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/MachineScheduler.h" 19 #include "llvm/CodeGen/PseudoSourceValue.h" 20 #include "llvm/CodeGen/ScoreboardHazardRecognizer.h" 21 #include "llvm/CodeGen/StackMaps.h" 22 #include "llvm/CodeGen/TargetFrameLowering.h" 23 #include "llvm/CodeGen/TargetLowering.h" 24 #include "llvm/CodeGen/TargetRegisterInfo.h" 25 #include "llvm/CodeGen/TargetSchedule.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/DebugInfoMetadata.h" 28 #include "llvm/MC/MCAsmInfo.h" 29 #include "llvm/MC/MCInstrItineraries.h" 30 #include "llvm/Support/CommandLine.h" 31 #include "llvm/Support/ErrorHandling.h" 32 #include "llvm/Support/raw_ostream.h" 33 #include "llvm/Target/TargetMachine.h" 34 #include <cctype> 35 36 using namespace llvm; 37 38 static cl::opt<bool> DisableHazardRecognizer( 39 "disable-sched-hazard", cl::Hidden, cl::init(false), 40 cl::desc("Disable hazard detection during preRA scheduling")); 41 42 TargetInstrInfo::~TargetInstrInfo() { 43 } 44 45 const TargetRegisterClass* 46 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum, 47 const TargetRegisterInfo *TRI, 48 const MachineFunction &MF) const { 49 if (OpNum >= MCID.getNumOperands()) 50 return nullptr; 51 52 short RegClass = MCID.OpInfo[OpNum].RegClass; 53 if (MCID.OpInfo[OpNum].isLookupPtrRegClass()) 54 return TRI->getPointerRegClass(MF, RegClass); 55 56 // Instructions like INSERT_SUBREG do not have fixed register classes. 57 if (RegClass < 0) 58 return nullptr; 59 60 // Otherwise just look it up normally. 61 return TRI->getRegClass(RegClass); 62 } 63 64 /// insertNoop - Insert a noop into the instruction stream at the specified 65 /// point. 66 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB, 67 MachineBasicBlock::iterator MI) const { 68 llvm_unreachable("Target didn't implement insertNoop!"); 69 } 70 71 static bool isAsmComment(const char *Str, const MCAsmInfo &MAI) { 72 return strncmp(Str, MAI.getCommentString().data(), 73 MAI.getCommentString().size()) == 0; 74 } 75 76 /// Measure the specified inline asm to determine an approximation of its 77 /// length. 78 /// Comments (which run till the next SeparatorString or newline) do not 79 /// count as an instruction. 80 /// Any other non-whitespace text is considered an instruction, with 81 /// multiple instructions separated by SeparatorString or newlines. 82 /// Variable-length instructions are not handled here; this function 83 /// may be overloaded in the target code to do that. 84 /// We implement a special case of the .space directive which takes only a 85 /// single integer argument in base 10 that is the size in bytes. This is a 86 /// restricted form of the GAS directive in that we only interpret 87 /// simple--i.e. not a logical or arithmetic expression--size values without 88 /// the optional fill value. This is primarily used for creating arbitrary 89 /// sized inline asm blocks for testing purposes. 90 unsigned TargetInstrInfo::getInlineAsmLength( 91 const char *Str, 92 const MCAsmInfo &MAI, const TargetSubtargetInfo *STI) const { 93 // Count the number of instructions in the asm. 94 bool AtInsnStart = true; 95 unsigned Length = 0; 96 const unsigned MaxInstLength = MAI.getMaxInstLength(STI); 97 for (; *Str; ++Str) { 98 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(), 99 strlen(MAI.getSeparatorString())) == 0) { 100 AtInsnStart = true; 101 } else if (isAsmComment(Str, MAI)) { 102 // Stop counting as an instruction after a comment until the next 103 // separator. 104 AtInsnStart = false; 105 } 106 107 if (AtInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) { 108 unsigned AddLength = MaxInstLength; 109 if (strncmp(Str, ".space", 6) == 0) { 110 char *EStr; 111 int SpaceSize; 112 SpaceSize = strtol(Str + 6, &EStr, 10); 113 SpaceSize = SpaceSize < 0 ? 0 : SpaceSize; 114 while (*EStr != '\n' && std::isspace(static_cast<unsigned char>(*EStr))) 115 ++EStr; 116 if (*EStr == '\0' || *EStr == '\n' || 117 isAsmComment(EStr, MAI)) // Successfully parsed .space argument 118 AddLength = SpaceSize; 119 } 120 Length += AddLength; 121 AtInsnStart = false; 122 } 123 } 124 125 return Length; 126 } 127 128 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 129 /// after it, replacing it with an unconditional branch to NewDest. 130 void 131 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail, 132 MachineBasicBlock *NewDest) const { 133 MachineBasicBlock *MBB = Tail->getParent(); 134 135 // Remove all the old successors of MBB from the CFG. 136 while (!MBB->succ_empty()) 137 MBB->removeSuccessor(MBB->succ_begin()); 138 139 // Save off the debug loc before erasing the instruction. 140 DebugLoc DL = Tail->getDebugLoc(); 141 142 // Update call site info and remove all the dead instructions 143 // from the end of MBB. 144 while (Tail != MBB->end()) { 145 auto MI = Tail++; 146 if (MI->isCall()) 147 MBB->getParent()->eraseCallSiteInfo(&*MI); 148 MBB->erase(MI); 149 } 150 151 // If MBB isn't immediately before MBB, insert a branch to it. 152 if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest)) 153 insertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(), DL); 154 MBB->addSuccessor(NewDest); 155 } 156 157 MachineInstr *TargetInstrInfo::commuteInstructionImpl(MachineInstr &MI, 158 bool NewMI, unsigned Idx1, 159 unsigned Idx2) const { 160 const MCInstrDesc &MCID = MI.getDesc(); 161 bool HasDef = MCID.getNumDefs(); 162 if (HasDef && !MI.getOperand(0).isReg()) 163 // No idea how to commute this instruction. Target should implement its own. 164 return nullptr; 165 166 unsigned CommutableOpIdx1 = Idx1; (void)CommutableOpIdx1; 167 unsigned CommutableOpIdx2 = Idx2; (void)CommutableOpIdx2; 168 assert(findCommutedOpIndices(MI, CommutableOpIdx1, CommutableOpIdx2) && 169 CommutableOpIdx1 == Idx1 && CommutableOpIdx2 == Idx2 && 170 "TargetInstrInfo::CommuteInstructionImpl(): not commutable operands."); 171 assert(MI.getOperand(Idx1).isReg() && MI.getOperand(Idx2).isReg() && 172 "This only knows how to commute register operands so far"); 173 174 Register Reg0 = HasDef ? MI.getOperand(0).getReg() : Register(); 175 Register Reg1 = MI.getOperand(Idx1).getReg(); 176 Register Reg2 = MI.getOperand(Idx2).getReg(); 177 unsigned SubReg0 = HasDef ? MI.getOperand(0).getSubReg() : 0; 178 unsigned SubReg1 = MI.getOperand(Idx1).getSubReg(); 179 unsigned SubReg2 = MI.getOperand(Idx2).getSubReg(); 180 bool Reg1IsKill = MI.getOperand(Idx1).isKill(); 181 bool Reg2IsKill = MI.getOperand(Idx2).isKill(); 182 bool Reg1IsUndef = MI.getOperand(Idx1).isUndef(); 183 bool Reg2IsUndef = MI.getOperand(Idx2).isUndef(); 184 bool Reg1IsInternal = MI.getOperand(Idx1).isInternalRead(); 185 bool Reg2IsInternal = MI.getOperand(Idx2).isInternalRead(); 186 // Avoid calling isRenamable for virtual registers since we assert that 187 // renamable property is only queried/set for physical registers. 188 bool Reg1IsRenamable = Register::isPhysicalRegister(Reg1) 189 ? MI.getOperand(Idx1).isRenamable() 190 : false; 191 bool Reg2IsRenamable = Register::isPhysicalRegister(Reg2) 192 ? MI.getOperand(Idx2).isRenamable() 193 : false; 194 // If destination is tied to either of the commuted source register, then 195 // it must be updated. 196 if (HasDef && Reg0 == Reg1 && 197 MI.getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) { 198 Reg2IsKill = false; 199 Reg0 = Reg2; 200 SubReg0 = SubReg2; 201 } else if (HasDef && Reg0 == Reg2 && 202 MI.getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) { 203 Reg1IsKill = false; 204 Reg0 = Reg1; 205 SubReg0 = SubReg1; 206 } 207 208 MachineInstr *CommutedMI = nullptr; 209 if (NewMI) { 210 // Create a new instruction. 211 MachineFunction &MF = *MI.getMF(); 212 CommutedMI = MF.CloneMachineInstr(&MI); 213 } else { 214 CommutedMI = &MI; 215 } 216 217 if (HasDef) { 218 CommutedMI->getOperand(0).setReg(Reg0); 219 CommutedMI->getOperand(0).setSubReg(SubReg0); 220 } 221 CommutedMI->getOperand(Idx2).setReg(Reg1); 222 CommutedMI->getOperand(Idx1).setReg(Reg2); 223 CommutedMI->getOperand(Idx2).setSubReg(SubReg1); 224 CommutedMI->getOperand(Idx1).setSubReg(SubReg2); 225 CommutedMI->getOperand(Idx2).setIsKill(Reg1IsKill); 226 CommutedMI->getOperand(Idx1).setIsKill(Reg2IsKill); 227 CommutedMI->getOperand(Idx2).setIsUndef(Reg1IsUndef); 228 CommutedMI->getOperand(Idx1).setIsUndef(Reg2IsUndef); 229 CommutedMI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal); 230 CommutedMI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal); 231 // Avoid calling setIsRenamable for virtual registers since we assert that 232 // renamable property is only queried/set for physical registers. 233 if (Register::isPhysicalRegister(Reg1)) 234 CommutedMI->getOperand(Idx2).setIsRenamable(Reg1IsRenamable); 235 if (Register::isPhysicalRegister(Reg2)) 236 CommutedMI->getOperand(Idx1).setIsRenamable(Reg2IsRenamable); 237 return CommutedMI; 238 } 239 240 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr &MI, bool NewMI, 241 unsigned OpIdx1, 242 unsigned OpIdx2) const { 243 // If OpIdx1 or OpIdx2 is not specified, then this method is free to choose 244 // any commutable operand, which is done in findCommutedOpIndices() method 245 // called below. 246 if ((OpIdx1 == CommuteAnyOperandIndex || OpIdx2 == CommuteAnyOperandIndex) && 247 !findCommutedOpIndices(MI, OpIdx1, OpIdx2)) { 248 assert(MI.isCommutable() && 249 "Precondition violation: MI must be commutable."); 250 return nullptr; 251 } 252 return commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 253 } 254 255 bool TargetInstrInfo::fixCommutedOpIndices(unsigned &ResultIdx1, 256 unsigned &ResultIdx2, 257 unsigned CommutableOpIdx1, 258 unsigned CommutableOpIdx2) { 259 if (ResultIdx1 == CommuteAnyOperandIndex && 260 ResultIdx2 == CommuteAnyOperandIndex) { 261 ResultIdx1 = CommutableOpIdx1; 262 ResultIdx2 = CommutableOpIdx2; 263 } else if (ResultIdx1 == CommuteAnyOperandIndex) { 264 if (ResultIdx2 == CommutableOpIdx1) 265 ResultIdx1 = CommutableOpIdx2; 266 else if (ResultIdx2 == CommutableOpIdx2) 267 ResultIdx1 = CommutableOpIdx1; 268 else 269 return false; 270 } else if (ResultIdx2 == CommuteAnyOperandIndex) { 271 if (ResultIdx1 == CommutableOpIdx1) 272 ResultIdx2 = CommutableOpIdx2; 273 else if (ResultIdx1 == CommutableOpIdx2) 274 ResultIdx2 = CommutableOpIdx1; 275 else 276 return false; 277 } else 278 // Check that the result operand indices match the given commutable 279 // operand indices. 280 return (ResultIdx1 == CommutableOpIdx1 && ResultIdx2 == CommutableOpIdx2) || 281 (ResultIdx1 == CommutableOpIdx2 && ResultIdx2 == CommutableOpIdx1); 282 283 return true; 284 } 285 286 bool TargetInstrInfo::findCommutedOpIndices(const MachineInstr &MI, 287 unsigned &SrcOpIdx1, 288 unsigned &SrcOpIdx2) const { 289 assert(!MI.isBundle() && 290 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles"); 291 292 const MCInstrDesc &MCID = MI.getDesc(); 293 if (!MCID.isCommutable()) 294 return false; 295 296 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this 297 // is not true, then the target must implement this. 298 unsigned CommutableOpIdx1 = MCID.getNumDefs(); 299 unsigned CommutableOpIdx2 = CommutableOpIdx1 + 1; 300 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 301 CommutableOpIdx1, CommutableOpIdx2)) 302 return false; 303 304 if (!MI.getOperand(SrcOpIdx1).isReg() || !MI.getOperand(SrcOpIdx2).isReg()) 305 // No idea. 306 return false; 307 return true; 308 } 309 310 bool TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const { 311 if (!MI.isTerminator()) return false; 312 313 // Conditional branch is a special case. 314 if (MI.isBranch() && !MI.isBarrier()) 315 return true; 316 if (!MI.isPredicable()) 317 return true; 318 return !isPredicated(MI); 319 } 320 321 bool TargetInstrInfo::PredicateInstruction( 322 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { 323 bool MadeChange = false; 324 325 assert(!MI.isBundle() && 326 "TargetInstrInfo::PredicateInstruction() can't handle bundles"); 327 328 const MCInstrDesc &MCID = MI.getDesc(); 329 if (!MI.isPredicable()) 330 return false; 331 332 for (unsigned j = 0, i = 0, e = MI.getNumOperands(); i != e; ++i) { 333 if (MCID.OpInfo[i].isPredicate()) { 334 MachineOperand &MO = MI.getOperand(i); 335 if (MO.isReg()) { 336 MO.setReg(Pred[j].getReg()); 337 MadeChange = true; 338 } else if (MO.isImm()) { 339 MO.setImm(Pred[j].getImm()); 340 MadeChange = true; 341 } else if (MO.isMBB()) { 342 MO.setMBB(Pred[j].getMBB()); 343 MadeChange = true; 344 } 345 ++j; 346 } 347 } 348 return MadeChange; 349 } 350 351 bool TargetInstrInfo::hasLoadFromStackSlot( 352 const MachineInstr &MI, 353 SmallVectorImpl<const MachineMemOperand *> &Accesses) const { 354 size_t StartSize = Accesses.size(); 355 for (MachineInstr::mmo_iterator o = MI.memoperands_begin(), 356 oe = MI.memoperands_end(); 357 o != oe; ++o) { 358 if ((*o)->isLoad() && 359 dyn_cast_or_null<FixedStackPseudoSourceValue>((*o)->getPseudoValue())) 360 Accesses.push_back(*o); 361 } 362 return Accesses.size() != StartSize; 363 } 364 365 bool TargetInstrInfo::hasStoreToStackSlot( 366 const MachineInstr &MI, 367 SmallVectorImpl<const MachineMemOperand *> &Accesses) const { 368 size_t StartSize = Accesses.size(); 369 for (MachineInstr::mmo_iterator o = MI.memoperands_begin(), 370 oe = MI.memoperands_end(); 371 o != oe; ++o) { 372 if ((*o)->isStore() && 373 dyn_cast_or_null<FixedStackPseudoSourceValue>((*o)->getPseudoValue())) 374 Accesses.push_back(*o); 375 } 376 return Accesses.size() != StartSize; 377 } 378 379 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC, 380 unsigned SubIdx, unsigned &Size, 381 unsigned &Offset, 382 const MachineFunction &MF) const { 383 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 384 if (!SubIdx) { 385 Size = TRI->getSpillSize(*RC); 386 Offset = 0; 387 return true; 388 } 389 unsigned BitSize = TRI->getSubRegIdxSize(SubIdx); 390 // Convert bit size to byte size. 391 if (BitSize % 8) 392 return false; 393 394 int BitOffset = TRI->getSubRegIdxOffset(SubIdx); 395 if (BitOffset < 0 || BitOffset % 8) 396 return false; 397 398 Size = BitSize / 8; 399 Offset = (unsigned)BitOffset / 8; 400 401 assert(TRI->getSpillSize(*RC) >= (Offset + Size) && "bad subregister range"); 402 403 if (!MF.getDataLayout().isLittleEndian()) { 404 Offset = TRI->getSpillSize(*RC) - (Offset + Size); 405 } 406 return true; 407 } 408 409 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB, 410 MachineBasicBlock::iterator I, 411 unsigned DestReg, unsigned SubIdx, 412 const MachineInstr &Orig, 413 const TargetRegisterInfo &TRI) const { 414 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(&Orig); 415 MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI); 416 MBB.insert(I, MI); 417 } 418 419 bool TargetInstrInfo::produceSameValue(const MachineInstr &MI0, 420 const MachineInstr &MI1, 421 const MachineRegisterInfo *MRI) const { 422 return MI0.isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs); 423 } 424 425 MachineInstr &TargetInstrInfo::duplicate(MachineBasicBlock &MBB, 426 MachineBasicBlock::iterator InsertBefore, const MachineInstr &Orig) const { 427 assert(!Orig.isNotDuplicable() && "Instruction cannot be duplicated"); 428 MachineFunction &MF = *MBB.getParent(); 429 return MF.CloneMachineInstrBundle(MBB, InsertBefore, Orig); 430 } 431 432 // If the COPY instruction in MI can be folded to a stack operation, return 433 // the register class to use. 434 static const TargetRegisterClass *canFoldCopy(const MachineInstr &MI, 435 unsigned FoldIdx) { 436 assert(MI.isCopy() && "MI must be a COPY instruction"); 437 if (MI.getNumOperands() != 2) 438 return nullptr; 439 assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand"); 440 441 const MachineOperand &FoldOp = MI.getOperand(FoldIdx); 442 const MachineOperand &LiveOp = MI.getOperand(1 - FoldIdx); 443 444 if (FoldOp.getSubReg() || LiveOp.getSubReg()) 445 return nullptr; 446 447 Register FoldReg = FoldOp.getReg(); 448 Register LiveReg = LiveOp.getReg(); 449 450 assert(Register::isVirtualRegister(FoldReg) && "Cannot fold physregs"); 451 452 const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo(); 453 const TargetRegisterClass *RC = MRI.getRegClass(FoldReg); 454 455 if (Register::isPhysicalRegister(LiveOp.getReg())) 456 return RC->contains(LiveOp.getReg()) ? RC : nullptr; 457 458 if (RC->hasSubClassEq(MRI.getRegClass(LiveReg))) 459 return RC; 460 461 // FIXME: Allow folding when register classes are memory compatible. 462 return nullptr; 463 } 464 465 void TargetInstrInfo::getNoop(MCInst &NopInst) const { 466 llvm_unreachable("Not implemented"); 467 } 468 469 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr &MI, 470 ArrayRef<unsigned> Ops, int FrameIndex, 471 const TargetInstrInfo &TII) { 472 unsigned StartIdx = 0; 473 switch (MI.getOpcode()) { 474 case TargetOpcode::STACKMAP: { 475 // StackMapLiveValues are foldable 476 StartIdx = StackMapOpers(&MI).getVarIdx(); 477 break; 478 } 479 case TargetOpcode::PATCHPOINT: { 480 // For PatchPoint, the call args are not foldable (even if reported in the 481 // stackmap e.g. via anyregcc). 482 StartIdx = PatchPointOpers(&MI).getVarIdx(); 483 break; 484 } 485 case TargetOpcode::STATEPOINT: { 486 // For statepoints, fold deopt and gc arguments, but not call arguments. 487 StartIdx = StatepointOpers(&MI).getVarIdx(); 488 break; 489 } 490 default: 491 llvm_unreachable("unexpected stackmap opcode"); 492 } 493 494 // Return false if any operands requested for folding are not foldable (not 495 // part of the stackmap's live values). 496 for (unsigned Op : Ops) { 497 if (Op < StartIdx) 498 return nullptr; 499 } 500 501 MachineInstr *NewMI = 502 MF.CreateMachineInstr(TII.get(MI.getOpcode()), MI.getDebugLoc(), true); 503 MachineInstrBuilder MIB(MF, NewMI); 504 505 // No need to fold return, the meta data, and function arguments 506 for (unsigned i = 0; i < StartIdx; ++i) 507 MIB.add(MI.getOperand(i)); 508 509 for (unsigned i = StartIdx; i < MI.getNumOperands(); ++i) { 510 MachineOperand &MO = MI.getOperand(i); 511 if (is_contained(Ops, i)) { 512 unsigned SpillSize; 513 unsigned SpillOffset; 514 // Compute the spill slot size and offset. 515 const TargetRegisterClass *RC = 516 MF.getRegInfo().getRegClass(MO.getReg()); 517 bool Valid = 518 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF); 519 if (!Valid) 520 report_fatal_error("cannot spill patchpoint subregister operand"); 521 MIB.addImm(StackMaps::IndirectMemRefOp); 522 MIB.addImm(SpillSize); 523 MIB.addFrameIndex(FrameIndex); 524 MIB.addImm(SpillOffset); 525 } 526 else 527 MIB.add(MO); 528 } 529 return NewMI; 530 } 531 532 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI, 533 ArrayRef<unsigned> Ops, int FI, 534 LiveIntervals *LIS, 535 VirtRegMap *VRM) const { 536 auto Flags = MachineMemOperand::MONone; 537 for (unsigned OpIdx : Ops) 538 Flags |= MI.getOperand(OpIdx).isDef() ? MachineMemOperand::MOStore 539 : MachineMemOperand::MOLoad; 540 541 MachineBasicBlock *MBB = MI.getParent(); 542 assert(MBB && "foldMemoryOperand needs an inserted instruction"); 543 MachineFunction &MF = *MBB->getParent(); 544 545 // If we're not folding a load into a subreg, the size of the load is the 546 // size of the spill slot. But if we are, we need to figure out what the 547 // actual load size is. 548 int64_t MemSize = 0; 549 const MachineFrameInfo &MFI = MF.getFrameInfo(); 550 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 551 552 if (Flags & MachineMemOperand::MOStore) { 553 MemSize = MFI.getObjectSize(FI); 554 } else { 555 for (unsigned OpIdx : Ops) { 556 int64_t OpSize = MFI.getObjectSize(FI); 557 558 if (auto SubReg = MI.getOperand(OpIdx).getSubReg()) { 559 unsigned SubRegSize = TRI->getSubRegIdxSize(SubReg); 560 if (SubRegSize > 0 && !(SubRegSize % 8)) 561 OpSize = SubRegSize / 8; 562 } 563 564 MemSize = std::max(MemSize, OpSize); 565 } 566 } 567 568 assert(MemSize && "Did not expect a zero-sized stack slot"); 569 570 MachineInstr *NewMI = nullptr; 571 572 if (MI.getOpcode() == TargetOpcode::STACKMAP || 573 MI.getOpcode() == TargetOpcode::PATCHPOINT || 574 MI.getOpcode() == TargetOpcode::STATEPOINT) { 575 // Fold stackmap/patchpoint. 576 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this); 577 if (NewMI) 578 MBB->insert(MI, NewMI); 579 } else { 580 // Ask the target to do the actual folding. 581 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI, LIS, VRM); 582 } 583 584 if (NewMI) { 585 NewMI->setMemRefs(MF, MI.memoperands()); 586 // Add a memory operand, foldMemoryOperandImpl doesn't do that. 587 assert((!(Flags & MachineMemOperand::MOStore) || 588 NewMI->mayStore()) && 589 "Folded a def to a non-store!"); 590 assert((!(Flags & MachineMemOperand::MOLoad) || 591 NewMI->mayLoad()) && 592 "Folded a use to a non-load!"); 593 assert(MFI.getObjectOffset(FI) != -1); 594 MachineMemOperand *MMO = MF.getMachineMemOperand( 595 MachinePointerInfo::getFixedStack(MF, FI), Flags, MemSize, 596 MFI.getObjectAlignment(FI)); 597 NewMI->addMemOperand(MF, MMO); 598 599 return NewMI; 600 } 601 602 // Straight COPY may fold as load/store. 603 if (!MI.isCopy() || Ops.size() != 1) 604 return nullptr; 605 606 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]); 607 if (!RC) 608 return nullptr; 609 610 const MachineOperand &MO = MI.getOperand(1 - Ops[0]); 611 MachineBasicBlock::iterator Pos = MI; 612 613 if (Flags == MachineMemOperand::MOStore) 614 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI); 615 else 616 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI); 617 return &*--Pos; 618 } 619 620 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI, 621 ArrayRef<unsigned> Ops, 622 MachineInstr &LoadMI, 623 LiveIntervals *LIS) const { 624 assert(LoadMI.canFoldAsLoad() && "LoadMI isn't foldable!"); 625 #ifndef NDEBUG 626 for (unsigned OpIdx : Ops) 627 assert(MI.getOperand(OpIdx).isUse() && "Folding load into def!"); 628 #endif 629 630 MachineBasicBlock &MBB = *MI.getParent(); 631 MachineFunction &MF = *MBB.getParent(); 632 633 // Ask the target to do the actual folding. 634 MachineInstr *NewMI = nullptr; 635 int FrameIndex = 0; 636 637 if ((MI.getOpcode() == TargetOpcode::STACKMAP || 638 MI.getOpcode() == TargetOpcode::PATCHPOINT || 639 MI.getOpcode() == TargetOpcode::STATEPOINT) && 640 isLoadFromStackSlot(LoadMI, FrameIndex)) { 641 // Fold stackmap/patchpoint. 642 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this); 643 if (NewMI) 644 NewMI = &*MBB.insert(MI, NewMI); 645 } else { 646 // Ask the target to do the actual folding. 647 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI, LIS); 648 } 649 650 if (!NewMI) 651 return nullptr; 652 653 // Copy the memoperands from the load to the folded instruction. 654 if (MI.memoperands_empty()) { 655 NewMI->setMemRefs(MF, LoadMI.memoperands()); 656 } else { 657 // Handle the rare case of folding multiple loads. 658 NewMI->setMemRefs(MF, MI.memoperands()); 659 for (MachineInstr::mmo_iterator I = LoadMI.memoperands_begin(), 660 E = LoadMI.memoperands_end(); 661 I != E; ++I) { 662 NewMI->addMemOperand(MF, *I); 663 } 664 } 665 return NewMI; 666 } 667 668 bool TargetInstrInfo::hasReassociableOperands( 669 const MachineInstr &Inst, const MachineBasicBlock *MBB) const { 670 const MachineOperand &Op1 = Inst.getOperand(1); 671 const MachineOperand &Op2 = Inst.getOperand(2); 672 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 673 674 // We need virtual register definitions for the operands that we will 675 // reassociate. 676 MachineInstr *MI1 = nullptr; 677 MachineInstr *MI2 = nullptr; 678 if (Op1.isReg() && Register::isVirtualRegister(Op1.getReg())) 679 MI1 = MRI.getUniqueVRegDef(Op1.getReg()); 680 if (Op2.isReg() && Register::isVirtualRegister(Op2.getReg())) 681 MI2 = MRI.getUniqueVRegDef(Op2.getReg()); 682 683 // And they need to be in the trace (otherwise, they won't have a depth). 684 return MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB; 685 } 686 687 bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst, 688 bool &Commuted) const { 689 const MachineBasicBlock *MBB = Inst.getParent(); 690 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 691 MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg()); 692 MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg()); 693 unsigned AssocOpcode = Inst.getOpcode(); 694 695 // If only one operand has the same opcode and it's the second source operand, 696 // the operands must be commuted. 697 Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode; 698 if (Commuted) 699 std::swap(MI1, MI2); 700 701 // 1. The previous instruction must be the same type as Inst. 702 // 2. The previous instruction must have virtual register definitions for its 703 // operands in the same basic block as Inst. 704 // 3. The previous instruction's result must only be used by Inst. 705 return MI1->getOpcode() == AssocOpcode && 706 hasReassociableOperands(*MI1, MBB) && 707 MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg()); 708 } 709 710 // 1. The operation must be associative and commutative. 711 // 2. The instruction must have virtual register definitions for its 712 // operands in the same basic block. 713 // 3. The instruction must have a reassociable sibling. 714 bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst, 715 bool &Commuted) const { 716 return isAssociativeAndCommutative(Inst) && 717 hasReassociableOperands(Inst, Inst.getParent()) && 718 hasReassociableSibling(Inst, Commuted); 719 } 720 721 // The concept of the reassociation pass is that these operations can benefit 722 // from this kind of transformation: 723 // 724 // A = ? op ? 725 // B = A op X (Prev) 726 // C = B op Y (Root) 727 // --> 728 // A = ? op ? 729 // B = X op Y 730 // C = A op B 731 // 732 // breaking the dependency between A and B, allowing them to be executed in 733 // parallel (or back-to-back in a pipeline) instead of depending on each other. 734 735 // FIXME: This has the potential to be expensive (compile time) while not 736 // improving the code at all. Some ways to limit the overhead: 737 // 1. Track successful transforms; bail out if hit rate gets too low. 738 // 2. Only enable at -O3 or some other non-default optimization level. 739 // 3. Pre-screen pattern candidates here: if an operand of the previous 740 // instruction is known to not increase the critical path, then don't match 741 // that pattern. 742 bool TargetInstrInfo::getMachineCombinerPatterns( 743 MachineInstr &Root, 744 SmallVectorImpl<MachineCombinerPattern> &Patterns) const { 745 bool Commute; 746 if (isReassociationCandidate(Root, Commute)) { 747 // We found a sequence of instructions that may be suitable for a 748 // reassociation of operands to increase ILP. Specify each commutation 749 // possibility for the Prev instruction in the sequence and let the 750 // machine combiner decide if changing the operands is worthwhile. 751 if (Commute) { 752 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_YB); 753 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_YB); 754 } else { 755 Patterns.push_back(MachineCombinerPattern::REASSOC_AX_BY); 756 Patterns.push_back(MachineCombinerPattern::REASSOC_XA_BY); 757 } 758 return true; 759 } 760 761 return false; 762 } 763 764 /// Return true when a code sequence can improve loop throughput. 765 bool 766 TargetInstrInfo::isThroughputPattern(MachineCombinerPattern Pattern) const { 767 return false; 768 } 769 770 /// Attempt the reassociation transformation to reduce critical path length. 771 /// See the above comments before getMachineCombinerPatterns(). 772 void TargetInstrInfo::reassociateOps( 773 MachineInstr &Root, MachineInstr &Prev, 774 MachineCombinerPattern Pattern, 775 SmallVectorImpl<MachineInstr *> &InsInstrs, 776 SmallVectorImpl<MachineInstr *> &DelInstrs, 777 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const { 778 MachineFunction *MF = Root.getMF(); 779 MachineRegisterInfo &MRI = MF->getRegInfo(); 780 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); 781 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); 782 const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI); 783 784 // This array encodes the operand index for each parameter because the 785 // operands may be commuted. Each row corresponds to a pattern value, 786 // and each column specifies the index of A, B, X, Y. 787 unsigned OpIdx[4][4] = { 788 { 1, 1, 2, 2 }, 789 { 1, 2, 2, 1 }, 790 { 2, 1, 1, 2 }, 791 { 2, 2, 1, 1 } 792 }; 793 794 int Row; 795 switch (Pattern) { 796 case MachineCombinerPattern::REASSOC_AX_BY: Row = 0; break; 797 case MachineCombinerPattern::REASSOC_AX_YB: Row = 1; break; 798 case MachineCombinerPattern::REASSOC_XA_BY: Row = 2; break; 799 case MachineCombinerPattern::REASSOC_XA_YB: Row = 3; break; 800 default: llvm_unreachable("unexpected MachineCombinerPattern"); 801 } 802 803 MachineOperand &OpA = Prev.getOperand(OpIdx[Row][0]); 804 MachineOperand &OpB = Root.getOperand(OpIdx[Row][1]); 805 MachineOperand &OpX = Prev.getOperand(OpIdx[Row][2]); 806 MachineOperand &OpY = Root.getOperand(OpIdx[Row][3]); 807 MachineOperand &OpC = Root.getOperand(0); 808 809 Register RegA = OpA.getReg(); 810 Register RegB = OpB.getReg(); 811 Register RegX = OpX.getReg(); 812 Register RegY = OpY.getReg(); 813 Register RegC = OpC.getReg(); 814 815 if (Register::isVirtualRegister(RegA)) 816 MRI.constrainRegClass(RegA, RC); 817 if (Register::isVirtualRegister(RegB)) 818 MRI.constrainRegClass(RegB, RC); 819 if (Register::isVirtualRegister(RegX)) 820 MRI.constrainRegClass(RegX, RC); 821 if (Register::isVirtualRegister(RegY)) 822 MRI.constrainRegClass(RegY, RC); 823 if (Register::isVirtualRegister(RegC)) 824 MRI.constrainRegClass(RegC, RC); 825 826 // Create a new virtual register for the result of (X op Y) instead of 827 // recycling RegB because the MachineCombiner's computation of the critical 828 // path requires a new register definition rather than an existing one. 829 Register NewVR = MRI.createVirtualRegister(RC); 830 InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0)); 831 832 unsigned Opcode = Root.getOpcode(); 833 bool KillA = OpA.isKill(); 834 bool KillX = OpX.isKill(); 835 bool KillY = OpY.isKill(); 836 837 // Create new instructions for insertion. 838 MachineInstrBuilder MIB1 = 839 BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR) 840 .addReg(RegX, getKillRegState(KillX)) 841 .addReg(RegY, getKillRegState(KillY)); 842 MachineInstrBuilder MIB2 = 843 BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC) 844 .addReg(RegA, getKillRegState(KillA)) 845 .addReg(NewVR, getKillRegState(true)); 846 847 setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2); 848 849 // Record new instructions for insertion and old instructions for deletion. 850 InsInstrs.push_back(MIB1); 851 InsInstrs.push_back(MIB2); 852 DelInstrs.push_back(&Prev); 853 DelInstrs.push_back(&Root); 854 } 855 856 void TargetInstrInfo::genAlternativeCodeSequence( 857 MachineInstr &Root, MachineCombinerPattern Pattern, 858 SmallVectorImpl<MachineInstr *> &InsInstrs, 859 SmallVectorImpl<MachineInstr *> &DelInstrs, 860 DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const { 861 MachineRegisterInfo &MRI = Root.getMF()->getRegInfo(); 862 863 // Select the previous instruction in the sequence based on the input pattern. 864 MachineInstr *Prev = nullptr; 865 switch (Pattern) { 866 case MachineCombinerPattern::REASSOC_AX_BY: 867 case MachineCombinerPattern::REASSOC_XA_BY: 868 Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg()); 869 break; 870 case MachineCombinerPattern::REASSOC_AX_YB: 871 case MachineCombinerPattern::REASSOC_XA_YB: 872 Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg()); 873 break; 874 default: 875 break; 876 } 877 878 assert(Prev && "Unknown pattern for machine combiner"); 879 880 reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg); 881 } 882 883 bool TargetInstrInfo::isReallyTriviallyReMaterializableGeneric( 884 const MachineInstr &MI, AAResults *AA) const { 885 const MachineFunction &MF = *MI.getMF(); 886 const MachineRegisterInfo &MRI = MF.getRegInfo(); 887 888 // Remat clients assume operand 0 is the defined register. 889 if (!MI.getNumOperands() || !MI.getOperand(0).isReg()) 890 return false; 891 Register DefReg = MI.getOperand(0).getReg(); 892 893 // A sub-register definition can only be rematerialized if the instruction 894 // doesn't read the other parts of the register. Otherwise it is really a 895 // read-modify-write operation on the full virtual register which cannot be 896 // moved safely. 897 if (Register::isVirtualRegister(DefReg) && MI.getOperand(0).getSubReg() && 898 MI.readsVirtualRegister(DefReg)) 899 return false; 900 901 // A load from a fixed stack slot can be rematerialized. This may be 902 // redundant with subsequent checks, but it's target-independent, 903 // simple, and a common case. 904 int FrameIdx = 0; 905 if (isLoadFromStackSlot(MI, FrameIdx) && 906 MF.getFrameInfo().isImmutableObjectIndex(FrameIdx)) 907 return true; 908 909 // Avoid instructions obviously unsafe for remat. 910 if (MI.isNotDuplicable() || MI.mayStore() || MI.mayRaiseFPException() || 911 MI.hasUnmodeledSideEffects()) 912 return false; 913 914 // Don't remat inline asm. We have no idea how expensive it is 915 // even if it's side effect free. 916 if (MI.isInlineAsm()) 917 return false; 918 919 // Avoid instructions which load from potentially varying memory. 920 if (MI.mayLoad() && !MI.isDereferenceableInvariantLoad(AA)) 921 return false; 922 923 // If any of the registers accessed are non-constant, conservatively assume 924 // the instruction is not rematerializable. 925 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 926 const MachineOperand &MO = MI.getOperand(i); 927 if (!MO.isReg()) continue; 928 Register Reg = MO.getReg(); 929 if (Reg == 0) 930 continue; 931 932 // Check for a well-behaved physical register. 933 if (Register::isPhysicalRegister(Reg)) { 934 if (MO.isUse()) { 935 // If the physreg has no defs anywhere, it's just an ambient register 936 // and we can freely move its uses. Alternatively, if it's allocatable, 937 // it could get allocated to something with a def during allocation. 938 if (!MRI.isConstantPhysReg(Reg)) 939 return false; 940 } else { 941 // A physreg def. We can't remat it. 942 return false; 943 } 944 continue; 945 } 946 947 // Only allow one virtual-register def. There may be multiple defs of the 948 // same virtual register, though. 949 if (MO.isDef() && Reg != DefReg) 950 return false; 951 952 // Don't allow any virtual-register uses. Rematting an instruction with 953 // virtual register uses would length the live ranges of the uses, which 954 // is not necessarily a good idea, certainly not "trivial". 955 if (MO.isUse()) 956 return false; 957 } 958 959 // Everything checked out. 960 return true; 961 } 962 963 int TargetInstrInfo::getSPAdjust(const MachineInstr &MI) const { 964 const MachineFunction *MF = MI.getMF(); 965 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering(); 966 bool StackGrowsDown = 967 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 968 969 unsigned FrameSetupOpcode = getCallFrameSetupOpcode(); 970 unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode(); 971 972 if (!isFrameInstr(MI)) 973 return 0; 974 975 int SPAdj = TFI->alignSPAdjust(getFrameSize(MI)); 976 977 if ((!StackGrowsDown && MI.getOpcode() == FrameSetupOpcode) || 978 (StackGrowsDown && MI.getOpcode() == FrameDestroyOpcode)) 979 SPAdj = -SPAdj; 980 981 return SPAdj; 982 } 983 984 /// isSchedulingBoundary - Test if the given instruction should be 985 /// considered a scheduling boundary. This primarily includes labels 986 /// and terminators. 987 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr &MI, 988 const MachineBasicBlock *MBB, 989 const MachineFunction &MF) const { 990 // Terminators and labels can't be scheduled around. 991 if (MI.isTerminator() || MI.isPosition()) 992 return true; 993 994 // Don't attempt to schedule around any instruction that defines 995 // a stack-oriented pointer, as it's unlikely to be profitable. This 996 // saves compile time, because it doesn't require every single 997 // stack slot reference to depend on the instruction that does the 998 // modification. 999 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering(); 1000 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1001 return MI.modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI); 1002 } 1003 1004 // Provide a global flag for disabling the PreRA hazard recognizer that targets 1005 // may choose to honor. 1006 bool TargetInstrInfo::usePreRAHazardRecognizer() const { 1007 return !DisableHazardRecognizer; 1008 } 1009 1010 // Default implementation of CreateTargetRAHazardRecognizer. 1011 ScheduleHazardRecognizer *TargetInstrInfo:: 1012 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, 1013 const ScheduleDAG *DAG) const { 1014 // Dummy hazard recognizer allows all instructions to issue. 1015 return new ScheduleHazardRecognizer(); 1016 } 1017 1018 // Default implementation of CreateTargetMIHazardRecognizer. 1019 ScheduleHazardRecognizer *TargetInstrInfo::CreateTargetMIHazardRecognizer( 1020 const InstrItineraryData *II, const ScheduleDAGMI *DAG) const { 1021 return new ScoreboardHazardRecognizer(II, DAG, "machine-scheduler"); 1022 } 1023 1024 // Default implementation of CreateTargetPostRAHazardRecognizer. 1025 ScheduleHazardRecognizer *TargetInstrInfo:: 1026 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 1027 const ScheduleDAG *DAG) const { 1028 return new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched"); 1029 } 1030 1031 //===----------------------------------------------------------------------===// 1032 // SelectionDAG latency interface. 1033 //===----------------------------------------------------------------------===// 1034 1035 int 1036 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 1037 SDNode *DefNode, unsigned DefIdx, 1038 SDNode *UseNode, unsigned UseIdx) const { 1039 if (!ItinData || ItinData->isEmpty()) 1040 return -1; 1041 1042 if (!DefNode->isMachineOpcode()) 1043 return -1; 1044 1045 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass(); 1046 if (!UseNode->isMachineOpcode()) 1047 return ItinData->getOperandCycle(DefClass, DefIdx); 1048 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass(); 1049 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 1050 } 1051 1052 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 1053 SDNode *N) const { 1054 if (!ItinData || ItinData->isEmpty()) 1055 return 1; 1056 1057 if (!N->isMachineOpcode()) 1058 return 1; 1059 1060 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass()); 1061 } 1062 1063 //===----------------------------------------------------------------------===// 1064 // MachineInstr latency interface. 1065 //===----------------------------------------------------------------------===// 1066 1067 unsigned TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, 1068 const MachineInstr &MI) const { 1069 if (!ItinData || ItinData->isEmpty()) 1070 return 1; 1071 1072 unsigned Class = MI.getDesc().getSchedClass(); 1073 int UOps = ItinData->Itineraries[Class].NumMicroOps; 1074 if (UOps >= 0) 1075 return UOps; 1076 1077 // The # of u-ops is dynamically determined. The specific target should 1078 // override this function to return the right number. 1079 return 1; 1080 } 1081 1082 /// Return the default expected latency for a def based on it's opcode. 1083 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel, 1084 const MachineInstr &DefMI) const { 1085 if (DefMI.isTransient()) 1086 return 0; 1087 if (DefMI.mayLoad()) 1088 return SchedModel.LoadLatency; 1089 if (isHighLatencyDef(DefMI.getOpcode())) 1090 return SchedModel.HighLatency; 1091 return 1; 1092 } 1093 1094 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr &) const { 1095 return 0; 1096 } 1097 1098 unsigned TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 1099 const MachineInstr &MI, 1100 unsigned *PredCost) const { 1101 // Default to one cycle for no itinerary. However, an "empty" itinerary may 1102 // still have a MinLatency property, which getStageLatency checks. 1103 if (!ItinData) 1104 return MI.mayLoad() ? 2 : 1; 1105 1106 return ItinData->getStageLatency(MI.getDesc().getSchedClass()); 1107 } 1108 1109 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel, 1110 const MachineInstr &DefMI, 1111 unsigned DefIdx) const { 1112 const InstrItineraryData *ItinData = SchedModel.getInstrItineraries(); 1113 if (!ItinData || ItinData->isEmpty()) 1114 return false; 1115 1116 unsigned DefClass = DefMI.getDesc().getSchedClass(); 1117 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx); 1118 return (DefCycle != -1 && DefCycle <= 1); 1119 } 1120 1121 Optional<ParamLoadedValue> 1122 TargetInstrInfo::describeLoadedValue(const MachineInstr &MI, 1123 Register Reg) const { 1124 const MachineFunction *MF = MI.getMF(); 1125 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); 1126 DIExpression *Expr = DIExpression::get(MF->getFunction().getContext(), {}); 1127 int64_t Offset; 1128 1129 // To simplify the sub-register handling, verify that we only need to 1130 // consider physical registers. 1131 assert(MF->getProperties().hasProperty( 1132 MachineFunctionProperties::Property::NoVRegs)); 1133 1134 if (auto DestSrc = isCopyInstr(MI)) { 1135 Register DestReg = DestSrc->Destination->getReg(); 1136 1137 if (Reg == DestReg) 1138 return ParamLoadedValue(*DestSrc->Source, Expr); 1139 1140 // Cases where super- or sub-registers needs to be described should 1141 // be handled by the target's hook implementation. 1142 assert(!TRI->isSuperOrSubRegisterEq(Reg, DestReg) && 1143 "TargetInstrInfo::describeLoadedValue can't describe super- or " 1144 "sub-regs for copy instructions"); 1145 return None; 1146 } else if (auto RegImm = isAddImmediate(MI, Reg)) { 1147 Register SrcReg = RegImm->Reg; 1148 Offset = RegImm->Imm; 1149 Expr = DIExpression::prepend(Expr, DIExpression::ApplyOffset, Offset); 1150 return ParamLoadedValue(MachineOperand::CreateReg(SrcReg, false), Expr); 1151 } else if (MI.hasOneMemOperand()) { 1152 // Only describe memory which provably does not escape the function. As 1153 // described in llvm.org/PR43343, escaped memory may be clobbered by the 1154 // callee (or by another thread). 1155 const auto &TII = MF->getSubtarget().getInstrInfo(); 1156 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1157 const MachineMemOperand *MMO = MI.memoperands()[0]; 1158 const PseudoSourceValue *PSV = MMO->getPseudoValue(); 1159 1160 // If the address points to "special" memory (e.g. a spill slot), it's 1161 // sufficient to check that it isn't aliased by any high-level IR value. 1162 if (!PSV || PSV->mayAlias(&MFI)) 1163 return None; 1164 1165 const MachineOperand *BaseOp; 1166 if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, TRI)) 1167 return None; 1168 1169 assert(MI.getNumExplicitDefs() == 1 && 1170 "Can currently only handle mem instructions with a single define"); 1171 1172 // TODO: In what way do we need to take Reg into consideration here? 1173 1174 SmallVector<uint64_t, 8> Ops; 1175 DIExpression::appendOffset(Ops, Offset); 1176 Ops.push_back(dwarf::DW_OP_deref_size); 1177 Ops.push_back(MMO->getSize()); 1178 Expr = DIExpression::prependOpcodes(Expr, Ops); 1179 return ParamLoadedValue(*BaseOp, Expr); 1180 } 1181 1182 return None; 1183 } 1184 1185 /// Both DefMI and UseMI must be valid. By default, call directly to the 1186 /// itinerary. This may be overriden by the target. 1187 int TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 1188 const MachineInstr &DefMI, 1189 unsigned DefIdx, 1190 const MachineInstr &UseMI, 1191 unsigned UseIdx) const { 1192 unsigned DefClass = DefMI.getDesc().getSchedClass(); 1193 unsigned UseClass = UseMI.getDesc().getSchedClass(); 1194 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); 1195 } 1196 1197 /// If we can determine the operand latency from the def only, without itinerary 1198 /// lookup, do so. Otherwise return -1. 1199 int TargetInstrInfo::computeDefOperandLatency( 1200 const InstrItineraryData *ItinData, const MachineInstr &DefMI) const { 1201 1202 // Let the target hook getInstrLatency handle missing itineraries. 1203 if (!ItinData) 1204 return getInstrLatency(ItinData, DefMI); 1205 1206 if(ItinData->isEmpty()) 1207 return defaultDefLatency(ItinData->SchedModel, DefMI); 1208 1209 // ...operand lookup required 1210 return -1; 1211 } 1212 1213 bool TargetInstrInfo::getRegSequenceInputs( 1214 const MachineInstr &MI, unsigned DefIdx, 1215 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const { 1216 assert((MI.isRegSequence() || 1217 MI.isRegSequenceLike()) && "Instruction do not have the proper type"); 1218 1219 if (!MI.isRegSequence()) 1220 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs); 1221 1222 // We are looking at: 1223 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ... 1224 assert(DefIdx == 0 && "REG_SEQUENCE only has one def"); 1225 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx; 1226 OpIdx += 2) { 1227 const MachineOperand &MOReg = MI.getOperand(OpIdx); 1228 if (MOReg.isUndef()) 1229 continue; 1230 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1); 1231 assert(MOSubIdx.isImm() && 1232 "One of the subindex of the reg_sequence is not an immediate"); 1233 // Record Reg:SubReg, SubIdx. 1234 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(), 1235 (unsigned)MOSubIdx.getImm())); 1236 } 1237 return true; 1238 } 1239 1240 bool TargetInstrInfo::getExtractSubregInputs( 1241 const MachineInstr &MI, unsigned DefIdx, 1242 RegSubRegPairAndIdx &InputReg) const { 1243 assert((MI.isExtractSubreg() || 1244 MI.isExtractSubregLike()) && "Instruction do not have the proper type"); 1245 1246 if (!MI.isExtractSubreg()) 1247 return getExtractSubregLikeInputs(MI, DefIdx, InputReg); 1248 1249 // We are looking at: 1250 // Def = EXTRACT_SUBREG v0.sub1, sub0. 1251 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def"); 1252 const MachineOperand &MOReg = MI.getOperand(1); 1253 if (MOReg.isUndef()) 1254 return false; 1255 const MachineOperand &MOSubIdx = MI.getOperand(2); 1256 assert(MOSubIdx.isImm() && 1257 "The subindex of the extract_subreg is not an immediate"); 1258 1259 InputReg.Reg = MOReg.getReg(); 1260 InputReg.SubReg = MOReg.getSubReg(); 1261 InputReg.SubIdx = (unsigned)MOSubIdx.getImm(); 1262 return true; 1263 } 1264 1265 bool TargetInstrInfo::getInsertSubregInputs( 1266 const MachineInstr &MI, unsigned DefIdx, 1267 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const { 1268 assert((MI.isInsertSubreg() || 1269 MI.isInsertSubregLike()) && "Instruction do not have the proper type"); 1270 1271 if (!MI.isInsertSubreg()) 1272 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg); 1273 1274 // We are looking at: 1275 // Def = INSERT_SEQUENCE v0, v1, sub0. 1276 assert(DefIdx == 0 && "INSERT_SUBREG only has one def"); 1277 const MachineOperand &MOBaseReg = MI.getOperand(1); 1278 const MachineOperand &MOInsertedReg = MI.getOperand(2); 1279 if (MOInsertedReg.isUndef()) 1280 return false; 1281 const MachineOperand &MOSubIdx = MI.getOperand(3); 1282 assert(MOSubIdx.isImm() && 1283 "One of the subindex of the reg_sequence is not an immediate"); 1284 BaseReg.Reg = MOBaseReg.getReg(); 1285 BaseReg.SubReg = MOBaseReg.getSubReg(); 1286 1287 InsertedReg.Reg = MOInsertedReg.getReg(); 1288 InsertedReg.SubReg = MOInsertedReg.getSubReg(); 1289 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm(); 1290 return true; 1291 } 1292 1293 TargetInstrInfo::PipelinerLoopInfo::~PipelinerLoopInfo() {} 1294