1 //===-- HexagonHardwareLoops.cpp - Identify and generate hardware loops ---===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This pass identifies loops where we can generate the Hexagon hardware 11 // loop instruction. The hardware loop can perform loop branches with a 12 // zero-cycle overhead. 13 // 14 // The pattern that defines the induction variable can changed depending on 15 // prior optimizations. For example, the IndVarSimplify phase run by 'opt' 16 // normalizes induction variables, and the Loop Strength Reduction pass 17 // run by 'llc' may also make changes to the induction variable. 18 // The pattern detected by this phase is due to running Strength Reduction. 19 // 20 // Criteria for hardware loops: 21 // - Countable loops (w/ ind. var for a trip count) 22 // - Assumes loops are normalized by IndVarSimplify 23 // - Try inner-most loops first 24 // - No nested hardware loops. 25 // - No function calls in loops. 26 // 27 //===----------------------------------------------------------------------===// 28 29 #include "llvm/ADT/SmallSet.h" 30 #include "Hexagon.h" 31 #include "HexagonTargetMachine.h" 32 #include "llvm/ADT/Statistic.h" 33 #include "llvm/CodeGen/MachineDominators.h" 34 #include "llvm/CodeGen/MachineFunction.h" 35 #include "llvm/CodeGen/MachineFunctionPass.h" 36 #include "llvm/CodeGen/MachineInstrBuilder.h" 37 #include "llvm/CodeGen/MachineLoopInfo.h" 38 #include "llvm/CodeGen/MachineRegisterInfo.h" 39 #include "llvm/PassSupport.h" 40 #include "llvm/Support/CommandLine.h" 41 #include "llvm/Support/Debug.h" 42 #include "llvm/Support/raw_ostream.h" 43 #include "llvm/Target/TargetInstrInfo.h" 44 #include <algorithm> 45 #include <vector> 46 47 using namespace llvm; 48 49 #define DEBUG_TYPE "hwloops" 50 51 #ifndef NDEBUG 52 static cl::opt<int> HWLoopLimit("max-hwloop", cl::Hidden, cl::init(-1)); 53 #endif 54 55 STATISTIC(NumHWLoops, "Number of loops converted to hardware loops"); 56 57 namespace llvm { 58 void initializeHexagonHardwareLoopsPass(PassRegistry&); 59 } 60 61 namespace { 62 class CountValue; 63 struct HexagonHardwareLoops : public MachineFunctionPass { 64 MachineLoopInfo *MLI; 65 MachineRegisterInfo *MRI; 66 MachineDominatorTree *MDT; 67 const HexagonTargetMachine *TM; 68 const HexagonInstrInfo *TII; 69 const HexagonRegisterInfo *TRI; 70 #ifndef NDEBUG 71 static int Counter; 72 #endif 73 74 public: 75 static char ID; 76 77 HexagonHardwareLoops() : MachineFunctionPass(ID) { 78 initializeHexagonHardwareLoopsPass(*PassRegistry::getPassRegistry()); 79 } 80 81 bool runOnMachineFunction(MachineFunction &MF) override; 82 83 const char *getPassName() const override { return "Hexagon Hardware Loops"; } 84 85 void getAnalysisUsage(AnalysisUsage &AU) const override { 86 AU.addRequired<MachineDominatorTree>(); 87 AU.addRequired<MachineLoopInfo>(); 88 MachineFunctionPass::getAnalysisUsage(AU); 89 } 90 91 private: 92 /// Kinds of comparisons in the compare instructions. 93 struct Comparison { 94 enum Kind { 95 EQ = 0x01, 96 NE = 0x02, 97 L = 0x04, // Less-than property. 98 G = 0x08, // Greater-than property. 99 U = 0x40, // Unsigned property. 100 LTs = L, 101 LEs = L | EQ, 102 GTs = G, 103 GEs = G | EQ, 104 LTu = L | U, 105 LEu = L | EQ | U, 106 GTu = G | U, 107 GEu = G | EQ | U 108 }; 109 110 static Kind getSwappedComparison(Kind Cmp) { 111 assert ((!((Cmp & L) && (Cmp & G))) && "Malformed comparison operator"); 112 if ((Cmp & L) || (Cmp & G)) 113 return (Kind)(Cmp ^ (L|G)); 114 return Cmp; 115 } 116 }; 117 118 /// \brief Find the register that contains the loop controlling 119 /// induction variable. 120 /// If successful, it will return true and set the \p Reg, \p IVBump 121 /// and \p IVOp arguments. Otherwise it will return false. 122 /// The returned induction register is the register R that follows the 123 /// following induction pattern: 124 /// loop: 125 /// R = phi ..., [ R.next, LatchBlock ] 126 /// R.next = R + #bump 127 /// if (R.next < #N) goto loop 128 /// IVBump is the immediate value added to R, and IVOp is the instruction 129 /// "R.next = R + #bump". 130 bool findInductionRegister(MachineLoop *L, unsigned &Reg, 131 int64_t &IVBump, MachineInstr *&IVOp) const; 132 133 /// \brief Analyze the statements in a loop to determine if the loop 134 /// has a computable trip count and, if so, return a value that represents 135 /// the trip count expression. 136 CountValue *getLoopTripCount(MachineLoop *L, 137 SmallVectorImpl<MachineInstr *> &OldInsts); 138 139 /// \brief Return the expression that represents the number of times 140 /// a loop iterates. The function takes the operands that represent the 141 /// loop start value, loop end value, and induction value. Based upon 142 /// these operands, the function attempts to compute the trip count. 143 /// If the trip count is not directly available (as an immediate value, 144 /// or a register), the function will attempt to insert computation of it 145 /// to the loop's preheader. 146 CountValue *computeCount(MachineLoop *Loop, 147 const MachineOperand *Start, 148 const MachineOperand *End, 149 unsigned IVReg, 150 int64_t IVBump, 151 Comparison::Kind Cmp) const; 152 153 /// \brief Return true if the instruction is not valid within a hardware 154 /// loop. 155 bool isInvalidLoopOperation(const MachineInstr *MI) const; 156 157 /// \brief Return true if the loop contains an instruction that inhibits 158 /// using the hardware loop. 159 bool containsInvalidInstruction(MachineLoop *L) const; 160 161 /// \brief Given a loop, check if we can convert it to a hardware loop. 162 /// If so, then perform the conversion and return true. 163 bool convertToHardwareLoop(MachineLoop *L); 164 165 /// \brief Return true if the instruction is now dead. 166 bool isDead(const MachineInstr *MI, 167 SmallVectorImpl<MachineInstr *> &DeadPhis) const; 168 169 /// \brief Remove the instruction if it is now dead. 170 void removeIfDead(MachineInstr *MI); 171 172 /// \brief Make sure that the "bump" instruction executes before the 173 /// compare. We need that for the IV fixup, so that the compare 174 /// instruction would not use a bumped value that has not yet been 175 /// defined. If the instructions are out of order, try to reorder them. 176 bool orderBumpCompare(MachineInstr *BumpI, MachineInstr *CmpI); 177 178 /// \brief Get the instruction that loads an immediate value into \p R, 179 /// or 0 if such an instruction does not exist. 180 MachineInstr *defWithImmediate(unsigned R); 181 182 /// \brief Get the immediate value referenced to by \p MO, either for 183 /// immediate operands, or for register operands, where the register 184 /// was defined with an immediate value. 185 int64_t getImmediate(MachineOperand &MO); 186 187 /// \brief Reset the given machine operand to now refer to a new immediate 188 /// value. Assumes that the operand was already referencing an immediate 189 /// value, either directly, or via a register. 190 void setImmediate(MachineOperand &MO, int64_t Val); 191 192 /// \brief Fix the data flow of the induction varible. 193 /// The desired flow is: phi ---> bump -+-> comparison-in-latch. 194 /// | 195 /// +-> back to phi 196 /// where "bump" is the increment of the induction variable: 197 /// iv = iv + #const. 198 /// Due to some prior code transformations, the actual flow may look 199 /// like this: 200 /// phi -+-> bump ---> back to phi 201 /// | 202 /// +-> comparison-in-latch (against upper_bound-bump), 203 /// i.e. the comparison that controls the loop execution may be using 204 /// the value of the induction variable from before the increment. 205 /// 206 /// Return true if the loop's flow is the desired one (i.e. it's 207 /// either been fixed, or no fixing was necessary). 208 /// Otherwise, return false. This can happen if the induction variable 209 /// couldn't be identified, or if the value in the latch's comparison 210 /// cannot be adjusted to reflect the post-bump value. 211 bool fixupInductionVariable(MachineLoop *L); 212 213 /// \brief Given a loop, if it does not have a preheader, create one. 214 /// Return the block that is the preheader. 215 MachineBasicBlock *createPreheaderForLoop(MachineLoop *L); 216 }; 217 218 char HexagonHardwareLoops::ID = 0; 219 #ifndef NDEBUG 220 int HexagonHardwareLoops::Counter = 0; 221 #endif 222 223 /// \brief Abstraction for a trip count of a loop. A smaller vesrsion 224 /// of the MachineOperand class without the concerns of changing the 225 /// operand representation. 226 class CountValue { 227 public: 228 enum CountValueType { 229 CV_Register, 230 CV_Immediate 231 }; 232 private: 233 CountValueType Kind; 234 union Values { 235 struct { 236 unsigned Reg; 237 unsigned Sub; 238 } R; 239 unsigned ImmVal; 240 } Contents; 241 242 public: 243 explicit CountValue(CountValueType t, unsigned v, unsigned u = 0) { 244 Kind = t; 245 if (Kind == CV_Register) { 246 Contents.R.Reg = v; 247 Contents.R.Sub = u; 248 } else { 249 Contents.ImmVal = v; 250 } 251 } 252 bool isReg() const { return Kind == CV_Register; } 253 bool isImm() const { return Kind == CV_Immediate; } 254 255 unsigned getReg() const { 256 assert(isReg() && "Wrong CountValue accessor"); 257 return Contents.R.Reg; 258 } 259 unsigned getSubReg() const { 260 assert(isReg() && "Wrong CountValue accessor"); 261 return Contents.R.Sub; 262 } 263 unsigned getImm() const { 264 assert(isImm() && "Wrong CountValue accessor"); 265 return Contents.ImmVal; 266 } 267 268 void print(raw_ostream &OS, const TargetMachine *TM = nullptr) const { 269 const TargetRegisterInfo *TRI = TM ? TM->getRegisterInfo() : nullptr; 270 if (isReg()) { OS << PrintReg(Contents.R.Reg, TRI, Contents.R.Sub); } 271 if (isImm()) { OS << Contents.ImmVal; } 272 } 273 }; 274 } // end anonymous namespace 275 276 277 INITIALIZE_PASS_BEGIN(HexagonHardwareLoops, "hwloops", 278 "Hexagon Hardware Loops", false, false) 279 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 280 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 281 INITIALIZE_PASS_END(HexagonHardwareLoops, "hwloops", 282 "Hexagon Hardware Loops", false, false) 283 284 285 /// \brief Returns true if the instruction is a hardware loop instruction. 286 static bool isHardwareLoop(const MachineInstr *MI) { 287 return MI->getOpcode() == Hexagon::LOOP0_r || 288 MI->getOpcode() == Hexagon::LOOP0_i; 289 } 290 291 FunctionPass *llvm::createHexagonHardwareLoops() { 292 return new HexagonHardwareLoops(); 293 } 294 295 296 bool HexagonHardwareLoops::runOnMachineFunction(MachineFunction &MF) { 297 DEBUG(dbgs() << "********* Hexagon Hardware Loops *********\n"); 298 299 bool Changed = false; 300 301 MLI = &getAnalysis<MachineLoopInfo>(); 302 MRI = &MF.getRegInfo(); 303 MDT = &getAnalysis<MachineDominatorTree>(); 304 TM = static_cast<const HexagonTargetMachine*>(&MF.getTarget()); 305 TII = static_cast<const HexagonInstrInfo*>(TM->getInstrInfo()); 306 TRI = static_cast<const HexagonRegisterInfo*>(TM->getRegisterInfo()); 307 308 for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); 309 I != E; ++I) { 310 MachineLoop *L = *I; 311 if (!L->getParentLoop()) 312 Changed |= convertToHardwareLoop(L); 313 } 314 315 return Changed; 316 } 317 318 319 bool HexagonHardwareLoops::findInductionRegister(MachineLoop *L, 320 unsigned &Reg, 321 int64_t &IVBump, 322 MachineInstr *&IVOp 323 ) const { 324 MachineBasicBlock *Header = L->getHeader(); 325 MachineBasicBlock *Preheader = L->getLoopPreheader(); 326 MachineBasicBlock *Latch = L->getLoopLatch(); 327 if (!Header || !Preheader || !Latch) 328 return false; 329 330 // This pair represents an induction register together with an immediate 331 // value that will be added to it in each loop iteration. 332 typedef std::pair<unsigned,int64_t> RegisterBump; 333 334 // Mapping: R.next -> (R, bump), where R, R.next and bump are derived 335 // from an induction operation 336 // R.next = R + bump 337 // where bump is an immediate value. 338 typedef std::map<unsigned,RegisterBump> InductionMap; 339 340 InductionMap IndMap; 341 342 typedef MachineBasicBlock::instr_iterator instr_iterator; 343 for (instr_iterator I = Header->instr_begin(), E = Header->instr_end(); 344 I != E && I->isPHI(); ++I) { 345 MachineInstr *Phi = &*I; 346 347 // Have a PHI instruction. Get the operand that corresponds to the 348 // latch block, and see if is a result of an addition of form "reg+imm", 349 // where the "reg" is defined by the PHI node we are looking at. 350 for (unsigned i = 1, n = Phi->getNumOperands(); i < n; i += 2) { 351 if (Phi->getOperand(i+1).getMBB() != Latch) 352 continue; 353 354 unsigned PhiOpReg = Phi->getOperand(i).getReg(); 355 MachineInstr *DI = MRI->getVRegDef(PhiOpReg); 356 unsigned UpdOpc = DI->getOpcode(); 357 bool isAdd = (UpdOpc == Hexagon::ADD_ri); 358 359 if (isAdd) { 360 // If the register operand to the add is the PHI we're 361 // looking at, this meets the induction pattern. 362 unsigned IndReg = DI->getOperand(1).getReg(); 363 if (MRI->getVRegDef(IndReg) == Phi) { 364 unsigned UpdReg = DI->getOperand(0).getReg(); 365 int64_t V = DI->getOperand(2).getImm(); 366 IndMap.insert(std::make_pair(UpdReg, std::make_pair(IndReg, V))); 367 } 368 } 369 } // for (i) 370 } // for (instr) 371 372 SmallVector<MachineOperand,2> Cond; 373 MachineBasicBlock *TB = nullptr, *FB = nullptr; 374 bool NotAnalyzed = TII->AnalyzeBranch(*Latch, TB, FB, Cond, false); 375 if (NotAnalyzed) 376 return false; 377 378 unsigned CSz = Cond.size(); 379 assert (CSz == 1 || CSz == 2); 380 unsigned PredR = Cond[CSz-1].getReg(); 381 382 MachineInstr *PredI = MRI->getVRegDef(PredR); 383 if (!PredI->isCompare()) 384 return false; 385 386 unsigned CmpReg1 = 0, CmpReg2 = 0; 387 int CmpImm = 0, CmpMask = 0; 388 bool CmpAnalyzed = TII->analyzeCompare(PredI, CmpReg1, CmpReg2, 389 CmpMask, CmpImm); 390 // Fail if the compare was not analyzed, or it's not comparing a register 391 // with an immediate value. Not checking the mask here, since we handle 392 // the individual compare opcodes (including CMPb) later on. 393 if (!CmpAnalyzed) 394 return false; 395 396 // Exactly one of the input registers to the comparison should be among 397 // the induction registers. 398 InductionMap::iterator IndMapEnd = IndMap.end(); 399 InductionMap::iterator F = IndMapEnd; 400 if (CmpReg1 != 0) { 401 InductionMap::iterator F1 = IndMap.find(CmpReg1); 402 if (F1 != IndMapEnd) 403 F = F1; 404 } 405 if (CmpReg2 != 0) { 406 InductionMap::iterator F2 = IndMap.find(CmpReg2); 407 if (F2 != IndMapEnd) { 408 if (F != IndMapEnd) 409 return false; 410 F = F2; 411 } 412 } 413 if (F == IndMapEnd) 414 return false; 415 416 Reg = F->second.first; 417 IVBump = F->second.second; 418 IVOp = MRI->getVRegDef(F->first); 419 return true; 420 } 421 422 423 /// \brief Analyze the statements in a loop to determine if the loop has 424 /// a computable trip count and, if so, return a value that represents 425 /// the trip count expression. 426 /// 427 /// This function iterates over the phi nodes in the loop to check for 428 /// induction variable patterns that are used in the calculation for 429 /// the number of time the loop is executed. 430 CountValue *HexagonHardwareLoops::getLoopTripCount(MachineLoop *L, 431 SmallVectorImpl<MachineInstr *> &OldInsts) { 432 MachineBasicBlock *TopMBB = L->getTopBlock(); 433 MachineBasicBlock::pred_iterator PI = TopMBB->pred_begin(); 434 assert(PI != TopMBB->pred_end() && 435 "Loop must have more than one incoming edge!"); 436 MachineBasicBlock *Backedge = *PI++; 437 if (PI == TopMBB->pred_end()) // dead loop? 438 return nullptr; 439 MachineBasicBlock *Incoming = *PI++; 440 if (PI != TopMBB->pred_end()) // multiple backedges? 441 return nullptr; 442 443 // Make sure there is one incoming and one backedge and determine which 444 // is which. 445 if (L->contains(Incoming)) { 446 if (L->contains(Backedge)) 447 return nullptr; 448 std::swap(Incoming, Backedge); 449 } else if (!L->contains(Backedge)) 450 return nullptr; 451 452 // Look for the cmp instruction to determine if we can get a useful trip 453 // count. The trip count can be either a register or an immediate. The 454 // location of the value depends upon the type (reg or imm). 455 MachineBasicBlock *Latch = L->getLoopLatch(); 456 if (!Latch) 457 return nullptr; 458 459 unsigned IVReg = 0; 460 int64_t IVBump = 0; 461 MachineInstr *IVOp; 462 bool FoundIV = findInductionRegister(L, IVReg, IVBump, IVOp); 463 if (!FoundIV) 464 return nullptr; 465 466 MachineBasicBlock *Preheader = L->getLoopPreheader(); 467 468 MachineOperand *InitialValue = nullptr; 469 MachineInstr *IV_Phi = MRI->getVRegDef(IVReg); 470 for (unsigned i = 1, n = IV_Phi->getNumOperands(); i < n; i += 2) { 471 MachineBasicBlock *MBB = IV_Phi->getOperand(i+1).getMBB(); 472 if (MBB == Preheader) 473 InitialValue = &IV_Phi->getOperand(i); 474 else if (MBB == Latch) 475 IVReg = IV_Phi->getOperand(i).getReg(); // Want IV reg after bump. 476 } 477 if (!InitialValue) 478 return nullptr; 479 480 SmallVector<MachineOperand,2> Cond; 481 MachineBasicBlock *TB = nullptr, *FB = nullptr; 482 bool NotAnalyzed = TII->AnalyzeBranch(*Latch, TB, FB, Cond, false); 483 if (NotAnalyzed) 484 return nullptr; 485 486 MachineBasicBlock *Header = L->getHeader(); 487 // TB must be non-null. If FB is also non-null, one of them must be 488 // the header. Otherwise, branch to TB could be exiting the loop, and 489 // the fall through can go to the header. 490 assert (TB && "Latch block without a branch?"); 491 assert ((!FB || TB == Header || FB == Header) && "Branches not to header?"); 492 if (!TB || (FB && TB != Header && FB != Header)) 493 return nullptr; 494 495 // Branches of form "if (!P) ..." cause HexagonInstrInfo::AnalyzeBranch 496 // to put imm(0), followed by P in the vector Cond. 497 // If TB is not the header, it means that the "not-taken" path must lead 498 // to the header. 499 bool Negated = (Cond.size() > 1) ^ (TB != Header); 500 unsigned PredReg = Cond[Cond.size()-1].getReg(); 501 MachineInstr *CondI = MRI->getVRegDef(PredReg); 502 unsigned CondOpc = CondI->getOpcode(); 503 504 unsigned CmpReg1 = 0, CmpReg2 = 0; 505 int Mask = 0, ImmValue = 0; 506 bool AnalyzedCmp = TII->analyzeCompare(CondI, CmpReg1, CmpReg2, 507 Mask, ImmValue); 508 if (!AnalyzedCmp) 509 return nullptr; 510 511 // The comparison operator type determines how we compute the loop 512 // trip count. 513 OldInsts.push_back(CondI); 514 OldInsts.push_back(IVOp); 515 516 // Sadly, the following code gets information based on the position 517 // of the operands in the compare instruction. This has to be done 518 // this way, because the comparisons check for a specific relationship 519 // between the operands (e.g. is-less-than), rather than to find out 520 // what relationship the operands are in (as on PPC). 521 Comparison::Kind Cmp; 522 bool isSwapped = false; 523 const MachineOperand &Op1 = CondI->getOperand(1); 524 const MachineOperand &Op2 = CondI->getOperand(2); 525 const MachineOperand *EndValue = nullptr; 526 527 if (Op1.isReg()) { 528 if (Op2.isImm() || Op1.getReg() == IVReg) 529 EndValue = &Op2; 530 else { 531 EndValue = &Op1; 532 isSwapped = true; 533 } 534 } 535 536 if (!EndValue) 537 return nullptr; 538 539 switch (CondOpc) { 540 case Hexagon::CMPEQri: 541 case Hexagon::CMPEQrr: 542 Cmp = !Negated ? Comparison::EQ : Comparison::NE; 543 break; 544 case Hexagon::CMPGTUri: 545 case Hexagon::CMPGTUrr: 546 Cmp = !Negated ? Comparison::GTu : Comparison::LEu; 547 break; 548 case Hexagon::CMPGTri: 549 case Hexagon::CMPGTrr: 550 Cmp = !Negated ? Comparison::GTs : Comparison::LEs; 551 break; 552 // Very limited support for byte/halfword compares. 553 case Hexagon::CMPbEQri_V4: 554 case Hexagon::CMPhEQri_V4: { 555 if (IVBump != 1) 556 return nullptr; 557 558 int64_t InitV, EndV; 559 // Since the comparisons are "ri", the EndValue should be an 560 // immediate. Check it just in case. 561 assert(EndValue->isImm() && "Unrecognized latch comparison"); 562 EndV = EndValue->getImm(); 563 // Allow InitialValue to be a register defined with an immediate. 564 if (InitialValue->isReg()) { 565 if (!defWithImmediate(InitialValue->getReg())) 566 return nullptr; 567 InitV = getImmediate(*InitialValue); 568 } else { 569 assert(InitialValue->isImm()); 570 InitV = InitialValue->getImm(); 571 } 572 if (InitV >= EndV) 573 return nullptr; 574 if (CondOpc == Hexagon::CMPbEQri_V4) { 575 if (!isInt<8>(InitV) || !isInt<8>(EndV)) 576 return nullptr; 577 } else { // Hexagon::CMPhEQri_V4 578 if (!isInt<16>(InitV) || !isInt<16>(EndV)) 579 return nullptr; 580 } 581 Cmp = !Negated ? Comparison::EQ : Comparison::NE; 582 break; 583 } 584 default: 585 return nullptr; 586 } 587 588 if (isSwapped) 589 Cmp = Comparison::getSwappedComparison(Cmp); 590 591 if (InitialValue->isReg()) { 592 unsigned R = InitialValue->getReg(); 593 MachineBasicBlock *DefBB = MRI->getVRegDef(R)->getParent(); 594 if (!MDT->properlyDominates(DefBB, Header)) 595 return nullptr; 596 OldInsts.push_back(MRI->getVRegDef(R)); 597 } 598 if (EndValue->isReg()) { 599 unsigned R = EndValue->getReg(); 600 MachineBasicBlock *DefBB = MRI->getVRegDef(R)->getParent(); 601 if (!MDT->properlyDominates(DefBB, Header)) 602 return nullptr; 603 } 604 605 return computeCount(L, InitialValue, EndValue, IVReg, IVBump, Cmp); 606 } 607 608 /// \brief Helper function that returns the expression that represents the 609 /// number of times a loop iterates. The function takes the operands that 610 /// represent the loop start value, loop end value, and induction value. 611 /// Based upon these operands, the function attempts to compute the trip count. 612 CountValue *HexagonHardwareLoops::computeCount(MachineLoop *Loop, 613 const MachineOperand *Start, 614 const MachineOperand *End, 615 unsigned IVReg, 616 int64_t IVBump, 617 Comparison::Kind Cmp) const { 618 // Cannot handle comparison EQ, i.e. while (A == B). 619 if (Cmp == Comparison::EQ) 620 return nullptr; 621 622 // Check if either the start or end values are an assignment of an immediate. 623 // If so, use the immediate value rather than the register. 624 if (Start->isReg()) { 625 const MachineInstr *StartValInstr = MRI->getVRegDef(Start->getReg()); 626 if (StartValInstr && StartValInstr->getOpcode() == Hexagon::TFRI) 627 Start = &StartValInstr->getOperand(1); 628 } 629 if (End->isReg()) { 630 const MachineInstr *EndValInstr = MRI->getVRegDef(End->getReg()); 631 if (EndValInstr && EndValInstr->getOpcode() == Hexagon::TFRI) 632 End = &EndValInstr->getOperand(1); 633 } 634 635 assert (Start->isReg() || Start->isImm()); 636 assert (End->isReg() || End->isImm()); 637 638 bool CmpLess = Cmp & Comparison::L; 639 bool CmpGreater = Cmp & Comparison::G; 640 bool CmpHasEqual = Cmp & Comparison::EQ; 641 642 // Avoid certain wrap-arounds. This doesn't detect all wrap-arounds. 643 // If loop executes while iv is "less" with the iv value going down, then 644 // the iv must wrap. 645 if (CmpLess && IVBump < 0) 646 return nullptr; 647 // If loop executes while iv is "greater" with the iv value going up, then 648 // the iv must wrap. 649 if (CmpGreater && IVBump > 0) 650 return nullptr; 651 652 if (Start->isImm() && End->isImm()) { 653 // Both, start and end are immediates. 654 int64_t StartV = Start->getImm(); 655 int64_t EndV = End->getImm(); 656 int64_t Dist = EndV - StartV; 657 if (Dist == 0) 658 return nullptr; 659 660 bool Exact = (Dist % IVBump) == 0; 661 662 if (Cmp == Comparison::NE) { 663 if (!Exact) 664 return nullptr; 665 if ((Dist < 0) ^ (IVBump < 0)) 666 return nullptr; 667 } 668 669 // For comparisons that include the final value (i.e. include equality 670 // with the final value), we need to increase the distance by 1. 671 if (CmpHasEqual) 672 Dist = Dist > 0 ? Dist+1 : Dist-1; 673 674 // assert (CmpLess => Dist > 0); 675 assert ((!CmpLess || Dist > 0) && "Loop should never iterate!"); 676 // assert (CmpGreater => Dist < 0); 677 assert ((!CmpGreater || Dist < 0) && "Loop should never iterate!"); 678 679 // "Normalized" distance, i.e. with the bump set to +-1. 680 int64_t Dist1 = (IVBump > 0) ? (Dist + (IVBump-1)) / IVBump 681 : (-Dist + (-IVBump-1)) / (-IVBump); 682 assert (Dist1 > 0 && "Fishy thing. Both operands have the same sign."); 683 684 uint64_t Count = Dist1; 685 686 if (Count > 0xFFFFFFFFULL) 687 return nullptr; 688 689 return new CountValue(CountValue::CV_Immediate, Count); 690 } 691 692 // A general case: Start and End are some values, but the actual 693 // iteration count may not be available. If it is not, insert 694 // a computation of it into the preheader. 695 696 // If the induction variable bump is not a power of 2, quit. 697 // Othwerise we'd need a general integer division. 698 if (!isPowerOf2_64(abs64(IVBump))) 699 return nullptr; 700 701 MachineBasicBlock *PH = Loop->getLoopPreheader(); 702 assert (PH && "Should have a preheader by now"); 703 MachineBasicBlock::iterator InsertPos = PH->getFirstTerminator(); 704 DebugLoc DL = (InsertPos != PH->end()) ? InsertPos->getDebugLoc() 705 : DebugLoc(); 706 707 // If Start is an immediate and End is a register, the trip count 708 // will be "reg - imm". Hexagon's "subtract immediate" instruction 709 // is actually "reg + -imm". 710 711 // If the loop IV is going downwards, i.e. if the bump is negative, 712 // then the iteration count (computed as End-Start) will need to be 713 // negated. To avoid the negation, just swap Start and End. 714 if (IVBump < 0) { 715 std::swap(Start, End); 716 IVBump = -IVBump; 717 } 718 // Cmp may now have a wrong direction, e.g. LEs may now be GEs. 719 // Signedness, and "including equality" are preserved. 720 721 bool RegToImm = Start->isReg() && End->isImm(); // for (reg..imm) 722 bool RegToReg = Start->isReg() && End->isReg(); // for (reg..reg) 723 724 int64_t StartV = 0, EndV = 0; 725 if (Start->isImm()) 726 StartV = Start->getImm(); 727 if (End->isImm()) 728 EndV = End->getImm(); 729 730 int64_t AdjV = 0; 731 // To compute the iteration count, we would need this computation: 732 // Count = (End - Start + (IVBump-1)) / IVBump 733 // or, when CmpHasEqual: 734 // Count = (End - Start + (IVBump-1)+1) / IVBump 735 // The "IVBump-1" part is the adjustment (AdjV). We can avoid 736 // generating an instruction specifically to add it if we can adjust 737 // the immediate values for Start or End. 738 739 if (CmpHasEqual) { 740 // Need to add 1 to the total iteration count. 741 if (Start->isImm()) 742 StartV--; 743 else if (End->isImm()) 744 EndV++; 745 else 746 AdjV += 1; 747 } 748 749 if (Cmp != Comparison::NE) { 750 if (Start->isImm()) 751 StartV -= (IVBump-1); 752 else if (End->isImm()) 753 EndV += (IVBump-1); 754 else 755 AdjV += (IVBump-1); 756 } 757 758 unsigned R = 0, SR = 0; 759 if (Start->isReg()) { 760 R = Start->getReg(); 761 SR = Start->getSubReg(); 762 } else { 763 R = End->getReg(); 764 SR = End->getSubReg(); 765 } 766 const TargetRegisterClass *RC = MRI->getRegClass(R); 767 // Hardware loops cannot handle 64-bit registers. If it's a double 768 // register, it has to have a subregister. 769 if (!SR && RC == &Hexagon::DoubleRegsRegClass) 770 return nullptr; 771 const TargetRegisterClass *IntRC = &Hexagon::IntRegsRegClass; 772 773 // Compute DistR (register with the distance between Start and End). 774 unsigned DistR, DistSR; 775 776 // Avoid special case, where the start value is an imm(0). 777 if (Start->isImm() && StartV == 0) { 778 DistR = End->getReg(); 779 DistSR = End->getSubReg(); 780 } else { 781 const MCInstrDesc &SubD = RegToReg ? TII->get(Hexagon::SUB_rr) : 782 (RegToImm ? TII->get(Hexagon::SUB_ri) : 783 TII->get(Hexagon::ADD_ri)); 784 unsigned SubR = MRI->createVirtualRegister(IntRC); 785 MachineInstrBuilder SubIB = 786 BuildMI(*PH, InsertPos, DL, SubD, SubR); 787 788 if (RegToReg) { 789 SubIB.addReg(End->getReg(), 0, End->getSubReg()) 790 .addReg(Start->getReg(), 0, Start->getSubReg()); 791 } else if (RegToImm) { 792 SubIB.addImm(EndV) 793 .addReg(Start->getReg(), 0, Start->getSubReg()); 794 } else { // ImmToReg 795 SubIB.addReg(End->getReg(), 0, End->getSubReg()) 796 .addImm(-StartV); 797 } 798 DistR = SubR; 799 DistSR = 0; 800 } 801 802 // From DistR, compute AdjR (register with the adjusted distance). 803 unsigned AdjR, AdjSR; 804 805 if (AdjV == 0) { 806 AdjR = DistR; 807 AdjSR = DistSR; 808 } else { 809 // Generate CountR = ADD DistR, AdjVal 810 unsigned AddR = MRI->createVirtualRegister(IntRC); 811 const MCInstrDesc &AddD = TII->get(Hexagon::ADD_ri); 812 BuildMI(*PH, InsertPos, DL, AddD, AddR) 813 .addReg(DistR, 0, DistSR) 814 .addImm(AdjV); 815 816 AdjR = AddR; 817 AdjSR = 0; 818 } 819 820 // From AdjR, compute CountR (register with the final count). 821 unsigned CountR, CountSR; 822 823 if (IVBump == 1) { 824 CountR = AdjR; 825 CountSR = AdjSR; 826 } else { 827 // The IV bump is a power of two. Log_2(IV bump) is the shift amount. 828 unsigned Shift = Log2_32(IVBump); 829 830 // Generate NormR = LSR DistR, Shift. 831 unsigned LsrR = MRI->createVirtualRegister(IntRC); 832 const MCInstrDesc &LsrD = TII->get(Hexagon::LSR_ri); 833 BuildMI(*PH, InsertPos, DL, LsrD, LsrR) 834 .addReg(AdjR, 0, AdjSR) 835 .addImm(Shift); 836 837 CountR = LsrR; 838 CountSR = 0; 839 } 840 841 return new CountValue(CountValue::CV_Register, CountR, CountSR); 842 } 843 844 845 /// \brief Return true if the operation is invalid within hardware loop. 846 bool HexagonHardwareLoops::isInvalidLoopOperation( 847 const MachineInstr *MI) const { 848 849 // call is not allowed because the callee may use a hardware loop 850 if (MI->getDesc().isCall()) 851 return true; 852 853 // do not allow nested hardware loops 854 if (isHardwareLoop(MI)) 855 return true; 856 857 // check if the instruction defines a hardware loop register 858 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 859 const MachineOperand &MO = MI->getOperand(i); 860 if (!MO.isReg() || !MO.isDef()) 861 continue; 862 unsigned R = MO.getReg(); 863 if (R == Hexagon::LC0 || R == Hexagon::LC1 || 864 R == Hexagon::SA0 || R == Hexagon::SA1) 865 return true; 866 } 867 return false; 868 } 869 870 871 /// \brief - Return true if the loop contains an instruction that inhibits 872 /// the use of the hardware loop function. 873 bool HexagonHardwareLoops::containsInvalidInstruction(MachineLoop *L) const { 874 const std::vector<MachineBasicBlock *> &Blocks = L->getBlocks(); 875 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 876 MachineBasicBlock *MBB = Blocks[i]; 877 for (MachineBasicBlock::iterator 878 MII = MBB->begin(), E = MBB->end(); MII != E; ++MII) { 879 const MachineInstr *MI = &*MII; 880 if (isInvalidLoopOperation(MI)) 881 return true; 882 } 883 } 884 return false; 885 } 886 887 888 /// \brief Returns true if the instruction is dead. This was essentially 889 /// copied from DeadMachineInstructionElim::isDead, but with special cases 890 /// for inline asm, physical registers and instructions with side effects 891 /// removed. 892 bool HexagonHardwareLoops::isDead(const MachineInstr *MI, 893 SmallVectorImpl<MachineInstr *> &DeadPhis) const { 894 // Examine each operand. 895 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 896 const MachineOperand &MO = MI->getOperand(i); 897 if (!MO.isReg() || !MO.isDef()) 898 continue; 899 900 unsigned Reg = MO.getReg(); 901 if (MRI->use_nodbg_empty(Reg)) 902 continue; 903 904 typedef MachineRegisterInfo::use_nodbg_iterator use_nodbg_iterator; 905 906 // This instruction has users, but if the only user is the phi node for the 907 // parent block, and the only use of that phi node is this instruction, then 908 // this instruction is dead: both it (and the phi node) can be removed. 909 use_nodbg_iterator I = MRI->use_nodbg_begin(Reg); 910 use_nodbg_iterator End = MRI->use_nodbg_end(); 911 if (std::next(I) != End || !I->getParent()->isPHI()) 912 return false; 913 914 MachineInstr *OnePhi = I->getParent(); 915 for (unsigned j = 0, f = OnePhi->getNumOperands(); j != f; ++j) { 916 const MachineOperand &OPO = OnePhi->getOperand(j); 917 if (!OPO.isReg() || !OPO.isDef()) 918 continue; 919 920 unsigned OPReg = OPO.getReg(); 921 use_nodbg_iterator nextJ; 922 for (use_nodbg_iterator J = MRI->use_nodbg_begin(OPReg); 923 J != End; J = nextJ) { 924 nextJ = std::next(J); 925 MachineOperand &Use = *J; 926 MachineInstr *UseMI = Use.getParent(); 927 928 // If the phi node has a user that is not MI, bail... 929 if (MI != UseMI) 930 return false; 931 } 932 } 933 DeadPhis.push_back(OnePhi); 934 } 935 936 // If there are no defs with uses, the instruction is dead. 937 return true; 938 } 939 940 void HexagonHardwareLoops::removeIfDead(MachineInstr *MI) { 941 // This procedure was essentially copied from DeadMachineInstructionElim. 942 943 SmallVector<MachineInstr*, 1> DeadPhis; 944 if (isDead(MI, DeadPhis)) { 945 DEBUG(dbgs() << "HW looping will remove: " << *MI); 946 947 // It is possible that some DBG_VALUE instructions refer to this 948 // instruction. Examine each def operand for such references; 949 // if found, mark the DBG_VALUE as undef (but don't delete it). 950 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 951 const MachineOperand &MO = MI->getOperand(i); 952 if (!MO.isReg() || !MO.isDef()) 953 continue; 954 unsigned Reg = MO.getReg(); 955 MachineRegisterInfo::use_iterator nextI; 956 for (MachineRegisterInfo::use_iterator I = MRI->use_begin(Reg), 957 E = MRI->use_end(); I != E; I = nextI) { 958 nextI = std::next(I); // I is invalidated by the setReg 959 MachineOperand &Use = *I; 960 MachineInstr *UseMI = I->getParent(); 961 if (UseMI == MI) 962 continue; 963 if (Use.isDebug()) 964 UseMI->getOperand(0).setReg(0U); 965 // This may also be a "instr -> phi -> instr" case which can 966 // be removed too. 967 } 968 } 969 970 MI->eraseFromParent(); 971 for (unsigned i = 0; i < DeadPhis.size(); ++i) 972 DeadPhis[i]->eraseFromParent(); 973 } 974 } 975 976 /// \brief Check if the loop is a candidate for converting to a hardware 977 /// loop. If so, then perform the transformation. 978 /// 979 /// This function works on innermost loops first. A loop can be converted 980 /// if it is a counting loop; either a register value or an immediate. 981 /// 982 /// The code makes several assumptions about the representation of the loop 983 /// in llvm. 984 bool HexagonHardwareLoops::convertToHardwareLoop(MachineLoop *L) { 985 // This is just for sanity. 986 assert(L->getHeader() && "Loop without a header?"); 987 988 bool Changed = false; 989 // Process nested loops first. 990 for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I) 991 Changed |= convertToHardwareLoop(*I); 992 993 // If a nested loop has been converted, then we can't convert this loop. 994 if (Changed) 995 return Changed; 996 997 #ifndef NDEBUG 998 // Stop trying after reaching the limit (if any). 999 int Limit = HWLoopLimit; 1000 if (Limit >= 0) { 1001 if (Counter >= HWLoopLimit) 1002 return false; 1003 Counter++; 1004 } 1005 #endif 1006 1007 // Does the loop contain any invalid instructions? 1008 if (containsInvalidInstruction(L)) 1009 return false; 1010 1011 // Is the induction variable bump feeding the latch condition? 1012 if (!fixupInductionVariable(L)) 1013 return false; 1014 1015 MachineBasicBlock *LastMBB = L->getExitingBlock(); 1016 // Don't generate hw loop if the loop has more than one exit. 1017 if (!LastMBB) 1018 return false; 1019 1020 MachineBasicBlock::iterator LastI = LastMBB->getFirstTerminator(); 1021 if (LastI == LastMBB->end()) 1022 return false; 1023 1024 // Ensure the loop has a preheader: the loop instruction will be 1025 // placed there. 1026 bool NewPreheader = false; 1027 MachineBasicBlock *Preheader = L->getLoopPreheader(); 1028 if (!Preheader) { 1029 Preheader = createPreheaderForLoop(L); 1030 if (!Preheader) 1031 return false; 1032 NewPreheader = true; 1033 } 1034 MachineBasicBlock::iterator InsertPos = Preheader->getFirstTerminator(); 1035 1036 SmallVector<MachineInstr*, 2> OldInsts; 1037 // Are we able to determine the trip count for the loop? 1038 CountValue *TripCount = getLoopTripCount(L, OldInsts); 1039 if (!TripCount) 1040 return false; 1041 1042 // Is the trip count available in the preheader? 1043 if (TripCount->isReg()) { 1044 // There will be a use of the register inserted into the preheader, 1045 // so make sure that the register is actually defined at that point. 1046 MachineInstr *TCDef = MRI->getVRegDef(TripCount->getReg()); 1047 MachineBasicBlock *BBDef = TCDef->getParent(); 1048 if (!NewPreheader) { 1049 if (!MDT->dominates(BBDef, Preheader)) 1050 return false; 1051 } else { 1052 // If we have just created a preheader, the dominator tree won't be 1053 // aware of it. Check if the definition of the register dominates 1054 // the header, but is not the header itself. 1055 if (!MDT->properlyDominates(BBDef, L->getHeader())) 1056 return false; 1057 } 1058 } 1059 1060 // Determine the loop start. 1061 MachineBasicBlock *LoopStart = L->getTopBlock(); 1062 if (L->getLoopLatch() != LastMBB) { 1063 // When the exit and latch are not the same, use the latch block as the 1064 // start. 1065 // The loop start address is used only after the 1st iteration, and the 1066 // loop latch may contains instrs. that need to be executed after the 1067 // first iteration. 1068 LoopStart = L->getLoopLatch(); 1069 // Make sure the latch is a successor of the exit, otherwise it won't work. 1070 if (!LastMBB->isSuccessor(LoopStart)) 1071 return false; 1072 } 1073 1074 // Convert the loop to a hardware loop. 1075 DEBUG(dbgs() << "Change to hardware loop at "; L->dump()); 1076 DebugLoc DL; 1077 if (InsertPos != Preheader->end()) 1078 DL = InsertPos->getDebugLoc(); 1079 1080 if (TripCount->isReg()) { 1081 // Create a copy of the loop count register. 1082 unsigned CountReg = MRI->createVirtualRegister(&Hexagon::IntRegsRegClass); 1083 BuildMI(*Preheader, InsertPos, DL, TII->get(TargetOpcode::COPY), CountReg) 1084 .addReg(TripCount->getReg(), 0, TripCount->getSubReg()); 1085 // Add the Loop instruction to the beginning of the loop. 1086 BuildMI(*Preheader, InsertPos, DL, TII->get(Hexagon::LOOP0_r)) 1087 .addMBB(LoopStart) 1088 .addReg(CountReg); 1089 } else { 1090 assert(TripCount->isImm() && "Expecting immediate value for trip count"); 1091 // Add the Loop immediate instruction to the beginning of the loop, 1092 // if the immediate fits in the instructions. Otherwise, we need to 1093 // create a new virtual register. 1094 int64_t CountImm = TripCount->getImm(); 1095 if (!TII->isValidOffset(Hexagon::LOOP0_i, CountImm)) { 1096 unsigned CountReg = MRI->createVirtualRegister(&Hexagon::IntRegsRegClass); 1097 BuildMI(*Preheader, InsertPos, DL, TII->get(Hexagon::TFRI), CountReg) 1098 .addImm(CountImm); 1099 BuildMI(*Preheader, InsertPos, DL, TII->get(Hexagon::LOOP0_r)) 1100 .addMBB(LoopStart).addReg(CountReg); 1101 } else 1102 BuildMI(*Preheader, InsertPos, DL, TII->get(Hexagon::LOOP0_i)) 1103 .addMBB(LoopStart).addImm(CountImm); 1104 } 1105 1106 // Make sure the loop start always has a reference in the CFG. We need 1107 // to create a BlockAddress operand to get this mechanism to work both the 1108 // MachineBasicBlock and BasicBlock objects need the flag set. 1109 LoopStart->setHasAddressTaken(); 1110 // This line is needed to set the hasAddressTaken flag on the BasicBlock 1111 // object. 1112 BlockAddress::get(const_cast<BasicBlock *>(LoopStart->getBasicBlock())); 1113 1114 // Replace the loop branch with an endloop instruction. 1115 DebugLoc LastIDL = LastI->getDebugLoc(); 1116 BuildMI(*LastMBB, LastI, LastIDL, 1117 TII->get(Hexagon::ENDLOOP0)).addMBB(LoopStart); 1118 1119 // The loop ends with either: 1120 // - a conditional branch followed by an unconditional branch, or 1121 // - a conditional branch to the loop start. 1122 if (LastI->getOpcode() == Hexagon::JMP_t || 1123 LastI->getOpcode() == Hexagon::JMP_f) { 1124 // Delete one and change/add an uncond. branch to out of the loop. 1125 MachineBasicBlock *BranchTarget = LastI->getOperand(1).getMBB(); 1126 LastI = LastMBB->erase(LastI); 1127 if (!L->contains(BranchTarget)) { 1128 if (LastI != LastMBB->end()) 1129 LastI = LastMBB->erase(LastI); 1130 SmallVector<MachineOperand, 0> Cond; 1131 TII->InsertBranch(*LastMBB, BranchTarget, nullptr, Cond, LastIDL); 1132 } 1133 } else { 1134 // Conditional branch to loop start; just delete it. 1135 LastMBB->erase(LastI); 1136 } 1137 delete TripCount; 1138 1139 // The induction operation and the comparison may now be 1140 // unneeded. If these are unneeded, then remove them. 1141 for (unsigned i = 0; i < OldInsts.size(); ++i) 1142 removeIfDead(OldInsts[i]); 1143 1144 ++NumHWLoops; 1145 return true; 1146 } 1147 1148 1149 bool HexagonHardwareLoops::orderBumpCompare(MachineInstr *BumpI, 1150 MachineInstr *CmpI) { 1151 assert (BumpI != CmpI && "Bump and compare in the same instruction?"); 1152 1153 MachineBasicBlock *BB = BumpI->getParent(); 1154 if (CmpI->getParent() != BB) 1155 return false; 1156 1157 typedef MachineBasicBlock::instr_iterator instr_iterator; 1158 // Check if things are in order to begin with. 1159 for (instr_iterator I = BumpI, E = BB->instr_end(); I != E; ++I) 1160 if (&*I == CmpI) 1161 return true; 1162 1163 // Out of order. 1164 unsigned PredR = CmpI->getOperand(0).getReg(); 1165 bool FoundBump = false; 1166 instr_iterator CmpIt = CmpI, NextIt = std::next(CmpIt); 1167 for (instr_iterator I = NextIt, E = BB->instr_end(); I != E; ++I) { 1168 MachineInstr *In = &*I; 1169 for (unsigned i = 0, n = In->getNumOperands(); i < n; ++i) { 1170 MachineOperand &MO = In->getOperand(i); 1171 if (MO.isReg() && MO.isUse()) { 1172 if (MO.getReg() == PredR) // Found an intervening use of PredR. 1173 return false; 1174 } 1175 } 1176 1177 if (In == BumpI) { 1178 instr_iterator After = BumpI; 1179 instr_iterator From = CmpI; 1180 BB->splice(std::next(After), BB, From); 1181 FoundBump = true; 1182 break; 1183 } 1184 } 1185 assert (FoundBump && "Cannot determine instruction order"); 1186 return FoundBump; 1187 } 1188 1189 1190 MachineInstr *HexagonHardwareLoops::defWithImmediate(unsigned R) { 1191 MachineInstr *DI = MRI->getVRegDef(R); 1192 unsigned DOpc = DI->getOpcode(); 1193 switch (DOpc) { 1194 case Hexagon::TFRI: 1195 case Hexagon::TFRI64: 1196 case Hexagon::CONST32_Int_Real: 1197 case Hexagon::CONST64_Int_Real: 1198 return DI; 1199 } 1200 return nullptr; 1201 } 1202 1203 1204 int64_t HexagonHardwareLoops::getImmediate(MachineOperand &MO) { 1205 if (MO.isImm()) 1206 return MO.getImm(); 1207 assert(MO.isReg()); 1208 unsigned R = MO.getReg(); 1209 MachineInstr *DI = defWithImmediate(R); 1210 assert(DI && "Need an immediate operand"); 1211 // All currently supported "define-with-immediate" instructions have the 1212 // actual immediate value in the operand(1). 1213 int64_t v = DI->getOperand(1).getImm(); 1214 return v; 1215 } 1216 1217 1218 void HexagonHardwareLoops::setImmediate(MachineOperand &MO, int64_t Val) { 1219 if (MO.isImm()) { 1220 MO.setImm(Val); 1221 return; 1222 } 1223 1224 assert(MO.isReg()); 1225 unsigned R = MO.getReg(); 1226 MachineInstr *DI = defWithImmediate(R); 1227 if (MRI->hasOneNonDBGUse(R)) { 1228 // If R has only one use, then just change its defining instruction to 1229 // the new immediate value. 1230 DI->getOperand(1).setImm(Val); 1231 return; 1232 } 1233 1234 const TargetRegisterClass *RC = MRI->getRegClass(R); 1235 unsigned NewR = MRI->createVirtualRegister(RC); 1236 MachineBasicBlock &B = *DI->getParent(); 1237 DebugLoc DL = DI->getDebugLoc(); 1238 BuildMI(B, DI, DL, TII->get(DI->getOpcode()), NewR) 1239 .addImm(Val); 1240 MO.setReg(NewR); 1241 } 1242 1243 1244 bool HexagonHardwareLoops::fixupInductionVariable(MachineLoop *L) { 1245 MachineBasicBlock *Header = L->getHeader(); 1246 MachineBasicBlock *Preheader = L->getLoopPreheader(); 1247 MachineBasicBlock *Latch = L->getLoopLatch(); 1248 1249 if (!Header || !Preheader || !Latch) 1250 return false; 1251 1252 // These data structures follow the same concept as the corresponding 1253 // ones in findInductionRegister (where some comments are). 1254 typedef std::pair<unsigned,int64_t> RegisterBump; 1255 typedef std::pair<unsigned,RegisterBump> RegisterInduction; 1256 typedef std::set<RegisterInduction> RegisterInductionSet; 1257 1258 // Register candidates for induction variables, with their associated bumps. 1259 RegisterInductionSet IndRegs; 1260 1261 // Look for induction patterns: 1262 // vreg1 = PHI ..., [ latch, vreg2 ] 1263 // vreg2 = ADD vreg1, imm 1264 typedef MachineBasicBlock::instr_iterator instr_iterator; 1265 for (instr_iterator I = Header->instr_begin(), E = Header->instr_end(); 1266 I != E && I->isPHI(); ++I) { 1267 MachineInstr *Phi = &*I; 1268 1269 // Have a PHI instruction. 1270 for (unsigned i = 1, n = Phi->getNumOperands(); i < n; i += 2) { 1271 if (Phi->getOperand(i+1).getMBB() != Latch) 1272 continue; 1273 1274 unsigned PhiReg = Phi->getOperand(i).getReg(); 1275 MachineInstr *DI = MRI->getVRegDef(PhiReg); 1276 unsigned UpdOpc = DI->getOpcode(); 1277 bool isAdd = (UpdOpc == Hexagon::ADD_ri); 1278 1279 if (isAdd) { 1280 // If the register operand to the add/sub is the PHI we are looking 1281 // at, this meets the induction pattern. 1282 unsigned IndReg = DI->getOperand(1).getReg(); 1283 if (MRI->getVRegDef(IndReg) == Phi) { 1284 unsigned UpdReg = DI->getOperand(0).getReg(); 1285 int64_t V = DI->getOperand(2).getImm(); 1286 IndRegs.insert(std::make_pair(UpdReg, std::make_pair(IndReg, V))); 1287 } 1288 } 1289 } // for (i) 1290 } // for (instr) 1291 1292 if (IndRegs.empty()) 1293 return false; 1294 1295 MachineBasicBlock *TB = nullptr, *FB = nullptr; 1296 SmallVector<MachineOperand,2> Cond; 1297 // AnalyzeBranch returns true if it fails to analyze branch. 1298 bool NotAnalyzed = TII->AnalyzeBranch(*Latch, TB, FB, Cond, false); 1299 if (NotAnalyzed) 1300 return false; 1301 1302 // Check if the latch branch is unconditional. 1303 if (Cond.empty()) 1304 return false; 1305 1306 if (TB != Header && FB != Header) 1307 // The latch does not go back to the header. Not a latch we know and love. 1308 return false; 1309 1310 // Expecting a predicate register as a condition. It won't be a hardware 1311 // predicate register at this point yet, just a vreg. 1312 // HexagonInstrInfo::AnalyzeBranch for negated branches inserts imm(0) 1313 // into Cond, followed by the predicate register. For non-negated branches 1314 // it's just the register. 1315 unsigned CSz = Cond.size(); 1316 if (CSz != 1 && CSz != 2) 1317 return false; 1318 1319 unsigned P = Cond[CSz-1].getReg(); 1320 MachineInstr *PredDef = MRI->getVRegDef(P); 1321 1322 if (!PredDef->isCompare()) 1323 return false; 1324 1325 SmallSet<unsigned,2> CmpRegs; 1326 MachineOperand *CmpImmOp = nullptr; 1327 1328 // Go over all operands to the compare and look for immediate and register 1329 // operands. Assume that if the compare has a single register use and a 1330 // single immediate operand, then the register is being compared with the 1331 // immediate value. 1332 for (unsigned i = 0, n = PredDef->getNumOperands(); i < n; ++i) { 1333 MachineOperand &MO = PredDef->getOperand(i); 1334 if (MO.isReg()) { 1335 // Skip all implicit references. In one case there was: 1336 // %vreg140<def> = FCMPUGT32_rr %vreg138, %vreg139, %USR<imp-use> 1337 if (MO.isImplicit()) 1338 continue; 1339 if (MO.isUse()) { 1340 unsigned R = MO.getReg(); 1341 if (!defWithImmediate(R)) { 1342 CmpRegs.insert(MO.getReg()); 1343 continue; 1344 } 1345 // Consider the register to be the "immediate" operand. 1346 if (CmpImmOp) 1347 return false; 1348 CmpImmOp = &MO; 1349 } 1350 } else if (MO.isImm()) { 1351 if (CmpImmOp) // A second immediate argument? Confusing. Bail out. 1352 return false; 1353 CmpImmOp = &MO; 1354 } 1355 } 1356 1357 if (CmpRegs.empty()) 1358 return false; 1359 1360 // Check if the compared register follows the order we want. Fix if needed. 1361 for (RegisterInductionSet::iterator I = IndRegs.begin(), E = IndRegs.end(); 1362 I != E; ++I) { 1363 // This is a success. If the register used in the comparison is one that 1364 // we have identified as a bumped (updated) induction register, there is 1365 // nothing to do. 1366 if (CmpRegs.count(I->first)) 1367 return true; 1368 1369 // Otherwise, if the register being compared comes out of a PHI node, 1370 // and has been recognized as following the induction pattern, and is 1371 // compared against an immediate, we can fix it. 1372 const RegisterBump &RB = I->second; 1373 if (CmpRegs.count(RB.first)) { 1374 if (!CmpImmOp) 1375 return false; 1376 1377 int64_t CmpImm = getImmediate(*CmpImmOp); 1378 int64_t V = RB.second; 1379 if (V > 0 && CmpImm+V < CmpImm) // Overflow (64-bit). 1380 return false; 1381 if (V < 0 && CmpImm+V > CmpImm) // Overflow (64-bit). 1382 return false; 1383 CmpImm += V; 1384 // Some forms of cmp-immediate allow u9 and s10. Assume the worst case 1385 // scenario, i.e. an 8-bit value. 1386 if (CmpImmOp->isImm() && !isInt<8>(CmpImm)) 1387 return false; 1388 1389 // Make sure that the compare happens after the bump. Otherwise, 1390 // after the fixup, the compare would use a yet-undefined register. 1391 MachineInstr *BumpI = MRI->getVRegDef(I->first); 1392 bool Order = orderBumpCompare(BumpI, PredDef); 1393 if (!Order) 1394 return false; 1395 1396 // Finally, fix the compare instruction. 1397 setImmediate(*CmpImmOp, CmpImm); 1398 for (unsigned i = 0, n = PredDef->getNumOperands(); i < n; ++i) { 1399 MachineOperand &MO = PredDef->getOperand(i); 1400 if (MO.isReg() && MO.getReg() == RB.first) { 1401 MO.setReg(I->first); 1402 return true; 1403 } 1404 } 1405 } 1406 } 1407 1408 return false; 1409 } 1410 1411 1412 /// \brief Create a preheader for a given loop. 1413 MachineBasicBlock *HexagonHardwareLoops::createPreheaderForLoop( 1414 MachineLoop *L) { 1415 if (MachineBasicBlock *TmpPH = L->getLoopPreheader()) 1416 return TmpPH; 1417 1418 MachineBasicBlock *Header = L->getHeader(); 1419 MachineBasicBlock *Latch = L->getLoopLatch(); 1420 MachineFunction *MF = Header->getParent(); 1421 DebugLoc DL; 1422 1423 if (!Latch || Header->hasAddressTaken()) 1424 return nullptr; 1425 1426 typedef MachineBasicBlock::instr_iterator instr_iterator; 1427 1428 // Verify that all existing predecessors have analyzable branches 1429 // (or no branches at all). 1430 typedef std::vector<MachineBasicBlock*> MBBVector; 1431 MBBVector Preds(Header->pred_begin(), Header->pred_end()); 1432 SmallVector<MachineOperand,2> Tmp1; 1433 MachineBasicBlock *TB = nullptr, *FB = nullptr; 1434 1435 if (TII->AnalyzeBranch(*Latch, TB, FB, Tmp1, false)) 1436 return nullptr; 1437 1438 for (MBBVector::iterator I = Preds.begin(), E = Preds.end(); I != E; ++I) { 1439 MachineBasicBlock *PB = *I; 1440 if (PB != Latch) { 1441 bool NotAnalyzed = TII->AnalyzeBranch(*PB, TB, FB, Tmp1, false); 1442 if (NotAnalyzed) 1443 return nullptr; 1444 } 1445 } 1446 1447 MachineBasicBlock *NewPH = MF->CreateMachineBasicBlock(); 1448 MF->insert(Header, NewPH); 1449 1450 if (Header->pred_size() > 2) { 1451 // Ensure that the header has only two predecessors: the preheader and 1452 // the loop latch. Any additional predecessors of the header should 1453 // join at the newly created preheader. Inspect all PHI nodes from the 1454 // header and create appropriate corresponding PHI nodes in the preheader. 1455 1456 for (instr_iterator I = Header->instr_begin(), E = Header->instr_end(); 1457 I != E && I->isPHI(); ++I) { 1458 MachineInstr *PN = &*I; 1459 1460 const MCInstrDesc &PD = TII->get(TargetOpcode::PHI); 1461 MachineInstr *NewPN = MF->CreateMachineInstr(PD, DL); 1462 NewPH->insert(NewPH->end(), NewPN); 1463 1464 unsigned PR = PN->getOperand(0).getReg(); 1465 const TargetRegisterClass *RC = MRI->getRegClass(PR); 1466 unsigned NewPR = MRI->createVirtualRegister(RC); 1467 NewPN->addOperand(MachineOperand::CreateReg(NewPR, true)); 1468 1469 // Copy all non-latch operands of a header's PHI node to the newly 1470 // created PHI node in the preheader. 1471 for (unsigned i = 1, n = PN->getNumOperands(); i < n; i += 2) { 1472 unsigned PredR = PN->getOperand(i).getReg(); 1473 MachineBasicBlock *PredB = PN->getOperand(i+1).getMBB(); 1474 if (PredB == Latch) 1475 continue; 1476 1477 NewPN->addOperand(MachineOperand::CreateReg(PredR, false)); 1478 NewPN->addOperand(MachineOperand::CreateMBB(PredB)); 1479 } 1480 1481 // Remove copied operands from the old PHI node and add the value 1482 // coming from the preheader's PHI. 1483 for (int i = PN->getNumOperands()-2; i > 0; i -= 2) { 1484 MachineBasicBlock *PredB = PN->getOperand(i+1).getMBB(); 1485 if (PredB != Latch) { 1486 PN->RemoveOperand(i+1); 1487 PN->RemoveOperand(i); 1488 } 1489 } 1490 PN->addOperand(MachineOperand::CreateReg(NewPR, false)); 1491 PN->addOperand(MachineOperand::CreateMBB(NewPH)); 1492 } 1493 1494 } else { 1495 assert(Header->pred_size() == 2); 1496 1497 // The header has only two predecessors, but the non-latch predecessor 1498 // is not a preheader (e.g. it has other successors, etc.) 1499 // In such a case we don't need any extra PHI nodes in the new preheader, 1500 // all we need is to adjust existing PHIs in the header to now refer to 1501 // the new preheader. 1502 for (instr_iterator I = Header->instr_begin(), E = Header->instr_end(); 1503 I != E && I->isPHI(); ++I) { 1504 MachineInstr *PN = &*I; 1505 for (unsigned i = 1, n = PN->getNumOperands(); i < n; i += 2) { 1506 MachineOperand &MO = PN->getOperand(i+1); 1507 if (MO.getMBB() != Latch) 1508 MO.setMBB(NewPH); 1509 } 1510 } 1511 } 1512 1513 // "Reroute" the CFG edges to link in the new preheader. 1514 // If any of the predecessors falls through to the header, insert a branch 1515 // to the new preheader in that place. 1516 SmallVector<MachineOperand,1> Tmp2; 1517 SmallVector<MachineOperand,1> EmptyCond; 1518 1519 TB = FB = nullptr; 1520 1521 for (MBBVector::iterator I = Preds.begin(), E = Preds.end(); I != E; ++I) { 1522 MachineBasicBlock *PB = *I; 1523 if (PB != Latch) { 1524 Tmp2.clear(); 1525 bool NotAnalyzed = TII->AnalyzeBranch(*PB, TB, FB, Tmp2, false); 1526 (void)NotAnalyzed; // suppress compiler warning 1527 assert (!NotAnalyzed && "Should be analyzable!"); 1528 if (TB != Header && (Tmp2.empty() || FB != Header)) 1529 TII->InsertBranch(*PB, NewPH, nullptr, EmptyCond, DL); 1530 PB->ReplaceUsesOfBlockWith(Header, NewPH); 1531 } 1532 } 1533 1534 // It can happen that the latch block will fall through into the header. 1535 // Insert an unconditional branch to the header. 1536 TB = FB = nullptr; 1537 bool LatchNotAnalyzed = TII->AnalyzeBranch(*Latch, TB, FB, Tmp2, false); 1538 (void)LatchNotAnalyzed; // suppress compiler warning 1539 assert (!LatchNotAnalyzed && "Should be analyzable!"); 1540 if (!TB && !FB) 1541 TII->InsertBranch(*Latch, Header, nullptr, EmptyCond, DL); 1542 1543 // Finally, the branch from the preheader to the header. 1544 TII->InsertBranch(*NewPH, Header, nullptr, EmptyCond, DL); 1545 NewPH->addSuccessor(Header); 1546 1547 return NewPH; 1548 } 1549