsigned Op : Ops) { 457 if (Op < StartIdx) 458 return nullptr; 459 } 460 461 MachineInstr *NewMI = 462 MF.CreateMachineInstr(TII.get(MI->getOpcode()), MI->getDebugLoc(), true); 463 MachineInstrBuilder MIB(MF, NewMI); 464 465 // No need to fold return, the meta data, and function arguments 466 for (unsigned i = 0; i < StartIdx; ++i) 467 MIB.addOperand(MI->getOperand(i)); 468 469 for (unsigned i = StartIdx; i < MI->getNumOperands(); ++i) { 470 MachineOperand &MO = MI->getOperand(i); 471 if (std::find(Ops.begin(), Ops.end(), i) != Ops.end()) { 472 unsigned SpillSize; 473 unsigned SpillOffset; 474 // Compute the spill slot size and offset. 475 const TargetRegisterClass *RC = 476 MF.getRegInfo().getRegClass(MO.getReg()); 477 bool Valid = 478 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF); 479 if (!Valid) 480 report_fatal_error("cannot spill patchpoint subregister operand"); 481 MIB.addImm(StackMaps::IndirectMemRefOp); 482 MIB.addImm(SpillSize); 483 MIB.addFrameIndex(FrameIndex); 484 MIB.addImm(SpillOffset); 485 } 486 else 487 MIB.addOperand(MO); 488 } 489 return NewMI; 490 } 491 492 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack 493 /// slot into the specified machine instruction for the specified operand(s). 494 /// If this is possible, a new instruction is returned with the specified 495 /// operand folded, otherwise NULL is returned. The client is responsible for 496 /// removing the old instruction and adding the new one in the instruction 497 /// stream. 498 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI, 499 ArrayRef<unsigned> Ops, 500 int FI) const { 501 unsigned Flags = 0; 502 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 503 if (MI->getOperand(Ops[i]).isDef()) 504 Flags |= MachineMemOperand::MOStore; 505 else 506 Flags |= MachineMemOperand::MOLoad; 507 508 MachineBasicBlock *MBB = MI->getParent(); 509 assert(MBB && "foldMemoryOperand needs an inserted instruction"); 510 MachineFunction &MF = *MBB->getParent(); 511 512 MachineInstr *NewMI = nullptr; 513 514 if (MI->getOpcode() == TargetOpcode::STACKMAP || 515 MI->getOpcode() == TargetOpcode::PATCHPOINT) { 516 // Fold stackmap/patchpoint. 517 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this); 518 if (NewMI) 519 MBB->insert(MI, NewMI); 520 } else { 521 // Ask the target to do the actual folding. 522 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI); 523 } 524 525 if (NewMI) { 526 NewMI->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 527 // Add a memory operand, foldMemoryOperandImpl doesn't do that. 528 assert((!(Flags & MachineMemOperand::MOStore) || 529 NewMI->mayStore()) && 530 "Folded a def to a non-store!"); 531 assert((!(Flags & MachineMemOperand::MOLoad) || 532 NewMI->mayLoad()) && 533 "Folded a use to a non-load!"); 534 const MachineFrameInfo &MFI = *MF.getFrameInfo(); 535 assert(MFI.getObjectOffset(FI) != -1); 536 MachineMemOperand *MMO = MF.getMachineMemOperand( 537 MachinePointerInfo::getFixedStack(MF, FI), Flags, MFI.getObjectSize(FI), 538 MFI.getObjectAlignment(FI)); 539 NewMI->addMemOperand(MF, MMO); 540 541 return NewMI; 542 } 543 544 // Straight COPY may fold as load/store. 545 if (!MI->isCopy() || Ops.size() != 1) 546 return nullptr; 547 548 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]); 549 if (!RC) 550 return nullptr; 551 552 const MachineOperand &MO = MI->getOperand(1-Ops[0]); 553 MachineBasicBlock::iterator Pos = MI; 554 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 555 556 if (Flags == MachineMemOperand::MOStore) 557 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI); 558 else 559 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI); 560 return --Pos; 561 } 562 563 bool TargetInstrInfo::hasReassociableOperands( 564 const MachineInstr &Inst, const MachineBasicBlock *MBB) const { 565 const MachineOperand &Op1 = Inst.getOperand(1); 566 const MachineOperand &Op2 = Inst.getOperand(2); 567 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 568 569 // We need virtual register definitions for the operands that we will 570 // reassociate. 571 MachineInstr *MI1 = nullptr; 572 MachineInstr *MI2 = nullptr; 573 if (Op1.isReg() && TargetRegisterInfo::isVirtualRegister(Op1.getReg())) 574 MI1 = MRI.getUniqueVRegDef(Op1.getReg()); 575 if (Op2.isReg() && TargetRegisterInfo::isVirtualRegister(Op2.getReg())) 576 MI2 = MRI.getUniqueVRegDef(Op2.getReg()); 577 578 // And they need to be in the trace (otherwise, they won't have a depth). 579 return MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB; 580 } 581 582 bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst, 583 bool &Commuted) const { 584 const MachineBasicBlock *MBB = Inst.getParent(); 585 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 586 MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg()); 587 MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg()); 588 unsigned AssocOpcode = Inst.getOpcode(); 589 590 // If only one operand has the same opcode and it's the second source operand, 591 // the operands must be commuted. 592 Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode; 593 if (Commuted) 594 std::swap(MI1, MI2); 595 596 // 1. The previous instruction must be the same type as Inst. 597 // 2. The previous instruction must have virtual register definitions for its 598 // operands in the same basic block as Inst. 599 // 3. The previous instruction's result must only be used by Inst. 600 return MI1->getOpcode() == AssocOpcode && 601 hasReassociableOperands(*MI1, MBB) && 602 MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg()); 603 } 604 605 // 1. The operation must be associative and commutative. 606 // 2. The instruction must have virtual register definitions for its 607 // operands in the same basic block. 608 // 3. The instruction must have a reassociable sibling. 609 bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst, 610 bool &Commuted) const { 611 return isAssociativeAndCommutative(Inst) && 612 hasReassociableOperands(Inst, Inst.getParent()) && 613 hasReassociableSibling(Inst, Commuted); 614 } 615 616 // The concept of the reassociation pass is that these operations can benefit 617 // from this kind of transformation: 618 // 619 // A = ? op ? 620 // B = A op X (Prev) 621 // C = B op Y (Root) 622 // --> 623 // A = ? op ? 624 // B = X op Y 625 // C = A op B 626 // 627 // breaking the dependency between A and B, allowing them to be executed in 628 // parallel (or back-to-back in a pipeline) instead of depending on each other. 629 630 // FIXME: This has the potential to be expensive (compile time) while not 631 // improving the code at all. Some ways to limit the overhead: 632 // 1. Track successful transforms; bail out if hit rate gets too low. 633 // 2. Only enable at -O3 or some other non-default optimization level. 634 // 3. Pre-screen pattern candidates here: if an operand of the previous 635 // instruction is known to not increase the critical path, then don't match 636 // that pattern. 637 bool TargetInstrInfo::getMachineCombinerPatterns( 638 MachineInstr &Root, 639 SmallVectorImpl<MachineCombinerPattern> &Patterns) const { 640 bool Commute; 641 if (isReassociationCandidate(Root, Commute)) { 642 // We found a sequence of instructions that may be suitable for a 643 // reassociation of operands to increase ILP. Specify each commutation 644 // possibility for the Prev instruction in the sequence and let the 645 // machine combiner decide if changing the operands is worthwhile. 646 if (Commute) { 647 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_YB); 648 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_YB); 649 } else { 650 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_BY); 651 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_BY); 652 } 653 return true; 654 } 655 656 return false; 657 } 658 659 /// Attempt the reassociation transformation to reduce critical path length. 660 /// See the above comments before getMachineCombinerPatterns(). 661 void TargetInstrInfo::reassociateOps( 662 MachineInstr &Root, MachineInstr &Prev, 663 MachineCombinerPattern Pattern, 664 SmallVectorImpl<MachineInstr *> &InsInstrs, 665 SmallVectorImpl<MachineInstr *> &DelInstrs, 666 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const { 667 MachineFunction *MF = Root.getParent()->getParent(); 668 MachineRegisterInfo &MRI = MF->getRegInfo(); 669 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); 670 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); 671 const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI); 672 673 // This array encodes the operand index for each parameter because the 674 // operands may be commuted. Each row corresponds to a pattern value, 675 // and each column specifies the index of A, B, X, Y. 676 unsigned OpIdx[4][4] = { 677 { 1, 1, 2, 2 }, 678 { 1, 2, 2, 1 }, 679 { 2, 1, 1, 2 }, 680 { 2, 2, 1, 1 } 681 }; 682 683 int Row; 684 switch (Pattern) { 685 case MachineCombinerPattern::REASSOC_AX_BY: Row = 0; break; 686 case MachineCombinerPattern::REASSOC_AX_YB: Row = 1; break; 687 case MachineCombinerPattern::REASSOC_XA_BY: Row = 2; break; 688 case MachineCombinerPattern::REASSOC_XA_YB: Row = 3; break; 689 default: llvm_unreachable("unexpected MachineCombinerPattern"); 690 } 691 692 MachineOperand &OpA = Prev.getOperand(OpIdx[Row][0]); 693 MachineOperand &OpB = Root.getOperand(OpIdx[Row][1]); 694 MachineOperand &OpX = Prev.getOperand(OpIdx[Row][2]); 695 MachineOperand &OpY = Root.getOperand(OpIdx[Row][3]); 696 MachineOperand &OpC = Root.getOperand(0); 697 698 unsigned RegA = OpA.getReg(); 699 unsigned RegB = OpB.getReg(); 700 unsigned RegX = OpX.getReg(); 701 unsigned RegY = OpY.getReg(); 702 unsigned RegC = OpC.getReg(); 703 704 if (TargetRegisterInfo::isVirtualRegister(RegA)) 705 MRI.constrainRegClass(RegA, RC); 706 if (TargetRegisterInfo::isVirtualRegister(RegB)) 707 MRI.constrainRegClass(RegB, RC); 708 if (TargetRegisterInfo::isVirtualRegister(RegX)) 709 MRI.constrainRegClass(RegX, RC); 710 if (TargetRegisterInfo::isVirtualRegister(RegY)) 711 MRI.constrainRegClass(RegY, RC); 712 if (TargetRegisterInfo::isVirtualRegister(RegC)) 713 MRI.constrainRegClass(RegC, RC); 714 715 // Create a new virtual register for the result of (X op Y) instead of 716 // recycling RegB because the MachineCombiner's computation of the critical 717 // path requires a new register definition rather than an existing one. 718 unsigned NewVR = MRI.createVirtualRegister(RC); 719 InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0)); 720 721 unsigned Opcode = Root.getOpcode(); 722 bool KillA = OpA.isKill(); 723 bool KillX = OpX.isKill(); 724 bool KillY = OpY.isKill(); 725 726 // Create new instructions for insertion. 727 MachineInstrBuilder MIB1 = 728 BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR) 729 .addReg(RegX, getKillRegState(KillX)) 730 .addReg(RegY, getKillRegState(KillY)); 731 MachineInstrBuilder MIB2 = 732 BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC) 733 .addReg(RegA, getKillRegState(KillA)) 734 .addReg(NewVR, getKillRegState(true)); 735 736 setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2); 737 738 // Record new instructions for insertion and old instructions for deletion. 739 InsInstrs.push_back(MIB1); 740 InsInstrs.push_back(MIB2); 741 DelInstrs.push_back(&Prev); 742 DelInstrs.push_back(&Root); 743 } 744 745 void TargetInstrInfo::genAlternativeCodeSequence( 746 MachineInstr &Root, MachineCombinerPattern Pattern, 747 SmallVectorImpl<MachineInstr *> &InsInstrs, 748 SmallVectorImpl<MachineInstr *> &DelInstrs, 749 DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const { 750 MachineRegisterInfo &MRI = Root.getParent()->getParent()->getRegInfo(); 751 752 // Select the previous instruction in the sequence based on the input pattern. 753 MachineInstr *Prev = nullptr; 754 switch (Pattern) { 755 case MachineCombinerPattern::REASSOC_AX_BY: 756 case MachineCombinerPattern::REASSOC_XA_BY: 757 Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg()); 758 break; 759 case MachineCombinerPattern::REASSOC_AX_YB: 760 case MachineCombinerPattern::REASSOC_XA_YB: 761 Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg()); 762 break; 763 default: 764 break; 765 } 766 767 assert(Prev && "Unknown pattern for machine combiner"); 768 769 reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg); 770 } 771 772 /// foldMemoryOperand - Same as the previous version except it allows folding 773 /// of any load and store from / to any address, not just from a specific 774 /// stack slot. 775 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI, 776 ArrayRef<unsigned> Ops, 777 MachineInstr *LoadMI) const { 778 assert(LoadMI->canFoldAsLoad() && "LoadMI isn't foldable!"); 779 #ifndef NDEBUG 780 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 781 assert(MI->getOperand(Ops[i]).isUse() && "Folding load into def!"); 782 #endif 783 MachineBasicBlock &MBB = *MI->getParent(); 784 MachineFunction &MF = *MBB.getParent(); 785 786 // Ask the target to do the actual folding. 787 MachineInstr *NewMI = nullptr; 788 int FrameIndex = 0; 789 790 if ((MI->getOpcode() == TargetOpcode::STACKMAP || 791 MI->getOpcode() == TargetOpcode::PATCHPOINT) && 792 isLoadFromStackSlot(LoadMI, FrameIndex)) { 793 // Fold stackmap/patchpoint. 794 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this); 795 if (NewMI) 796 NewMI = MBB.insert(MI, NewMI); 797 } else { 798 // Ask the target to do the actual folding. 799 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI); 800 } 801 802 if (!NewMI) return nullptr; 803 804 // Copy the memoperands from the load to the folded instruction. 805 if (MI->memoperands_empty()) { 806 NewMI->setMemRefs(LoadMI->memoperands_begin(), 807 LoadMI->memoperands_end()); 808 } 809 else { 810 // Handle the rare case of folding multiple loads. 811 NewMI->setMemRefs(MI->memoperands_begin(), 812 MI->memoperands_end()); 813 for (MachineInstr::mmo_iterator I = LoadMI->memoperands_begin(), 814 E = LoadMI->memoperands_end(); I != E; ++I) { 815 NewMI->addMemOperand(MF, *I); 816 } 817 } 818 return NewMI; 819 } 820 821 bool TargetInstrInfo:: 822 isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI, 823 AliasAnalysis *AA) const { 824 const MachineFunction &MF = *MI->getParent()->getParent(); 825 const MachineRegisterInfo &MRI = MF.getRegInfo(); 826 827 // Remat clients assume operand 0 is the defined register. 828 if (!MI->getNumOperands() || !MI->getOperand(0).isReg()) 829 return false; 830 unsigned DefReg = MI->getOperand(0).getReg(); 831 832 // A sub-register definition can only be rematerialized if the instruction 833 // doesn't read the other parts of the register. Otherwise it is really a 834 // read-modify-write operation on the full virtual register which cannot be 835 // moved safely. 836 if (TargetRegisterInfo::isVirtualRegister(DefReg) && 837 MI->getOperand(0).getSubReg() && MI->readsVirtualRegister(DefReg)) 838 return false; 839 840 // A load from a fixed stack slot can be rematerialized. This may be 841 // redundant with subsequent checks, but it's target-independent, 842 // simple, and a common case. 843 int FrameIdx = 0; 844 if (isLoadFromStackSlot(MI, FrameIdx) && 845 MF.getFrameInfo()->isImmutableObjectIndex(FrameIdx)) 846 return true; 847 848 // Avoid instructions obviously unsafe for remat. 849 if (MI->isNotDuplicable() || MI->mayStore() || 850 MI->hasUnmodeledSideEffects()) 851 return false; 852 853 // Don't remat inline asm. We have no idea how expensive it is 854 // even if it's side effect free. 855 if (MI->isInlineAsm()) 856 return false; 857 858 // Avoid instructions which load from potentially varying memory. 859 if (MI->mayLoad() && !MI->isInvariantLoad(AA)) 860 return false; 861 862 // If any of the registers accessed are non-constant, conservatively assume 863 // the instruction is not rematerializable. 864 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 865 const MachineOperand &MO = MI->getOperand(i); 866 if (!MO.isReg()) continue; 867 unsigned Reg = MO.getReg(); 868 if (Reg == 0) 869 continue; 870 871 // Check for a well-behaved physical register. 872 if (TargetRegisterInfo::isPhysicalRegister(Reg)) { 873 if (MO.isUse()) { 874 // If the physreg has no defs anywhere, it's just an ambient register 875 // and we can freely move its uses. Alternatively, if it's allocatable, 876 // it could get allocated to something with a def during allocation. 877 if (!MRI.isConstantPhysReg(Reg, MF)) 878 return false; 879 } else { 880 // A physreg def. We can't remat it. 881 return false; 882 } 883 continue; 884 } 885 886 // Only allow one virtual-register def. There may be multiple defs of the 887 // same virtual register, though. 888 if (MO.isDef() && Reg != DefReg) 889 return false; 890 891 // Don't allow any virtual-register uses. Rematting an instruction with 892 // virtual register uses would length the live ranges of the uses, which 893 // is not necessarily a good idea, certainly not "trivial". 894 if (MO.isUse()) 895 return false; 896 } 897 898 // Everything checked out. 899 return true; 900 } 901 902 int TargetInstrInfo::getSPAdjust(const MachineInstr *MI) const { 903 const MachineFunction *MF = MI->getParent()->getParent(); 904 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering(); 905 bool StackGrowsDown = 906 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 907 908 unsigned FrameSetupOpcode = getCallFrameSetupOpcode(); 909 unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode(); 910 911 if (MI->getOpcode() != FrameSetupOpcode && 912 MI->getOpcode() != FrameDestroyOpcode) 913 return 0; 914 915 int SPAdj = MI->getOperand(0).getImm(); 916 SPAdj = TFI->alignSPAdjust(SPAdj); 917 918 if ((!StackGrowsDown && MI->getOpcode() == FrameSetupOpcode) || 919 (StackGrowsDown && MI->getOpcode() == FrameDestroyOpcode)) 920 SPAdj = -SPAdj; 921 922 return SPAdj; 923 } 924 925 /// isSchedulingBoundary - Test if the given instruction should be 926 /// considered a scheduling boundary. This primarily includes labels 927 /// and terminators. 928 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr *MI, 929 const MachineBasicBlock *MBB, 930 const MachineFunction &MF) const { 931 // Terminators and labels can't be scheduled around. 932 if (MI->isTerminator() || MI->isPosition()) 933 return true; 934 935 // Don't attempt to schedule around any instruction that defines 936 // a stack-oriented pointer, as it's unlikely to be profitable. This 937 // saves compile time, because it doesn't require every single 938 // stack slot reference to depend on the instruction that does the 939 // modification. 940 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering(); 941 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 942 return MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI); 943 } 944 945 // Provide a global flag for disabling the PreRA hazard recognizer that targets 946 // may choose to honor. 947 bool TargetInstrInfo::usePreRAHazardRecognizer() const { 948 return !DisableHazardRecognizer; 949 } 950 951 // Default implementation of CreateTargetRAHazardRecognizer. 952 ScheduleHazardRecognizer *TargetInstrInfo:: 953 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, 954 const ScheduleDAG *DAG) const { 955 // Dummy hazard recognizer allows all instructions to issue. 956 return new ScheduleHazardRecognizer(); 957 } 958 959 // Default implementation of CreateTargetMIHazardRecognizer. 960 ScheduleHazardRecognizer *TargetInstrInfo:: 961 CreateTargetMIHazardRecognizer(const InstrItineraryData *II, 962 const ScheduleDAG *DAG) const { 963 return (ScheduleHazardRecognizer *) 964 new ScoreboardHazardRecognizer(II, DAG, "misched"); 965 } 966 967 // Default implementation of CreateTargetPostRAHazardRecognizer. 968 ScheduleHazardRecognizer *TargetInstrInfo:: 969 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 970 const ScheduleDAG *DAG) const { 971 return (ScheduleHazardRecognizer *) 972 new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched"); 973 } 974 975 //===----------------------------------------------------------------------===// 976 // SelectionDAG latency interface. 977 //===----------------------------------------------------------------------===// 978 979 int 980 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 981 SDNode *DefNode, unsigned DefIdx, 982 SDNode *UseNode, unsigned UseIdx) const { 983 if (!ItinData || ItinData->isEmpty()) 984 return -1; 985 986 if (!DefNode->isMachineOpcode()) 987 return -1; 988 989 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass(); 990 if (!UseNode->isMachineOpcode()) 991 return ItinData->getOperandCycle(DefClass, DefIdx); 992 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass(); 993 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 994 } 995 996 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 997 SDNode *N) const { 998 if (!ItinData || ItinData->isEmpty()) 999 return 1; 1000 1001 if (!N->isMachineOpcode()) 1002 return 1; 1003 1004 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass()); 1005 } 1006 1007 //===----------------------------------------------------------------------===// 1008 // MachineInstr latency interface. 1009 //===----------------------------------------------------------------------===// 1010 1011 unsigned 1012 TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, 1013 const MachineInstr *MI) const { 1014 if (!ItinData || ItinData->isEmpty()) 1015 return 1; 1016 1017 unsigned Class = MI->getDesc().getSchedClass(); 1018 int UOps = ItinData->Itineraries[Class].NumMicroOps; 1019 if (UOps >= 0) 1020 return UOps; 1021 1022 // The # of u-ops is dynamically determined. The specific target should 1023 // override this function to return the right number. 1024 return 1; 1025 } 1026 1027 /// Return the default expected latency for a def based on it's opcode. 1028 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel, 1029 const MachineInstr *DefMI) const { 1030 if (DefMI->isTransient()) 1031 return 0; 1032 if (DefMI->mayLoad()) 1033 return SchedModel.LoadLatency; 1034 if (isHighLatencyDef(DefMI->getOpcode())) 1035 return SchedModel.HighLatency; 1036 return 1; 1037 } 1038 1039 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr &) const { 1040 return 0; 1041 } 1042 1043 unsigned TargetInstrInfo:: 1044 getInstrLatency(const InstrItineraryData *ItinData, 1045 const MachineInstr *MI, 1046 unsigned *PredCost) const { 1047 // Default to one cycle for no itinerary. However, an "empty" itinerary may 1048 // still have a MinLatency property, which getStageLatency checks. 1049 if (!ItinData) 1050 return MI->mayLoad() ? 2 : 1; 1051 1052 return ItinData->getStageLatency(MI->getDesc().getSchedClass()); 1053 } 1054 1055 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel, 1056 const MachineInstr *DefMI, 1057 unsigned DefIdx) const { 1058 const InstrItineraryData *ItinData = SchedModel.getInstrItineraries(); 1059 if (!ItinData || ItinData->isEmpty()) 1060 return false; 1061 1062 unsigned DefClass = DefMI->getDesc().getSchedClass(); 1063 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx); 1064 return (DefCycle != -1 && DefCycle <= 1); 1065 } 1066 1067 /// Both DefMI and UseMI must be valid. By default, call directly to the 1068 /// itinerary. This may be overriden by the target. 1069 int TargetInstrInfo:: 1070 getOperandLatency(const InstrItineraryData *ItinData, 1071 const MachineInstr *DefMI, unsigned DefIdx, 1072 const MachineInstr *UseMI, unsigned UseIdx) const { 1073 unsigned DefClass = DefMI->getDesc().getSchedClass(); 1074 unsigned UseClass = UseMI->getDesc().getSchedClass(); 1075 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 1076 } 1077 1078 /// If we can determine the operand latency from the def only, without itinerary 1079 /// lookup, do so. Otherwise return -1. 1080 int TargetInstrInfo::computeDefOperandLatency( 1081 const InstrItineraryData *ItinData, 1082 const MachineInstr *DefMI) const { 1083 1084 // Let the target hook getInstrLatency handle missing itineraries. 1085 if (!ItinData) 1086 return getInstrLatency(ItinData, DefMI); 1087 1088 if(ItinData->isEmpty()) 1089 return defaultDefLatency(ItinData->SchedModel, DefMI); 1090 1091 // ...operand lookup required 1092 return -1; 1093 } 1094 1095 /// computeOperandLatency - Compute and return the latency of the given data 1096 /// dependent def and use when the operand indices are already known. UseMI may 1097 /// be NULL for an unknown use. 1098 /// 1099 /// FindMin may be set to get the minimum vs. expected latency. Minimum 1100 /// latency is used for scheduling groups, while expected latency is for 1101 /// instruction cost and critical path. 1102 /// 1103 /// Depending on the subtarget's itinerary properties, this may or may not need 1104 /// to call getOperandLatency(). For most subtargets, we don't need DefIdx or 1105 /// UseIdx to compute min latency. 1106 unsigned TargetInstrInfo:: 1107 computeOperandLatency(const InstrItineraryData *ItinData, 1108 const MachineInstr *DefMI, unsigned DefIdx, 1109 const MachineInstr *UseMI, unsigned UseIdx) const { 1110 1111 int DefLatency = computeDefOperandLatency(ItinData, DefMI); 1112 if (DefLatency >= 0) 1113 return DefLatency; 1114 1115 assert(ItinData && !ItinData->isEmpty() && "computeDefOperandLatency fail"); 1116 1117 int OperLatency = 0; 1118 if (UseMI) 1119 OperLatency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx); 1120 else { 1121 unsigned DefClass = DefMI->getDesc().getSchedClass(); 1122 OperLatency = ItinData->getOperandCycle(DefClass, DefIdx); 1123 } 1124 if (OperLatency >= 0) 1125 return OperLatency; 1126 1127 // No operand latency was found. 1128 unsigned InstrLatency = getInstrLatency(ItinData, DefMI); 1129 1130 // Expected latency is the max of the stage latency and itinerary props. 1131 InstrLatency = std::max(InstrLatency, 1132 defaultDefLatency(ItinData->SchedModel, DefMI)); 1133 return InstrLatency; 1134 } 1135 1136 bool TargetInstrInfo::getRegSequenceInputs( 1137 const MachineInstr &MI, unsigned DefIdx, 1138 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const { 1139 assert((MI.isRegSequence() || 1140 MI.isRegSequenceLike()) && "Instruction do not have the proper type"); 1141 1142 if (!MI.isRegSequence()) 1143 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs); 1144 1145 // We are looking at: 1146 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ... 1147 assert(DefIdx == 0 && "REG_SEQUENCE only has one def"); 1148 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx; 1149 OpIdx += 2) { 1150 const MachineOperand &MOReg = MI.getOperand(OpIdx); 1151 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1); 1152 assert(MOSubIdx.isImm() && 1153 "One of the subindex of the reg_sequence is not an immediate"); 1154 // Record Reg:SubReg, SubIdx. 1155 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(), 1156 (unsigned)MOSubIdx.getImm())); 1157 } 1158 return true; 1159 } 1160 1161 bool TargetInstrInfo::getExtractSubregInputs( 1162 const MachineInstr &MI, unsigned DefIdx, 1163 RegSubRegPairAndIdx &InputReg) const { 1164 assert((MI.isExtractSubreg() || 1165 MI.isExtractSubregLike()) && "Instruction do not have the proper type"); 1166 1167 if (!MI.isExtractSubreg()) 1168 return getExtractSubregLikeInputs(MI, DefIdx, InputReg); 1169 1170 // We are looking at: 1171 // Def = EXTRACT_SUBREG v0.sub1, sub0. 1172 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def"); 1173 const MachineOperand &MOReg = MI.getOperand(1); 1174 const MachineOperand &MOSubIdx = MI.getOperand(2); 1175 assert(MOSubIdx.isImm() && 1176 "The subindex of the extract_subreg is not an immediate"); 1177 1178 InputReg.Reg = MOReg.getReg(); 1179 InputReg.SubReg = MOReg.getSubReg(); 1180 InputReg.SubIdx = (unsigned)MOSubIdx.getImm(); 1181 return true; 1182 } 1183 1184 bool TargetInstrInfo::getInsertSubregInputs( 1185 const MachineInstr &MI, unsigned DefIdx, 1186 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const { 1187 assert((MI.isInsertSubreg() || 1188 MI.isInsertSubregLike()) && "Instruction do not have the proper type"); 1189 1190 if (!MI.isInsertSubreg()) 1191 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg); 1192 1193 // We are looking at: 1194 // Def = INSERT_SEQUENCE v0, v1, sub0. 1195 assert(DefIdx == 0 && "INSERT_SUBREG only has one def"); 1196 const MachineOperand &MOBaseReg = MI.getOperand(1); 1197 const MachineOperand &MOInsertedReg = MI.getOperand(2); 1198 const MachineOperand &MOSubIdx = MI.getOperand(3); 1199 assert(MOSubIdx.isImm() && 1200 "One of the subindex of the reg_sequence is not an immediate"); 1201 BaseReg.Reg = MOBaseReg.getReg(); 1202 BaseReg.SubReg = MOBaseReg.getSubReg(); 1203 1204 InsertedReg.Reg = MOInsertedReg.getReg(); 1205 InsertedReg.SubReg = MOInsertedReg.getSubReg(); 1206 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm(); 1207 return true; 1208 } 1209