1 //===- BranchFolding.cpp - Fold machine code branch instructions ----------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This pass forwards branches to unconditional branches to make them branch 10 // directly to the target block. This pass often results in dead MBB's, which 11 // it then removes. 12 // 13 // Note that this pass must be run after register allocation, it cannot handle 14 // SSA form. It also must handle virtual registers for targets that emit virtual 15 // ISA (e.g. NVPTX). 16 // 17 //===----------------------------------------------------------------------===// 18 19 #include "BranchFolding.h" 20 #include "llvm/ADT/BitVector.h" 21 #include "llvm/ADT/DenseMap.h" 22 #include "llvm/ADT/STLExtras.h" 23 #include "llvm/ADT/SmallPtrSet.h" 24 #include "llvm/ADT/SmallSet.h" 25 #include "llvm/ADT/SmallVector.h" 26 #include "llvm/ADT/Statistic.h" 27 #include "llvm/Analysis/ProfileSummaryInfo.h" 28 #include "llvm/CodeGen/Analysis.h" 29 #include "llvm/CodeGen/LivePhysRegs.h" 30 #include "llvm/CodeGen/MachineBasicBlock.h" 31 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 32 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 33 #include "llvm/CodeGen/MachineFunction.h" 34 #include "llvm/CodeGen/MachineFunctionPass.h" 35 #include "llvm/CodeGen/MachineInstr.h" 36 #include "llvm/CodeGen/MachineInstrBuilder.h" 37 #include "llvm/CodeGen/MachineJumpTableInfo.h" 38 #include "llvm/CodeGen/MachineLoopInfo.h" 39 #include "llvm/CodeGen/MachineModuleInfo.h" 40 #include "llvm/CodeGen/MachineOperand.h" 41 #include "llvm/CodeGen/MachineRegisterInfo.h" 42 #include "llvm/CodeGen/MachineSizeOpts.h" 43 #include "llvm/CodeGen/MBFIWrapper.h" 44 #include "llvm/CodeGen/TargetInstrInfo.h" 45 #include "llvm/CodeGen/TargetOpcodes.h" 46 #include "llvm/CodeGen/TargetPassConfig.h" 47 #include "llvm/CodeGen/TargetRegisterInfo.h" 48 #include "llvm/CodeGen/TargetSubtargetInfo.h" 49 #include "llvm/IR/DebugInfoMetadata.h" 50 #include "llvm/IR/DebugLoc.h" 51 #include "llvm/IR/Function.h" 52 #include "llvm/InitializePasses.h" 53 #include "llvm/MC/LaneBitmask.h" 54 #include "llvm/MC/MCRegisterInfo.h" 55 #include "llvm/Pass.h" 56 #include "llvm/Support/BlockFrequency.h" 57 #include "llvm/Support/BranchProbability.h" 58 #include "llvm/Support/CommandLine.h" 59 #include "llvm/Support/Debug.h" 60 #include "llvm/Support/ErrorHandling.h" 61 #include "llvm/Support/raw_ostream.h" 62 #include "llvm/Target/TargetMachine.h" 63 #include <cassert> 64 #include <cstddef> 65 #include <iterator> 66 #include <numeric> 67 #include <vector> 68 69 using namespace llvm; 70 71 #define DEBUG_TYPE "branch-folder" 72 73 STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 74 STATISTIC(NumBranchOpts, "Number of branches optimized"); 75 STATISTIC(NumTailMerge , "Number of block tails merged"); 76 STATISTIC(NumHoist , "Number of times common instructions are hoisted"); 77 STATISTIC(NumTailCalls, "Number of tail calls optimized"); 78 79 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 80 cl::init(cl::BOU_UNSET), cl::Hidden); 81 82 // Throttle for huge numbers of predecessors (compile speed problems) 83 static cl::opt<unsigned> 84 TailMergeThreshold("tail-merge-threshold", 85 cl::desc("Max number of predecessors to consider tail merging"), 86 cl::init(150), cl::Hidden); 87 88 // Heuristic for tail merging (and, inversely, tail duplication). 89 // TODO: This should be replaced with a target query. 90 static cl::opt<unsigned> 91 TailMergeSize("tail-merge-size", 92 cl::desc("Min number of instructions to consider tail merging"), 93 cl::init(3), cl::Hidden); 94 95 namespace { 96 97 /// BranchFolderPass - Wrap branch folder in a machine function pass. 98 class BranchFolderPass : public MachineFunctionPass { 99 public: 100 static char ID; 101 102 explicit BranchFolderPass(): MachineFunctionPass(ID) {} 103 104 bool runOnMachineFunction(MachineFunction &MF) override; 105 106 void getAnalysisUsage(AnalysisUsage &AU) const override { 107 AU.addRequired<MachineBlockFrequencyInfo>(); 108 AU.addRequired<MachineBranchProbabilityInfo>(); 109 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 110 AU.addRequired<TargetPassConfig>(); 111 MachineFunctionPass::getAnalysisUsage(AU); 112 } 113 }; 114 115 } // end anonymous namespace 116 117 char BranchFolderPass::ID = 0; 118 119 char &llvm::BranchFolderPassID = BranchFolderPass::ID; 120 121 INITIALIZE_PASS(BranchFolderPass, DEBUG_TYPE, 122 "Control Flow Optimizer", false, false) 123 124 bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { 125 if (skipFunction(MF.getFunction())) 126 return false; 127 128 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>(); 129 // TailMerge can create jump into if branches that make CFG irreducible for 130 // HW that requires structurized CFG. 131 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() && 132 PassConfig->getEnableTailMerge(); 133 MBFIWrapper MBBFreqInfo( 134 getAnalysis<MachineBlockFrequencyInfo>()); 135 BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo, 136 getAnalysis<MachineBranchProbabilityInfo>(), 137 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI()); 138 return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(), 139 MF.getSubtarget().getRegisterInfo()); 140 } 141 142 BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist, 143 MBFIWrapper &FreqInfo, 144 const MachineBranchProbabilityInfo &ProbInfo, 145 ProfileSummaryInfo *PSI, 146 unsigned MinTailLength) 147 : EnableHoistCommonCode(CommonHoist), MinCommonTailLength(MinTailLength), 148 MBBFreqInfo(FreqInfo), MBPI(ProbInfo), PSI(PSI) { 149 if (MinCommonTailLength == 0) 150 MinCommonTailLength = TailMergeSize; 151 switch (FlagEnableTailMerge) { 152 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 153 case cl::BOU_TRUE: EnableTailMerge = true; break; 154 case cl::BOU_FALSE: EnableTailMerge = false; break; 155 } 156 } 157 158 void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 159 assert(MBB->pred_empty() && "MBB must be dead!"); 160 LLVM_DEBUG(dbgs() << "\nRemoving MBB: " << *MBB); 161 162 MachineFunction *MF = MBB->getParent(); 163 // drop all successors. 164 while (!MBB->succ_empty()) 165 MBB->removeSuccessor(MBB->succ_end()-1); 166 167 // Avoid matching if this pointer gets reused. 168 TriedMerging.erase(MBB); 169 170 // Update call site info. 171 std::for_each(MBB->begin(), MBB->end(), [MF](const MachineInstr &MI) { 172 if (MI.shouldUpdateCallSiteInfo()) 173 MF->eraseCallSiteInfo(&MI); 174 }); 175 // Remove the block. 176 MF->erase(MBB); 177 EHScopeMembership.erase(MBB); 178 if (MLI) 179 MLI->removeBlock(MBB); 180 } 181 182 bool BranchFolder::OptimizeFunction(MachineFunction &MF, 183 const TargetInstrInfo *tii, 184 const TargetRegisterInfo *tri, 185 MachineLoopInfo *mli, bool AfterPlacement) { 186 if (!tii) return false; 187 188 TriedMerging.clear(); 189 190 MachineRegisterInfo &MRI = MF.getRegInfo(); 191 AfterBlockPlacement = AfterPlacement; 192 TII = tii; 193 TRI = tri; 194 MLI = mli; 195 this->MRI = &MRI; 196 197 UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF); 198 if (!UpdateLiveIns) 199 MRI.invalidateLiveness(); 200 201 bool MadeChange = false; 202 203 // Recalculate EH scope membership. 204 EHScopeMembership = getEHScopeMembership(MF); 205 206 bool MadeChangeThisIteration = true; 207 while (MadeChangeThisIteration) { 208 MadeChangeThisIteration = TailMergeBlocks(MF); 209 // No need to clean up if tail merging does not change anything after the 210 // block placement. 211 if (!AfterBlockPlacement || MadeChangeThisIteration) 212 MadeChangeThisIteration |= OptimizeBranches(MF); 213 if (EnableHoistCommonCode) 214 MadeChangeThisIteration |= HoistCommonCode(MF); 215 MadeChange |= MadeChangeThisIteration; 216 } 217 218 // See if any jump tables have become dead as the code generator 219 // did its thing. 220 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 221 if (!JTI) 222 return MadeChange; 223 224 // Walk the function to find jump tables that are live. 225 BitVector JTIsLive(JTI->getJumpTables().size()); 226 for (const MachineBasicBlock &BB : MF) { 227 for (const MachineInstr &I : BB) 228 for (const MachineOperand &Op : I.operands()) { 229 if (!Op.isJTI()) continue; 230 231 // Remember that this JT is live. 232 JTIsLive.set(Op.getIndex()); 233 } 234 } 235 236 // Finally, remove dead jump tables. This happens when the 237 // indirect jump was unreachable (and thus deleted). 238 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 239 if (!JTIsLive.test(i)) { 240 JTI->RemoveJumpTable(i); 241 MadeChange = true; 242 } 243 244 return MadeChange; 245 } 246 247 //===----------------------------------------------------------------------===// 248 // Tail Merging of Blocks 249 //===----------------------------------------------------------------------===// 250 251 /// HashMachineInstr - Compute a hash value for MI and its operands. 252 static unsigned HashMachineInstr(const MachineInstr &MI) { 253 unsigned Hash = MI.getOpcode(); 254 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 255 const MachineOperand &Op = MI.getOperand(i); 256 257 // Merge in bits from the operand if easy. We can't use MachineOperand's 258 // hash_code here because it's not deterministic and we sort by hash value 259 // later. 260 unsigned OperandHash = 0; 261 switch (Op.getType()) { 262 case MachineOperand::MO_Register: 263 OperandHash = Op.getReg(); 264 break; 265 case MachineOperand::MO_Immediate: 266 OperandHash = Op.getImm(); 267 break; 268 case MachineOperand::MO_MachineBasicBlock: 269 OperandHash = Op.getMBB()->getNumber(); 270 break; 271 case MachineOperand::MO_FrameIndex: 272 case MachineOperand::MO_ConstantPoolIndex: 273 case MachineOperand::MO_JumpTableIndex: 274 OperandHash = Op.getIndex(); 275 break; 276 case MachineOperand::MO_GlobalAddress: 277 case MachineOperand::MO_ExternalSymbol: 278 // Global address / external symbol are too hard, don't bother, but do 279 // pull in the offset. 280 OperandHash = Op.getOffset(); 281 break; 282 default: 283 break; 284 } 285 286 Hash += ((OperandHash << 3) | Op.getType()) << (i & 31); 287 } 288 return Hash; 289 } 290 291 /// HashEndOfMBB - Hash the last instruction in the MBB. 292 static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) { 293 MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr(); 294 if (I == MBB.end()) 295 return 0; 296 297 return HashMachineInstr(*I); 298 } 299 300 /// Whether MI should be counted as an instruction when calculating common tail. 301 static bool countsAsInstruction(const MachineInstr &MI) { 302 return !(MI.isDebugInstr() || MI.isCFIInstruction()); 303 } 304 305 /// Iterate backwards from the given iterator \p I, towards the beginning of the 306 /// block. If a MI satisfying 'countsAsInstruction' is found, return an iterator 307 /// pointing to that MI. If no such MI is found, return the end iterator. 308 static MachineBasicBlock::iterator 309 skipBackwardPastNonInstructions(MachineBasicBlock::iterator I, 310 MachineBasicBlock *MBB) { 311 while (I != MBB->begin()) { 312 --I; 313 if (countsAsInstruction(*I)) 314 return I; 315 } 316 return MBB->end(); 317 } 318 319 /// Given two machine basic blocks, return the number of instructions they 320 /// actually have in common together at their end. If a common tail is found (at 321 /// least by one instruction), then iterators for the first shared instruction 322 /// in each block are returned as well. 323 /// 324 /// Non-instructions according to countsAsInstruction are ignored. 325 static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 326 MachineBasicBlock *MBB2, 327 MachineBasicBlock::iterator &I1, 328 MachineBasicBlock::iterator &I2) { 329 MachineBasicBlock::iterator MBBI1 = MBB1->end(); 330 MachineBasicBlock::iterator MBBI2 = MBB2->end(); 331 332 unsigned TailLen = 0; 333 while (true) { 334 MBBI1 = skipBackwardPastNonInstructions(MBBI1, MBB1); 335 MBBI2 = skipBackwardPastNonInstructions(MBBI2, MBB2); 336 if (MBBI1 == MBB1->end() || MBBI2 == MBB2->end()) 337 break; 338 if (!MBBI1->isIdenticalTo(*MBBI2) || 339 // FIXME: This check is dubious. It's used to get around a problem where 340 // people incorrectly expect inline asm directives to remain in the same 341 // relative order. This is untenable because normal compiler 342 // optimizations (like this one) may reorder and/or merge these 343 // directives. 344 MBBI1->isInlineAsm()) { 345 break; 346 } 347 if (MBBI1->getFlag(MachineInstr::NoMerge) || 348 MBBI2->getFlag(MachineInstr::NoMerge)) 349 break; 350 ++TailLen; 351 I1 = MBBI1; 352 I2 = MBBI2; 353 } 354 355 return TailLen; 356 } 357 358 void BranchFolder::replaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 359 MachineBasicBlock &NewDest) { 360 if (UpdateLiveIns) { 361 // OldInst should always point to an instruction. 362 MachineBasicBlock &OldMBB = *OldInst->getParent(); 363 LiveRegs.clear(); 364 LiveRegs.addLiveOuts(OldMBB); 365 // Move backward to the place where will insert the jump. 366 MachineBasicBlock::iterator I = OldMBB.end(); 367 do { 368 --I; 369 LiveRegs.stepBackward(*I); 370 } while (I != OldInst); 371 372 // Merging the tails may have switched some undef operand to non-undef ones. 373 // Add IMPLICIT_DEFS into OldMBB as necessary to have a definition of the 374 // register. 375 for (MachineBasicBlock::RegisterMaskPair P : NewDest.liveins()) { 376 // We computed the liveins with computeLiveIn earlier and should only see 377 // full registers: 378 assert(P.LaneMask == LaneBitmask::getAll() && 379 "Can only handle full register."); 380 MCPhysReg Reg = P.PhysReg; 381 if (!LiveRegs.available(*MRI, Reg)) 382 continue; 383 DebugLoc DL; 384 BuildMI(OldMBB, OldInst, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Reg); 385 } 386 } 387 388 TII->ReplaceTailWithBranchTo(OldInst, &NewDest); 389 ++NumTailMerge; 390 } 391 392 MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 393 MachineBasicBlock::iterator BBI1, 394 const BasicBlock *BB) { 395 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1)) 396 return nullptr; 397 398 MachineFunction &MF = *CurMBB.getParent(); 399 400 // Create the fall-through block. 401 MachineFunction::iterator MBBI = CurMBB.getIterator(); 402 MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(BB); 403 CurMBB.getParent()->insert(++MBBI, NewMBB); 404 405 // Move all the successors of this block to the specified block. 406 NewMBB->transferSuccessors(&CurMBB); 407 408 // Add an edge from CurMBB to NewMBB for the fall-through. 409 CurMBB.addSuccessor(NewMBB); 410 411 // Splice the code over. 412 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 413 414 // NewMBB belongs to the same loop as CurMBB. 415 if (MLI) 416 if (MachineLoop *ML = MLI->getLoopFor(&CurMBB)) 417 ML->addBasicBlockToLoop(NewMBB, MLI->getBase()); 418 419 // NewMBB inherits CurMBB's block frequency. 420 MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB)); 421 422 if (UpdateLiveIns) 423 computeAndAddLiveIns(LiveRegs, *NewMBB); 424 425 // Add the new block to the EH scope. 426 const auto &EHScopeI = EHScopeMembership.find(&CurMBB); 427 if (EHScopeI != EHScopeMembership.end()) { 428 auto n = EHScopeI->second; 429 EHScopeMembership[NewMBB] = n; 430 } 431 432 return NewMBB; 433 } 434 435 /// EstimateRuntime - Make a rough estimate for how long it will take to run 436 /// the specified code. 437 static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 438 MachineBasicBlock::iterator E) { 439 unsigned Time = 0; 440 for (; I != E; ++I) { 441 if (!countsAsInstruction(*I)) 442 continue; 443 if (I->isCall()) 444 Time += 10; 445 else if (I->mayLoadOrStore()) 446 Time += 2; 447 else 448 ++Time; 449 } 450 return Time; 451 } 452 453 // CurMBB needs to add an unconditional branch to SuccMBB (we removed these 454 // branches temporarily for tail merging). In the case where CurMBB ends 455 // with a conditional branch to the next block, optimize by reversing the 456 // test and conditionally branching to SuccMBB instead. 457 static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 458 const TargetInstrInfo *TII) { 459 MachineFunction *MF = CurMBB->getParent(); 460 MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB)); 461 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 462 SmallVector<MachineOperand, 4> Cond; 463 DebugLoc dl = CurMBB->findBranchDebugLoc(); 464 if (I != MF->end() && !TII->analyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 465 MachineBasicBlock *NextBB = &*I; 466 if (TBB == NextBB && !Cond.empty() && !FBB) { 467 if (!TII->reverseBranchCondition(Cond)) { 468 TII->removeBranch(*CurMBB); 469 TII->insertBranch(*CurMBB, SuccBB, nullptr, Cond, dl); 470 return; 471 } 472 } 473 } 474 TII->insertBranch(*CurMBB, SuccBB, nullptr, 475 SmallVector<MachineOperand, 0>(), dl); 476 } 477 478 bool 479 BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 480 if (getHash() < o.getHash()) 481 return true; 482 if (getHash() > o.getHash()) 483 return false; 484 if (getBlock()->getNumber() < o.getBlock()->getNumber()) 485 return true; 486 if (getBlock()->getNumber() > o.getBlock()->getNumber()) 487 return false; 488 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 489 // an object with itself. 490 #ifndef _GLIBCXX_DEBUG 491 llvm_unreachable("Predecessor appears twice"); 492 #else 493 return false; 494 #endif 495 } 496 497 /// CountTerminators - Count the number of terminators in the given 498 /// block and set I to the position of the first non-terminator, if there 499 /// is one, or MBB->end() otherwise. 500 static unsigned CountTerminators(MachineBasicBlock *MBB, 501 MachineBasicBlock::iterator &I) { 502 I = MBB->end(); 503 unsigned NumTerms = 0; 504 while (true) { 505 if (I == MBB->begin()) { 506 I = MBB->end(); 507 break; 508 } 509 --I; 510 if (!I->isTerminator()) break; 511 ++NumTerms; 512 } 513 return NumTerms; 514 } 515 516 /// A no successor, non-return block probably ends in unreachable and is cold. 517 /// Also consider a block that ends in an indirect branch to be a return block, 518 /// since many targets use plain indirect branches to return. 519 static bool blockEndsInUnreachable(const MachineBasicBlock *MBB) { 520 if (!MBB->succ_empty()) 521 return false; 522 if (MBB->empty()) 523 return true; 524 return !(MBB->back().isReturn() || MBB->back().isIndirectBranch()); 525 } 526 527 /// ProfitableToMerge - Check if two machine basic blocks have a common tail 528 /// and decide if it would be profitable to merge those tails. Return the 529 /// length of the common tail and iterators to the first common instruction 530 /// in each block. 531 /// MBB1, MBB2 The blocks to check 532 /// MinCommonTailLength Minimum size of tail block to be merged. 533 /// CommonTailLen Out parameter to record the size of the shared tail between 534 /// MBB1 and MBB2 535 /// I1, I2 Iterator references that will be changed to point to the first 536 /// instruction in the common tail shared by MBB1,MBB2 537 /// SuccBB A common successor of MBB1, MBB2 which are in a canonical form 538 /// relative to SuccBB 539 /// PredBB The layout predecessor of SuccBB, if any. 540 /// EHScopeMembership map from block to EH scope #. 541 /// AfterPlacement True if we are merging blocks after layout. Stricter 542 /// thresholds apply to prevent undoing tail-duplication. 543 static bool 544 ProfitableToMerge(MachineBasicBlock *MBB1, MachineBasicBlock *MBB2, 545 unsigned MinCommonTailLength, unsigned &CommonTailLen, 546 MachineBasicBlock::iterator &I1, 547 MachineBasicBlock::iterator &I2, MachineBasicBlock *SuccBB, 548 MachineBasicBlock *PredBB, 549 DenseMap<const MachineBasicBlock *, int> &EHScopeMembership, 550 bool AfterPlacement, 551 MBFIWrapper &MBBFreqInfo, 552 ProfileSummaryInfo *PSI) { 553 // It is never profitable to tail-merge blocks from two different EH scopes. 554 if (!EHScopeMembership.empty()) { 555 auto EHScope1 = EHScopeMembership.find(MBB1); 556 assert(EHScope1 != EHScopeMembership.end()); 557 auto EHScope2 = EHScopeMembership.find(MBB2); 558 assert(EHScope2 != EHScopeMembership.end()); 559 if (EHScope1->second != EHScope2->second) 560 return false; 561 } 562 563 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 564 if (CommonTailLen == 0) 565 return false; 566 LLVM_DEBUG(dbgs() << "Common tail length of " << printMBBReference(*MBB1) 567 << " and " << printMBBReference(*MBB2) << " is " 568 << CommonTailLen << '\n'); 569 570 // Move the iterators to the beginning of the MBB if we only got debug 571 // instructions before the tail. This is to avoid splitting a block when we 572 // only got debug instructions before the tail (to be invariant on -g). 573 if (skipDebugInstructionsForward(MBB1->begin(), MBB1->end()) == I1) 574 I1 = MBB1->begin(); 575 if (skipDebugInstructionsForward(MBB2->begin(), MBB2->end()) == I2) 576 I2 = MBB2->begin(); 577 578 bool FullBlockTail1 = I1 == MBB1->begin(); 579 bool FullBlockTail2 = I2 == MBB2->begin(); 580 581 // It's almost always profitable to merge any number of non-terminator 582 // instructions with the block that falls through into the common successor. 583 // This is true only for a single successor. For multiple successors, we are 584 // trading a conditional branch for an unconditional one. 585 // TODO: Re-visit successor size for non-layout tail merging. 586 if ((MBB1 == PredBB || MBB2 == PredBB) && 587 (!AfterPlacement || MBB1->succ_size() == 1)) { 588 MachineBasicBlock::iterator I; 589 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 590 if (CommonTailLen > NumTerms) 591 return true; 592 } 593 594 // If these are identical non-return blocks with no successors, merge them. 595 // Such blocks are typically cold calls to noreturn functions like abort, and 596 // are unlikely to become a fallthrough target after machine block placement. 597 // Tail merging these blocks is unlikely to create additional unconditional 598 // branches, and will reduce the size of this cold code. 599 if (FullBlockTail1 && FullBlockTail2 && 600 blockEndsInUnreachable(MBB1) && blockEndsInUnreachable(MBB2)) 601 return true; 602 603 // If one of the blocks can be completely merged and happens to be in 604 // a position where the other could fall through into it, merge any number 605 // of instructions, because it can be done without a branch. 606 // TODO: If the blocks are not adjacent, move one of them so that they are? 607 if (MBB1->isLayoutSuccessor(MBB2) && FullBlockTail2) 608 return true; 609 if (MBB2->isLayoutSuccessor(MBB1) && FullBlockTail1) 610 return true; 611 612 // If both blocks are identical and end in a branch, merge them unless they 613 // both have a fallthrough predecessor and successor. 614 // We can only do this after block placement because it depends on whether 615 // there are fallthroughs, and we don't know until after layout. 616 if (AfterPlacement && FullBlockTail1 && FullBlockTail2) { 617 auto BothFallThrough = [](MachineBasicBlock *MBB) { 618 if (MBB->succ_size() != 0 && !MBB->canFallThrough()) 619 return false; 620 MachineFunction::iterator I(MBB); 621 MachineFunction *MF = MBB->getParent(); 622 return (MBB != &*MF->begin()) && std::prev(I)->canFallThrough(); 623 }; 624 if (!BothFallThrough(MBB1) || !BothFallThrough(MBB2)) 625 return true; 626 } 627 628 // If both blocks have an unconditional branch temporarily stripped out, 629 // count that as an additional common instruction for the following 630 // heuristics. This heuristic is only accurate for single-succ blocks, so to 631 // make sure that during layout merging and duplicating don't crash, we check 632 // for that when merging during layout. 633 unsigned EffectiveTailLen = CommonTailLen; 634 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 635 (MBB1->succ_size() == 1 || !AfterPlacement) && 636 !MBB1->back().isBarrier() && 637 !MBB2->back().isBarrier()) 638 ++EffectiveTailLen; 639 640 // Check if the common tail is long enough to be worthwhile. 641 if (EffectiveTailLen >= MinCommonTailLength) 642 return true; 643 644 // If we are optimizing for code size, 2 instructions in common is enough if 645 // we don't have to split a block. At worst we will be introducing 1 new 646 // branch instruction, which is likely to be smaller than the 2 647 // instructions that would be deleted in the merge. 648 MachineFunction *MF = MBB1->getParent(); 649 bool OptForSize = 650 MF->getFunction().hasOptSize() || 651 (llvm::shouldOptimizeForSize(MBB1, PSI, &MBBFreqInfo) && 652 llvm::shouldOptimizeForSize(MBB2, PSI, &MBBFreqInfo)); 653 return EffectiveTailLen >= 2 && OptForSize && 654 (FullBlockTail1 || FullBlockTail2); 655 } 656 657 unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 658 unsigned MinCommonTailLength, 659 MachineBasicBlock *SuccBB, 660 MachineBasicBlock *PredBB) { 661 unsigned maxCommonTailLength = 0U; 662 SameTails.clear(); 663 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 664 MPIterator HighestMPIter = std::prev(MergePotentials.end()); 665 for (MPIterator CurMPIter = std::prev(MergePotentials.end()), 666 B = MergePotentials.begin(); 667 CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) { 668 for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) { 669 unsigned CommonTailLen; 670 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 671 MinCommonTailLength, 672 CommonTailLen, TrialBBI1, TrialBBI2, 673 SuccBB, PredBB, 674 EHScopeMembership, 675 AfterBlockPlacement, MBBFreqInfo, PSI)) { 676 if (CommonTailLen > maxCommonTailLength) { 677 SameTails.clear(); 678 maxCommonTailLength = CommonTailLen; 679 HighestMPIter = CurMPIter; 680 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 681 } 682 if (HighestMPIter == CurMPIter && 683 CommonTailLen == maxCommonTailLength) 684 SameTails.push_back(SameTailElt(I, TrialBBI2)); 685 } 686 if (I == B) 687 break; 688 } 689 } 690 return maxCommonTailLength; 691 } 692 693 void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 694 MachineBasicBlock *SuccBB, 695 MachineBasicBlock *PredBB) { 696 MPIterator CurMPIter, B; 697 for (CurMPIter = std::prev(MergePotentials.end()), 698 B = MergePotentials.begin(); 699 CurMPIter->getHash() == CurHash; --CurMPIter) { 700 // Put the unconditional branch back, if we need one. 701 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 702 if (SuccBB && CurMBB != PredBB) 703 FixTail(CurMBB, SuccBB, TII); 704 if (CurMPIter == B) 705 break; 706 } 707 if (CurMPIter->getHash() != CurHash) 708 CurMPIter++; 709 MergePotentials.erase(CurMPIter, MergePotentials.end()); 710 } 711 712 bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 713 MachineBasicBlock *SuccBB, 714 unsigned maxCommonTailLength, 715 unsigned &commonTailIndex) { 716 commonTailIndex = 0; 717 unsigned TimeEstimate = ~0U; 718 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 719 // Use PredBB if possible; that doesn't require a new branch. 720 if (SameTails[i].getBlock() == PredBB) { 721 commonTailIndex = i; 722 break; 723 } 724 // Otherwise, make a (fairly bogus) choice based on estimate of 725 // how long it will take the various blocks to execute. 726 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 727 SameTails[i].getTailStartPos()); 728 if (t <= TimeEstimate) { 729 TimeEstimate = t; 730 commonTailIndex = i; 731 } 732 } 733 734 MachineBasicBlock::iterator BBI = 735 SameTails[commonTailIndex].getTailStartPos(); 736 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 737 738 LLVM_DEBUG(dbgs() << "\nSplitting " << printMBBReference(*MBB) << ", size " 739 << maxCommonTailLength); 740 741 // If the split block unconditionally falls-thru to SuccBB, it will be 742 // merged. In control flow terms it should then take SuccBB's name. e.g. If 743 // SuccBB is an inner loop, the common tail is still part of the inner loop. 744 const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ? 745 SuccBB->getBasicBlock() : MBB->getBasicBlock(); 746 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB); 747 if (!newMBB) { 748 LLVM_DEBUG(dbgs() << "... failed!"); 749 return false; 750 } 751 752 SameTails[commonTailIndex].setBlock(newMBB); 753 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 754 755 // If we split PredBB, newMBB is the new predecessor. 756 if (PredBB == MBB) 757 PredBB = newMBB; 758 759 return true; 760 } 761 762 static void 763 mergeOperations(MachineBasicBlock::iterator MBBIStartPos, 764 MachineBasicBlock &MBBCommon) { 765 MachineBasicBlock *MBB = MBBIStartPos->getParent(); 766 // Note CommonTailLen does not necessarily matches the size of 767 // the common BB nor all its instructions because of debug 768 // instructions differences. 769 unsigned CommonTailLen = 0; 770 for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos) 771 ++CommonTailLen; 772 773 MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin(); 774 MachineBasicBlock::reverse_iterator MBBIE = MBB->rend(); 775 MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin(); 776 MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend(); 777 778 while (CommonTailLen--) { 779 assert(MBBI != MBBIE && "Reached BB end within common tail length!"); 780 (void)MBBIE; 781 782 if (!countsAsInstruction(*MBBI)) { 783 ++MBBI; 784 continue; 785 } 786 787 while ((MBBICommon != MBBIECommon) && !countsAsInstruction(*MBBICommon)) 788 ++MBBICommon; 789 790 assert(MBBICommon != MBBIECommon && 791 "Reached BB end within common tail length!"); 792 assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!"); 793 794 // Merge MMOs from memory operations in the common block. 795 if (MBBICommon->mayLoadOrStore()) 796 MBBICommon->cloneMergedMemRefs(*MBB->getParent(), {&*MBBICommon, &*MBBI}); 797 // Drop undef flags if they aren't present in all merged instructions. 798 for (unsigned I = 0, E = MBBICommon->getNumOperands(); I != E; ++I) { 799 MachineOperand &MO = MBBICommon->getOperand(I); 800 if (MO.isReg() && MO.isUndef()) { 801 const MachineOperand &OtherMO = MBBI->getOperand(I); 802 if (!OtherMO.isUndef()) 803 MO.setIsUndef(false); 804 } 805 } 806 807 ++MBBI; 808 ++MBBICommon; 809 } 810 } 811 812 void BranchFolder::mergeCommonTails(unsigned commonTailIndex) { 813 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 814 815 std::vector<MachineBasicBlock::iterator> NextCommonInsts(SameTails.size()); 816 for (unsigned int i = 0 ; i != SameTails.size() ; ++i) { 817 if (i != commonTailIndex) { 818 NextCommonInsts[i] = SameTails[i].getTailStartPos(); 819 mergeOperations(SameTails[i].getTailStartPos(), *MBB); 820 } else { 821 assert(SameTails[i].getTailStartPos() == MBB->begin() && 822 "MBB is not a common tail only block"); 823 } 824 } 825 826 for (auto &MI : *MBB) { 827 if (!countsAsInstruction(MI)) 828 continue; 829 DebugLoc DL = MI.getDebugLoc(); 830 for (unsigned int i = 0 ; i < NextCommonInsts.size() ; i++) { 831 if (i == commonTailIndex) 832 continue; 833 834 auto &Pos = NextCommonInsts[i]; 835 assert(Pos != SameTails[i].getBlock()->end() && 836 "Reached BB end within common tail"); 837 while (!countsAsInstruction(*Pos)) { 838 ++Pos; 839 assert(Pos != SameTails[i].getBlock()->end() && 840 "Reached BB end within common tail"); 841 } 842 assert(MI.isIdenticalTo(*Pos) && "Expected matching MIIs!"); 843 DL = DILocation::getMergedLocation(DL, Pos->getDebugLoc()); 844 NextCommonInsts[i] = ++Pos; 845 } 846 MI.setDebugLoc(DL); 847 } 848 849 if (UpdateLiveIns) { 850 LivePhysRegs NewLiveIns(*TRI); 851 computeLiveIns(NewLiveIns, *MBB); 852 LiveRegs.init(*TRI); 853 854 // The flag merging may lead to some register uses no longer using the 855 // <undef> flag, add IMPLICIT_DEFs in the predecessors as necessary. 856 for (MachineBasicBlock *Pred : MBB->predecessors()) { 857 LiveRegs.clear(); 858 LiveRegs.addLiveOuts(*Pred); 859 MachineBasicBlock::iterator InsertBefore = Pred->getFirstTerminator(); 860 for (Register Reg : NewLiveIns) { 861 if (!LiveRegs.available(*MRI, Reg)) 862 continue; 863 DebugLoc DL; 864 BuildMI(*Pred, InsertBefore, DL, TII->get(TargetOpcode::IMPLICIT_DEF), 865 Reg); 866 } 867 } 868 869 MBB->clearLiveIns(); 870 addLiveIns(*MBB, NewLiveIns); 871 } 872 } 873 874 // See if any of the blocks in MergePotentials (which all have SuccBB as a 875 // successor, or all have no successor if it is null) can be tail-merged. 876 // If there is a successor, any blocks in MergePotentials that are not 877 // tail-merged and are not immediately before Succ must have an unconditional 878 // branch to Succ added (but the predecessor/successor lists need no 879 // adjustment). The lone predecessor of Succ that falls through into Succ, 880 // if any, is given in PredBB. 881 // MinCommonTailLength - Except for the special cases below, tail-merge if 882 // there are at least this many instructions in common. 883 bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 884 MachineBasicBlock *PredBB, 885 unsigned MinCommonTailLength) { 886 bool MadeChange = false; 887 888 LLVM_DEBUG( 889 dbgs() << "\nTryTailMergeBlocks: "; 890 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) dbgs() 891 << printMBBReference(*MergePotentials[i].getBlock()) 892 << (i == e - 1 ? "" : ", "); 893 dbgs() << "\n"; if (SuccBB) { 894 dbgs() << " with successor " << printMBBReference(*SuccBB) << '\n'; 895 if (PredBB) 896 dbgs() << " which has fall-through from " 897 << printMBBReference(*PredBB) << "\n"; 898 } dbgs() << "Looking for common tails of at least " 899 << MinCommonTailLength << " instruction" 900 << (MinCommonTailLength == 1 ? "" : "s") << '\n';); 901 902 // Sort by hash value so that blocks with identical end sequences sort 903 // together. 904 array_pod_sort(MergePotentials.begin(), MergePotentials.end()); 905 906 // Walk through equivalence sets looking for actual exact matches. 907 while (MergePotentials.size() > 1) { 908 unsigned CurHash = MergePotentials.back().getHash(); 909 910 // Build SameTails, identifying the set of blocks with this hash code 911 // and with the maximum number of instructions in common. 912 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 913 MinCommonTailLength, 914 SuccBB, PredBB); 915 916 // If we didn't find any pair that has at least MinCommonTailLength 917 // instructions in common, remove all blocks with this hash code and retry. 918 if (SameTails.empty()) { 919 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 920 continue; 921 } 922 923 // If one of the blocks is the entire common tail (and is not the entry 924 // block/an EH pad, which we can't jump to), we can treat all blocks with 925 // this same tail at once. Use PredBB if that is one of the possibilities, 926 // as that will not introduce any extra branches. 927 MachineBasicBlock *EntryBB = 928 &MergePotentials.front().getBlock()->getParent()->front(); 929 unsigned commonTailIndex = SameTails.size(); 930 // If there are two blocks, check to see if one can be made to fall through 931 // into the other. 932 if (SameTails.size() == 2 && 933 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 934 SameTails[1].tailIsWholeBlock() && !SameTails[1].getBlock()->isEHPad()) 935 commonTailIndex = 1; 936 else if (SameTails.size() == 2 && 937 SameTails[1].getBlock()->isLayoutSuccessor( 938 SameTails[0].getBlock()) && 939 SameTails[0].tailIsWholeBlock() && 940 !SameTails[0].getBlock()->isEHPad()) 941 commonTailIndex = 0; 942 else { 943 // Otherwise just pick one, favoring the fall-through predecessor if 944 // there is one. 945 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 946 MachineBasicBlock *MBB = SameTails[i].getBlock(); 947 if ((MBB == EntryBB || MBB->isEHPad()) && 948 SameTails[i].tailIsWholeBlock()) 949 continue; 950 if (MBB == PredBB) { 951 commonTailIndex = i; 952 break; 953 } 954 if (SameTails[i].tailIsWholeBlock()) 955 commonTailIndex = i; 956 } 957 } 958 959 if (commonTailIndex == SameTails.size() || 960 (SameTails[commonTailIndex].getBlock() == PredBB && 961 !SameTails[commonTailIndex].tailIsWholeBlock())) { 962 // None of the blocks consist entirely of the common tail. 963 // Split a block so that one does. 964 if (!CreateCommonTailOnlyBlock(PredBB, SuccBB, 965 maxCommonTailLength, commonTailIndex)) { 966 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 967 continue; 968 } 969 } 970 971 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 972 973 // Recompute common tail MBB's edge weights and block frequency. 974 setCommonTailEdgeWeights(*MBB); 975 976 // Merge debug locations, MMOs and undef flags across identical instructions 977 // for common tail. 978 mergeCommonTails(commonTailIndex); 979 980 // MBB is common tail. Adjust all other BB's to jump to this one. 981 // Traversal must be forwards so erases work. 982 LLVM_DEBUG(dbgs() << "\nUsing common tail in " << printMBBReference(*MBB) 983 << " for "); 984 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 985 if (commonTailIndex == i) 986 continue; 987 LLVM_DEBUG(dbgs() << printMBBReference(*SameTails[i].getBlock()) 988 << (i == e - 1 ? "" : ", ")); 989 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 990 replaceTailWithBranchTo(SameTails[i].getTailStartPos(), *MBB); 991 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 992 MergePotentials.erase(SameTails[i].getMPIter()); 993 } 994 LLVM_DEBUG(dbgs() << "\n"); 995 // We leave commonTailIndex in the worklist in case there are other blocks 996 // that match it with a smaller number of instructions. 997 MadeChange = true; 998 } 999 return MadeChange; 1000 } 1001 1002 bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 1003 bool MadeChange = false; 1004 if (!EnableTailMerge) 1005 return MadeChange; 1006 1007 // First find blocks with no successors. 1008 // Block placement may create new tail merging opportunities for these blocks. 1009 MergePotentials.clear(); 1010 for (MachineBasicBlock &MBB : MF) { 1011 if (MergePotentials.size() == TailMergeThreshold) 1012 break; 1013 if (!TriedMerging.count(&MBB) && MBB.succ_empty()) 1014 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB)); 1015 } 1016 1017 // If this is a large problem, avoid visiting the same basic blocks 1018 // multiple times. 1019 if (MergePotentials.size() == TailMergeThreshold) 1020 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 1021 TriedMerging.insert(MergePotentials[i].getBlock()); 1022 1023 // See if we can do any tail merging on those. 1024 if (MergePotentials.size() >= 2) 1025 MadeChange |= TryTailMergeBlocks(nullptr, nullptr, MinCommonTailLength); 1026 1027 // Look at blocks (IBB) with multiple predecessors (PBB). 1028 // We change each predecessor to a canonical form, by 1029 // (1) temporarily removing any unconditional branch from the predecessor 1030 // to IBB, and 1031 // (2) alter conditional branches so they branch to the other block 1032 // not IBB; this may require adding back an unconditional branch to IBB 1033 // later, where there wasn't one coming in. E.g. 1034 // Bcc IBB 1035 // fallthrough to QBB 1036 // here becomes 1037 // Bncc QBB 1038 // with a conceptual B to IBB after that, which never actually exists. 1039 // With those changes, we see whether the predecessors' tails match, 1040 // and merge them if so. We change things out of canonical form and 1041 // back to the way they were later in the process. (OptimizeBranches 1042 // would undo some of this, but we can't use it, because we'd get into 1043 // a compile-time infinite loop repeatedly doing and undoing the same 1044 // transformations.) 1045 1046 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end(); 1047 I != E; ++I) { 1048 if (I->pred_size() < 2) continue; 1049 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 1050 MachineBasicBlock *IBB = &*I; 1051 MachineBasicBlock *PredBB = &*std::prev(I); 1052 MergePotentials.clear(); 1053 MachineLoop *ML; 1054 1055 // Bail if merging after placement and IBB is the loop header because 1056 // -- If merging predecessors that belong to the same loop as IBB, the 1057 // common tail of merged predecessors may become the loop top if block 1058 // placement is called again and the predecessors may branch to this common 1059 // tail and require more branches. This can be relaxed if 1060 // MachineBlockPlacement::findBestLoopTop is more flexible. 1061 // --If merging predecessors that do not belong to the same loop as IBB, the 1062 // loop info of IBB's loop and the other loops may be affected. Calling the 1063 // block placement again may make big change to the layout and eliminate the 1064 // reason to do tail merging here. 1065 if (AfterBlockPlacement && MLI) { 1066 ML = MLI->getLoopFor(IBB); 1067 if (ML && IBB == ML->getHeader()) 1068 continue; 1069 } 1070 1071 for (MachineBasicBlock *PBB : I->predecessors()) { 1072 if (MergePotentials.size() == TailMergeThreshold) 1073 break; 1074 1075 if (TriedMerging.count(PBB)) 1076 continue; 1077 1078 // Skip blocks that loop to themselves, can't tail merge these. 1079 if (PBB == IBB) 1080 continue; 1081 1082 // Visit each predecessor only once. 1083 if (!UniquePreds.insert(PBB).second) 1084 continue; 1085 1086 // Skip blocks which may jump to a landing pad or jump from an asm blob. 1087 // Can't tail merge these. 1088 if (PBB->hasEHPadSuccessor() || PBB->mayHaveInlineAsmBr()) 1089 continue; 1090 1091 // After block placement, only consider predecessors that belong to the 1092 // same loop as IBB. The reason is the same as above when skipping loop 1093 // header. 1094 if (AfterBlockPlacement && MLI) 1095 if (ML != MLI->getLoopFor(PBB)) 1096 continue; 1097 1098 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 1099 SmallVector<MachineOperand, 4> Cond; 1100 if (!TII->analyzeBranch(*PBB, TBB, FBB, Cond, true)) { 1101 // Failing case: IBB is the target of a cbr, and we cannot reverse the 1102 // branch. 1103 SmallVector<MachineOperand, 4> NewCond(Cond); 1104 if (!Cond.empty() && TBB == IBB) { 1105 if (TII->reverseBranchCondition(NewCond)) 1106 continue; 1107 // This is the QBB case described above 1108 if (!FBB) { 1109 auto Next = ++PBB->getIterator(); 1110 if (Next != MF.end()) 1111 FBB = &*Next; 1112 } 1113 } 1114 1115 // Remove the unconditional branch at the end, if any. 1116 if (TBB && (Cond.empty() || FBB)) { 1117 DebugLoc dl = PBB->findBranchDebugLoc(); 1118 TII->removeBranch(*PBB); 1119 if (!Cond.empty()) 1120 // reinsert conditional branch only, for now 1121 TII->insertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr, 1122 NewCond, dl); 1123 } 1124 1125 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(*PBB), PBB)); 1126 } 1127 } 1128 1129 // If this is a large problem, avoid visiting the same basic blocks multiple 1130 // times. 1131 if (MergePotentials.size() == TailMergeThreshold) 1132 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 1133 TriedMerging.insert(MergePotentials[i].getBlock()); 1134 1135 if (MergePotentials.size() >= 2) 1136 MadeChange |= TryTailMergeBlocks(IBB, PredBB, MinCommonTailLength); 1137 1138 // Reinsert an unconditional branch if needed. The 1 below can occur as a 1139 // result of removing blocks in TryTailMergeBlocks. 1140 PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks 1141 if (MergePotentials.size() == 1 && 1142 MergePotentials.begin()->getBlock() != PredBB) 1143 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 1144 } 1145 1146 return MadeChange; 1147 } 1148 1149 void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) { 1150 SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size()); 1151 BlockFrequency AccumulatedMBBFreq; 1152 1153 // Aggregate edge frequency of successor edge j: 1154 // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)), 1155 // where bb is a basic block that is in SameTails. 1156 for (const auto &Src : SameTails) { 1157 const MachineBasicBlock *SrcMBB = Src.getBlock(); 1158 BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB); 1159 AccumulatedMBBFreq += BlockFreq; 1160 1161 // It is not necessary to recompute edge weights if TailBB has less than two 1162 // successors. 1163 if (TailMBB.succ_size() <= 1) 1164 continue; 1165 1166 auto EdgeFreq = EdgeFreqLs.begin(); 1167 1168 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end(); 1169 SuccI != SuccE; ++SuccI, ++EdgeFreq) 1170 *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI); 1171 } 1172 1173 MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq); 1174 1175 if (TailMBB.succ_size() <= 1) 1176 return; 1177 1178 auto SumEdgeFreq = 1179 std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0)) 1180 .getFrequency(); 1181 auto EdgeFreq = EdgeFreqLs.begin(); 1182 1183 if (SumEdgeFreq > 0) { 1184 for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end(); 1185 SuccI != SuccE; ++SuccI, ++EdgeFreq) { 1186 auto Prob = BranchProbability::getBranchProbability( 1187 EdgeFreq->getFrequency(), SumEdgeFreq); 1188 TailMBB.setSuccProbability(SuccI, Prob); 1189 } 1190 } 1191 } 1192 1193 //===----------------------------------------------------------------------===// 1194 // Branch Optimization 1195 //===----------------------------------------------------------------------===// 1196 1197 bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 1198 bool MadeChange = false; 1199 1200 // Make sure blocks are numbered in order 1201 MF.RenumberBlocks(); 1202 // Renumbering blocks alters EH scope membership, recalculate it. 1203 EHScopeMembership = getEHScopeMembership(MF); 1204 1205 for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end(); 1206 I != E; ) { 1207 MachineBasicBlock *MBB = &*I++; 1208 MadeChange |= OptimizeBlock(MBB); 1209 1210 // If it is dead, remove it. 1211 if (MBB->pred_empty()) { 1212 RemoveDeadBlock(MBB); 1213 MadeChange = true; 1214 ++NumDeadBlocks; 1215 } 1216 } 1217 1218 return MadeChange; 1219 } 1220 1221 // Blocks should be considered empty if they contain only debug info; 1222 // else the debug info would affect codegen. 1223 static bool IsEmptyBlock(MachineBasicBlock *MBB) { 1224 return MBB->getFirstNonDebugInstr() == MBB->end(); 1225 } 1226 1227 // Blocks with only debug info and branches should be considered the same 1228 // as blocks with only branches. 1229 static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { 1230 MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr(); 1231 assert(I != MBB->end() && "empty block!"); 1232 return I->isBranch(); 1233 } 1234 1235 /// IsBetterFallthrough - Return true if it would be clearly better to 1236 /// fall-through to MBB1 than to fall through into MBB2. This has to return 1237 /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 1238 /// result in infinite loops. 1239 static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 1240 MachineBasicBlock *MBB2) { 1241 assert(MBB1 && MBB2 && "Unknown MachineBasicBlock"); 1242 1243 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 1244 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 1245 // optimize branches that branch to either a return block or an assert block 1246 // into a fallthrough to the return. 1247 MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr(); 1248 MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr(); 1249 if (MBB1I == MBB1->end() || MBB2I == MBB2->end()) 1250 return false; 1251 1252 // If there is a clear successor ordering we make sure that one block 1253 // will fall through to the next 1254 if (MBB1->isSuccessor(MBB2)) return true; 1255 if (MBB2->isSuccessor(MBB1)) return false; 1256 1257 return MBB2I->isCall() && !MBB1I->isCall(); 1258 } 1259 1260 /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch 1261 /// instructions on the block. 1262 static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) { 1263 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); 1264 if (I != MBB.end() && I->isBranch()) 1265 return I->getDebugLoc(); 1266 return DebugLoc(); 1267 } 1268 1269 static void copyDebugInfoToPredecessor(const TargetInstrInfo *TII, 1270 MachineBasicBlock &MBB, 1271 MachineBasicBlock &PredMBB) { 1272 auto InsertBefore = PredMBB.getFirstTerminator(); 1273 for (MachineInstr &MI : MBB.instrs()) 1274 if (MI.isDebugInstr()) { 1275 TII->duplicate(PredMBB, InsertBefore, MI); 1276 LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to pred: " 1277 << MI); 1278 } 1279 } 1280 1281 static void copyDebugInfoToSuccessor(const TargetInstrInfo *TII, 1282 MachineBasicBlock &MBB, 1283 MachineBasicBlock &SuccMBB) { 1284 auto InsertBefore = SuccMBB.SkipPHIsAndLabels(SuccMBB.begin()); 1285 for (MachineInstr &MI : MBB.instrs()) 1286 if (MI.isDebugInstr()) { 1287 TII->duplicate(SuccMBB, InsertBefore, MI); 1288 LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to succ: " 1289 << MI); 1290 } 1291 } 1292 1293 // Try to salvage DBG_VALUE instructions from an otherwise empty block. If such 1294 // a basic block is removed we would lose the debug information unless we have 1295 // copied the information to a predecessor/successor. 1296 // 1297 // TODO: This function only handles some simple cases. An alternative would be 1298 // to run a heavier analysis, such as the LiveDebugValues pass, before we do 1299 // branch folding. 1300 static void salvageDebugInfoFromEmptyBlock(const TargetInstrInfo *TII, 1301 MachineBasicBlock &MBB) { 1302 assert(IsEmptyBlock(&MBB) && "Expected an empty block (except debug info)."); 1303 // If this MBB is the only predecessor of a successor it is legal to copy 1304 // DBG_VALUE instructions to the beginning of the successor. 1305 for (MachineBasicBlock *SuccBB : MBB.successors()) 1306 if (SuccBB->pred_size() == 1) 1307 copyDebugInfoToSuccessor(TII, MBB, *SuccBB); 1308 // If this MBB is the only successor of a predecessor it is legal to copy the 1309 // DBG_VALUE instructions to the end of the predecessor (just before the 1310 // terminators, assuming that the terminator isn't affecting the DBG_VALUE). 1311 for (MachineBasicBlock *PredBB : MBB.predecessors()) 1312 if (PredBB->succ_size() == 1) 1313 copyDebugInfoToPredecessor(TII, MBB, *PredBB); 1314 } 1315 1316 bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1317 bool MadeChange = false; 1318 MachineFunction &MF = *MBB->getParent(); 1319 ReoptimizeBlock: 1320 1321 MachineFunction::iterator FallThrough = MBB->getIterator(); 1322 ++FallThrough; 1323 1324 // Make sure MBB and FallThrough belong to the same EH scope. 1325 bool SameEHScope = true; 1326 if (!EHScopeMembership.empty() && FallThrough != MF.end()) { 1327 auto MBBEHScope = EHScopeMembership.find(MBB); 1328 assert(MBBEHScope != EHScopeMembership.end()); 1329 auto FallThroughEHScope = EHScopeMembership.find(&*FallThrough); 1330 assert(FallThroughEHScope != EHScopeMembership.end()); 1331 SameEHScope = MBBEHScope->second == FallThroughEHScope->second; 1332 } 1333 1334 // Analyze the branch in the current block. As a side-effect, this may cause 1335 // the block to become empty. 1336 MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr; 1337 SmallVector<MachineOperand, 4> CurCond; 1338 bool CurUnAnalyzable = 1339 TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1340 1341 // If this block is empty, make everyone use its fall-through, not the block 1342 // explicitly. Landing pads should not do this since the landing-pad table 1343 // points to this block. Blocks with their addresses taken shouldn't be 1344 // optimized away. 1345 if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() && 1346 SameEHScope) { 1347 salvageDebugInfoFromEmptyBlock(TII, *MBB); 1348 // Dead block? Leave for cleanup later. 1349 if (MBB->pred_empty()) return MadeChange; 1350 1351 if (FallThrough == MF.end()) { 1352 // TODO: Simplify preds to not branch here if possible! 1353 } else if (FallThrough->isEHPad()) { 1354 // Don't rewrite to a landing pad fallthough. That could lead to the case 1355 // where a BB jumps to more than one landing pad. 1356 // TODO: Is it ever worth rewriting predecessors which don't already 1357 // jump to a landing pad, and so can safely jump to the fallthrough? 1358 } else if (MBB->isSuccessor(&*FallThrough)) { 1359 // Rewrite all predecessors of the old block to go to the fallthrough 1360 // instead. 1361 while (!MBB->pred_empty()) { 1362 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1363 Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough); 1364 } 1365 // If MBB was the target of a jump table, update jump tables to go to the 1366 // fallthrough instead. 1367 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1368 MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough); 1369 MadeChange = true; 1370 } 1371 return MadeChange; 1372 } 1373 1374 // Check to see if we can simplify the terminator of the block before this 1375 // one. 1376 MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB)); 1377 1378 MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr; 1379 SmallVector<MachineOperand, 4> PriorCond; 1380 bool PriorUnAnalyzable = 1381 TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1382 if (!PriorUnAnalyzable) { 1383 // If the previous branch is conditional and both conditions go to the same 1384 // destination, remove the branch, replacing it with an unconditional one or 1385 // a fall-through. 1386 if (PriorTBB && PriorTBB == PriorFBB) { 1387 DebugLoc dl = getBranchDebugLoc(PrevBB); 1388 TII->removeBranch(PrevBB); 1389 PriorCond.clear(); 1390 if (PriorTBB != MBB) 1391 TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl); 1392 MadeChange = true; 1393 ++NumBranchOpts; 1394 goto ReoptimizeBlock; 1395 } 1396 1397 // If the previous block unconditionally falls through to this block and 1398 // this block has no other predecessors, move the contents of this block 1399 // into the prior block. This doesn't usually happen when SimplifyCFG 1400 // has been used, but it can happen if tail merging splits a fall-through 1401 // predecessor of a block. 1402 // This has to check PrevBB->succ_size() because EH edges are ignored by 1403 // analyzeBranch. 1404 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1405 PrevBB.succ_size() == 1 && 1406 !MBB->hasAddressTaken() && !MBB->isEHPad()) { 1407 LLVM_DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1408 << "From MBB: " << *MBB); 1409 // Remove redundant DBG_VALUEs first. 1410 if (PrevBB.begin() != PrevBB.end()) { 1411 MachineBasicBlock::iterator PrevBBIter = PrevBB.end(); 1412 --PrevBBIter; 1413 MachineBasicBlock::iterator MBBIter = MBB->begin(); 1414 // Check if DBG_VALUE at the end of PrevBB is identical to the 1415 // DBG_VALUE at the beginning of MBB. 1416 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end() 1417 && PrevBBIter->isDebugInstr() && MBBIter->isDebugInstr()) { 1418 if (!MBBIter->isIdenticalTo(*PrevBBIter)) 1419 break; 1420 MachineInstr &DuplicateDbg = *MBBIter; 1421 ++MBBIter; -- PrevBBIter; 1422 DuplicateDbg.eraseFromParent(); 1423 } 1424 } 1425 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1426 PrevBB.removeSuccessor(PrevBB.succ_begin()); 1427 assert(PrevBB.succ_empty()); 1428 PrevBB.transferSuccessors(MBB); 1429 MadeChange = true; 1430 return MadeChange; 1431 } 1432 1433 // If the previous branch *only* branches to *this* block (conditional or 1434 // not) remove the branch. 1435 if (PriorTBB == MBB && !PriorFBB) { 1436 TII->removeBranch(PrevBB); 1437 MadeChange = true; 1438 ++NumBranchOpts; 1439 goto ReoptimizeBlock; 1440 } 1441 1442 // If the prior block branches somewhere else on the condition and here if 1443 // the condition is false, remove the uncond second branch. 1444 if (PriorFBB == MBB) { 1445 DebugLoc dl = getBranchDebugLoc(PrevBB); 1446 TII->removeBranch(PrevBB); 1447 TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl); 1448 MadeChange = true; 1449 ++NumBranchOpts; 1450 goto ReoptimizeBlock; 1451 } 1452 1453 // If the prior block branches here on true and somewhere else on false, and 1454 // if the branch condition is reversible, reverse the branch to create a 1455 // fall-through. 1456 if (PriorTBB == MBB) { 1457 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1458 if (!TII->reverseBranchCondition(NewPriorCond)) { 1459 DebugLoc dl = getBranchDebugLoc(PrevBB); 1460 TII->removeBranch(PrevBB); 1461 TII->insertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl); 1462 MadeChange = true; 1463 ++NumBranchOpts; 1464 goto ReoptimizeBlock; 1465 } 1466 } 1467 1468 // If this block has no successors (e.g. it is a return block or ends with 1469 // a call to a no-return function like abort or __cxa_throw) and if the pred 1470 // falls through into this block, and if it would otherwise fall through 1471 // into the block after this, move this block to the end of the function. 1472 // 1473 // We consider it more likely that execution will stay in the function (e.g. 1474 // due to loops) than it is to exit it. This asserts in loops etc, moving 1475 // the assert condition out of the loop body. 1476 if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB && 1477 MachineFunction::iterator(PriorTBB) == FallThrough && 1478 !MBB->canFallThrough()) { 1479 bool DoTransform = true; 1480 1481 // We have to be careful that the succs of PredBB aren't both no-successor 1482 // blocks. If neither have successors and if PredBB is the second from 1483 // last block in the function, we'd just keep swapping the two blocks for 1484 // last. Only do the swap if one is clearly better to fall through than 1485 // the other. 1486 if (FallThrough == --MF.end() && 1487 !IsBetterFallthrough(PriorTBB, MBB)) 1488 DoTransform = false; 1489 1490 if (DoTransform) { 1491 // Reverse the branch so we will fall through on the previous true cond. 1492 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1493 if (!TII->reverseBranchCondition(NewPriorCond)) { 1494 LLVM_DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1495 << "To make fallthrough to: " << *PriorTBB << "\n"); 1496 1497 DebugLoc dl = getBranchDebugLoc(PrevBB); 1498 TII->removeBranch(PrevBB); 1499 TII->insertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl); 1500 1501 // Move this block to the end of the function. 1502 MBB->moveAfter(&MF.back()); 1503 MadeChange = true; 1504 ++NumBranchOpts; 1505 return MadeChange; 1506 } 1507 } 1508 } 1509 } 1510 1511 bool OptForSize = 1512 MF.getFunction().hasOptSize() || 1513 llvm::shouldOptimizeForSize(MBB, PSI, &MBBFreqInfo); 1514 if (!IsEmptyBlock(MBB) && MBB->pred_size() == 1 && OptForSize) { 1515 // Changing "Jcc foo; foo: jmp bar;" into "Jcc bar;" might change the branch 1516 // direction, thereby defeating careful block placement and regressing 1517 // performance. Therefore, only consider this for optsize functions. 1518 MachineInstr &TailCall = *MBB->getFirstNonDebugInstr(); 1519 if (TII->isUnconditionalTailCall(TailCall)) { 1520 MachineBasicBlock *Pred = *MBB->pred_begin(); 1521 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr; 1522 SmallVector<MachineOperand, 4> PredCond; 1523 bool PredAnalyzable = 1524 !TII->analyzeBranch(*Pred, PredTBB, PredFBB, PredCond, true); 1525 1526 if (PredAnalyzable && !PredCond.empty() && PredTBB == MBB && 1527 PredTBB != PredFBB) { 1528 // The predecessor has a conditional branch to this block which consists 1529 // of only a tail call. Try to fold the tail call into the conditional 1530 // branch. 1531 if (TII->canMakeTailCallConditional(PredCond, TailCall)) { 1532 // TODO: It would be nice if analyzeBranch() could provide a pointer 1533 // to the branch instruction so replaceBranchWithTailCall() doesn't 1534 // have to search for it. 1535 TII->replaceBranchWithTailCall(*Pred, PredCond, TailCall); 1536 ++NumTailCalls; 1537 Pred->removeSuccessor(MBB); 1538 MadeChange = true; 1539 return MadeChange; 1540 } 1541 } 1542 // If the predecessor is falling through to this block, we could reverse 1543 // the branch condition and fold the tail call into that. However, after 1544 // that we might have to re-arrange the CFG to fall through to the other 1545 // block and there is a high risk of regressing code size rather than 1546 // improving it. 1547 } 1548 } 1549 1550 if (!CurUnAnalyzable) { 1551 // If this is a two-way branch, and the FBB branches to this block, reverse 1552 // the condition so the single-basic-block loop is faster. Instead of: 1553 // Loop: xxx; jcc Out; jmp Loop 1554 // we want: 1555 // Loop: xxx; jncc Loop; jmp Out 1556 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1557 SmallVector<MachineOperand, 4> NewCond(CurCond); 1558 if (!TII->reverseBranchCondition(NewCond)) { 1559 DebugLoc dl = getBranchDebugLoc(*MBB); 1560 TII->removeBranch(*MBB); 1561 TII->insertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); 1562 MadeChange = true; 1563 ++NumBranchOpts; 1564 goto ReoptimizeBlock; 1565 } 1566 } 1567 1568 // If this branch is the only thing in its block, see if we can forward 1569 // other blocks across it. 1570 if (CurTBB && CurCond.empty() && !CurFBB && 1571 IsBranchOnlyBlock(MBB) && CurTBB != MBB && 1572 !MBB->hasAddressTaken() && !MBB->isEHPad()) { 1573 DebugLoc dl = getBranchDebugLoc(*MBB); 1574 // This block may contain just an unconditional branch. Because there can 1575 // be 'non-branch terminators' in the block, try removing the branch and 1576 // then seeing if the block is empty. 1577 TII->removeBranch(*MBB); 1578 // If the only things remaining in the block are debug info, remove these 1579 // as well, so this will behave the same as an empty block in non-debug 1580 // mode. 1581 if (IsEmptyBlock(MBB)) { 1582 // Make the block empty, losing the debug info (we could probably 1583 // improve this in some cases.) 1584 MBB->erase(MBB->begin(), MBB->end()); 1585 } 1586 // If this block is just an unconditional branch to CurTBB, we can 1587 // usually completely eliminate the block. The only case we cannot 1588 // completely eliminate the block is when the block before this one 1589 // falls through into MBB and we can't understand the prior block's branch 1590 // condition. 1591 if (MBB->empty()) { 1592 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1593 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1594 !PrevBB.isSuccessor(MBB)) { 1595 // If the prior block falls through into us, turn it into an 1596 // explicit branch to us to make updates simpler. 1597 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1598 PriorTBB != MBB && PriorFBB != MBB) { 1599 if (!PriorTBB) { 1600 assert(PriorCond.empty() && !PriorFBB && 1601 "Bad branch analysis"); 1602 PriorTBB = MBB; 1603 } else { 1604 assert(!PriorFBB && "Machine CFG out of date!"); 1605 PriorFBB = MBB; 1606 } 1607 DebugLoc pdl = getBranchDebugLoc(PrevBB); 1608 TII->removeBranch(PrevBB); 1609 TII->insertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl); 1610 } 1611 1612 // Iterate through all the predecessors, revectoring each in-turn. 1613 size_t PI = 0; 1614 bool DidChange = false; 1615 bool HasBranchToSelf = false; 1616 while(PI != MBB->pred_size()) { 1617 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1618 if (PMBB == MBB) { 1619 // If this block has an uncond branch to itself, leave it. 1620 ++PI; 1621 HasBranchToSelf = true; 1622 } else { 1623 DidChange = true; 1624 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1625 // If this change resulted in PMBB ending in a conditional 1626 // branch where both conditions go to the same destination, 1627 // change this to an unconditional branch. 1628 MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr; 1629 SmallVector<MachineOperand, 4> NewCurCond; 1630 bool NewCurUnAnalyzable = TII->analyzeBranch( 1631 *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true); 1632 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1633 DebugLoc pdl = getBranchDebugLoc(*PMBB); 1634 TII->removeBranch(*PMBB); 1635 NewCurCond.clear(); 1636 TII->insertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl); 1637 MadeChange = true; 1638 ++NumBranchOpts; 1639 } 1640 } 1641 } 1642 1643 // Change any jumptables to go to the new MBB. 1644 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1645 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); 1646 if (DidChange) { 1647 ++NumBranchOpts; 1648 MadeChange = true; 1649 if (!HasBranchToSelf) return MadeChange; 1650 } 1651 } 1652 } 1653 1654 // Add the branch back if the block is more than just an uncond branch. 1655 TII->insertBranch(*MBB, CurTBB, nullptr, CurCond, dl); 1656 } 1657 } 1658 1659 // If the prior block doesn't fall through into this block, and if this 1660 // block doesn't fall through into some other block, see if we can find a 1661 // place to move this block where a fall-through will happen. 1662 if (!PrevBB.canFallThrough()) { 1663 // Now we know that there was no fall-through into this block, check to 1664 // see if it has a fall-through into its successor. 1665 bool CurFallsThru = MBB->canFallThrough(); 1666 1667 if (!MBB->isEHPad()) { 1668 // Check all the predecessors of this block. If one of them has no fall 1669 // throughs, and analyzeBranch thinks it _could_ fallthrough to this 1670 // block, move this block right after it. 1671 for (MachineBasicBlock *PredBB : MBB->predecessors()) { 1672 // Analyze the branch at the end of the pred. 1673 MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr; 1674 SmallVector<MachineOperand, 4> PredCond; 1675 if (PredBB != MBB && !PredBB->canFallThrough() && 1676 !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) && 1677 (PredTBB == MBB || PredFBB == MBB) && 1678 (!CurFallsThru || !CurTBB || !CurFBB) && 1679 (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1680 // If the current block doesn't fall through, just move it. 1681 // If the current block can fall through and does not end with a 1682 // conditional branch, we need to append an unconditional jump to 1683 // the (current) next block. To avoid a possible compile-time 1684 // infinite loop, move blocks only backward in this case. 1685 // Also, if there are already 2 branches here, we cannot add a third; 1686 // this means we have the case 1687 // Bcc next 1688 // B elsewhere 1689 // next: 1690 if (CurFallsThru) { 1691 MachineBasicBlock *NextBB = &*std::next(MBB->getIterator()); 1692 CurCond.clear(); 1693 TII->insertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc()); 1694 } 1695 MBB->moveAfter(PredBB); 1696 MadeChange = true; 1697 goto ReoptimizeBlock; 1698 } 1699 } 1700 } 1701 1702 if (!CurFallsThru) { 1703 // Check analyzable branch-successors to see if we can move this block 1704 // before one. 1705 if (!CurUnAnalyzable) { 1706 for (MachineBasicBlock *SuccBB : {CurFBB, CurTBB}) { 1707 if (!SuccBB) 1708 continue; 1709 // Analyze the branch at the end of the block before the succ. 1710 MachineFunction::iterator SuccPrev = --SuccBB->getIterator(); 1711 1712 // If this block doesn't already fall-through to that successor, and 1713 // if the succ doesn't already have a block that can fall through into 1714 // it, we can arrange for the fallthrough to happen. 1715 if (SuccBB != MBB && &*SuccPrev != MBB && 1716 !SuccPrev->canFallThrough()) { 1717 MBB->moveBefore(SuccBB); 1718 MadeChange = true; 1719 goto ReoptimizeBlock; 1720 } 1721 } 1722 } 1723 1724 // Okay, there is no really great place to put this block. If, however, 1725 // the block before this one would be a fall-through if this block were 1726 // removed, move this block to the end of the function. There is no real 1727 // advantage in "falling through" to an EH block, so we don't want to 1728 // perform this transformation for that case. 1729 // 1730 // Also, Windows EH introduced the possibility of an arbitrary number of 1731 // successors to a given block. The analyzeBranch call does not consider 1732 // exception handling and so we can get in a state where a block 1733 // containing a call is followed by multiple EH blocks that would be 1734 // rotated infinitely at the end of the function if the transformation 1735 // below were performed for EH "FallThrough" blocks. Therefore, even if 1736 // that appears not to be happening anymore, we should assume that it is 1737 // possible and not remove the "!FallThrough()->isEHPad" condition below. 1738 MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr; 1739 SmallVector<MachineOperand, 4> PrevCond; 1740 if (FallThrough != MF.end() && 1741 !FallThrough->isEHPad() && 1742 !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1743 PrevBB.isSuccessor(&*FallThrough)) { 1744 MBB->moveAfter(&MF.back()); 1745 MadeChange = true; 1746 return MadeChange; 1747 } 1748 } 1749 } 1750 1751 return MadeChange; 1752 } 1753 1754 //===----------------------------------------------------------------------===// 1755 // Hoist Common Code 1756 //===----------------------------------------------------------------------===// 1757 1758 bool BranchFolder::HoistCommonCode(MachineFunction &MF) { 1759 bool MadeChange = false; 1760 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) { 1761 MachineBasicBlock *MBB = &*I++; 1762 MadeChange |= HoistCommonCodeInSuccs(MBB); 1763 } 1764 1765 return MadeChange; 1766 } 1767 1768 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 1769 /// its 'true' successor. 1770 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 1771 MachineBasicBlock *TrueBB) { 1772 for (MachineBasicBlock *SuccBB : BB->successors()) 1773 if (SuccBB != TrueBB) 1774 return SuccBB; 1775 return nullptr; 1776 } 1777 1778 template <class Container> 1779 static void addRegAndItsAliases(Register Reg, const TargetRegisterInfo *TRI, 1780 Container &Set) { 1781 if (Reg.isPhysical()) { 1782 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1783 Set.insert(*AI); 1784 } else { 1785 Set.insert(Reg); 1786 } 1787 } 1788 1789 /// findHoistingInsertPosAndDeps - Find the location to move common instructions 1790 /// in successors to. The location is usually just before the terminator, 1791 /// however if the terminator is a conditional branch and its previous 1792 /// instruction is the flag setting instruction, the previous instruction is 1793 /// the preferred location. This function also gathers uses and defs of the 1794 /// instructions from the insertion point to the end of the block. The data is 1795 /// used by HoistCommonCodeInSuccs to ensure safety. 1796 static 1797 MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB, 1798 const TargetInstrInfo *TII, 1799 const TargetRegisterInfo *TRI, 1800 SmallSet<Register, 4> &Uses, 1801 SmallSet<Register, 4> &Defs) { 1802 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator(); 1803 if (!TII->isUnpredicatedTerminator(*Loc)) 1804 return MBB->end(); 1805 1806 for (const MachineOperand &MO : Loc->operands()) { 1807 if (!MO.isReg()) 1808 continue; 1809 Register Reg = MO.getReg(); 1810 if (!Reg) 1811 continue; 1812 if (MO.isUse()) { 1813 addRegAndItsAliases(Reg, TRI, Uses); 1814 } else { 1815 if (!MO.isDead()) 1816 // Don't try to hoist code in the rare case the terminator defines a 1817 // register that is later used. 1818 return MBB->end(); 1819 1820 // If the terminator defines a register, make sure we don't hoist 1821 // the instruction whose def might be clobbered by the terminator. 1822 addRegAndItsAliases(Reg, TRI, Defs); 1823 } 1824 } 1825 1826 if (Uses.empty()) 1827 return Loc; 1828 // If the terminator is the only instruction in the block and Uses is not 1829 // empty (or we would have returned above), we can still safely hoist 1830 // instructions just before the terminator as long as the Defs/Uses are not 1831 // violated (which is checked in HoistCommonCodeInSuccs). 1832 if (Loc == MBB->begin()) 1833 return Loc; 1834 1835 // The terminator is probably a conditional branch, try not to separate the 1836 // branch from condition setting instruction. 1837 MachineBasicBlock::iterator PI = prev_nodbg(Loc, MBB->begin()); 1838 1839 bool IsDef = false; 1840 for (const MachineOperand &MO : PI->operands()) { 1841 // If PI has a regmask operand, it is probably a call. Separate away. 1842 if (MO.isRegMask()) 1843 return Loc; 1844 if (!MO.isReg() || MO.isUse()) 1845 continue; 1846 Register Reg = MO.getReg(); 1847 if (!Reg) 1848 continue; 1849 if (Uses.count(Reg)) { 1850 IsDef = true; 1851 break; 1852 } 1853 } 1854 if (!IsDef) 1855 // The condition setting instruction is not just before the conditional 1856 // branch. 1857 return Loc; 1858 1859 // Be conservative, don't insert instruction above something that may have 1860 // side-effects. And since it's potentially bad to separate flag setting 1861 // instruction from the conditional branch, just abort the optimization 1862 // completely. 1863 // Also avoid moving code above predicated instruction since it's hard to 1864 // reason about register liveness with predicated instruction. 1865 bool DontMoveAcrossStore = true; 1866 if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI)) 1867 return MBB->end(); 1868 1869 // Find out what registers are live. Note this routine is ignoring other live 1870 // registers which are only used by instructions in successor blocks. 1871 for (const MachineOperand &MO : PI->operands()) { 1872 if (!MO.isReg()) 1873 continue; 1874 Register Reg = MO.getReg(); 1875 if (!Reg) 1876 continue; 1877 if (MO.isUse()) { 1878 addRegAndItsAliases(Reg, TRI, Uses); 1879 } else { 1880 if (Uses.erase(Reg)) { 1881 if (Register::isPhysicalRegister(Reg)) { 1882 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) 1883 Uses.erase(*SubRegs); // Use sub-registers to be conservative 1884 } 1885 } 1886 addRegAndItsAliases(Reg, TRI, Defs); 1887 } 1888 } 1889 1890 return PI; 1891 } 1892 1893 bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) { 1894 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 1895 SmallVector<MachineOperand, 4> Cond; 1896 if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty()) 1897 return false; 1898 1899 if (!FBB) FBB = findFalseBlock(MBB, TBB); 1900 if (!FBB) 1901 // Malformed bcc? True and false blocks are the same? 1902 return false; 1903 1904 // Restrict the optimization to cases where MBB is the only predecessor, 1905 // it is an obvious win. 1906 if (TBB->pred_size() > 1 || FBB->pred_size() > 1) 1907 return false; 1908 1909 // Find a suitable position to hoist the common instructions to. Also figure 1910 // out which registers are used or defined by instructions from the insertion 1911 // point to the end of the block. 1912 SmallSet<Register, 4> Uses, Defs; 1913 MachineBasicBlock::iterator Loc = 1914 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs); 1915 if (Loc == MBB->end()) 1916 return false; 1917 1918 bool HasDups = false; 1919 SmallSet<Register, 4> ActiveDefsSet, AllDefsSet; 1920 MachineBasicBlock::iterator TIB = TBB->begin(); 1921 MachineBasicBlock::iterator FIB = FBB->begin(); 1922 MachineBasicBlock::iterator TIE = TBB->end(); 1923 MachineBasicBlock::iterator FIE = FBB->end(); 1924 while (TIB != TIE && FIB != FIE) { 1925 // Skip dbg_value instructions. These do not count. 1926 TIB = skipDebugInstructionsForward(TIB, TIE); 1927 FIB = skipDebugInstructionsForward(FIB, FIE); 1928 if (TIB == TIE || FIB == FIE) 1929 break; 1930 1931 if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead)) 1932 break; 1933 1934 if (TII->isPredicated(*TIB)) 1935 // Hard to reason about register liveness with predicated instruction. 1936 break; 1937 1938 bool IsSafe = true; 1939 for (MachineOperand &MO : TIB->operands()) { 1940 // Don't attempt to hoist instructions with register masks. 1941 if (MO.isRegMask()) { 1942 IsSafe = false; 1943 break; 1944 } 1945 if (!MO.isReg()) 1946 continue; 1947 Register Reg = MO.getReg(); 1948 if (!Reg) 1949 continue; 1950 if (MO.isDef()) { 1951 if (Uses.count(Reg)) { 1952 // Avoid clobbering a register that's used by the instruction at 1953 // the point of insertion. 1954 IsSafe = false; 1955 break; 1956 } 1957 1958 if (Defs.count(Reg) && !MO.isDead()) { 1959 // Don't hoist the instruction if the def would be clobber by the 1960 // instruction at the point insertion. FIXME: This is overly 1961 // conservative. It should be possible to hoist the instructions 1962 // in BB2 in the following example: 1963 // BB1: 1964 // r1, eflag = op1 r2, r3 1965 // brcc eflag 1966 // 1967 // BB2: 1968 // r1 = op2, ... 1969 // = op3, killed r1 1970 IsSafe = false; 1971 break; 1972 } 1973 } else if (!ActiveDefsSet.count(Reg)) { 1974 if (Defs.count(Reg)) { 1975 // Use is defined by the instruction at the point of insertion. 1976 IsSafe = false; 1977 break; 1978 } 1979 1980 if (MO.isKill() && Uses.count(Reg)) 1981 // Kills a register that's read by the instruction at the point of 1982 // insertion. Remove the kill marker. 1983 MO.setIsKill(false); 1984 } 1985 } 1986 if (!IsSafe) 1987 break; 1988 1989 bool DontMoveAcrossStore = true; 1990 if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore)) 1991 break; 1992 1993 // Remove kills from ActiveDefsSet, these registers had short live ranges. 1994 for (const MachineOperand &MO : TIB->operands()) { 1995 if (!MO.isReg() || !MO.isUse() || !MO.isKill()) 1996 continue; 1997 Register Reg = MO.getReg(); 1998 if (!Reg) 1999 continue; 2000 if (!AllDefsSet.count(Reg)) { 2001 continue; 2002 } 2003 if (Reg.isPhysical()) { 2004 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 2005 ActiveDefsSet.erase(*AI); 2006 } else { 2007 ActiveDefsSet.erase(Reg); 2008 } 2009 } 2010 2011 // Track local defs so we can update liveins. 2012 for (const MachineOperand &MO : TIB->operands()) { 2013 if (!MO.isReg() || !MO.isDef() || MO.isDead()) 2014 continue; 2015 Register Reg = MO.getReg(); 2016 if (!Reg || Reg.isVirtual()) 2017 continue; 2018 addRegAndItsAliases(Reg, TRI, ActiveDefsSet); 2019 addRegAndItsAliases(Reg, TRI, AllDefsSet); 2020 } 2021 2022 HasDups = true; 2023 ++TIB; 2024 ++FIB; 2025 } 2026 2027 if (!HasDups) 2028 return false; 2029 2030 MBB->splice(Loc, TBB, TBB->begin(), TIB); 2031 FBB->erase(FBB->begin(), FIB); 2032 2033 if (UpdateLiveIns) { 2034 recomputeLiveIns(*TBB); 2035 recomputeLiveIns(*FBB); 2036 } 2037 2038 ++NumHoist; 2039 return true; 2040 } 2041