BranchFolding.cpp revision 154e1047184384afd0a701a9f8816459cf0b3490
1//===-- BranchFolding.cpp - Fold machine code branch instructions ---------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source 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. 16// 17//===----------------------------------------------------------------------===// 18 19#include "llvm/CodeGen/Passes.h" 20#include "llvm/CodeGen/MachineDebugInfo.h" 21#include "llvm/CodeGen/MachineFunctionPass.h" 22#include "llvm/CodeGen/MachineJumpTableInfo.h" 23#include "llvm/Target/TargetInstrInfo.h" 24#include "llvm/Target/TargetMachine.h" 25#include "llvm/Support/CommandLine.h" 26#include "llvm/ADT/Statistic.h" 27#include "llvm/ADT/STLExtras.h" 28#include <algorithm> 29using namespace llvm; 30 31static Statistic<> NumDeadBlocks("branchfold", "Number of dead blocks removed"); 32static Statistic<> NumBranchOpts("branchfold", "Number of branches optimized"); 33static Statistic<> NumTailMerge ("branchfold", "Number of block tails merged"); 34static cl::opt<bool> EnableTailMerge("enable-tail-merge", cl::Hidden); 35 36namespace { 37 struct BranchFolder : public MachineFunctionPass { 38 virtual bool runOnMachineFunction(MachineFunction &MF); 39 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 40 const TargetInstrInfo *TII; 41 MachineDebugInfo *MDI; 42 bool MadeChange; 43 private: 44 // Tail Merging. 45 bool TailMergeBlocks(MachineFunction &MF); 46 void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 47 MachineBasicBlock *NewDest); 48 MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB, 49 MachineBasicBlock::iterator BBI1); 50 51 // Branch optzn. 52 bool OptimizeBranches(MachineFunction &MF); 53 void OptimizeBlock(MachineBasicBlock *MBB); 54 void RemoveDeadBlock(MachineBasicBlock *MBB); 55 56 bool CanFallThrough(MachineBasicBlock *CurBB); 57 bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable, 58 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 59 const std::vector<MachineOperand> &Cond); 60 }; 61} 62 63FunctionPass *llvm::createBranchFoldingPass() { return new BranchFolder(); } 64 65/// RemoveDeadBlock - Remove the specified dead machine basic block from the 66/// function, updating the CFG. 67void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 68 assert(MBB->pred_empty() && "MBB must be dead!"); 69 70 MachineFunction *MF = MBB->getParent(); 71 // drop all successors. 72 while (!MBB->succ_empty()) 73 MBB->removeSuccessor(MBB->succ_end()-1); 74 75 // If there is DWARF info to active, check to see if there are any DWARF_LABEL 76 // records in the basic block. If so, unregister them from MachineDebugInfo. 77 if (MDI && !MBB->empty()) { 78 unsigned DWARF_LABELOpc = TII->getDWARF_LABELOpcode(); 79 assert(DWARF_LABELOpc && 80 "Target supports dwarf but didn't implement getDWARF_LABELOpcode!"); 81 82 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 83 I != E; ++I) { 84 if ((unsigned)I->getOpcode() == DWARF_LABELOpc) { 85 // The label ID # is always operand #0, an immediate. 86 MDI->InvalidateLabel(I->getOperand(0).getImm()); 87 } 88 } 89 } 90 91 // Remove the block. 92 MF->getBasicBlockList().erase(MBB); 93} 94 95bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { 96 TII = MF.getTarget().getInstrInfo(); 97 if (!TII) return false; 98 99 MDI = getAnalysisToUpdate<MachineDebugInfo>(); 100 101 bool EverMadeChange = false; 102 bool MadeChangeThisIteration = true; 103 while (MadeChangeThisIteration) { 104 MadeChangeThisIteration = false; 105 MadeChangeThisIteration |= TailMergeBlocks(MF); 106 MadeChangeThisIteration |= OptimizeBranches(MF); 107 EverMadeChange |= MadeChangeThisIteration; 108 } 109 110 // See if any jump tables have become mergable or dead as the code generator 111 // did its thing. 112 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 113 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 114 if (!JTs.empty()) { 115 // Figure out how these jump tables should be merged. 116 std::vector<unsigned> JTMapping; 117 JTMapping.reserve(JTs.size()); 118 119 // We always keep the 0th jump table. 120 JTMapping.push_back(0); 121 122 // Scan the jump tables, seeing if there are any duplicates. Note that this 123 // is N^2, which should be fixed someday. 124 for (unsigned i = 1, e = JTs.size(); i != e; ++i) 125 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 126 127 // If a jump table was merge with another one, walk the function rewriting 128 // references to jump tables to reference the new JT ID's. Keep track of 129 // whether we see a jump table idx, if not, we can delete the JT. 130 std::vector<bool> JTIsLive; 131 JTIsLive.resize(JTs.size()); 132 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 133 BB != E; ++BB) { 134 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 135 I != E; ++I) 136 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 137 MachineOperand &Op = I->getOperand(op); 138 if (!Op.isJumpTableIndex()) continue; 139 unsigned NewIdx = JTMapping[Op.getJumpTableIndex()]; 140 Op.setJumpTableIndex(NewIdx); 141 142 // Remember that this JT is live. 143 JTIsLive[NewIdx] = true; 144 } 145 } 146 147 // Finally, remove dead jump tables. This happens either because the 148 // indirect jump was unreachable (and thus deleted) or because the jump 149 // table was merged with some other one. 150 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 151 if (!JTIsLive[i]) { 152 JTI->RemoveJumpTable(i); 153 EverMadeChange = true; 154 } 155 } 156 157 return EverMadeChange; 158} 159 160//===----------------------------------------------------------------------===// 161// Tail Merging of Blocks 162//===----------------------------------------------------------------------===// 163 164/// HashMachineInstr - Compute a hash value for MI and its operands. 165static unsigned HashMachineInstr(const MachineInstr *MI) { 166 unsigned Hash = MI->getOpcode(); 167 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 168 const MachineOperand &Op = MI->getOperand(i); 169 170 // Merge in bits from the operand if easy. 171 unsigned OperandHash = 0; 172 switch (Op.getType()) { 173 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 174 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 175 case MachineOperand::MO_MachineBasicBlock: 176 OperandHash = Op.getMachineBasicBlock()->getNumber(); 177 break; 178 case MachineOperand::MO_FrameIndex: OperandHash = Op.getFrameIndex(); break; 179 case MachineOperand::MO_ConstantPoolIndex: 180 OperandHash = Op.getConstantPoolIndex(); 181 break; 182 case MachineOperand::MO_JumpTableIndex: 183 OperandHash = Op.getJumpTableIndex(); 184 break; 185 case MachineOperand::MO_GlobalAddress: 186 case MachineOperand::MO_ExternalSymbol: 187 // Global address / external symbol are too hard, don't bother, but do 188 // pull in the offset. 189 OperandHash = Op.getOffset(); 190 break; 191 default: break; 192 } 193 194 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 195 } 196 return Hash; 197} 198 199/// HashEndOfMBB - Hash the last two instructions in the MBB. We hash two 200/// instructions, because cross-jumping only saves code when at least two 201/// instructions are removed (since a branch must be inserted). 202static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) { 203 MachineBasicBlock::const_iterator I = MBB->end(); 204 if (I == MBB->begin()) 205 return 0; // Empty MBB. 206 207 --I; 208 unsigned Hash = HashMachineInstr(I); 209 210 if (I == MBB->begin()) 211 return Hash; // Single instr MBB. 212 213 --I; 214 // Hash in the second-to-last instruction. 215 Hash ^= HashMachineInstr(I) << 2; 216 return Hash; 217} 218 219/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 220/// of instructions they actually have in common together at their end. Return 221/// iterators for the first shared instruction in each block. 222static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 223 MachineBasicBlock *MBB2, 224 MachineBasicBlock::iterator &I1, 225 MachineBasicBlock::iterator &I2) { 226 I1 = MBB1->end(); 227 I2 = MBB2->end(); 228 229 unsigned TailLen = 0; 230 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 231 --I1; --I2; 232 if (!I1->isIdenticalTo(I2)) { 233 ++I1; ++I2; 234 break; 235 } 236 ++TailLen; 237 } 238 return TailLen; 239} 240 241/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 242/// after it, replacing it with an unconditional branch to NewDest. This 243/// returns true if OldInst's block is modified, false if NewDest is modified. 244void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 245 MachineBasicBlock *NewDest) { 246 MachineBasicBlock *OldBB = OldInst->getParent(); 247 248 // Remove all the old successors of OldBB from the CFG. 249 while (!OldBB->succ_empty()) 250 OldBB->removeSuccessor(OldBB->succ_begin()); 251 252 // Remove all the dead instructions from the end of OldBB. 253 OldBB->erase(OldInst, OldBB->end()); 254 255 // If OldBB isn't immediately before OldBB, insert a branch to it. 256 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 257 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>()); 258 OldBB->addSuccessor(NewDest); 259 ++NumTailMerge; 260} 261 262/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 263/// MBB so that the part before the iterator falls into the part starting at the 264/// iterator. This returns the new MBB. 265MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 266 MachineBasicBlock::iterator BBI1) { 267 // Create the fall-through block. 268 MachineFunction::iterator MBBI = &CurMBB; 269 MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock()); 270 CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB); 271 272 // Move all the successors of this block to the specified block. 273 while (!CurMBB.succ_empty()) { 274 MachineBasicBlock *S = *(CurMBB.succ_end()-1); 275 NewMBB->addSuccessor(S); 276 CurMBB.removeSuccessor(S); 277 } 278 279 // Add an edge from CurMBB to NewMBB for the fall-through. 280 CurMBB.addSuccessor(NewMBB); 281 282 // Splice the code over. 283 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 284 return NewMBB; 285} 286 287/// EstimateRuntime - Make a rough estimate for how long it will take to run 288/// the specified code. 289static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 290 MachineBasicBlock::iterator E, 291 const TargetInstrInfo *TII) { 292 unsigned Time = 0; 293 for (; I != E; ++I) { 294 const TargetInstrDescriptor &TID = TII->get(I->getOpcode()); 295 if (TID.Flags & M_CALL_FLAG) 296 Time += 10; 297 else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG)) 298 Time += 2; 299 else 300 ++Time; 301 } 302 return Time; 303} 304 305/// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at 306/// MBB2I and then insert an unconditional branch in the other block. Determine 307/// which is the best to split 308static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1, 309 MachineBasicBlock::iterator MBB1I, 310 MachineBasicBlock *MBB2, 311 MachineBasicBlock::iterator MBB2I, 312 const TargetInstrInfo *TII) { 313 // TODO: if we had some notion of which block was hotter, we could split 314 // the hot block, so it is the fall-through. Since we don't have profile info 315 // make a decision based on which will hurt most to split. 316 unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I, TII); 317 unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I, TII); 318 319 // If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the 320 // MBB1 block so it falls through. This will penalize the MBB2 path, but will 321 // have a lower overall impact on the program execution. 322 return MBB1Time < MBB2Time; 323} 324 325bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 326 MadeChange = false; 327 328 if (!EnableTailMerge) return false; 329 330 // Find blocks with no successors. 331 std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials; 332 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 333 if (I->succ_empty()) 334 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I), I)); 335 } 336 337 // Sort by hash value so that blocks with identical end sequences sort 338 // together. 339 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 340 341 // Walk through equivalence sets looking for actual exact matches. 342 while (MergePotentials.size() > 1) { 343 unsigned CurHash = (MergePotentials.end()-1)->first; 344 unsigned PrevHash = (MergePotentials.end()-2)->first; 345 MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second; 346 347 // If there is nothing that matches the hash of the current basic block, 348 // give up. 349 if (CurHash != PrevHash) { 350 MergePotentials.pop_back(); 351 continue; 352 } 353 354 // Determine the actual length of the shared tail between these two basic 355 // blocks. Because the hash can have collisions, it's possible that this is 356 // less than 2. 357 MachineBasicBlock::iterator BBI1, BBI2; 358 unsigned CommonTailLen = 359 ComputeCommonTailLength(CurMBB, (MergePotentials.end()-2)->second, 360 BBI1, BBI2); 361 362 // If the tails don't have at least two instructions in common, see if there 363 // is anything else in the equivalence class that does match. 364 if (CommonTailLen < 2) { 365 unsigned FoundMatch = ~0U; 366 for (int i = MergePotentials.size()-2; 367 i != -1 && MergePotentials[i].first == CurHash; --i) { 368 CommonTailLen = ComputeCommonTailLength(CurMBB, 369 MergePotentials[i].second, 370 BBI1, BBI2); 371 if (CommonTailLen >= 2) { 372 FoundMatch = i; 373 break; 374 } 375 } 376 377 // If we didn't find anything that has at least two instructions matching 378 // this one, bail out. 379 if (FoundMatch == ~0U) { 380 MergePotentials.pop_back(); 381 continue; 382 } 383 384 // Otherwise, move the matching block to the right position. 385 std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2)); 386 } 387 388 MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second; 389 390 // If neither block is the entire common tail, split the tail of one block 391 // to make it redundant with the other tail. 392 if (CurMBB->begin() != BBI1 && MBB2->begin() != BBI2) { 393 if (0) { // Enable this to disable partial tail merges. 394 MergePotentials.pop_back(); 395 continue; 396 } 397 398 // Decide whether we want to split CurMBB or MBB2. 399 if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII)) { 400 CurMBB = SplitMBBAt(*CurMBB, BBI1); 401 BBI1 = CurMBB->begin(); 402 MergePotentials.back().second = CurMBB; 403 } else { 404 MBB2 = SplitMBBAt(*MBB2, BBI2); 405 BBI2 = MBB2->begin(); 406 (MergePotentials.end()-2)->second = MBB2; 407 } 408 } 409 410 if (MBB2->begin() == BBI2) { 411 // Hack the end off CurMBB, making it jump to MBBI@ instead. 412 ReplaceTailWithBranchTo(BBI1, MBB2); 413 // This modifies CurMBB, so remove it from the worklist. 414 MergePotentials.pop_back(); 415 } else { 416 assert(CurMBB->begin() == BBI1 && "Didn't split block correctly?"); 417 // Hack the end off MBB2, making it jump to CurMBB instead. 418 ReplaceTailWithBranchTo(BBI2, CurMBB); 419 // This modifies MBB2, so remove it from the worklist. 420 MergePotentials.erase(MergePotentials.end()-2); 421 } 422 MadeChange = true; 423 } 424 425 return MadeChange; 426} 427 428 429//===----------------------------------------------------------------------===// 430// Branch Optimization 431//===----------------------------------------------------------------------===// 432 433bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 434 MadeChange = false; 435 436 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 437 MachineBasicBlock *MBB = I++; 438 OptimizeBlock(MBB); 439 440 // If it is dead, remove it. 441 if (MBB->pred_empty()) { 442 RemoveDeadBlock(MBB); 443 MadeChange = true; 444 ++NumDeadBlocks; 445 } 446 } 447 return MadeChange; 448} 449 450 451/// CorrectExtraCFGEdges - Various pieces of code can cause excess edges in the 452/// CFG to be inserted. If we have proven that MBB can only branch to DestA and 453/// DestB, remove any other MBB successors from the CFG. DestA and DestB can 454/// be null. 455static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB, 456 MachineBasicBlock *DestA, 457 MachineBasicBlock *DestB, 458 bool isCond, 459 MachineFunction::iterator FallThru) { 460 bool MadeChange = false; 461 bool AddedFallThrough = false; 462 463 // If this block ends with a conditional branch that falls through to its 464 // successor, set DestB as the successor. 465 if (isCond) { 466 if (DestB == 0 && FallThru != MBB.getParent()->end()) { 467 DestB = FallThru; 468 AddedFallThrough = true; 469 } 470 } else { 471 // If this is an unconditional branch with no explicit dest, it must just be 472 // a fallthrough into DestB. 473 if (DestA == 0 && FallThru != MBB.getParent()->end()) { 474 DestA = FallThru; 475 AddedFallThrough = true; 476 } 477 } 478 479 MachineBasicBlock::pred_iterator SI = MBB.succ_begin(); 480 while (SI != MBB.succ_end()) { 481 if (*SI == DestA) { 482 DestA = 0; 483 ++SI; 484 } else if (*SI == DestB) { 485 DestB = 0; 486 ++SI; 487 } else { 488 // Otherwise, this is a superfluous edge, remove it. 489 MBB.removeSuccessor(SI); 490 MadeChange = true; 491 } 492 } 493 if (!AddedFallThrough) { 494 assert(DestA == 0 && DestB == 0 && 495 "MachineCFG is missing edges!"); 496 } else if (isCond) { 497 assert(DestA == 0 && "MachineCFG is missing edges!"); 498 } 499 return MadeChange; 500} 501 502 503/// ReplaceUsesOfBlockWith - Given a machine basic block 'BB' that branched to 504/// 'Old', change the code and CFG so that it branches to 'New' instead. 505static void ReplaceUsesOfBlockWith(MachineBasicBlock *BB, 506 MachineBasicBlock *Old, 507 MachineBasicBlock *New, 508 const TargetInstrInfo *TII) { 509 assert(Old != New && "Cannot replace self with self!"); 510 511 MachineBasicBlock::iterator I = BB->end(); 512 while (I != BB->begin()) { 513 --I; 514 if (!TII->isTerminatorInstr(I->getOpcode())) break; 515 516 // Scan the operands of this machine instruction, replacing any uses of Old 517 // with New. 518 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 519 if (I->getOperand(i).isMachineBasicBlock() && 520 I->getOperand(i).getMachineBasicBlock() == Old) 521 I->getOperand(i).setMachineBasicBlock(New); 522 } 523 524 // Update the successor information. 525 std::vector<MachineBasicBlock*> Succs(BB->succ_begin(), BB->succ_end()); 526 for (int i = Succs.size()-1; i >= 0; --i) 527 if (Succs[i] == Old) { 528 BB->removeSuccessor(Old); 529 BB->addSuccessor(New); 530 } 531} 532 533/// CanFallThrough - Return true if the specified block (with the specified 534/// branch condition) can implicitly transfer control to the block after it by 535/// falling off the end of it. This should return false if it can reach the 536/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 537/// 538/// True is a conservative answer. 539/// 540bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 541 bool BranchUnAnalyzable, 542 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 543 const std::vector<MachineOperand> &Cond) { 544 MachineFunction::iterator Fallthrough = CurBB; 545 ++Fallthrough; 546 // If FallthroughBlock is off the end of the function, it can't fall through. 547 if (Fallthrough == CurBB->getParent()->end()) 548 return false; 549 550 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 551 if (!CurBB->isSuccessor(Fallthrough)) 552 return false; 553 554 // If we couldn't analyze the branch, assume it could fall through. 555 if (BranchUnAnalyzable) return true; 556 557 // If there is no branch, control always falls through. 558 if (TBB == 0) return true; 559 560 // If there is some explicit branch to the fallthrough block, it can obviously 561 // reach, even though the branch should get folded to fall through implicitly. 562 if (MachineFunction::iterator(TBB) == Fallthrough || 563 MachineFunction::iterator(FBB) == Fallthrough) 564 return true; 565 566 // If it's an unconditional branch to some block not the fall through, it 567 // doesn't fall through. 568 if (Cond.empty()) return false; 569 570 // Otherwise, if it is conditional and has no explicit false block, it falls 571 // through. 572 return FBB == 0; 573} 574 575/// CanFallThrough - Return true if the specified can implicitly transfer 576/// control to the block after it by falling off the end of it. This should 577/// return false if it can reach the block after it, but it uses an explicit 578/// branch to do so (e.g. a table jump). 579/// 580/// True is a conservative answer. 581/// 582bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 583 MachineBasicBlock *TBB = 0, *FBB = 0; 584 std::vector<MachineOperand> Cond; 585 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond); 586 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 587} 588 589/// IsBetterFallthrough - Return true if it would be clearly better to 590/// fall-through to MBB1 than to fall through into MBB2. This has to return 591/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 592/// result in infinite loops. 593static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 594 MachineBasicBlock *MBB2, 595 const TargetInstrInfo &TII) { 596 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 597 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 598 // optimize branches that branch to either a return block or an assert block 599 // into a fallthrough to the return. 600 if (MBB1->empty() || MBB2->empty()) return false; 601 602 MachineInstr *MBB1I = --MBB1->end(); 603 MachineInstr *MBB2I = --MBB2->end(); 604 return TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode()); 605} 606 607/// OptimizeBlock - Analyze and optimize control flow related to the specified 608/// block. This is never called on the entry block. 609void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 610 MachineFunction::iterator FallThrough = MBB; 611 ++FallThrough; 612 613 // If this block is empty, make everyone use its fall-through, not the block 614 // explicitly. 615 if (MBB->empty()) { 616 // Dead block? Leave for cleanup later. 617 if (MBB->pred_empty()) return; 618 619 if (FallThrough == MBB->getParent()->end()) { 620 // TODO: Simplify preds to not branch here if possible! 621 } else { 622 // Rewrite all predecessors of the old block to go to the fallthrough 623 // instead. 624 while (!MBB->pred_empty()) { 625 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 626 ReplaceUsesOfBlockWith(Pred, MBB, FallThrough, TII); 627 } 628 629 // If MBB was the target of a jump table, update jump tables to go to the 630 // fallthrough instead. 631 MBB->getParent()->getJumpTableInfo()-> 632 ReplaceMBBInJumpTables(MBB, FallThrough); 633 MadeChange = true; 634 } 635 return; 636 } 637 638 // Check to see if we can simplify the terminator of the block before this 639 // one. 640 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 641 642 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 643 std::vector<MachineOperand> PriorCond; 644 bool PriorUnAnalyzable = 645 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 646 if (!PriorUnAnalyzable) { 647 // If the CFG for the prior block has extra edges, remove them. 648 MadeChange |= CorrectExtraCFGEdges(PrevBB, PriorTBB, PriorFBB, 649 !PriorCond.empty(), MBB); 650 651 // If the previous branch is conditional and both conditions go to the same 652 // destination, remove the branch, replacing it with an unconditional one or 653 // a fall-through. 654 if (PriorTBB && PriorTBB == PriorFBB) { 655 TII->RemoveBranch(PrevBB); 656 PriorCond.clear(); 657 if (PriorTBB != MBB) 658 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 659 MadeChange = true; 660 ++NumBranchOpts; 661 return OptimizeBlock(MBB); 662 } 663 664 // If the previous branch *only* branches to *this* block (conditional or 665 // not) remove the branch. 666 if (PriorTBB == MBB && PriorFBB == 0) { 667 TII->RemoveBranch(PrevBB); 668 MadeChange = true; 669 ++NumBranchOpts; 670 return OptimizeBlock(MBB); 671 } 672 673 // If the prior block branches somewhere else on the condition and here if 674 // the condition is false, remove the uncond second branch. 675 if (PriorFBB == MBB) { 676 TII->RemoveBranch(PrevBB); 677 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 678 MadeChange = true; 679 ++NumBranchOpts; 680 return OptimizeBlock(MBB); 681 } 682 683 // If the prior block branches here on true and somewhere else on false, and 684 // if the branch condition is reversible, reverse the branch to create a 685 // fall-through. 686 if (PriorTBB == MBB) { 687 std::vector<MachineOperand> NewPriorCond(PriorCond); 688 if (!TII->ReverseBranchCondition(NewPriorCond)) { 689 TII->RemoveBranch(PrevBB); 690 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 691 MadeChange = true; 692 ++NumBranchOpts; 693 return OptimizeBlock(MBB); 694 } 695 } 696 697 // If this block doesn't fall through (e.g. it ends with an uncond branch or 698 // has no successors) and if the pred falls through into this block, and if 699 // it would otherwise fall through into the block after this, move this 700 // block to the end of the function. 701 // 702 // We consider it more likely that execution will stay in the function (e.g. 703 // due to loops) than it is to exit it. This asserts in loops etc, moving 704 // the assert condition out of the loop body. 705 if (!PriorCond.empty() && PriorFBB == 0 && 706 MachineFunction::iterator(PriorTBB) == FallThrough && 707 !CanFallThrough(MBB)) { 708 // We have to be careful that the succs of PredBB aren't both no-successor 709 // blocks. If neither have successors and if PredBB is the second from 710 // last block in the function, we'd just keep swapping the two blocks for 711 // last. Only do the swap if one is clearly better to fall through than 712 // the other. 713 if (FallThrough != --MBB->getParent()->end() || 714 IsBetterFallthrough(PriorTBB, MBB, *TII)) { 715 716 // Reverse the branch so we will fall through on the previous true cond. 717 std::vector<MachineOperand> NewPriorCond(PriorCond); 718 if (!TII->ReverseBranchCondition(NewPriorCond)) { 719 TII->RemoveBranch(PrevBB); 720 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 721 722 // Move this block to the end of the function. 723 MBB->moveAfter(--MBB->getParent()->end()); 724 MadeChange = true; 725 ++NumBranchOpts; 726 return; 727 } 728 } 729 } 730 } 731 732 // Analyze the branch in the current block. 733 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 734 std::vector<MachineOperand> CurCond; 735 bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond); 736 if (!CurUnAnalyzable) { 737 // If the CFG for the prior block has extra edges, remove them. 738 MadeChange |= CorrectExtraCFGEdges(*MBB, CurTBB, CurFBB, 739 !CurCond.empty(), 740 ++MachineFunction::iterator(MBB)); 741 742 // If this is a two-way branch, and the FBB branches to this block, reverse 743 // the condition so the single-basic-block loop is faster. Instead of: 744 // Loop: xxx; jcc Out; jmp Loop 745 // we want: 746 // Loop: xxx; jncc Loop; jmp Out 747 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 748 std::vector<MachineOperand> NewCond(CurCond); 749 if (!TII->ReverseBranchCondition(NewCond)) { 750 TII->RemoveBranch(*MBB); 751 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 752 MadeChange = true; 753 ++NumBranchOpts; 754 return OptimizeBlock(MBB); 755 } 756 } 757 758 759 // If this branch is the only thing in its block, see if we can forward 760 // other blocks across it. 761 if (CurTBB && CurCond.empty() && CurFBB == 0 && 762 TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) { 763 // This block may contain just an unconditional branch. Because there can 764 // be 'non-branch terminators' in the block, try removing the branch and 765 // then seeing if the block is empty. 766 TII->RemoveBranch(*MBB); 767 768 // If this block is just an unconditional branch to CurTBB, we can 769 // usually completely eliminate the block. The only case we cannot 770 // completely eliminate the block is when the block before this one 771 // falls through into MBB and we can't understand the prior block's branch 772 // condition. 773 if (MBB->empty()) { 774 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 775 if (PredHasNoFallThrough || !PriorUnAnalyzable || 776 !PrevBB.isSuccessor(MBB)) { 777 // If the prior block falls through into us, turn it into an 778 // explicit branch to us to make updates simpler. 779 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 780 PriorTBB != MBB && PriorFBB != MBB) { 781 if (PriorTBB == 0) { 782 assert(PriorCond.empty() && PriorFBB == 0 && 783 "Bad branch analysis"); 784 PriorTBB = MBB; 785 } else { 786 assert(PriorFBB == 0 && "Machine CFG out of date!"); 787 PriorFBB = MBB; 788 } 789 TII->RemoveBranch(PrevBB); 790 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 791 } 792 793 // Iterate through all the predecessors, revectoring each in-turn. 794 MachineBasicBlock::pred_iterator PI = MBB->pred_begin(); 795 bool DidChange = false; 796 bool HasBranchToSelf = false; 797 while (PI != MBB->pred_end()) { 798 if (*PI == MBB) { 799 // If this block has an uncond branch to itself, leave it. 800 ++PI; 801 HasBranchToSelf = true; 802 } else { 803 DidChange = true; 804 ReplaceUsesOfBlockWith(*PI, MBB, CurTBB, TII); 805 } 806 } 807 808 // Change any jumptables to go to the new MBB. 809 MBB->getParent()->getJumpTableInfo()-> 810 ReplaceMBBInJumpTables(MBB, CurTBB); 811 if (DidChange) { 812 ++NumBranchOpts; 813 MadeChange = true; 814 if (!HasBranchToSelf) return; 815 } 816 } 817 } 818 819 // Add the branch back if the block is more than just an uncond branch. 820 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 821 } 822 } 823 824 // If the prior block doesn't fall through into this block, and if this 825 // block doesn't fall through into some other block, see if we can find a 826 // place to move this block where a fall-through will happen. 827 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 828 PriorTBB, PriorFBB, PriorCond)) { 829 // Now we know that there was no fall-through into this block, check to 830 // see if it has a fall-through into its successor. 831 if (!CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, CurCond)) { 832 // Check all the predecessors of this block. If one of them has no fall 833 // throughs, move this block right after it. 834 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 835 E = MBB->pred_end(); PI != E; ++PI) { 836 // Analyze the branch at the end of the pred. 837 MachineBasicBlock *PredBB = *PI; 838 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 839 std::vector<MachineOperand> PredCond; 840 if (PredBB != MBB && !CanFallThrough(PredBB)) { 841 MBB->moveAfter(PredBB); 842 MadeChange = true; 843 return OptimizeBlock(MBB); 844 } 845 } 846 847 // Check all successors to see if we can move this block before it. 848 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 849 E = MBB->succ_end(); SI != E; ++SI) { 850 // Analyze the branch at the end of the block before the succ. 851 MachineBasicBlock *SuccBB = *SI; 852 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 853 std::vector<MachineOperand> SuccPrevCond; 854 if (SuccBB != MBB && !CanFallThrough(SuccPrev)) { 855 MBB->moveBefore(SuccBB); 856 MadeChange = true; 857 return OptimizeBlock(MBB); 858 } 859 } 860 861 // Okay, there is no really great place to put this block. If, however, 862 // the block before this one would be a fall-through if this block were 863 // removed, move this block to the end of the function. 864 if (FallThrough != MBB->getParent()->end() && 865 PrevBB.isSuccessor(FallThrough)) { 866 MBB->moveAfter(--MBB->getParent()->end()); 867 MadeChange = true; 868 return; 869 } 870 } 871 } 872} 873