BranchFolding.cpp revision 465e2b950d61c870fb3120c80191973e8282a3ef
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/SetVector.h" 36#include "llvm/ADT/Statistic.h" 37#include "llvm/ADT/STLExtras.h" 38#include <algorithm> 39using namespace llvm; 40 41STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 42STATISTIC(NumBranchOpts, "Number of branches optimized"); 43STATISTIC(NumTailMerge , "Number of block tails merged"); 44 45static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 46 cl::init(cl::BOU_UNSET), cl::Hidden); 47 48// Throttle for huge numbers of predecessors (compile speed problems) 49static cl::opt<unsigned> 50TailMergeThreshold("tail-merge-threshold", 51 cl::desc("Max number of predecessors to consider tail merging"), 52 cl::init(150), cl::Hidden); 53 54// Heuristic for tail merging (and, inversely, tail duplication). 55// TODO: This should be replaced with a target query. 56static cl::opt<unsigned> 57TailMergeSize("tail-merge-size", 58 cl::desc("Min number of instructions to consider tail merging"), 59 cl::init(3), cl::Hidden); 60 61namespace { 62 /// BranchFolderPass - Wrap branch folder in a machine function pass. 63 class BranchFolderPass : public MachineFunctionPass, 64 public BranchFolder { 65 public: 66 static char ID; 67 explicit BranchFolderPass(bool defaultEnableTailMerge) 68 : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {} 69 70 virtual bool runOnMachineFunction(MachineFunction &MF); 71 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 72 }; 73} 74 75char BranchFolderPass::ID = 0; 76 77FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 78 return new BranchFolderPass(DefaultEnableTailMerge); 79} 80 81bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { 82 return OptimizeFunction(MF, 83 MF.getTarget().getInstrInfo(), 84 MF.getTarget().getRegisterInfo(), 85 getAnalysisIfAvailable<MachineModuleInfo>()); 86} 87 88 89BranchFolder::BranchFolder(bool defaultEnableTailMerge) { 90 switch (FlagEnableTailMerge) { 91 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 92 case cl::BOU_TRUE: EnableTailMerge = true; break; 93 case cl::BOU_FALSE: EnableTailMerge = false; break; 94 } 95} 96 97/// RemoveDeadBlock - Remove the specified dead machine basic block from the 98/// function, updating the CFG. 99void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 100 assert(MBB->pred_empty() && "MBB must be dead!"); 101 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB); 102 103 MachineFunction *MF = MBB->getParent(); 104 // drop all successors. 105 while (!MBB->succ_empty()) 106 MBB->removeSuccessor(MBB->succ_end()-1); 107 108 // If there are any labels in the basic block, unregister them from 109 // MachineModuleInfo. 110 if (MMI && !MBB->empty()) { 111 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 112 I != E; ++I) { 113 if (I->isLabel()) 114 // The label ID # is always operand #0, an immediate. 115 MMI->InvalidateLabel(I->getOperand(0).getImm()); 116 } 117 } 118 119 // Remove the block. 120 MF->erase(MBB); 121} 122 123/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 124/// followed by terminators, and if the implicitly defined registers are not 125/// used by the terminators, remove those implicit_def's. e.g. 126/// BB1: 127/// r0 = implicit_def 128/// r1 = implicit_def 129/// br 130/// This block can be optimized away later if the implicit instructions are 131/// removed. 132bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 133 SmallSet<unsigned, 4> ImpDefRegs; 134 MachineBasicBlock::iterator I = MBB->begin(); 135 while (I != MBB->end()) { 136 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) 137 break; 138 unsigned Reg = I->getOperand(0).getReg(); 139 ImpDefRegs.insert(Reg); 140 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg); 141 unsigned SubReg = *SubRegs; ++SubRegs) 142 ImpDefRegs.insert(SubReg); 143 ++I; 144 } 145 if (ImpDefRegs.empty()) 146 return false; 147 148 MachineBasicBlock::iterator FirstTerm = I; 149 while (I != MBB->end()) { 150 if (!TII->isUnpredicatedTerminator(I)) 151 return false; 152 // See if it uses any of the implicitly defined registers. 153 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 154 MachineOperand &MO = I->getOperand(i); 155 if (!MO.isReg() || !MO.isUse()) 156 continue; 157 unsigned Reg = MO.getReg(); 158 if (ImpDefRegs.count(Reg)) 159 return false; 160 } 161 ++I; 162 } 163 164 I = MBB->begin(); 165 while (I != FirstTerm) { 166 MachineInstr *ImpDefMI = &*I; 167 ++I; 168 MBB->erase(ImpDefMI); 169 } 170 171 return true; 172} 173 174/// OptimizeFunction - Perhaps branch folding, tail merging and other 175/// CFG optimizations on the given function. 176bool BranchFolder::OptimizeFunction(MachineFunction &MF, 177 const TargetInstrInfo *tii, 178 const TargetRegisterInfo *tri, 179 MachineModuleInfo *mmi) { 180 if (!tii) return false; 181 182 TII = tii; 183 TRI = tri; 184 MMI = mmi; 185 186 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 187 188 // Fix CFG. The later algorithms expect it to be right. 189 bool MadeChange = false; 190 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 191 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 192 SmallVector<MachineOperand, 4> Cond; 193 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 194 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 195 MadeChange |= OptimizeImpDefsBlock(MBB); 196 } 197 198 bool MadeChangeThisIteration = true; 199 while (MadeChangeThisIteration) { 200 MadeChangeThisIteration = false; 201 MadeChangeThisIteration |= TailMergeBlocks(MF); 202 MadeChangeThisIteration |= OptimizeBranches(MF); 203 MadeChange |= MadeChangeThisIteration; 204 } 205 206 // See if any jump tables have become mergable or dead as the code generator 207 // did its thing. 208 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 209 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 210 if (!JTs.empty()) { 211 // Figure out how these jump tables should be merged. 212 std::vector<unsigned> JTMapping; 213 JTMapping.reserve(JTs.size()); 214 215 // We always keep the 0th jump table. 216 JTMapping.push_back(0); 217 218 // Scan the jump tables, seeing if there are any duplicates. Note that this 219 // is N^2, which should be fixed someday. 220 for (unsigned i = 1, e = JTs.size(); i != e; ++i) { 221 if (JTs[i].MBBs.empty()) 222 JTMapping.push_back(i); 223 else 224 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 225 } 226 227 // If a jump table was merge with another one, walk the function rewriting 228 // references to jump tables to reference the new JT ID's. Keep track of 229 // whether we see a jump table idx, if not, we can delete the JT. 230 BitVector JTIsLive(JTs.size()); 231 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 232 BB != E; ++BB) { 233 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 234 I != E; ++I) 235 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 236 MachineOperand &Op = I->getOperand(op); 237 if (!Op.isJTI()) continue; 238 unsigned NewIdx = JTMapping[Op.getIndex()]; 239 Op.setIndex(NewIdx); 240 241 // Remember that this JT is live. 242 JTIsLive.set(NewIdx); 243 } 244 } 245 246 // Finally, remove dead jump tables. This happens either because the 247 // indirect jump was unreachable (and thus deleted) or because the jump 248 // table was merged with some other one. 249 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 250 if (!JTIsLive.test(i)) { 251 JTI->RemoveJumpTable(i); 252 MadeChange = true; 253 } 254 } 255 256 delete RS; 257 return MadeChange; 258} 259 260//===----------------------------------------------------------------------===// 261// Tail Merging of Blocks 262//===----------------------------------------------------------------------===// 263 264/// HashMachineInstr - Compute a hash value for MI and its operands. 265static unsigned HashMachineInstr(const MachineInstr *MI) { 266 unsigned Hash = MI->getOpcode(); 267 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 268 const MachineOperand &Op = MI->getOperand(i); 269 270 // Merge in bits from the operand if easy. 271 unsigned OperandHash = 0; 272 switch (Op.getType()) { 273 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 274 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 275 case MachineOperand::MO_MachineBasicBlock: 276 OperandHash = Op.getMBB()->getNumber(); 277 break; 278 case MachineOperand::MO_FrameIndex: 279 case MachineOperand::MO_ConstantPoolIndex: 280 case MachineOperand::MO_JumpTableIndex: 281 OperandHash = Op.getIndex(); 282 break; 283 case MachineOperand::MO_GlobalAddress: 284 case MachineOperand::MO_ExternalSymbol: 285 // Global address / external symbol are too hard, don't bother, but do 286 // pull in the offset. 287 OperandHash = Op.getOffset(); 288 break; 289 default: break; 290 } 291 292 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 293 } 294 return Hash; 295} 296 297/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 298/// with no successors, we hash two instructions, because cross-jumping 299/// only saves code when at least two instructions are removed (since a 300/// branch must be inserted). For blocks with a successor, one of the 301/// two blocks to be tail-merged will end with a branch already, so 302/// it gains to cross-jump even for one instruction. 303static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 304 unsigned minCommonTailLength) { 305 MachineBasicBlock::const_iterator I = MBB->end(); 306 if (I == MBB->begin()) 307 return 0; // Empty MBB. 308 309 --I; 310 unsigned Hash = HashMachineInstr(I); 311 312 if (I == MBB->begin() || minCommonTailLength == 1) 313 return Hash; // Single instr MBB. 314 315 --I; 316 // Hash in the second-to-last instruction. 317 Hash ^= HashMachineInstr(I) << 2; 318 return Hash; 319} 320 321/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 322/// of instructions they actually have in common together at their end. Return 323/// iterators for the first shared instruction in each block. 324static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 325 MachineBasicBlock *MBB2, 326 MachineBasicBlock::iterator &I1, 327 MachineBasicBlock::iterator &I2) { 328 I1 = MBB1->end(); 329 I2 = MBB2->end(); 330 331 unsigned TailLen = 0; 332 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 333 --I1; --I2; 334 if (!I1->isIdenticalTo(I2) || 335 // FIXME: This check is dubious. It's used to get around a problem where 336 // people incorrectly expect inline asm directives to remain in the same 337 // relative order. This is untenable because normal compiler 338 // optimizations (like this one) may reorder and/or merge these 339 // directives. 340 I1->getOpcode() == TargetInstrInfo::INLINEASM) { 341 ++I1; ++I2; 342 break; 343 } 344 ++TailLen; 345 } 346 return TailLen; 347} 348 349/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 350/// after it, replacing it with an unconditional branch to NewDest. This 351/// returns true if OldInst's block is modified, false if NewDest is modified. 352void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 353 MachineBasicBlock *NewDest) { 354 MachineBasicBlock *OldBB = OldInst->getParent(); 355 356 // Remove all the old successors of OldBB from the CFG. 357 while (!OldBB->succ_empty()) 358 OldBB->removeSuccessor(OldBB->succ_begin()); 359 360 // Remove all the dead instructions from the end of OldBB. 361 OldBB->erase(OldInst, OldBB->end()); 362 363 // If OldBB isn't immediately before OldBB, insert a branch to it. 364 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 365 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>()); 366 OldBB->addSuccessor(NewDest); 367 ++NumTailMerge; 368} 369 370/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 371/// MBB so that the part before the iterator falls into the part starting at the 372/// iterator. This returns the new MBB. 373MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 374 MachineBasicBlock::iterator BBI1) { 375 MachineFunction &MF = *CurMBB.getParent(); 376 377 // Create the fall-through block. 378 MachineFunction::iterator MBBI = &CurMBB; 379 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 380 CurMBB.getParent()->insert(++MBBI, NewMBB); 381 382 // Move all the successors of this block to the specified block. 383 NewMBB->transferSuccessors(&CurMBB); 384 385 // Add an edge from CurMBB to NewMBB for the fall-through. 386 CurMBB.addSuccessor(NewMBB); 387 388 // Splice the code over. 389 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 390 391 // For targets that use the register scavenger, we must maintain LiveIns. 392 if (RS) { 393 RS->enterBasicBlock(&CurMBB); 394 if (!CurMBB.empty()) 395 RS->forward(prior(CurMBB.end())); 396 BitVector RegsLiveAtExit(TRI->getNumRegs()); 397 RS->getRegsUsed(RegsLiveAtExit, false); 398 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 399 if (RegsLiveAtExit[i]) 400 NewMBB->addLiveIn(i); 401 } 402 403 return NewMBB; 404} 405 406/// EstimateRuntime - Make a rough estimate for how long it will take to run 407/// the specified code. 408static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 409 MachineBasicBlock::iterator E) { 410 unsigned Time = 0; 411 for (; I != E; ++I) { 412 const TargetInstrDesc &TID = I->getDesc(); 413 if (TID.isCall()) 414 Time += 10; 415 else if (TID.mayLoad() || TID.mayStore()) 416 Time += 2; 417 else 418 ++Time; 419 } 420 return Time; 421} 422 423// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 424// branches temporarily for tail merging). In the case where CurMBB ends 425// with a conditional branch to the next block, optimize by reversing the 426// test and conditionally branching to SuccMBB instead. 427static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 428 const TargetInstrInfo *TII) { 429 MachineFunction *MF = CurMBB->getParent(); 430 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB)); 431 MachineBasicBlock *TBB = 0, *FBB = 0; 432 SmallVector<MachineOperand, 4> Cond; 433 if (I != MF->end() && 434 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 435 MachineBasicBlock *NextBB = I; 436 if (TBB == NextBB && !Cond.empty() && !FBB) { 437 if (!TII->ReverseBranchCondition(Cond)) { 438 TII->RemoveBranch(*CurMBB); 439 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 440 return; 441 } 442 } 443 } 444 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>()); 445} 446 447bool 448BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 449 if (getHash() < o.getHash()) 450 return true; 451 else if (getHash() > o.getHash()) 452 return false; 453 else if (getBlock()->getNumber() < o.getBlock()->getNumber()) 454 return true; 455 else if (getBlock()->getNumber() > o.getBlock()->getNumber()) 456 return false; 457 else { 458 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 459 // an object with itself. 460#ifndef _GLIBCXX_DEBUG 461 llvm_unreachable("Predecessor appears twice"); 462#endif 463 return false; 464 } 465} 466 467/// CountTerminators - Count the number of terminators in the given 468/// block and set I to the position of the first non-terminator, if there 469/// is one, or MBB->end() otherwise. 470static unsigned CountTerminators(MachineBasicBlock *MBB, 471 MachineBasicBlock::iterator &I) { 472 I = MBB->end(); 473 unsigned NumTerms = 0; 474 for (;;) { 475 if (I == MBB->begin()) { 476 I = MBB->end(); 477 break; 478 } 479 --I; 480 if (!I->getDesc().isTerminator()) break; 481 ++NumTerms; 482 } 483 return NumTerms; 484} 485 486/// ProfitableToMerge - Check if two machine basic blocks have a common tail 487/// and decide if it would be profitable to merge those tails. Return the 488/// length of the common tail and iterators to the first common instruction 489/// in each block. 490static bool ProfitableToMerge(MachineBasicBlock *MBB1, 491 MachineBasicBlock *MBB2, 492 unsigned minCommonTailLength, 493 unsigned &CommonTailLen, 494 MachineBasicBlock::iterator &I1, 495 MachineBasicBlock::iterator &I2, 496 MachineBasicBlock *SuccBB, 497 MachineBasicBlock *PredBB) { 498 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 499 MachineFunction *MF = MBB1->getParent(); 500 501 if (CommonTailLen == 0) 502 return false; 503 504 // It's almost always profitable to merge any number of non-terminator 505 // instructions with the block that falls through into the common successor. 506 if (MBB1 == PredBB || MBB2 == PredBB) { 507 MachineBasicBlock::iterator I; 508 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 509 if (CommonTailLen > NumTerms) 510 return true; 511 } 512 513 // If one of the blocks can be completely merged and happens to be in 514 // a position where the other could fall through into it, merge any number 515 // of instructions, because it can be done without a branch. 516 // TODO: If the blocks are not adjacent, move one of them so that they are? 517 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 518 return true; 519 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 520 return true; 521 522 // If both blocks have an unconditional branch temporarily stripped out, 523 // count that as an additional common instruction for the following 524 // heuristics. 525 unsigned EffectiveTailLen = CommonTailLen; 526 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 527 !MBB1->back().getDesc().isBarrier() && 528 !MBB2->back().getDesc().isBarrier()) 529 ++EffectiveTailLen; 530 531 // Check if the common tail is long enough to be worthwhile. 532 if (EffectiveTailLen >= minCommonTailLength) 533 return true; 534 535 // If we are optimizing for code size, 2 instructions in common is enough if 536 // we don't have to split a block. At worst we will be introducing 1 new 537 // branch instruction, which is likely to be smaller than the 2 538 // instructions that would be deleted in the merge. 539 if (EffectiveTailLen >= 2 && 540 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 541 (I1 == MBB1->begin() || I2 == MBB2->begin())) 542 return true; 543 544 return false; 545} 546 547/// ComputeSameTails - Look through all the blocks in MergePotentials that have 548/// hash CurHash (guaranteed to match the last element). Build the vector 549/// SameTails of all those that have the (same) largest number of instructions 550/// in common of any pair of these blocks. SameTails entries contain an 551/// iterator into MergePotentials (from which the MachineBasicBlock can be 552/// found) and a MachineBasicBlock::iterator into that MBB indicating the 553/// instruction where the matching code sequence begins. 554/// Order of elements in SameTails is the reverse of the order in which 555/// those blocks appear in MergePotentials (where they are not necessarily 556/// consecutive). 557unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 558 unsigned minCommonTailLength, 559 MachineBasicBlock *SuccBB, 560 MachineBasicBlock *PredBB) { 561 unsigned maxCommonTailLength = 0U; 562 SameTails.clear(); 563 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 564 MPIterator HighestMPIter = prior(MergePotentials.end()); 565 for (MPIterator CurMPIter = prior(MergePotentials.end()), 566 B = MergePotentials.begin(); 567 CurMPIter != B && CurMPIter->getHash() == CurHash; 568 --CurMPIter) { 569 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 570 unsigned CommonTailLen; 571 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 572 minCommonTailLength, 573 CommonTailLen, TrialBBI1, TrialBBI2, 574 SuccBB, PredBB)) { 575 if (CommonTailLen > maxCommonTailLength) { 576 SameTails.clear(); 577 maxCommonTailLength = CommonTailLen; 578 HighestMPIter = CurMPIter; 579 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 580 } 581 if (HighestMPIter == CurMPIter && 582 CommonTailLen == maxCommonTailLength) 583 SameTails.push_back(SameTailElt(I, TrialBBI2)); 584 } 585 if (I == B) 586 break; 587 } 588 } 589 return maxCommonTailLength; 590} 591 592/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 593/// MergePotentials, restoring branches at ends of blocks as appropriate. 594void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 595 MachineBasicBlock *SuccBB, 596 MachineBasicBlock *PredBB) { 597 MPIterator CurMPIter, B; 598 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 599 CurMPIter->getHash() == CurHash; 600 --CurMPIter) { 601 // Put the unconditional branch back, if we need one. 602 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 603 if (SuccBB && CurMBB != PredBB) 604 FixTail(CurMBB, SuccBB, TII); 605 if (CurMPIter == B) 606 break; 607 } 608 if (CurMPIter->getHash() != CurHash) 609 CurMPIter++; 610 MergePotentials.erase(CurMPIter, MergePotentials.end()); 611} 612 613/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 614/// only of the common tail. Create a block that does by splitting one. 615unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 616 unsigned maxCommonTailLength) { 617 unsigned commonTailIndex = 0; 618 unsigned TimeEstimate = ~0U; 619 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 620 // Use PredBB if possible; that doesn't require a new branch. 621 if (SameTails[i].getBlock() == PredBB) { 622 commonTailIndex = i; 623 break; 624 } 625 // Otherwise, make a (fairly bogus) choice based on estimate of 626 // how long it will take the various blocks to execute. 627 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 628 SameTails[i].getTailStartPos()); 629 if (t <= TimeEstimate) { 630 TimeEstimate = t; 631 commonTailIndex = i; 632 } 633 } 634 635 MachineBasicBlock::iterator BBI = 636 SameTails[commonTailIndex].getTailStartPos(); 637 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 638 639 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 640 << maxCommonTailLength); 641 642 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 643 SameTails[commonTailIndex].setBlock(newMBB); 644 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 645 646 // If we split PredBB, newMBB is the new predecessor. 647 if (PredBB == MBB) 648 PredBB = newMBB; 649 650 return commonTailIndex; 651} 652 653// See if any of the blocks in MergePotentials (which all have a common single 654// successor, or all have no successor) can be tail-merged. If there is a 655// successor, any blocks in MergePotentials that are not tail-merged and 656// are not immediately before Succ must have an unconditional branch to 657// Succ added (but the predecessor/successor lists need no adjustment). 658// The lone predecessor of Succ that falls through into Succ, 659// if any, is given in PredBB. 660 661bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 662 MachineBasicBlock *PredBB) { 663 bool MadeChange = false; 664 665 // Except for the special cases below, tail-merge if there are at least 666 // this many instructions in common. 667 unsigned minCommonTailLength = TailMergeSize; 668 669 DEBUG(dbgs() << "\nTryTailMergeBlocks: "; 670 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 671 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 672 << (i == e-1 ? "" : ", "); 673 dbgs() << "\n"; 674 if (SuccBB) { 675 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 676 if (PredBB) 677 dbgs() << " which has fall-through from BB#" 678 << PredBB->getNumber() << "\n"; 679 } 680 dbgs() << "Looking for common tails of at least " 681 << minCommonTailLength << " instruction" 682 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 683 ); 684 685 // Sort by hash value so that blocks with identical end sequences sort 686 // together. 687 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 688 689 // Walk through equivalence sets looking for actual exact matches. 690 while (MergePotentials.size() > 1) { 691 unsigned CurHash = MergePotentials.back().getHash(); 692 693 // Build SameTails, identifying the set of blocks with this hash code 694 // and with the maximum number of instructions in common. 695 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 696 minCommonTailLength, 697 SuccBB, PredBB); 698 699 // If we didn't find any pair that has at least minCommonTailLength 700 // instructions in common, remove all blocks with this hash code and retry. 701 if (SameTails.empty()) { 702 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 703 continue; 704 } 705 706 // If one of the blocks is the entire common tail (and not the entry 707 // block, which we can't jump to), we can treat all blocks with this same 708 // tail at once. Use PredBB if that is one of the possibilities, as that 709 // will not introduce any extra branches. 710 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 711 getParent()->begin(); 712 unsigned commonTailIndex = SameTails.size(); 713 // If there are two blocks, check to see if one can be made to fall through 714 // into the other. 715 if (SameTails.size() == 2 && 716 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 717 SameTails[1].tailIsWholeBlock()) 718 commonTailIndex = 1; 719 else if (SameTails.size() == 2 && 720 SameTails[1].getBlock()->isLayoutSuccessor( 721 SameTails[0].getBlock()) && 722 SameTails[0].tailIsWholeBlock()) 723 commonTailIndex = 0; 724 else { 725 // Otherwise just pick one, favoring the fall-through predecessor if 726 // there is one. 727 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 728 MachineBasicBlock *MBB = SameTails[i].getBlock(); 729 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 730 continue; 731 if (MBB == PredBB) { 732 commonTailIndex = i; 733 break; 734 } 735 if (SameTails[i].tailIsWholeBlock()) 736 commonTailIndex = i; 737 } 738 } 739 740 if (commonTailIndex == SameTails.size() || 741 (SameTails[commonTailIndex].getBlock() == PredBB && 742 !SameTails[commonTailIndex].tailIsWholeBlock())) { 743 // None of the blocks consist entirely of the common tail. 744 // Split a block so that one does. 745 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); 746 } 747 748 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 749 // MBB is common tail. Adjust all other BB's to jump to this one. 750 // Traversal must be forwards so erases work. 751 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber() 752 << " for "); 753 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 754 if (commonTailIndex == i) 755 continue; 756 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() 757 << (i == e-1 ? "" : ", ")); 758 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 759 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 760 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 761 MergePotentials.erase(SameTails[i].getMPIter()); 762 } 763 DEBUG(dbgs() << "\n"); 764 // We leave commonTailIndex in the worklist in case there are other blocks 765 // that match it with a smaller number of instructions. 766 MadeChange = true; 767 } 768 return MadeChange; 769} 770 771bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 772 773 if (!EnableTailMerge) return false; 774 775 bool MadeChange = false; 776 777 // First find blocks with no successors. 778 MergePotentials.clear(); 779 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 780 if (I->succ_empty()) 781 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I)); 782 } 783 784 // See if we can do any tail merging on those. 785 if (MergePotentials.size() < TailMergeThreshold && 786 MergePotentials.size() >= 2) 787 MadeChange |= TryTailMergeBlocks(NULL, NULL); 788 789 // Look at blocks (IBB) with multiple predecessors (PBB). 790 // We change each predecessor to a canonical form, by 791 // (1) temporarily removing any unconditional branch from the predecessor 792 // to IBB, and 793 // (2) alter conditional branches so they branch to the other block 794 // not IBB; this may require adding back an unconditional branch to IBB 795 // later, where there wasn't one coming in. E.g. 796 // Bcc IBB 797 // fallthrough to QBB 798 // here becomes 799 // Bncc QBB 800 // with a conceptual B to IBB after that, which never actually exists. 801 // With those changes, we see whether the predecessors' tails match, 802 // and merge them if so. We change things out of canonical form and 803 // back to the way they were later in the process. (OptimizeBranches 804 // would undo some of this, but we can't use it, because we'd get into 805 // a compile-time infinite loop repeatedly doing and undoing the same 806 // transformations.) 807 808 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 809 I != E; ++I) { 810 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 811 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 812 MachineBasicBlock *IBB = I; 813 MachineBasicBlock *PredBB = prior(I); 814 MergePotentials.clear(); 815 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 816 E2 = I->pred_end(); 817 P != E2; ++P) { 818 MachineBasicBlock *PBB = *P; 819 // Skip blocks that loop to themselves, can't tail merge these. 820 if (PBB == IBB) 821 continue; 822 // Visit each predecessor only once. 823 if (!UniquePreds.insert(PBB)) 824 continue; 825 MachineBasicBlock *TBB = 0, *FBB = 0; 826 SmallVector<MachineOperand, 4> Cond; 827 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 828 // Failing case: IBB is the target of a cbr, and 829 // we cannot reverse the branch. 830 SmallVector<MachineOperand, 4> NewCond(Cond); 831 if (!Cond.empty() && TBB == IBB) { 832 if (TII->ReverseBranchCondition(NewCond)) 833 continue; 834 // This is the QBB case described above 835 if (!FBB) 836 FBB = llvm::next(MachineFunction::iterator(PBB)); 837 } 838 // Failing case: the only way IBB can be reached from PBB is via 839 // exception handling. Happens for landing pads. Would be nice 840 // to have a bit in the edge so we didn't have to do all this. 841 if (IBB->isLandingPad()) { 842 MachineFunction::iterator IP = PBB; IP++; 843 MachineBasicBlock *PredNextBB = NULL; 844 if (IP != MF.end()) 845 PredNextBB = IP; 846 if (TBB == NULL) { 847 if (IBB != PredNextBB) // fallthrough 848 continue; 849 } else if (FBB) { 850 if (TBB != IBB && FBB != IBB) // cbr then ubr 851 continue; 852 } else if (Cond.empty()) { 853 if (TBB != IBB) // ubr 854 continue; 855 } else { 856 if (TBB != IBB && IBB != PredNextBB) // cbr 857 continue; 858 } 859 } 860 // Remove the unconditional branch at the end, if any. 861 if (TBB && (Cond.empty() || FBB)) { 862 TII->RemoveBranch(*PBB); 863 if (!Cond.empty()) 864 // reinsert conditional branch only, for now 865 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond); 866 } 867 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U), 868 *P)); 869 } 870 } 871 if (MergePotentials.size() >= 2) 872 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 873 // Reinsert an unconditional branch if needed. 874 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks. 875 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 876 if (MergePotentials.size() == 1 && 877 MergePotentials.begin()->getBlock() != PredBB) 878 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 879 } 880 } 881 return MadeChange; 882} 883 884//===----------------------------------------------------------------------===// 885// Branch Optimization 886//===----------------------------------------------------------------------===// 887 888bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 889 bool MadeChange = false; 890 891 // Make sure blocks are numbered in order 892 MF.RenumberBlocks(); 893 894 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 895 MachineBasicBlock *MBB = I++; 896 MadeChange |= OptimizeBlock(MBB); 897 898 // If it is dead, remove it. 899 if (MBB->pred_empty()) { 900 RemoveDeadBlock(MBB); 901 MadeChange = true; 902 ++NumDeadBlocks; 903 } 904 } 905 return MadeChange; 906} 907 908 909/// IsBetterFallthrough - Return true if it would be clearly better to 910/// fall-through to MBB1 than to fall through into MBB2. This has to return 911/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 912/// result in infinite loops. 913static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 914 MachineBasicBlock *MBB2) { 915 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 916 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 917 // optimize branches that branch to either a return block or an assert block 918 // into a fallthrough to the return. 919 if (MBB1->empty() || MBB2->empty()) return false; 920 921 // If there is a clear successor ordering we make sure that one block 922 // will fall through to the next 923 if (MBB1->isSuccessor(MBB2)) return true; 924 if (MBB2->isSuccessor(MBB1)) return false; 925 926 MachineInstr *MBB1I = --MBB1->end(); 927 MachineInstr *MBB2I = --MBB2->end(); 928 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 929} 930 931/// OptimizeBlock - Analyze and optimize control flow related to the specified 932/// block. This is never called on the entry block. 933bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 934 bool MadeChange = false; 935 MachineFunction &MF = *MBB->getParent(); 936ReoptimizeBlock: 937 938 MachineFunction::iterator FallThrough = MBB; 939 ++FallThrough; 940 941 // If this block is empty, make everyone use its fall-through, not the block 942 // explicitly. Landing pads should not do this since the landing-pad table 943 // points to this block. Blocks with their addresses taken shouldn't be 944 // optimized away. 945 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 946 // Dead block? Leave for cleanup later. 947 if (MBB->pred_empty()) return MadeChange; 948 949 if (FallThrough == MF.end()) { 950 // TODO: Simplify preds to not branch here if possible! 951 } else { 952 // Rewrite all predecessors of the old block to go to the fallthrough 953 // instead. 954 while (!MBB->pred_empty()) { 955 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 956 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 957 } 958 // If MBB was the target of a jump table, update jump tables to go to the 959 // fallthrough instead. 960 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough); 961 MadeChange = true; 962 } 963 return MadeChange; 964 } 965 966 // Check to see if we can simplify the terminator of the block before this 967 // one. 968 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 969 970 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 971 SmallVector<MachineOperand, 4> PriorCond; 972 bool PriorUnAnalyzable = 973 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 974 if (!PriorUnAnalyzable) { 975 // If the CFG for the prior block has extra edges, remove them. 976 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 977 !PriorCond.empty()); 978 979 // If the previous branch is conditional and both conditions go to the same 980 // destination, remove the branch, replacing it with an unconditional one or 981 // a fall-through. 982 if (PriorTBB && PriorTBB == PriorFBB) { 983 TII->RemoveBranch(PrevBB); 984 PriorCond.clear(); 985 if (PriorTBB != MBB) 986 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 987 MadeChange = true; 988 ++NumBranchOpts; 989 goto ReoptimizeBlock; 990 } 991 992 // If the previous block unconditionally falls through to this block and 993 // this block has no other predecessors, move the contents of this block 994 // into the prior block. This doesn't usually happen when SimplifyCFG 995 // has been used, but it can happen if tail merging splits a fall-through 996 // predecessor of a block. 997 // This has to check PrevBB->succ_size() because EH edges are ignored by 998 // AnalyzeBranch. 999 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1000 PrevBB.succ_size() == 1 && 1001 !MBB->hasAddressTaken()) { 1002 DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1003 << "From MBB: " << *MBB); 1004 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1005 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1006 assert(PrevBB.succ_empty()); 1007 PrevBB.transferSuccessors(MBB); 1008 MadeChange = true; 1009 return MadeChange; 1010 } 1011 1012 // If the previous branch *only* branches to *this* block (conditional or 1013 // not) remove the branch. 1014 if (PriorTBB == MBB && PriorFBB == 0) { 1015 TII->RemoveBranch(PrevBB); 1016 MadeChange = true; 1017 ++NumBranchOpts; 1018 goto ReoptimizeBlock; 1019 } 1020 1021 // If the prior block branches somewhere else on the condition and here if 1022 // the condition is false, remove the uncond second branch. 1023 if (PriorFBB == MBB) { 1024 TII->RemoveBranch(PrevBB); 1025 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 1026 MadeChange = true; 1027 ++NumBranchOpts; 1028 goto ReoptimizeBlock; 1029 } 1030 1031 // If the prior block branches here on true and somewhere else on false, and 1032 // if the branch condition is reversible, reverse the branch to create a 1033 // fall-through. 1034 if (PriorTBB == MBB) { 1035 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1036 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1037 TII->RemoveBranch(PrevBB); 1038 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 1039 MadeChange = true; 1040 ++NumBranchOpts; 1041 goto ReoptimizeBlock; 1042 } 1043 } 1044 1045 // If this block has no successors (e.g. it is a return block or ends with 1046 // a call to a no-return function like abort or __cxa_throw) and if the pred 1047 // falls through into this block, and if it would otherwise fall through 1048 // into the block after this, move this block to the end of the function. 1049 // 1050 // We consider it more likely that execution will stay in the function (e.g. 1051 // due to loops) than it is to exit it. This asserts in loops etc, moving 1052 // the assert condition out of the loop body. 1053 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1054 MachineFunction::iterator(PriorTBB) == FallThrough && 1055 !MBB->canFallThrough()) { 1056 bool DoTransform = true; 1057 1058 // We have to be careful that the succs of PredBB aren't both no-successor 1059 // blocks. If neither have successors and if PredBB is the second from 1060 // last block in the function, we'd just keep swapping the two blocks for 1061 // last. Only do the swap if one is clearly better to fall through than 1062 // the other. 1063 if (FallThrough == --MF.end() && 1064 !IsBetterFallthrough(PriorTBB, MBB)) 1065 DoTransform = false; 1066 1067 // We don't want to do this transformation if we have control flow like: 1068 // br cond BB2 1069 // BB1: 1070 // .. 1071 // jmp BBX 1072 // BB2: 1073 // .. 1074 // ret 1075 // 1076 // In this case, we could actually be moving the return block *into* a 1077 // loop! 1078 if (DoTransform && !MBB->succ_empty() && 1079 (!PriorTBB->canFallThrough() || PriorTBB->empty())) 1080 DoTransform = false; 1081 1082 1083 if (DoTransform) { 1084 // Reverse the branch so we will fall through on the previous true cond. 1085 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1086 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1087 DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1088 << "To make fallthrough to: " << *PriorTBB << "\n"); 1089 1090 TII->RemoveBranch(PrevBB); 1091 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 1092 1093 // Move this block to the end of the function. 1094 MBB->moveAfter(--MF.end()); 1095 MadeChange = true; 1096 ++NumBranchOpts; 1097 return MadeChange; 1098 } 1099 } 1100 } 1101 } 1102 1103 // Analyze the branch in the current block. 1104 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1105 SmallVector<MachineOperand, 4> CurCond; 1106 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1107 if (!CurUnAnalyzable) { 1108 // If the CFG for the prior block has extra edges, remove them. 1109 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1110 1111 // If this is a two-way branch, and the FBB branches to this block, reverse 1112 // the condition so the single-basic-block loop is faster. Instead of: 1113 // Loop: xxx; jcc Out; jmp Loop 1114 // we want: 1115 // Loop: xxx; jncc Loop; jmp Out 1116 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1117 SmallVector<MachineOperand, 4> NewCond(CurCond); 1118 if (!TII->ReverseBranchCondition(NewCond)) { 1119 TII->RemoveBranch(*MBB); 1120 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 1121 MadeChange = true; 1122 ++NumBranchOpts; 1123 goto ReoptimizeBlock; 1124 } 1125 } 1126 1127 // If this branch is the only thing in its block, see if we can forward 1128 // other blocks across it. 1129 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1130 MBB->begin()->getDesc().isBranch() && CurTBB != MBB && 1131 !MBB->hasAddressTaken()) { 1132 // This block may contain just an unconditional branch. Because there can 1133 // be 'non-branch terminators' in the block, try removing the branch and 1134 // then seeing if the block is empty. 1135 TII->RemoveBranch(*MBB); 1136 1137 // If this block is just an unconditional branch to CurTBB, we can 1138 // usually completely eliminate the block. The only case we cannot 1139 // completely eliminate the block is when the block before this one 1140 // falls through into MBB and we can't understand the prior block's branch 1141 // condition. 1142 if (MBB->empty()) { 1143 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1144 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1145 !PrevBB.isSuccessor(MBB)) { 1146 // If the prior block falls through into us, turn it into an 1147 // explicit branch to us to make updates simpler. 1148 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1149 PriorTBB != MBB && PriorFBB != MBB) { 1150 if (PriorTBB == 0) { 1151 assert(PriorCond.empty() && PriorFBB == 0 && 1152 "Bad branch analysis"); 1153 PriorTBB = MBB; 1154 } else { 1155 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1156 PriorFBB = MBB; 1157 } 1158 TII->RemoveBranch(PrevBB); 1159 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1160 } 1161 1162 // Iterate through all the predecessors, revectoring each in-turn. 1163 size_t PI = 0; 1164 bool DidChange = false; 1165 bool HasBranchToSelf = false; 1166 while(PI != MBB->pred_size()) { 1167 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1168 if (PMBB == MBB) { 1169 // If this block has an uncond branch to itself, leave it. 1170 ++PI; 1171 HasBranchToSelf = true; 1172 } else { 1173 DidChange = true; 1174 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1175 // If this change resulted in PMBB ending in a conditional 1176 // branch where both conditions go to the same destination, 1177 // change this to an unconditional branch (and fix the CFG). 1178 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1179 SmallVector<MachineOperand, 4> NewCurCond; 1180 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1181 NewCurFBB, NewCurCond, true); 1182 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1183 TII->RemoveBranch(*PMBB); 1184 NewCurCond.clear(); 1185 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond); 1186 MadeChange = true; 1187 ++NumBranchOpts; 1188 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1189 } 1190 } 1191 } 1192 1193 // Change any jumptables to go to the new MBB. 1194 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB); 1195 if (DidChange) { 1196 ++NumBranchOpts; 1197 MadeChange = true; 1198 if (!HasBranchToSelf) return MadeChange; 1199 } 1200 } 1201 } 1202 1203 // Add the branch back if the block is more than just an uncond branch. 1204 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1205 } 1206 } 1207 1208 // If the prior block doesn't fall through into this block, and if this 1209 // block doesn't fall through into some other block, see if we can find a 1210 // place to move this block where a fall-through will happen. 1211 if (!PrevBB.canFallThrough()) { 1212 1213 // Now we know that there was no fall-through into this block, check to 1214 // see if it has a fall-through into its successor. 1215 bool CurFallsThru = MBB->canFallThrough(); 1216 1217 if (!MBB->isLandingPad()) { 1218 // Check all the predecessors of this block. If one of them has no fall 1219 // throughs, move this block right after it. 1220 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1221 E = MBB->pred_end(); PI != E; ++PI) { 1222 // Analyze the branch at the end of the pred. 1223 MachineBasicBlock *PredBB = *PI; 1224 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1225 MachineBasicBlock *PredTBB = 0, *PredFBB = 0; 1226 SmallVector<MachineOperand, 4> PredCond; 1227 if (PredBB != MBB && !PredBB->canFallThrough() && 1228 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1229 && (!CurFallsThru || !CurTBB || !CurFBB) 1230 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1231 // If the current block doesn't fall through, just move it. 1232 // If the current block can fall through and does not end with a 1233 // conditional branch, we need to append an unconditional jump to 1234 // the (current) next block. To avoid a possible compile-time 1235 // infinite loop, move blocks only backward in this case. 1236 // Also, if there are already 2 branches here, we cannot add a third; 1237 // this means we have the case 1238 // Bcc next 1239 // B elsewhere 1240 // next: 1241 if (CurFallsThru) { 1242 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); 1243 CurCond.clear(); 1244 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1245 } 1246 MBB->moveAfter(PredBB); 1247 MadeChange = true; 1248 goto ReoptimizeBlock; 1249 } 1250 } 1251 } 1252 1253 if (!CurFallsThru) { 1254 // Check all successors to see if we can move this block before it. 1255 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1256 E = MBB->succ_end(); SI != E; ++SI) { 1257 // Analyze the branch at the end of the block before the succ. 1258 MachineBasicBlock *SuccBB = *SI; 1259 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1260 1261 // If this block doesn't already fall-through to that successor, and if 1262 // the succ doesn't already have a block that can fall through into it, 1263 // and if the successor isn't an EH destination, we can arrange for the 1264 // fallthrough to happen. 1265 if (SuccBB != MBB && &*SuccPrev != MBB && 1266 !SuccPrev->canFallThrough() && !CurUnAnalyzable && 1267 !SuccBB->isLandingPad()) { 1268 MBB->moveBefore(SuccBB); 1269 MadeChange = true; 1270 goto ReoptimizeBlock; 1271 } 1272 } 1273 1274 // Okay, there is no really great place to put this block. If, however, 1275 // the block before this one would be a fall-through if this block were 1276 // removed, move this block to the end of the function. 1277 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0; 1278 SmallVector<MachineOperand, 4> PrevCond; 1279 if (FallThrough != MF.end() && 1280 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1281 PrevBB.isSuccessor(FallThrough)) { 1282 MBB->moveAfter(--MF.end()); 1283 MadeChange = true; 1284 return MadeChange; 1285 } 1286 } 1287 } 1288 1289 return MadeChange; 1290} 1291