BranchFolding.cpp revision 14ba0cc42959a3fcc9b6781aea614b01877fb55f
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#define DEBUG_TYPE "branchfolding" 20#include "llvm/CodeGen/Passes.h" 21#include "llvm/CodeGen/MachineModuleInfo.h" 22#include "llvm/CodeGen/MachineFunctionPass.h" 23#include "llvm/CodeGen/MachineJumpTableInfo.h" 24#include "llvm/CodeGen/RegisterScavenging.h" 25#include "llvm/Target/TargetInstrInfo.h" 26#include "llvm/Target/TargetMachine.h" 27#include "llvm/Target/MRegisterInfo.h" 28#include "llvm/Support/CommandLine.h" 29#include "llvm/Support/Debug.h" 30#include "llvm/ADT/Statistic.h" 31#include "llvm/ADT/STLExtras.h" 32#include <algorithm> 33using namespace llvm; 34 35STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 36STATISTIC(NumBranchOpts, "Number of branches optimized"); 37STATISTIC(NumTailMerge , "Number of block tails merged"); 38static cl::opt<bool> EnableTailMerge("enable-tail-merge", cl::Hidden); 39 40namespace { 41 struct BranchFolder : public MachineFunctionPass { 42 static char ID; 43 BranchFolder() : MachineFunctionPass((intptr_t)&ID) {} 44 45 virtual bool runOnMachineFunction(MachineFunction &MF); 46 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 47 const TargetInstrInfo *TII; 48 MachineModuleInfo *MMI; 49 bool MadeChange; 50 private: 51 // Tail Merging. 52 bool TailMergeBlocks(MachineFunction &MF); 53 bool TryMergeBlocks(MachineBasicBlock* SuccBB, 54 MachineBasicBlock* PredBB); 55 void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 56 MachineBasicBlock *NewDest); 57 MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB, 58 MachineBasicBlock::iterator BBI1); 59 60 std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials; 61 const MRegisterInfo *RegInfo; 62 RegScavenger *RS; 63 // Branch optzn. 64 bool OptimizeBranches(MachineFunction &MF); 65 void OptimizeBlock(MachineBasicBlock *MBB); 66 void RemoveDeadBlock(MachineBasicBlock *MBB); 67 68 bool CanFallThrough(MachineBasicBlock *CurBB); 69 bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable, 70 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 71 const std::vector<MachineOperand> &Cond); 72 }; 73 char BranchFolder::ID = 0; 74} 75 76static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB, 77 MachineBasicBlock *DestA, 78 MachineBasicBlock *DestB, 79 bool isCond, 80 MachineFunction::iterator FallThru); 81 82FunctionPass *llvm::createBranchFoldingPass() { return new BranchFolder(); } 83 84/// RemoveDeadBlock - Remove the specified dead machine basic block from the 85/// function, updating the CFG. 86void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 87 assert(MBB->pred_empty() && "MBB must be dead!"); 88 DOUT << "\nRemoving MBB: " << *MBB; 89 90 MachineFunction *MF = MBB->getParent(); 91 // drop all successors. 92 while (!MBB->succ_empty()) 93 MBB->removeSuccessor(MBB->succ_end()-1); 94 95 // If there is DWARF info to active, check to see if there are any LABEL 96 // records in the basic block. If so, unregister them from MachineModuleInfo. 97 if (MMI && !MBB->empty()) { 98 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 99 I != E; ++I) { 100 if ((unsigned)I->getOpcode() == TargetInstrInfo::LABEL) { 101 // The label ID # is always operand #0, an immediate. 102 MMI->InvalidateLabel(I->getOperand(0).getImm()); 103 } 104 } 105 } 106 107 // Remove the block. 108 MF->getBasicBlockList().erase(MBB); 109} 110 111bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { 112 TII = MF.getTarget().getInstrInfo(); 113 if (!TII) return false; 114 115 // Fix CFG. The later algorithms expect it to be right. 116 bool EverMadeChange = false; 117 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 118 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 119 std::vector<MachineOperand> Cond; 120 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond)) 121 EverMadeChange |= CorrectExtraCFGEdges(*MBB, TBB, FBB, 122 !Cond.empty(), next(I)); 123 } 124 125 RegInfo = MF.getTarget().getRegisterInfo(); 126 RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 127 128 MMI = getAnalysisToUpdate<MachineModuleInfo>(); 129 130 bool MadeChangeThisIteration = true; 131 while (MadeChangeThisIteration) { 132 MadeChangeThisIteration = false; 133 MadeChangeThisIteration |= TailMergeBlocks(MF); 134 MadeChangeThisIteration |= OptimizeBranches(MF); 135 EverMadeChange |= MadeChangeThisIteration; 136 } 137 138 // See if any jump tables have become mergable or dead as the code generator 139 // did its thing. 140 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 141 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 142 if (!JTs.empty()) { 143 // Figure out how these jump tables should be merged. 144 std::vector<unsigned> JTMapping; 145 JTMapping.reserve(JTs.size()); 146 147 // We always keep the 0th jump table. 148 JTMapping.push_back(0); 149 150 // Scan the jump tables, seeing if there are any duplicates. Note that this 151 // is N^2, which should be fixed someday. 152 for (unsigned i = 1, e = JTs.size(); i != e; ++i) 153 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 154 155 // If a jump table was merge with another one, walk the function rewriting 156 // references to jump tables to reference the new JT ID's. Keep track of 157 // whether we see a jump table idx, if not, we can delete the JT. 158 std::vector<bool> JTIsLive; 159 JTIsLive.resize(JTs.size()); 160 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 161 BB != E; ++BB) { 162 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 163 I != E; ++I) 164 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 165 MachineOperand &Op = I->getOperand(op); 166 if (!Op.isJumpTableIndex()) continue; 167 unsigned NewIdx = JTMapping[Op.getJumpTableIndex()]; 168 Op.setJumpTableIndex(NewIdx); 169 170 // Remember that this JT is live. 171 JTIsLive[NewIdx] = true; 172 } 173 } 174 175 // Finally, remove dead jump tables. This happens either because the 176 // indirect jump was unreachable (and thus deleted) or because the jump 177 // table was merged with some other one. 178 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 179 if (!JTIsLive[i]) { 180 JTI->RemoveJumpTable(i); 181 EverMadeChange = true; 182 } 183 } 184 185 delete RS; 186 return EverMadeChange; 187} 188 189//===----------------------------------------------------------------------===// 190// Tail Merging of Blocks 191//===----------------------------------------------------------------------===// 192 193/// HashMachineInstr - Compute a hash value for MI and its operands. 194static unsigned HashMachineInstr(const MachineInstr *MI) { 195 unsigned Hash = MI->getOpcode(); 196 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 197 const MachineOperand &Op = MI->getOperand(i); 198 199 // Merge in bits from the operand if easy. 200 unsigned OperandHash = 0; 201 switch (Op.getType()) { 202 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 203 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 204 case MachineOperand::MO_MachineBasicBlock: 205 OperandHash = Op.getMachineBasicBlock()->getNumber(); 206 break; 207 case MachineOperand::MO_FrameIndex: OperandHash = Op.getFrameIndex(); break; 208 case MachineOperand::MO_ConstantPoolIndex: 209 OperandHash = Op.getConstantPoolIndex(); 210 break; 211 case MachineOperand::MO_JumpTableIndex: 212 OperandHash = Op.getJumpTableIndex(); 213 break; 214 case MachineOperand::MO_GlobalAddress: 215 case MachineOperand::MO_ExternalSymbol: 216 // Global address / external symbol are too hard, don't bother, but do 217 // pull in the offset. 218 OperandHash = Op.getOffset(); 219 break; 220 default: break; 221 } 222 223 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 224 } 225 return Hash; 226} 227 228/// HashEndOfMBB - Hash the last two instructions in the MBB. We hash two 229/// instructions, because cross-jumping only saves code when at least two 230/// instructions are removed (since a branch must be inserted). 231static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) { 232 MachineBasicBlock::const_iterator I = MBB->end(); 233 if (I == MBB->begin()) 234 return 0; // Empty MBB. 235 236 --I; 237 unsigned Hash = HashMachineInstr(I); 238 239 if (I == MBB->begin()) 240 return Hash; // Single instr MBB. 241 242 --I; 243 // Hash in the second-to-last instruction. 244 Hash ^= HashMachineInstr(I) << 2; 245 return Hash; 246} 247 248/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 249/// of instructions they actually have in common together at their end. Return 250/// iterators for the first shared instruction in each block. 251static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 252 MachineBasicBlock *MBB2, 253 MachineBasicBlock::iterator &I1, 254 MachineBasicBlock::iterator &I2) { 255 I1 = MBB1->end(); 256 I2 = MBB2->end(); 257 258 unsigned TailLen = 0; 259 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 260 --I1; --I2; 261 if (!I1->isIdenticalTo(I2)) { 262 ++I1; ++I2; 263 break; 264 } 265 ++TailLen; 266 } 267 return TailLen; 268} 269 270/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 271/// after it, replacing it with an unconditional branch to NewDest. This 272/// returns true if OldInst's block is modified, false if NewDest is modified. 273void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 274 MachineBasicBlock *NewDest) { 275 MachineBasicBlock *OldBB = OldInst->getParent(); 276 277 // Remove all the old successors of OldBB from the CFG. 278 while (!OldBB->succ_empty()) 279 OldBB->removeSuccessor(OldBB->succ_begin()); 280 281 // Remove all the dead instructions from the end of OldBB. 282 OldBB->erase(OldInst, OldBB->end()); 283 284 // If OldBB isn't immediately before OldBB, insert a branch to it. 285 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 286 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>()); 287 OldBB->addSuccessor(NewDest); 288 ++NumTailMerge; 289} 290 291/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 292/// MBB so that the part before the iterator falls into the part starting at the 293/// iterator. This returns the new MBB. 294MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 295 MachineBasicBlock::iterator BBI1) { 296 // Create the fall-through block. 297 MachineFunction::iterator MBBI = &CurMBB; 298 MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock()); 299 CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB); 300 301 // Move all the successors of this block to the specified block. 302 while (!CurMBB.succ_empty()) { 303 MachineBasicBlock *S = *(CurMBB.succ_end()-1); 304 NewMBB->addSuccessor(S); 305 CurMBB.removeSuccessor(S); 306 } 307 308 // Add an edge from CurMBB to NewMBB for the fall-through. 309 CurMBB.addSuccessor(NewMBB); 310 311 // Splice the code over. 312 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 313 314 // For targets that use the register scavenger, we must maintain LiveIns. 315 if (RS) { 316 RS->enterBasicBlock(&CurMBB); 317 if (!CurMBB.empty()) 318 RS->forward(prior(CurMBB.end())); 319 BitVector RegsLiveAtExit(RegInfo->getNumRegs()); 320 RS->getRegsUsed(RegsLiveAtExit, false); 321 for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++) 322 if (RegsLiveAtExit[i]) 323 NewMBB->addLiveIn(i); 324 } 325 326 return NewMBB; 327} 328 329/// EstimateRuntime - Make a rough estimate for how long it will take to run 330/// the specified code. 331static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 332 MachineBasicBlock::iterator E, 333 const TargetInstrInfo *TII) { 334 unsigned Time = 0; 335 for (; I != E; ++I) { 336 const TargetInstrDescriptor &TID = TII->get(I->getOpcode()); 337 if (TID.Flags & M_CALL_FLAG) 338 Time += 10; 339 else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG)) 340 Time += 2; 341 else 342 ++Time; 343 } 344 return Time; 345} 346 347/// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at 348/// MBB2I and then insert an unconditional branch in the other block. Determine 349/// which is the best to split 350static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1, 351 MachineBasicBlock::iterator MBB1I, 352 MachineBasicBlock *MBB2, 353 MachineBasicBlock::iterator MBB2I, 354 const TargetInstrInfo *TII, 355 MachineBasicBlock *PredBB) { 356 // If one block is the entry block, split the other one; we can't generate 357 // a branch to the entry block, as its label is not emitted. 358 MachineBasicBlock *Entry = MBB1->getParent()->begin(); 359 if (MBB1 == Entry) 360 return false; 361 if (MBB2 == Entry) 362 return true; 363 364 // If one block falls through into the common successor, choose that 365 // one to split; it is one instruction less to do that. 366 if (PredBB) { 367 if (MBB1 == PredBB) 368 return true; 369 else if (MBB2 == PredBB) 370 return false; 371 } 372 // TODO: if we had some notion of which block was hotter, we could split 373 // the hot block, so it is the fall-through. Since we don't have profile info 374 // make a decision based on which will hurt most to split. 375 unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I, TII); 376 unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I, TII); 377 378 // If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the 379 // MBB1 block so it falls through. This will penalize the MBB2 path, but will 380 // have a lower overall impact on the program execution. 381 return MBB1Time < MBB2Time; 382} 383 384// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 385// branches temporarily for tail merging). In the case where CurMBB ends 386// with a conditional branch to the next block, optimize by reversing the 387// test and conditionally branching to SuccMBB instead. 388 389static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, 390 const TargetInstrInfo *TII) { 391 MachineFunction *MF = CurMBB->getParent(); 392 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 393 MachineBasicBlock *TBB = 0, *FBB = 0; 394 std::vector<MachineOperand> Cond; 395 if (I != MF->end() && 396 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) { 397 MachineBasicBlock *NextBB = I; 398 if (TBB == NextBB && Cond.size() && !FBB) { 399 if (!TII->ReverseBranchCondition(Cond)) { 400 TII->RemoveBranch(*CurMBB); 401 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 402 return; 403 } 404 } 405 } 406 TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector<MachineOperand>()); 407} 408 409// See if any of the blocks in MergePotentials (which all have a common single 410// successor, or all have no successor) can be tail-merged. If there is a 411// successor, any blocks in MergePotentials that are not tail-merged and 412// are not immediately before Succ must have an unconditional branch to 413// Succ added (but the predecessor/successor lists need no adjustment). 414// The lone predecessor of Succ that falls through into Succ, 415// if any, is given in PredBB. 416 417bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, 418 MachineBasicBlock* PredBB) { 419 MadeChange = false; 420 421 // Sort by hash value so that blocks with identical end sequences sort 422 // together. 423 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 424 425 // Walk through equivalence sets looking for actual exact matches. 426 while (MergePotentials.size() > 1) { 427 unsigned CurHash = (MergePotentials.end()-1)->first; 428 unsigned PrevHash = (MergePotentials.end()-2)->first; 429 MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second; 430 431 // If there is nothing that matches the hash of the current basic block, 432 // give up. 433 if (CurHash != PrevHash) { 434 if (SuccBB && CurMBB != PredBB) 435 FixTail(CurMBB, SuccBB, TII); 436 MergePotentials.pop_back(); 437 continue; 438 } 439 440 // Determine the actual length of the shared tail between these two basic 441 // blocks. Because the hash can have collisions, it's possible that this is 442 // less than 2. 443 MachineBasicBlock::iterator BBI1, BBI2; 444 unsigned CommonTailLen = 445 ComputeCommonTailLength(CurMBB, (MergePotentials.end()-2)->second, 446 BBI1, BBI2); 447 448 // If the tails don't have at least two instructions in common, see if there 449 // is anything else in the equivalence class that does match. 450 // Since instructions may get combined later (e.g. single stores into 451 // store multiple) this measure is not particularly accurate. 452 if (CommonTailLen < 2) { 453 unsigned FoundMatch = ~0U; 454 for (int i = MergePotentials.size()-2; 455 i != -1 && MergePotentials[i].first == CurHash; --i) { 456 CommonTailLen = ComputeCommonTailLength(CurMBB, 457 MergePotentials[i].second, 458 BBI1, BBI2); 459 if (CommonTailLen >= 2) { 460 FoundMatch = i; 461 break; 462 } 463 } 464 465 // If we didn't find anything that has at least two instructions matching 466 // this one, bail out. 467 if (FoundMatch == ~0U) { 468 // Put the unconditional branch back, if we need one. 469 if (SuccBB && CurMBB != PredBB) 470 FixTail(CurMBB, SuccBB, TII); 471 MergePotentials.pop_back(); 472 continue; 473 } 474 475 // Otherwise, move the matching block to the right position. 476 std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2)); 477 } 478 479 MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second; 480 481 // If neither block is the entire common tail, split the tail of one block 482 // to make it redundant with the other tail. Also, we cannot jump to the 483 // entry block, so if one block is the entry block, split the other one. 484 MachineBasicBlock *Entry = CurMBB->getParent()->begin(); 485 if (CurMBB->begin() == BBI1 && CurMBB != Entry) 486 ; // CurMBB is common tail 487 else if (MBB2->begin() == BBI2 && MBB2 != Entry) 488 ; // MBB2 is common tail 489 else { 490 if (0) { // Enable this to disable partial tail merges. 491 MergePotentials.pop_back(); 492 continue; 493 } 494 495 // Decide whether we want to split CurMBB or MBB2. 496 if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII, PredBB)) { 497 CurMBB = SplitMBBAt(*CurMBB, BBI1); 498 BBI1 = CurMBB->begin(); 499 MergePotentials.back().second = CurMBB; 500 } else { 501 MBB2 = SplitMBBAt(*MBB2, BBI2); 502 BBI2 = MBB2->begin(); 503 (MergePotentials.end()-2)->second = MBB2; 504 } 505 } 506 507 if (MBB2->begin() == BBI2 && MBB2 != Entry) { 508 // Hack the end off CurMBB, making it jump to MBBI@ instead. 509 ReplaceTailWithBranchTo(BBI1, MBB2); 510 // This modifies CurMBB, so remove it from the worklist. 511 MergePotentials.pop_back(); 512 } else { 513 assert(CurMBB->begin() == BBI1 && CurMBB != Entry && 514 "Didn't split block correctly?"); 515 // Hack the end off MBB2, making it jump to CurMBB instead. 516 ReplaceTailWithBranchTo(BBI2, CurMBB); 517 // This modifies MBB2, so remove it from the worklist. 518 MergePotentials.erase(MergePotentials.end()-2); 519 } 520 MadeChange = true; 521 } 522 return MadeChange; 523} 524 525bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 526 527 if (!EnableTailMerge) return false; 528 529 MadeChange = false; 530 531 // First find blocks with no successors. 532 MergePotentials.clear(); 533 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 534 if (I->succ_empty()) 535 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I), I)); 536 } 537 // See if we can do any tail merging on those. 538 MadeChange |= TryMergeBlocks(NULL, NULL); 539 540 // Look at blocks (IBB) with multiple predecessors (PBB). 541 // We change each predecessor to a canonical form, by 542 // (1) temporarily removing any unconditional branch from the predecessor 543 // to IBB, and 544 // (2) alter conditional branches so they branch to the other block 545 // not IBB; this may require adding back an unconditional branch to IBB 546 // later, where there wasn't one coming in. E.g. 547 // Bcc IBB 548 // fallthrough to QBB 549 // here becomes 550 // Bncc QBB 551 // with a conceptual B to IBB after that, which never actually exists. 552 // With those changes, we see whether the predecessors' tails match, 553 // and merge them if so. We change things out of canonical form and 554 // back to the way they were later in the process. (OptimizeBranches 555 // would undo some of this, but we can't use it, because we'd get into 556 // a compile-time infinite loop repeatedly doing and undoing the same 557 // transformations.) 558 559 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 560 if (!I->succ_empty() && I->pred_size() >= 2) { 561 MachineBasicBlock *IBB = I; 562 MachineBasicBlock *PredBB = prior(I); 563 MergePotentials.clear(); 564 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), E2 = I->pred_end(); 565 P != E2; ++P) { 566 MachineBasicBlock* PBB = *P; 567 // Skip blocks that loop to themselves, can't tail merge these. 568 if (PBB==IBB) 569 continue; 570 MachineBasicBlock *TBB = 0, *FBB = 0; 571 std::vector<MachineOperand> Cond; 572 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond)) { 573 // Failing case: IBB is the target of a cbr, and 574 // we cannot reverse the branch. 575 std::vector<MachineOperand> NewCond(Cond); 576 if (Cond.size() && TBB==IBB) { 577 if (TII->ReverseBranchCondition(NewCond)) 578 continue; 579 // This is the QBB case described above 580 if (!FBB) 581 FBB = next(MachineFunction::iterator(PBB)); 582 } 583 // Remove the unconditional branch at the end, if any. 584 if (TBB && (Cond.size()==0 || FBB)) { 585 TII->RemoveBranch(*PBB); 586 if (Cond.size()) 587 // reinsert conditional branch only, for now 588 TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond); 589 } 590 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB), *P)); 591 } 592 } 593 if (MergePotentials.size() >= 2) 594 MadeChange |= TryMergeBlocks(I, PredBB); 595 // Reinsert an unconditional branch if needed. 596 // The 1 below can be either an original single predecessor, or a result 597 // of removing blocks in TryMergeBlocks. 598 if (MergePotentials.size()==1 && 599 (MergePotentials.begin())->second != PredBB) 600 FixTail((MergePotentials.begin())->second, I, TII); 601 } 602 } 603 return MadeChange; 604} 605 606//===----------------------------------------------------------------------===// 607// Branch Optimization 608//===----------------------------------------------------------------------===// 609 610bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 611 MadeChange = false; 612 613 // Make sure blocks are numbered in order 614 MF.RenumberBlocks(); 615 616 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 617 MachineBasicBlock *MBB = I++; 618 OptimizeBlock(MBB); 619 620 // If it is dead, remove it. 621 if (MBB->pred_empty()) { 622 RemoveDeadBlock(MBB); 623 MadeChange = true; 624 ++NumDeadBlocks; 625 } 626 } 627 return MadeChange; 628} 629 630 631/// CorrectExtraCFGEdges - Various pieces of code can cause excess edges in the 632/// CFG to be inserted. If we have proven that MBB can only branch to DestA and 633/// DestB, remove any other MBB successors from the CFG. DestA and DestB can 634/// be null. 635static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB, 636 MachineBasicBlock *DestA, 637 MachineBasicBlock *DestB, 638 bool isCond, 639 MachineFunction::iterator FallThru) { 640 bool MadeChange = false; 641 bool AddedFallThrough = false; 642 643 // If this block ends with a conditional branch that falls through to its 644 // successor, set DestB as the successor. 645 if (isCond) { 646 if (DestB == 0 && FallThru != MBB.getParent()->end()) { 647 DestB = FallThru; 648 AddedFallThrough = true; 649 } 650 } else { 651 // If this is an unconditional branch with no explicit dest, it must just be 652 // a fallthrough into DestB. 653 if (DestA == 0 && FallThru != MBB.getParent()->end()) { 654 DestA = FallThru; 655 AddedFallThrough = true; 656 } 657 } 658 659 MachineBasicBlock::pred_iterator SI = MBB.succ_begin(); 660 while (SI != MBB.succ_end()) { 661 if (*SI == DestA) { 662 DestA = 0; 663 ++SI; 664 } else if (*SI == DestB) { 665 DestB = 0; 666 ++SI; 667 } else if ((*SI)->isLandingPad()) { 668 ++SI; 669 } else { 670 // Otherwise, this is a superfluous edge, remove it. 671 MBB.removeSuccessor(SI); 672 MadeChange = true; 673 } 674 } 675 if (!AddedFallThrough) { 676 assert(DestA == 0 && DestB == 0 && 677 "MachineCFG is missing edges!"); 678 } else if (isCond) { 679 assert(DestA == 0 && "MachineCFG is missing edges!"); 680 } 681 return MadeChange; 682} 683 684 685/// ReplaceUsesOfBlockWith - Given a machine basic block 'BB' that branched to 686/// 'Old', change the code and CFG so that it branches to 'New' instead. 687static void ReplaceUsesOfBlockWith(MachineBasicBlock *BB, 688 MachineBasicBlock *Old, 689 MachineBasicBlock *New, 690 const TargetInstrInfo *TII) { 691 assert(Old != New && "Cannot replace self with self!"); 692 693 MachineBasicBlock::iterator I = BB->end(); 694 while (I != BB->begin()) { 695 --I; 696 if (!TII->isTerminatorInstr(I->getOpcode())) break; 697 698 // Scan the operands of this machine instruction, replacing any uses of Old 699 // with New. 700 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 701 if (I->getOperand(i).isMachineBasicBlock() && 702 I->getOperand(i).getMachineBasicBlock() == Old) 703 I->getOperand(i).setMachineBasicBlock(New); 704 } 705 706 // Update the successor information. 707 std::vector<MachineBasicBlock*> Succs(BB->succ_begin(), BB->succ_end()); 708 for (int i = Succs.size()-1; i >= 0; --i) 709 if (Succs[i] == Old) { 710 BB->removeSuccessor(Old); 711 BB->addSuccessor(New); 712 } 713} 714 715/// CanFallThrough - Return true if the specified block (with the specified 716/// branch condition) can implicitly transfer control to the block after it by 717/// falling off the end of it. This should return false if it can reach the 718/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 719/// 720/// True is a conservative answer. 721/// 722bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 723 bool BranchUnAnalyzable, 724 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 725 const std::vector<MachineOperand> &Cond) { 726 MachineFunction::iterator Fallthrough = CurBB; 727 ++Fallthrough; 728 // If FallthroughBlock is off the end of the function, it can't fall through. 729 if (Fallthrough == CurBB->getParent()->end()) 730 return false; 731 732 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 733 if (!CurBB->isSuccessor(Fallthrough)) 734 return false; 735 736 // If we couldn't analyze the branch, assume it could fall through. 737 if (BranchUnAnalyzable) return true; 738 739 // If there is no branch, control always falls through. 740 if (TBB == 0) return true; 741 742 // If there is some explicit branch to the fallthrough block, it can obviously 743 // reach, even though the branch should get folded to fall through implicitly. 744 if (MachineFunction::iterator(TBB) == Fallthrough || 745 MachineFunction::iterator(FBB) == Fallthrough) 746 return true; 747 748 // If it's an unconditional branch to some block not the fall through, it 749 // doesn't fall through. 750 if (Cond.empty()) return false; 751 752 // Otherwise, if it is conditional and has no explicit false block, it falls 753 // through. 754 return FBB == 0; 755} 756 757/// CanFallThrough - Return true if the specified can implicitly transfer 758/// control to the block after it by falling off the end of it. This should 759/// return false if it can reach the block after it, but it uses an explicit 760/// branch to do so (e.g. a table jump). 761/// 762/// True is a conservative answer. 763/// 764bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 765 MachineBasicBlock *TBB = 0, *FBB = 0; 766 std::vector<MachineOperand> Cond; 767 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond); 768 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 769} 770 771/// IsBetterFallthrough - Return true if it would be clearly better to 772/// fall-through to MBB1 than to fall through into MBB2. This has to return 773/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 774/// result in infinite loops. 775static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 776 MachineBasicBlock *MBB2, 777 const TargetInstrInfo &TII) { 778 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 779 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 780 // optimize branches that branch to either a return block or an assert block 781 // into a fallthrough to the return. 782 if (MBB1->empty() || MBB2->empty()) return false; 783 784 MachineInstr *MBB1I = --MBB1->end(); 785 MachineInstr *MBB2I = --MBB2->end(); 786 return TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode()); 787} 788 789/// OptimizeBlock - Analyze and optimize control flow related to the specified 790/// block. This is never called on the entry block. 791void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 792 MachineFunction::iterator FallThrough = MBB; 793 ++FallThrough; 794 795 // If this block is empty, make everyone use its fall-through, not the block 796 // explicitly. 797 if (MBB->empty()) { 798 // Dead block? Leave for cleanup later. 799 if (MBB->pred_empty()) return; 800 801 if (FallThrough == MBB->getParent()->end()) { 802 // TODO: Simplify preds to not branch here if possible! 803 } else { 804 // Rewrite all predecessors of the old block to go to the fallthrough 805 // instead. 806 while (!MBB->pred_empty()) { 807 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 808 ReplaceUsesOfBlockWith(Pred, MBB, FallThrough, TII); 809 } 810 811 // If MBB was the target of a jump table, update jump tables to go to the 812 // fallthrough instead. 813 MBB->getParent()->getJumpTableInfo()-> 814 ReplaceMBBInJumpTables(MBB, FallThrough); 815 MadeChange = true; 816 } 817 return; 818 } 819 820 // Check to see if we can simplify the terminator of the block before this 821 // one. 822 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 823 824 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 825 std::vector<MachineOperand> PriorCond; 826 bool PriorUnAnalyzable = 827 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 828 if (!PriorUnAnalyzable) { 829 // If the CFG for the prior block has extra edges, remove them. 830 MadeChange |= CorrectExtraCFGEdges(PrevBB, PriorTBB, PriorFBB, 831 !PriorCond.empty(), MBB); 832 833 // If the previous branch is conditional and both conditions go to the same 834 // destination, remove the branch, replacing it with an unconditional one or 835 // a fall-through. 836 if (PriorTBB && PriorTBB == PriorFBB) { 837 TII->RemoveBranch(PrevBB); 838 PriorCond.clear(); 839 if (PriorTBB != MBB) 840 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 841 MadeChange = true; 842 ++NumBranchOpts; 843 return OptimizeBlock(MBB); 844 } 845 846 // If the previous branch *only* branches to *this* block (conditional or 847 // not) remove the branch. 848 if (PriorTBB == MBB && PriorFBB == 0) { 849 TII->RemoveBranch(PrevBB); 850 MadeChange = true; 851 ++NumBranchOpts; 852 return OptimizeBlock(MBB); 853 } 854 855 // If the prior block branches somewhere else on the condition and here if 856 // the condition is false, remove the uncond second branch. 857 if (PriorFBB == MBB) { 858 TII->RemoveBranch(PrevBB); 859 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 860 MadeChange = true; 861 ++NumBranchOpts; 862 return OptimizeBlock(MBB); 863 } 864 865 // If the prior block branches here on true and somewhere else on false, and 866 // if the branch condition is reversible, reverse the branch to create a 867 // fall-through. 868 if (PriorTBB == MBB) { 869 std::vector<MachineOperand> NewPriorCond(PriorCond); 870 if (!TII->ReverseBranchCondition(NewPriorCond)) { 871 TII->RemoveBranch(PrevBB); 872 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 873 MadeChange = true; 874 ++NumBranchOpts; 875 return OptimizeBlock(MBB); 876 } 877 } 878 879 // If this block doesn't fall through (e.g. it ends with an uncond branch or 880 // has no successors) and if the pred falls through into this block, and if 881 // it would otherwise fall through into the block after this, move this 882 // block to the end of the function. 883 // 884 // We consider it more likely that execution will stay in the function (e.g. 885 // due to loops) than it is to exit it. This asserts in loops etc, moving 886 // the assert condition out of the loop body. 887 if (!PriorCond.empty() && PriorFBB == 0 && 888 MachineFunction::iterator(PriorTBB) == FallThrough && 889 !CanFallThrough(MBB)) { 890 bool DoTransform = true; 891 892 // We have to be careful that the succs of PredBB aren't both no-successor 893 // blocks. If neither have successors and if PredBB is the second from 894 // last block in the function, we'd just keep swapping the two blocks for 895 // last. Only do the swap if one is clearly better to fall through than 896 // the other. 897 if (FallThrough == --MBB->getParent()->end() && 898 !IsBetterFallthrough(PriorTBB, MBB, *TII)) 899 DoTransform = false; 900 901 // We don't want to do this transformation if we have control flow like: 902 // br cond BB2 903 // BB1: 904 // .. 905 // jmp BBX 906 // BB2: 907 // .. 908 // ret 909 // 910 // In this case, we could actually be moving the return block *into* a 911 // loop! 912 if (DoTransform && !MBB->succ_empty() && 913 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 914 DoTransform = false; 915 916 917 if (DoTransform) { 918 // Reverse the branch so we will fall through on the previous true cond. 919 std::vector<MachineOperand> NewPriorCond(PriorCond); 920 if (!TII->ReverseBranchCondition(NewPriorCond)) { 921 DOUT << "\nMoving MBB: " << *MBB; 922 DOUT << "To make fallthrough to: " << *PriorTBB << "\n"; 923 924 TII->RemoveBranch(PrevBB); 925 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 926 927 // Move this block to the end of the function. 928 MBB->moveAfter(--MBB->getParent()->end()); 929 MadeChange = true; 930 ++NumBranchOpts; 931 return; 932 } 933 } 934 } 935 } 936 937 // Analyze the branch in the current block. 938 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 939 std::vector<MachineOperand> CurCond; 940 bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond); 941 if (!CurUnAnalyzable) { 942 // If the CFG for the prior block has extra edges, remove them. 943 MadeChange |= CorrectExtraCFGEdges(*MBB, CurTBB, CurFBB, 944 !CurCond.empty(), 945 ++MachineFunction::iterator(MBB)); 946 947 // If this is a two-way branch, and the FBB branches to this block, reverse 948 // the condition so the single-basic-block loop is faster. Instead of: 949 // Loop: xxx; jcc Out; jmp Loop 950 // we want: 951 // Loop: xxx; jncc Loop; jmp Out 952 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 953 std::vector<MachineOperand> NewCond(CurCond); 954 if (!TII->ReverseBranchCondition(NewCond)) { 955 TII->RemoveBranch(*MBB); 956 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 957 MadeChange = true; 958 ++NumBranchOpts; 959 return OptimizeBlock(MBB); 960 } 961 } 962 963 964 // If this branch is the only thing in its block, see if we can forward 965 // other blocks across it. 966 if (CurTBB && CurCond.empty() && CurFBB == 0 && 967 TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) { 968 // This block may contain just an unconditional branch. Because there can 969 // be 'non-branch terminators' in the block, try removing the branch and 970 // then seeing if the block is empty. 971 TII->RemoveBranch(*MBB); 972 973 // If this block is just an unconditional branch to CurTBB, we can 974 // usually completely eliminate the block. The only case we cannot 975 // completely eliminate the block is when the block before this one 976 // falls through into MBB and we can't understand the prior block's branch 977 // condition. 978 if (MBB->empty()) { 979 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 980 if (PredHasNoFallThrough || !PriorUnAnalyzable || 981 !PrevBB.isSuccessor(MBB)) { 982 // If the prior block falls through into us, turn it into an 983 // explicit branch to us to make updates simpler. 984 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 985 PriorTBB != MBB && PriorFBB != MBB) { 986 if (PriorTBB == 0) { 987 assert(PriorCond.empty() && PriorFBB == 0 && 988 "Bad branch analysis"); 989 PriorTBB = MBB; 990 } else { 991 assert(PriorFBB == 0 && "Machine CFG out of date!"); 992 PriorFBB = MBB; 993 } 994 TII->RemoveBranch(PrevBB); 995 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 996 } 997 998 // Iterate through all the predecessors, revectoring each in-turn. 999 MachineBasicBlock::pred_iterator PI = MBB->pred_begin(); 1000 bool DidChange = false; 1001 bool HasBranchToSelf = false; 1002 while (PI != MBB->pred_end()) { 1003 if (*PI == MBB) { 1004 // If this block has an uncond branch to itself, leave it. 1005 ++PI; 1006 HasBranchToSelf = true; 1007 } else { 1008 DidChange = true; 1009 ReplaceUsesOfBlockWith(*PI, MBB, CurTBB, TII); 1010 } 1011 } 1012 1013 // Change any jumptables to go to the new MBB. 1014 MBB->getParent()->getJumpTableInfo()-> 1015 ReplaceMBBInJumpTables(MBB, CurTBB); 1016 if (DidChange) { 1017 ++NumBranchOpts; 1018 MadeChange = true; 1019 if (!HasBranchToSelf) return; 1020 } 1021 } 1022 } 1023 1024 // Add the branch back if the block is more than just an uncond branch. 1025 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1026 } 1027 } 1028 1029 // If the prior block doesn't fall through into this block, and if this 1030 // block doesn't fall through into some other block, see if we can find a 1031 // place to move this block where a fall-through will happen. 1032 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 1033 PriorTBB, PriorFBB, PriorCond)) { 1034 // Now we know that there was no fall-through into this block, check to 1035 // see if it has a fall-through into its successor. 1036 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1037 CurCond); 1038 1039 if (!MBB->isLandingPad()) { 1040 // Check all the predecessors of this block. If one of them has no fall 1041 // throughs, move this block right after it. 1042 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1043 E = MBB->pred_end(); PI != E; ++PI) { 1044 // Analyze the branch at the end of the pred. 1045 MachineBasicBlock *PredBB = *PI; 1046 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1047 if (PredBB != MBB && !CanFallThrough(PredBB) 1048 && (!CurFallsThru || !CurTBB || !CurFBB) 1049 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1050 // If the current block doesn't fall through, just move it. 1051 // If the current block can fall through and does not end with a 1052 // conditional branch, we need to append an unconditional jump to 1053 // the (current) next block. To avoid a possible compile-time 1054 // infinite loop, move blocks only backward in this case. 1055 // Also, if there are already 2 branches here, we cannot add a third; 1056 // this means we have the case 1057 // Bcc next 1058 // B elsewhere 1059 // next: 1060 if (CurFallsThru) { 1061 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1062 CurCond.clear(); 1063 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1064 } 1065 MBB->moveAfter(PredBB); 1066 MadeChange = true; 1067 return OptimizeBlock(MBB); 1068 } 1069 } 1070 } 1071 1072 if (!CurFallsThru) { 1073 // Check all successors to see if we can move this block before it. 1074 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1075 E = MBB->succ_end(); SI != E; ++SI) { 1076 // Analyze the branch at the end of the block before the succ. 1077 MachineBasicBlock *SuccBB = *SI; 1078 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1079 std::vector<MachineOperand> SuccPrevCond; 1080 1081 // If this block doesn't already fall-through to that successor, and if 1082 // the succ doesn't already have a block that can fall through into it, 1083 // and if the successor isn't an EH destination, we can arrange for the 1084 // fallthrough to happen. 1085 if (SuccBB != MBB && !CanFallThrough(SuccPrev) && 1086 !SuccBB->isLandingPad()) { 1087 MBB->moveBefore(SuccBB); 1088 MadeChange = true; 1089 return OptimizeBlock(MBB); 1090 } 1091 } 1092 1093 // Okay, there is no really great place to put this block. If, however, 1094 // the block before this one would be a fall-through if this block were 1095 // removed, move this block to the end of the function. 1096 if (FallThrough != MBB->getParent()->end() && 1097 PrevBB.isSuccessor(FallThrough)) { 1098 MBB->moveAfter(--MBB->getParent()->end()); 1099 MadeChange = true; 1100 return; 1101 } 1102 } 1103 } 1104} 1105