BranchFolding.cpp revision a0045b0723b852027e2a6cf3fe8a0876a7e06280
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 "BranchFolding.h" 21#include "llvm/Function.h" 22#include "llvm/CodeGen/Passes.h" 23#include "llvm/CodeGen/MachineModuleInfo.h" 24#include "llvm/CodeGen/MachineFunctionPass.h" 25#include "llvm/CodeGen/MachineJumpTableInfo.h" 26#include "llvm/CodeGen/RegisterScavenging.h" 27#include "llvm/Target/TargetInstrInfo.h" 28#include "llvm/Target/TargetMachine.h" 29#include "llvm/Target/TargetRegisterInfo.h" 30#include "llvm/Support/CommandLine.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/ErrorHandling.h" 33#include "llvm/Support/raw_ostream.h" 34#include "llvm/ADT/SmallSet.h" 35#include "llvm/ADT/Statistic.h" 36#include "llvm/ADT/STLExtras.h" 37#include <algorithm> 38using namespace llvm; 39 40STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 41STATISTIC(NumBranchOpts, "Number of branches optimized"); 42STATISTIC(NumTailMerge , "Number of block tails merged"); 43static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 44 cl::init(cl::BOU_UNSET), cl::Hidden); 45// Throttle for huge numbers of predecessors (compile speed problems) 46static cl::opt<unsigned> 47TailMergeThreshold("tail-merge-threshold", 48 cl::desc("Max number of predecessors to consider tail merging"), 49 cl::init(150), cl::Hidden); 50 51 52char BranchFolderPass::ID = 0; 53 54FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 55 return new BranchFolderPass(DefaultEnableTailMerge); 56} 57 58bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { 59 return OptimizeFunction(MF, 60 MF.getTarget().getInstrInfo(), 61 MF.getTarget().getRegisterInfo(), 62 getAnalysisIfAvailable<MachineModuleInfo>()); 63} 64 65 66 67BranchFolder::BranchFolder(bool defaultEnableTailMerge) { 68 switch (FlagEnableTailMerge) { 69 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 70 case cl::BOU_TRUE: EnableTailMerge = true; break; 71 case cl::BOU_FALSE: EnableTailMerge = false; break; 72 } 73} 74 75/// RemoveDeadBlock - Remove the specified dead machine basic block from the 76/// function, updating the CFG. 77void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 78 assert(MBB->pred_empty() && "MBB must be dead!"); 79 DEBUG(errs() << "\nRemoving MBB: " << *MBB); 80 81 MachineFunction *MF = MBB->getParent(); 82 // drop all successors. 83 while (!MBB->succ_empty()) 84 MBB->removeSuccessor(MBB->succ_end()-1); 85 86 // If there are any labels in the basic block, unregister them from 87 // MachineModuleInfo. 88 if (MMI && !MBB->empty()) { 89 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 90 I != E; ++I) { 91 if (I->isLabel()) 92 // The label ID # is always operand #0, an immediate. 93 MMI->InvalidateLabel(I->getOperand(0).getImm()); 94 } 95 } 96 97 // Remove the block. 98 MF->erase(MBB); 99} 100 101/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 102/// followed by terminators, and if the implicitly defined registers are not 103/// used by the terminators, remove those implicit_def's. e.g. 104/// BB1: 105/// r0 = implicit_def 106/// r1 = implicit_def 107/// br 108/// This block can be optimized away later if the implicit instructions are 109/// removed. 110bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 111 SmallSet<unsigned, 4> ImpDefRegs; 112 MachineBasicBlock::iterator I = MBB->begin(); 113 while (I != MBB->end()) { 114 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) 115 break; 116 unsigned Reg = I->getOperand(0).getReg(); 117 ImpDefRegs.insert(Reg); 118 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg); 119 unsigned SubReg = *SubRegs; ++SubRegs) 120 ImpDefRegs.insert(SubReg); 121 ++I; 122 } 123 if (ImpDefRegs.empty()) 124 return false; 125 126 MachineBasicBlock::iterator FirstTerm = I; 127 while (I != MBB->end()) { 128 if (!TII->isUnpredicatedTerminator(I)) 129 return false; 130 // See if it uses any of the implicitly defined registers. 131 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 132 MachineOperand &MO = I->getOperand(i); 133 if (!MO.isReg() || !MO.isUse()) 134 continue; 135 unsigned Reg = MO.getReg(); 136 if (ImpDefRegs.count(Reg)) 137 return false; 138 } 139 ++I; 140 } 141 142 I = MBB->begin(); 143 while (I != FirstTerm) { 144 MachineInstr *ImpDefMI = &*I; 145 ++I; 146 MBB->erase(ImpDefMI); 147 } 148 149 return true; 150} 151 152/// OptimizeFunction - Perhaps branch folding, tail merging and other 153/// CFG optimizations on the given function. 154bool BranchFolder::OptimizeFunction(MachineFunction &MF, 155 const TargetInstrInfo *tii, 156 const TargetRegisterInfo *tri, 157 MachineModuleInfo *mmi) { 158 if (!tii) return false; 159 160 TII = tii; 161 TRI = tri; 162 MMI = mmi; 163 164 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 165 166 // Fix CFG. The later algorithms expect it to be right. 167 bool MadeChange = false; 168 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 169 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 170 SmallVector<MachineOperand, 4> Cond; 171 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 172 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 173 MadeChange |= OptimizeImpDefsBlock(MBB); 174 } 175 176 177 bool MadeChangeThisIteration = true; 178 while (MadeChangeThisIteration) { 179 MadeChangeThisIteration = false; 180 MadeChangeThisIteration |= TailMergeBlocks(MF); 181 MadeChangeThisIteration |= OptimizeBranches(MF); 182 MadeChange |= MadeChangeThisIteration; 183 } 184 185 // See if any jump tables have become mergable or dead as the code generator 186 // did its thing. 187 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 188 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 189 if (!JTs.empty()) { 190 // Figure out how these jump tables should be merged. 191 std::vector<unsigned> JTMapping; 192 JTMapping.reserve(JTs.size()); 193 194 // We always keep the 0th jump table. 195 JTMapping.push_back(0); 196 197 // Scan the jump tables, seeing if there are any duplicates. Note that this 198 // is N^2, which should be fixed someday. 199 for (unsigned i = 1, e = JTs.size(); i != e; ++i) { 200 if (JTs[i].MBBs.empty()) 201 JTMapping.push_back(i); 202 else 203 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 204 } 205 206 // If a jump table was merge with another one, walk the function rewriting 207 // references to jump tables to reference the new JT ID's. Keep track of 208 // whether we see a jump table idx, if not, we can delete the JT. 209 BitVector JTIsLive(JTs.size()); 210 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 211 BB != E; ++BB) { 212 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 213 I != E; ++I) 214 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 215 MachineOperand &Op = I->getOperand(op); 216 if (!Op.isJTI()) continue; 217 unsigned NewIdx = JTMapping[Op.getIndex()]; 218 Op.setIndex(NewIdx); 219 220 // Remember that this JT is live. 221 JTIsLive.set(NewIdx); 222 } 223 } 224 225 // Finally, remove dead jump tables. This happens either because the 226 // indirect jump was unreachable (and thus deleted) or because the jump 227 // table was merged with some other one. 228 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 229 if (!JTIsLive.test(i)) { 230 JTI->RemoveJumpTable(i); 231 MadeChange = true; 232 } 233 } 234 235 delete RS; 236 return MadeChange; 237} 238 239//===----------------------------------------------------------------------===// 240// Tail Merging of Blocks 241//===----------------------------------------------------------------------===// 242 243/// HashMachineInstr - Compute a hash value for MI and its operands. 244static unsigned HashMachineInstr(const MachineInstr *MI) { 245 unsigned Hash = MI->getOpcode(); 246 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 247 const MachineOperand &Op = MI->getOperand(i); 248 249 // Merge in bits from the operand if easy. 250 unsigned OperandHash = 0; 251 switch (Op.getType()) { 252 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 253 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 254 case MachineOperand::MO_MachineBasicBlock: 255 OperandHash = Op.getMBB()->getNumber(); 256 break; 257 case MachineOperand::MO_FrameIndex: 258 case MachineOperand::MO_ConstantPoolIndex: 259 case MachineOperand::MO_JumpTableIndex: 260 OperandHash = Op.getIndex(); 261 break; 262 case MachineOperand::MO_GlobalAddress: 263 case MachineOperand::MO_ExternalSymbol: 264 // Global address / external symbol are too hard, don't bother, but do 265 // pull in the offset. 266 OperandHash = Op.getOffset(); 267 break; 268 default: break; 269 } 270 271 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 272 } 273 return Hash; 274} 275 276/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 277/// with no successors, we hash two instructions, because cross-jumping 278/// only saves code when at least two instructions are removed (since a 279/// branch must be inserted). For blocks with a successor, one of the 280/// two blocks to be tail-merged will end with a branch already, so 281/// it gains to cross-jump even for one instruction. 282 283static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 284 unsigned minCommonTailLength) { 285 MachineBasicBlock::const_iterator I = MBB->end(); 286 if (I == MBB->begin()) 287 return 0; // Empty MBB. 288 289 --I; 290 unsigned Hash = HashMachineInstr(I); 291 292 if (I == MBB->begin() || minCommonTailLength == 1) 293 return Hash; // Single instr MBB. 294 295 --I; 296 // Hash in the second-to-last instruction. 297 Hash ^= HashMachineInstr(I) << 2; 298 return Hash; 299} 300 301/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 302/// of instructions they actually have in common together at their end. Return 303/// iterators for the first shared instruction in each block. 304static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 305 MachineBasicBlock *MBB2, 306 MachineBasicBlock::iterator &I1, 307 MachineBasicBlock::iterator &I2) { 308 I1 = MBB1->end(); 309 I2 = MBB2->end(); 310 311 unsigned TailLen = 0; 312 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 313 --I1; --I2; 314 if (!I1->isIdenticalTo(I2) || 315 // FIXME: This check is dubious. It's used to get around a problem where 316 // people incorrectly expect inline asm directives to remain in the same 317 // relative order. This is untenable because normal compiler 318 // optimizations (like this one) may reorder and/or merge these 319 // directives. 320 I1->getOpcode() == TargetInstrInfo::INLINEASM) { 321 ++I1; ++I2; 322 break; 323 } 324 ++TailLen; 325 } 326 return TailLen; 327} 328 329/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 330/// after it, replacing it with an unconditional branch to NewDest. This 331/// returns true if OldInst's block is modified, false if NewDest is modified. 332void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 333 MachineBasicBlock *NewDest) { 334 MachineBasicBlock *OldBB = OldInst->getParent(); 335 336 // Remove all the old successors of OldBB from the CFG. 337 while (!OldBB->succ_empty()) 338 OldBB->removeSuccessor(OldBB->succ_begin()); 339 340 // Remove all the dead instructions from the end of OldBB. 341 OldBB->erase(OldInst, OldBB->end()); 342 343 // If OldBB isn't immediately before OldBB, insert a branch to it. 344 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 345 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>()); 346 OldBB->addSuccessor(NewDest); 347 ++NumTailMerge; 348} 349 350/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 351/// MBB so that the part before the iterator falls into the part starting at the 352/// iterator. This returns the new MBB. 353MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 354 MachineBasicBlock::iterator BBI1) { 355 MachineFunction &MF = *CurMBB.getParent(); 356 357 // Create the fall-through block. 358 MachineFunction::iterator MBBI = &CurMBB; 359 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 360 CurMBB.getParent()->insert(++MBBI, NewMBB); 361 362 // Move all the successors of this block to the specified block. 363 NewMBB->transferSuccessors(&CurMBB); 364 365 // Add an edge from CurMBB to NewMBB for the fall-through. 366 CurMBB.addSuccessor(NewMBB); 367 368 // Splice the code over. 369 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 370 371 // For targets that use the register scavenger, we must maintain LiveIns. 372 if (RS) { 373 RS->enterBasicBlock(&CurMBB); 374 if (!CurMBB.empty()) 375 RS->forward(prior(CurMBB.end())); 376 BitVector RegsLiveAtExit(TRI->getNumRegs()); 377 RS->getRegsUsed(RegsLiveAtExit, false); 378 for (unsigned int i=0, e=TRI->getNumRegs(); i!=e; i++) 379 if (RegsLiveAtExit[i]) 380 NewMBB->addLiveIn(i); 381 } 382 383 return NewMBB; 384} 385 386/// EstimateRuntime - Make a rough estimate for how long it will take to run 387/// the specified code. 388static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 389 MachineBasicBlock::iterator E) { 390 unsigned Time = 0; 391 for (; I != E; ++I) { 392 const TargetInstrDesc &TID = I->getDesc(); 393 if (TID.isCall()) 394 Time += 10; 395 else if (TID.mayLoad() || TID.mayStore()) 396 Time += 2; 397 else 398 ++Time; 399 } 400 return Time; 401} 402 403// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 404// branches temporarily for tail merging). In the case where CurMBB ends 405// with a conditional branch to the next block, optimize by reversing the 406// test and conditionally branching to SuccMBB instead. 407 408static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, 409 const TargetInstrInfo *TII) { 410 MachineFunction *MF = CurMBB->getParent(); 411 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 412 MachineBasicBlock *TBB = 0, *FBB = 0; 413 SmallVector<MachineOperand, 4> Cond; 414 if (I != MF->end() && 415 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 416 MachineBasicBlock *NextBB = I; 417 if (TBB == NextBB && !Cond.empty() && !FBB) { 418 if (!TII->ReverseBranchCondition(Cond)) { 419 TII->RemoveBranch(*CurMBB); 420 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 421 return; 422 } 423 } 424 } 425 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>()); 426} 427 428static bool MergeCompare(const std::pair<unsigned,MachineBasicBlock*> &p, 429 const std::pair<unsigned,MachineBasicBlock*> &q) { 430 if (p.first < q.first) 431 return true; 432 else if (p.first > q.first) 433 return false; 434 else if (p.second->getNumber() < q.second->getNumber()) 435 return true; 436 else if (p.second->getNumber() > q.second->getNumber()) 437 return false; 438 else { 439 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 440 // an object with itself. 441#ifndef _GLIBCXX_DEBUG 442 llvm_unreachable("Predecessor appears twice"); 443#endif 444 return false; 445 } 446} 447 448/// ProfitableToMerge - Check if two machine basic blocks have a common tail 449/// and decide if it would be profitable to merge those tails. Return the 450/// length of the common tail and iterators to the first common instruction 451/// in each block. 452static bool ProfitableToMerge(MachineBasicBlock *MBB1, 453 MachineBasicBlock *MBB2, 454 unsigned minCommonTailLength, 455 unsigned &CommonTailLen, 456 MachineBasicBlock::iterator &I1, 457 MachineBasicBlock::iterator &I2) { 458 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 459 MachineFunction *MF = MBB1->getParent(); 460 461 if (CommonTailLen >= minCommonTailLength) 462 return true; 463 464 if (CommonTailLen == 0) 465 return false; 466 467 // If we are optimizing for code size, 1 instruction in common is enough if 468 // we don't have to split a block. At worst we will be replacing a 469 // fallthrough into the common tail with a branch, which at worst breaks 470 // even with falling through into the duplicated common tail. 471 if (MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 472 (I1 == MBB1->begin() || I2 == MBB2->begin())) 473 return true; 474 475 return false; 476} 477 478/// ComputeSameTails - Look through all the blocks in MergePotentials that have 479/// hash CurHash (guaranteed to match the last element). Build the vector 480/// SameTails of all those that have the (same) largest number of instructions 481/// in common of any pair of these blocks. SameTails entries contain an 482/// iterator into MergePotentials (from which the MachineBasicBlock can be 483/// found) and a MachineBasicBlock::iterator into that MBB indicating the 484/// instruction where the matching code sequence begins. 485/// Order of elements in SameTails is the reverse of the order in which 486/// those blocks appear in MergePotentials (where they are not necessarily 487/// consecutive). 488unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 489 unsigned minCommonTailLength) { 490 unsigned maxCommonTailLength = 0U; 491 SameTails.clear(); 492 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 493 MPIterator HighestMPIter = prior(MergePotentials.end()); 494 for (MPIterator CurMPIter = prior(MergePotentials.end()), 495 B = MergePotentials.begin(); 496 CurMPIter!=B && CurMPIter->first==CurHash; 497 --CurMPIter) { 498 for (MPIterator I = prior(CurMPIter); I->first==CurHash ; --I) { 499 unsigned CommonTailLen; 500 if (ProfitableToMerge(CurMPIter->second, I->second, minCommonTailLength, 501 CommonTailLen, TrialBBI1, TrialBBI2)) { 502 if (CommonTailLen > maxCommonTailLength) { 503 SameTails.clear(); 504 maxCommonTailLength = CommonTailLen; 505 HighestMPIter = CurMPIter; 506 SameTails.push_back(std::make_pair(CurMPIter, TrialBBI1)); 507 } 508 if (HighestMPIter == CurMPIter && 509 CommonTailLen == maxCommonTailLength) 510 SameTails.push_back(std::make_pair(I, TrialBBI2)); 511 } 512 if (I==B) 513 break; 514 } 515 } 516 return maxCommonTailLength; 517} 518 519/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 520/// MergePotentials, restoring branches at ends of blocks as appropriate. 521void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 522 MachineBasicBlock* SuccBB, 523 MachineBasicBlock* PredBB) { 524 MPIterator CurMPIter, B; 525 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 526 CurMPIter->first==CurHash; 527 --CurMPIter) { 528 // Put the unconditional branch back, if we need one. 529 MachineBasicBlock *CurMBB = CurMPIter->second; 530 if (SuccBB && CurMBB != PredBB) 531 FixTail(CurMBB, SuccBB, TII); 532 if (CurMPIter==B) 533 break; 534 } 535 if (CurMPIter->first!=CurHash) 536 CurMPIter++; 537 MergePotentials.erase(CurMPIter, MergePotentials.end()); 538} 539 540/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 541/// only of the common tail. Create a block that does by splitting one. 542unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 543 unsigned maxCommonTailLength) { 544 unsigned i, commonTailIndex; 545 unsigned TimeEstimate = ~0U; 546 for (i=0, commonTailIndex=0; i<SameTails.size(); i++) { 547 // Use PredBB if possible; that doesn't require a new branch. 548 if (SameTails[i].first->second==PredBB) { 549 commonTailIndex = i; 550 break; 551 } 552 // Otherwise, make a (fairly bogus) choice based on estimate of 553 // how long it will take the various blocks to execute. 554 unsigned t = EstimateRuntime(SameTails[i].first->second->begin(), 555 SameTails[i].second); 556 if (t<=TimeEstimate) { 557 TimeEstimate = t; 558 commonTailIndex = i; 559 } 560 } 561 562 MachineBasicBlock::iterator BBI = SameTails[commonTailIndex].second; 563 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; 564 565 DEBUG(errs() << "\nSplitting " << MBB->getNumber() << ", size " 566 << maxCommonTailLength); 567 568 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 569 SameTails[commonTailIndex].first->second = newMBB; 570 SameTails[commonTailIndex].second = newMBB->begin(); 571 // If we split PredBB, newMBB is the new predecessor. 572 if (PredBB==MBB) 573 PredBB = newMBB; 574 575 return commonTailIndex; 576} 577 578// See if any of the blocks in MergePotentials (which all have a common single 579// successor, or all have no successor) can be tail-merged. If there is a 580// successor, any blocks in MergePotentials that are not tail-merged and 581// are not immediately before Succ must have an unconditional branch to 582// Succ added (but the predecessor/successor lists need no adjustment). 583// The lone predecessor of Succ that falls through into Succ, 584// if any, is given in PredBB. 585 586bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, 587 MachineBasicBlock* PredBB) { 588 bool MadeChange = false; 589 590 // It doesn't make sense to save a single instruction since tail merging 591 // will add a jump. 592 // FIXME: Ask the target to provide the threshold? 593 unsigned minCommonTailLength = (SuccBB ? 1 : 2) + 1; 594 595 DEBUG(errs() << "\nTryMergeBlocks " << MergePotentials.size() << '\n'); 596 597 // Sort by hash value so that blocks with identical end sequences sort 598 // together. 599 std::stable_sort(MergePotentials.begin(), MergePotentials.end(),MergeCompare); 600 601 // Walk through equivalence sets looking for actual exact matches. 602 while (MergePotentials.size() > 1) { 603 unsigned CurHash = MergePotentials.back().first; 604 605 // Build SameTails, identifying the set of blocks with this hash code 606 // and with the maximum number of instructions in common. 607 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 608 minCommonTailLength); 609 610 // If we didn't find any pair that has at least minCommonTailLength 611 // instructions in common, remove all blocks with this hash code and retry. 612 if (SameTails.empty()) { 613 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 614 continue; 615 } 616 617 // If one of the blocks is the entire common tail (and not the entry 618 // block, which we can't jump to), we can treat all blocks with this same 619 // tail at once. Use PredBB if that is one of the possibilities, as that 620 // will not introduce any extra branches. 621 MachineBasicBlock *EntryBB = MergePotentials.begin()->second-> 622 getParent()->begin(); 623 unsigned int commonTailIndex, i; 624 for (commonTailIndex=SameTails.size(), i=0; i<SameTails.size(); i++) { 625 MachineBasicBlock *MBB = SameTails[i].first->second; 626 if (MBB->begin() == SameTails[i].second && MBB != EntryBB) { 627 commonTailIndex = i; 628 if (MBB==PredBB) 629 break; 630 } 631 } 632 633 if (commonTailIndex==SameTails.size()) { 634 // None of the blocks consist entirely of the common tail. 635 // Split a block so that one does. 636 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); 637 } 638 639 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; 640 // MBB is common tail. Adjust all other BB's to jump to this one. 641 // Traversal must be forwards so erases work. 642 DEBUG(errs() << "\nUsing common tail " << MBB->getNumber() << " for "); 643 for (unsigned int i=0; i<SameTails.size(); ++i) { 644 if (commonTailIndex==i) 645 continue; 646 DEBUG(errs() << SameTails[i].first->second->getNumber() << ","); 647 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 648 ReplaceTailWithBranchTo(SameTails[i].second, MBB); 649 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 650 MergePotentials.erase(SameTails[i].first); 651 } 652 DEBUG(errs() << "\n"); 653 // We leave commonTailIndex in the worklist in case there are other blocks 654 // that match it with a smaller number of instructions. 655 MadeChange = true; 656 } 657 return MadeChange; 658} 659 660bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 661 662 if (!EnableTailMerge) return false; 663 664 bool MadeChange = false; 665 666 // First find blocks with no successors. 667 MergePotentials.clear(); 668 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 669 if (I->succ_empty()) 670 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I)); 671 } 672 // See if we can do any tail merging on those. 673 if (MergePotentials.size() < TailMergeThreshold && 674 MergePotentials.size() >= 2) 675 MadeChange |= TryMergeBlocks(NULL, NULL); 676 677 // Look at blocks (IBB) with multiple predecessors (PBB). 678 // We change each predecessor to a canonical form, by 679 // (1) temporarily removing any unconditional branch from the predecessor 680 // to IBB, and 681 // (2) alter conditional branches so they branch to the other block 682 // not IBB; this may require adding back an unconditional branch to IBB 683 // later, where there wasn't one coming in. E.g. 684 // Bcc IBB 685 // fallthrough to QBB 686 // here becomes 687 // Bncc QBB 688 // with a conceptual B to IBB after that, which never actually exists. 689 // With those changes, we see whether the predecessors' tails match, 690 // and merge them if so. We change things out of canonical form and 691 // back to the way they were later in the process. (OptimizeBranches 692 // would undo some of this, but we can't use it, because we'd get into 693 // a compile-time infinite loop repeatedly doing and undoing the same 694 // transformations.) 695 696 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 697 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 698 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 699 MachineBasicBlock *IBB = I; 700 MachineBasicBlock *PredBB = prior(I); 701 MergePotentials.clear(); 702 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 703 E2 = I->pred_end(); 704 P != E2; ++P) { 705 MachineBasicBlock* PBB = *P; 706 // Skip blocks that loop to themselves, can't tail merge these. 707 if (PBB==IBB) 708 continue; 709 // Visit each predecessor only once. 710 if (!UniquePreds.insert(PBB)) 711 continue; 712 MachineBasicBlock *TBB = 0, *FBB = 0; 713 SmallVector<MachineOperand, 4> Cond; 714 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 715 // Failing case: IBB is the target of a cbr, and 716 // we cannot reverse the branch. 717 SmallVector<MachineOperand, 4> 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 bool 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 MadeChange |= 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 SmallVectorImpl<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 SmallVector<MachineOperand, 4> Cond; 848 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true); 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. 876bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 877 bool MadeChange = false; 878 MachineFunction &MF = *MBB->getParent(); 879 880 MachineFunction::iterator FallThrough = MBB; 881 ++FallThrough; 882 883 // If this block is empty, make everyone use its fall-through, not the block 884 // explicitly. Landing pads should not do this since the landing-pad table 885 // points to this block. Blocks with their addresses taken shouldn't be 886 // optimized away. 887 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 888 // Dead block? Leave for cleanup later. 889 if (MBB->pred_empty()) return MadeChange; 890 891 if (FallThrough == MF.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 // If MBB was the target of a jump table, update jump tables to go to the 901 // fallthrough instead. 902 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough); 903 MadeChange = true; 904 } 905 return MadeChange; 906 } 907 908 // Check to see if we can simplify the terminator of the block before this 909 // one. 910 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 911 912 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 913 SmallVector<MachineOperand, 4> PriorCond; 914 bool PriorUnAnalyzable = 915 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 916 if (!PriorUnAnalyzable) { 917 // If the CFG for the prior block has extra edges, remove them. 918 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 919 !PriorCond.empty()); 920 921 // If the previous branch is conditional and both conditions go to the same 922 // destination, remove the branch, replacing it with an unconditional one or 923 // a fall-through. 924 if (PriorTBB && PriorTBB == PriorFBB) { 925 TII->RemoveBranch(PrevBB); 926 PriorCond.clear(); 927 if (PriorTBB != MBB) 928 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 929 MadeChange = true; 930 ++NumBranchOpts; 931 return OptimizeBlock(MBB); 932 } 933 934 // If the previous branch *only* branches to *this* block (conditional or 935 // not) remove the branch. 936 if (PriorTBB == MBB && PriorFBB == 0) { 937 TII->RemoveBranch(PrevBB); 938 MadeChange = true; 939 ++NumBranchOpts; 940 return OptimizeBlock(MBB); 941 } 942 943 // If the prior block branches somewhere else on the condition and here if 944 // the condition is false, remove the uncond second branch. 945 if (PriorFBB == MBB) { 946 TII->RemoveBranch(PrevBB); 947 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 948 MadeChange = true; 949 ++NumBranchOpts; 950 return OptimizeBlock(MBB); 951 } 952 953 // If the prior block branches here on true and somewhere else on false, and 954 // if the branch condition is reversible, reverse the branch to create a 955 // fall-through. 956 if (PriorTBB == MBB) { 957 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 958 if (!TII->ReverseBranchCondition(NewPriorCond)) { 959 TII->RemoveBranch(PrevBB); 960 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 961 MadeChange = true; 962 ++NumBranchOpts; 963 return OptimizeBlock(MBB); 964 } 965 } 966 967 // If this block has no successors (e.g. it is a return block or ends with 968 // a call to a no-return function like abort or __cxa_throw) and if the pred 969 // falls through into this block, and if it would otherwise fall through 970 // into the block after this, move this block to the end of the function. 971 // 972 // We consider it more likely that execution will stay in the function (e.g. 973 // due to loops) than it is to exit it. This asserts in loops etc, moving 974 // the assert condition out of the loop body. 975 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 976 MachineFunction::iterator(PriorTBB) == FallThrough && 977 !CanFallThrough(MBB)) { 978 bool DoTransform = true; 979 980 // We have to be careful that the succs of PredBB aren't both no-successor 981 // blocks. If neither have successors and if PredBB is the second from 982 // last block in the function, we'd just keep swapping the two blocks for 983 // last. Only do the swap if one is clearly better to fall through than 984 // the other. 985 if (FallThrough == --MF.end() && 986 !IsBetterFallthrough(PriorTBB, MBB)) 987 DoTransform = false; 988 989 // We don't want to do this transformation if we have control flow like: 990 // br cond BB2 991 // BB1: 992 // .. 993 // jmp BBX 994 // BB2: 995 // .. 996 // ret 997 // 998 // In this case, we could actually be moving the return block *into* a 999 // loop! 1000 if (DoTransform && !MBB->succ_empty() && 1001 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 1002 DoTransform = false; 1003 1004 1005 if (DoTransform) { 1006 // Reverse the branch so we will fall through on the previous true cond. 1007 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1008 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1009 DEBUG(errs() << "\nMoving MBB: " << *MBB 1010 << "To make fallthrough to: " << *PriorTBB << "\n"); 1011 1012 TII->RemoveBranch(PrevBB); 1013 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 1014 1015 // Move this block to the end of the function. 1016 MBB->moveAfter(--MF.end()); 1017 MadeChange = true; 1018 ++NumBranchOpts; 1019 return MadeChange; 1020 } 1021 } 1022 } 1023 } 1024 1025 // Analyze the branch in the current block. 1026 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1027 SmallVector<MachineOperand, 4> CurCond; 1028 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1029 if (!CurUnAnalyzable) { 1030 // If the CFG for the prior block has extra edges, remove them. 1031 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1032 1033 // If this is a two-way branch, and the FBB branches to this block, reverse 1034 // the condition so the single-basic-block loop is faster. Instead of: 1035 // Loop: xxx; jcc Out; jmp Loop 1036 // we want: 1037 // Loop: xxx; jncc Loop; jmp Out 1038 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1039 SmallVector<MachineOperand, 4> NewCond(CurCond); 1040 if (!TII->ReverseBranchCondition(NewCond)) { 1041 TII->RemoveBranch(*MBB); 1042 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 1043 MadeChange = true; 1044 ++NumBranchOpts; 1045 return OptimizeBlock(MBB); 1046 } 1047 } 1048 1049 1050 // If this branch is the only thing in its block, see if we can forward 1051 // other blocks across it. 1052 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1053 MBB->begin()->getDesc().isBranch() && CurTBB != MBB && 1054 !MBB->hasAddressTaken()) { 1055 // This block may contain just an unconditional branch. Because there can 1056 // be 'non-branch terminators' in the block, try removing the branch and 1057 // then seeing if the block is empty. 1058 TII->RemoveBranch(*MBB); 1059 1060 // If this block is just an unconditional branch to CurTBB, we can 1061 // usually completely eliminate the block. The only case we cannot 1062 // completely eliminate the block is when the block before this one 1063 // falls through into MBB and we can't understand the prior block's branch 1064 // condition. 1065 if (MBB->empty()) { 1066 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 1067 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1068 !PrevBB.isSuccessor(MBB)) { 1069 // If the prior block falls through into us, turn it into an 1070 // explicit branch to us to make updates simpler. 1071 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1072 PriorTBB != MBB && PriorFBB != MBB) { 1073 if (PriorTBB == 0) { 1074 assert(PriorCond.empty() && PriorFBB == 0 && 1075 "Bad branch analysis"); 1076 PriorTBB = MBB; 1077 } else { 1078 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1079 PriorFBB = MBB; 1080 } 1081 TII->RemoveBranch(PrevBB); 1082 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1083 } 1084 1085 // Iterate through all the predecessors, revectoring each in-turn. 1086 size_t PI = 0; 1087 bool DidChange = false; 1088 bool HasBranchToSelf = false; 1089 while(PI != MBB->pred_size()) { 1090 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1091 if (PMBB == MBB) { 1092 // If this block has an uncond branch to itself, leave it. 1093 ++PI; 1094 HasBranchToSelf = true; 1095 } else { 1096 DidChange = true; 1097 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1098 // If this change resulted in PMBB ending in a conditional 1099 // branch where both conditions go to the same destination, 1100 // change this to an unconditional branch (and fix the CFG). 1101 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1102 SmallVector<MachineOperand, 4> NewCurCond; 1103 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1104 NewCurFBB, NewCurCond, true); 1105 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1106 TII->RemoveBranch(*PMBB); 1107 NewCurCond.clear(); 1108 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond); 1109 MadeChange = true; 1110 ++NumBranchOpts; 1111 PMBB->CorrectExtraCFGEdges(NewCurTBB, NewCurFBB, false); 1112 } 1113 } 1114 } 1115 1116 // Change any jumptables to go to the new MBB. 1117 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB); 1118 if (DidChange) { 1119 ++NumBranchOpts; 1120 MadeChange = true; 1121 if (!HasBranchToSelf) return MadeChange; 1122 } 1123 } 1124 } 1125 1126 // Add the branch back if the block is more than just an uncond branch. 1127 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1128 } 1129 } 1130 1131 // If the prior block doesn't fall through into this block, and if this 1132 // block doesn't fall through into some other block, see if we can find a 1133 // place to move this block where a fall-through will happen. 1134 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 1135 PriorTBB, PriorFBB, PriorCond)) { 1136 // Now we know that there was no fall-through into this block, check to 1137 // see if it has a fall-through into its successor. 1138 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1139 CurCond); 1140 1141 if (!MBB->isLandingPad()) { 1142 // Check all the predecessors of this block. If one of them has no fall 1143 // throughs, move this block right after it. 1144 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1145 E = MBB->pred_end(); PI != E; ++PI) { 1146 // Analyze the branch at the end of the pred. 1147 MachineBasicBlock *PredBB = *PI; 1148 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1149 if (PredBB != MBB && !CanFallThrough(PredBB) 1150 && (!CurFallsThru || !CurTBB || !CurFBB) 1151 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1152 // If the current block doesn't fall through, just move it. 1153 // If the current block can fall through and does not end with a 1154 // conditional branch, we need to append an unconditional jump to 1155 // the (current) next block. To avoid a possible compile-time 1156 // infinite loop, move blocks only backward in this case. 1157 // Also, if there are already 2 branches here, we cannot add a third; 1158 // this means we have the case 1159 // Bcc next 1160 // B elsewhere 1161 // next: 1162 if (CurFallsThru) { 1163 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1164 CurCond.clear(); 1165 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1166 } 1167 MBB->moveAfter(PredBB); 1168 MadeChange = true; 1169 return OptimizeBlock(MBB); 1170 } 1171 } 1172 } 1173 1174 if (!CurFallsThru) { 1175 // Check all successors to see if we can move this block before it. 1176 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1177 E = MBB->succ_end(); SI != E; ++SI) { 1178 // Analyze the branch at the end of the block before the succ. 1179 MachineBasicBlock *SuccBB = *SI; 1180 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1181 1182 // If this block doesn't already fall-through to that successor, and if 1183 // the succ doesn't already have a block that can fall through into it, 1184 // and if the successor isn't an EH destination, we can arrange for the 1185 // fallthrough to happen. 1186 if (SuccBB != MBB && !CanFallThrough(SuccPrev) && 1187 !SuccBB->isLandingPad()) { 1188 MBB->moveBefore(SuccBB); 1189 MadeChange = true; 1190 return OptimizeBlock(MBB); 1191 } 1192 } 1193 1194 // Okay, there is no really great place to put this block. If, however, 1195 // the block before this one would be a fall-through if this block were 1196 // removed, move this block to the end of the function. 1197 if (FallThrough != MF.end() && 1198 PrevBB.isSuccessor(FallThrough)) { 1199 MBB->moveAfter(--MF.end()); 1200 MadeChange = true; 1201 return MadeChange; 1202 } 1203 } 1204 } 1205 1206 return MadeChange; 1207} 1208