1 //===-- BranchFolding.cpp - Fold machine code branch instructions ---------===// 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 forwards branches to unconditional branches to make them branch 11 // directly to the target block. This pass often results in dead MBB's, which 12 // it then removes. 13 // 14 // Note that this pass must be run after register allocation, it cannot handle 15 // SSA form. It also must handle virtual registers for targets that emit virtual 16 // ISA (e.g. NVPTX). 17 // 18 //===----------------------------------------------------------------------===// 19 20 #include "BranchFolding.h" 21 #include "llvm/ADT/STLExtras.h" 22 #include "llvm/ADT/SmallSet.h" 23 #include "llvm/ADT/Statistic.h" 24 #include "llvm/CodeGen/Analysis.h" 25 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 26 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 27 #include "llvm/CodeGen/MachineFunctionPass.h" 28 #include "llvm/CodeGen/MachineJumpTableInfo.h" 29 #include "llvm/CodeGen/MachineMemOperand.h" 30 #include "llvm/CodeGen/MachineLoopInfo.h" 31 #include "llvm/CodeGen/MachineModuleInfo.h" 32 #include "llvm/CodeGen/MachineRegisterInfo.h" 33 #include "llvm/CodeGen/Passes.h" 34 #include "llvm/CodeGen/TargetPassConfig.h" 35 #include "llvm/IR/Function.h" 36 #include "llvm/Support/CommandLine.h" 37 #include "llvm/Support/Debug.h" 38 #include "llvm/Support/ErrorHandling.h" 39 #include "llvm/Support/raw_ostream.h" 40 #include "llvm/Target/TargetInstrInfo.h" 41 #include "llvm/Target/TargetRegisterInfo.h" 42 #include "llvm/Target/TargetSubtargetInfo.h" 43 #include <algorithm> 44 using namespace llvm; 45 46 #define DEBUG_TYPE "branchfolding" 47 48 STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 49 STATISTIC(NumBranchOpts, "Number of branches optimized"); 50 STATISTIC(NumTailMerge , "Number of block tails merged"); 51 STATISTIC(NumHoist , "Number of times common instructions are hoisted"); 52 53 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 54 cl::init(cl::BOU_UNSET), cl::Hidden); 55 56 // Throttle for huge numbers of predecessors (compile speed problems) 57 static cl::opt<unsigned> 58 TailMergeThreshold("tail-merge-threshold", 59 cl::desc("Max number of predecessors to consider tail merging"), 60 cl::init(150), cl::Hidden); 61 62 // Heuristic for tail merging (and, inversely, tail duplication). 63 // TODO: This should be replaced with a target query. 64 static cl::opt<unsigned> 65 TailMergeSize("tail-merge-size", 66 cl::desc("Min number of instructions to consider tail merging"), 67 cl::init(3), cl::Hidden); 68 69 namespace { 70 /// BranchFolderPass - Wrap branch folder in a machine function pass. 71 class BranchFolderPass : public MachineFunctionPass { 72 public: 73 static char ID; 74 explicit BranchFolderPass(): MachineFunctionPass(ID) {} 75 76 bool runOnMachineFunction(MachineFunction &MF) override; 77 78 void getAnalysisUsage(AnalysisUsage &AU) const override { 79 AU.addRequired<MachineBlockFrequencyInfo>(); 80 AU.addRequired<MachineBranchProbabilityInfo>(); 81 AU.addRequired<TargetPassConfig>(); 82 MachineFunctionPass::getAnalysisUsage(AU); 83 } 84 }; 85 } 86 87 char BranchFolderPass::ID = 0; 88 char &llvm::BranchFolderPassID = BranchFolderPass::ID; 89 90 INITIALIZE_PASS(BranchFolderPass, "branch-folder", 91 "Control Flow Optimizer", false, false) 92 93 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { 94 if (skipFunction(*MF.getFunction())) 95 return false; 96 97 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>(); 98 // TailMerge can create jump into if branches that make CFG irreducible for 99 // HW that requires structurized CFG. 100 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() && 101 PassConfig->getEnableTailMerge(); 102 BranchFolder::MBFIWrapper MBBFreqInfo( 103 getAnalysis<MachineBlockFrequencyInfo>()); 104 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo, 105 getAnalysis<MachineBranchProbabilityInfo>()); 106 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(), 107 MF.getSubtarget().getRegisterInfo(), 108 getAnalysisIfAvailable<MachineModuleInfo>()); 109 } 110 111 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist, 112 MBFIWrapper &FreqInfo, 113 const MachineBranchProbabilityInfo &ProbInfo) 114 : EnableHoistCommonCode(CommonHoist), MBBFreqInfo(FreqInfo), 115 MBPI(ProbInfo) { 116 switch (FlagEnableTailMerge) { 117 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 118 case cl::BOU_TRUE: EnableTailMerge = true; break; 119 case cl::BOU_FALSE: EnableTailMerge = false; break; 120 } 121 } 122 123 /// RemoveDeadBlock - Remove the specified dead machine basic block from the 124 /// function, updating the CFG. 125 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 126 assert(MBB->pred_empty() && "MBB must be dead!"); 127 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB); 128 129 MachineFunction *MF = MBB->getParent(); 130 // drop all successors. 131 while (!MBB->succ_empty()) 132 MBB->removeSuccessor(MBB->succ_end()-1); 133 134 // Avoid matching if this pointer gets reused. 135 TriedMerging.erase(MBB); 136 137 // Remove the block. 138 MF->erase(MBB); 139 FuncletMembership.erase(MBB); 140 if (MLI) 141 MLI->removeBlock(MBB); 142 } 143 144 /// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 145 /// followed by terminators, and if the implicitly defined registers are not 146 /// used by the terminators, remove those implicit_def's. e.g. 147 /// BB1: 148 /// r0 = implicit_def 149 /// r1 = implicit_def 150 /// br 151 /// This block can be optimized away later if the implicit instructions are 152 /// removed. 153 bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 154 SmallSet<unsigned, 4> ImpDefRegs; 155 MachineBasicBlock::iterator I = MBB->begin(); 156 while (I != MBB->end()) { 157 if (!I->isImplicitDef()) 158 break; 159 unsigned Reg = I->getOperand(0).getReg(); 160 if (TargetRegisterInfo::isPhysicalRegister(Reg)) { 161 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true); 162 SubRegs.isValid(); ++SubRegs) 163 ImpDefRegs.insert(*SubRegs); 164 } else { 165 ImpDefRegs.insert(Reg); 166 } 167 ++I; 168 } 169 if (ImpDefRegs.empty()) 170 return false; 171 172 MachineBasicBlock::iterator FirstTerm = I; 173 while (I != MBB->end()) { 174 if (!TII->isUnpredicatedTerminator(*I)) 175 return false; 176 // See if it uses any of the implicitly defined registers. 177 for (const MachineOperand &MO : I->operands()) { 178 if (!MO.isReg() || !MO.isUse()) 179 continue; 180 unsigned Reg = MO.getReg(); 181 if (ImpDefRegs.count(Reg)) 182 return false; 183 } 184 ++I; 185 } 186 187 I = MBB->begin(); 188 while (I != FirstTerm) { 189 MachineInstr *ImpDefMI = &*I; 190 ++I; 191 MBB->erase(ImpDefMI); 192 } 193 194 return true; 195 } 196 197 /// OptimizeFunction - Perhaps branch folding, tail merging and other 198 /// CFG optimizations on the given function. Block placement changes the layout 199 /// and may create new tail merging opportunities. 200 bool BranchFolder::OptimizeFunction(MachineFunction &MF, 201 const TargetInstrInfo *tii, 202 const TargetRegisterInfo *tri, 203 MachineModuleInfo *mmi, 204 MachineLoopInfo *mli, bool AfterPlacement) { 205 if (!tii) return false; 206 207 TriedMerging.clear(); 208 209 AfterBlockPlacement = AfterPlacement; 210 TII = tii; 211 TRI = tri; 212 MMI = mmi; 213 MLI = mli; 214 215 MachineRegisterInfo &MRI = MF.getRegInfo(); 216 UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF); 217 if (!UpdateLiveIns) 218 MRI.invalidateLiveness(); 219 220 // Fix CFG. The later algorithms expect it to be right. 221 bool MadeChange = false; 222 for (MachineBasicBlock &MBB : MF) { 223 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 224 SmallVector<MachineOperand, 4> Cond; 225 if (!TII->analyzeBranch(MBB, TBB, FBB, Cond, true)) 226 MadeChange |= MBB.CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 227 MadeChange |= OptimizeImpDefsBlock(&MBB); 228 } 229 230 // Recalculate funclet membership. 231 FuncletMembership = getFuncletMembership(MF); 232 233 bool MadeChangeThisIteration = true; 234 while (MadeChangeThisIteration) { 235 MadeChangeThisIteration = TailMergeBlocks(MF); 236 // No need to clean up if tail merging does not change anything after the 237 // block placement. 238 if (!AfterBlockPlacement || MadeChangeThisIteration) 239 MadeChangeThisIteration |= OptimizeBranches(MF); 240 if (EnableHoistCommonCode) 241 MadeChangeThisIteration |= HoistCommonCode(MF); 242 MadeChange |= MadeChangeThisIteration; 243 } 244 245 // See if any jump tables have become dead as the code generator 246 // did its thing. 247 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 248 if (!JTI) 249 return MadeChange; 250 251 // Walk the function to find jump tables that are live. 252 BitVector JTIsLive(JTI->getJumpTables().size()); 253 for (const MachineBasicBlock &BB : MF) { 254 for (const MachineInstr &I : BB) 255 for (const MachineOperand &Op : I.operands()) { 256 if (!Op.isJTI()) continue; 257 258 // Remember that this JT is live. 259 JTIsLive.set(Op.getIndex()); 260 } 261 } 262 263 // Finally, remove dead jump tables. This happens when the 264 // indirect jump was unreachable (and thus deleted). 265 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 266 if (!JTIsLive.test(i)) { 267 JTI->RemoveJumpTable(i); 268 MadeChange = true; 269 } 270 271 return MadeChange; 272 } 273 274 //===----------------------------------------------------------------------===// 275 // Tail Merging of Blocks 276 //===----------------------------------------------------------------------===// 277 278 /// HashMachineInstr - Compute a hash value for MI and its operands. 279 static unsigned HashMachineInstr(const MachineInstr &MI) { 280 unsigned Hash = MI.getOpcode(); 281 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 282 const MachineOperand &Op = MI.getOperand(i); 283 284 // Merge in bits from the operand if easy. We can't use MachineOperand's 285 // hash_code here because it's not deterministic and we sort by hash value 286 // later. 287 unsigned OperandHash = 0; 288 switch (Op.getType()) { 289 case MachineOperand::MO_Register: 290 OperandHash = Op.getReg(); 291 break; 292 case MachineOperand::MO_Immediate: 293 OperandHash = Op.getImm(); 294 break; 295 case MachineOperand::MO_MachineBasicBlock: 296 OperandHash = Op.getMBB()->getNumber(); 297 break; 298 case MachineOperand::MO_FrameIndex: 299 case MachineOperand::MO_ConstantPoolIndex: 300 case MachineOperand::MO_JumpTableIndex: 301 OperandHash = Op.getIndex(); 302 break; 303 case MachineOperand::MO_GlobalAddress: 304 case MachineOperand::MO_ExternalSymbol: 305 // Global address / external symbol are too hard, don't bother, but do 306 // pull in the offset. 307 OperandHash = Op.getOffset(); 308 break; 309 default: 310 break; 311 } 312 313 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31); 314 } 315 return Hash; 316 } 317 318 /// HashEndOfMBB - Hash the last instruction in the MBB. 319 static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) { 320 MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr(); 321 if (I == MBB.end()) 322 return 0; 323 324 return HashMachineInstr(*I); 325 } 326 327 /// ComputeCommonTailLength - Given two machine basic blocks, compute the number 328 /// of instructions they actually have in common together at their end. Return 329 /// iterators for the first shared instruction in each block. 330 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 331 MachineBasicBlock *MBB2, 332 MachineBasicBlock::iterator &I1, 333 MachineBasicBlock::iterator &I2) { 334 I1 = MBB1->end(); 335 I2 = MBB2->end(); 336 337 unsigned TailLen = 0; 338 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 339 --I1; --I2; 340 // Skip debugging pseudos; necessary to avoid changing the code. 341 while (I1->isDebugValue()) { 342 if (I1==MBB1->begin()) { 343 while (I2->isDebugValue()) { 344 if (I2==MBB2->begin()) 345 // I1==DBG at begin; I2==DBG at begin 346 return TailLen; 347 --I2; 348 } 349 ++I2; 350 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin 351 return TailLen; 352 } 353 --I1; 354 } 355 // I1==first (untested) non-DBG preceding known match 356 while (I2->isDebugValue()) { 357 if (I2==MBB2->begin()) { 358 ++I1; 359 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin 360 return TailLen; 361 } 362 --I2; 363 } 364 // I1, I2==first (untested) non-DBGs preceding known match 365 if (!I1->isIdenticalTo(*I2) || 366 // FIXME: This check is dubious. It's used to get around a problem where 367 // people incorrectly expect inline asm directives to remain in the same 368 // relative order. This is untenable because normal compiler 369 // optimizations (like this one) may reorder and/or merge these 370 // directives. 371 I1->isInlineAsm()) { 372 ++I1; ++I2; 373 break; 374 } 375 ++TailLen; 376 } 377 // Back past possible debugging pseudos at beginning of block. This matters 378 // when one block differs from the other only by whether debugging pseudos 379 // are present at the beginning. (This way, the various checks later for 380 // I1==MBB1->begin() work as expected.) 381 if (I1 == MBB1->begin() && I2 != MBB2->begin()) { 382 --I2; 383 while (I2->isDebugValue()) { 384 if (I2 == MBB2->begin()) 385 return TailLen; 386 --I2; 387 } 388 ++I2; 389 } 390 if (I2 == MBB2->begin() && I1 != MBB1->begin()) { 391 --I1; 392 while (I1->isDebugValue()) { 393 if (I1 == MBB1->begin()) 394 return TailLen; 395 --I1; 396 } 397 ++I1; 398 } 399 return TailLen; 400 } 401 402 void BranchFolder::computeLiveIns(MachineBasicBlock &MBB) { 403 if (!UpdateLiveIns) 404 return; 405 406 LiveRegs.init(TRI); 407 LiveRegs.addLiveOutsNoPristines(MBB); 408 for (MachineInstr &MI : make_range(MBB.rbegin(), MBB.rend())) 409 LiveRegs.stepBackward(MI); 410 411 for (unsigned Reg : LiveRegs) { 412 // Skip the register if we are about to add one of its super registers. 413 bool ContainsSuperReg = false; 414 for (MCSuperRegIterator SReg(Reg, TRI); SReg.isValid(); ++SReg) { 415 if (LiveRegs.contains(*SReg)) { 416 ContainsSuperReg = true; 417 break; 418 } 419 } 420 if (ContainsSuperReg) 421 continue; 422 MBB.addLiveIn(Reg); 423 } 424 } 425 426 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 427 /// after it, replacing it with an unconditional branch to NewDest. 428 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 429 MachineBasicBlock *NewDest) { 430 TII->ReplaceTailWithBranchTo(OldInst, NewDest); 431 432 computeLiveIns(*NewDest); 433 434 ++NumTailMerge; 435 } 436 437 /// SplitMBBAt - Given a machine basic block and an iterator into it, split the 438 /// MBB so that the part before the iterator falls into the part starting at the 439 /// iterator. This returns the new MBB. 440 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 441 MachineBasicBlock::iterator BBI1, 442 const BasicBlock *BB) { 443 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1)) 444 return nullptr; 445 446 MachineFunction &MF = *CurMBB.getParent(); 447 448 // Create the fall-through block. 449 MachineFunction::iterator MBBI = CurMBB.getIterator(); 450 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(BB); 451 CurMBB.getParent()->insert(++MBBI, NewMBB); 452 453 // Move all the successors of this block to the specified block. 454 NewMBB->transferSuccessors(&CurMBB); 455 456 // Add an edge from CurMBB to NewMBB for the fall-through. 457 CurMBB.addSuccessor(NewMBB); 458 459 // Splice the code over. 460 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 461 462 // NewMBB belongs to the same loop as CurMBB. 463 if (MLI) 464 if (MachineLoop *ML = MLI->getLoopFor(&CurMBB)) 465 ML->addBasicBlockToLoop(NewMBB, MLI->getBase()); 466 467 // NewMBB inherits CurMBB's block frequency. 468 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB)); 469 470 computeLiveIns(*NewMBB); 471 472 // Add the new block to the funclet. 473 const auto &FuncletI = FuncletMembership.find(&CurMBB); 474 if (FuncletI != FuncletMembership.end()) { 475 auto n = FuncletI->second; 476 FuncletMembership[NewMBB] = n; 477 } 478 479 return NewMBB; 480 } 481 482 /// EstimateRuntime - Make a rough estimate for how long it will take to run 483 /// the specified code. 484 static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 485 MachineBasicBlock::iterator E) { 486 unsigned Time = 0; 487 for (; I != E; ++I) { 488 if (I->isDebugValue()) 489 continue; 490 if (I->isCall()) 491 Time += 10; 492 else if (I->mayLoad() || I->mayStore()) 493 Time += 2; 494 else 495 ++Time; 496 } 497 return Time; 498 } 499 500 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these 501 // branches temporarily for tail merging). In the case where CurMBB ends 502 // with a conditional branch to the next block, optimize by reversing the 503 // test and conditionally branching to SuccMBB instead. 504 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 505 const TargetInstrInfo *TII) { 506 MachineFunction *MF = CurMBB->getParent(); 507 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB)); 508 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 509 SmallVector<MachineOperand, 4> Cond; 510 DebugLoc dl; // FIXME: this is nowhere 511 if (I != MF->end() && !TII->analyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 512 MachineBasicBlock *NextBB = &*I; 513 if (TBB == NextBB && !Cond.empty() && !FBB) { 514 if (!TII->ReverseBranchCondition(Cond)) { 515 TII->RemoveBranch(*CurMBB); 516 TII->InsertBranch(*CurMBB, SuccBB, nullptr, Cond, dl); 517 return; 518 } 519 } 520 } 521 TII->InsertBranch(*CurMBB, SuccBB, nullptr, 522 SmallVector<MachineOperand, 0>(), dl); 523 } 524 525 bool 526 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 527 if (getHash() < o.getHash()) 528 return true; 529 if (getHash() > o.getHash()) 530 return false; 531 if (getBlock()->getNumber() < o.getBlock()->getNumber()) 532 return true; 533 if (getBlock()->getNumber() > o.getBlock()->getNumber()) 534 return false; 535 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 536 // an object with itself. 537 #ifndef _GLIBCXX_DEBUG 538 llvm_unreachable("Predecessor appears twice"); 539 #else 540 return false; 541 #endif 542 } 543 544 BlockFrequency 545 BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const { 546 auto I = MergedBBFreq.find(MBB); 547 548 if (I != MergedBBFreq.end()) 549 return I->second; 550 551 return MBFI.getBlockFreq(MBB); 552 } 553 554 void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB, 555 BlockFrequency F) { 556 MergedBBFreq[MBB] = F; 557 } 558 559 raw_ostream & 560 BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS, 561 const MachineBasicBlock *MBB) const { 562 return MBFI.printBlockFreq(OS, getBlockFreq(MBB)); 563 } 564 565 raw_ostream & 566 BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS, 567 const BlockFrequency Freq) const { 568 return MBFI.printBlockFreq(OS, Freq); 569 } 570 571 /// CountTerminators - Count the number of terminators in the given 572 /// block and set I to the position of the first non-terminator, if there 573 /// is one, or MBB->end() otherwise. 574 static unsigned CountTerminators(MachineBasicBlock *MBB, 575 MachineBasicBlock::iterator &I) { 576 I = MBB->end(); 577 unsigned NumTerms = 0; 578 for (;;) { 579 if (I == MBB->begin()) { 580 I = MBB->end(); 581 break; 582 } 583 --I; 584 if (!I->isTerminator()) break; 585 ++NumTerms; 586 } 587 return NumTerms; 588 } 589 590 /// ProfitableToMerge - Check if two machine basic blocks have a common tail 591 /// and decide if it would be profitable to merge those tails. Return the 592 /// length of the common tail and iterators to the first common instruction 593 /// in each block. 594 static bool 595 ProfitableToMerge(MachineBasicBlock *MBB1, MachineBasicBlock *MBB2, 596 unsigned minCommonTailLength, unsigned &CommonTailLen, 597 MachineBasicBlock::iterator &I1, 598 MachineBasicBlock::iterator &I2, MachineBasicBlock *SuccBB, 599 MachineBasicBlock *PredBB, 600 DenseMap<const MachineBasicBlock *, int> &FuncletMembership) { 601 // It is never profitable to tail-merge blocks from two different funclets. 602 if (!FuncletMembership.empty()) { 603 auto Funclet1 = FuncletMembership.find(MBB1); 604 assert(Funclet1 != FuncletMembership.end()); 605 auto Funclet2 = FuncletMembership.find(MBB2); 606 assert(Funclet2 != FuncletMembership.end()); 607 if (Funclet1->second != Funclet2->second) 608 return false; 609 } 610 611 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 612 if (CommonTailLen == 0) 613 return false; 614 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber() 615 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen 616 << '\n'); 617 618 // It's almost always profitable to merge any number of non-terminator 619 // instructions with the block that falls through into the common successor. 620 if (MBB1 == PredBB || MBB2 == PredBB) { 621 MachineBasicBlock::iterator I; 622 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 623 if (CommonTailLen > NumTerms) 624 return true; 625 } 626 627 // If one of the blocks can be completely merged and happens to be in 628 // a position where the other could fall through into it, merge any number 629 // of instructions, because it can be done without a branch. 630 // TODO: If the blocks are not adjacent, move one of them so that they are? 631 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 632 return true; 633 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 634 return true; 635 636 // If both blocks have an unconditional branch temporarily stripped out, 637 // count that as an additional common instruction for the following 638 // heuristics. 639 unsigned EffectiveTailLen = CommonTailLen; 640 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 641 !MBB1->back().isBarrier() && 642 !MBB2->back().isBarrier()) 643 ++EffectiveTailLen; 644 645 // Check if the common tail is long enough to be worthwhile. 646 if (EffectiveTailLen >= minCommonTailLength) 647 return true; 648 649 // If we are optimizing for code size, 2 instructions in common is enough if 650 // we don't have to split a block. At worst we will be introducing 1 new 651 // branch instruction, which is likely to be smaller than the 2 652 // instructions that would be deleted in the merge. 653 MachineFunction *MF = MBB1->getParent(); 654 return EffectiveTailLen >= 2 && MF->getFunction()->optForSize() && 655 (I1 == MBB1->begin() || I2 == MBB2->begin()); 656 } 657 658 /// ComputeSameTails - Look through all the blocks in MergePotentials that have 659 /// hash CurHash (guaranteed to match the last element). Build the vector 660 /// SameTails of all those that have the (same) largest number of instructions 661 /// in common of any pair of these blocks. SameTails entries contain an 662 /// iterator into MergePotentials (from which the MachineBasicBlock can be 663 /// found) and a MachineBasicBlock::iterator into that MBB indicating the 664 /// instruction where the matching code sequence begins. 665 /// Order of elements in SameTails is the reverse of the order in which 666 /// those blocks appear in MergePotentials (where they are not necessarily 667 /// consecutive). 668 unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 669 unsigned minCommonTailLength, 670 MachineBasicBlock *SuccBB, 671 MachineBasicBlock *PredBB) { 672 unsigned maxCommonTailLength = 0U; 673 SameTails.clear(); 674 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 675 MPIterator HighestMPIter = std::prev(MergePotentials.end()); 676 for (MPIterator CurMPIter = std::prev(MergePotentials.end()), 677 B = MergePotentials.begin(); 678 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) { 679 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) { 680 unsigned CommonTailLen; 681 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 682 minCommonTailLength, 683 CommonTailLen, TrialBBI1, TrialBBI2, 684 SuccBB, PredBB, 685 FuncletMembership)) { 686 if (CommonTailLen > maxCommonTailLength) { 687 SameTails.clear(); 688 maxCommonTailLength = CommonTailLen; 689 HighestMPIter = CurMPIter; 690 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 691 } 692 if (HighestMPIter == CurMPIter && 693 CommonTailLen == maxCommonTailLength) 694 SameTails.push_back(SameTailElt(I, TrialBBI2)); 695 } 696 if (I == B) 697 break; 698 } 699 } 700 return maxCommonTailLength; 701 } 702 703 /// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 704 /// MergePotentials, restoring branches at ends of blocks as appropriate. 705 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 706 MachineBasicBlock *SuccBB, 707 MachineBasicBlock *PredBB) { 708 MPIterator CurMPIter, B; 709 for (CurMPIter = std::prev(MergePotentials.end()), 710 B = MergePotentials.begin(); 711 CurMPIter->getHash() == CurHash; --CurMPIter) { 712 // Put the unconditional branch back, if we need one. 713 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 714 if (SuccBB && CurMBB != PredBB) 715 FixTail(CurMBB, SuccBB, TII); 716 if (CurMPIter == B) 717 break; 718 } 719 if (CurMPIter->getHash() != CurHash) 720 CurMPIter++; 721 MergePotentials.erase(CurMPIter, MergePotentials.end()); 722 } 723 724 /// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 725 /// only of the common tail. Create a block that does by splitting one. 726 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 727 MachineBasicBlock *SuccBB, 728 unsigned maxCommonTailLength, 729 unsigned &commonTailIndex) { 730 commonTailIndex = 0; 731 unsigned TimeEstimate = ~0U; 732 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 733 // Use PredBB if possible; that doesn't require a new branch. 734 if (SameTails[i].getBlock() == PredBB) { 735 commonTailIndex = i; 736 break; 737 } 738 // Otherwise, make a (fairly bogus) choice based on estimate of 739 // how long it will take the various blocks to execute. 740 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 741 SameTails[i].getTailStartPos()); 742 if (t <= TimeEstimate) { 743 TimeEstimate = t; 744 commonTailIndex = i; 745 } 746 } 747 748 MachineBasicBlock::iterator BBI = 749 SameTails[commonTailIndex].getTailStartPos(); 750 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 751 752 // If the common tail includes any debug info we will take it pretty 753 // randomly from one of the inputs. Might be better to remove it? 754 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 755 << maxCommonTailLength); 756 757 // If the split block unconditionally falls-thru to SuccBB, it will be 758 // merged. In control flow terms it should then take SuccBB's name. e.g. If 759 // SuccBB is an inner loop, the common tail is still part of the inner loop. 760 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ? 761 SuccBB->getBasicBlock() : MBB->getBasicBlock(); 762 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB); 763 if (!newMBB) { 764 DEBUG(dbgs() << "... failed!"); 765 return false; 766 } 767 768 SameTails[commonTailIndex].setBlock(newMBB); 769 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 770 771 // If we split PredBB, newMBB is the new predecessor. 772 if (PredBB == MBB) 773 PredBB = newMBB; 774 775 return true; 776 } 777 778 static void 779 mergeMMOsFromMemoryOperations(MachineBasicBlock::iterator MBBIStartPos, 780 MachineBasicBlock &MBBCommon) { 781 // Merge MMOs from memory operations in the common block. 782 MachineBasicBlock *MBB = MBBIStartPos->getParent(); 783 // Note CommonTailLen does not necessarily matches the size of 784 // the common BB nor all its instructions because of debug 785 // instructions differences. 786 unsigned CommonTailLen = 0; 787 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos) 788 ++CommonTailLen; 789 790 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin(); 791 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend(); 792 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin(); 793 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend(); 794 795 while (CommonTailLen--) { 796 assert(MBBI != MBBIE && "Reached BB end within common tail length!"); 797 (void)MBBIE; 798 799 if (MBBI->isDebugValue()) { 800 ++MBBI; 801 continue; 802 } 803 804 while ((MBBICommon != MBBIECommon) && MBBICommon->isDebugValue()) 805 ++MBBICommon; 806 807 assert(MBBICommon != MBBIECommon && 808 "Reached BB end within common tail length!"); 809 assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!"); 810 811 if (MBBICommon->mayLoad() || MBBICommon->mayStore()) 812 MBBICommon->setMemRefs(MBBICommon->mergeMemRefsWith(*MBBI)); 813 814 ++MBBI; 815 ++MBBICommon; 816 } 817 } 818 819 // See if any of the blocks in MergePotentials (which all have SuccBB as a 820 // successor, or all have no successor if it is null) can be tail-merged. 821 // If there is a successor, any blocks in MergePotentials that are not 822 // tail-merged and are not immediately before Succ must have an unconditional 823 // branch to Succ added (but the predecessor/successor lists need no 824 // adjustment). The lone predecessor of Succ that falls through into Succ, 825 // if any, is given in PredBB. 826 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 827 MachineBasicBlock *PredBB) { 828 bool MadeChange = false; 829 830 // Except for the special cases below, tail-merge if there are at least 831 // this many instructions in common. 832 unsigned minCommonTailLength = TailMergeSize; 833 834 DEBUG(dbgs() << "\nTryTailMergeBlocks: "; 835 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 836 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 837 << (i == e-1 ? "" : ", "); 838 dbgs() << "\n"; 839 if (SuccBB) { 840 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 841 if (PredBB) 842 dbgs() << " which has fall-through from BB#" 843 << PredBB->getNumber() << "\n"; 844 } 845 dbgs() << "Looking for common tails of at least " 846 << minCommonTailLength << " instruction" 847 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 848 ); 849 850 // Sort by hash value so that blocks with identical end sequences sort 851 // together. 852 array_pod_sort(MergePotentials.begin(), MergePotentials.end()); 853 854 // Walk through equivalence sets looking for actual exact matches. 855 while (MergePotentials.size() > 1) { 856 unsigned CurHash = MergePotentials.back().getHash(); 857 858 // Build SameTails, identifying the set of blocks with this hash code 859 // and with the maximum number of instructions in common. 860 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 861 minCommonTailLength, 862 SuccBB, PredBB); 863 864 // If we didn't find any pair that has at least minCommonTailLength 865 // instructions in common, remove all blocks with this hash code and retry. 866 if (SameTails.empty()) { 867 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 868 continue; 869 } 870 871 // If one of the blocks is the entire common tail (and not the entry 872 // block, which we can't jump to), we can treat all blocks with this same 873 // tail at once. Use PredBB if that is one of the possibilities, as that 874 // will not introduce any extra branches. 875 MachineBasicBlock *EntryBB = 876 &MergePotentials.front().getBlock()->getParent()->front(); 877 unsigned commonTailIndex = SameTails.size(); 878 // If there are two blocks, check to see if one can be made to fall through 879 // into the other. 880 if (SameTails.size() == 2 && 881 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 882 SameTails[1].tailIsWholeBlock()) 883 commonTailIndex = 1; 884 else if (SameTails.size() == 2 && 885 SameTails[1].getBlock()->isLayoutSuccessor( 886 SameTails[0].getBlock()) && 887 SameTails[0].tailIsWholeBlock()) 888 commonTailIndex = 0; 889 else { 890 // Otherwise just pick one, favoring the fall-through predecessor if 891 // there is one. 892 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 893 MachineBasicBlock *MBB = SameTails[i].getBlock(); 894 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 895 continue; 896 if (MBB == PredBB) { 897 commonTailIndex = i; 898 break; 899 } 900 if (SameTails[i].tailIsWholeBlock()) 901 commonTailIndex = i; 902 } 903 } 904 905 if (commonTailIndex == SameTails.size() || 906 (SameTails[commonTailIndex].getBlock() == PredBB && 907 !SameTails[commonTailIndex].tailIsWholeBlock())) { 908 // None of the blocks consist entirely of the common tail. 909 // Split a block so that one does. 910 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB, 911 maxCommonTailLength, commonTailIndex)) { 912 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 913 continue; 914 } 915 } 916 917 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 918 919 // Recompute common tail MBB's edge weights and block frequency. 920 setCommonTailEdgeWeights(*MBB); 921 922 // MBB is common tail. Adjust all other BB's to jump to this one. 923 // Traversal must be forwards so erases work. 924 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber() 925 << " for "); 926 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 927 if (commonTailIndex == i) 928 continue; 929 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() 930 << (i == e-1 ? "" : ", ")); 931 // Merge MMOs from memory operations as needed. 932 mergeMMOsFromMemoryOperations(SameTails[i].getTailStartPos(), *MBB); 933 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 934 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 935 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 936 MergePotentials.erase(SameTails[i].getMPIter()); 937 } 938 DEBUG(dbgs() << "\n"); 939 // We leave commonTailIndex in the worklist in case there are other blocks 940 // that match it with a smaller number of instructions. 941 MadeChange = true; 942 } 943 return MadeChange; 944 } 945 946 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 947 bool MadeChange = false; 948 if (!EnableTailMerge) return MadeChange; 949 950 // First find blocks with no successors. 951 // Block placement does not create new tail merging opportunities for these 952 // blocks. 953 if (!AfterBlockPlacement) { 954 MergePotentials.clear(); 955 for (MachineBasicBlock &MBB : MF) { 956 if (MergePotentials.size() == TailMergeThreshold) 957 break; 958 if (!TriedMerging.count(&MBB) && MBB.succ_empty()) 959 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB)); 960 } 961 962 // If this is a large problem, avoid visiting the same basic blocks 963 // multiple times. 964 if (MergePotentials.size() == TailMergeThreshold) 965 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 966 TriedMerging.insert(MergePotentials[i].getBlock()); 967 968 // See if we can do any tail merging on those. 969 if (MergePotentials.size() >= 2) 970 MadeChange |= TryTailMergeBlocks(nullptr, nullptr); 971 } 972 973 // Look at blocks (IBB) with multiple predecessors (PBB). 974 // We change each predecessor to a canonical form, by 975 // (1) temporarily removing any unconditional branch from the predecessor 976 // to IBB, and 977 // (2) alter conditional branches so they branch to the other block 978 // not IBB; this may require adding back an unconditional branch to IBB 979 // later, where there wasn't one coming in. E.g. 980 // Bcc IBB 981 // fallthrough to QBB 982 // here becomes 983 // Bncc QBB 984 // with a conceptual B to IBB after that, which never actually exists. 985 // With those changes, we see whether the predecessors' tails match, 986 // and merge them if so. We change things out of canonical form and 987 // back to the way they were later in the process. (OptimizeBranches 988 // would undo some of this, but we can't use it, because we'd get into 989 // a compile-time infinite loop repeatedly doing and undoing the same 990 // transformations.) 991 992 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end(); 993 I != E; ++I) { 994 if (I->pred_size() < 2) continue; 995 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 996 MachineBasicBlock *IBB = &*I; 997 MachineBasicBlock *PredBB = &*std::prev(I); 998 MergePotentials.clear(); 999 for (MachineBasicBlock *PBB : I->predecessors()) { 1000 if (MergePotentials.size() == TailMergeThreshold) 1001 break; 1002 1003 if (TriedMerging.count(PBB)) 1004 continue; 1005 1006 // Skip blocks that loop to themselves, can't tail merge these. 1007 if (PBB == IBB) 1008 continue; 1009 1010 // Visit each predecessor only once. 1011 if (!UniquePreds.insert(PBB).second) 1012 continue; 1013 1014 // Skip blocks which may jump to a landing pad. Can't tail merge these. 1015 if (PBB->hasEHPadSuccessor()) 1016 continue; 1017 1018 // Bail out if the loop header (IBB) is not the top of the loop chain 1019 // after the block placement. Otherwise, the common tail of IBB's 1020 // predecessors may become the loop top if block placement is called again 1021 // and the predecessors may branch to this common tail. 1022 // FIXME: Relaxed this check if the algorithm of finding loop top is 1023 // changed in MBP. 1024 if (AfterBlockPlacement && MLI) 1025 if (MachineLoop *ML = MLI->getLoopFor(IBB)) 1026 if (IBB == ML->getHeader() && ML == MLI->getLoopFor(PBB)) 1027 continue; 1028 1029 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 1030 SmallVector<MachineOperand, 4> Cond; 1031 if (!TII->analyzeBranch(*PBB, TBB, FBB, Cond, true)) { 1032 // Failing case: IBB is the target of a cbr, and we cannot reverse the 1033 // branch. 1034 SmallVector<MachineOperand, 4> NewCond(Cond); 1035 if (!Cond.empty() && TBB == IBB) { 1036 if (TII->ReverseBranchCondition(NewCond)) 1037 continue; 1038 // This is the QBB case described above 1039 if (!FBB) { 1040 auto Next = ++PBB->getIterator(); 1041 if (Next != MF.end()) 1042 FBB = &*Next; 1043 } 1044 } 1045 1046 // Failing case: the only way IBB can be reached from PBB is via 1047 // exception handling. Happens for landing pads. Would be nice to have 1048 // a bit in the edge so we didn't have to do all this. 1049 if (IBB->isEHPad()) { 1050 MachineFunction::iterator IP = ++PBB->getIterator(); 1051 MachineBasicBlock *PredNextBB = nullptr; 1052 if (IP != MF.end()) 1053 PredNextBB = &*IP; 1054 if (!TBB) { 1055 if (IBB != PredNextBB) // fallthrough 1056 continue; 1057 } else if (FBB) { 1058 if (TBB != IBB && FBB != IBB) // cbr then ubr 1059 continue; 1060 } else if (Cond.empty()) { 1061 if (TBB != IBB) // ubr 1062 continue; 1063 } else { 1064 if (TBB != IBB && IBB != PredNextBB) // cbr 1065 continue; 1066 } 1067 } 1068 1069 // Remove the unconditional branch at the end, if any. 1070 if (TBB && (Cond.empty() || FBB)) { 1071 DebugLoc dl; // FIXME: this is nowhere 1072 TII->RemoveBranch(*PBB); 1073 if (!Cond.empty()) 1074 // reinsert conditional branch only, for now 1075 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr, 1076 NewCond, dl); 1077 } 1078 1079 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(*PBB), PBB)); 1080 } 1081 } 1082 1083 // If this is a large problem, avoid visiting the same basic blocks multiple 1084 // times. 1085 if (MergePotentials.size() == TailMergeThreshold) 1086 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 1087 TriedMerging.insert(MergePotentials[i].getBlock()); 1088 1089 if (MergePotentials.size() >= 2) 1090 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 1091 1092 // Reinsert an unconditional branch if needed. The 1 below can occur as a 1093 // result of removing blocks in TryTailMergeBlocks. 1094 PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks 1095 if (MergePotentials.size() == 1 && 1096 MergePotentials.begin()->getBlock() != PredBB) 1097 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 1098 } 1099 1100 return MadeChange; 1101 } 1102 1103 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) { 1104 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size()); 1105 BlockFrequency AccumulatedMBBFreq; 1106 1107 // Aggregate edge frequency of successor edge j: 1108 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)), 1109 // where bb is a basic block that is in SameTails. 1110 for (const auto &Src : SameTails) { 1111 const MachineBasicBlock *SrcMBB = Src.getBlock(); 1112 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB); 1113 AccumulatedMBBFreq += BlockFreq; 1114 1115 // It is not necessary to recompute edge weights if TailBB has less than two 1116 // successors. 1117 if (TailMBB.succ_size() <= 1) 1118 continue; 1119 1120 auto EdgeFreq = EdgeFreqLs.begin(); 1121 1122 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end(); 1123 SuccI != SuccE; ++SuccI, ++EdgeFreq) 1124 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI); 1125 } 1126 1127 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq); 1128 1129 if (TailMBB.succ_size() <= 1) 1130 return; 1131 1132 auto SumEdgeFreq = 1133 std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0)) 1134 .getFrequency(); 1135 auto EdgeFreq = EdgeFreqLs.begin(); 1136 1137 if (SumEdgeFreq > 0) { 1138 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end(); 1139 SuccI != SuccE; ++SuccI, ++EdgeFreq) { 1140 auto Prob = BranchProbability::getBranchProbability( 1141 EdgeFreq->getFrequency(), SumEdgeFreq); 1142 TailMBB.setSuccProbability(SuccI, Prob); 1143 } 1144 } 1145 } 1146 1147 //===----------------------------------------------------------------------===// 1148 // Branch Optimization 1149 //===----------------------------------------------------------------------===// 1150 1151 bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 1152 bool MadeChange = false; 1153 1154 // Make sure blocks are numbered in order 1155 MF.RenumberBlocks(); 1156 // Renumbering blocks alters funclet membership, recalculate it. 1157 FuncletMembership = getFuncletMembership(MF); 1158 1159 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end(); 1160 I != E; ) { 1161 MachineBasicBlock *MBB = &*I++; 1162 MadeChange |= OptimizeBlock(MBB); 1163 1164 // If it is dead, remove it. 1165 if (MBB->pred_empty()) { 1166 RemoveDeadBlock(MBB); 1167 MadeChange = true; 1168 ++NumDeadBlocks; 1169 } 1170 } 1171 1172 return MadeChange; 1173 } 1174 1175 // Blocks should be considered empty if they contain only debug info; 1176 // else the debug info would affect codegen. 1177 static bool IsEmptyBlock(MachineBasicBlock *MBB) { 1178 return MBB->getFirstNonDebugInstr() == MBB->end(); 1179 } 1180 1181 // Blocks with only debug info and branches should be considered the same 1182 // as blocks with only branches. 1183 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { 1184 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr(); 1185 assert(I != MBB->end() && "empty block!"); 1186 return I->isBranch(); 1187 } 1188 1189 /// IsBetterFallthrough - Return true if it would be clearly better to 1190 /// fall-through to MBB1 than to fall through into MBB2. This has to return 1191 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 1192 /// result in infinite loops. 1193 static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 1194 MachineBasicBlock *MBB2) { 1195 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 1196 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 1197 // optimize branches that branch to either a return block or an assert block 1198 // into a fallthrough to the return. 1199 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr(); 1200 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr(); 1201 if (MBB1I == MBB1->end() || MBB2I == MBB2->end()) 1202 return false; 1203 1204 // If there is a clear successor ordering we make sure that one block 1205 // will fall through to the next 1206 if (MBB1->isSuccessor(MBB2)) return true; 1207 if (MBB2->isSuccessor(MBB1)) return false; 1208 1209 return MBB2I->isCall() && !MBB1I->isCall(); 1210 } 1211 1212 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch 1213 /// instructions on the block. 1214 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) { 1215 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); 1216 if (I != MBB.end() && I->isBranch()) 1217 return I->getDebugLoc(); 1218 return DebugLoc(); 1219 } 1220 1221 /// OptimizeBlock - Analyze and optimize control flow related to the specified 1222 /// block. This is never called on the entry block. 1223 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1224 bool MadeChange = false; 1225 MachineFunction &MF = *MBB->getParent(); 1226 ReoptimizeBlock: 1227 1228 MachineFunction::iterator FallThrough = MBB->getIterator(); 1229 ++FallThrough; 1230 1231 // Make sure MBB and FallThrough belong to the same funclet. 1232 bool SameFunclet = true; 1233 if (!FuncletMembership.empty() && FallThrough != MF.end()) { 1234 auto MBBFunclet = FuncletMembership.find(MBB); 1235 assert(MBBFunclet != FuncletMembership.end()); 1236 auto FallThroughFunclet = FuncletMembership.find(&*FallThrough); 1237 assert(FallThroughFunclet != FuncletMembership.end()); 1238 SameFunclet = MBBFunclet->second == FallThroughFunclet->second; 1239 } 1240 1241 // If this block is empty, make everyone use its fall-through, not the block 1242 // explicitly. Landing pads should not do this since the landing-pad table 1243 // points to this block. Blocks with their addresses taken shouldn't be 1244 // optimized away. 1245 if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() && 1246 SameFunclet) { 1247 // Dead block? Leave for cleanup later. 1248 if (MBB->pred_empty()) return MadeChange; 1249 1250 if (FallThrough == MF.end()) { 1251 // TODO: Simplify preds to not branch here if possible! 1252 } else if (FallThrough->isEHPad()) { 1253 // Don't rewrite to a landing pad fallthough. That could lead to the case 1254 // where a BB jumps to more than one landing pad. 1255 // TODO: Is it ever worth rewriting predecessors which don't already 1256 // jump to a landing pad, and so can safely jump to the fallthrough? 1257 } else if (MBB->isSuccessor(&*FallThrough)) { 1258 // Rewrite all predecessors of the old block to go to the fallthrough 1259 // instead. 1260 while (!MBB->pred_empty()) { 1261 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1262 Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough); 1263 } 1264 // If MBB was the target of a jump table, update jump tables to go to the 1265 // fallthrough instead. 1266 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1267 MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough); 1268 MadeChange = true; 1269 } 1270 return MadeChange; 1271 } 1272 1273 // Check to see if we can simplify the terminator of the block before this 1274 // one. 1275 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB)); 1276 1277 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr; 1278 SmallVector<MachineOperand, 4> PriorCond; 1279 bool PriorUnAnalyzable = 1280 TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1281 if (!PriorUnAnalyzable) { 1282 // If the CFG for the prior block has extra edges, remove them. 1283 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1284 !PriorCond.empty()); 1285 1286 // If the previous branch is conditional and both conditions go to the same 1287 // destination, remove the branch, replacing it with an unconditional one or 1288 // a fall-through. 1289 if (PriorTBB && PriorTBB == PriorFBB) { 1290 DebugLoc dl = getBranchDebugLoc(PrevBB); 1291 TII->RemoveBranch(PrevBB); 1292 PriorCond.clear(); 1293 if (PriorTBB != MBB) 1294 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl); 1295 MadeChange = true; 1296 ++NumBranchOpts; 1297 goto ReoptimizeBlock; 1298 } 1299 1300 // If the previous block unconditionally falls through to this block and 1301 // this block has no other predecessors, move the contents of this block 1302 // into the prior block. This doesn't usually happen when SimplifyCFG 1303 // has been used, but it can happen if tail merging splits a fall-through 1304 // predecessor of a block. 1305 // This has to check PrevBB->succ_size() because EH edges are ignored by 1306 // AnalyzeBranch. 1307 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1308 PrevBB.succ_size() == 1 && 1309 !MBB->hasAddressTaken() && !MBB->isEHPad()) { 1310 DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1311 << "From MBB: " << *MBB); 1312 // Remove redundant DBG_VALUEs first. 1313 if (PrevBB.begin() != PrevBB.end()) { 1314 MachineBasicBlock::iterator PrevBBIter = PrevBB.end(); 1315 --PrevBBIter; 1316 MachineBasicBlock::iterator MBBIter = MBB->begin(); 1317 // Check if DBG_VALUE at the end of PrevBB is identical to the 1318 // DBG_VALUE at the beginning of MBB. 1319 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end() 1320 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) { 1321 if (!MBBIter->isIdenticalTo(*PrevBBIter)) 1322 break; 1323 MachineInstr &DuplicateDbg = *MBBIter; 1324 ++MBBIter; -- PrevBBIter; 1325 DuplicateDbg.eraseFromParent(); 1326 } 1327 } 1328 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1329 PrevBB.removeSuccessor(PrevBB.succ_begin()); 1330 assert(PrevBB.succ_empty()); 1331 PrevBB.transferSuccessors(MBB); 1332 MadeChange = true; 1333 return MadeChange; 1334 } 1335 1336 // If the previous branch *only* branches to *this* block (conditional or 1337 // not) remove the branch. 1338 if (PriorTBB == MBB && !PriorFBB) { 1339 TII->RemoveBranch(PrevBB); 1340 MadeChange = true; 1341 ++NumBranchOpts; 1342 goto ReoptimizeBlock; 1343 } 1344 1345 // If the prior block branches somewhere else on the condition and here if 1346 // the condition is false, remove the uncond second branch. 1347 if (PriorFBB == MBB) { 1348 DebugLoc dl = getBranchDebugLoc(PrevBB); 1349 TII->RemoveBranch(PrevBB); 1350 TII->InsertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl); 1351 MadeChange = true; 1352 ++NumBranchOpts; 1353 goto ReoptimizeBlock; 1354 } 1355 1356 // If the prior block branches here on true and somewhere else on false, and 1357 // if the branch condition is reversible, reverse the branch to create a 1358 // fall-through. 1359 if (PriorTBB == MBB) { 1360 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1361 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1362 DebugLoc dl = getBranchDebugLoc(PrevBB); 1363 TII->RemoveBranch(PrevBB); 1364 TII->InsertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl); 1365 MadeChange = true; 1366 ++NumBranchOpts; 1367 goto ReoptimizeBlock; 1368 } 1369 } 1370 1371 // If this block has no successors (e.g. it is a return block or ends with 1372 // a call to a no-return function like abort or __cxa_throw) and if the pred 1373 // falls through into this block, and if it would otherwise fall through 1374 // into the block after this, move this block to the end of the function. 1375 // 1376 // We consider it more likely that execution will stay in the function (e.g. 1377 // due to loops) than it is to exit it. This asserts in loops etc, moving 1378 // the assert condition out of the loop body. 1379 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB && 1380 MachineFunction::iterator(PriorTBB) == FallThrough && 1381 !MBB->canFallThrough()) { 1382 bool DoTransform = true; 1383 1384 // We have to be careful that the succs of PredBB aren't both no-successor 1385 // blocks. If neither have successors and if PredBB is the second from 1386 // last block in the function, we'd just keep swapping the two blocks for 1387 // last. Only do the swap if one is clearly better to fall through than 1388 // the other. 1389 if (FallThrough == --MF.end() && 1390 !IsBetterFallthrough(PriorTBB, MBB)) 1391 DoTransform = false; 1392 1393 if (DoTransform) { 1394 // Reverse the branch so we will fall through on the previous true cond. 1395 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1396 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1397 DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1398 << "To make fallthrough to: " << *PriorTBB << "\n"); 1399 1400 DebugLoc dl = getBranchDebugLoc(PrevBB); 1401 TII->RemoveBranch(PrevBB); 1402 TII->InsertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl); 1403 1404 // Move this block to the end of the function. 1405 MBB->moveAfter(&MF.back()); 1406 MadeChange = true; 1407 ++NumBranchOpts; 1408 return MadeChange; 1409 } 1410 } 1411 } 1412 } 1413 1414 // Analyze the branch in the current block. 1415 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr; 1416 SmallVector<MachineOperand, 4> CurCond; 1417 bool CurUnAnalyzable = 1418 TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1419 if (!CurUnAnalyzable) { 1420 // If the CFG for the prior block has extra edges, remove them. 1421 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1422 1423 // If this is a two-way branch, and the FBB branches to this block, reverse 1424 // the condition so the single-basic-block loop is faster. Instead of: 1425 // Loop: xxx; jcc Out; jmp Loop 1426 // we want: 1427 // Loop: xxx; jncc Loop; jmp Out 1428 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1429 SmallVector<MachineOperand, 4> NewCond(CurCond); 1430 if (!TII->ReverseBranchCondition(NewCond)) { 1431 DebugLoc dl = getBranchDebugLoc(*MBB); 1432 TII->RemoveBranch(*MBB); 1433 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); 1434 MadeChange = true; 1435 ++NumBranchOpts; 1436 goto ReoptimizeBlock; 1437 } 1438 } 1439 1440 // If this branch is the only thing in its block, see if we can forward 1441 // other blocks across it. 1442 if (CurTBB && CurCond.empty() && !CurFBB && 1443 IsBranchOnlyBlock(MBB) && CurTBB != MBB && 1444 !MBB->hasAddressTaken() && !MBB->isEHPad()) { 1445 DebugLoc dl = getBranchDebugLoc(*MBB); 1446 // This block may contain just an unconditional branch. Because there can 1447 // be 'non-branch terminators' in the block, try removing the branch and 1448 // then seeing if the block is empty. 1449 TII->RemoveBranch(*MBB); 1450 // If the only things remaining in the block are debug info, remove these 1451 // as well, so this will behave the same as an empty block in non-debug 1452 // mode. 1453 if (IsEmptyBlock(MBB)) { 1454 // Make the block empty, losing the debug info (we could probably 1455 // improve this in some cases.) 1456 MBB->erase(MBB->begin(), MBB->end()); 1457 } 1458 // If this block is just an unconditional branch to CurTBB, we can 1459 // usually completely eliminate the block. The only case we cannot 1460 // completely eliminate the block is when the block before this one 1461 // falls through into MBB and we can't understand the prior block's branch 1462 // condition. 1463 if (MBB->empty()) { 1464 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1465 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1466 !PrevBB.isSuccessor(MBB)) { 1467 // If the prior block falls through into us, turn it into an 1468 // explicit branch to us to make updates simpler. 1469 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1470 PriorTBB != MBB && PriorFBB != MBB) { 1471 if (!PriorTBB) { 1472 assert(PriorCond.empty() && !PriorFBB && 1473 "Bad branch analysis"); 1474 PriorTBB = MBB; 1475 } else { 1476 assert(!PriorFBB && "Machine CFG out of date!"); 1477 PriorFBB = MBB; 1478 } 1479 DebugLoc pdl = getBranchDebugLoc(PrevBB); 1480 TII->RemoveBranch(PrevBB); 1481 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl); 1482 } 1483 1484 // Iterate through all the predecessors, revectoring each in-turn. 1485 size_t PI = 0; 1486 bool DidChange = false; 1487 bool HasBranchToSelf = false; 1488 while(PI != MBB->pred_size()) { 1489 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1490 if (PMBB == MBB) { 1491 // If this block has an uncond branch to itself, leave it. 1492 ++PI; 1493 HasBranchToSelf = true; 1494 } else { 1495 DidChange = true; 1496 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1497 // If this change resulted in PMBB ending in a conditional 1498 // branch where both conditions go to the same destination, 1499 // change this to an unconditional branch (and fix the CFG). 1500 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr; 1501 SmallVector<MachineOperand, 4> NewCurCond; 1502 bool NewCurUnAnalyzable = TII->analyzeBranch( 1503 *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true); 1504 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1505 DebugLoc pdl = getBranchDebugLoc(*PMBB); 1506 TII->RemoveBranch(*PMBB); 1507 NewCurCond.clear(); 1508 TII->InsertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl); 1509 MadeChange = true; 1510 ++NumBranchOpts; 1511 PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false); 1512 } 1513 } 1514 } 1515 1516 // Change any jumptables to go to the new MBB. 1517 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1518 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); 1519 if (DidChange) { 1520 ++NumBranchOpts; 1521 MadeChange = true; 1522 if (!HasBranchToSelf) return MadeChange; 1523 } 1524 } 1525 } 1526 1527 // Add the branch back if the block is more than just an uncond branch. 1528 TII->InsertBranch(*MBB, CurTBB, nullptr, CurCond, dl); 1529 } 1530 } 1531 1532 // If the prior block doesn't fall through into this block, and if this 1533 // block doesn't fall through into some other block, see if we can find a 1534 // place to move this block where a fall-through will happen. 1535 if (!PrevBB.canFallThrough()) { 1536 1537 // Now we know that there was no fall-through into this block, check to 1538 // see if it has a fall-through into its successor. 1539 bool CurFallsThru = MBB->canFallThrough(); 1540 1541 if (!MBB->isEHPad()) { 1542 // Check all the predecessors of this block. If one of them has no fall 1543 // throughs, move this block right after it. 1544 for (MachineBasicBlock *PredBB : MBB->predecessors()) { 1545 // Analyze the branch at the end of the pred. 1546 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr; 1547 SmallVector<MachineOperand, 4> PredCond; 1548 if (PredBB != MBB && !PredBB->canFallThrough() && 1549 !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) && 1550 (!CurFallsThru || !CurTBB || !CurFBB) && 1551 (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1552 // If the current block doesn't fall through, just move it. 1553 // If the current block can fall through and does not end with a 1554 // conditional branch, we need to append an unconditional jump to 1555 // the (current) next block. To avoid a possible compile-time 1556 // infinite loop, move blocks only backward in this case. 1557 // Also, if there are already 2 branches here, we cannot add a third; 1558 // this means we have the case 1559 // Bcc next 1560 // B elsewhere 1561 // next: 1562 if (CurFallsThru) { 1563 MachineBasicBlock *NextBB = &*std::next(MBB->getIterator()); 1564 CurCond.clear(); 1565 TII->InsertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc()); 1566 } 1567 MBB->moveAfter(PredBB); 1568 MadeChange = true; 1569 goto ReoptimizeBlock; 1570 } 1571 } 1572 } 1573 1574 if (!CurFallsThru) { 1575 // Check all successors to see if we can move this block before it. 1576 for (MachineBasicBlock *SuccBB : MBB->successors()) { 1577 // Analyze the branch at the end of the block before the succ. 1578 MachineFunction::iterator SuccPrev = --SuccBB->getIterator(); 1579 1580 // If this block doesn't already fall-through to that successor, and if 1581 // the succ doesn't already have a block that can fall through into it, 1582 // and if the successor isn't an EH destination, we can arrange for the 1583 // fallthrough to happen. 1584 if (SuccBB != MBB && &*SuccPrev != MBB && 1585 !SuccPrev->canFallThrough() && !CurUnAnalyzable && 1586 !SuccBB->isEHPad()) { 1587 MBB->moveBefore(SuccBB); 1588 MadeChange = true; 1589 goto ReoptimizeBlock; 1590 } 1591 } 1592 1593 // Okay, there is no really great place to put this block. If, however, 1594 // the block before this one would be a fall-through if this block were 1595 // removed, move this block to the end of the function. 1596 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr; 1597 SmallVector<MachineOperand, 4> PrevCond; 1598 // We're looking for cases where PrevBB could possibly fall through to 1599 // FallThrough, but if FallThrough is an EH pad that wouldn't be useful 1600 // so here we skip over any EH pads so we might have a chance to find 1601 // a branch target from PrevBB. 1602 while (FallThrough != MF.end() && FallThrough->isEHPad()) 1603 ++FallThrough; 1604 // Now check to see if the current block is sitting between PrevBB and 1605 // a block to which it could fall through. 1606 if (FallThrough != MF.end() && 1607 !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1608 PrevBB.isSuccessor(&*FallThrough)) { 1609 MBB->moveAfter(&MF.back()); 1610 MadeChange = true; 1611 return MadeChange; 1612 } 1613 } 1614 } 1615 1616 return MadeChange; 1617 } 1618 1619 //===----------------------------------------------------------------------===// 1620 // Hoist Common Code 1621 //===----------------------------------------------------------------------===// 1622 1623 /// HoistCommonCode - Hoist common instruction sequences at the start of basic 1624 /// blocks to their common predecessor. 1625 bool BranchFolder::HoistCommonCode(MachineFunction &MF) { 1626 bool MadeChange = false; 1627 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) { 1628 MachineBasicBlock *MBB = &*I++; 1629 MadeChange |= HoistCommonCodeInSuccs(MBB); 1630 } 1631 1632 return MadeChange; 1633 } 1634 1635 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 1636 /// its 'true' successor. 1637 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 1638 MachineBasicBlock *TrueBB) { 1639 for (MachineBasicBlock *SuccBB : BB->successors()) 1640 if (SuccBB != TrueBB) 1641 return SuccBB; 1642 return nullptr; 1643 } 1644 1645 template <class Container> 1646 static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI, 1647 Container &Set) { 1648 if (TargetRegisterInfo::isPhysicalRegister(Reg)) { 1649 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1650 Set.insert(*AI); 1651 } else { 1652 Set.insert(Reg); 1653 } 1654 } 1655 1656 /// findHoistingInsertPosAndDeps - Find the location to move common instructions 1657 /// in successors to. The location is usually just before the terminator, 1658 /// however if the terminator is a conditional branch and its previous 1659 /// instruction is the flag setting instruction, the previous instruction is 1660 /// the preferred location. This function also gathers uses and defs of the 1661 /// instructions from the insertion point to the end of the block. The data is 1662 /// used by HoistCommonCodeInSuccs to ensure safety. 1663 static 1664 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB, 1665 const TargetInstrInfo *TII, 1666 const TargetRegisterInfo *TRI, 1667 SmallSet<unsigned,4> &Uses, 1668 SmallSet<unsigned,4> &Defs) { 1669 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator(); 1670 if (!TII->isUnpredicatedTerminator(*Loc)) 1671 return MBB->end(); 1672 1673 for (const MachineOperand &MO : Loc->operands()) { 1674 if (!MO.isReg()) 1675 continue; 1676 unsigned Reg = MO.getReg(); 1677 if (!Reg) 1678 continue; 1679 if (MO.isUse()) { 1680 addRegAndItsAliases(Reg, TRI, Uses); 1681 } else { 1682 if (!MO.isDead()) 1683 // Don't try to hoist code in the rare case the terminator defines a 1684 // register that is later used. 1685 return MBB->end(); 1686 1687 // If the terminator defines a register, make sure we don't hoist 1688 // the instruction whose def might be clobbered by the terminator. 1689 addRegAndItsAliases(Reg, TRI, Defs); 1690 } 1691 } 1692 1693 if (Uses.empty()) 1694 return Loc; 1695 if (Loc == MBB->begin()) 1696 return MBB->end(); 1697 1698 // The terminator is probably a conditional branch, try not to separate the 1699 // branch from condition setting instruction. 1700 MachineBasicBlock::iterator PI = Loc; 1701 --PI; 1702 while (PI != MBB->begin() && PI->isDebugValue()) 1703 --PI; 1704 1705 bool IsDef = false; 1706 for (const MachineOperand &MO : PI->operands()) { 1707 // If PI has a regmask operand, it is probably a call. Separate away. 1708 if (MO.isRegMask()) 1709 return Loc; 1710 if (!MO.isReg() || MO.isUse()) 1711 continue; 1712 unsigned Reg = MO.getReg(); 1713 if (!Reg) 1714 continue; 1715 if (Uses.count(Reg)) { 1716 IsDef = true; 1717 break; 1718 } 1719 } 1720 if (!IsDef) 1721 // The condition setting instruction is not just before the conditional 1722 // branch. 1723 return Loc; 1724 1725 // Be conservative, don't insert instruction above something that may have 1726 // side-effects. And since it's potentially bad to separate flag setting 1727 // instruction from the conditional branch, just abort the optimization 1728 // completely. 1729 // Also avoid moving code above predicated instruction since it's hard to 1730 // reason about register liveness with predicated instruction. 1731 bool DontMoveAcrossStore = true; 1732 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI)) 1733 return MBB->end(); 1734 1735 1736 // Find out what registers are live. Note this routine is ignoring other live 1737 // registers which are only used by instructions in successor blocks. 1738 for (const MachineOperand &MO : PI->operands()) { 1739 if (!MO.isReg()) 1740 continue; 1741 unsigned Reg = MO.getReg(); 1742 if (!Reg) 1743 continue; 1744 if (MO.isUse()) { 1745 addRegAndItsAliases(Reg, TRI, Uses); 1746 } else { 1747 if (Uses.erase(Reg)) { 1748 if (TargetRegisterInfo::isPhysicalRegister(Reg)) { 1749 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) 1750 Uses.erase(*SubRegs); // Use sub-registers to be conservative 1751 } 1752 } 1753 addRegAndItsAliases(Reg, TRI, Defs); 1754 } 1755 } 1756 1757 return PI; 1758 } 1759 1760 /// HoistCommonCodeInSuccs - If the successors of MBB has common instruction 1761 /// sequence at the start of the function, move the instructions before MBB 1762 /// terminator if it's legal. 1763 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) { 1764 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 1765 SmallVector<MachineOperand, 4> Cond; 1766 if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty()) 1767 return false; 1768 1769 if (!FBB) FBB = findFalseBlock(MBB, TBB); 1770 if (!FBB) 1771 // Malformed bcc? True and false blocks are the same? 1772 return false; 1773 1774 // Restrict the optimization to cases where MBB is the only predecessor, 1775 // it is an obvious win. 1776 if (TBB->pred_size() > 1 || FBB->pred_size() > 1) 1777 return false; 1778 1779 // Find a suitable position to hoist the common instructions to. Also figure 1780 // out which registers are used or defined by instructions from the insertion 1781 // point to the end of the block. 1782 SmallSet<unsigned, 4> Uses, Defs; 1783 MachineBasicBlock::iterator Loc = 1784 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs); 1785 if (Loc == MBB->end()) 1786 return false; 1787 1788 bool HasDups = false; 1789 SmallVector<unsigned, 4> LocalDefs; 1790 SmallSet<unsigned, 4> LocalDefsSet; 1791 MachineBasicBlock::iterator TIB = TBB->begin(); 1792 MachineBasicBlock::iterator FIB = FBB->begin(); 1793 MachineBasicBlock::iterator TIE = TBB->end(); 1794 MachineBasicBlock::iterator FIE = FBB->end(); 1795 while (TIB != TIE && FIB != FIE) { 1796 // Skip dbg_value instructions. These do not count. 1797 if (TIB->isDebugValue()) { 1798 while (TIB != TIE && TIB->isDebugValue()) 1799 ++TIB; 1800 if (TIB == TIE) 1801 break; 1802 } 1803 if (FIB->isDebugValue()) { 1804 while (FIB != FIE && FIB->isDebugValue()) 1805 ++FIB; 1806 if (FIB == FIE) 1807 break; 1808 } 1809 if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead)) 1810 break; 1811 1812 if (TII->isPredicated(*TIB)) 1813 // Hard to reason about register liveness with predicated instruction. 1814 break; 1815 1816 bool IsSafe = true; 1817 for (MachineOperand &MO : TIB->operands()) { 1818 // Don't attempt to hoist instructions with register masks. 1819 if (MO.isRegMask()) { 1820 IsSafe = false; 1821 break; 1822 } 1823 if (!MO.isReg()) 1824 continue; 1825 unsigned Reg = MO.getReg(); 1826 if (!Reg) 1827 continue; 1828 if (MO.isDef()) { 1829 if (Uses.count(Reg)) { 1830 // Avoid clobbering a register that's used by the instruction at 1831 // the point of insertion. 1832 IsSafe = false; 1833 break; 1834 } 1835 1836 if (Defs.count(Reg) && !MO.isDead()) { 1837 // Don't hoist the instruction if the def would be clobber by the 1838 // instruction at the point insertion. FIXME: This is overly 1839 // conservative. It should be possible to hoist the instructions 1840 // in BB2 in the following example: 1841 // BB1: 1842 // r1, eflag = op1 r2, r3 1843 // brcc eflag 1844 // 1845 // BB2: 1846 // r1 = op2, ... 1847 // = op3, r1<kill> 1848 IsSafe = false; 1849 break; 1850 } 1851 } else if (!LocalDefsSet.count(Reg)) { 1852 if (Defs.count(Reg)) { 1853 // Use is defined by the instruction at the point of insertion. 1854 IsSafe = false; 1855 break; 1856 } 1857 1858 if (MO.isKill() && Uses.count(Reg)) 1859 // Kills a register that's read by the instruction at the point of 1860 // insertion. Remove the kill marker. 1861 MO.setIsKill(false); 1862 } 1863 } 1864 if (!IsSafe) 1865 break; 1866 1867 bool DontMoveAcrossStore = true; 1868 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore)) 1869 break; 1870 1871 // Remove kills from LocalDefsSet, these registers had short live ranges. 1872 for (const MachineOperand &MO : TIB->operands()) { 1873 if (!MO.isReg() || !MO.isUse() || !MO.isKill()) 1874 continue; 1875 unsigned Reg = MO.getReg(); 1876 if (!Reg || !LocalDefsSet.count(Reg)) 1877 continue; 1878 if (TargetRegisterInfo::isPhysicalRegister(Reg)) { 1879 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1880 LocalDefsSet.erase(*AI); 1881 } else { 1882 LocalDefsSet.erase(Reg); 1883 } 1884 } 1885 1886 // Track local defs so we can update liveins. 1887 for (const MachineOperand &MO : TIB->operands()) { 1888 if (!MO.isReg() || !MO.isDef() || MO.isDead()) 1889 continue; 1890 unsigned Reg = MO.getReg(); 1891 if (!Reg || TargetRegisterInfo::isVirtualRegister(Reg)) 1892 continue; 1893 LocalDefs.push_back(Reg); 1894 addRegAndItsAliases(Reg, TRI, LocalDefsSet); 1895 } 1896 1897 HasDups = true; 1898 ++TIB; 1899 ++FIB; 1900 } 1901 1902 if (!HasDups) 1903 return false; 1904 1905 MBB->splice(Loc, TBB, TBB->begin(), TIB); 1906 FBB->erase(FBB->begin(), FIB); 1907 1908 // Update livein's. 1909 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) { 1910 unsigned Def = LocalDefs[i]; 1911 if (LocalDefsSet.count(Def)) { 1912 TBB->addLiveIn(Def); 1913 FBB->addLiveIn(Def); 1914 } 1915 } 1916 1917 ++NumHoist; 1918 return true; 1919 } 1920