1 //===-- TargetInstrInfo.cpp - Target Instruction Information --------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements the TargetInstrInfo class. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/CodeGen/TargetInstrInfo.h" 14 #include "llvm/CodeGen/MachineFrameInfo.h" 15 #include "llvm/CodeGen/MachineInstrBuilder.h" 16 #include "llvm/CodeGen/MachineMemOperand.h" 17 #include "llvm/CodeGen/MachineRegisterInfo.h" 18 #include "llvm/CodeGen/PseudoSourceValue.h" 19 #include "llvm/CodeGen/ScoreboardHazardRecognizer.h" 20 #include "llvm/CodeGen/StackMaps.h" 21 #include "llvm/CodeGen/TargetFrameLowering.h" 22 #include "llvm/CodeGen/TargetLowering.h" 23 #include "llvm/CodeGen/TargetRegisterInfo.h" 24 #include "llvm/CodeGen/TargetSchedule.h" 25 #include "llvm/IR/DataLayout.h" 26 #include "llvm/MC/MCAsmInfo.h" 27 #include "llvm/MC/MCInstrItineraries.h" 28 #include "llvm/Support/CommandLine.h" 29 #include "llvm/Support/ErrorHandling.h" 30 #include "llvm/Support/raw_ostream.h" 31 #include "llvm/Target/TargetMachine.h" 32 #include <cctype> 33 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 static bool isAsmComment(const char *Str, const MCAsmInfo &MAI) { 70 return strncmp(Str, MAI.getCommentString().data(), 71 MAI.getCommentString().size()) == 0; 72 } 73 74 /// Measure the specified inline asm to determine an approximation of its 75 /// length. 76 /// Comments (which run till the next SeparatorString or newline) do not 77 /// count as an instruction. 78 /// Any other non-whitespace text is considered an instruction, with 79 /// multiple instructions separated by SeparatorString or newlines. 80 /// Variable-length instructions are not handled here; this function 81 /// may be overloaded in the target code to do that. 82 /// We implement a special case of the .space directive which takes only a 83 /// single integer argument in base 10 that is the size in bytes. This is a 84 /// restricted form of the GAS directive in that we only interpret 85 /// simple--i.e. not a logical or arithmetic expression--size values without 86 /// the optional fill value. This is primarily used for creating arbitrary 87 /// sized inline asm blocks for testing purposes. 88 unsigned TargetInstrInfo::getInlineAsmLength( 89 const char *Str, 90 const MCAsmInfo &MAI, const TargetSubtargetInfo *STI) const { 91 // Count the number of instructions in the asm. 92 bool AtInsnStart = true; 93 unsigned Length = 0; 94 const unsigned MaxInstLength = MAI.getMaxInstLength(STI); 95 for (; *Str; ++Str) { 96 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(), 97 strlen(MAI.getSeparatorString())) == 0) { 98 AtInsnStart = true; 99 } else if (isAsmComment(Str, MAI)) { 100 // Stop counting as an instruction after a comment until the next 101 // separator. 102 AtInsnStart = false; 103 } 104 105 if (AtInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) { 106 unsigned AddLength = MaxInstLength; 107 if (strncmp(Str, ".space", 6) == 0) { 108 char *EStr; 109 int SpaceSize; 110 SpaceSize = strtol(Str + 6, &EStr, 10); 111 SpaceSize = SpaceSize < 0 ? 0 : SpaceSize; 112 while (*EStr != '\n' && std::isspace(static_cast<unsigned char>(*EStr))) 113 ++EStr; 114 if (*EStr == '\0' || *EStr == '\n' || 115 isAsmComment(EStr, MAI)) // Successfully parsed .space argument 116 AddLength = SpaceSize; 117 } 118 Length += AddLength; 119 AtInsnStart = false; 120 } 121 } 122 123 return Length; 124 } 125 126 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 127 /// after it, replacing it with an unconditional branch to NewDest. 128 void 129 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail, 130 MachineBasicBlock *NewDest) const { 131 MachineBasicBlock *MBB = Tail->getParent(); 132 133 // Remove all the old successors of MBB from the CFG. 134 while (!MBB->succ_empty()) 135 MBB->removeSuccessor(MBB->succ_begin()); 136 137 // Save off the debug loc before erasing the instruction. 138 DebugLoc DL = Tail->getDebugLoc(); 139 140 // Remove all the dead instructions from the end of MBB. 141 MBB->erase(Tail, MBB->end()); 142 143 // If MBB isn't immediately before MBB, insert a branch to it. 144 if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest)) 145 insertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(), DL); 146 MBB->addSuccessor(NewDest); 147 } 148 149 MachineInstr *TargetInstrInfo::commuteInstructionImpl(MachineInstr &MI, 150 bool NewMI, unsigned Idx1, 151 unsigned Idx2) const { 152 const MCInstrDesc &MCID = MI.getDesc(); 153 bool HasDef = MCID.getNumDefs(); 154 if (HasDef && !MI.getOperand(0).isReg()) 155 // No idea how to commute this instruction. Target should implement its own. 156 return nullptr; 157 158 unsigned CommutableOpIdx1 = Idx1; (void)CommutableOpIdx1; 159 unsigned CommutableOpIdx2 = Idx2; (void)CommutableOpIdx2; 160 assert(findCommutedOpIndices(MI, CommutableOpIdx1, CommutableOpIdx2) && 161 CommutableOpIdx1 == Idx1 && CommutableOpIdx2 == Idx2 && 162 "TargetInstrInfo::CommuteInstructionImpl(): not commutable operands."); 163 assert(MI.getOperand(Idx1).isReg() && MI.getOperand(Idx2).isReg() && 164 "This only knows how to commute register operands so far"); 165 166 unsigned Reg0 = HasDef ? MI.getOperand(0).getReg() : 0; 167 unsigned Reg1 = MI.getOperand(Idx1).getReg(); 168 unsigned Reg2 = MI.getOperand(Idx2).getReg(); 169 unsigned SubReg0 = HasDef ? MI.getOperand(0).getSubReg() : 0; 170 unsigned SubReg1 = MI.getOperand(Idx1).getSubReg(); 171 unsigned SubReg2 = MI.getOperand(Idx2).getSubReg(); 172 bool Reg1IsKill = MI.getOperand(Idx1).isKill(); 173 bool Reg2IsKill = MI.getOperand(Idx2).isKill(); 174 bool Reg1IsUndef = MI.getOperand(Idx1).isUndef(); 175 bool Reg2IsUndef = MI.getOperand(Idx2).isUndef(); 176 bool Reg1IsInternal = MI.getOperand(Idx1).isInternalRead(); 177 bool Reg2IsInternal = MI.getOperand(Idx2).isInternalRead(); 178 // Avoid calling isRenamable for virtual registers since we assert that 179 // renamable property is only queried/set for physical registers. 180 bool Reg1IsRenamable = TargetRegisterInfo::isPhysicalRegister(Reg1) 181 ? MI.getOperand(Idx1).isRenamable() 182 : false; 183 bool Reg2IsRenamable = TargetRegisterInfo::isPhysicalRegister(Reg2) 184 ? MI.getOperand(Idx2).isRenamable() 185 : false; 186 // If destination is tied to either of the commuted source register, then 187 // it must be updated. 188 if (HasDef && Reg0 == Reg1 && 189 MI.getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) { 190 Reg2IsKill = false; 191 Reg0 = Reg2; 192 SubReg0 = SubReg2; 193 } else if (HasDef && Reg0 == Reg2 && 194 MI.getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) { 195 Reg1IsKill = false; 196 Reg0 = Reg1; 197 SubReg0 = SubReg1; 198 } 199 200 MachineInstr *CommutedMI = nullptr; 201 if (NewMI) { 202 // Create a new instruction. 203 MachineFunction &MF = *MI.getMF(); 204 CommutedMI = MF.CloneMachineInstr(&MI); 205 } else { 206 CommutedMI = &MI; 207 } 208 209 if (HasDef) { 210 CommutedMI->getOperand(0).setReg(Reg0); 211 CommutedMI->getOperand(0).setSubReg(SubReg0); 212 } 213 CommutedMI->getOperand(Idx2).setReg(Reg1); 214 CommutedMI->getOperand(Idx1).setReg(Reg2); 215 CommutedMI->getOperand(Idx2).setSubReg(SubReg1); 216 CommutedMI->getOperand(Idx1).setSubReg(SubReg2); 217 CommutedMI->getOperand(Idx2).setIsKill(Reg1IsKill); 218 CommutedMI->getOperand(Idx1).setIsKill(Reg2IsKill); 219 CommutedMI->getOperand(Idx2).setIsUndef(Reg1IsUndef); 220 CommutedMI->getOperand(Idx1).setIsUndef(Reg2IsUndef); 221 CommutedMI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal); 222 CommutedMI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal); 223 // Avoid calling setIsRenamable for virtual registers since we assert that 224 // renamable property is only queried/set for physical registers. 225 if (TargetRegisterInfo::isPhysicalRegister(Reg1)) 226 CommutedMI->getOperand(Idx2).setIsRenamable(Reg1IsRenamable); 227 if (TargetRegisterInfo::isPhysicalRegister(Reg2)) 228 CommutedMI->getOperand(Idx1).setIsRenamable(Reg2IsRenamable); 229 return CommutedMI; 230 } 231 232 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr &MI, bool NewMI, 233 unsigned OpIdx1, 234 unsigned OpIdx2) const { 235 // If OpIdx1 or OpIdx2 is not specified, then this method is free to choose 236 // any commutable operand, which is done in findCommutedOpIndices() method 237 // called below. 238 if ((OpIdx1 == CommuteAnyOperandIndex || OpIdx2 == CommuteAnyOperandIndex) && 239 !findCommutedOpIndices(MI, OpIdx1, OpIdx2)) { 240 assert(MI.isCommutable() && 241 "Precondition violation: MI must be commutable."); 242 return nullptr; 243 } 244 return commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 245 } 246 247 bool TargetInstrInfo::fixCommutedOpIndices(unsigned &ResultIdx1, 248 unsigned &ResultIdx2, 249 unsigned CommutableOpIdx1, 250 unsigned CommutableOpIdx2) { 251 if (ResultIdx1 == CommuteAnyOperandIndex && 252 ResultIdx2 == CommuteAnyOperandIndex) { 253 ResultIdx1 = CommutableOpIdx1; 254 ResultIdx2 = CommutableOpIdx2; 255 } else if (ResultIdx1 == CommuteAnyOperandIndex) { 256 if (ResultIdx2 == CommutableOpIdx1) 257 ResultIdx1 = CommutableOpIdx2; 258 else if (ResultIdx2 == CommutableOpIdx2) 259 ResultIdx1 = CommutableOpIdx1; 260 else 261 return false; 262 } else if (ResultIdx2 == CommuteAnyOperandIndex) { 263 if (ResultIdx1 == CommutableOpIdx1) 264 ResultIdx2 = CommutableOpIdx2; 265 else if (ResultIdx1 == CommutableOpIdx2) 266 ResultIdx2 = CommutableOpIdx1; 267 else 268 return false; 269 } else 270 // Check that the result operand indices match the given commutable 271 // operand indices. 272 return (ResultIdx1 == CommutableOpIdx1 && ResultIdx2 == CommutableOpIdx2) || 273 (ResultIdx1 == CommutableOpIdx2 && ResultIdx2 == CommutableOpIdx1); 274 275 return true; 276 } 277 278 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr &MI, 279 unsigned &SrcOpIdx1, 280 unsigned &SrcOpIdx2) const { 281 assert(!MI.isBundle() && 282 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles"); 283 284 const MCInstrDesc &MCID = MI.getDesc(); 285 if (!MCID.isCommutable()) 286 return false; 287 288 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this 289 // is not true, then the target must implement this. 290 unsigned CommutableOpIdx1 = MCID.getNumDefs(); 291 unsigned CommutableOpIdx2 = CommutableOpIdx1 + 1; 292 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 293 CommutableOpIdx1, CommutableOpIdx2)) 294 return false; 295 296 if (!MI.getOperand(SrcOpIdx1).isReg() || !MI.getOperand(SrcOpIdx2).isReg()) 297 // No idea. 298 return false; 299 return true; 300 } 301 302 bool TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const { 303 if (!MI.isTerminator()) return false; 304 305 // Conditional branch is a special case. 306 if (MI.isBranch() && !MI.isBarrier()) 307 return true; 308 if (!MI.isPredicable()) 309 return true; 310 return !isPredicated(MI); 311 } 312 313 bool TargetInstrInfo::PredicateInstruction( 314 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { 315 bool MadeChange = false; 316 317 assert(!MI.isBundle() && 318 "TargetInstrInfo::PredicateInstruction() can't handle bundles"); 319 320 const MCInstrDesc &MCID = MI.getDesc(); 321 if (!MI.isPredicable()) 322 return false; 323 324 for (unsigned j = 0, i = 0, e = MI.getNumOperands(); i != e; ++i) { 325 if (MCID.OpInfo[i].isPredicate()) { 326 MachineOperand &MO = MI.getOperand(i); 327 if (MO.isReg()) { 328 MO.setReg(Pred[j].getReg()); 329 MadeChange = true; 330 } else if (MO.isImm()) { 331 MO.setImm(Pred[j].getImm()); 332 MadeChange = true; 333 } else if (MO.isMBB()) { 334 MO.setMBB(Pred[j].getMBB()); 335 MadeChange = true; 336 } 337 ++j; 338 } 339 } 340 return MadeChange; 341 } 342 343 bool TargetInstrInfo::hasLoadFromStackSlot( 344 const MachineInstr &MI, 345 SmallVectorImpl<const MachineMemOperand *> &Accesses) const { 346 size_t StartSize = Accesses.size(); 347 for (MachineInstr::mmo_iterator o = MI.memoperands_begin(), 348 oe = MI.memoperands_end(); 349 o != oe; ++o) { 350 if ((*o)->isLoad() && 351 dyn_cast_or_null<FixedStackPseudoSourceValue>((*o)->getPseudoValue())) 352 Accesses.push_back(*o); 353 } 354 return Accesses.size() != StartSize; 355 } 356 357 bool TargetInstrInfo::hasStoreToStackSlot( 358 const MachineInstr &MI, 359 SmallVectorImpl<const MachineMemOperand *> &Accesses) const { 360 size_t StartSize = Accesses.size(); 361 for (MachineInstr::mmo_iterator o = MI.memoperands_begin(), 362 oe = MI.memoperands_end(); 363 o != oe; ++o) { 364 if ((*o)->isStore() && 365 dyn_cast_or_null<FixedStackPseudoSourceValue>((*o)->getPseudoValue())) 366 Accesses.push_back(*o); 367 } 368 return Accesses.size() != StartSize; 369 } 370 371 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC, 372 unsigned SubIdx, unsigned &Size, 373 unsigned &Offset, 374 const MachineFunction &MF) const { 375 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 376 if (!SubIdx) { 377 Size = TRI->getSpillSize(*RC); 378 Offset = 0; 379 return true; 380 } 381 unsigned BitSize = TRI->getSubRegIdxSize(SubIdx); 382 // Convert bit size to byte size. 383 if (BitSize % 8) 384 return false; 385 386 int BitOffset = TRI->getSubRegIdxOffset(SubIdx); 387 if (BitOffset < 0 || BitOffset % 8) 388 return false; 389 390 Size = BitSize /= 8; 391 Offset = (unsigned)BitOffset / 8; 392 393 assert(TRI->getSpillSize(*RC) >= (Offset + Size) && "bad subregister range"); 394 395 if (!MF.getDataLayout().isLittleEndian()) { 396 Offset = TRI->getSpillSize(*RC) - (Offset + Size); 397 } 398 return true; 399 } 400 401 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB, 402 MachineBasicBlock::iterator I, 403 unsigned DestReg, unsigned SubIdx, 404 const MachineInstr &Orig, 405 const TargetRegisterInfo &TRI) const { 406 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig); 407 MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI); 408 MBB.insert(I, MI); 409 } 410 411 bool TargetInstrInfo::produceSameValue(const MachineInstr &MI0, 412 const MachineInstr &MI1, 413 const MachineRegisterInfo *MRI) const { 414 return MI0.isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs); 415 } 416 417 MachineInstr &TargetInstrInfo::duplicate(MachineBasicBlock &MBB, 418 MachineBasicBlock::iterator InsertBefore, const MachineInstr &Orig) const { 419 assert(!Orig.isNotDuplicable() && "Instruction cannot be duplicated"); 420 MachineFunction &MF = *MBB.getParent(); 421 return MF.CloneMachineInstrBundle(MBB, InsertBefore, Orig); 422 } 423 424 // If the COPY instruction in MI can be folded to a stack operation, return 425 // the register class to use. 426 static const TargetRegisterClass *canFoldCopy(const MachineInstr &MI, 427 unsigned FoldIdx) { 428 assert(MI.isCopy() && "MI must be a COPY instruction"); 429 if (MI.getNumOperands() != 2) 430 return nullptr; 431 assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand"); 432 433 const MachineOperand &FoldOp = MI.getOperand(FoldIdx); 434 const MachineOperand &LiveOp = MI.getOperand(1 - FoldIdx); 435 436 if (FoldOp.getSubReg() || LiveOp.getSubReg()) 437 return nullptr; 438 439 unsigned FoldReg = FoldOp.getReg(); 440 unsigned LiveReg = LiveOp.getReg(); 441 442 assert(TargetRegisterInfo::isVirtualRegister(FoldReg) && 443 "Cannot fold physregs"); 444 445 const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo(); 446 const TargetRegisterClass *RC = MRI.getRegClass(FoldReg); 447 448 if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg())) 449 return RC->contains(LiveOp.getReg()) ? RC : nullptr; 450 451 if (RC->hasSubClassEq(MRI.getRegClass(LiveReg))) 452 return RC; 453 454 // FIXME: Allow folding when register classes are memory compatible. 455 return nullptr; 456 } 457 458 void TargetInstrInfo::getNoop(MCInst &NopInst) const { 459 llvm_unreachable("Not implemented"); 460 } 461 462 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr &MI, 463 ArrayRef<unsigned> Ops, int FrameIndex, 464 const TargetInstrInfo &TII) { 465 unsigned StartIdx = 0; 466 switch (MI.getOpcode()) { 467 case TargetOpcode::STACKMAP: { 468 // StackMapLiveValues are foldable 469 StartIdx = StackMapOpers(&MI).getVarIdx(); 470 break; 471 } 472 case TargetOpcode::PATCHPOINT: { 473 // For PatchPoint, the call args are not foldable (even if reported in the 474 // stackmap e.g. via anyregcc). 475 StartIdx = PatchPointOpers(&MI).getVarIdx(); 476 break; 477 } 478 case TargetOpcode::STATEPOINT: { 479 // For statepoints, fold deopt and gc arguments, but not call arguments. 480 StartIdx = StatepointOpers(&MI).getVarIdx(); 481 break; 482 } 483 default: 484 llvm_unreachable("unexpected stackmap opcode"); 485 } 486 487 // Return false if any operands requested for folding are not foldable (not 488 // part of the stackmap's live values). 489 for (unsigned Op : Ops) { 490 if (Op < StartIdx) 491 return nullptr; 492 } 493 494 MachineInstr *NewMI = 495 MF.CreateMachineInstr(TII.get(MI.getOpcode()), MI.getDebugLoc(), true); 496 MachineInstrBuilder MIB(MF, NewMI); 497 498 // No need to fold return, the meta data, and function arguments 499 for (unsigned i = 0; i < StartIdx; ++i) 500 MIB.add(MI.getOperand(i)); 501 502 for (unsigned i = StartIdx; i < MI.getNumOperands(); ++i) { 503 MachineOperand &MO = MI.getOperand(i); 504 if (is_contained(Ops, i)) { 505 unsigned SpillSize; 506 unsigned SpillOffset; 507 // Compute the spill slot size and offset. 508 const TargetRegisterClass *RC = 509 MF.getRegInfo().getRegClass(MO.getReg()); 510 bool Valid = 511 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF); 512 if (!Valid) 513 report_fatal_error("cannot spill patchpoint subregister operand"); 514 MIB.addImm(StackMaps::IndirectMemRefOp); 515 MIB.addImm(SpillSize); 516 MIB.addFrameIndex(FrameIndex); 517 MIB.addImm(SpillOffset); 518 } 519 else 520 MIB.add(MO); 521 } 522 return NewMI; 523 } 524 525 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI, 526 ArrayRef<unsigned> Ops, int FI, 527 LiveIntervals *LIS, 528 VirtRegMap *VRM) const { 529 auto Flags = MachineMemOperand::MONone; 530 for (unsigned OpIdx : Ops) 531 Flags |= MI.getOperand(OpIdx).isDef() ? MachineMemOperand::MOStore 532 : MachineMemOperand::MOLoad; 533 534 MachineBasicBlock *MBB = MI.getParent(); 535 assert(MBB && "foldMemoryOperand needs an inserted instruction"); 536 MachineFunction &MF = *MBB->getParent(); 537 538 // If we're not folding a load into a subreg, the size of the load is the 539 // size of the spill slot. But if we are, we need to figure out what the 540 // actual load size is. 541 int64_t MemSize = 0; 542 const MachineFrameInfo &MFI = MF.getFrameInfo(); 543 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 544 545 if (Flags & MachineMemOperand::MOStore) { 546 MemSize = MFI.getObjectSize(FI); 547 } else { 548 for (unsigned OpIdx : Ops) { 549 int64_t OpSize = MFI.getObjectSize(FI); 550 551 if (auto SubReg = MI.getOperand(OpIdx).getSubReg()) { 552 unsigned SubRegSize = TRI->getSubRegIdxSize(SubReg); 553 if (SubRegSize > 0 && !(SubRegSize % 8)) 554 OpSize = SubRegSize / 8; 555 } 556 557 MemSize = std::max(MemSize, OpSize); 558 } 559 } 560 561 assert(MemSize && "Did not expect a zero-sized stack slot"); 562 563 MachineInstr *NewMI = nullptr; 564 565 if (MI.getOpcode() == TargetOpcode::STACKMAP || 566 MI.getOpcode() == TargetOpcode::PATCHPOINT || 567 MI.getOpcode() == TargetOpcode::STATEPOINT) { 568 // Fold stackmap/patchpoint. 569 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this); 570 if (NewMI) 571 MBB->insert(MI, NewMI); 572 } else { 573 // Ask the target to do the actual folding. 574 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI, LIS, VRM); 575 } 576 577 if (NewMI) { 578 NewMI->setMemRefs(MF, MI.memoperands()); 579 // Add a memory operand, foldMemoryOperandImpl doesn't do that. 580 assert((!(Flags & MachineMemOperand::MOStore) || 581 NewMI->mayStore()) && 582 "Folded a def to a non-store!"); 583 assert((!(Flags & MachineMemOperand::MOLoad) || 584 NewMI->mayLoad()) && 585 "Folded a use to a non-load!"); 586 assert(MFI.getObjectOffset(FI) != -1); 587 MachineMemOperand *MMO = MF.getMachineMemOperand( 588 MachinePointerInfo::getFixedStack(MF, FI), Flags, MemSize, 589 MFI.getObjectAlignment(FI)); 590 NewMI->addMemOperand(MF, MMO); 591 592 return NewMI; 593 } 594 595 // Straight COPY may fold as load/store. 596 if (!MI.isCopy() || Ops.size() != 1) 597 return nullptr; 598 599 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]); 600 if (!RC) 601 return nullptr; 602 603 const MachineOperand &MO = MI.getOperand(1 - Ops[0]); 604 MachineBasicBlock::iterator Pos = MI; 605 606 if (Flags == MachineMemOperand::MOStore) 607 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI); 608 else 609 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI); 610 return &*--Pos; 611 } 612 613 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI, 614 ArrayRef<unsigned> Ops, 615 MachineInstr &LoadMI, 616 LiveIntervals *LIS) const { 617 assert(LoadMI.canFoldAsLoad() && "LoadMI isn't foldable!"); 618 #ifndef NDEBUG 619 for (unsigned OpIdx : Ops) 620 assert(MI.getOperand(OpIdx).isUse() && "Folding load into def!"); 621 #endif 622 623 MachineBasicBlock &MBB = *MI.getParent(); 624 MachineFunction &MF = *MBB.getParent(); 625 626 // Ask the target to do the actual folding. 627 MachineInstr *NewMI = nullptr; 628 int FrameIndex = 0; 629 630 if ((MI.getOpcode() == TargetOpcode::STACKMAP || 631 MI.getOpcode() == TargetOpcode::PATCHPOINT || 632 MI.getOpcode() == TargetOpcode::STATEPOINT) && 633 isLoadFromStackSlot(LoadMI, FrameIndex)) { 634 // Fold stackmap/patchpoint. 635 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this); 636 if (NewMI) 637 NewMI = &*MBB.insert(MI, NewMI); 638 } else { 639 // Ask the target to do the actual folding. 640 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI, LIS); 641 } 642 643 if (!NewMI) 644 return nullptr; 645 646 // Copy the memoperands from the load to the folded instruction. 647 if (MI.memoperands_empty()) { 648 NewMI->setMemRefs(MF, LoadMI.memoperands()); 649 } else { 650 // Handle the rare case of folding multiple loads. 651 NewMI->setMemRefs(MF, MI.memoperands()); 652 for (MachineInstr::mmo_iterator I = LoadMI.memoperands_begin(), 653 E = LoadMI.memoperands_end(); 654 I != E; ++I) { 655 NewMI->addMemOperand(MF, *I); 656 } 657 } 658 return NewMI; 659 } 660 661 bool TargetInstrInfo::hasReassociableOperands( 662 const MachineInstr &Inst, const MachineBasicBlock *MBB) const { 663 const MachineOperand &Op1 = Inst.getOperand(1); 664 const MachineOperand &Op2 = Inst.getOperand(2); 665 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 666 667 // We need virtual register definitions for the operands that we will 668 // reassociate. 669 MachineInstr *MI1 = nullptr; 670 MachineInstr *MI2 = nullptr; 671 if (Op1.isReg() && TargetRegisterInfo::isVirtualRegister(Op1.getReg())) 672 MI1 = MRI.getUniqueVRegDef(Op1.getReg()); 673 if (Op2.isReg() && TargetRegisterInfo::isVirtualRegister(Op2.getReg())) 674 MI2 = MRI.getUniqueVRegDef(Op2.getReg()); 675 676 // And they need to be in the trace (otherwise, they won't have a depth). 677 return MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB; 678 } 679 680 bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst, 681 bool &Commuted) const { 682 const MachineBasicBlock *MBB = Inst.getParent(); 683 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 684 MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg()); 685 MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg()); 686 unsigned AssocOpcode = Inst.getOpcode(); 687 688 // If only one operand has the same opcode and it's the second source operand, 689 // the operands must be commuted. 690 Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode; 691 if (Commuted) 692 std::swap(MI1, MI2); 693 694 // 1. The previous instruction must be the same type as Inst. 695 // 2. The previous instruction must have virtual register definitions for its 696 // operands in the same basic block as Inst. 697 // 3. The previous instruction's result must only be used by Inst. 698 return MI1->getOpcode() == AssocOpcode && 699 hasReassociableOperands(*MI1, MBB) && 700 MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg()); 701 } 702 703 // 1. The operation must be associative and commutative. 704 // 2. The instruction must have virtual register definitions for its 705 // operands in the same basic block. 706 // 3. The instruction must have a reassociable sibling. 707 bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst, 708 bool &Commuted) const { 709 return isAssociativeAndCommutative(Inst) && 710 hasReassociableOperands(Inst, Inst.getParent()) && 711 hasReassociableSibling(Inst, Commuted); 712 } 713 714 // The concept of the reassociation pass is that these operations can benefit 715 // from this kind of transformation: 716 // 717 // A = ? op ? 718 // B = A op X (Prev) 719 // C = B op Y (Root) 720 // --> 721 // A = ? op ? 722 // B = X op Y 723 // C = A op B 724 // 725 // breaking the dependency between A and B, allowing them to be executed in 726 // parallel (or back-to-back in a pipeline) instead of depending on each other. 727 728 // FIXME: This has the potential to be expensive (compile time) while not 729 // improving the code at all. Some ways to limit the overhead: 730 // 1. Track successful transforms; bail out if hit rate gets too low. 731 // 2. Only enable at -O3 or some other non-default optimization level. 732 // 3. Pre-screen pattern candidates here: if an operand of the previous 733 // instruction is known to not increase the critical path, then don't match 734 // that pattern. 735 bool TargetInstrInfo::getMachineCombinerPatterns( 736 MachineInstr &Root, 737 SmallVectorImpl<MachineCombinerPattern> &Patterns) const { 738 bool Commute; 739 if (isReassociationCandidate(Root, Commute)) { 740 // We found a sequence of instructions that may be suitable for a 741 // reassociation of operands to increase ILP. Specify each commutation 742 // possibility for the Prev instruction in the sequence and let the 743 // machine combiner decide if changing the operands is worthwhile. 744 if (Commute) { 745 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_YB); 746 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_YB); 747 } else { 748 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_BY); 749 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_BY); 750 } 751 return true; 752 } 753 754 return false; 755 } 756 757 /// Return true when a code sequence can improve loop throughput. 758 bool 759 TargetInstrInfo::isThroughputPattern(MachineCombinerPattern Pattern) const { 760 return false; 761 } 762 763 /// Attempt the reassociation transformation to reduce critical path length. 764 /// See the above comments before getMachineCombinerPatterns(). 765 void TargetInstrInfo::reassociateOps( 766 MachineInstr &Root, MachineInstr &Prev, 767 MachineCombinerPattern Pattern, 768 SmallVectorImpl<MachineInstr *> &InsInstrs, 769 SmallVectorImpl<MachineInstr *> &DelInstrs, 770 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const { 771 MachineFunction *MF = Root.getMF(); 772 MachineRegisterInfo &MRI = MF->getRegInfo(); 773 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); 774 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); 775 const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI); 776 777 // This array encodes the operand index for each parameter because the 778 // operands may be commuted. Each row corresponds to a pattern value, 779 // and each column specifies the index of A, B, X, Y. 780 unsigned OpIdx[4][4] = { 781 { 1, 1, 2, 2 }, 782 { 1, 2, 2, 1 }, 783 { 2, 1, 1, 2 }, 784 { 2, 2, 1, 1 } 785 }; 786 787 int Row; 788 switch (Pattern) { 789 case MachineCombinerPattern::REASSOC_AX_BY: Row = 0; break; 790 case MachineCombinerPattern::REASSOC_AX_YB: Row = 1; break; 791 case MachineCombinerPattern::REASSOC_XA_BY: Row = 2; break; 792 case MachineCombinerPattern::REASSOC_XA_YB: Row = 3; break; 793 default: llvm_unreachable("unexpected MachineCombinerPattern"); 794 } 795 796 MachineOperand &OpA = Prev.getOperand(OpIdx[Row][0]); 797 MachineOperand &OpB = Root.getOperand(OpIdx[Row][1]); 798 MachineOperand &OpX = Prev.getOperand(OpIdx[Row][2]); 799 MachineOperand &OpY = Root.getOperand(OpIdx[Row][3]); 800 MachineOperand &OpC = Root.getOperand(0); 801 802 unsigned RegA = OpA.getReg(); 803 unsigned RegB = OpB.getReg(); 804 unsigned RegX = OpX.getReg(); 805 unsigned RegY = OpY.getReg(); 806 unsigned RegC = OpC.getReg(); 807 808 if (TargetRegisterInfo::isVirtualRegister(RegA)) 809 MRI.constrainRegClass(RegA, RC); 810 if (TargetRegisterInfo::isVirtualRegister(RegB)) 811 MRI.constrainRegClass(RegB, RC); 812 if (TargetRegisterInfo::isVirtualRegister(RegX)) 813 MRI.constrainRegClass(RegX, RC); 814 if (TargetRegisterInfo::isVirtualRegister(RegY)) 815 MRI.constrainRegClass(RegY, RC); 816 if (TargetRegisterInfo::isVirtualRegister(RegC)) 817 MRI.constrainRegClass(RegC, RC); 818 819 // Create a new virtual register for the result of (X op Y) instead of 820 // recycling RegB because the MachineCombiner's computation of the critical 821 // path requires a new register definition rather than an existing one. 822 unsigned NewVR = MRI.createVirtualRegister(RC); 823 InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0)); 824 825 unsigned Opcode = Root.getOpcode(); 826 bool KillA = OpA.isKill(); 827 bool KillX = OpX.isKill(); 828 bool KillY = OpY.isKill(); 829 830 // Create new instructions for insertion. 831 MachineInstrBuilder MIB1 = 832 BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR) 833 .addReg(RegX, getKillRegState(KillX)) 834 .addReg(RegY, getKillRegState(KillY)); 835 MachineInstrBuilder MIB2 = 836 BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC) 837 .addReg(RegA, getKillRegState(KillA)) 838 .addReg(NewVR, getKillRegState(true)); 839 840 setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2); 841 842 // Record new instructions for insertion and old instructions for deletion. 843 InsInstrs.push_back(MIB1); 844 InsInstrs.push_back(MIB2); 845 DelInstrs.push_back(&Prev); 846 DelInstrs.push_back(&Root); 847 } 848 849 void TargetInstrInfo::genAlternativeCodeSequence( 850 MachineInstr &Root, MachineCombinerPattern Pattern, 851 SmallVectorImpl<MachineInstr *> &InsInstrs, 852 SmallVectorImpl<MachineInstr *> &DelInstrs, 853 DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const { 854 MachineRegisterInfo &MRI = Root.getMF()->getRegInfo(); 855 856 // Select the previous instruction in the sequence based on the input pattern. 857 MachineInstr *Prev = nullptr; 858 switch (Pattern) { 859 case MachineCombinerPattern::REASSOC_AX_BY: 860 case MachineCombinerPattern::REASSOC_XA_BY: 861 Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg()); 862 break; 863 case MachineCombinerPattern::REASSOC_AX_YB: 864 case MachineCombinerPattern::REASSOC_XA_YB: 865 Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg()); 866 break; 867 default: 868 break; 869 } 870 871 assert(Prev && "Unknown pattern for machine combiner"); 872 873 reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg); 874 } 875 876 bool TargetInstrInfo::isReallyTriviallyReMaterializableGeneric( 877 const MachineInstr &MI, AliasAnalysis *AA) const { 878 const MachineFunction &MF = *MI.getMF(); 879 const MachineRegisterInfo &MRI = MF.getRegInfo(); 880 881 // Remat clients assume operand 0 is the defined register. 882 if (!MI.getNumOperands() || !MI.getOperand(0).isReg()) 883 return false; 884 unsigned DefReg = MI.getOperand(0).getReg(); 885 886 // A sub-register definition can only be rematerialized if the instruction 887 // doesn't read the other parts of the register. Otherwise it is really a 888 // read-modify-write operation on the full virtual register which cannot be 889 // moved safely. 890 if (TargetRegisterInfo::isVirtualRegister(DefReg) && 891 MI.getOperand(0).getSubReg() && MI.readsVirtualRegister(DefReg)) 892 return false; 893 894 // A load from a fixed stack slot can be rematerialized. This may be 895 // redundant with subsequent checks, but it's target-independent, 896 // simple, and a common case. 897 int FrameIdx = 0; 898 if (isLoadFromStackSlot(MI, FrameIdx) && 899 MF.getFrameInfo().isImmutableObjectIndex(FrameIdx)) 900 return true; 901 902 // Avoid instructions obviously unsafe for remat. 903 if (MI.isNotDuplicable() || MI.mayStore() || MI.mayRaiseFPException() || 904 MI.hasUnmodeledSideEffects()) 905 return false; 906 907 // Don't remat inline asm. We have no idea how expensive it is 908 // even if it's side effect free. 909 if (MI.isInlineAsm()) 910 return false; 911 912 // Avoid instructions which load from potentially varying memory. 913 if (MI.mayLoad() && !MI.isDereferenceableInvariantLoad(AA)) 914 return false; 915 916 // If any of the registers accessed are non-constant, conservatively assume 917 // the instruction is not rematerializable. 918 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 919 const MachineOperand &MO = MI.getOperand(i); 920 if (!MO.isReg()) continue; 921 unsigned Reg = MO.getReg(); 922 if (Reg == 0) 923 continue; 924 925 // Check for a well-behaved physical register. 926 if (TargetRegisterInfo::isPhysicalRegister(Reg)) { 927 if (MO.isUse()) { 928 // If the physreg has no defs anywhere, it's just an ambient register 929 // and we can freely move its uses. Alternatively, if it's allocatable, 930 // it could get allocated to something with a def during allocation. 931 if (!MRI.isConstantPhysReg(Reg)) 932 return false; 933 } else { 934 // A physreg def. We can't remat it. 935 return false; 936 } 937 continue; 938 } 939 940 // Only allow one virtual-register def. There may be multiple defs of the 941 // same virtual register, though. 942 if (MO.isDef() && Reg != DefReg) 943 return false; 944 945 // Don't allow any virtual-register uses. Rematting an instruction with 946 // virtual register uses would length the live ranges of the uses, which 947 // is not necessarily a good idea, certainly not "trivial". 948 if (MO.isUse()) 949 return false; 950 } 951 952 // Everything checked out. 953 return true; 954 } 955 956 int TargetInstrInfo::getSPAdjust(const MachineInstr &MI) const { 957 const MachineFunction *MF = MI.getMF(); 958 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering(); 959 bool StackGrowsDown = 960 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 961 962 unsigned FrameSetupOpcode = getCallFrameSetupOpcode(); 963 unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode(); 964 965 if (!isFrameInstr(MI)) 966 return 0; 967 968 int SPAdj = TFI->alignSPAdjust(getFrameSize(MI)); 969 970 if ((!StackGrowsDown && MI.getOpcode() == FrameSetupOpcode) || 971 (StackGrowsDown && MI.getOpcode() == FrameDestroyOpcode)) 972 SPAdj = -SPAdj; 973 974 return SPAdj; 975 } 976 977 /// isSchedulingBoundary - Test if the given instruction should be 978 /// considered a scheduling boundary. This primarily includes labels 979 /// and terminators. 980 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr &MI, 981 const MachineBasicBlock *MBB, 982 const MachineFunction &MF) const { 983 // Terminators and labels can't be scheduled around. 984 if (MI.isTerminator() || MI.isPosition()) 985 return true; 986 987 // Don't attempt to schedule around any instruction that defines 988 // a stack-oriented pointer, as it's unlikely to be profitable. This 989 // saves compile time, because it doesn't require every single 990 // stack slot reference to depend on the instruction that does the 991 // modification. 992 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering(); 993 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 994 return MI.modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI); 995 } 996 997 // Provide a global flag for disabling the PreRA hazard recognizer that targets 998 // may choose to honor. 999 bool TargetInstrInfo::usePreRAHazardRecognizer() const { 1000 return !DisableHazardRecognizer; 1001 } 1002 1003 // Default implementation of CreateTargetRAHazardRecognizer. 1004 ScheduleHazardRecognizer *TargetInstrInfo:: 1005 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, 1006 const ScheduleDAG *DAG) const { 1007 // Dummy hazard recognizer allows all instructions to issue. 1008 return new ScheduleHazardRecognizer(); 1009 } 1010 1011 // Default implementation of CreateTargetMIHazardRecognizer. 1012 ScheduleHazardRecognizer *TargetInstrInfo:: 1013 CreateTargetMIHazardRecognizer(const InstrItineraryData *II, 1014 const ScheduleDAG *DAG) const { 1015 return (ScheduleHazardRecognizer *) 1016 new ScoreboardHazardRecognizer(II, DAG, "machine-scheduler"); 1017 } 1018 1019 // Default implementation of CreateTargetPostRAHazardRecognizer. 1020 ScheduleHazardRecognizer *TargetInstrInfo:: 1021 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 1022 const ScheduleDAG *DAG) const { 1023 return (ScheduleHazardRecognizer *) 1024 new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched"); 1025 } 1026 1027 //===----------------------------------------------------------------------===// 1028 // SelectionDAG latency interface. 1029 //===----------------------------------------------------------------------===// 1030 1031 int 1032 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 1033 SDNode *DefNode, unsigned DefIdx, 1034 SDNode *UseNode, unsigned UseIdx) const { 1035 if (!ItinData || ItinData->isEmpty()) 1036 return -1; 1037 1038 if (!DefNode->isMachineOpcode()) 1039 return -1; 1040 1041 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass(); 1042 if (!UseNode->isMachineOpcode()) 1043 return ItinData->getOperandCycle(DefClass, DefIdx); 1044 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass(); 1045 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 1046 } 1047 1048 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 1049 SDNode *N) const { 1050 if (!ItinData || ItinData->isEmpty()) 1051 return 1; 1052 1053 if (!N->isMachineOpcode()) 1054 return 1; 1055 1056 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass()); 1057 } 1058 1059 //===----------------------------------------------------------------------===// 1060 // MachineInstr latency interface. 1061 //===----------------------------------------------------------------------===// 1062 1063 unsigned TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, 1064 const MachineInstr &MI) const { 1065 if (!ItinData || ItinData->isEmpty()) 1066 return 1; 1067 1068 unsigned Class = MI.getDesc().getSchedClass(); 1069 int UOps = ItinData->Itineraries[Class].NumMicroOps; 1070 if (UOps >= 0) 1071 return UOps; 1072 1073 // The # of u-ops is dynamically determined. The specific target should 1074 // override this function to return the right number. 1075 return 1; 1076 } 1077 1078 /// Return the default expected latency for a def based on it's opcode. 1079 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel, 1080 const MachineInstr &DefMI) const { 1081 if (DefMI.isTransient()) 1082 return 0; 1083 if (DefMI.mayLoad()) 1084 return SchedModel.LoadLatency; 1085 if (isHighLatencyDef(DefMI.getOpcode())) 1086 return SchedModel.HighLatency; 1087 return 1; 1088 } 1089 1090 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr &) const { 1091 return 0; 1092 } 1093 1094 unsigned TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 1095 const MachineInstr &MI, 1096 unsigned *PredCost) const { 1097 // Default to one cycle for no itinerary. However, an "empty" itinerary may 1098 // still have a MinLatency property, which getStageLatency checks. 1099 if (!ItinData) 1100 return MI.mayLoad() ? 2 : 1; 1101 1102 return ItinData->getStageLatency(MI.getDesc().getSchedClass()); 1103 } 1104 1105 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel, 1106 const MachineInstr &DefMI, 1107 unsigned DefIdx) const { 1108 const InstrItineraryData *ItinData = SchedModel.getInstrItineraries(); 1109 if (!ItinData || ItinData->isEmpty()) 1110 return false; 1111 1112 unsigned DefClass = DefMI.getDesc().getSchedClass(); 1113 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx); 1114 return (DefCycle != -1 && DefCycle <= 1); 1115 } 1116 1117 /// Both DefMI and UseMI must be valid. By default, call directly to the 1118 /// itinerary. This may be overriden by the target. 1119 int TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 1120 const MachineInstr &DefMI, 1121 unsigned DefIdx, 1122 const MachineInstr &UseMI, 1123 unsigned UseIdx) const { 1124 unsigned DefClass = DefMI.getDesc().getSchedClass(); 1125 unsigned UseClass = UseMI.getDesc().getSchedClass(); 1126 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 1127 } 1128 1129 /// If we can determine the operand latency from the def only, without itinerary 1130 /// lookup, do so. Otherwise return -1. 1131 int TargetInstrInfo::computeDefOperandLatency( 1132 const InstrItineraryData *ItinData, const MachineInstr &DefMI) const { 1133 1134 // Let the target hook getInstrLatency handle missing itineraries. 1135 if (!ItinData) 1136 return getInstrLatency(ItinData, DefMI); 1137 1138 if(ItinData->isEmpty()) 1139 return defaultDefLatency(ItinData->SchedModel, DefMI); 1140 1141 // ...operand lookup required 1142 return -1; 1143 } 1144 1145 bool TargetInstrInfo::getRegSequenceInputs( 1146 const MachineInstr &MI, unsigned DefIdx, 1147 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const { 1148 assert((MI.isRegSequence() || 1149 MI.isRegSequenceLike()) && "Instruction do not have the proper type"); 1150 1151 if (!MI.isRegSequence()) 1152 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs); 1153 1154 // We are looking at: 1155 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ... 1156 assert(DefIdx == 0 && "REG_SEQUENCE only has one def"); 1157 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx; 1158 OpIdx += 2) { 1159 const MachineOperand &MOReg = MI.getOperand(OpIdx); 1160 if (MOReg.isUndef()) 1161 continue; 1162 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1); 1163 assert(MOSubIdx.isImm() && 1164 "One of the subindex of the reg_sequence is not an immediate"); 1165 // Record Reg:SubReg, SubIdx. 1166 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(), 1167 (unsigned)MOSubIdx.getImm())); 1168 } 1169 return true; 1170 } 1171 1172 bool TargetInstrInfo::getExtractSubregInputs( 1173 const MachineInstr &MI, unsigned DefIdx, 1174 RegSubRegPairAndIdx &InputReg) const { 1175 assert((MI.isExtractSubreg() || 1176 MI.isExtractSubregLike()) && "Instruction do not have the proper type"); 1177 1178 if (!MI.isExtractSubreg()) 1179 return getExtractSubregLikeInputs(MI, DefIdx, InputReg); 1180 1181 // We are looking at: 1182 // Def = EXTRACT_SUBREG v0.sub1, sub0. 1183 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def"); 1184 const MachineOperand &MOReg = MI.getOperand(1); 1185 if (MOReg.isUndef()) 1186 return false; 1187 const MachineOperand &MOSubIdx = MI.getOperand(2); 1188 assert(MOSubIdx.isImm() && 1189 "The subindex of the extract_subreg is not an immediate"); 1190 1191 InputReg.Reg = MOReg.getReg(); 1192 InputReg.SubReg = MOReg.getSubReg(); 1193 InputReg.SubIdx = (unsigned)MOSubIdx.getImm(); 1194 return true; 1195 } 1196 1197 bool TargetInstrInfo::getInsertSubregInputs( 1198 const MachineInstr &MI, unsigned DefIdx, 1199 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const { 1200 assert((MI.isInsertSubreg() || 1201 MI.isInsertSubregLike()) && "Instruction do not have the proper type"); 1202 1203 if (!MI.isInsertSubreg()) 1204 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg); 1205 1206 // We are looking at: 1207 // Def = INSERT_SEQUENCE v0, v1, sub0. 1208 assert(DefIdx == 0 && "INSERT_SUBREG only has one def"); 1209 const MachineOperand &MOBaseReg = MI.getOperand(1); 1210 const MachineOperand &MOInsertedReg = MI.getOperand(2); 1211 if (MOInsertedReg.isUndef()) 1212 return false; 1213 const MachineOperand &MOSubIdx = MI.getOperand(3); 1214 assert(MOSubIdx.isImm() && 1215 "One of the subindex of the reg_sequence is not an immediate"); 1216 BaseReg.Reg = MOBaseReg.getReg(); 1217 BaseReg.SubReg = MOBaseReg.getSubReg(); 1218 1219 InsertedReg.Reg = MOInsertedReg.getReg(); 1220 InsertedReg.SubReg = MOInsertedReg.getSubReg(); 1221 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm(); 1222 return true; 1223 } 1224