BranchFolding.cpp revision 4e3f125e184f96ae72f2c44d16cafe0d44158283
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 66BranchFolder::BranchFolder(bool defaultEnableTailMerge) { 67 switch (FlagEnableTailMerge) { 68 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 69 case cl::BOU_TRUE: EnableTailMerge = true; break; 70 case cl::BOU_FALSE: EnableTailMerge = false; break; 71 } 72} 73 74/// RemoveDeadBlock - Remove the specified dead machine basic block from the 75/// function, updating the CFG. 76void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 77 assert(MBB->pred_empty() && "MBB must be dead!"); 78 DEBUG(errs() << "\nRemoving MBB: " << *MBB); 79 80 MachineFunction *MF = MBB->getParent(); 81 // drop all successors. 82 while (!MBB->succ_empty()) 83 MBB->removeSuccessor(MBB->succ_end()-1); 84 85 // If there are any labels in the basic block, unregister them from 86 // MachineModuleInfo. 87 if (MMI && !MBB->empty()) { 88 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 89 I != E; ++I) { 90 if (I->isLabel()) 91 // The label ID # is always operand #0, an immediate. 92 MMI->InvalidateLabel(I->getOperand(0).getImm()); 93 } 94 } 95 96 // Remove the block. 97 MF->erase(MBB); 98} 99 100/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 101/// followed by terminators, and if the implicitly defined registers are not 102/// used by the terminators, remove those implicit_def's. e.g. 103/// BB1: 104/// r0 = implicit_def 105/// r1 = implicit_def 106/// br 107/// This block can be optimized away later if the implicit instructions are 108/// removed. 109bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 110 SmallSet<unsigned, 4> ImpDefRegs; 111 MachineBasicBlock::iterator I = MBB->begin(); 112 while (I != MBB->end()) { 113 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) 114 break; 115 unsigned Reg = I->getOperand(0).getReg(); 116 ImpDefRegs.insert(Reg); 117 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg); 118 unsigned SubReg = *SubRegs; ++SubRegs) 119 ImpDefRegs.insert(SubReg); 120 ++I; 121 } 122 if (ImpDefRegs.empty()) 123 return false; 124 125 MachineBasicBlock::iterator FirstTerm = I; 126 while (I != MBB->end()) { 127 if (!TII->isUnpredicatedTerminator(I)) 128 return false; 129 // See if it uses any of the implicitly defined registers. 130 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 131 MachineOperand &MO = I->getOperand(i); 132 if (!MO.isReg() || !MO.isUse()) 133 continue; 134 unsigned Reg = MO.getReg(); 135 if (ImpDefRegs.count(Reg)) 136 return false; 137 } 138 ++I; 139 } 140 141 I = MBB->begin(); 142 while (I != FirstTerm) { 143 MachineInstr *ImpDefMI = &*I; 144 ++I; 145 MBB->erase(ImpDefMI); 146 } 147 148 return true; 149} 150 151/// OptimizeFunction - Perhaps branch folding, tail merging and other 152/// CFG optimizations on the given function. 153bool BranchFolder::OptimizeFunction(MachineFunction &MF, 154 const TargetInstrInfo *tii, 155 const TargetRegisterInfo *tri, 156 MachineModuleInfo *mmi) { 157 if (!tii) return false; 158 159 TII = tii; 160 TRI = tri; 161 MMI = mmi; 162 163 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 164 165 // Fix CFG. The later algorithms expect it to be right. 166 bool MadeChange = false; 167 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 168 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 169 SmallVector<MachineOperand, 4> Cond; 170 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 171 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 172 MadeChange |= OptimizeImpDefsBlock(MBB); 173 } 174 175 176 bool MadeChangeThisIteration = true; 177 while (MadeChangeThisIteration) { 178 MadeChangeThisIteration = false; 179 MadeChangeThisIteration |= TailMergeBlocks(MF); 180 MadeChangeThisIteration |= OptimizeBranches(MF); 181 MadeChange |= MadeChangeThisIteration; 182 } 183 184 // See if any jump tables have become mergable or dead as the code generator 185 // did its thing. 186 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 187 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 188 if (!JTs.empty()) { 189 // Figure out how these jump tables should be merged. 190 std::vector<unsigned> JTMapping; 191 JTMapping.reserve(JTs.size()); 192 193 // We always keep the 0th jump table. 194 JTMapping.push_back(0); 195 196 // Scan the jump tables, seeing if there are any duplicates. Note that this 197 // is N^2, which should be fixed someday. 198 for (unsigned i = 1, e = JTs.size(); i != e; ++i) { 199 if (JTs[i].MBBs.empty()) 200 JTMapping.push_back(i); 201 else 202 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 203 } 204 205 // If a jump table was merge with another one, walk the function rewriting 206 // references to jump tables to reference the new JT ID's. Keep track of 207 // whether we see a jump table idx, if not, we can delete the JT. 208 BitVector JTIsLive(JTs.size()); 209 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 210 BB != E; ++BB) { 211 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 212 I != E; ++I) 213 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 214 MachineOperand &Op = I->getOperand(op); 215 if (!Op.isJTI()) continue; 216 unsigned NewIdx = JTMapping[Op.getIndex()]; 217 Op.setIndex(NewIdx); 218 219 // Remember that this JT is live. 220 JTIsLive.set(NewIdx); 221 } 222 } 223 224 // Finally, remove dead jump tables. This happens either because the 225 // indirect jump was unreachable (and thus deleted) or because the jump 226 // table was merged with some other one. 227 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 228 if (!JTIsLive.test(i)) { 229 JTI->RemoveJumpTable(i); 230 MadeChange = true; 231 } 232 } 233 234 delete RS; 235 return MadeChange; 236} 237 238//===----------------------------------------------------------------------===// 239// Tail Merging of Blocks 240//===----------------------------------------------------------------------===// 241 242/// HashMachineInstr - Compute a hash value for MI and its operands. 243static unsigned HashMachineInstr(const MachineInstr *MI) { 244 unsigned Hash = MI->getOpcode(); 245 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 246 const MachineOperand &Op = MI->getOperand(i); 247 248 // Merge in bits from the operand if easy. 249 unsigned OperandHash = 0; 250 switch (Op.getType()) { 251 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 252 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 253 case MachineOperand::MO_MachineBasicBlock: 254 OperandHash = Op.getMBB()->getNumber(); 255 break; 256 case MachineOperand::MO_FrameIndex: 257 case MachineOperand::MO_ConstantPoolIndex: 258 case MachineOperand::MO_JumpTableIndex: 259 OperandHash = Op.getIndex(); 260 break; 261 case MachineOperand::MO_GlobalAddress: 262 case MachineOperand::MO_ExternalSymbol: 263 // Global address / external symbol are too hard, don't bother, but do 264 // pull in the offset. 265 OperandHash = Op.getOffset(); 266 break; 267 default: break; 268 } 269 270 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 271 } 272 return Hash; 273} 274 275/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 276/// with no successors, we hash two instructions, because cross-jumping 277/// only saves code when at least two instructions are removed (since a 278/// branch must be inserted). For blocks with a successor, one of the 279/// two blocks to be tail-merged will end with a branch already, so 280/// it gains to cross-jump even for one instruction. 281static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 282 unsigned minCommonTailLength) { 283 MachineBasicBlock::const_iterator I = MBB->end(); 284 if (I == MBB->begin()) 285 return 0; // Empty MBB. 286 287 --I; 288 unsigned Hash = HashMachineInstr(I); 289 290 if (I == MBB->begin() || minCommonTailLength == 1) 291 return Hash; // Single instr MBB. 292 293 --I; 294 // Hash in the second-to-last instruction. 295 Hash ^= HashMachineInstr(I) << 2; 296 return Hash; 297} 298 299/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 300/// of instructions they actually have in common together at their end. Return 301/// iterators for the first shared instruction in each block. 302static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 303 MachineBasicBlock *MBB2, 304 MachineBasicBlock::iterator &I1, 305 MachineBasicBlock::iterator &I2) { 306 I1 = MBB1->end(); 307 I2 = MBB2->end(); 308 309 unsigned TailLen = 0; 310 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 311 --I1; --I2; 312 if (!I1->isIdenticalTo(I2) || 313 // FIXME: This check is dubious. It's used to get around a problem where 314 // people incorrectly expect inline asm directives to remain in the same 315 // relative order. This is untenable because normal compiler 316 // optimizations (like this one) may reorder and/or merge these 317 // directives. 318 I1->getOpcode() == TargetInstrInfo::INLINEASM) { 319 ++I1; ++I2; 320 break; 321 } 322 ++TailLen; 323 } 324 return TailLen; 325} 326 327/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 328/// after it, replacing it with an unconditional branch to NewDest. This 329/// returns true if OldInst's block is modified, false if NewDest is modified. 330void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 331 MachineBasicBlock *NewDest) { 332 MachineBasicBlock *OldBB = OldInst->getParent(); 333 334 // Remove all the old successors of OldBB from the CFG. 335 while (!OldBB->succ_empty()) 336 OldBB->removeSuccessor(OldBB->succ_begin()); 337 338 // Remove all the dead instructions from the end of OldBB. 339 OldBB->erase(OldInst, OldBB->end()); 340 341 // If OldBB isn't immediately before OldBB, insert a branch to it. 342 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 343 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>()); 344 OldBB->addSuccessor(NewDest); 345 ++NumTailMerge; 346} 347 348/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 349/// MBB so that the part before the iterator falls into the part starting at the 350/// iterator. This returns the new MBB. 351MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 352 MachineBasicBlock::iterator BBI1) { 353 MachineFunction &MF = *CurMBB.getParent(); 354 355 // Create the fall-through block. 356 MachineFunction::iterator MBBI = &CurMBB; 357 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 358 CurMBB.getParent()->insert(++MBBI, NewMBB); 359 360 // Move all the successors of this block to the specified block. 361 NewMBB->transferSuccessors(&CurMBB); 362 363 // Add an edge from CurMBB to NewMBB for the fall-through. 364 CurMBB.addSuccessor(NewMBB); 365 366 // Splice the code over. 367 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 368 369 // For targets that use the register scavenger, we must maintain LiveIns. 370 if (RS) { 371 RS->enterBasicBlock(&CurMBB); 372 if (!CurMBB.empty()) 373 RS->forward(prior(CurMBB.end())); 374 BitVector RegsLiveAtExit(TRI->getNumRegs()); 375 RS->getRegsUsed(RegsLiveAtExit, false); 376 for (unsigned int i=0, e=TRI->getNumRegs(); i!=e; i++) 377 if (RegsLiveAtExit[i]) 378 NewMBB->addLiveIn(i); 379 } 380 381 return NewMBB; 382} 383 384/// EstimateRuntime - Make a rough estimate for how long it will take to run 385/// the specified code. 386static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 387 MachineBasicBlock::iterator E) { 388 unsigned Time = 0; 389 for (; I != E; ++I) { 390 const TargetInstrDesc &TID = I->getDesc(); 391 if (TID.isCall()) 392 Time += 10; 393 else if (TID.mayLoad() || TID.mayStore()) 394 Time += 2; 395 else 396 ++Time; 397 } 398 return Time; 399} 400 401// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 402// branches temporarily for tail merging). In the case where CurMBB ends 403// with a conditional branch to the next block, optimize by reversing the 404// test and conditionally branching to SuccMBB instead. 405static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, 406 const TargetInstrInfo *TII) { 407 MachineFunction *MF = CurMBB->getParent(); 408 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 409 MachineBasicBlock *TBB = 0, *FBB = 0; 410 SmallVector<MachineOperand, 4> Cond; 411 if (I != MF->end() && 412 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 413 MachineBasicBlock *NextBB = I; 414 if (TBB == NextBB && !Cond.empty() && !FBB) { 415 if (!TII->ReverseBranchCondition(Cond)) { 416 TII->RemoveBranch(*CurMBB); 417 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 418 return; 419 } 420 } 421 } 422 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>()); 423} 424 425static bool MergeCompare(const std::pair<unsigned,MachineBasicBlock*> &p, 426 const std::pair<unsigned,MachineBasicBlock*> &q) { 427 if (p.first < q.first) 428 return true; 429 else if (p.first > q.first) 430 return false; 431 else if (p.second->getNumber() < q.second->getNumber()) 432 return true; 433 else if (p.second->getNumber() > q.second->getNumber()) 434 return false; 435 else { 436 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 437 // an object with itself. 438#ifndef _GLIBCXX_DEBUG 439 llvm_unreachable("Predecessor appears twice"); 440#endif 441 return false; 442 } 443} 444 445/// ProfitableToMerge - Check if two machine basic blocks have a common tail 446/// and decide if it would be profitable to merge those tails. Return the 447/// length of the common tail and iterators to the first common instruction 448/// in each block. 449static bool ProfitableToMerge(MachineBasicBlock *MBB1, 450 MachineBasicBlock *MBB2, 451 unsigned minCommonTailLength, 452 unsigned &CommonTailLen, 453 MachineBasicBlock::iterator &I1, 454 MachineBasicBlock::iterator &I2) { 455 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 456 MachineFunction *MF = MBB1->getParent(); 457 458 if (CommonTailLen >= minCommonTailLength) 459 return true; 460 461 if (CommonTailLen == 0) 462 return false; 463 464 // If we are optimizing for code size, 1 instruction in common is enough if 465 // we don't have to split a block. At worst we will be replacing a 466 // fallthrough into the common tail with a branch, which at worst breaks 467 // even with falling through into the duplicated common tail. 468 if (MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 469 (I1 == MBB1->begin() || I2 == MBB2->begin())) 470 return true; 471 472 return false; 473} 474 475/// ComputeSameTails - Look through all the blocks in MergePotentials that have 476/// hash CurHash (guaranteed to match the last element). Build the vector 477/// SameTails of all those that have the (same) largest number of instructions 478/// in common of any pair of these blocks. SameTails entries contain an 479/// iterator into MergePotentials (from which the MachineBasicBlock can be 480/// found) and a MachineBasicBlock::iterator into that MBB indicating the 481/// instruction where the matching code sequence begins. 482/// Order of elements in SameTails is the reverse of the order in which 483/// those blocks appear in MergePotentials (where they are not necessarily 484/// consecutive). 485unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 486 unsigned minCommonTailLength) { 487 unsigned maxCommonTailLength = 0U; 488 SameTails.clear(); 489 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 490 MPIterator HighestMPIter = prior(MergePotentials.end()); 491 for (MPIterator CurMPIter = prior(MergePotentials.end()), 492 B = MergePotentials.begin(); 493 CurMPIter!=B && CurMPIter->first == CurHash; 494 --CurMPIter) { 495 for (MPIterator I = prior(CurMPIter); I->first == CurHash ; --I) { 496 unsigned CommonTailLen; 497 if (ProfitableToMerge(CurMPIter->second, I->second, minCommonTailLength, 498 CommonTailLen, TrialBBI1, TrialBBI2)) { 499 if (CommonTailLen > maxCommonTailLength) { 500 SameTails.clear(); 501 maxCommonTailLength = CommonTailLen; 502 HighestMPIter = CurMPIter; 503 SameTails.push_back(std::make_pair(CurMPIter, TrialBBI1)); 504 } 505 if (HighestMPIter == CurMPIter && 506 CommonTailLen == maxCommonTailLength) 507 SameTails.push_back(std::make_pair(I, TrialBBI2)); 508 } 509 if (I == B) 510 break; 511 } 512 } 513 return maxCommonTailLength; 514} 515 516/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 517/// MergePotentials, restoring branches at ends of blocks as appropriate. 518void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 519 MachineBasicBlock* SuccBB, 520 MachineBasicBlock* PredBB) { 521 MPIterator CurMPIter, B; 522 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 523 CurMPIter->first == CurHash; 524 --CurMPIter) { 525 // Put the unconditional branch back, if we need one. 526 MachineBasicBlock *CurMBB = CurMPIter->second; 527 if (SuccBB && CurMBB != PredBB) 528 FixTail(CurMBB, SuccBB, TII); 529 if (CurMPIter == B) 530 break; 531 } 532 if (CurMPIter->first!=CurHash) 533 CurMPIter++; 534 MergePotentials.erase(CurMPIter, MergePotentials.end()); 535} 536 537/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 538/// only of the common tail. Create a block that does by splitting one. 539unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 540 unsigned maxCommonTailLength) { 541 unsigned i, commonTailIndex; 542 unsigned TimeEstimate = ~0U; 543 for (i=0, commonTailIndex=0; i<SameTails.size(); i++) { 544 // Use PredBB if possible; that doesn't require a new branch. 545 if (SameTails[i].first->second == PredBB) { 546 commonTailIndex = i; 547 break; 548 } 549 // Otherwise, make a (fairly bogus) choice based on estimate of 550 // how long it will take the various blocks to execute. 551 unsigned t = EstimateRuntime(SameTails[i].first->second->begin(), 552 SameTails[i].second); 553 if (t <= TimeEstimate) { 554 TimeEstimate = t; 555 commonTailIndex = i; 556 } 557 } 558 559 MachineBasicBlock::iterator BBI = SameTails[commonTailIndex].second; 560 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; 561 562 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 563 << maxCommonTailLength); 564 565 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 566 SameTails[commonTailIndex].first->second = newMBB; 567 SameTails[commonTailIndex].second = newMBB->begin(); 568 569 // If we split PredBB, newMBB is the new predecessor. 570 if (PredBB == MBB) 571 PredBB = newMBB; 572 573 return commonTailIndex; 574} 575 576// See if any of the blocks in MergePotentials (which all have a common single 577// successor, or all have no successor) can be tail-merged. If there is a 578// successor, any blocks in MergePotentials that are not tail-merged and 579// are not immediately before Succ must have an unconditional branch to 580// Succ added (but the predecessor/successor lists need no adjustment). 581// The lone predecessor of Succ that falls through into Succ, 582// if any, is given in PredBB. 583 584bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, 585 MachineBasicBlock* PredBB) { 586 bool MadeChange = false; 587 588 // It doesn't make sense to save a single instruction since tail merging 589 // will add a jump. 590 // FIXME: Ask the target to provide the threshold? 591 unsigned minCommonTailLength = (SuccBB ? 1 : 2) + 1; 592 593 DEBUG(errs() << "\nTryMergeBlocks " << MergePotentials.size() << '\n'); 594 595 // Sort by hash value so that blocks with identical end sequences sort 596 // together. 597 std::stable_sort(MergePotentials.begin(), MergePotentials.end(),MergeCompare); 598 599 // Walk through equivalence sets looking for actual exact matches. 600 while (MergePotentials.size() > 1) { 601 unsigned CurHash = MergePotentials.back().first; 602 603 // Build SameTails, identifying the set of blocks with this hash code 604 // and with the maximum number of instructions in common. 605 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 606 minCommonTailLength); 607 608 // If we didn't find any pair that has at least minCommonTailLength 609 // instructions in common, remove all blocks with this hash code and retry. 610 if (SameTails.empty()) { 611 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 612 continue; 613 } 614 615 // If one of the blocks is the entire common tail (and not the entry 616 // block, which we can't jump to), we can treat all blocks with this same 617 // tail at once. Use PredBB if that is one of the possibilities, as that 618 // will not introduce any extra branches. 619 MachineBasicBlock *EntryBB = MergePotentials.begin()->second-> 620 getParent()->begin(); 621 unsigned int commonTailIndex, i; 622 for (commonTailIndex=SameTails.size(), i=0; i<SameTails.size(); i++) { 623 MachineBasicBlock *MBB = SameTails[i].first->second; 624 if (MBB->begin() == SameTails[i].second && MBB != EntryBB) { 625 commonTailIndex = i; 626 if (MBB == PredBB) 627 break; 628 } 629 } 630 631 if (commonTailIndex == SameTails.size()) { 632 // None of the blocks consist entirely of the common tail. 633 // Split a block so that one does. 634 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); 635 } 636 637 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; 638 // MBB is common tail. Adjust all other BB's to jump to this one. 639 // Traversal must be forwards so erases work. 640 DEBUG(errs() << "\nUsing common tail BB#" << MBB->getNumber() << " for "); 641 for (unsigned int i=0; i<SameTails.size(); ++i) { 642 if (commonTailIndex == i) 643 continue; 644 DEBUG(errs() << "BB#" << SameTails[i].first->second->getNumber() << ", "); 645 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 646 ReplaceTailWithBranchTo(SameTails[i].second, MBB); 647 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 648 MergePotentials.erase(SameTails[i].first); 649 } 650 DEBUG(errs() << "\n"); 651 // We leave commonTailIndex in the worklist in case there are other blocks 652 // that match it with a smaller number of instructions. 653 MadeChange = true; 654 } 655 return MadeChange; 656} 657 658bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 659 660 if (!EnableTailMerge) return false; 661 662 bool MadeChange = false; 663 664 // First find blocks with no successors. 665 MergePotentials.clear(); 666 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 667 if (I->succ_empty()) 668 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I)); 669 } 670 671 // See if we can do any tail merging on those. 672 if (MergePotentials.size() < TailMergeThreshold && 673 MergePotentials.size() >= 2) 674 MadeChange |= TryMergeBlocks(NULL, NULL); 675 676 // Look at blocks (IBB) with multiple predecessors (PBB). 677 // We change each predecessor to a canonical form, by 678 // (1) temporarily removing any unconditional branch from the predecessor 679 // to IBB, and 680 // (2) alter conditional branches so they branch to the other block 681 // not IBB; this may require adding back an unconditional branch to IBB 682 // later, where there wasn't one coming in. E.g. 683 // Bcc IBB 684 // fallthrough to QBB 685 // here becomes 686 // Bncc QBB 687 // with a conceptual B to IBB after that, which never actually exists. 688 // With those changes, we see whether the predecessors' tails match, 689 // and merge them if so. We change things out of canonical form and 690 // back to the way they were later in the process. (OptimizeBranches 691 // would undo some of this, but we can't use it, because we'd get into 692 // a compile-time infinite loop repeatedly doing and undoing the same 693 // transformations.) 694 695 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 696 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 697 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 698 MachineBasicBlock *IBB = I; 699 MachineBasicBlock *PredBB = prior(I); 700 MergePotentials.clear(); 701 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 702 E2 = I->pred_end(); 703 P != E2; ++P) { 704 MachineBasicBlock* PBB = *P; 705 // Skip blocks that loop to themselves, can't tail merge these. 706 if (PBB == IBB) 707 continue; 708 // Visit each predecessor only once. 709 if (!UniquePreds.insert(PBB)) 710 continue; 711 MachineBasicBlock *TBB = 0, *FBB = 0; 712 SmallVector<MachineOperand, 4> Cond; 713 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 714 // Failing case: IBB is the target of a cbr, and 715 // we cannot reverse the branch. 716 SmallVector<MachineOperand, 4> NewCond(Cond); 717 if (!Cond.empty() && TBB == IBB) { 718 if (TII->ReverseBranchCondition(NewCond)) 719 continue; 720 // This is the QBB case described above 721 if (!FBB) 722 FBB = next(MachineFunction::iterator(PBB)); 723 } 724 // Failing case: the only way IBB can be reached from PBB is via 725 // exception handling. Happens for landing pads. Would be nice 726 // to have a bit in the edge so we didn't have to do all this. 727 if (IBB->isLandingPad()) { 728 MachineFunction::iterator IP = PBB; IP++; 729 MachineBasicBlock* PredNextBB = NULL; 730 if (IP!=MF.end()) 731 PredNextBB = IP; 732 if (TBB == NULL) { 733 if (IBB!=PredNextBB) // fallthrough 734 continue; 735 } else if (FBB) { 736 if (TBB!=IBB && FBB!=IBB) // cbr then ubr 737 continue; 738 } else if (Cond.empty()) { 739 if (TBB!=IBB) // ubr 740 continue; 741 } else { 742 if (TBB!=IBB && IBB!=PredNextBB) // cbr 743 continue; 744 } 745 } 746 // Remove the unconditional branch at the end, if any. 747 if (TBB && (Cond.empty() || FBB)) { 748 TII->RemoveBranch(*PBB); 749 if (!Cond.empty()) 750 // reinsert conditional branch only, for now 751 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond); 752 } 753 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P)); 754 } 755 } 756 if (MergePotentials.size() >= 2) 757 MadeChange |= TryMergeBlocks(I, PredBB); 758 // Reinsert an unconditional branch if needed. 759 // The 1 below can occur as a result of removing blocks in TryMergeBlocks. 760 PredBB = prior(I); // this may have been changed in TryMergeBlocks 761 if (MergePotentials.size() == 1 && 762 MergePotentials.begin()->second != PredBB) 763 FixTail(MergePotentials.begin()->second, I, TII); 764 } 765 } 766 return MadeChange; 767} 768 769//===----------------------------------------------------------------------===// 770// Branch Optimization 771//===----------------------------------------------------------------------===// 772 773bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 774 bool MadeChange = false; 775 776 // Make sure blocks are numbered in order 777 MF.RenumberBlocks(); 778 779 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 780 MachineBasicBlock *MBB = I++; 781 MadeChange |= OptimizeBlock(MBB); 782 783 // If it is dead, remove it. 784 if (MBB->pred_empty()) { 785 RemoveDeadBlock(MBB); 786 MadeChange = true; 787 ++NumDeadBlocks; 788 } 789 } 790 return MadeChange; 791} 792 793 794/// CanFallThrough - Return true if the specified block (with the specified 795/// branch condition) can implicitly transfer control to the block after it by 796/// falling off the end of it. This should return false if it can reach the 797/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 798/// 799/// True is a conservative answer. 800/// 801bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 802 bool BranchUnAnalyzable, 803 MachineBasicBlock *TBB, 804 MachineBasicBlock *FBB, 805 const SmallVectorImpl<MachineOperand> &Cond) { 806 MachineFunction::iterator Fallthrough = CurBB; 807 ++Fallthrough; 808 // If FallthroughBlock is off the end of the function, it can't fall through. 809 if (Fallthrough == CurBB->getParent()->end()) 810 return false; 811 812 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 813 if (!CurBB->isSuccessor(Fallthrough)) 814 return false; 815 816 // If we couldn't analyze the branch, assume it could fall through. 817 if (BranchUnAnalyzable) return true; 818 819 // If there is no branch, control always falls through. 820 if (TBB == 0) return true; 821 822 // If there is some explicit branch to the fallthrough block, it can obviously 823 // reach, even though the branch should get folded to fall through implicitly. 824 if (MachineFunction::iterator(TBB) == Fallthrough || 825 MachineFunction::iterator(FBB) == Fallthrough) 826 return true; 827 828 // If it's an unconditional branch to some block not the fall through, it 829 // doesn't fall through. 830 if (Cond.empty()) return false; 831 832 // Otherwise, if it is conditional and has no explicit false block, it falls 833 // through. 834 return FBB == 0; 835} 836 837/// CanFallThrough - Return true if the specified can implicitly transfer 838/// control to the block after it by falling off the end of it. This should 839/// return false if it can reach the block after it, but it uses an explicit 840/// branch to do so (e.g. a table jump). 841/// 842/// True is a conservative answer. 843/// 844bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 845 MachineBasicBlock *TBB = 0, *FBB = 0; 846 SmallVector<MachineOperand, 4> Cond; 847 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true); 848 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 849} 850 851/// IsBetterFallthrough - Return true if it would be clearly better to 852/// fall-through to MBB1 than to fall through into MBB2. This has to return 853/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 854/// result in infinite loops. 855static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 856 MachineBasicBlock *MBB2) { 857 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 858 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 859 // optimize branches that branch to either a return block or an assert block 860 // into a fallthrough to the return. 861 if (MBB1->empty() || MBB2->empty()) return false; 862 863 // If there is a clear successor ordering we make sure that one block 864 // will fall through to the next 865 if (MBB1->isSuccessor(MBB2)) return true; 866 if (MBB2->isSuccessor(MBB1)) return false; 867 868 MachineInstr *MBB1I = --MBB1->end(); 869 MachineInstr *MBB2I = --MBB2->end(); 870 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 871} 872 873/// OptimizeBlock - Analyze and optimize control flow related to the specified 874/// block. This is never called on the entry block. 875bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 876 bool MadeChange = false; 877 MachineFunction &MF = *MBB->getParent(); 878 879 MachineFunction::iterator FallThrough = MBB; 880 ++FallThrough; 881 882 // If this block is empty, make everyone use its fall-through, not the block 883 // explicitly. Landing pads should not do this since the landing-pad table 884 // points to this block. Blocks with their addresses taken shouldn't be 885 // optimized away. 886 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 887 // Dead block? Leave for cleanup later. 888 if (MBB->pred_empty()) return MadeChange; 889 890 if (FallThrough == MF.end()) { 891 // TODO: Simplify preds to not branch here if possible! 892 } else { 893 // Rewrite all predecessors of the old block to go to the fallthrough 894 // instead. 895 while (!MBB->pred_empty()) { 896 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 897 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 898 } 899 // If MBB was the target of a jump table, update jump tables to go to the 900 // fallthrough instead. 901 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough); 902 MadeChange = true; 903 } 904 return MadeChange; 905 } 906 907 // Check to see if we can simplify the terminator of the block before this 908 // one. 909 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 910 911 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 912 SmallVector<MachineOperand, 4> PriorCond; 913 bool PriorUnAnalyzable = 914 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 915 if (!PriorUnAnalyzable) { 916 // If the CFG for the prior block has extra edges, remove them. 917 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 918 !PriorCond.empty()); 919 920 // If the previous branch is conditional and both conditions go to the same 921 // destination, remove the branch, replacing it with an unconditional one or 922 // a fall-through. 923 if (PriorTBB && PriorTBB == PriorFBB) { 924 TII->RemoveBranch(PrevBB); 925 PriorCond.clear(); 926 if (PriorTBB != MBB) 927 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 928 MadeChange = true; 929 ++NumBranchOpts; 930 return OptimizeBlock(MBB); 931 } 932 933 // If the previous branch *only* branches to *this* block (conditional or 934 // not) remove the branch. 935 if (PriorTBB == MBB && PriorFBB == 0) { 936 TII->RemoveBranch(PrevBB); 937 MadeChange = true; 938 ++NumBranchOpts; 939 return OptimizeBlock(MBB); 940 } 941 942 // If the prior block branches somewhere else on the condition and here if 943 // the condition is false, remove the uncond second branch. 944 if (PriorFBB == MBB) { 945 TII->RemoveBranch(PrevBB); 946 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 947 MadeChange = true; 948 ++NumBranchOpts; 949 return OptimizeBlock(MBB); 950 } 951 952 // If the prior block branches here on true and somewhere else on false, and 953 // if the branch condition is reversible, reverse the branch to create a 954 // fall-through. 955 if (PriorTBB == MBB) { 956 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 957 if (!TII->ReverseBranchCondition(NewPriorCond)) { 958 TII->RemoveBranch(PrevBB); 959 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 960 MadeChange = true; 961 ++NumBranchOpts; 962 return OptimizeBlock(MBB); 963 } 964 } 965 966 // If this block has no successors (e.g. it is a return block or ends with 967 // a call to a no-return function like abort or __cxa_throw) and if the pred 968 // falls through into this block, and if it would otherwise fall through 969 // into the block after this, move this block to the end of the function. 970 // 971 // We consider it more likely that execution will stay in the function (e.g. 972 // due to loops) than it is to exit it. This asserts in loops etc, moving 973 // the assert condition out of the loop body. 974 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 975 MachineFunction::iterator(PriorTBB) == FallThrough && 976 !CanFallThrough(MBB)) { 977 bool DoTransform = true; 978 979 // We have to be careful that the succs of PredBB aren't both no-successor 980 // blocks. If neither have successors and if PredBB is the second from 981 // last block in the function, we'd just keep swapping the two blocks for 982 // last. Only do the swap if one is clearly better to fall through than 983 // the other. 984 if (FallThrough == --MF.end() && 985 !IsBetterFallthrough(PriorTBB, MBB)) 986 DoTransform = false; 987 988 // We don't want to do this transformation if we have control flow like: 989 // br cond BB2 990 // BB1: 991 // .. 992 // jmp BBX 993 // BB2: 994 // .. 995 // ret 996 // 997 // In this case, we could actually be moving the return block *into* a 998 // loop! 999 if (DoTransform && !MBB->succ_empty() && 1000 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 1001 DoTransform = false; 1002 1003 1004 if (DoTransform) { 1005 // Reverse the branch so we will fall through on the previous true cond. 1006 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1007 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1008 DEBUG(errs() << "\nMoving MBB: " << *MBB 1009 << "To make fallthrough to: " << *PriorTBB << "\n"); 1010 1011 TII->RemoveBranch(PrevBB); 1012 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 1013 1014 // Move this block to the end of the function. 1015 MBB->moveAfter(--MF.end()); 1016 MadeChange = true; 1017 ++NumBranchOpts; 1018 return MadeChange; 1019 } 1020 } 1021 } 1022 } 1023 1024 // Analyze the branch in the current block. 1025 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1026 SmallVector<MachineOperand, 4> CurCond; 1027 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1028 if (!CurUnAnalyzable) { 1029 // If the CFG for the prior block has extra edges, remove them. 1030 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1031 1032 // If this is a two-way branch, and the FBB branches to this block, reverse 1033 // the condition so the single-basic-block loop is faster. Instead of: 1034 // Loop: xxx; jcc Out; jmp Loop 1035 // we want: 1036 // Loop: xxx; jncc Loop; jmp Out 1037 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1038 SmallVector<MachineOperand, 4> NewCond(CurCond); 1039 if (!TII->ReverseBranchCondition(NewCond)) { 1040 TII->RemoveBranch(*MBB); 1041 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 1042 MadeChange = true; 1043 ++NumBranchOpts; 1044 return OptimizeBlock(MBB); 1045 } 1046 } 1047 1048 1049 // If this branch is the only thing in its block, see if we can forward 1050 // other blocks across it. 1051 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1052 MBB->begin()->getDesc().isBranch() && CurTBB != MBB && 1053 !MBB->hasAddressTaken()) { 1054 // This block may contain just an unconditional branch. Because there can 1055 // be 'non-branch terminators' in the block, try removing the branch and 1056 // then seeing if the block is empty. 1057 TII->RemoveBranch(*MBB); 1058 1059 // If this block is just an unconditional branch to CurTBB, we can 1060 // usually completely eliminate the block. The only case we cannot 1061 // completely eliminate the block is when the block before this one 1062 // falls through into MBB and we can't understand the prior block's branch 1063 // condition. 1064 if (MBB->empty()) { 1065 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 1066 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1067 !PrevBB.isSuccessor(MBB)) { 1068 // If the prior block falls through into us, turn it into an 1069 // explicit branch to us to make updates simpler. 1070 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1071 PriorTBB != MBB && PriorFBB != MBB) { 1072 if (PriorTBB == 0) { 1073 assert(PriorCond.empty() && PriorFBB == 0 && 1074 "Bad branch analysis"); 1075 PriorTBB = MBB; 1076 } else { 1077 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1078 PriorFBB = MBB; 1079 } 1080 TII->RemoveBranch(PrevBB); 1081 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1082 } 1083 1084 // Iterate through all the predecessors, revectoring each in-turn. 1085 size_t PI = 0; 1086 bool DidChange = false; 1087 bool HasBranchToSelf = false; 1088 while(PI != MBB->pred_size()) { 1089 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1090 if (PMBB == MBB) { 1091 // If this block has an uncond branch to itself, leave it. 1092 ++PI; 1093 HasBranchToSelf = true; 1094 } else { 1095 DidChange = true; 1096 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1097 // If this change resulted in PMBB ending in a conditional 1098 // branch where both conditions go to the same destination, 1099 // change this to an unconditional branch (and fix the CFG). 1100 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1101 SmallVector<MachineOperand, 4> NewCurCond; 1102 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1103 NewCurFBB, NewCurCond, true); 1104 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1105 TII->RemoveBranch(*PMBB); 1106 NewCurCond.clear(); 1107 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond); 1108 MadeChange = true; 1109 ++NumBranchOpts; 1110 PMBB->CorrectExtraCFGEdges(NewCurTBB, NewCurFBB, false); 1111 } 1112 } 1113 } 1114 1115 // Change any jumptables to go to the new MBB. 1116 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB); 1117 if (DidChange) { 1118 ++NumBranchOpts; 1119 MadeChange = true; 1120 if (!HasBranchToSelf) return MadeChange; 1121 } 1122 } 1123 } 1124 1125 // Add the branch back if the block is more than just an uncond branch. 1126 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1127 } 1128 } 1129 1130 // If the prior block doesn't fall through into this block, and if this 1131 // block doesn't fall through into some other block, see if we can find a 1132 // place to move this block where a fall-through will happen. 1133 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 1134 PriorTBB, PriorFBB, PriorCond)) { 1135 // Now we know that there was no fall-through into this block, check to 1136 // see if it has a fall-through into its successor. 1137 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1138 CurCond); 1139 1140 if (!MBB->isLandingPad()) { 1141 // Check all the predecessors of this block. If one of them has no fall 1142 // throughs, move this block right after it. 1143 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1144 E = MBB->pred_end(); PI != E; ++PI) { 1145 // Analyze the branch at the end of the pred. 1146 MachineBasicBlock *PredBB = *PI; 1147 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1148 if (PredBB != MBB && !CanFallThrough(PredBB) 1149 && (!CurFallsThru || !CurTBB || !CurFBB) 1150 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1151 // If the current block doesn't fall through, just move it. 1152 // If the current block can fall through and does not end with a 1153 // conditional branch, we need to append an unconditional jump to 1154 // the (current) next block. To avoid a possible compile-time 1155 // infinite loop, move blocks only backward in this case. 1156 // Also, if there are already 2 branches here, we cannot add a third; 1157 // this means we have the case 1158 // Bcc next 1159 // B elsewhere 1160 // next: 1161 if (CurFallsThru) { 1162 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1163 CurCond.clear(); 1164 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1165 } 1166 MBB->moveAfter(PredBB); 1167 MadeChange = true; 1168 return OptimizeBlock(MBB); 1169 } 1170 } 1171 } 1172 1173 if (!CurFallsThru) { 1174 // Check all successors to see if we can move this block before it. 1175 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1176 E = MBB->succ_end(); SI != E; ++SI) { 1177 // Analyze the branch at the end of the block before the succ. 1178 MachineBasicBlock *SuccBB = *SI; 1179 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1180 1181 // If this block doesn't already fall-through to that successor, and if 1182 // the succ doesn't already have a block that can fall through into it, 1183 // and if the successor isn't an EH destination, we can arrange for the 1184 // fallthrough to happen. 1185 if (SuccBB != MBB && !CanFallThrough(SuccPrev) && 1186 !SuccBB->isLandingPad()) { 1187 MBB->moveBefore(SuccBB); 1188 MadeChange = true; 1189 return OptimizeBlock(MBB); 1190 } 1191 } 1192 1193 // Okay, there is no really great place to put this block. If, however, 1194 // the block before this one would be a fall-through if this block were 1195 // removed, move this block to the end of the function. 1196 if (FallThrough != MF.end() && 1197 PrevBB.isSuccessor(FallThrough)) { 1198 MBB->moveAfter(--MF.end()); 1199 MadeChange = true; 1200 return MadeChange; 1201 } 1202 } 1203 } 1204 1205 return MadeChange; 1206} 1207