BranchFolding.cpp revision 679860e31bd7b8f043ba1ccdc5990cb9bafd9055
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 is DWARF info to active, check to see if there are any LABEL 118 // records in the basic block. If so, unregister them from MachineModuleInfo. 119 if (MMI && !MBB->empty()) { 120 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 121 I != E; ++I) { 122 if ((unsigned)I->getOpcode() == TargetInstrInfo::LABEL) { 123 // The label ID # is always operand #0, an immediate. 124 MMI->InvalidateLabel(I->getOperand(0).getImm()); 125 } 126 } 127 } 128 129 // Remove the block. 130 MF->getBasicBlockList().erase(MBB); 131} 132 133/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 134/// followed by terminators, and if the implicitly defined registers are not 135/// used by the terminators, remove those implicit_def's. e.g. 136/// BB1: 137/// r0 = implicit_def 138/// r1 = implicit_def 139/// br 140/// This block can be optimized away later if the implicit instructions are 141/// removed. 142bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 143 SmallSet<unsigned, 4> ImpDefRegs; 144 MachineBasicBlock::iterator I = MBB->begin(); 145 while (I != MBB->end()) { 146 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) 147 break; 148 unsigned Reg = I->getOperand(0).getReg(); 149 ImpDefRegs.insert(Reg); 150 for (const unsigned *SubRegs = RegInfo->getSubRegisters(Reg); 151 unsigned SubReg = *SubRegs; ++SubRegs) 152 ImpDefRegs.insert(SubReg); 153 ++I; 154 } 155 if (ImpDefRegs.empty()) 156 return false; 157 158 MachineBasicBlock::iterator FirstTerm = I; 159 while (I != MBB->end()) { 160 if (!TII->isUnpredicatedTerminator(I)) 161 return false; 162 // See if it uses any of the implicitly defined registers. 163 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 164 MachineOperand &MO = I->getOperand(i); 165 if (!MO.isReg() || !MO.isUse()) 166 continue; 167 unsigned Reg = MO.getReg(); 168 if (ImpDefRegs.count(Reg)) 169 return false; 170 } 171 ++I; 172 } 173 174 I = MBB->begin(); 175 while (I != FirstTerm) { 176 MachineInstr *ImpDefMI = &*I; 177 ++I; 178 MBB->erase(ImpDefMI); 179 } 180 181 return true; 182} 183 184bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { 185 TII = MF.getTarget().getInstrInfo(); 186 if (!TII) return false; 187 188 RegInfo = MF.getTarget().getRegisterInfo(); 189 190 // Fix CFG. The later algorithms expect it to be right. 191 bool EverMadeChange = false; 192 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 193 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 194 std::vector<MachineOperand> Cond; 195 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond)) 196 EverMadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 197 EverMadeChange |= OptimizeImpDefsBlock(MBB); 198 } 199 200 RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 201 202 MMI = getAnalysisToUpdate<MachineModuleInfo>(); 203 204 bool MadeChangeThisIteration = true; 205 while (MadeChangeThisIteration) { 206 MadeChangeThisIteration = false; 207 MadeChangeThisIteration |= TailMergeBlocks(MF); 208 MadeChangeThisIteration |= OptimizeBranches(MF); 209 EverMadeChange |= MadeChangeThisIteration; 210 } 211 212 // See if any jump tables have become mergable or dead as the code generator 213 // did its thing. 214 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 215 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 216 if (!JTs.empty()) { 217 // Figure out how these jump tables should be merged. 218 std::vector<unsigned> JTMapping; 219 JTMapping.reserve(JTs.size()); 220 221 // We always keep the 0th jump table. 222 JTMapping.push_back(0); 223 224 // Scan the jump tables, seeing if there are any duplicates. Note that this 225 // is N^2, which should be fixed someday. 226 for (unsigned i = 1, e = JTs.size(); i != e; ++i) 227 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 228 229 // If a jump table was merge with another one, walk the function rewriting 230 // references to jump tables to reference the new JT ID's. Keep track of 231 // whether we see a jump table idx, if not, we can delete the JT. 232 BitVector JTIsLive(JTs.size()); 233 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 234 BB != E; ++BB) { 235 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 236 I != E; ++I) 237 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 238 MachineOperand &Op = I->getOperand(op); 239 if (!Op.isJumpTableIndex()) continue; 240 unsigned NewIdx = JTMapping[Op.getIndex()]; 241 Op.setIndex(NewIdx); 242 243 // Remember that this JT is live. 244 JTIsLive.set(NewIdx); 245 } 246 } 247 248 // Finally, remove dead jump tables. This happens either because the 249 // indirect jump was unreachable (and thus deleted) or because the jump 250 // table was merged with some other one. 251 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 252 if (!JTIsLive.test(i)) { 253 JTI->RemoveJumpTable(i); 254 EverMadeChange = true; 255 } 256 } 257 258 delete RS; 259 return EverMadeChange; 260} 261 262//===----------------------------------------------------------------------===// 263// Tail Merging of Blocks 264//===----------------------------------------------------------------------===// 265 266/// HashMachineInstr - Compute a hash value for MI and its operands. 267static unsigned HashMachineInstr(const MachineInstr *MI) { 268 unsigned Hash = MI->getOpcode(); 269 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 270 const MachineOperand &Op = MI->getOperand(i); 271 272 // Merge in bits from the operand if easy. 273 unsigned OperandHash = 0; 274 switch (Op.getType()) { 275 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 276 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 277 case MachineOperand::MO_MachineBasicBlock: 278 OperandHash = Op.getMBB()->getNumber(); 279 break; 280 case MachineOperand::MO_FrameIndex: 281 case MachineOperand::MO_ConstantPoolIndex: 282 case MachineOperand::MO_JumpTableIndex: 283 OperandHash = Op.getIndex(); 284 break; 285 case MachineOperand::MO_GlobalAddress: 286 case MachineOperand::MO_ExternalSymbol: 287 // Global address / external symbol are too hard, don't bother, but do 288 // pull in the offset. 289 OperandHash = Op.getOffset(); 290 break; 291 default: break; 292 } 293 294 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 295 } 296 return Hash; 297} 298 299/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 300/// with no successors, we hash two instructions, because cross-jumping 301/// only saves code when at least two instructions are removed (since a 302/// branch must be inserted). For blocks with a successor, one of the 303/// two blocks to be tail-merged will end with a branch already, so 304/// it gains to cross-jump even for one instruction. 305 306static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 307 unsigned minCommonTailLength) { 308 MachineBasicBlock::const_iterator I = MBB->end(); 309 if (I == MBB->begin()) 310 return 0; // Empty MBB. 311 312 --I; 313 unsigned Hash = HashMachineInstr(I); 314 315 if (I == MBB->begin() || minCommonTailLength == 1) 316 return Hash; // Single instr MBB. 317 318 --I; 319 // Hash in the second-to-last instruction. 320 Hash ^= HashMachineInstr(I) << 2; 321 return Hash; 322} 323 324/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 325/// of instructions they actually have in common together at their end. Return 326/// iterators for the first shared instruction in each block. 327static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 328 MachineBasicBlock *MBB2, 329 MachineBasicBlock::iterator &I1, 330 MachineBasicBlock::iterator &I2) { 331 I1 = MBB1->end(); 332 I2 = MBB2->end(); 333 334 unsigned TailLen = 0; 335 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 336 --I1; --I2; 337 if (!I1->isIdenticalTo(I2) || 338 // FIXME: This check is dubious. It's used to get around a problem where 339 // people incorrectly expect inline asm directives to remain in the same 340 // relative order. This is untenable because normal compiler 341 // optimizations (like this one) may reorder and/or merge these 342 // directives. 343 I1->getOpcode() == TargetInstrInfo::INLINEASM) { 344 ++I1; ++I2; 345 break; 346 } 347 ++TailLen; 348 } 349 return TailLen; 350} 351 352/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 353/// after it, replacing it with an unconditional branch to NewDest. This 354/// returns true if OldInst's block is modified, false if NewDest is modified. 355void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 356 MachineBasicBlock *NewDest) { 357 MachineBasicBlock *OldBB = OldInst->getParent(); 358 359 // Remove all the old successors of OldBB from the CFG. 360 while (!OldBB->succ_empty()) 361 OldBB->removeSuccessor(OldBB->succ_begin()); 362 363 // Remove all the dead instructions from the end of OldBB. 364 OldBB->erase(OldInst, OldBB->end()); 365 366 // If OldBB isn't immediately before OldBB, insert a branch to it. 367 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 368 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>()); 369 OldBB->addSuccessor(NewDest); 370 ++NumTailMerge; 371} 372 373/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 374/// MBB so that the part before the iterator falls into the part starting at the 375/// iterator. This returns the new MBB. 376MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 377 MachineBasicBlock::iterator BBI1) { 378 // Create the fall-through block. 379 MachineFunction::iterator MBBI = &CurMBB; 380 MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock()); 381 CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB); 382 383 // Move all the successors of this block to the specified block. 384 while (!CurMBB.succ_empty()) { 385 MachineBasicBlock *S = *(CurMBB.succ_end()-1); 386 NewMBB->addSuccessor(S); 387 CurMBB.removeSuccessor(S); 388 } 389 390 // Add an edge from CurMBB to NewMBB for the fall-through. 391 CurMBB.addSuccessor(NewMBB); 392 393 // Splice the code over. 394 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 395 396 // For targets that use the register scavenger, we must maintain LiveIns. 397 if (RS) { 398 RS->enterBasicBlock(&CurMBB); 399 if (!CurMBB.empty()) 400 RS->forward(prior(CurMBB.end())); 401 BitVector RegsLiveAtExit(RegInfo->getNumRegs()); 402 RS->getRegsUsed(RegsLiveAtExit, false); 403 for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++) 404 if (RegsLiveAtExit[i]) 405 NewMBB->addLiveIn(i); 406 } 407 408 return NewMBB; 409} 410 411/// EstimateRuntime - Make a rough estimate for how long it will take to run 412/// the specified code. 413static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 414 MachineBasicBlock::iterator E) { 415 unsigned Time = 0; 416 for (; I != E; ++I) { 417 const TargetInstrDesc &TID = I->getDesc(); 418 if (TID.isCall()) 419 Time += 10; 420 else if (TID.isSimpleLoad() || TID.mayStore()) 421 Time += 2; 422 else 423 ++Time; 424 } 425 return Time; 426} 427 428// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 429// branches temporarily for tail merging). In the case where CurMBB ends 430// with a conditional branch to the next block, optimize by reversing the 431// test and conditionally branching to SuccMBB instead. 432 433static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, 434 const TargetInstrInfo *TII) { 435 MachineFunction *MF = CurMBB->getParent(); 436 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 437 MachineBasicBlock *TBB = 0, *FBB = 0; 438 std::vector<MachineOperand> Cond; 439 if (I != MF->end() && 440 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) { 441 MachineBasicBlock *NextBB = I; 442 if (TBB == NextBB && !Cond.empty() && !FBB) { 443 if (!TII->ReverseBranchCondition(Cond)) { 444 TII->RemoveBranch(*CurMBB); 445 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 446 return; 447 } 448 } 449 } 450 TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector<MachineOperand>()); 451} 452 453static bool MergeCompare(const std::pair<unsigned,MachineBasicBlock*> &p, 454 const std::pair<unsigned,MachineBasicBlock*> &q) { 455 if (p.first < q.first) 456 return true; 457 else if (p.first > q.first) 458 return false; 459 else if (p.second->getNumber() < q.second->getNumber()) 460 return true; 461 else if (p.second->getNumber() > q.second->getNumber()) 462 return false; 463 else { 464 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 465 // an object with itself. 466#ifndef _GLIBCXX_DEBUG 467 assert(0 && "Predecessor appears twice"); 468#endif 469 return(false); 470 } 471} 472 473/// ComputeSameTails - Look through all the blocks in MergePotentials that have 474/// hash CurHash (guaranteed to match the last element). Build the vector 475/// SameTails of all those that have the (same) largest number of instructions 476/// in common of any pair of these blocks. SameTails entries contain an 477/// iterator into MergePotentials (from which the MachineBasicBlock can be 478/// found) and a MachineBasicBlock::iterator into that MBB indicating the 479/// instruction where the matching code sequence begins. 480/// Order of elements in SameTails is the reverse of the order in which 481/// those blocks appear in MergePotentials (where they are not necessarily 482/// consecutive). 483unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 484 unsigned minCommonTailLength) { 485 unsigned maxCommonTailLength = 0U; 486 SameTails.clear(); 487 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 488 MPIterator HighestMPIter = prior(MergePotentials.end()); 489 for (MPIterator CurMPIter = prior(MergePotentials.end()), 490 B = MergePotentials.begin(); 491 CurMPIter!=B && CurMPIter->first==CurHash; 492 --CurMPIter) { 493 for (MPIterator I = prior(CurMPIter); I->first==CurHash ; --I) { 494 unsigned CommonTailLen = ComputeCommonTailLength( 495 CurMPIter->second, 496 I->second, 497 TrialBBI1, TrialBBI2); 498 // If we will have to split a block, there should be at least 499 // minCommonTailLength instructions in common; if not, at worst 500 // we will be replacing a fallthrough into the common tail with a 501 // branch, which at worst breaks even with falling through into 502 // the duplicated common tail, so 1 instruction in common is enough. 503 // We will always pick a block we do not have to split as the common 504 // tail if there is one. 505 // (Empty blocks will get forwarded and need not be considered.) 506 if (CommonTailLen >= minCommonTailLength || 507 (CommonTailLen > 0 && 508 (TrialBBI1==CurMPIter->second->begin() || 509 TrialBBI2==I->second->begin()))) { 510 if (CommonTailLen > maxCommonTailLength) { 511 SameTails.clear(); 512 maxCommonTailLength = CommonTailLen; 513 HighestMPIter = CurMPIter; 514 SameTails.push_back(std::make_pair(CurMPIter, TrialBBI1)); 515 } 516 if (HighestMPIter == CurMPIter && 517 CommonTailLen == maxCommonTailLength) 518 SameTails.push_back(std::make_pair(I, TrialBBI2)); 519 } 520 if (I==B) 521 break; 522 } 523 } 524 return maxCommonTailLength; 525} 526 527/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 528/// MergePotentials, restoring branches at ends of blocks as appropriate. 529void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 530 MachineBasicBlock* SuccBB, 531 MachineBasicBlock* PredBB) { 532 MPIterator CurMPIter, B; 533 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 534 CurMPIter->first==CurHash; 535 --CurMPIter) { 536 // Put the unconditional branch back, if we need one. 537 MachineBasicBlock *CurMBB = CurMPIter->second; 538 if (SuccBB && CurMBB != PredBB) 539 FixTail(CurMBB, SuccBB, TII); 540 if (CurMPIter==B) 541 break; 542 } 543 if (CurMPIter->first!=CurHash) 544 CurMPIter++; 545 MergePotentials.erase(CurMPIter, MergePotentials.end()); 546} 547 548/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 549/// only of the common tail. Create a block that does by splitting one. 550unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 551 unsigned maxCommonTailLength) { 552 unsigned i, commonTailIndex; 553 unsigned TimeEstimate = ~0U; 554 for (i=0, commonTailIndex=0; i<SameTails.size(); i++) { 555 // Use PredBB if possible; that doesn't require a new branch. 556 if (SameTails[i].first->second==PredBB) { 557 commonTailIndex = i; 558 break; 559 } 560 // Otherwise, make a (fairly bogus) choice based on estimate of 561 // how long it will take the various blocks to execute. 562 unsigned t = EstimateRuntime(SameTails[i].first->second->begin(), 563 SameTails[i].second); 564 if (t<=TimeEstimate) { 565 TimeEstimate = t; 566 commonTailIndex = i; 567 } 568 } 569 570 MachineBasicBlock::iterator BBI = SameTails[commonTailIndex].second; 571 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; 572 573 DOUT << "\nSplitting " << MBB->getNumber() << ", size " << 574 maxCommonTailLength; 575 576 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 577 SameTails[commonTailIndex].first->second = newMBB; 578 SameTails[commonTailIndex].second = newMBB->begin(); 579 // If we split PredBB, newMBB is the new predecessor. 580 if (PredBB==MBB) 581 PredBB = newMBB; 582 583 return commonTailIndex; 584} 585 586// See if any of the blocks in MergePotentials (which all have a common single 587// successor, or all have no successor) can be tail-merged. If there is a 588// successor, any blocks in MergePotentials that are not tail-merged and 589// are not immediately before Succ must have an unconditional branch to 590// Succ added (but the predecessor/successor lists need no adjustment). 591// The lone predecessor of Succ that falls through into Succ, 592// if any, is given in PredBB. 593 594bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, 595 MachineBasicBlock* PredBB) { 596 // It doesn't make sense to save a single instruction since tail merging 597 // will add a jump. 598 // FIXME: Ask the target to provide the threshold? 599 unsigned minCommonTailLength = (SuccBB ? 1 : 2) + 1; 600 MadeChange = false; 601 602 DOUT << "\nTryMergeBlocks " << MergePotentials.size(); 603 604 // Sort by hash value so that blocks with identical end sequences sort 605 // together. 606 std::stable_sort(MergePotentials.begin(), MergePotentials.end(),MergeCompare); 607 608 // Walk through equivalence sets looking for actual exact matches. 609 while (MergePotentials.size() > 1) { 610 unsigned CurHash = prior(MergePotentials.end())->first; 611 612 // Build SameTails, identifying the set of blocks with this hash code 613 // and with the maximum number of instructions in common. 614 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 615 minCommonTailLength); 616 617 // If we didn't find any pair that has at least minCommonTailLength 618 // instructions in common, remove all blocks with this hash code and retry. 619 if (SameTails.empty()) { 620 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 621 continue; 622 } 623 624 // If one of the blocks is the entire common tail (and not the entry 625 // block, which we can't jump to), we can treat all blocks with this same 626 // tail at once. Use PredBB if that is one of the possibilities, as that 627 // will not introduce any extra branches. 628 MachineBasicBlock *EntryBB = MergePotentials.begin()->second-> 629 getParent()->begin(); 630 unsigned int commonTailIndex, i; 631 for (commonTailIndex=SameTails.size(), i=0; i<SameTails.size(); i++) { 632 MachineBasicBlock *MBB = SameTails[i].first->second; 633 if (MBB->begin() == SameTails[i].second && MBB != EntryBB) { 634 commonTailIndex = i; 635 if (MBB==PredBB) 636 break; 637 } 638 } 639 640 if (commonTailIndex==SameTails.size()) { 641 // None of the blocks consist entirely of the common tail. 642 // Split a block so that one does. 643 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); 644 } 645 646 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; 647 // MBB is common tail. Adjust all other BB's to jump to this one. 648 // Traversal must be forwards so erases work. 649 DOUT << "\nUsing common tail " << MBB->getNumber() << " for "; 650 for (unsigned int i=0; i<SameTails.size(); ++i) { 651 if (commonTailIndex==i) 652 continue; 653 DOUT << SameTails[i].first->second->getNumber() << ","; 654 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 655 ReplaceTailWithBranchTo(SameTails[i].second, MBB); 656 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 657 MergePotentials.erase(SameTails[i].first); 658 } 659 DOUT << "\n"; 660 // We leave commonTailIndex in the worklist in case there are other blocks 661 // that match it with a smaller number of instructions. 662 MadeChange = true; 663 } 664 return MadeChange; 665} 666 667bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 668 669 if (!EnableTailMerge) return false; 670 671 MadeChange = false; 672 673 // First find blocks with no successors. 674 MergePotentials.clear(); 675 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 676 if (I->succ_empty()) 677 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I)); 678 } 679 // See if we can do any tail merging on those. 680 if (MergePotentials.size() < TailMergeThreshold && 681 MergePotentials.size() >= 2) 682 MadeChange |= TryMergeBlocks(NULL, NULL); 683 684 // Look at blocks (IBB) with multiple predecessors (PBB). 685 // We change each predecessor to a canonical form, by 686 // (1) temporarily removing any unconditional branch from the predecessor 687 // to IBB, and 688 // (2) alter conditional branches so they branch to the other block 689 // not IBB; this may require adding back an unconditional branch to IBB 690 // later, where there wasn't one coming in. E.g. 691 // Bcc IBB 692 // fallthrough to QBB 693 // here becomes 694 // Bncc QBB 695 // with a conceptual B to IBB after that, which never actually exists. 696 // With those changes, we see whether the predecessors' tails match, 697 // and merge them if so. We change things out of canonical form and 698 // back to the way they were later in the process. (OptimizeBranches 699 // would undo some of this, but we can't use it, because we'd get into 700 // a compile-time infinite loop repeatedly doing and undoing the same 701 // transformations.) 702 703 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 704 if (!I->succ_empty() && I->pred_size() >= 2 && 705 I->pred_size() < TailMergeThreshold) { 706 MachineBasicBlock *IBB = I; 707 MachineBasicBlock *PredBB = prior(I); 708 MergePotentials.clear(); 709 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 710 E2 = I->pred_end(); 711 P != E2; ++P) { 712 MachineBasicBlock* PBB = *P; 713 // Skip blocks that loop to themselves, can't tail merge these. 714 if (PBB==IBB) 715 continue; 716 MachineBasicBlock *TBB = 0, *FBB = 0; 717 std::vector<MachineOperand> Cond; 718 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond)) { 719 // Failing case: IBB is the target of a cbr, and 720 // we cannot reverse the branch. 721 std::vector<MachineOperand> NewCond(Cond); 722 if (!Cond.empty() && TBB==IBB) { 723 if (TII->ReverseBranchCondition(NewCond)) 724 continue; 725 // This is the QBB case described above 726 if (!FBB) 727 FBB = next(MachineFunction::iterator(PBB)); 728 } 729 // Failing case: the only way IBB can be reached from PBB is via 730 // exception handling. Happens for landing pads. Would be nice 731 // to have a bit in the edge so we didn't have to do all this. 732 if (IBB->isLandingPad()) { 733 MachineFunction::iterator IP = PBB; IP++; 734 MachineBasicBlock* PredNextBB = NULL; 735 if (IP!=MF.end()) 736 PredNextBB = IP; 737 if (TBB==NULL) { 738 if (IBB!=PredNextBB) // fallthrough 739 continue; 740 } else if (FBB) { 741 if (TBB!=IBB && FBB!=IBB) // cbr then ubr 742 continue; 743 } else if (Cond.empty()) { 744 if (TBB!=IBB) // ubr 745 continue; 746 } else { 747 if (TBB!=IBB && IBB!=PredNextBB) // cbr 748 continue; 749 } 750 } 751 // Remove the unconditional branch at the end, if any. 752 if (TBB && (Cond.empty() || FBB)) { 753 TII->RemoveBranch(*PBB); 754 if (!Cond.empty()) 755 // reinsert conditional branch only, for now 756 TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond); 757 } 758 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P)); 759 } 760 } 761 if (MergePotentials.size() >= 2) 762 MadeChange |= TryMergeBlocks(I, PredBB); 763 // Reinsert an unconditional branch if needed. 764 // The 1 below can occur as a result of removing blocks in TryMergeBlocks. 765 PredBB = prior(I); // this may have been changed in TryMergeBlocks 766 if (MergePotentials.size()==1 && 767 MergePotentials.begin()->second != PredBB) 768 FixTail(MergePotentials.begin()->second, I, TII); 769 } 770 } 771 return MadeChange; 772} 773 774//===----------------------------------------------------------------------===// 775// Branch Optimization 776//===----------------------------------------------------------------------===// 777 778bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 779 MadeChange = false; 780 781 // Make sure blocks are numbered in order 782 MF.RenumberBlocks(); 783 784 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 785 MachineBasicBlock *MBB = I++; 786 OptimizeBlock(MBB); 787 788 // If it is dead, remove it. 789 if (MBB->pred_empty()) { 790 RemoveDeadBlock(MBB); 791 MadeChange = true; 792 ++NumDeadBlocks; 793 } 794 } 795 return MadeChange; 796} 797 798 799/// CanFallThrough - Return true if the specified block (with the specified 800/// branch condition) can implicitly transfer control to the block after it by 801/// falling off the end of it. This should return false if it can reach the 802/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 803/// 804/// True is a conservative answer. 805/// 806bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 807 bool BranchUnAnalyzable, 808 MachineBasicBlock *TBB, 809 MachineBasicBlock *FBB, 810 const std::vector<MachineOperand> &Cond) { 811 MachineFunction::iterator Fallthrough = CurBB; 812 ++Fallthrough; 813 // If FallthroughBlock is off the end of the function, it can't fall through. 814 if (Fallthrough == CurBB->getParent()->end()) 815 return false; 816 817 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 818 if (!CurBB->isSuccessor(Fallthrough)) 819 return false; 820 821 // If we couldn't analyze the branch, assume it could fall through. 822 if (BranchUnAnalyzable) return true; 823 824 // If there is no branch, control always falls through. 825 if (TBB == 0) return true; 826 827 // If there is some explicit branch to the fallthrough block, it can obviously 828 // reach, even though the branch should get folded to fall through implicitly. 829 if (MachineFunction::iterator(TBB) == Fallthrough || 830 MachineFunction::iterator(FBB) == Fallthrough) 831 return true; 832 833 // If it's an unconditional branch to some block not the fall through, it 834 // doesn't fall through. 835 if (Cond.empty()) return false; 836 837 // Otherwise, if it is conditional and has no explicit false block, it falls 838 // through. 839 return FBB == 0; 840} 841 842/// CanFallThrough - Return true if the specified can implicitly transfer 843/// control to the block after it by falling off the end of it. This should 844/// return false if it can reach the block after it, but it uses an explicit 845/// branch to do so (e.g. a table jump). 846/// 847/// True is a conservative answer. 848/// 849bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 850 MachineBasicBlock *TBB = 0, *FBB = 0; 851 std::vector<MachineOperand> Cond; 852 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond); 853 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 854} 855 856/// IsBetterFallthrough - Return true if it would be clearly better to 857/// fall-through to MBB1 than to fall through into MBB2. This has to return 858/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 859/// result in infinite loops. 860static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 861 MachineBasicBlock *MBB2) { 862 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 863 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 864 // optimize branches that branch to either a return block or an assert block 865 // into a fallthrough to the return. 866 if (MBB1->empty() || MBB2->empty()) return false; 867 868 // If there is a clear successor ordering we make sure that one block 869 // will fall through to the next 870 if (MBB1->isSuccessor(MBB2)) return true; 871 if (MBB2->isSuccessor(MBB1)) return false; 872 873 MachineInstr *MBB1I = --MBB1->end(); 874 MachineInstr *MBB2I = --MBB2->end(); 875 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 876} 877 878/// OptimizeBlock - Analyze and optimize control flow related to the specified 879/// block. This is never called on the entry block. 880void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 881 MachineFunction::iterator FallThrough = MBB; 882 ++FallThrough; 883 884 // If this block is empty, make everyone use its fall-through, not the block 885 // explicitly. Landing pads should not do this since the landing-pad table 886 // points to this block. 887 if (MBB->empty() && !MBB->isLandingPad()) { 888 // Dead block? Leave for cleanup later. 889 if (MBB->pred_empty()) return; 890 891 if (FallThrough == MBB->getParent()->end()) { 892 // TODO: Simplify preds to not branch here if possible! 893 } else { 894 // Rewrite all predecessors of the old block to go to the fallthrough 895 // instead. 896 while (!MBB->pred_empty()) { 897 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 898 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 899 } 900 901 // If MBB was the target of a jump table, update jump tables to go to the 902 // fallthrough instead. 903 MBB->getParent()->getJumpTableInfo()-> 904 ReplaceMBBInJumpTables(MBB, FallThrough); 905 MadeChange = true; 906 } 907 return; 908 } 909 910 // Check to see if we can simplify the terminator of the block before this 911 // one. 912 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 913 914 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 915 std::vector<MachineOperand> PriorCond; 916 bool PriorUnAnalyzable = 917 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 918 if (!PriorUnAnalyzable) { 919 // If the CFG for the prior block has extra edges, remove them. 920 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 921 !PriorCond.empty()); 922 923 // If the previous branch is conditional and both conditions go to the same 924 // destination, remove the branch, replacing it with an unconditional one or 925 // a fall-through. 926 if (PriorTBB && PriorTBB == PriorFBB) { 927 TII->RemoveBranch(PrevBB); 928 PriorCond.clear(); 929 if (PriorTBB != MBB) 930 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 931 MadeChange = true; 932 ++NumBranchOpts; 933 return OptimizeBlock(MBB); 934 } 935 936 // If the previous branch *only* branches to *this* block (conditional or 937 // not) remove the branch. 938 if (PriorTBB == MBB && PriorFBB == 0) { 939 TII->RemoveBranch(PrevBB); 940 MadeChange = true; 941 ++NumBranchOpts; 942 return OptimizeBlock(MBB); 943 } 944 945 // If the prior block branches somewhere else on the condition and here if 946 // the condition is false, remove the uncond second branch. 947 if (PriorFBB == MBB) { 948 TII->RemoveBranch(PrevBB); 949 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 950 MadeChange = true; 951 ++NumBranchOpts; 952 return OptimizeBlock(MBB); 953 } 954 955 // If the prior block branches here on true and somewhere else on false, and 956 // if the branch condition is reversible, reverse the branch to create a 957 // fall-through. 958 if (PriorTBB == MBB) { 959 std::vector<MachineOperand> NewPriorCond(PriorCond); 960 if (!TII->ReverseBranchCondition(NewPriorCond)) { 961 TII->RemoveBranch(PrevBB); 962 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 963 MadeChange = true; 964 ++NumBranchOpts; 965 return OptimizeBlock(MBB); 966 } 967 } 968 969 // If this block doesn't fall through (e.g. it ends with an uncond branch or 970 // has no successors) and if the pred falls through into this block, and if 971 // it would otherwise fall through into the block after this, move this 972 // block to the end of the function. 973 // 974 // We consider it more likely that execution will stay in the function (e.g. 975 // due to loops) than it is to exit it. This asserts in loops etc, moving 976 // the assert condition out of the loop body. 977 if (!PriorCond.empty() && PriorFBB == 0 && 978 MachineFunction::iterator(PriorTBB) == FallThrough && 979 !CanFallThrough(MBB)) { 980 bool DoTransform = true; 981 982 // We have to be careful that the succs of PredBB aren't both no-successor 983 // blocks. If neither have successors and if PredBB is the second from 984 // last block in the function, we'd just keep swapping the two blocks for 985 // last. Only do the swap if one is clearly better to fall through than 986 // the other. 987 if (FallThrough == --MBB->getParent()->end() && 988 !IsBetterFallthrough(PriorTBB, MBB)) 989 DoTransform = false; 990 991 // We don't want to do this transformation if we have control flow like: 992 // br cond BB2 993 // BB1: 994 // .. 995 // jmp BBX 996 // BB2: 997 // .. 998 // ret 999 // 1000 // In this case, we could actually be moving the return block *into* a 1001 // loop! 1002 if (DoTransform && !MBB->succ_empty() && 1003 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 1004 DoTransform = false; 1005 1006 1007 if (DoTransform) { 1008 // Reverse the branch so we will fall through on the previous true cond. 1009 std::vector<MachineOperand> NewPriorCond(PriorCond); 1010 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1011 DOUT << "\nMoving MBB: " << *MBB; 1012 DOUT << "To make fallthrough to: " << *PriorTBB << "\n"; 1013 1014 TII->RemoveBranch(PrevBB); 1015 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 1016 1017 // Move this block to the end of the function. 1018 MBB->moveAfter(--MBB->getParent()->end()); 1019 MadeChange = true; 1020 ++NumBranchOpts; 1021 return; 1022 } 1023 } 1024 } 1025 } 1026 1027 // Analyze the branch in the current block. 1028 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1029 std::vector<MachineOperand> CurCond; 1030 bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond); 1031 if (!CurUnAnalyzable) { 1032 // If the CFG for the prior block has extra edges, remove them. 1033 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1034 1035 // If this is a two-way branch, and the FBB branches to this block, reverse 1036 // the condition so the single-basic-block loop is faster. Instead of: 1037 // Loop: xxx; jcc Out; jmp Loop 1038 // we want: 1039 // Loop: xxx; jncc Loop; jmp Out 1040 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1041 std::vector<MachineOperand> NewCond(CurCond); 1042 if (!TII->ReverseBranchCondition(NewCond)) { 1043 TII->RemoveBranch(*MBB); 1044 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 1045 MadeChange = true; 1046 ++NumBranchOpts; 1047 return OptimizeBlock(MBB); 1048 } 1049 } 1050 1051 1052 // If this branch is the only thing in its block, see if we can forward 1053 // other blocks across it. 1054 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1055 MBB->begin()->getDesc().isBranch() && CurTBB != MBB) { 1056 // This block may contain just an unconditional branch. Because there can 1057 // be 'non-branch terminators' in the block, try removing the branch and 1058 // then seeing if the block is empty. 1059 TII->RemoveBranch(*MBB); 1060 1061 // If this block is just an unconditional branch to CurTBB, we can 1062 // usually completely eliminate the block. The only case we cannot 1063 // completely eliminate the block is when the block before this one 1064 // falls through into MBB and we can't understand the prior block's branch 1065 // condition. 1066 if (MBB->empty()) { 1067 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 1068 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1069 !PrevBB.isSuccessor(MBB)) { 1070 // If the prior block falls through into us, turn it into an 1071 // explicit branch to us to make updates simpler. 1072 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1073 PriorTBB != MBB && PriorFBB != MBB) { 1074 if (PriorTBB == 0) { 1075 assert(PriorCond.empty() && PriorFBB == 0 && 1076 "Bad branch analysis"); 1077 PriorTBB = MBB; 1078 } else { 1079 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1080 PriorFBB = MBB; 1081 } 1082 TII->RemoveBranch(PrevBB); 1083 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1084 } 1085 1086 // Iterate through all the predecessors, revectoring each in-turn. 1087 size_t PI = 0; 1088 bool DidChange = false; 1089 bool HasBranchToSelf = false; 1090 while(PI != MBB->pred_size()) { 1091 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1092 if (PMBB == MBB) { 1093 // If this block has an uncond branch to itself, leave it. 1094 ++PI; 1095 HasBranchToSelf = true; 1096 } else { 1097 DidChange = true; 1098 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1099 } 1100 } 1101 1102 // Change any jumptables to go to the new MBB. 1103 MBB->getParent()->getJumpTableInfo()-> 1104 ReplaceMBBInJumpTables(MBB, CurTBB); 1105 if (DidChange) { 1106 ++NumBranchOpts; 1107 MadeChange = true; 1108 if (!HasBranchToSelf) return; 1109 } 1110 } 1111 } 1112 1113 // Add the branch back if the block is more than just an uncond branch. 1114 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1115 } 1116 } 1117 1118 // If the prior block doesn't fall through into this block, and if this 1119 // block doesn't fall through into some other block, see if we can find a 1120 // place to move this block where a fall-through will happen. 1121 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 1122 PriorTBB, PriorFBB, PriorCond)) { 1123 // Now we know that there was no fall-through into this block, check to 1124 // see if it has a fall-through into its successor. 1125 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1126 CurCond); 1127 1128 if (!MBB->isLandingPad()) { 1129 // Check all the predecessors of this block. If one of them has no fall 1130 // throughs, move this block right after it. 1131 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1132 E = MBB->pred_end(); PI != E; ++PI) { 1133 // Analyze the branch at the end of the pred. 1134 MachineBasicBlock *PredBB = *PI; 1135 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1136 if (PredBB != MBB && !CanFallThrough(PredBB) 1137 && (!CurFallsThru || !CurTBB || !CurFBB) 1138 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1139 // If the current block doesn't fall through, just move it. 1140 // If the current block can fall through and does not end with a 1141 // conditional branch, we need to append an unconditional jump to 1142 // the (current) next block. To avoid a possible compile-time 1143 // infinite loop, move blocks only backward in this case. 1144 // Also, if there are already 2 branches here, we cannot add a third; 1145 // this means we have the case 1146 // Bcc next 1147 // B elsewhere 1148 // next: 1149 if (CurFallsThru) { 1150 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1151 CurCond.clear(); 1152 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1153 } 1154 MBB->moveAfter(PredBB); 1155 MadeChange = true; 1156 return OptimizeBlock(MBB); 1157 } 1158 } 1159 } 1160 1161 if (!CurFallsThru) { 1162 // Check all successors to see if we can move this block before it. 1163 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1164 E = MBB->succ_end(); SI != E; ++SI) { 1165 // Analyze the branch at the end of the block before the succ. 1166 MachineBasicBlock *SuccBB = *SI; 1167 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1168 std::vector<MachineOperand> SuccPrevCond; 1169 1170 // If this block doesn't already fall-through to that successor, and if 1171 // the succ doesn't already have a block that can fall through into it, 1172 // and if the successor isn't an EH destination, we can arrange for the 1173 // fallthrough to happen. 1174 if (SuccBB != MBB && !CanFallThrough(SuccPrev) && 1175 !SuccBB->isLandingPad()) { 1176 MBB->moveBefore(SuccBB); 1177 MadeChange = true; 1178 return OptimizeBlock(MBB); 1179 } 1180 } 1181 1182 // Okay, there is no really great place to put this block. If, however, 1183 // the block before this one would be a fall-through if this block were 1184 // removed, move this block to the end of the function. 1185 if (FallThrough != MBB->getParent()->end() && 1186 PrevBB.isSuccessor(FallThrough)) { 1187 MBB->moveAfter(--MBB->getParent()->end()); 1188 MadeChange = true; 1189 return; 1190 } 1191 } 1192 } 1193} 1194