BranchFolding.cpp revision de0963df420f408b2d813bd16f3c040c2d189f63
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<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 39 cl::init(cl::BOU_UNSET), cl::Hidden); 40namespace { 41 struct BranchFolder : public MachineFunctionPass { 42 static char ID; 43 BranchFolder(bool defaultEnableTailMerge) : 44 MachineFunctionPass((intptr_t)&ID) { 45 switch (FlagEnableTailMerge) { 46 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 47 case cl::BOU_TRUE: EnableTailMerge = true; break; 48 case cl::BOU_FALSE: EnableTailMerge = false; break; 49 } 50 } 51 52 virtual bool runOnMachineFunction(MachineFunction &MF); 53 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 54 const TargetInstrInfo *TII; 55 MachineModuleInfo *MMI; 56 bool MadeChange; 57 private: 58 // Tail Merging. 59 bool EnableTailMerge; 60 bool TailMergeBlocks(MachineFunction &MF); 61 bool TryMergeBlocks(MachineBasicBlock* SuccBB, 62 MachineBasicBlock* PredBB); 63 void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 64 MachineBasicBlock *NewDest); 65 MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB, 66 MachineBasicBlock::iterator BBI1); 67 68 std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials; 69 const MRegisterInfo *RegInfo; 70 RegScavenger *RS; 71 // Branch optzn. 72 bool OptimizeBranches(MachineFunction &MF); 73 void OptimizeBlock(MachineBasicBlock *MBB); 74 void RemoveDeadBlock(MachineBasicBlock *MBB); 75 76 bool CanFallThrough(MachineBasicBlock *CurBB); 77 bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable, 78 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 79 const std::vector<MachineOperand> &Cond); 80 }; 81 char BranchFolder::ID = 0; 82} 83 84static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB, 85 MachineBasicBlock *DestA, 86 MachineBasicBlock *DestB, 87 bool isCond, 88 MachineFunction::iterator FallThru); 89 90FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 91 return new BranchFolder(DefaultEnableTailMerge); } 92 93/// RemoveDeadBlock - Remove the specified dead machine basic block from the 94/// function, updating the CFG. 95void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 96 assert(MBB->pred_empty() && "MBB must be dead!"); 97 DOUT << "\nRemoving MBB: " << *MBB; 98 99 MachineFunction *MF = MBB->getParent(); 100 // drop all successors. 101 while (!MBB->succ_empty()) 102 MBB->removeSuccessor(MBB->succ_end()-1); 103 104 // If there is DWARF info to active, check to see if there are any LABEL 105 // records in the basic block. If so, unregister them from MachineModuleInfo. 106 if (MMI && !MBB->empty()) { 107 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 108 I != E; ++I) { 109 if ((unsigned)I->getOpcode() == TargetInstrInfo::LABEL) { 110 // The label ID # is always operand #0, an immediate. 111 MMI->InvalidateLabel(I->getOperand(0).getImm()); 112 } 113 } 114 } 115 116 // Remove the block. 117 MF->getBasicBlockList().erase(MBB); 118} 119 120bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { 121 TII = MF.getTarget().getInstrInfo(); 122 if (!TII) return false; 123 124 // Fix CFG. The later algorithms expect it to be right. 125 bool EverMadeChange = false; 126 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 127 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 128 std::vector<MachineOperand> Cond; 129 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond)) 130 EverMadeChange |= CorrectExtraCFGEdges(*MBB, TBB, FBB, 131 !Cond.empty(), next(I)); 132 } 133 134 RegInfo = MF.getTarget().getRegisterInfo(); 135 RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 136 137 MMI = getAnalysisToUpdate<MachineModuleInfo>(); 138 139 bool MadeChangeThisIteration = true; 140 while (MadeChangeThisIteration) { 141 MadeChangeThisIteration = false; 142 MadeChangeThisIteration |= TailMergeBlocks(MF); 143 MadeChangeThisIteration |= OptimizeBranches(MF); 144 EverMadeChange |= MadeChangeThisIteration; 145 } 146 147 // See if any jump tables have become mergable or dead as the code generator 148 // did its thing. 149 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 150 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 151 if (!JTs.empty()) { 152 // Figure out how these jump tables should be merged. 153 std::vector<unsigned> JTMapping; 154 JTMapping.reserve(JTs.size()); 155 156 // We always keep the 0th jump table. 157 JTMapping.push_back(0); 158 159 // Scan the jump tables, seeing if there are any duplicates. Note that this 160 // is N^2, which should be fixed someday. 161 for (unsigned i = 1, e = JTs.size(); i != e; ++i) 162 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 163 164 // If a jump table was merge with another one, walk the function rewriting 165 // references to jump tables to reference the new JT ID's. Keep track of 166 // whether we see a jump table idx, if not, we can delete the JT. 167 std::vector<bool> JTIsLive; 168 JTIsLive.resize(JTs.size()); 169 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 170 BB != E; ++BB) { 171 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 172 I != E; ++I) 173 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 174 MachineOperand &Op = I->getOperand(op); 175 if (!Op.isJumpTableIndex()) continue; 176 unsigned NewIdx = JTMapping[Op.getJumpTableIndex()]; 177 Op.setJumpTableIndex(NewIdx); 178 179 // Remember that this JT is live. 180 JTIsLive[NewIdx] = true; 181 } 182 } 183 184 // Finally, remove dead jump tables. This happens either because the 185 // indirect jump was unreachable (and thus deleted) or because the jump 186 // table was merged with some other one. 187 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 188 if (!JTIsLive[i]) { 189 JTI->RemoveJumpTable(i); 190 EverMadeChange = true; 191 } 192 } 193 194 delete RS; 195 return EverMadeChange; 196} 197 198//===----------------------------------------------------------------------===// 199// Tail Merging of Blocks 200//===----------------------------------------------------------------------===// 201 202/// HashMachineInstr - Compute a hash value for MI and its operands. 203static unsigned HashMachineInstr(const MachineInstr *MI) { 204 unsigned Hash = MI->getOpcode(); 205 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 206 const MachineOperand &Op = MI->getOperand(i); 207 208 // Merge in bits from the operand if easy. 209 unsigned OperandHash = 0; 210 switch (Op.getType()) { 211 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 212 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 213 case MachineOperand::MO_MachineBasicBlock: 214 OperandHash = Op.getMachineBasicBlock()->getNumber(); 215 break; 216 case MachineOperand::MO_FrameIndex: OperandHash = Op.getFrameIndex(); break; 217 case MachineOperand::MO_ConstantPoolIndex: 218 OperandHash = Op.getConstantPoolIndex(); 219 break; 220 case MachineOperand::MO_JumpTableIndex: 221 OperandHash = Op.getJumpTableIndex(); 222 break; 223 case MachineOperand::MO_GlobalAddress: 224 case MachineOperand::MO_ExternalSymbol: 225 // Global address / external symbol are too hard, don't bother, but do 226 // pull in the offset. 227 OperandHash = Op.getOffset(); 228 break; 229 default: break; 230 } 231 232 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 233 } 234 return Hash; 235} 236 237/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 238/// with no successors, we hash two instructions, because cross-jumping 239/// only saves code when at least two instructions are removed (since a 240/// branch must be inserted). For blocks with a successor, one of the 241/// two blocks to be tail-merged will end with a branch already, so 242/// it gains to cross-jump even for one instruction. 243 244static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 245 unsigned minCommonTailLength) { 246 MachineBasicBlock::const_iterator I = MBB->end(); 247 if (I == MBB->begin()) 248 return 0; // Empty MBB. 249 250 --I; 251 unsigned Hash = HashMachineInstr(I); 252 253 if (I == MBB->begin() || minCommonTailLength == 1) 254 return Hash; // Single instr MBB. 255 256 --I; 257 // Hash in the second-to-last instruction. 258 Hash ^= HashMachineInstr(I) << 2; 259 return Hash; 260} 261 262/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 263/// of instructions they actually have in common together at their end. Return 264/// iterators for the first shared instruction in each block. 265static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 266 MachineBasicBlock *MBB2, 267 MachineBasicBlock::iterator &I1, 268 MachineBasicBlock::iterator &I2) { 269 I1 = MBB1->end(); 270 I2 = MBB2->end(); 271 272 unsigned TailLen = 0; 273 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 274 --I1; --I2; 275 if (!I1->isIdenticalTo(I2)) { 276 ++I1; ++I2; 277 break; 278 } 279 ++TailLen; 280 } 281 return TailLen; 282} 283 284/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 285/// after it, replacing it with an unconditional branch to NewDest. This 286/// returns true if OldInst's block is modified, false if NewDest is modified. 287void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 288 MachineBasicBlock *NewDest) { 289 MachineBasicBlock *OldBB = OldInst->getParent(); 290 291 // Remove all the old successors of OldBB from the CFG. 292 while (!OldBB->succ_empty()) 293 OldBB->removeSuccessor(OldBB->succ_begin()); 294 295 // Remove all the dead instructions from the end of OldBB. 296 OldBB->erase(OldInst, OldBB->end()); 297 298 // If OldBB isn't immediately before OldBB, insert a branch to it. 299 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 300 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>()); 301 OldBB->addSuccessor(NewDest); 302 ++NumTailMerge; 303} 304 305/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 306/// MBB so that the part before the iterator falls into the part starting at the 307/// iterator. This returns the new MBB. 308MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 309 MachineBasicBlock::iterator BBI1) { 310 // Create the fall-through block. 311 MachineFunction::iterator MBBI = &CurMBB; 312 MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock()); 313 CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB); 314 315 // Move all the successors of this block to the specified block. 316 while (!CurMBB.succ_empty()) { 317 MachineBasicBlock *S = *(CurMBB.succ_end()-1); 318 NewMBB->addSuccessor(S); 319 CurMBB.removeSuccessor(S); 320 } 321 322 // Add an edge from CurMBB to NewMBB for the fall-through. 323 CurMBB.addSuccessor(NewMBB); 324 325 // Splice the code over. 326 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 327 328 // For targets that use the register scavenger, we must maintain LiveIns. 329 if (RS) { 330 RS->enterBasicBlock(&CurMBB); 331 if (!CurMBB.empty()) 332 RS->forward(prior(CurMBB.end())); 333 BitVector RegsLiveAtExit(RegInfo->getNumRegs()); 334 RS->getRegsUsed(RegsLiveAtExit, false); 335 for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++) 336 if (RegsLiveAtExit[i]) 337 NewMBB->addLiveIn(i); 338 } 339 340 return NewMBB; 341} 342 343/// EstimateRuntime - Make a rough estimate for how long it will take to run 344/// the specified code. 345static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 346 MachineBasicBlock::iterator E, 347 const TargetInstrInfo *TII) { 348 unsigned Time = 0; 349 for (; I != E; ++I) { 350 const TargetInstrDescriptor &TID = TII->get(I->getOpcode()); 351 if (TID.Flags & M_CALL_FLAG) 352 Time += 10; 353 else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG)) 354 Time += 2; 355 else 356 ++Time; 357 } 358 return Time; 359} 360 361/// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at 362/// MBB2I and then insert an unconditional branch in the other block. Determine 363/// which is the best to split 364static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1, 365 MachineBasicBlock::iterator MBB1I, 366 MachineBasicBlock *MBB2, 367 MachineBasicBlock::iterator MBB2I, 368 const TargetInstrInfo *TII, 369 MachineBasicBlock *PredBB) { 370 // If one block is the entry block, split the other one; we can't generate 371 // a branch to the entry block, as its label is not emitted. 372 MachineBasicBlock *Entry = MBB1->getParent()->begin(); 373 if (MBB1 == Entry) 374 return false; 375 if (MBB2 == Entry) 376 return true; 377 378 // If one block falls through into the common successor, choose that 379 // one to split; it is one instruction less to do that. 380 if (PredBB) { 381 if (MBB1 == PredBB) 382 return true; 383 else if (MBB2 == PredBB) 384 return false; 385 } 386 // TODO: if we had some notion of which block was hotter, we could split 387 // the hot block, so it is the fall-through. Since we don't have profile info 388 // make a decision based on which will hurt most to split. 389 unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I, TII); 390 unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I, TII); 391 392 // If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the 393 // MBB1 block so it falls through. This will penalize the MBB2 path, but will 394 // have a lower overall impact on the program execution. 395 return MBB1Time < MBB2Time; 396} 397 398// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 399// branches temporarily for tail merging). In the case where CurMBB ends 400// with a conditional branch to the next block, optimize by reversing the 401// test and conditionally branching to SuccMBB instead. 402 403static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, 404 const TargetInstrInfo *TII) { 405 MachineFunction *MF = CurMBB->getParent(); 406 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 407 MachineBasicBlock *TBB = 0, *FBB = 0; 408 std::vector<MachineOperand> Cond; 409 if (I != MF->end() && 410 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) { 411 MachineBasicBlock *NextBB = I; 412 if (TBB == NextBB && Cond.size() && !FBB) { 413 if (!TII->ReverseBranchCondition(Cond)) { 414 TII->RemoveBranch(*CurMBB); 415 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 416 return; 417 } 418 } 419 } 420 TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector<MachineOperand>()); 421} 422 423// See if any of the blocks in MergePotentials (which all have a common single 424// successor, or all have no successor) can be tail-merged. If there is a 425// successor, any blocks in MergePotentials that are not tail-merged and 426// are not immediately before Succ must have an unconditional branch to 427// Succ added (but the predecessor/successor lists need no adjustment). 428// The lone predecessor of Succ that falls through into Succ, 429// if any, is given in PredBB. 430 431bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, 432 MachineBasicBlock* PredBB) { 433 unsigned minCommonTailLength = (SuccBB ? 1 : 2); 434 MadeChange = false; 435 436 // Sort by hash value so that blocks with identical end sequences sort 437 // together. 438 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 439 440 // Walk through equivalence sets looking for actual exact matches. 441 while (MergePotentials.size() > 1) { 442 unsigned CurHash = (MergePotentials.end()-1)->first; 443 unsigned PrevHash = (MergePotentials.end()-2)->first; 444 MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second; 445 446 // If there is nothing that matches the hash of the current basic block, 447 // give up. 448 if (CurHash != PrevHash) { 449 if (SuccBB && CurMBB != PredBB) 450 FixTail(CurMBB, SuccBB, TII); 451 MergePotentials.pop_back(); 452 continue; 453 } 454 455 // Look through all the blocks that have the same hash as this one, and 456 // find the one that has the largest number of instructions in common. 457 // Since instructions may get combined later (e.g. single stores into 458 // store multiple) this measure is not particularly accurate. 459 MachineBasicBlock::iterator BBI1, BBI2; 460 461 unsigned FoundMatch = ~0U; 462 unsigned maxCommonTailLength = 0U; 463 for (int i = MergePotentials.size()-2; 464 i != -1 && MergePotentials[i].first == CurHash; --i) { 465 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 466 unsigned CommonTailLen = ComputeCommonTailLength(CurMBB, 467 MergePotentials[i].second, 468 TrialBBI1, TrialBBI2); 469 if (CommonTailLen >= minCommonTailLength && 470 CommonTailLen >= maxCommonTailLength) { 471 FoundMatch = i; 472 maxCommonTailLength = CommonTailLen; 473 BBI1 = TrialBBI1; 474 BBI2 = TrialBBI2; 475 } 476 } 477 478 // If we didn't find anything that has at least minCommonTailLength 479 // instructions matching this one, bail out. 480 if (FoundMatch == ~0U) { 481 // Put the unconditional branch back, if we need one. 482 if (SuccBB && CurMBB != PredBB) 483 FixTail(CurMBB, SuccBB, TII); 484 MergePotentials.pop_back(); 485 continue; 486 } 487 488 // Otherwise, move the matching block to the right position. 489 if (FoundMatch != MergePotentials.size()-2) 490 std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2)); 491 492 MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second; 493 494 // If neither block is the entire common tail, split the tail of one block 495 // to make it redundant with the other tail. Also, we cannot jump to the 496 // entry block, so if one block is the entry block, split the other one. 497 MachineBasicBlock *Entry = CurMBB->getParent()->begin(); 498 if (CurMBB->begin() == BBI1 && CurMBB != Entry) 499 ; // CurMBB is common tail 500 else if (MBB2->begin() == BBI2 && MBB2 != Entry) 501 ; // MBB2 is common tail 502 else { 503 if (0) { // Enable this to disable partial tail merges. 504 MergePotentials.pop_back(); 505 continue; 506 } 507 508 // Decide whether we want to split CurMBB or MBB2. 509 if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII, PredBB)) { 510 CurMBB = SplitMBBAt(*CurMBB, BBI1); 511 BBI1 = CurMBB->begin(); 512 MergePotentials.back().second = CurMBB; 513 } else { 514 MBB2 = SplitMBBAt(*MBB2, BBI2); 515 BBI2 = MBB2->begin(); 516 (MergePotentials.end()-2)->second = MBB2; 517 } 518 } 519 520 if (MBB2->begin() == BBI2 && MBB2 != Entry) { 521 // Hack the end off CurMBB, making it jump to MBBI@ instead. 522 ReplaceTailWithBranchTo(BBI1, MBB2); 523 // This modifies CurMBB, so remove it from the worklist. 524 MergePotentials.pop_back(); 525 } else { 526 assert(CurMBB->begin() == BBI1 && CurMBB != Entry && 527 "Didn't split block correctly?"); 528 // Hack the end off MBB2, making it jump to CurMBB instead. 529 ReplaceTailWithBranchTo(BBI2, CurMBB); 530 // This modifies MBB2, so remove it from the worklist. 531 MergePotentials.erase(MergePotentials.end()-2); 532 } 533 MadeChange = true; 534 } 535 return MadeChange; 536} 537 538bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 539 540 if (!EnableTailMerge) return false; 541 542 MadeChange = false; 543 544 // First find blocks with no successors. 545 MergePotentials.clear(); 546 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 547 if (I->succ_empty()) 548 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I)); 549 } 550 // See if we can do any tail merging on those. 551 MadeChange |= TryMergeBlocks(NULL, NULL); 552 553 // Look at blocks (IBB) with multiple predecessors (PBB). 554 // We change each predecessor to a canonical form, by 555 // (1) temporarily removing any unconditional branch from the predecessor 556 // to IBB, and 557 // (2) alter conditional branches so they branch to the other block 558 // not IBB; this may require adding back an unconditional branch to IBB 559 // later, where there wasn't one coming in. E.g. 560 // Bcc IBB 561 // fallthrough to QBB 562 // here becomes 563 // Bncc QBB 564 // with a conceptual B to IBB after that, which never actually exists. 565 // With those changes, we see whether the predecessors' tails match, 566 // and merge them if so. We change things out of canonical form and 567 // back to the way they were later in the process. (OptimizeBranches 568 // would undo some of this, but we can't use it, because we'd get into 569 // a compile-time infinite loop repeatedly doing and undoing the same 570 // transformations.) 571 572 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 573 if (!I->succ_empty() && I->pred_size() >= 2) { 574 MachineBasicBlock *IBB = I; 575 MachineBasicBlock *PredBB = prior(I); 576 MergePotentials.clear(); 577 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), E2 = I->pred_end(); 578 P != E2; ++P) { 579 MachineBasicBlock* PBB = *P; 580 // Skip blocks that loop to themselves, can't tail merge these. 581 if (PBB==IBB) 582 continue; 583 MachineBasicBlock *TBB = 0, *FBB = 0; 584 std::vector<MachineOperand> Cond; 585 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond)) { 586 // Failing case: IBB is the target of a cbr, and 587 // we cannot reverse the branch. 588 std::vector<MachineOperand> NewCond(Cond); 589 if (Cond.size() && TBB==IBB) { 590 if (TII->ReverseBranchCondition(NewCond)) 591 continue; 592 // This is the QBB case described above 593 if (!FBB) 594 FBB = next(MachineFunction::iterator(PBB)); 595 } 596 // Remove the unconditional branch at the end, if any. 597 if (TBB && (Cond.size()==0 || FBB)) { 598 TII->RemoveBranch(*PBB); 599 if (Cond.size()) 600 // reinsert conditional branch only, for now 601 TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond); 602 } 603 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P)); 604 } 605 } 606 if (MergePotentials.size() >= 2) 607 MadeChange |= TryMergeBlocks(I, PredBB); 608 // Reinsert an unconditional branch if needed. 609 // The 1 below can be either an original single predecessor, or a result 610 // of removing blocks in TryMergeBlocks. 611 PredBB = prior(I); // this may have been changed in TryMergeBlocks 612 if (MergePotentials.size()==1 && 613 (MergePotentials.begin())->second != PredBB) 614 FixTail((MergePotentials.begin())->second, I, TII); 615 } 616 } 617 return MadeChange; 618} 619 620//===----------------------------------------------------------------------===// 621// Branch Optimization 622//===----------------------------------------------------------------------===// 623 624bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 625 MadeChange = false; 626 627 // Make sure blocks are numbered in order 628 MF.RenumberBlocks(); 629 630 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 631 MachineBasicBlock *MBB = I++; 632 OptimizeBlock(MBB); 633 634 // If it is dead, remove it. 635 if (MBB->pred_empty()) { 636 RemoveDeadBlock(MBB); 637 MadeChange = true; 638 ++NumDeadBlocks; 639 } 640 } 641 return MadeChange; 642} 643 644 645/// CorrectExtraCFGEdges - Various pieces of code can cause excess edges in the 646/// CFG to be inserted. If we have proven that MBB can only branch to DestA and 647/// DestB, remove any other MBB successors from the CFG. DestA and DestB can 648/// be null. 649static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB, 650 MachineBasicBlock *DestA, 651 MachineBasicBlock *DestB, 652 bool isCond, 653 MachineFunction::iterator FallThru) { 654 bool MadeChange = false; 655 bool AddedFallThrough = false; 656 657 // If this block ends with a conditional branch that falls through to its 658 // successor, set DestB as the successor. 659 if (isCond) { 660 if (DestB == 0 && FallThru != MBB.getParent()->end()) { 661 DestB = FallThru; 662 AddedFallThrough = true; 663 } 664 } else { 665 // If this is an unconditional branch with no explicit dest, it must just be 666 // a fallthrough into DestB. 667 if (DestA == 0 && FallThru != MBB.getParent()->end()) { 668 DestA = FallThru; 669 AddedFallThrough = true; 670 } 671 } 672 673 MachineBasicBlock::pred_iterator SI = MBB.succ_begin(); 674 while (SI != MBB.succ_end()) { 675 if (*SI == DestA) { 676 DestA = 0; 677 ++SI; 678 } else if (*SI == DestB) { 679 DestB = 0; 680 ++SI; 681 } else if ((*SI)->isLandingPad()) { 682 ++SI; 683 } else { 684 // Otherwise, this is a superfluous edge, remove it. 685 MBB.removeSuccessor(SI); 686 MadeChange = true; 687 } 688 } 689 if (!AddedFallThrough) { 690 assert(DestA == 0 && DestB == 0 && 691 "MachineCFG is missing edges!"); 692 } else if (isCond) { 693 assert(DestA == 0 && "MachineCFG is missing edges!"); 694 } 695 return MadeChange; 696} 697 698 699/// ReplaceUsesOfBlockWith - Given a machine basic block 'BB' that branched to 700/// 'Old', change the code and CFG so that it branches to 'New' instead. 701static void ReplaceUsesOfBlockWith(MachineBasicBlock *BB, 702 MachineBasicBlock *Old, 703 MachineBasicBlock *New, 704 const TargetInstrInfo *TII) { 705 assert(Old != New && "Cannot replace self with self!"); 706 707 MachineBasicBlock::iterator I = BB->end(); 708 while (I != BB->begin()) { 709 --I; 710 if (!TII->isTerminatorInstr(I->getOpcode())) break; 711 712 // Scan the operands of this machine instruction, replacing any uses of Old 713 // with New. 714 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 715 if (I->getOperand(i).isMachineBasicBlock() && 716 I->getOperand(i).getMachineBasicBlock() == Old) 717 I->getOperand(i).setMachineBasicBlock(New); 718 } 719 720 // Update the successor information. If New was already a successor, just 721 // remove the link to Old instead of creating another one. PR 1444. 722 bool HadSuccessorNew = false; 723 std::vector<MachineBasicBlock*> Succs(BB->succ_begin(), BB->succ_end()); 724 for (int i = Succs.size()-1; i >= 0; --i) 725 if (Succs[i] == New) { 726 HadSuccessorNew = true; 727 break; 728 } 729 for (int i = Succs.size()-1; i >= 0; --i) 730 if (Succs[i] == Old) { 731 BB->removeSuccessor(Old); 732 if (!HadSuccessorNew) 733 BB->addSuccessor(New); 734 } 735} 736 737/// CanFallThrough - Return true if the specified block (with the specified 738/// branch condition) can implicitly transfer control to the block after it by 739/// falling off the end of it. This should return false if it can reach the 740/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 741/// 742/// True is a conservative answer. 743/// 744bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 745 bool BranchUnAnalyzable, 746 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 747 const std::vector<MachineOperand> &Cond) { 748 MachineFunction::iterator Fallthrough = CurBB; 749 ++Fallthrough; 750 // If FallthroughBlock is off the end of the function, it can't fall through. 751 if (Fallthrough == CurBB->getParent()->end()) 752 return false; 753 754 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 755 if (!CurBB->isSuccessor(Fallthrough)) 756 return false; 757 758 // If we couldn't analyze the branch, assume it could fall through. 759 if (BranchUnAnalyzable) return true; 760 761 // If there is no branch, control always falls through. 762 if (TBB == 0) return true; 763 764 // If there is some explicit branch to the fallthrough block, it can obviously 765 // reach, even though the branch should get folded to fall through implicitly. 766 if (MachineFunction::iterator(TBB) == Fallthrough || 767 MachineFunction::iterator(FBB) == Fallthrough) 768 return true; 769 770 // If it's an unconditional branch to some block not the fall through, it 771 // doesn't fall through. 772 if (Cond.empty()) return false; 773 774 // Otherwise, if it is conditional and has no explicit false block, it falls 775 // through. 776 return FBB == 0; 777} 778 779/// CanFallThrough - Return true if the specified can implicitly transfer 780/// control to the block after it by falling off the end of it. This should 781/// return false if it can reach the block after it, but it uses an explicit 782/// branch to do so (e.g. a table jump). 783/// 784/// True is a conservative answer. 785/// 786bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 787 MachineBasicBlock *TBB = 0, *FBB = 0; 788 std::vector<MachineOperand> Cond; 789 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond); 790 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 791} 792 793/// IsBetterFallthrough - Return true if it would be clearly better to 794/// fall-through to MBB1 than to fall through into MBB2. This has to return 795/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 796/// result in infinite loops. 797static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 798 MachineBasicBlock *MBB2, 799 const TargetInstrInfo &TII) { 800 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 801 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 802 // optimize branches that branch to either a return block or an assert block 803 // into a fallthrough to the return. 804 if (MBB1->empty() || MBB2->empty()) return false; 805 806 MachineInstr *MBB1I = --MBB1->end(); 807 MachineInstr *MBB2I = --MBB2->end(); 808 return TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode()); 809} 810 811/// OptimizeBlock - Analyze and optimize control flow related to the specified 812/// block. This is never called on the entry block. 813void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 814 MachineFunction::iterator FallThrough = MBB; 815 ++FallThrough; 816 817 // If this block is empty, make everyone use its fall-through, not the block 818 // explicitly. 819 if (MBB->empty()) { 820 // Dead block? Leave for cleanup later. 821 if (MBB->pred_empty()) return; 822 823 if (FallThrough == MBB->getParent()->end()) { 824 // TODO: Simplify preds to not branch here if possible! 825 } else { 826 // Rewrite all predecessors of the old block to go to the fallthrough 827 // instead. 828 while (!MBB->pred_empty()) { 829 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 830 ReplaceUsesOfBlockWith(Pred, MBB, FallThrough, TII); 831 } 832 833 // If MBB was the target of a jump table, update jump tables to go to the 834 // fallthrough instead. 835 MBB->getParent()->getJumpTableInfo()-> 836 ReplaceMBBInJumpTables(MBB, FallThrough); 837 MadeChange = true; 838 } 839 return; 840 } 841 842 // Check to see if we can simplify the terminator of the block before this 843 // one. 844 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 845 846 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 847 std::vector<MachineOperand> PriorCond; 848 bool PriorUnAnalyzable = 849 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 850 if (!PriorUnAnalyzable) { 851 // If the CFG for the prior block has extra edges, remove them. 852 MadeChange |= CorrectExtraCFGEdges(PrevBB, PriorTBB, PriorFBB, 853 !PriorCond.empty(), MBB); 854 855 // If the previous branch is conditional and both conditions go to the same 856 // destination, remove the branch, replacing it with an unconditional one or 857 // a fall-through. 858 if (PriorTBB && PriorTBB == PriorFBB) { 859 TII->RemoveBranch(PrevBB); 860 PriorCond.clear(); 861 if (PriorTBB != MBB) 862 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 863 MadeChange = true; 864 ++NumBranchOpts; 865 return OptimizeBlock(MBB); 866 } 867 868 // If the previous branch *only* branches to *this* block (conditional or 869 // not) remove the branch. 870 if (PriorTBB == MBB && PriorFBB == 0) { 871 TII->RemoveBranch(PrevBB); 872 MadeChange = true; 873 ++NumBranchOpts; 874 return OptimizeBlock(MBB); 875 } 876 877 // If the prior block branches somewhere else on the condition and here if 878 // the condition is false, remove the uncond second branch. 879 if (PriorFBB == MBB) { 880 TII->RemoveBranch(PrevBB); 881 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 882 MadeChange = true; 883 ++NumBranchOpts; 884 return OptimizeBlock(MBB); 885 } 886 887 // If the prior block branches here on true and somewhere else on false, and 888 // if the branch condition is reversible, reverse the branch to create a 889 // fall-through. 890 if (PriorTBB == MBB) { 891 std::vector<MachineOperand> NewPriorCond(PriorCond); 892 if (!TII->ReverseBranchCondition(NewPriorCond)) { 893 TII->RemoveBranch(PrevBB); 894 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 895 MadeChange = true; 896 ++NumBranchOpts; 897 return OptimizeBlock(MBB); 898 } 899 } 900 901 // If this block doesn't fall through (e.g. it ends with an uncond branch or 902 // has no successors) and if the pred falls through into this block, and if 903 // it would otherwise fall through into the block after this, move this 904 // block to the end of the function. 905 // 906 // We consider it more likely that execution will stay in the function (e.g. 907 // due to loops) than it is to exit it. This asserts in loops etc, moving 908 // the assert condition out of the loop body. 909 if (!PriorCond.empty() && PriorFBB == 0 && 910 MachineFunction::iterator(PriorTBB) == FallThrough && 911 !CanFallThrough(MBB)) { 912 bool DoTransform = true; 913 914 // We have to be careful that the succs of PredBB aren't both no-successor 915 // blocks. If neither have successors and if PredBB is the second from 916 // last block in the function, we'd just keep swapping the two blocks for 917 // last. Only do the swap if one is clearly better to fall through than 918 // the other. 919 if (FallThrough == --MBB->getParent()->end() && 920 !IsBetterFallthrough(PriorTBB, MBB, *TII)) 921 DoTransform = false; 922 923 // We don't want to do this transformation if we have control flow like: 924 // br cond BB2 925 // BB1: 926 // .. 927 // jmp BBX 928 // BB2: 929 // .. 930 // ret 931 // 932 // In this case, we could actually be moving the return block *into* a 933 // loop! 934 if (DoTransform && !MBB->succ_empty() && 935 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 936 DoTransform = false; 937 938 939 if (DoTransform) { 940 // Reverse the branch so we will fall through on the previous true cond. 941 std::vector<MachineOperand> NewPriorCond(PriorCond); 942 if (!TII->ReverseBranchCondition(NewPriorCond)) { 943 DOUT << "\nMoving MBB: " << *MBB; 944 DOUT << "To make fallthrough to: " << *PriorTBB << "\n"; 945 946 TII->RemoveBranch(PrevBB); 947 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 948 949 // Move this block to the end of the function. 950 MBB->moveAfter(--MBB->getParent()->end()); 951 MadeChange = true; 952 ++NumBranchOpts; 953 return; 954 } 955 } 956 } 957 } 958 959 // Analyze the branch in the current block. 960 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 961 std::vector<MachineOperand> CurCond; 962 bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond); 963 if (!CurUnAnalyzable) { 964 // If the CFG for the prior block has extra edges, remove them. 965 MadeChange |= CorrectExtraCFGEdges(*MBB, CurTBB, CurFBB, 966 !CurCond.empty(), 967 ++MachineFunction::iterator(MBB)); 968 969 // If this is a two-way branch, and the FBB branches to this block, reverse 970 // the condition so the single-basic-block loop is faster. Instead of: 971 // Loop: xxx; jcc Out; jmp Loop 972 // we want: 973 // Loop: xxx; jncc Loop; jmp Out 974 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 975 std::vector<MachineOperand> NewCond(CurCond); 976 if (!TII->ReverseBranchCondition(NewCond)) { 977 TII->RemoveBranch(*MBB); 978 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 979 MadeChange = true; 980 ++NumBranchOpts; 981 return OptimizeBlock(MBB); 982 } 983 } 984 985 986 // If this branch is the only thing in its block, see if we can forward 987 // other blocks across it. 988 if (CurTBB && CurCond.empty() && CurFBB == 0 && 989 TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) { 990 // This block may contain just an unconditional branch. Because there can 991 // be 'non-branch terminators' in the block, try removing the branch and 992 // then seeing if the block is empty. 993 TII->RemoveBranch(*MBB); 994 995 // If this block is just an unconditional branch to CurTBB, we can 996 // usually completely eliminate the block. The only case we cannot 997 // completely eliminate the block is when the block before this one 998 // falls through into MBB and we can't understand the prior block's branch 999 // condition. 1000 if (MBB->empty()) { 1001 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 1002 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1003 !PrevBB.isSuccessor(MBB)) { 1004 // If the prior block falls through into us, turn it into an 1005 // explicit branch to us to make updates simpler. 1006 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1007 PriorTBB != MBB && PriorFBB != MBB) { 1008 if (PriorTBB == 0) { 1009 assert(PriorCond.empty() && PriorFBB == 0 && 1010 "Bad branch analysis"); 1011 PriorTBB = MBB; 1012 } else { 1013 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1014 PriorFBB = MBB; 1015 } 1016 TII->RemoveBranch(PrevBB); 1017 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1018 } 1019 1020 // Iterate through all the predecessors, revectoring each in-turn. 1021 MachineBasicBlock::pred_iterator PI = MBB->pred_begin(); 1022 bool DidChange = false; 1023 bool HasBranchToSelf = false; 1024 while (PI != MBB->pred_end()) { 1025 if (*PI == MBB) { 1026 // If this block has an uncond branch to itself, leave it. 1027 ++PI; 1028 HasBranchToSelf = true; 1029 } else { 1030 DidChange = true; 1031 ReplaceUsesOfBlockWith(*PI, MBB, CurTBB, TII); 1032 } 1033 } 1034 1035 // Change any jumptables to go to the new MBB. 1036 MBB->getParent()->getJumpTableInfo()-> 1037 ReplaceMBBInJumpTables(MBB, CurTBB); 1038 if (DidChange) { 1039 ++NumBranchOpts; 1040 MadeChange = true; 1041 if (!HasBranchToSelf) return; 1042 } 1043 } 1044 } 1045 1046 // Add the branch back if the block is more than just an uncond branch. 1047 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1048 } 1049 } 1050 1051 // If the prior block doesn't fall through into this block, and if this 1052 // block doesn't fall through into some other block, see if we can find a 1053 // place to move this block where a fall-through will happen. 1054 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 1055 PriorTBB, PriorFBB, PriorCond)) { 1056 // Now we know that there was no fall-through into this block, check to 1057 // see if it has a fall-through into its successor. 1058 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1059 CurCond); 1060 1061 if (!MBB->isLandingPad()) { 1062 // Check all the predecessors of this block. If one of them has no fall 1063 // throughs, move this block right after it. 1064 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1065 E = MBB->pred_end(); PI != E; ++PI) { 1066 // Analyze the branch at the end of the pred. 1067 MachineBasicBlock *PredBB = *PI; 1068 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1069 if (PredBB != MBB && !CanFallThrough(PredBB) 1070 && (!CurFallsThru || !CurTBB || !CurFBB) 1071 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1072 // If the current block doesn't fall through, just move it. 1073 // If the current block can fall through and does not end with a 1074 // conditional branch, we need to append an unconditional jump to 1075 // the (current) next block. To avoid a possible compile-time 1076 // infinite loop, move blocks only backward in this case. 1077 // Also, if there are already 2 branches here, we cannot add a third; 1078 // this means we have the case 1079 // Bcc next 1080 // B elsewhere 1081 // next: 1082 if (CurFallsThru) { 1083 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1084 CurCond.clear(); 1085 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1086 } 1087 MBB->moveAfter(PredBB); 1088 MadeChange = true; 1089 return OptimizeBlock(MBB); 1090 } 1091 } 1092 } 1093 1094 if (!CurFallsThru) { 1095 // Check all successors to see if we can move this block before it. 1096 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1097 E = MBB->succ_end(); SI != E; ++SI) { 1098 // Analyze the branch at the end of the block before the succ. 1099 MachineBasicBlock *SuccBB = *SI; 1100 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1101 std::vector<MachineOperand> SuccPrevCond; 1102 1103 // If this block doesn't already fall-through to that successor, and if 1104 // the succ doesn't already have a block that can fall through into it, 1105 // and if the successor isn't an EH destination, we can arrange for the 1106 // fallthrough to happen. 1107 if (SuccBB != MBB && !CanFallThrough(SuccPrev) && 1108 !SuccBB->isLandingPad()) { 1109 MBB->moveBefore(SuccBB); 1110 MadeChange = true; 1111 return OptimizeBlock(MBB); 1112 } 1113 } 1114 1115 // Okay, there is no really great place to put this block. If, however, 1116 // the block before this one would be a fall-through if this block were 1117 // removed, move this block to the end of the function. 1118 if (FallThrough != MBB->getParent()->end() && 1119 PrevBB.isSuccessor(FallThrough)) { 1120 MBB->moveAfter(--MBB->getParent()->end()); 1121 MadeChange = true; 1122 return; 1123 } 1124 } 1125 } 1126} 1127