BranchFolding.cpp revision 68d4d1d49c813d10047ad116e897a17d67112c10
1//===-- BranchFolding.cpp - Fold machine code branch instructions ---------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass forwards branches to unconditional branches to make them branch 11// directly to the target block. This pass often results in dead MBB's, which 12// it then removes. 13// 14// Note that this pass must be run after register allocation, it cannot handle 15// SSA form. 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/TargetRegisterInfo.h" 28#include "llvm/Support/CommandLine.h" 29#include "llvm/Support/Debug.h" 30#include "llvm/ADT/SmallSet.h" 31#include "llvm/ADT/Statistic.h" 32#include "llvm/ADT/STLExtras.h" 33#include <algorithm> 34using namespace llvm; 35 36STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 37STATISTIC(NumBranchOpts, "Number of branches optimized"); 38STATISTIC(NumTailMerge , "Number of block tails merged"); 39static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 40 cl::init(cl::BOU_UNSET), cl::Hidden); 41// Throttle for huge numbers of predecessors (compile speed problems) 42static cl::opt<unsigned> 43TailMergeThreshold("tail-merge-threshold", 44 cl::desc("Max number of predecessors to consider tail merging"), 45 cl::init(100), cl::Hidden); 46 47namespace { 48 struct VISIBILITY_HIDDEN BranchFolder : public MachineFunctionPass { 49 static char ID; 50 explicit BranchFolder(bool defaultEnableTailMerge) : 51 MachineFunctionPass((intptr_t)&ID) { 52 switch (FlagEnableTailMerge) { 53 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 54 case cl::BOU_TRUE: EnableTailMerge = true; break; 55 case cl::BOU_FALSE: EnableTailMerge = false; break; 56 } 57 } 58 59 virtual bool runOnMachineFunction(MachineFunction &MF); 60 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 61 const TargetInstrInfo *TII; 62 MachineModuleInfo *MMI; 63 bool MadeChange; 64 private: 65 // Tail Merging. 66 bool EnableTailMerge; 67 bool TailMergeBlocks(MachineFunction &MF); 68 bool TryMergeBlocks(MachineBasicBlock* SuccBB, 69 MachineBasicBlock* PredBB); 70 void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 71 MachineBasicBlock *NewDest); 72 MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB, 73 MachineBasicBlock::iterator BBI1); 74 unsigned ComputeSameTails(unsigned CurHash, unsigned minCommonTailLength); 75 void RemoveBlocksWithHash(unsigned CurHash, MachineBasicBlock* SuccBB, 76 MachineBasicBlock* PredBB); 77 unsigned CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 78 unsigned maxCommonTailLength); 79 80 typedef std::pair<unsigned,MachineBasicBlock*> MergePotentialsElt; 81 typedef std::vector<MergePotentialsElt>::iterator MPIterator; 82 std::vector<MergePotentialsElt> MergePotentials; 83 84 typedef std::pair<MPIterator, MachineBasicBlock::iterator> SameTailElt; 85 std::vector<SameTailElt> SameTails; 86 87 const TargetRegisterInfo *RegInfo; 88 RegScavenger *RS; 89 // Branch optzn. 90 bool OptimizeBranches(MachineFunction &MF); 91 void OptimizeBlock(MachineBasicBlock *MBB); 92 void RemoveDeadBlock(MachineBasicBlock *MBB); 93 bool OptimizeImpDefsBlock(MachineBasicBlock *MBB); 94 95 bool CanFallThrough(MachineBasicBlock *CurBB); 96 bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable, 97 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 98 const std::vector<MachineOperand> &Cond); 99 }; 100 char BranchFolder::ID = 0; 101} 102 103FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 104 return new BranchFolder(DefaultEnableTailMerge); } 105 106/// RemoveDeadBlock - Remove the specified dead machine basic block from the 107/// function, updating the CFG. 108void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 109 assert(MBB->pred_empty() && "MBB must be dead!"); 110 DOUT << "\nRemoving MBB: " << *MBB; 111 112 MachineFunction *MF = MBB->getParent(); 113 // drop all successors. 114 while (!MBB->succ_empty()) 115 MBB->removeSuccessor(MBB->succ_end()-1); 116 117 // If there are any labels in the basic block, unregister them from 118 // MachineModuleInfo. 119 if (MMI && !MBB->empty()) { 120 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 121 I != E; ++I) { 122 if (I->isLabel()) 123 // The label ID # is always operand #0, an immediate. 124 MMI->InvalidateLabel(I->getOperand(0).getImm()); 125 } 126 } 127 128 // Remove the block. 129 MF->erase(MBB); 130} 131 132/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 133/// followed by terminators, and if the implicitly defined registers are not 134/// used by the terminators, remove those implicit_def's. e.g. 135/// BB1: 136/// r0 = implicit_def 137/// r1 = implicit_def 138/// br 139/// This block can be optimized away later if the implicit instructions are 140/// removed. 141bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 142 SmallSet<unsigned, 4> ImpDefRegs; 143 MachineBasicBlock::iterator I = MBB->begin(); 144 while (I != MBB->end()) { 145 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) 146 break; 147 unsigned Reg = I->getOperand(0).getReg(); 148 ImpDefRegs.insert(Reg); 149 for (const unsigned *SubRegs = RegInfo->getSubRegisters(Reg); 150 unsigned SubReg = *SubRegs; ++SubRegs) 151 ImpDefRegs.insert(SubReg); 152 ++I; 153 } 154 if (ImpDefRegs.empty()) 155 return false; 156 157 MachineBasicBlock::iterator FirstTerm = I; 158 while (I != MBB->end()) { 159 if (!TII->isUnpredicatedTerminator(I)) 160 return false; 161 // See if it uses any of the implicitly defined registers. 162 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 163 MachineOperand &MO = I->getOperand(i); 164 if (!MO.isReg() || !MO.isUse()) 165 continue; 166 unsigned Reg = MO.getReg(); 167 if (ImpDefRegs.count(Reg)) 168 return false; 169 } 170 ++I; 171 } 172 173 I = MBB->begin(); 174 while (I != FirstTerm) { 175 MachineInstr *ImpDefMI = &*I; 176 ++I; 177 MBB->erase(ImpDefMI); 178 } 179 180 return true; 181} 182 183bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { 184 TII = MF.getTarget().getInstrInfo(); 185 if (!TII) return false; 186 187 RegInfo = MF.getTarget().getRegisterInfo(); 188 189 // Fix CFG. The later algorithms expect it to be right. 190 bool EverMadeChange = false; 191 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 192 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 193 std::vector<MachineOperand> Cond; 194 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond)) 195 EverMadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 196 EverMadeChange |= OptimizeImpDefsBlock(MBB); 197 } 198 199 RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 200 201 MMI = getAnalysisToUpdate<MachineModuleInfo>(); 202 203 bool MadeChangeThisIteration = true; 204 while (MadeChangeThisIteration) { 205 MadeChangeThisIteration = false; 206 MadeChangeThisIteration |= TailMergeBlocks(MF); 207 MadeChangeThisIteration |= OptimizeBranches(MF); 208 EverMadeChange |= MadeChangeThisIteration; 209 } 210 211 // See if any jump tables have become mergable or dead as the code generator 212 // did its thing. 213 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 214 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 215 if (!JTs.empty()) { 216 // Figure out how these jump tables should be merged. 217 std::vector<unsigned> JTMapping; 218 JTMapping.reserve(JTs.size()); 219 220 // We always keep the 0th jump table. 221 JTMapping.push_back(0); 222 223 // Scan the jump tables, seeing if there are any duplicates. Note that this 224 // is N^2, which should be fixed someday. 225 for (unsigned i = 1, e = JTs.size(); i != e; ++i) 226 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 227 228 // If a jump table was merge with another one, walk the function rewriting 229 // references to jump tables to reference the new JT ID's. Keep track of 230 // whether we see a jump table idx, if not, we can delete the JT. 231 BitVector JTIsLive(JTs.size()); 232 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 233 BB != E; ++BB) { 234 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 235 I != E; ++I) 236 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 237 MachineOperand &Op = I->getOperand(op); 238 if (!Op.isJumpTableIndex()) continue; 239 unsigned NewIdx = JTMapping[Op.getIndex()]; 240 Op.setIndex(NewIdx); 241 242 // Remember that this JT is live. 243 JTIsLive.set(NewIdx); 244 } 245 } 246 247 // Finally, remove dead jump tables. This happens either because the 248 // indirect jump was unreachable (and thus deleted) or because the jump 249 // table was merged with some other one. 250 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 251 if (!JTIsLive.test(i)) { 252 JTI->RemoveJumpTable(i); 253 EverMadeChange = true; 254 } 255 } 256 257 delete RS; 258 return EverMadeChange; 259} 260 261//===----------------------------------------------------------------------===// 262// Tail Merging of Blocks 263//===----------------------------------------------------------------------===// 264 265/// HashMachineInstr - Compute a hash value for MI and its operands. 266static unsigned HashMachineInstr(const MachineInstr *MI) { 267 unsigned Hash = MI->getOpcode(); 268 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 269 const MachineOperand &Op = MI->getOperand(i); 270 271 // Merge in bits from the operand if easy. 272 unsigned OperandHash = 0; 273 switch (Op.getType()) { 274 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 275 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 276 case MachineOperand::MO_MachineBasicBlock: 277 OperandHash = Op.getMBB()->getNumber(); 278 break; 279 case MachineOperand::MO_FrameIndex: 280 case MachineOperand::MO_ConstantPoolIndex: 281 case MachineOperand::MO_JumpTableIndex: 282 OperandHash = Op.getIndex(); 283 break; 284 case MachineOperand::MO_GlobalAddress: 285 case MachineOperand::MO_ExternalSymbol: 286 // Global address / external symbol are too hard, don't bother, but do 287 // pull in the offset. 288 OperandHash = Op.getOffset(); 289 break; 290 default: break; 291 } 292 293 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 294 } 295 return Hash; 296} 297 298/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 299/// with no successors, we hash two instructions, because cross-jumping 300/// only saves code when at least two instructions are removed (since a 301/// branch must be inserted). For blocks with a successor, one of the 302/// two blocks to be tail-merged will end with a branch already, so 303/// it gains to cross-jump even for one instruction. 304 305static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 306 unsigned minCommonTailLength) { 307 MachineBasicBlock::const_iterator I = MBB->end(); 308 if (I == MBB->begin()) 309 return 0; // Empty MBB. 310 311 --I; 312 unsigned Hash = HashMachineInstr(I); 313 314 if (I == MBB->begin() || minCommonTailLength == 1) 315 return Hash; // Single instr MBB. 316 317 --I; 318 // Hash in the second-to-last instruction. 319 Hash ^= HashMachineInstr(I) << 2; 320 return Hash; 321} 322 323/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 324/// of instructions they actually have in common together at their end. Return 325/// iterators for the first shared instruction in each block. 326static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 327 MachineBasicBlock *MBB2, 328 MachineBasicBlock::iterator &I1, 329 MachineBasicBlock::iterator &I2) { 330 I1 = MBB1->end(); 331 I2 = MBB2->end(); 332 333 unsigned TailLen = 0; 334 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 335 --I1; --I2; 336 if (!I1->isIdenticalTo(I2) || 337 // FIXME: This check is dubious. It's used to get around a problem where 338 // people incorrectly expect inline asm directives to remain in the same 339 // relative order. This is untenable because normal compiler 340 // optimizations (like this one) may reorder and/or merge these 341 // directives. 342 I1->getOpcode() == TargetInstrInfo::INLINEASM) { 343 ++I1; ++I2; 344 break; 345 } 346 ++TailLen; 347 } 348 return TailLen; 349} 350 351/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 352/// after it, replacing it with an unconditional branch to NewDest. This 353/// returns true if OldInst's block is modified, false if NewDest is modified. 354void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 355 MachineBasicBlock *NewDest) { 356 MachineBasicBlock *OldBB = OldInst->getParent(); 357 358 // Remove all the old successors of OldBB from the CFG. 359 while (!OldBB->succ_empty()) 360 OldBB->removeSuccessor(OldBB->succ_begin()); 361 362 // Remove all the dead instructions from the end of OldBB. 363 OldBB->erase(OldInst, OldBB->end()); 364 365 // If OldBB isn't immediately before OldBB, insert a branch to it. 366 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 367 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>()); 368 OldBB->addSuccessor(NewDest); 369 ++NumTailMerge; 370} 371 372/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 373/// MBB so that the part before the iterator falls into the part starting at the 374/// iterator. This returns the new MBB. 375MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 376 MachineBasicBlock::iterator BBI1) { 377 MachineFunction &MF = *CurMBB.getParent(); 378 379 // Create the fall-through block. 380 MachineFunction::iterator MBBI = &CurMBB; 381 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 382 CurMBB.getParent()->insert(++MBBI, NewMBB); 383 384 // Move all the successors of this block to the specified block. 385 NewMBB->transferSuccessors(&CurMBB); 386 387 // Add an edge from CurMBB to NewMBB for the fall-through. 388 CurMBB.addSuccessor(NewMBB); 389 390 // Splice the code over. 391 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 392 393 // For targets that use the register scavenger, we must maintain LiveIns. 394 if (RS) { 395 RS->enterBasicBlock(&CurMBB); 396 if (!CurMBB.empty()) 397 RS->forward(prior(CurMBB.end())); 398 BitVector RegsLiveAtExit(RegInfo->getNumRegs()); 399 RS->getRegsUsed(RegsLiveAtExit, false); 400 for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++) 401 if (RegsLiveAtExit[i]) 402 NewMBB->addLiveIn(i); 403 } 404 405 return NewMBB; 406} 407 408/// EstimateRuntime - Make a rough estimate for how long it will take to run 409/// the specified code. 410static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 411 MachineBasicBlock::iterator E) { 412 unsigned Time = 0; 413 for (; I != E; ++I) { 414 const TargetInstrDesc &TID = I->getDesc(); 415 if (TID.isCall()) 416 Time += 10; 417 else if (TID.isSimpleLoad() || TID.mayStore()) 418 Time += 2; 419 else 420 ++Time; 421 } 422 return Time; 423} 424 425// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 426// branches temporarily for tail merging). In the case where CurMBB ends 427// with a conditional branch to the next block, optimize by reversing the 428// test and conditionally branching to SuccMBB instead. 429 430static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, 431 const TargetInstrInfo *TII) { 432 MachineFunction *MF = CurMBB->getParent(); 433 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 434 MachineBasicBlock *TBB = 0, *FBB = 0; 435 std::vector<MachineOperand> Cond; 436 if (I != MF->end() && 437 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) { 438 MachineBasicBlock *NextBB = I; 439 if (TBB == NextBB && !Cond.empty() && !FBB) { 440 if (!TII->ReverseBranchCondition(Cond)) { 441 TII->RemoveBranch(*CurMBB); 442 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 443 return; 444 } 445 } 446 } 447 TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector<MachineOperand>()); 448} 449 450static bool MergeCompare(const std::pair<unsigned,MachineBasicBlock*> &p, 451 const std::pair<unsigned,MachineBasicBlock*> &q) { 452 if (p.first < q.first) 453 return true; 454 else if (p.first > q.first) 455 return false; 456 else if (p.second->getNumber() < q.second->getNumber()) 457 return true; 458 else if (p.second->getNumber() > q.second->getNumber()) 459 return false; 460 else { 461 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 462 // an object with itself. 463#ifndef _GLIBCXX_DEBUG 464 assert(0 && "Predecessor appears twice"); 465#endif 466 return(false); 467 } 468} 469 470/// ComputeSameTails - Look through all the blocks in MergePotentials that have 471/// hash CurHash (guaranteed to match the last element). Build the vector 472/// SameTails of all those that have the (same) largest number of instructions 473/// in common of any pair of these blocks. SameTails entries contain an 474/// iterator into MergePotentials (from which the MachineBasicBlock can be 475/// found) and a MachineBasicBlock::iterator into that MBB indicating the 476/// instruction where the matching code sequence begins. 477/// Order of elements in SameTails is the reverse of the order in which 478/// those blocks appear in MergePotentials (where they are not necessarily 479/// consecutive). 480unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 481 unsigned minCommonTailLength) { 482 unsigned maxCommonTailLength = 0U; 483 SameTails.clear(); 484 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 485 MPIterator HighestMPIter = prior(MergePotentials.end()); 486 for (MPIterator CurMPIter = prior(MergePotentials.end()), 487 B = MergePotentials.begin(); 488 CurMPIter!=B && CurMPIter->first==CurHash; 489 --CurMPIter) { 490 for (MPIterator I = prior(CurMPIter); I->first==CurHash ; --I) { 491 unsigned CommonTailLen = ComputeCommonTailLength( 492 CurMPIter->second, 493 I->second, 494 TrialBBI1, TrialBBI2); 495 // If we will have to split a block, there should be at least 496 // minCommonTailLength instructions in common; if not, at worst 497 // we will be replacing a fallthrough into the common tail with a 498 // branch, which at worst breaks even with falling through into 499 // the duplicated common tail, so 1 instruction in common is enough. 500 // We will always pick a block we do not have to split as the common 501 // tail if there is one. 502 // (Empty blocks will get forwarded and need not be considered.) 503 if (CommonTailLen >= minCommonTailLength || 504 (CommonTailLen > 0 && 505 (TrialBBI1==CurMPIter->second->begin() || 506 TrialBBI2==I->second->begin()))) { 507 if (CommonTailLen > maxCommonTailLength) { 508 SameTails.clear(); 509 maxCommonTailLength = CommonTailLen; 510 HighestMPIter = CurMPIter; 511 SameTails.push_back(std::make_pair(CurMPIter, TrialBBI1)); 512 } 513 if (HighestMPIter == CurMPIter && 514 CommonTailLen == maxCommonTailLength) 515 SameTails.push_back(std::make_pair(I, TrialBBI2)); 516 } 517 if (I==B) 518 break; 519 } 520 } 521 return maxCommonTailLength; 522} 523 524/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 525/// MergePotentials, restoring branches at ends of blocks as appropriate. 526void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 527 MachineBasicBlock* SuccBB, 528 MachineBasicBlock* PredBB) { 529 MPIterator CurMPIter, B; 530 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 531 CurMPIter->first==CurHash; 532 --CurMPIter) { 533 // Put the unconditional branch back, if we need one. 534 MachineBasicBlock *CurMBB = CurMPIter->second; 535 if (SuccBB && CurMBB != PredBB) 536 FixTail(CurMBB, SuccBB, TII); 537 if (CurMPIter==B) 538 break; 539 } 540 if (CurMPIter->first!=CurHash) 541 CurMPIter++; 542 MergePotentials.erase(CurMPIter, MergePotentials.end()); 543} 544 545/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 546/// only of the common tail. Create a block that does by splitting one. 547unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 548 unsigned maxCommonTailLength) { 549 unsigned i, commonTailIndex; 550 unsigned TimeEstimate = ~0U; 551 for (i=0, commonTailIndex=0; i<SameTails.size(); i++) { 552 // Use PredBB if possible; that doesn't require a new branch. 553 if (SameTails[i].first->second==PredBB) { 554 commonTailIndex = i; 555 break; 556 } 557 // Otherwise, make a (fairly bogus) choice based on estimate of 558 // how long it will take the various blocks to execute. 559 unsigned t = EstimateRuntime(SameTails[i].first->second->begin(), 560 SameTails[i].second); 561 if (t<=TimeEstimate) { 562 TimeEstimate = t; 563 commonTailIndex = i; 564 } 565 } 566 567 MachineBasicBlock::iterator BBI = SameTails[commonTailIndex].second; 568 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; 569 570 DOUT << "\nSplitting " << MBB->getNumber() << ", size " << 571 maxCommonTailLength; 572 573 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 574 SameTails[commonTailIndex].first->second = newMBB; 575 SameTails[commonTailIndex].second = newMBB->begin(); 576 // If we split PredBB, newMBB is the new predecessor. 577 if (PredBB==MBB) 578 PredBB = newMBB; 579 580 return commonTailIndex; 581} 582 583// See if any of the blocks in MergePotentials (which all have a common single 584// successor, or all have no successor) can be tail-merged. If there is a 585// successor, any blocks in MergePotentials that are not tail-merged and 586// are not immediately before Succ must have an unconditional branch to 587// Succ added (but the predecessor/successor lists need no adjustment). 588// The lone predecessor of Succ that falls through into Succ, 589// if any, is given in PredBB. 590 591bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, 592 MachineBasicBlock* PredBB) { 593 // It doesn't make sense to save a single instruction since tail merging 594 // will add a jump. 595 // FIXME: Ask the target to provide the threshold? 596 unsigned minCommonTailLength = (SuccBB ? 1 : 2) + 1; 597 MadeChange = false; 598 599 DOUT << "\nTryMergeBlocks " << MergePotentials.size(); 600 601 // Sort by hash value so that blocks with identical end sequences sort 602 // together. 603 std::stable_sort(MergePotentials.begin(), MergePotentials.end(),MergeCompare); 604 605 // Walk through equivalence sets looking for actual exact matches. 606 while (MergePotentials.size() > 1) { 607 unsigned CurHash = prior(MergePotentials.end())->first; 608 609 // Build SameTails, identifying the set of blocks with this hash code 610 // and with the maximum number of instructions in common. 611 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 612 minCommonTailLength); 613 614 // If we didn't find any pair that has at least minCommonTailLength 615 // instructions in common, remove all blocks with this hash code and retry. 616 if (SameTails.empty()) { 617 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 618 continue; 619 } 620 621 // If one of the blocks is the entire common tail (and not the entry 622 // block, which we can't jump to), we can treat all blocks with this same 623 // tail at once. Use PredBB if that is one of the possibilities, as that 624 // will not introduce any extra branches. 625 MachineBasicBlock *EntryBB = MergePotentials.begin()->second-> 626 getParent()->begin(); 627 unsigned int commonTailIndex, i; 628 for (commonTailIndex=SameTails.size(), i=0; i<SameTails.size(); i++) { 629 MachineBasicBlock *MBB = SameTails[i].first->second; 630 if (MBB->begin() == SameTails[i].second && MBB != EntryBB) { 631 commonTailIndex = i; 632 if (MBB==PredBB) 633 break; 634 } 635 } 636 637 if (commonTailIndex==SameTails.size()) { 638 // None of the blocks consist entirely of the common tail. 639 // Split a block so that one does. 640 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); 641 } 642 643 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; 644 // MBB is common tail. Adjust all other BB's to jump to this one. 645 // Traversal must be forwards so erases work. 646 DOUT << "\nUsing common tail " << MBB->getNumber() << " for "; 647 for (unsigned int i=0; i<SameTails.size(); ++i) { 648 if (commonTailIndex==i) 649 continue; 650 DOUT << SameTails[i].first->second->getNumber() << ","; 651 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 652 ReplaceTailWithBranchTo(SameTails[i].second, MBB); 653 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 654 MergePotentials.erase(SameTails[i].first); 655 } 656 DOUT << "\n"; 657 // We leave commonTailIndex in the worklist in case there are other blocks 658 // that match it with a smaller number of instructions. 659 MadeChange = true; 660 } 661 return MadeChange; 662} 663 664bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 665 666 if (!EnableTailMerge) return false; 667 668 MadeChange = false; 669 670 // First find blocks with no successors. 671 MergePotentials.clear(); 672 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 673 if (I->succ_empty()) 674 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I)); 675 } 676 // See if we can do any tail merging on those. 677 if (MergePotentials.size() < TailMergeThreshold && 678 MergePotentials.size() >= 2) 679 MadeChange |= TryMergeBlocks(NULL, NULL); 680 681 // Look at blocks (IBB) with multiple predecessors (PBB). 682 // We change each predecessor to a canonical form, by 683 // (1) temporarily removing any unconditional branch from the predecessor 684 // to IBB, and 685 // (2) alter conditional branches so they branch to the other block 686 // not IBB; this may require adding back an unconditional branch to IBB 687 // later, where there wasn't one coming in. E.g. 688 // Bcc IBB 689 // fallthrough to QBB 690 // here becomes 691 // Bncc QBB 692 // with a conceptual B to IBB after that, which never actually exists. 693 // With those changes, we see whether the predecessors' tails match, 694 // and merge them if so. We change things out of canonical form and 695 // back to the way they were later in the process. (OptimizeBranches 696 // would undo some of this, but we can't use it, because we'd get into 697 // a compile-time infinite loop repeatedly doing and undoing the same 698 // transformations.) 699 700 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 701 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 702 MachineBasicBlock *IBB = I; 703 MachineBasicBlock *PredBB = prior(I); 704 MergePotentials.clear(); 705 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 706 E2 = I->pred_end(); 707 P != E2; ++P) { 708 MachineBasicBlock* PBB = *P; 709 // Skip blocks that loop to themselves, can't tail merge these. 710 if (PBB==IBB) 711 continue; 712 MachineBasicBlock *TBB = 0, *FBB = 0; 713 std::vector<MachineOperand> Cond; 714 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond)) { 715 // Failing case: IBB is the target of a cbr, and 716 // we cannot reverse the branch. 717 std::vector<MachineOperand> NewCond(Cond); 718 if (!Cond.empty() && TBB==IBB) { 719 if (TII->ReverseBranchCondition(NewCond)) 720 continue; 721 // This is the QBB case described above 722 if (!FBB) 723 FBB = next(MachineFunction::iterator(PBB)); 724 } 725 // Failing case: the only way IBB can be reached from PBB is via 726 // exception handling. Happens for landing pads. Would be nice 727 // to have a bit in the edge so we didn't have to do all this. 728 if (IBB->isLandingPad()) { 729 MachineFunction::iterator IP = PBB; IP++; 730 MachineBasicBlock* PredNextBB = NULL; 731 if (IP!=MF.end()) 732 PredNextBB = IP; 733 if (TBB==NULL) { 734 if (IBB!=PredNextBB) // fallthrough 735 continue; 736 } else if (FBB) { 737 if (TBB!=IBB && FBB!=IBB) // cbr then ubr 738 continue; 739 } else if (Cond.empty()) { 740 if (TBB!=IBB) // ubr 741 continue; 742 } else { 743 if (TBB!=IBB && IBB!=PredNextBB) // cbr 744 continue; 745 } 746 } 747 // Remove the unconditional branch at the end, if any. 748 if (TBB && (Cond.empty() || FBB)) { 749 TII->RemoveBranch(*PBB); 750 if (!Cond.empty()) 751 // reinsert conditional branch only, for now 752 TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond); 753 } 754 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P)); 755 } 756 } 757 if (MergePotentials.size() >= 2) 758 MadeChange |= TryMergeBlocks(I, PredBB); 759 // Reinsert an unconditional branch if needed. 760 // The 1 below can occur as a result of removing blocks in TryMergeBlocks. 761 PredBB = prior(I); // this may have been changed in TryMergeBlocks 762 if (MergePotentials.size()==1 && 763 MergePotentials.begin()->second != PredBB) 764 FixTail(MergePotentials.begin()->second, I, TII); 765 } 766 } 767 return MadeChange; 768} 769 770//===----------------------------------------------------------------------===// 771// Branch Optimization 772//===----------------------------------------------------------------------===// 773 774bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 775 MadeChange = false; 776 777 // Make sure blocks are numbered in order 778 MF.RenumberBlocks(); 779 780 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 781 MachineBasicBlock *MBB = I++; 782 OptimizeBlock(MBB); 783 784 // If it is dead, remove it. 785 if (MBB->pred_empty()) { 786 RemoveDeadBlock(MBB); 787 MadeChange = true; 788 ++NumDeadBlocks; 789 } 790 } 791 return MadeChange; 792} 793 794 795/// CanFallThrough - Return true if the specified block (with the specified 796/// branch condition) can implicitly transfer control to the block after it by 797/// falling off the end of it. This should return false if it can reach the 798/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 799/// 800/// True is a conservative answer. 801/// 802bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 803 bool BranchUnAnalyzable, 804 MachineBasicBlock *TBB, 805 MachineBasicBlock *FBB, 806 const std::vector<MachineOperand> &Cond) { 807 MachineFunction::iterator Fallthrough = CurBB; 808 ++Fallthrough; 809 // If FallthroughBlock is off the end of the function, it can't fall through. 810 if (Fallthrough == CurBB->getParent()->end()) 811 return false; 812 813 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 814 if (!CurBB->isSuccessor(Fallthrough)) 815 return false; 816 817 // If we couldn't analyze the branch, assume it could fall through. 818 if (BranchUnAnalyzable) return true; 819 820 // If there is no branch, control always falls through. 821 if (TBB == 0) return true; 822 823 // If there is some explicit branch to the fallthrough block, it can obviously 824 // reach, even though the branch should get folded to fall through implicitly. 825 if (MachineFunction::iterator(TBB) == Fallthrough || 826 MachineFunction::iterator(FBB) == Fallthrough) 827 return true; 828 829 // If it's an unconditional branch to some block not the fall through, it 830 // doesn't fall through. 831 if (Cond.empty()) return false; 832 833 // Otherwise, if it is conditional and has no explicit false block, it falls 834 // through. 835 return FBB == 0; 836} 837 838/// CanFallThrough - Return true if the specified can implicitly transfer 839/// control to the block after it by falling off the end of it. This should 840/// return false if it can reach the block after it, but it uses an explicit 841/// branch to do so (e.g. a table jump). 842/// 843/// True is a conservative answer. 844/// 845bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 846 MachineBasicBlock *TBB = 0, *FBB = 0; 847 std::vector<MachineOperand> Cond; 848 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond); 849 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 850} 851 852/// IsBetterFallthrough - Return true if it would be clearly better to 853/// fall-through to MBB1 than to fall through into MBB2. This has to return 854/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 855/// result in infinite loops. 856static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 857 MachineBasicBlock *MBB2) { 858 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 859 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 860 // optimize branches that branch to either a return block or an assert block 861 // into a fallthrough to the return. 862 if (MBB1->empty() || MBB2->empty()) return false; 863 864 // If there is a clear successor ordering we make sure that one block 865 // will fall through to the next 866 if (MBB1->isSuccessor(MBB2)) return true; 867 if (MBB2->isSuccessor(MBB1)) return false; 868 869 MachineInstr *MBB1I = --MBB1->end(); 870 MachineInstr *MBB2I = --MBB2->end(); 871 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 872} 873 874/// OptimizeBlock - Analyze and optimize control flow related to the specified 875/// block. This is never called on the entry block. 876void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 877 MachineFunction::iterator FallThrough = MBB; 878 ++FallThrough; 879 880 // If this block is empty, make everyone use its fall-through, not the block 881 // explicitly. Landing pads should not do this since the landing-pad table 882 // points to this block. 883 if (MBB->empty() && !MBB->isLandingPad()) { 884 // Dead block? Leave for cleanup later. 885 if (MBB->pred_empty()) return; 886 887 if (FallThrough == MBB->getParent()->end()) { 888 // TODO: Simplify preds to not branch here if possible! 889 } else { 890 // Rewrite all predecessors of the old block to go to the fallthrough 891 // instead. 892 while (!MBB->pred_empty()) { 893 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 894 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 895 } 896 897 // If MBB was the target of a jump table, update jump tables to go to the 898 // fallthrough instead. 899 MBB->getParent()->getJumpTableInfo()-> 900 ReplaceMBBInJumpTables(MBB, FallThrough); 901 MadeChange = true; 902 } 903 return; 904 } 905 906 // Check to see if we can simplify the terminator of the block before this 907 // one. 908 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 909 910 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 911 std::vector<MachineOperand> PriorCond; 912 bool PriorUnAnalyzable = 913 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 914 if (!PriorUnAnalyzable) { 915 // If the CFG for the prior block has extra edges, remove them. 916 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 917 !PriorCond.empty()); 918 919 // If the previous branch is conditional and both conditions go to the same 920 // destination, remove the branch, replacing it with an unconditional one or 921 // a fall-through. 922 if (PriorTBB && PriorTBB == PriorFBB) { 923 TII->RemoveBranch(PrevBB); 924 PriorCond.clear(); 925 if (PriorTBB != MBB) 926 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 927 MadeChange = true; 928 ++NumBranchOpts; 929 return OptimizeBlock(MBB); 930 } 931 932 // If the previous branch *only* branches to *this* block (conditional or 933 // not) remove the branch. 934 if (PriorTBB == MBB && PriorFBB == 0) { 935 TII->RemoveBranch(PrevBB); 936 MadeChange = true; 937 ++NumBranchOpts; 938 return OptimizeBlock(MBB); 939 } 940 941 // If the prior block branches somewhere else on the condition and here if 942 // the condition is false, remove the uncond second branch. 943 if (PriorFBB == MBB) { 944 TII->RemoveBranch(PrevBB); 945 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 946 MadeChange = true; 947 ++NumBranchOpts; 948 return OptimizeBlock(MBB); 949 } 950 951 // If the prior block branches here on true and somewhere else on false, and 952 // if the branch condition is reversible, reverse the branch to create a 953 // fall-through. 954 if (PriorTBB == MBB) { 955 std::vector<MachineOperand> NewPriorCond(PriorCond); 956 if (!TII->ReverseBranchCondition(NewPriorCond)) { 957 TII->RemoveBranch(PrevBB); 958 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 959 MadeChange = true; 960 ++NumBranchOpts; 961 return OptimizeBlock(MBB); 962 } 963 } 964 965 // If this block doesn't fall through (e.g. it ends with an uncond branch or 966 // has no successors) and if the pred falls through into this block, and if 967 // it would otherwise fall through into the block after this, move this 968 // block to the end of the function. 969 // 970 // We consider it more likely that execution will stay in the function (e.g. 971 // due to loops) than it is to exit it. This asserts in loops etc, moving 972 // the assert condition out of the loop body. 973 if (!PriorCond.empty() && PriorFBB == 0 && 974 MachineFunction::iterator(PriorTBB) == FallThrough && 975 !CanFallThrough(MBB)) { 976 bool DoTransform = true; 977 978 // We have to be careful that the succs of PredBB aren't both no-successor 979 // blocks. If neither have successors and if PredBB is the second from 980 // last block in the function, we'd just keep swapping the two blocks for 981 // last. Only do the swap if one is clearly better to fall through than 982 // the other. 983 if (FallThrough == --MBB->getParent()->end() && 984 !IsBetterFallthrough(PriorTBB, MBB)) 985 DoTransform = false; 986 987 // We don't want to do this transformation if we have control flow like: 988 // br cond BB2 989 // BB1: 990 // .. 991 // jmp BBX 992 // BB2: 993 // .. 994 // ret 995 // 996 // In this case, we could actually be moving the return block *into* a 997 // loop! 998 if (DoTransform && !MBB->succ_empty() && 999 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 1000 DoTransform = false; 1001 1002 1003 if (DoTransform) { 1004 // Reverse the branch so we will fall through on the previous true cond. 1005 std::vector<MachineOperand> NewPriorCond(PriorCond); 1006 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1007 DOUT << "\nMoving MBB: " << *MBB; 1008 DOUT << "To make fallthrough to: " << *PriorTBB << "\n"; 1009 1010 TII->RemoveBranch(PrevBB); 1011 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 1012 1013 // Move this block to the end of the function. 1014 MBB->moveAfter(--MBB->getParent()->end()); 1015 MadeChange = true; 1016 ++NumBranchOpts; 1017 return; 1018 } 1019 } 1020 } 1021 } 1022 1023 // Analyze the branch in the current block. 1024 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1025 std::vector<MachineOperand> CurCond; 1026 bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond); 1027 if (!CurUnAnalyzable) { 1028 // If the CFG for the prior block has extra edges, remove them. 1029 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1030 1031 // If this is a two-way branch, and the FBB branches to this block, reverse 1032 // the condition so the single-basic-block loop is faster. Instead of: 1033 // Loop: xxx; jcc Out; jmp Loop 1034 // we want: 1035 // Loop: xxx; jncc Loop; jmp Out 1036 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1037 std::vector<MachineOperand> NewCond(CurCond); 1038 if (!TII->ReverseBranchCondition(NewCond)) { 1039 TII->RemoveBranch(*MBB); 1040 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 1041 MadeChange = true; 1042 ++NumBranchOpts; 1043 return OptimizeBlock(MBB); 1044 } 1045 } 1046 1047 1048 // If this branch is the only thing in its block, see if we can forward 1049 // other blocks across it. 1050 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1051 MBB->begin()->getDesc().isBranch() && CurTBB != MBB) { 1052 // This block may contain just an unconditional branch. Because there can 1053 // be 'non-branch terminators' in the block, try removing the branch and 1054 // then seeing if the block is empty. 1055 TII->RemoveBranch(*MBB); 1056 1057 // If this block is just an unconditional branch to CurTBB, we can 1058 // usually completely eliminate the block. The only case we cannot 1059 // completely eliminate the block is when the block before this one 1060 // falls through into MBB and we can't understand the prior block's branch 1061 // condition. 1062 if (MBB->empty()) { 1063 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 1064 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1065 !PrevBB.isSuccessor(MBB)) { 1066 // If the prior block falls through into us, turn it into an 1067 // explicit branch to us to make updates simpler. 1068 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1069 PriorTBB != MBB && PriorFBB != MBB) { 1070 if (PriorTBB == 0) { 1071 assert(PriorCond.empty() && PriorFBB == 0 && 1072 "Bad branch analysis"); 1073 PriorTBB = MBB; 1074 } else { 1075 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1076 PriorFBB = MBB; 1077 } 1078 TII->RemoveBranch(PrevBB); 1079 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1080 } 1081 1082 // Iterate through all the predecessors, revectoring each in-turn. 1083 size_t PI = 0; 1084 bool DidChange = false; 1085 bool HasBranchToSelf = false; 1086 while(PI != MBB->pred_size()) { 1087 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1088 if (PMBB == MBB) { 1089 // If this block has an uncond branch to itself, leave it. 1090 ++PI; 1091 HasBranchToSelf = true; 1092 } else { 1093 DidChange = true; 1094 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1095 } 1096 } 1097 1098 // Change any jumptables to go to the new MBB. 1099 MBB->getParent()->getJumpTableInfo()-> 1100 ReplaceMBBInJumpTables(MBB, CurTBB); 1101 if (DidChange) { 1102 ++NumBranchOpts; 1103 MadeChange = true; 1104 if (!HasBranchToSelf) return; 1105 } 1106 } 1107 } 1108 1109 // Add the branch back if the block is more than just an uncond branch. 1110 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1111 } 1112 } 1113 1114 // If the prior block doesn't fall through into this block, and if this 1115 // block doesn't fall through into some other block, see if we can find a 1116 // place to move this block where a fall-through will happen. 1117 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 1118 PriorTBB, PriorFBB, PriorCond)) { 1119 // Now we know that there was no fall-through into this block, check to 1120 // see if it has a fall-through into its successor. 1121 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1122 CurCond); 1123 1124 if (!MBB->isLandingPad()) { 1125 // Check all the predecessors of this block. If one of them has no fall 1126 // throughs, move this block right after it. 1127 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1128 E = MBB->pred_end(); PI != E; ++PI) { 1129 // Analyze the branch at the end of the pred. 1130 MachineBasicBlock *PredBB = *PI; 1131 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1132 if (PredBB != MBB && !CanFallThrough(PredBB) 1133 && (!CurFallsThru || !CurTBB || !CurFBB) 1134 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1135 // If the current block doesn't fall through, just move it. 1136 // If the current block can fall through and does not end with a 1137 // conditional branch, we need to append an unconditional jump to 1138 // the (current) next block. To avoid a possible compile-time 1139 // infinite loop, move blocks only backward in this case. 1140 // Also, if there are already 2 branches here, we cannot add a third; 1141 // this means we have the case 1142 // Bcc next 1143 // B elsewhere 1144 // next: 1145 if (CurFallsThru) { 1146 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1147 CurCond.clear(); 1148 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1149 } 1150 MBB->moveAfter(PredBB); 1151 MadeChange = true; 1152 return OptimizeBlock(MBB); 1153 } 1154 } 1155 } 1156 1157 if (!CurFallsThru) { 1158 // Check all successors to see if we can move this block before it. 1159 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1160 E = MBB->succ_end(); SI != E; ++SI) { 1161 // Analyze the branch at the end of the block before the succ. 1162 MachineBasicBlock *SuccBB = *SI; 1163 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1164 std::vector<MachineOperand> SuccPrevCond; 1165 1166 // If this block doesn't already fall-through to that successor, and if 1167 // the succ doesn't already have a block that can fall through into it, 1168 // and if the successor isn't an EH destination, we can arrange for the 1169 // fallthrough to happen. 1170 if (SuccBB != MBB && !CanFallThrough(SuccPrev) && 1171 !SuccBB->isLandingPad()) { 1172 MBB->moveBefore(SuccBB); 1173 MadeChange = true; 1174 return OptimizeBlock(MBB); 1175 } 1176 } 1177 1178 // Okay, there is no really great place to put this block. If, however, 1179 // the block before this one would be a fall-through if this block were 1180 // removed, move this block to the end of the function. 1181 if (FallThrough != MBB->getParent()->end() && 1182 PrevBB.isSuccessor(FallThrough)) { 1183 MBB->moveAfter(--MBB->getParent()->end()); 1184 MadeChange = true; 1185 return; 1186 } 1187 } 1188 } 1189} 1190