BranchFolding.cpp revision 9ccaf53ada99c63737547c0235baeb8454b04e80
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 // Remove the block. 109 MF->erase(MBB); 110} 111 112/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 113/// followed by terminators, and if the implicitly defined registers are not 114/// used by the terminators, remove those implicit_def's. e.g. 115/// BB1: 116/// r0 = implicit_def 117/// r1 = implicit_def 118/// br 119/// This block can be optimized away later if the implicit instructions are 120/// removed. 121bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 122 SmallSet<unsigned, 4> ImpDefRegs; 123 MachineBasicBlock::iterator I = MBB->begin(); 124 while (I != MBB->end()) { 125 if (!I->isImplicitDef()) 126 break; 127 unsigned Reg = I->getOperand(0).getReg(); 128 ImpDefRegs.insert(Reg); 129 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg); 130 unsigned SubReg = *SubRegs; ++SubRegs) 131 ImpDefRegs.insert(SubReg); 132 ++I; 133 } 134 if (ImpDefRegs.empty()) 135 return false; 136 137 MachineBasicBlock::iterator FirstTerm = I; 138 while (I != MBB->end()) { 139 if (!TII->isUnpredicatedTerminator(I)) 140 return false; 141 // See if it uses any of the implicitly defined registers. 142 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 143 MachineOperand &MO = I->getOperand(i); 144 if (!MO.isReg() || !MO.isUse()) 145 continue; 146 unsigned Reg = MO.getReg(); 147 if (ImpDefRegs.count(Reg)) 148 return false; 149 } 150 ++I; 151 } 152 153 I = MBB->begin(); 154 while (I != FirstTerm) { 155 MachineInstr *ImpDefMI = &*I; 156 ++I; 157 MBB->erase(ImpDefMI); 158 } 159 160 return true; 161} 162 163/// OptimizeFunction - Perhaps branch folding, tail merging and other 164/// CFG optimizations on the given function. 165bool BranchFolder::OptimizeFunction(MachineFunction &MF, 166 const TargetInstrInfo *tii, 167 const TargetRegisterInfo *tri, 168 MachineModuleInfo *mmi) { 169 if (!tii) return false; 170 171 TII = tii; 172 TRI = tri; 173 MMI = mmi; 174 175 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 176 177 // Fix CFG. The later algorithms expect it to be right. 178 bool MadeChange = false; 179 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 180 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 181 SmallVector<MachineOperand, 4> Cond; 182 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 183 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 184 MadeChange |= OptimizeImpDefsBlock(MBB); 185 } 186 187 bool MadeChangeThisIteration = true; 188 while (MadeChangeThisIteration) { 189 MadeChangeThisIteration = false; 190 MadeChangeThisIteration |= TailMergeBlocks(MF); 191 MadeChangeThisIteration |= OptimizeBranches(MF); 192 MadeChange |= MadeChangeThisIteration; 193 } 194 195 // See if any jump tables have become dead as the code generator 196 // did its thing. 197 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 198 if (JTI == 0) { 199 delete RS; 200 return MadeChange; 201 } 202 203 // Walk the function to find jump tables that are live. 204 BitVector JTIsLive(JTI->getJumpTables().size()); 205 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 206 BB != E; ++BB) { 207 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 208 I != E; ++I) 209 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 210 MachineOperand &Op = I->getOperand(op); 211 if (!Op.isJTI()) continue; 212 213 // Remember that this JT is live. 214 JTIsLive.set(Op.getIndex()); 215 } 216 } 217 218 // Finally, remove dead jump tables. This happens when the 219 // indirect jump was unreachable (and thus deleted). 220 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 221 if (!JTIsLive.test(i)) { 222 JTI->RemoveJumpTable(i); 223 MadeChange = true; 224 } 225 226 delete RS; 227 return MadeChange; 228} 229 230//===----------------------------------------------------------------------===// 231// Tail Merging of Blocks 232//===----------------------------------------------------------------------===// 233 234/// HashMachineInstr - Compute a hash value for MI and its operands. 235static unsigned HashMachineInstr(const MachineInstr *MI) { 236 unsigned Hash = MI->getOpcode(); 237 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 238 const MachineOperand &Op = MI->getOperand(i); 239 240 // Merge in bits from the operand if easy. 241 unsigned OperandHash = 0; 242 switch (Op.getType()) { 243 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 244 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 245 case MachineOperand::MO_MachineBasicBlock: 246 OperandHash = Op.getMBB()->getNumber(); 247 break; 248 case MachineOperand::MO_FrameIndex: 249 case MachineOperand::MO_ConstantPoolIndex: 250 case MachineOperand::MO_JumpTableIndex: 251 OperandHash = Op.getIndex(); 252 break; 253 case MachineOperand::MO_GlobalAddress: 254 case MachineOperand::MO_ExternalSymbol: 255 // Global address / external symbol are too hard, don't bother, but do 256 // pull in the offset. 257 OperandHash = Op.getOffset(); 258 break; 259 default: break; 260 } 261 262 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 263 } 264 return Hash; 265} 266 267/// HashEndOfMBB - Hash the last instruction in the MBB. 268static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) { 269 MachineBasicBlock::const_iterator I = MBB->end(); 270 if (I == MBB->begin()) 271 return 0; // Empty MBB. 272 273 --I; 274 // Skip debug info so it will not affect codegen. 275 while (I->isDebugValue()) { 276 if (I==MBB->begin()) 277 return 0; // MBB empty except for debug info. 278 --I; 279 } 280 281 return HashMachineInstr(I); 282} 283 284/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 285/// of instructions they actually have in common together at their end. Return 286/// iterators for the first shared instruction in each block. 287static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 288 MachineBasicBlock *MBB2, 289 MachineBasicBlock::iterator &I1, 290 MachineBasicBlock::iterator &I2) { 291 I1 = MBB1->end(); 292 I2 = MBB2->end(); 293 294 unsigned TailLen = 0; 295 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 296 --I1; --I2; 297 // Skip debugging pseudos; necessary to avoid changing the code. 298 while (I1->isDebugValue()) { 299 if (I1==MBB1->begin()) { 300 while (I2->isDebugValue()) { 301 if (I2==MBB2->begin()) 302 // I1==DBG at begin; I2==DBG at begin 303 return TailLen; 304 --I2; 305 } 306 ++I2; 307 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin 308 return TailLen; 309 } 310 --I1; 311 } 312 // I1==first (untested) non-DBG preceding known match 313 while (I2->isDebugValue()) { 314 if (I2==MBB2->begin()) { 315 ++I1; 316 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin 317 return TailLen; 318 } 319 --I2; 320 } 321 // I1, I2==first (untested) non-DBGs preceding known match 322 if (!I1->isIdenticalTo(I2) || 323 // FIXME: This check is dubious. It's used to get around a problem where 324 // people incorrectly expect inline asm directives to remain in the same 325 // relative order. This is untenable because normal compiler 326 // optimizations (like this one) may reorder and/or merge these 327 // directives. 328 I1->isInlineAsm()) { 329 ++I1; ++I2; 330 break; 331 } 332 ++TailLen; 333 } 334 // Back past possible debugging pseudos at beginning of block. This matters 335 // when one block differs from the other only by whether debugging pseudos 336 // are present at the beginning. (This way, the various checks later for 337 // I1==MBB1->begin() work as expected.) 338 if (I1 == MBB1->begin() && I2 != MBB2->begin()) { 339 --I2; 340 while (I2->isDebugValue()) { 341 if (I2 == MBB2->begin()) { 342 return TailLen; 343 } 344 --I2; 345 } 346 ++I2; 347 } 348 if (I2 == MBB2->begin() && I1 != MBB1->begin()) { 349 --I1; 350 while (I1->isDebugValue()) { 351 if (I1 == MBB1->begin()) 352 return TailLen; 353 --I1; 354 } 355 ++I1; 356 } 357 return TailLen; 358} 359 360/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 361/// after it, replacing it with an unconditional branch to NewDest. 362void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 363 MachineBasicBlock *NewDest) { 364 TII->ReplaceTailWithBranchTo(OldInst, NewDest); 365 ++NumTailMerge; 366} 367 368/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 369/// MBB so that the part before the iterator falls into the part starting at the 370/// iterator. This returns the new MBB. 371MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 372 MachineBasicBlock::iterator BBI1) { 373 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1)) 374 return 0; 375 376 MachineFunction &MF = *CurMBB.getParent(); 377 378 // Create the fall-through block. 379 MachineFunction::iterator MBBI = &CurMBB; 380 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 381 CurMBB.getParent()->insert(++MBBI, NewMBB); 382 383 // Move all the successors of this block to the specified block. 384 NewMBB->transferSuccessors(&CurMBB); 385 386 // Add an edge from CurMBB to NewMBB for the fall-through. 387 CurMBB.addSuccessor(NewMBB); 388 389 // Splice the code over. 390 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 391 392 // For targets that use the register scavenger, we must maintain LiveIns. 393 if (RS) { 394 RS->enterBasicBlock(&CurMBB); 395 if (!CurMBB.empty()) 396 RS->forward(prior(CurMBB.end())); 397 BitVector RegsLiveAtExit(TRI->getNumRegs()); 398 RS->getRegsUsed(RegsLiveAtExit, false); 399 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 400 if (RegsLiveAtExit[i]) 401 NewMBB->addLiveIn(i); 402 } 403 404 return NewMBB; 405} 406 407/// EstimateRuntime - Make a rough estimate for how long it will take to run 408/// the specified code. 409static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 410 MachineBasicBlock::iterator E) { 411 unsigned Time = 0; 412 for (; I != E; ++I) { 413 if (I->isDebugValue()) 414 continue; 415 const TargetInstrDesc &TID = I->getDesc(); 416 if (TID.isCall()) 417 Time += 10; 418 else if (TID.mayLoad() || TID.mayStore()) 419 Time += 2; 420 else 421 ++Time; 422 } 423 return Time; 424} 425 426// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 427// branches temporarily for tail merging). In the case where CurMBB ends 428// with a conditional branch to the next block, optimize by reversing the 429// test and conditionally branching to SuccMBB instead. 430static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 431 const TargetInstrInfo *TII) { 432 MachineFunction *MF = CurMBB->getParent(); 433 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB)); 434 MachineBasicBlock *TBB = 0, *FBB = 0; 435 SmallVector<MachineOperand, 4> Cond; 436 DebugLoc dl; // FIXME: this is nowhere 437 if (I != MF->end() && 438 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 439 MachineBasicBlock *NextBB = I; 440 if (TBB == NextBB && !Cond.empty() && !FBB) { 441 if (!TII->ReverseBranchCondition(Cond)) { 442 TII->RemoveBranch(*CurMBB); 443 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl); 444 return; 445 } 446 } 447 } 448 TII->InsertBranch(*CurMBB, SuccBB, NULL, 449 SmallVector<MachineOperand, 0>(), dl); 450} 451 452bool 453BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 454 if (getHash() < o.getHash()) 455 return true; 456 else if (getHash() > o.getHash()) 457 return false; 458 else if (getBlock()->getNumber() < o.getBlock()->getNumber()) 459 return true; 460 else if (getBlock()->getNumber() > o.getBlock()->getNumber()) 461 return false; 462 else { 463 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 464 // an object with itself. 465#ifndef _GLIBCXX_DEBUG 466 llvm_unreachable("Predecessor appears twice"); 467#endif 468 return false; 469 } 470} 471 472/// CountTerminators - Count the number of terminators in the given 473/// block and set I to the position of the first non-terminator, if there 474/// is one, or MBB->end() otherwise. 475static unsigned CountTerminators(MachineBasicBlock *MBB, 476 MachineBasicBlock::iterator &I) { 477 I = MBB->end(); 478 unsigned NumTerms = 0; 479 for (;;) { 480 if (I == MBB->begin()) { 481 I = MBB->end(); 482 break; 483 } 484 --I; 485 if (!I->getDesc().isTerminator()) break; 486 ++NumTerms; 487 } 488 return NumTerms; 489} 490 491/// ProfitableToMerge - Check if two machine basic blocks have a common tail 492/// and decide if it would be profitable to merge those tails. Return the 493/// length of the common tail and iterators to the first common instruction 494/// in each block. 495static bool ProfitableToMerge(MachineBasicBlock *MBB1, 496 MachineBasicBlock *MBB2, 497 unsigned minCommonTailLength, 498 unsigned &CommonTailLen, 499 MachineBasicBlock::iterator &I1, 500 MachineBasicBlock::iterator &I2, 501 MachineBasicBlock *SuccBB, 502 MachineBasicBlock *PredBB) { 503 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 504 MachineFunction *MF = MBB1->getParent(); 505 506 if (CommonTailLen == 0) 507 return false; 508 509 // It's almost always profitable to merge any number of non-terminator 510 // instructions with the block that falls through into the common successor. 511 if (MBB1 == PredBB || MBB2 == PredBB) { 512 MachineBasicBlock::iterator I; 513 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 514 if (CommonTailLen > NumTerms) 515 return true; 516 } 517 518 // If one of the blocks can be completely merged and happens to be in 519 // a position where the other could fall through into it, merge any number 520 // of instructions, because it can be done without a branch. 521 // TODO: If the blocks are not adjacent, move one of them so that they are? 522 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 523 return true; 524 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 525 return true; 526 527 // If both blocks have an unconditional branch temporarily stripped out, 528 // count that as an additional common instruction for the following 529 // heuristics. 530 unsigned EffectiveTailLen = CommonTailLen; 531 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 532 !MBB1->back().getDesc().isBarrier() && 533 !MBB2->back().getDesc().isBarrier()) 534 ++EffectiveTailLen; 535 536 // Check if the common tail is long enough to be worthwhile. 537 if (EffectiveTailLen >= minCommonTailLength) 538 return true; 539 540 // If we are optimizing for code size, 2 instructions in common is enough if 541 // we don't have to split a block. At worst we will be introducing 1 new 542 // branch instruction, which is likely to be smaller than the 2 543 // instructions that would be deleted in the merge. 544 if (EffectiveTailLen >= 2 && 545 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 546 (I1 == MBB1->begin() || I2 == MBB2->begin())) 547 return true; 548 549 return false; 550} 551 552/// ComputeSameTails - Look through all the blocks in MergePotentials that have 553/// hash CurHash (guaranteed to match the last element). Build the vector 554/// SameTails of all those that have the (same) largest number of instructions 555/// in common of any pair of these blocks. SameTails entries contain an 556/// iterator into MergePotentials (from which the MachineBasicBlock can be 557/// found) and a MachineBasicBlock::iterator into that MBB indicating the 558/// instruction where the matching code sequence begins. 559/// Order of elements in SameTails is the reverse of the order in which 560/// those blocks appear in MergePotentials (where they are not necessarily 561/// consecutive). 562unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 563 unsigned minCommonTailLength, 564 MachineBasicBlock *SuccBB, 565 MachineBasicBlock *PredBB) { 566 unsigned maxCommonTailLength = 0U; 567 SameTails.clear(); 568 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 569 MPIterator HighestMPIter = prior(MergePotentials.end()); 570 for (MPIterator CurMPIter = prior(MergePotentials.end()), 571 B = MergePotentials.begin(); 572 CurMPIter != B && CurMPIter->getHash() == CurHash; 573 --CurMPIter) { 574 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 575 unsigned CommonTailLen; 576 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 577 minCommonTailLength, 578 CommonTailLen, TrialBBI1, TrialBBI2, 579 SuccBB, PredBB)) { 580 if (CommonTailLen > maxCommonTailLength) { 581 SameTails.clear(); 582 maxCommonTailLength = CommonTailLen; 583 HighestMPIter = CurMPIter; 584 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 585 } 586 if (HighestMPIter == CurMPIter && 587 CommonTailLen == maxCommonTailLength) 588 SameTails.push_back(SameTailElt(I, TrialBBI2)); 589 } 590 if (I == B) 591 break; 592 } 593 } 594 return maxCommonTailLength; 595} 596 597/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 598/// MergePotentials, restoring branches at ends of blocks as appropriate. 599void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 600 MachineBasicBlock *SuccBB, 601 MachineBasicBlock *PredBB) { 602 MPIterator CurMPIter, B; 603 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 604 CurMPIter->getHash() == CurHash; 605 --CurMPIter) { 606 // Put the unconditional branch back, if we need one. 607 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 608 if (SuccBB && CurMBB != PredBB) 609 FixTail(CurMBB, SuccBB, TII); 610 if (CurMPIter == B) 611 break; 612 } 613 if (CurMPIter->getHash() != CurHash) 614 CurMPIter++; 615 MergePotentials.erase(CurMPIter, MergePotentials.end()); 616} 617 618/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 619/// only of the common tail. Create a block that does by splitting one. 620bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 621 unsigned maxCommonTailLength, 622 unsigned &commonTailIndex) { 623 commonTailIndex = 0; 624 unsigned TimeEstimate = ~0U; 625 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 626 // Use PredBB if possible; that doesn't require a new branch. 627 if (SameTails[i].getBlock() == PredBB) { 628 commonTailIndex = i; 629 break; 630 } 631 // Otherwise, make a (fairly bogus) choice based on estimate of 632 // how long it will take the various blocks to execute. 633 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 634 SameTails[i].getTailStartPos()); 635 if (t <= TimeEstimate) { 636 TimeEstimate = t; 637 commonTailIndex = i; 638 } 639 } 640 641 MachineBasicBlock::iterator BBI = 642 SameTails[commonTailIndex].getTailStartPos(); 643 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 644 645 // If the common tail includes any debug info we will take it pretty 646 // randomly from one of the inputs. Might be better to remove it? 647 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 648 << maxCommonTailLength); 649 650 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 651 if (!newMBB) { 652 DEBUG(dbgs() << "... failed!"); 653 return false; 654 } 655 656 SameTails[commonTailIndex].setBlock(newMBB); 657 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 658 659 // If we split PredBB, newMBB is the new predecessor. 660 if (PredBB == MBB) 661 PredBB = newMBB; 662 663 return true; 664} 665 666// See if any of the blocks in MergePotentials (which all have a common single 667// successor, or all have no successor) can be tail-merged. If there is a 668// successor, any blocks in MergePotentials that are not tail-merged and 669// are not immediately before Succ must have an unconditional branch to 670// Succ added (but the predecessor/successor lists need no adjustment). 671// The lone predecessor of Succ that falls through into Succ, 672// if any, is given in PredBB. 673 674bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 675 MachineBasicBlock *PredBB) { 676 bool MadeChange = false; 677 678 // Except for the special cases below, tail-merge if there are at least 679 // this many instructions in common. 680 unsigned minCommonTailLength = TailMergeSize; 681 682 DEBUG(dbgs() << "\nTryTailMergeBlocks: "; 683 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 684 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 685 << (i == e-1 ? "" : ", "); 686 dbgs() << "\n"; 687 if (SuccBB) { 688 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 689 if (PredBB) 690 dbgs() << " which has fall-through from BB#" 691 << PredBB->getNumber() << "\n"; 692 } 693 dbgs() << "Looking for common tails of at least " 694 << minCommonTailLength << " instruction" 695 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 696 ); 697 698 // Sort by hash value so that blocks with identical end sequences sort 699 // together. 700 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 701 702 // Walk through equivalence sets looking for actual exact matches. 703 while (MergePotentials.size() > 1) { 704 unsigned CurHash = MergePotentials.back().getHash(); 705 706 // Build SameTails, identifying the set of blocks with this hash code 707 // and with the maximum number of instructions in common. 708 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 709 minCommonTailLength, 710 SuccBB, PredBB); 711 712 // If we didn't find any pair that has at least minCommonTailLength 713 // instructions in common, remove all blocks with this hash code and retry. 714 if (SameTails.empty()) { 715 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 716 continue; 717 } 718 719 // If one of the blocks is the entire common tail (and not the entry 720 // block, which we can't jump to), we can treat all blocks with this same 721 // tail at once. Use PredBB if that is one of the possibilities, as that 722 // will not introduce any extra branches. 723 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 724 getParent()->begin(); 725 unsigned commonTailIndex = SameTails.size(); 726 // If there are two blocks, check to see if one can be made to fall through 727 // into the other. 728 if (SameTails.size() == 2 && 729 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 730 SameTails[1].tailIsWholeBlock()) 731 commonTailIndex = 1; 732 else if (SameTails.size() == 2 && 733 SameTails[1].getBlock()->isLayoutSuccessor( 734 SameTails[0].getBlock()) && 735 SameTails[0].tailIsWholeBlock()) 736 commonTailIndex = 0; 737 else { 738 // Otherwise just pick one, favoring the fall-through predecessor if 739 // there is one. 740 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 741 MachineBasicBlock *MBB = SameTails[i].getBlock(); 742 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 743 continue; 744 if (MBB == PredBB) { 745 commonTailIndex = i; 746 break; 747 } 748 if (SameTails[i].tailIsWholeBlock()) 749 commonTailIndex = i; 750 } 751 } 752 753 if (commonTailIndex == SameTails.size() || 754 (SameTails[commonTailIndex].getBlock() == PredBB && 755 !SameTails[commonTailIndex].tailIsWholeBlock())) { 756 // None of the blocks consist entirely of the common tail. 757 // Split a block so that one does. 758 if (!CreateCommonTailOnlyBlock(PredBB, 759 maxCommonTailLength, commonTailIndex)) { 760 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 761 continue; 762 } 763 } 764 765 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 766 // MBB is common tail. Adjust all other BB's to jump to this one. 767 // Traversal must be forwards so erases work. 768 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber() 769 << " for "); 770 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 771 if (commonTailIndex == i) 772 continue; 773 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() 774 << (i == e-1 ? "" : ", ")); 775 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 776 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 777 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 778 MergePotentials.erase(SameTails[i].getMPIter()); 779 } 780 DEBUG(dbgs() << "\n"); 781 // We leave commonTailIndex in the worklist in case there are other blocks 782 // that match it with a smaller number of instructions. 783 MadeChange = true; 784 } 785 return MadeChange; 786} 787 788bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 789 790 if (!EnableTailMerge) return false; 791 792 bool MadeChange = false; 793 794 // First find blocks with no successors. 795 MergePotentials.clear(); 796 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 797 if (I->succ_empty()) 798 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I)); 799 } 800 801 // See if we can do any tail merging on those. 802 if (MergePotentials.size() < TailMergeThreshold && 803 MergePotentials.size() >= 2) 804 MadeChange |= TryTailMergeBlocks(NULL, NULL); 805 806 // Look at blocks (IBB) with multiple predecessors (PBB). 807 // We change each predecessor to a canonical form, by 808 // (1) temporarily removing any unconditional branch from the predecessor 809 // to IBB, and 810 // (2) alter conditional branches so they branch to the other block 811 // not IBB; this may require adding back an unconditional branch to IBB 812 // later, where there wasn't one coming in. E.g. 813 // Bcc IBB 814 // fallthrough to QBB 815 // here becomes 816 // Bncc QBB 817 // with a conceptual B to IBB after that, which never actually exists. 818 // With those changes, we see whether the predecessors' tails match, 819 // and merge them if so. We change things out of canonical form and 820 // back to the way they were later in the process. (OptimizeBranches 821 // would undo some of this, but we can't use it, because we'd get into 822 // a compile-time infinite loop repeatedly doing and undoing the same 823 // transformations.) 824 825 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 826 I != E; ++I) { 827 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 828 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 829 MachineBasicBlock *IBB = I; 830 MachineBasicBlock *PredBB = prior(I); 831 MergePotentials.clear(); 832 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 833 E2 = I->pred_end(); 834 P != E2; ++P) { 835 MachineBasicBlock *PBB = *P; 836 // Skip blocks that loop to themselves, can't tail merge these. 837 if (PBB == IBB) 838 continue; 839 // Visit each predecessor only once. 840 if (!UniquePreds.insert(PBB)) 841 continue; 842 MachineBasicBlock *TBB = 0, *FBB = 0; 843 SmallVector<MachineOperand, 4> Cond; 844 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 845 // Failing case: IBB is the target of a cbr, and 846 // we cannot reverse the branch. 847 SmallVector<MachineOperand, 4> NewCond(Cond); 848 if (!Cond.empty() && TBB == IBB) { 849 if (TII->ReverseBranchCondition(NewCond)) 850 continue; 851 // This is the QBB case described above 852 if (!FBB) 853 FBB = llvm::next(MachineFunction::iterator(PBB)); 854 } 855 // Failing case: the only way IBB can be reached from PBB is via 856 // exception handling. Happens for landing pads. Would be nice 857 // to have a bit in the edge so we didn't have to do all this. 858 if (IBB->isLandingPad()) { 859 MachineFunction::iterator IP = PBB; IP++; 860 MachineBasicBlock *PredNextBB = NULL; 861 if (IP != MF.end()) 862 PredNextBB = IP; 863 if (TBB == NULL) { 864 if (IBB != PredNextBB) // fallthrough 865 continue; 866 } else if (FBB) { 867 if (TBB != IBB && FBB != IBB) // cbr then ubr 868 continue; 869 } else if (Cond.empty()) { 870 if (TBB != IBB) // ubr 871 continue; 872 } else { 873 if (TBB != IBB && IBB != PredNextBB) // cbr 874 continue; 875 } 876 } 877 // Remove the unconditional branch at the end, if any. 878 if (TBB && (Cond.empty() || FBB)) { 879 DebugLoc dl; // FIXME: this is nowhere 880 TII->RemoveBranch(*PBB); 881 if (!Cond.empty()) 882 // reinsert conditional branch only, for now 883 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl); 884 } 885 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P)); 886 } 887 } 888 if (MergePotentials.size() >= 2) 889 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 890 // Reinsert an unconditional branch if needed. 891 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks. 892 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 893 if (MergePotentials.size() == 1 && 894 MergePotentials.begin()->getBlock() != PredBB) 895 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 896 } 897 } 898 return MadeChange; 899} 900 901//===----------------------------------------------------------------------===// 902// Branch Optimization 903//===----------------------------------------------------------------------===// 904 905bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 906 bool MadeChange = false; 907 908 // Make sure blocks are numbered in order 909 MF.RenumberBlocks(); 910 911 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 912 MachineBasicBlock *MBB = I++; 913 MadeChange |= OptimizeBlock(MBB); 914 915 // If it is dead, remove it. 916 if (MBB->pred_empty()) { 917 RemoveDeadBlock(MBB); 918 MadeChange = true; 919 ++NumDeadBlocks; 920 } 921 } 922 return MadeChange; 923} 924 925// Blocks should be considered empty if they contain only debug info; 926// else the debug info would affect codegen. 927static bool IsEmptyBlock(MachineBasicBlock *MBB) { 928 if (MBB->empty()) 929 return true; 930 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end(); 931 MBBI!=MBBE; ++MBBI) { 932 if (!MBBI->isDebugValue()) 933 return false; 934 } 935 return true; 936} 937 938// Blocks with only debug info and branches should be considered the same 939// as blocks with only branches. 940static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { 941 MachineBasicBlock::iterator MBBI, MBBE; 942 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) { 943 if (!MBBI->isDebugValue()) 944 break; 945 } 946 return (MBBI->getDesc().isBranch()); 947} 948 949/// IsBetterFallthrough - Return true if it would be clearly better to 950/// fall-through to MBB1 than to fall through into MBB2. This has to return 951/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 952/// result in infinite loops. 953static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 954 MachineBasicBlock *MBB2) { 955 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 956 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 957 // optimize branches that branch to either a return block or an assert block 958 // into a fallthrough to the return. 959 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false; 960 961 // If there is a clear successor ordering we make sure that one block 962 // will fall through to the next 963 if (MBB1->isSuccessor(MBB2)) return true; 964 if (MBB2->isSuccessor(MBB1)) return false; 965 966 // Neither block consists entirely of debug info (per IsEmptyBlock check), 967 // so we needn't test for falling off the beginning here. 968 MachineBasicBlock::iterator MBB1I = --MBB1->end(); 969 while (MBB1I->isDebugValue()) 970 --MBB1I; 971 MachineBasicBlock::iterator MBB2I = --MBB2->end(); 972 while (MBB2I->isDebugValue()) 973 --MBB2I; 974 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 975} 976 977/// OptimizeBlock - Analyze and optimize control flow related to the specified 978/// block. This is never called on the entry block. 979bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 980 bool MadeChange = false; 981 MachineFunction &MF = *MBB->getParent(); 982 DebugLoc dl; // FIXME: this is nowhere 983ReoptimizeBlock: 984 985 MachineFunction::iterator FallThrough = MBB; 986 ++FallThrough; 987 988 // If this block is empty, make everyone use its fall-through, not the block 989 // explicitly. Landing pads should not do this since the landing-pad table 990 // points to this block. Blocks with their addresses taken shouldn't be 991 // optimized away. 992 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 993 // Dead block? Leave for cleanup later. 994 if (MBB->pred_empty()) return MadeChange; 995 996 if (FallThrough == MF.end()) { 997 // TODO: Simplify preds to not branch here if possible! 998 } else { 999 // Rewrite all predecessors of the old block to go to the fallthrough 1000 // instead. 1001 while (!MBB->pred_empty()) { 1002 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1003 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1004 } 1005 // If MBB was the target of a jump table, update jump tables to go to the 1006 // fallthrough instead. 1007 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1008 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough); 1009 MadeChange = true; 1010 } 1011 return MadeChange; 1012 } 1013 1014 // Check to see if we can simplify the terminator of the block before this 1015 // one. 1016 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1017 1018 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1019 SmallVector<MachineOperand, 4> PriorCond; 1020 bool PriorUnAnalyzable = 1021 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1022 if (!PriorUnAnalyzable) { 1023 // If the CFG for the prior block has extra edges, remove them. 1024 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1025 !PriorCond.empty()); 1026 1027 // If the previous branch is conditional and both conditions go to the same 1028 // destination, remove the branch, replacing it with an unconditional one or 1029 // a fall-through. 1030 if (PriorTBB && PriorTBB == PriorFBB) { 1031 TII->RemoveBranch(PrevBB); 1032 PriorCond.clear(); 1033 if (PriorTBB != MBB) 1034 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1035 MadeChange = true; 1036 ++NumBranchOpts; 1037 goto ReoptimizeBlock; 1038 } 1039 1040 // If the previous block unconditionally falls through to this block and 1041 // this block has no other predecessors, move the contents of this block 1042 // into the prior block. This doesn't usually happen when SimplifyCFG 1043 // has been used, but it can happen if tail merging splits a fall-through 1044 // predecessor of a block. 1045 // This has to check PrevBB->succ_size() because EH edges are ignored by 1046 // AnalyzeBranch. 1047 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1048 PrevBB.succ_size() == 1 && 1049 !MBB->hasAddressTaken()) { 1050 DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1051 << "From MBB: " << *MBB); 1052 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1053 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1054 assert(PrevBB.succ_empty()); 1055 PrevBB.transferSuccessors(MBB); 1056 MadeChange = true; 1057 return MadeChange; 1058 } 1059 1060 // If the previous branch *only* branches to *this* block (conditional or 1061 // not) remove the branch. 1062 if (PriorTBB == MBB && PriorFBB == 0) { 1063 TII->RemoveBranch(PrevBB); 1064 MadeChange = true; 1065 ++NumBranchOpts; 1066 goto ReoptimizeBlock; 1067 } 1068 1069 // If the prior block branches somewhere else on the condition and here if 1070 // the condition is false, remove the uncond second branch. 1071 if (PriorFBB == MBB) { 1072 TII->RemoveBranch(PrevBB); 1073 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1074 MadeChange = true; 1075 ++NumBranchOpts; 1076 goto ReoptimizeBlock; 1077 } 1078 1079 // If the prior block branches here on true and somewhere else on false, and 1080 // if the branch condition is reversible, reverse the branch to create a 1081 // fall-through. 1082 if (PriorTBB == MBB) { 1083 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1084 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1085 TII->RemoveBranch(PrevBB); 1086 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl); 1087 MadeChange = true; 1088 ++NumBranchOpts; 1089 goto ReoptimizeBlock; 1090 } 1091 } 1092 1093 // If this block has no successors (e.g. it is a return block or ends with 1094 // a call to a no-return function like abort or __cxa_throw) and if the pred 1095 // falls through into this block, and if it would otherwise fall through 1096 // into the block after this, move this block to the end of the function. 1097 // 1098 // We consider it more likely that execution will stay in the function (e.g. 1099 // due to loops) than it is to exit it. This asserts in loops etc, moving 1100 // the assert condition out of the loop body. 1101 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1102 MachineFunction::iterator(PriorTBB) == FallThrough && 1103 !MBB->canFallThrough()) { 1104 bool DoTransform = true; 1105 1106 // We have to be careful that the succs of PredBB aren't both no-successor 1107 // blocks. If neither have successors and if PredBB is the second from 1108 // last block in the function, we'd just keep swapping the two blocks for 1109 // last. Only do the swap if one is clearly better to fall through than 1110 // the other. 1111 if (FallThrough == --MF.end() && 1112 !IsBetterFallthrough(PriorTBB, MBB)) 1113 DoTransform = false; 1114 1115 if (DoTransform) { 1116 // Reverse the branch so we will fall through on the previous true cond. 1117 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1118 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1119 DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1120 << "To make fallthrough to: " << *PriorTBB << "\n"); 1121 1122 TII->RemoveBranch(PrevBB); 1123 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl); 1124 1125 // Move this block to the end of the function. 1126 MBB->moveAfter(--MF.end()); 1127 MadeChange = true; 1128 ++NumBranchOpts; 1129 return MadeChange; 1130 } 1131 } 1132 } 1133 } 1134 1135 // Analyze the branch in the current block. 1136 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1137 SmallVector<MachineOperand, 4> CurCond; 1138 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1139 if (!CurUnAnalyzable) { 1140 // If the CFG for the prior block has extra edges, remove them. 1141 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1142 1143 // If this is a two-way branch, and the FBB branches to this block, reverse 1144 // the condition so the single-basic-block loop is faster. Instead of: 1145 // Loop: xxx; jcc Out; jmp Loop 1146 // we want: 1147 // Loop: xxx; jncc Loop; jmp Out 1148 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1149 SmallVector<MachineOperand, 4> NewCond(CurCond); 1150 if (!TII->ReverseBranchCondition(NewCond)) { 1151 TII->RemoveBranch(*MBB); 1152 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); 1153 MadeChange = true; 1154 ++NumBranchOpts; 1155 goto ReoptimizeBlock; 1156 } 1157 } 1158 1159 // If this branch is the only thing in its block, see if we can forward 1160 // other blocks across it. 1161 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1162 IsBranchOnlyBlock(MBB) && CurTBB != MBB && 1163 !MBB->hasAddressTaken()) { 1164 // This block may contain just an unconditional branch. Because there can 1165 // be 'non-branch terminators' in the block, try removing the branch and 1166 // then seeing if the block is empty. 1167 TII->RemoveBranch(*MBB); 1168 // If the only things remaining in the block are debug info, remove these 1169 // as well, so this will behave the same as an empty block in non-debug 1170 // mode. 1171 if (!MBB->empty()) { 1172 bool NonDebugInfoFound = false; 1173 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 1174 I != E; ++I) { 1175 if (!I->isDebugValue()) { 1176 NonDebugInfoFound = true; 1177 break; 1178 } 1179 } 1180 if (!NonDebugInfoFound) 1181 // Make the block empty, losing the debug info (we could probably 1182 // improve this in some cases.) 1183 MBB->erase(MBB->begin(), MBB->end()); 1184 } 1185 // If this block is just an unconditional branch to CurTBB, we can 1186 // usually completely eliminate the block. The only case we cannot 1187 // completely eliminate the block is when the block before this one 1188 // falls through into MBB and we can't understand the prior block's branch 1189 // condition. 1190 if (MBB->empty()) { 1191 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1192 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1193 !PrevBB.isSuccessor(MBB)) { 1194 // If the prior block falls through into us, turn it into an 1195 // explicit branch to us to make updates simpler. 1196 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1197 PriorTBB != MBB && PriorFBB != MBB) { 1198 if (PriorTBB == 0) { 1199 assert(PriorCond.empty() && PriorFBB == 0 && 1200 "Bad branch analysis"); 1201 PriorTBB = MBB; 1202 } else { 1203 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1204 PriorFBB = MBB; 1205 } 1206 TII->RemoveBranch(PrevBB); 1207 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, dl); 1208 } 1209 1210 // Iterate through all the predecessors, revectoring each in-turn. 1211 size_t PI = 0; 1212 bool DidChange = false; 1213 bool HasBranchToSelf = false; 1214 while(PI != MBB->pred_size()) { 1215 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1216 if (PMBB == MBB) { 1217 // If this block has an uncond branch to itself, leave it. 1218 ++PI; 1219 HasBranchToSelf = true; 1220 } else { 1221 DidChange = true; 1222 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1223 // If this change resulted in PMBB ending in a conditional 1224 // branch where both conditions go to the same destination, 1225 // change this to an unconditional branch (and fix the CFG). 1226 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1227 SmallVector<MachineOperand, 4> NewCurCond; 1228 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1229 NewCurFBB, NewCurCond, true); 1230 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1231 TII->RemoveBranch(*PMBB); 1232 NewCurCond.clear(); 1233 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, dl); 1234 MadeChange = true; 1235 ++NumBranchOpts; 1236 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1237 } 1238 } 1239 } 1240 1241 // Change any jumptables to go to the new MBB. 1242 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1243 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); 1244 if (DidChange) { 1245 ++NumBranchOpts; 1246 MadeChange = true; 1247 if (!HasBranchToSelf) return MadeChange; 1248 } 1249 } 1250 } 1251 1252 // Add the branch back if the block is more than just an uncond branch. 1253 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl); 1254 } 1255 } 1256 1257 // If the prior block doesn't fall through into this block, and if this 1258 // block doesn't fall through into some other block, see if we can find a 1259 // place to move this block where a fall-through will happen. 1260 if (!PrevBB.canFallThrough()) { 1261 1262 // Now we know that there was no fall-through into this block, check to 1263 // see if it has a fall-through into its successor. 1264 bool CurFallsThru = MBB->canFallThrough(); 1265 1266 if (!MBB->isLandingPad()) { 1267 // Check all the predecessors of this block. If one of them has no fall 1268 // throughs, move this block right after it. 1269 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1270 E = MBB->pred_end(); PI != E; ++PI) { 1271 // Analyze the branch at the end of the pred. 1272 MachineBasicBlock *PredBB = *PI; 1273 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1274 MachineBasicBlock *PredTBB = 0, *PredFBB = 0; 1275 SmallVector<MachineOperand, 4> PredCond; 1276 if (PredBB != MBB && !PredBB->canFallThrough() && 1277 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1278 && (!CurFallsThru || !CurTBB || !CurFBB) 1279 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1280 // If the current block doesn't fall through, just move it. 1281 // If the current block can fall through and does not end with a 1282 // conditional branch, we need to append an unconditional jump to 1283 // the (current) next block. To avoid a possible compile-time 1284 // infinite loop, move blocks only backward in this case. 1285 // Also, if there are already 2 branches here, we cannot add a third; 1286 // this means we have the case 1287 // Bcc next 1288 // B elsewhere 1289 // next: 1290 if (CurFallsThru) { 1291 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); 1292 CurCond.clear(); 1293 TII->InsertBranch(*MBB, NextBB, 0, CurCond, dl); 1294 } 1295 MBB->moveAfter(PredBB); 1296 MadeChange = true; 1297 goto ReoptimizeBlock; 1298 } 1299 } 1300 } 1301 1302 if (!CurFallsThru) { 1303 // Check all successors to see if we can move this block before it. 1304 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1305 E = MBB->succ_end(); SI != E; ++SI) { 1306 // Analyze the branch at the end of the block before the succ. 1307 MachineBasicBlock *SuccBB = *SI; 1308 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1309 1310 // If this block doesn't already fall-through to that successor, and if 1311 // the succ doesn't already have a block that can fall through into it, 1312 // and if the successor isn't an EH destination, we can arrange for the 1313 // fallthrough to happen. 1314 if (SuccBB != MBB && &*SuccPrev != MBB && 1315 !SuccPrev->canFallThrough() && !CurUnAnalyzable && 1316 !SuccBB->isLandingPad()) { 1317 MBB->moveBefore(SuccBB); 1318 MadeChange = true; 1319 goto ReoptimizeBlock; 1320 } 1321 } 1322 1323 // Okay, there is no really great place to put this block. If, however, 1324 // the block before this one would be a fall-through if this block were 1325 // removed, move this block to the end of the function. 1326 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0; 1327 SmallVector<MachineOperand, 4> PrevCond; 1328 if (FallThrough != MF.end() && 1329 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1330 PrevBB.isSuccessor(FallThrough)) { 1331 MBB->moveAfter(--MF.end()); 1332 MadeChange = true; 1333 return MadeChange; 1334 } 1335 } 1336 } 1337 1338 return MadeChange; 1339} 1340