BranchFolding.cpp revision 9fa5e3347d0fcd2a4626807dc2d8f7ceedf48ffd
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. This 362/// returns true if OldInst's block is modified, false if NewDest is modified. 363void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 364 MachineBasicBlock *NewDest) { 365 MachineBasicBlock *OldBB = OldInst->getParent(); 366 367 // Remove all the old successors of OldBB from the CFG. 368 while (!OldBB->succ_empty()) 369 OldBB->removeSuccessor(OldBB->succ_begin()); 370 371 // Remove all the dead instructions from the end of OldBB. 372 OldBB->erase(OldInst, OldBB->end()); 373 374 // If OldBB isn't immediately before OldBB, insert a branch to it. 375 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 376 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>(), 377 OldInst->getDebugLoc()); 378 OldBB->addSuccessor(NewDest); 379 ++NumTailMerge; 380} 381 382/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 383/// MBB so that the part before the iterator falls into the part starting at the 384/// iterator. This returns the new MBB. 385MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 386 MachineBasicBlock::iterator BBI1) { 387 MachineFunction &MF = *CurMBB.getParent(); 388 389 // Create the fall-through block. 390 MachineFunction::iterator MBBI = &CurMBB; 391 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 392 CurMBB.getParent()->insert(++MBBI, NewMBB); 393 394 // Move all the successors of this block to the specified block. 395 NewMBB->transferSuccessors(&CurMBB); 396 397 // Add an edge from CurMBB to NewMBB for the fall-through. 398 CurMBB.addSuccessor(NewMBB); 399 400 // Splice the code over. 401 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 402 403 // For targets that use the register scavenger, we must maintain LiveIns. 404 if (RS) { 405 RS->enterBasicBlock(&CurMBB); 406 if (!CurMBB.empty()) 407 RS->forward(prior(CurMBB.end())); 408 BitVector RegsLiveAtExit(TRI->getNumRegs()); 409 RS->getRegsUsed(RegsLiveAtExit, false); 410 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 411 if (RegsLiveAtExit[i]) 412 NewMBB->addLiveIn(i); 413 } 414 415 return NewMBB; 416} 417 418/// EstimateRuntime - Make a rough estimate for how long it will take to run 419/// the specified code. 420static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 421 MachineBasicBlock::iterator E) { 422 unsigned Time = 0; 423 for (; I != E; ++I) { 424 if (I->isDebugValue()) 425 continue; 426 const TargetInstrDesc &TID = I->getDesc(); 427 if (TID.isCall()) 428 Time += 10; 429 else if (TID.mayLoad() || TID.mayStore()) 430 Time += 2; 431 else 432 ++Time; 433 } 434 return Time; 435} 436 437// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 438// branches temporarily for tail merging). In the case where CurMBB ends 439// with a conditional branch to the next block, optimize by reversing the 440// test and conditionally branching to SuccMBB instead. 441static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 442 const TargetInstrInfo *TII) { 443 MachineFunction *MF = CurMBB->getParent(); 444 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB)); 445 MachineBasicBlock *TBB = 0, *FBB = 0; 446 SmallVector<MachineOperand, 4> Cond; 447 DebugLoc dl; // FIXME: this is nowhere 448 if (I != MF->end() && 449 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 450 MachineBasicBlock *NextBB = I; 451 if (TBB == NextBB && !Cond.empty() && !FBB) { 452 if (!TII->ReverseBranchCondition(Cond)) { 453 TII->RemoveBranch(*CurMBB); 454 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl); 455 return; 456 } 457 } 458 } 459 TII->InsertBranch(*CurMBB, SuccBB, NULL, 460 SmallVector<MachineOperand, 0>(), dl); 461} 462 463bool 464BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 465 if (getHash() < o.getHash()) 466 return true; 467 else if (getHash() > o.getHash()) 468 return false; 469 else if (getBlock()->getNumber() < o.getBlock()->getNumber()) 470 return true; 471 else if (getBlock()->getNumber() > o.getBlock()->getNumber()) 472 return false; 473 else { 474 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 475 // an object with itself. 476#ifndef _GLIBCXX_DEBUG 477 llvm_unreachable("Predecessor appears twice"); 478#endif 479 return false; 480 } 481} 482 483/// CountTerminators - Count the number of terminators in the given 484/// block and set I to the position of the first non-terminator, if there 485/// is one, or MBB->end() otherwise. 486static unsigned CountTerminators(MachineBasicBlock *MBB, 487 MachineBasicBlock::iterator &I) { 488 I = MBB->end(); 489 unsigned NumTerms = 0; 490 for (;;) { 491 if (I == MBB->begin()) { 492 I = MBB->end(); 493 break; 494 } 495 --I; 496 if (!I->getDesc().isTerminator()) break; 497 ++NumTerms; 498 } 499 return NumTerms; 500} 501 502/// ProfitableToMerge - Check if two machine basic blocks have a common tail 503/// and decide if it would be profitable to merge those tails. Return the 504/// length of the common tail and iterators to the first common instruction 505/// in each block. 506static bool ProfitableToMerge(MachineBasicBlock *MBB1, 507 MachineBasicBlock *MBB2, 508 unsigned minCommonTailLength, 509 unsigned &CommonTailLen, 510 MachineBasicBlock::iterator &I1, 511 MachineBasicBlock::iterator &I2, 512 MachineBasicBlock *SuccBB, 513 MachineBasicBlock *PredBB) { 514 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 515 MachineFunction *MF = MBB1->getParent(); 516 517 if (CommonTailLen == 0) 518 return false; 519 520 // It's almost always profitable to merge any number of non-terminator 521 // instructions with the block that falls through into the common successor. 522 if (MBB1 == PredBB || MBB2 == PredBB) { 523 MachineBasicBlock::iterator I; 524 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 525 if (CommonTailLen > NumTerms) 526 return true; 527 } 528 529 // If one of the blocks can be completely merged and happens to be in 530 // a position where the other could fall through into it, merge any number 531 // of instructions, because it can be done without a branch. 532 // TODO: If the blocks are not adjacent, move one of them so that they are? 533 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 534 return true; 535 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 536 return true; 537 538 // If both blocks have an unconditional branch temporarily stripped out, 539 // count that as an additional common instruction for the following 540 // heuristics. 541 unsigned EffectiveTailLen = CommonTailLen; 542 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 543 !MBB1->back().getDesc().isBarrier() && 544 !MBB2->back().getDesc().isBarrier()) 545 ++EffectiveTailLen; 546 547 // Check if the common tail is long enough to be worthwhile. 548 if (EffectiveTailLen >= minCommonTailLength) 549 return true; 550 551 // If we are optimizing for code size, 2 instructions in common is enough if 552 // we don't have to split a block. At worst we will be introducing 1 new 553 // branch instruction, which is likely to be smaller than the 2 554 // instructions that would be deleted in the merge. 555 if (EffectiveTailLen >= 2 && 556 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 557 (I1 == MBB1->begin() || I2 == MBB2->begin())) 558 return true; 559 560 return false; 561} 562 563/// ComputeSameTails - Look through all the blocks in MergePotentials that have 564/// hash CurHash (guaranteed to match the last element). Build the vector 565/// SameTails of all those that have the (same) largest number of instructions 566/// in common of any pair of these blocks. SameTails entries contain an 567/// iterator into MergePotentials (from which the MachineBasicBlock can be 568/// found) and a MachineBasicBlock::iterator into that MBB indicating the 569/// instruction where the matching code sequence begins. 570/// Order of elements in SameTails is the reverse of the order in which 571/// those blocks appear in MergePotentials (where they are not necessarily 572/// consecutive). 573unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 574 unsigned minCommonTailLength, 575 MachineBasicBlock *SuccBB, 576 MachineBasicBlock *PredBB) { 577 unsigned maxCommonTailLength = 0U; 578 SameTails.clear(); 579 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 580 MPIterator HighestMPIter = prior(MergePotentials.end()); 581 for (MPIterator CurMPIter = prior(MergePotentials.end()), 582 B = MergePotentials.begin(); 583 CurMPIter != B && CurMPIter->getHash() == CurHash; 584 --CurMPIter) { 585 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 586 unsigned CommonTailLen; 587 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 588 minCommonTailLength, 589 CommonTailLen, TrialBBI1, TrialBBI2, 590 SuccBB, PredBB)) { 591 if (CommonTailLen > maxCommonTailLength) { 592 SameTails.clear(); 593 maxCommonTailLength = CommonTailLen; 594 HighestMPIter = CurMPIter; 595 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 596 } 597 if (HighestMPIter == CurMPIter && 598 CommonTailLen == maxCommonTailLength) 599 SameTails.push_back(SameTailElt(I, TrialBBI2)); 600 } 601 if (I == B) 602 break; 603 } 604 } 605 return maxCommonTailLength; 606} 607 608/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 609/// MergePotentials, restoring branches at ends of blocks as appropriate. 610void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 611 MachineBasicBlock *SuccBB, 612 MachineBasicBlock *PredBB) { 613 MPIterator CurMPIter, B; 614 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 615 CurMPIter->getHash() == CurHash; 616 --CurMPIter) { 617 // Put the unconditional branch back, if we need one. 618 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 619 if (SuccBB && CurMBB != PredBB) 620 FixTail(CurMBB, SuccBB, TII); 621 if (CurMPIter == B) 622 break; 623 } 624 if (CurMPIter->getHash() != CurHash) 625 CurMPIter++; 626 MergePotentials.erase(CurMPIter, MergePotentials.end()); 627} 628 629/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 630/// only of the common tail. Create a block that does by splitting one. 631unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 632 unsigned maxCommonTailLength) { 633 unsigned commonTailIndex = 0; 634 unsigned TimeEstimate = ~0U; 635 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 636 // Use PredBB if possible; that doesn't require a new branch. 637 if (SameTails[i].getBlock() == PredBB) { 638 commonTailIndex = i; 639 break; 640 } 641 // Otherwise, make a (fairly bogus) choice based on estimate of 642 // how long it will take the various blocks to execute. 643 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 644 SameTails[i].getTailStartPos()); 645 if (t <= TimeEstimate) { 646 TimeEstimate = t; 647 commonTailIndex = i; 648 } 649 } 650 651 MachineBasicBlock::iterator BBI = 652 SameTails[commonTailIndex].getTailStartPos(); 653 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 654 655 // If the common tail includes any debug info we will take it pretty 656 // randomly from one of the inputs. Might be better to remove it? 657 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 658 << maxCommonTailLength); 659 660 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 661 SameTails[commonTailIndex].setBlock(newMBB); 662 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 663 664 // If we split PredBB, newMBB is the new predecessor. 665 if (PredBB == MBB) 666 PredBB = newMBB; 667 668 return commonTailIndex; 669} 670 671// See if any of the blocks in MergePotentials (which all have a common single 672// successor, or all have no successor) can be tail-merged. If there is a 673// successor, any blocks in MergePotentials that are not tail-merged and 674// are not immediately before Succ must have an unconditional branch to 675// Succ added (but the predecessor/successor lists need no adjustment). 676// The lone predecessor of Succ that falls through into Succ, 677// if any, is given in PredBB. 678 679bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 680 MachineBasicBlock *PredBB) { 681 bool MadeChange = false; 682 683 // Except for the special cases below, tail-merge if there are at least 684 // this many instructions in common. 685 unsigned minCommonTailLength = TailMergeSize; 686 687 DEBUG(dbgs() << "\nTryTailMergeBlocks: "; 688 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 689 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 690 << (i == e-1 ? "" : ", "); 691 dbgs() << "\n"; 692 if (SuccBB) { 693 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 694 if (PredBB) 695 dbgs() << " which has fall-through from BB#" 696 << PredBB->getNumber() << "\n"; 697 } 698 dbgs() << "Looking for common tails of at least " 699 << minCommonTailLength << " instruction" 700 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 701 ); 702 703 // Sort by hash value so that blocks with identical end sequences sort 704 // together. 705 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 706 707 // Walk through equivalence sets looking for actual exact matches. 708 while (MergePotentials.size() > 1) { 709 unsigned CurHash = MergePotentials.back().getHash(); 710 711 // Build SameTails, identifying the set of blocks with this hash code 712 // and with the maximum number of instructions in common. 713 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 714 minCommonTailLength, 715 SuccBB, PredBB); 716 717 // If we didn't find any pair that has at least minCommonTailLength 718 // instructions in common, remove all blocks with this hash code and retry. 719 if (SameTails.empty()) { 720 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 721 continue; 722 } 723 724 // If one of the blocks is the entire common tail (and not the entry 725 // block, which we can't jump to), we can treat all blocks with this same 726 // tail at once. Use PredBB if that is one of the possibilities, as that 727 // will not introduce any extra branches. 728 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 729 getParent()->begin(); 730 unsigned commonTailIndex = SameTails.size(); 731 // If there are two blocks, check to see if one can be made to fall through 732 // into the other. 733 if (SameTails.size() == 2 && 734 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 735 SameTails[1].tailIsWholeBlock()) 736 commonTailIndex = 1; 737 else if (SameTails.size() == 2 && 738 SameTails[1].getBlock()->isLayoutSuccessor( 739 SameTails[0].getBlock()) && 740 SameTails[0].tailIsWholeBlock()) 741 commonTailIndex = 0; 742 else { 743 // Otherwise just pick one, favoring the fall-through predecessor if 744 // there is one. 745 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 746 MachineBasicBlock *MBB = SameTails[i].getBlock(); 747 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 748 continue; 749 if (MBB == PredBB) { 750 commonTailIndex = i; 751 break; 752 } 753 if (SameTails[i].tailIsWholeBlock()) 754 commonTailIndex = i; 755 } 756 } 757 758 if (commonTailIndex == SameTails.size() || 759 (SameTails[commonTailIndex].getBlock() == PredBB && 760 !SameTails[commonTailIndex].tailIsWholeBlock())) { 761 // None of the blocks consist entirely of the common tail. 762 // Split a block so that one does. 763 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); 764 } 765 766 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 767 // MBB is common tail. Adjust all other BB's to jump to this one. 768 // Traversal must be forwards so erases work. 769 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber() 770 << " for "); 771 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 772 if (commonTailIndex == i) 773 continue; 774 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() 775 << (i == e-1 ? "" : ", ")); 776 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 777 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 778 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 779 MergePotentials.erase(SameTails[i].getMPIter()); 780 } 781 DEBUG(dbgs() << "\n"); 782 // We leave commonTailIndex in the worklist in case there are other blocks 783 // that match it with a smaller number of instructions. 784 MadeChange = true; 785 } 786 return MadeChange; 787} 788 789bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 790 791 if (!EnableTailMerge) return false; 792 793 bool MadeChange = false; 794 795 // First find blocks with no successors. 796 MergePotentials.clear(); 797 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 798 if (I->succ_empty()) 799 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I)); 800 } 801 802 // See if we can do any tail merging on those. 803 if (MergePotentials.size() < TailMergeThreshold && 804 MergePotentials.size() >= 2) 805 MadeChange |= TryTailMergeBlocks(NULL, NULL); 806 807 // Look at blocks (IBB) with multiple predecessors (PBB). 808 // We change each predecessor to a canonical form, by 809 // (1) temporarily removing any unconditional branch from the predecessor 810 // to IBB, and 811 // (2) alter conditional branches so they branch to the other block 812 // not IBB; this may require adding back an unconditional branch to IBB 813 // later, where there wasn't one coming in. E.g. 814 // Bcc IBB 815 // fallthrough to QBB 816 // here becomes 817 // Bncc QBB 818 // with a conceptual B to IBB after that, which never actually exists. 819 // With those changes, we see whether the predecessors' tails match, 820 // and merge them if so. We change things out of canonical form and 821 // back to the way they were later in the process. (OptimizeBranches 822 // would undo some of this, but we can't use it, because we'd get into 823 // a compile-time infinite loop repeatedly doing and undoing the same 824 // transformations.) 825 826 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 827 I != E; ++I) { 828 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 829 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 830 MachineBasicBlock *IBB = I; 831 MachineBasicBlock *PredBB = prior(I); 832 MergePotentials.clear(); 833 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 834 E2 = I->pred_end(); 835 P != E2; ++P) { 836 MachineBasicBlock *PBB = *P; 837 // Skip blocks that loop to themselves, can't tail merge these. 838 if (PBB == IBB) 839 continue; 840 // Visit each predecessor only once. 841 if (!UniquePreds.insert(PBB)) 842 continue; 843 MachineBasicBlock *TBB = 0, *FBB = 0; 844 SmallVector<MachineOperand, 4> Cond; 845 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 846 // Failing case: IBB is the target of a cbr, and 847 // we cannot reverse the branch. 848 SmallVector<MachineOperand, 4> NewCond(Cond); 849 if (!Cond.empty() && TBB == IBB) { 850 if (TII->ReverseBranchCondition(NewCond)) 851 continue; 852 // This is the QBB case described above 853 if (!FBB) 854 FBB = llvm::next(MachineFunction::iterator(PBB)); 855 } 856 // Failing case: the only way IBB can be reached from PBB is via 857 // exception handling. Happens for landing pads. Would be nice 858 // to have a bit in the edge so we didn't have to do all this. 859 if (IBB->isLandingPad()) { 860 MachineFunction::iterator IP = PBB; IP++; 861 MachineBasicBlock *PredNextBB = NULL; 862 if (IP != MF.end()) 863 PredNextBB = IP; 864 if (TBB == NULL) { 865 if (IBB != PredNextBB) // fallthrough 866 continue; 867 } else if (FBB) { 868 if (TBB != IBB && FBB != IBB) // cbr then ubr 869 continue; 870 } else if (Cond.empty()) { 871 if (TBB != IBB) // ubr 872 continue; 873 } else { 874 if (TBB != IBB && IBB != PredNextBB) // cbr 875 continue; 876 } 877 } 878 // Remove the unconditional branch at the end, if any. 879 if (TBB && (Cond.empty() || FBB)) { 880 DebugLoc dl; // FIXME: this is nowhere 881 TII->RemoveBranch(*PBB); 882 if (!Cond.empty()) 883 // reinsert conditional branch only, for now 884 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl); 885 } 886 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P)); 887 } 888 } 889 if (MergePotentials.size() >= 2) 890 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 891 // Reinsert an unconditional branch if needed. 892 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks. 893 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 894 if (MergePotentials.size() == 1 && 895 MergePotentials.begin()->getBlock() != PredBB) 896 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 897 } 898 } 899 return MadeChange; 900} 901 902//===----------------------------------------------------------------------===// 903// Branch Optimization 904//===----------------------------------------------------------------------===// 905 906bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 907 bool MadeChange = false; 908 909 // Make sure blocks are numbered in order 910 MF.RenumberBlocks(); 911 912 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 913 MachineBasicBlock *MBB = I++; 914 MadeChange |= OptimizeBlock(MBB); 915 916 // If it is dead, remove it. 917 if (MBB->pred_empty()) { 918 RemoveDeadBlock(MBB); 919 MadeChange = true; 920 ++NumDeadBlocks; 921 } 922 } 923 return MadeChange; 924} 925 926// Blocks should be considered empty if they contain only debug info; 927// else the debug info would affect codegen. 928static bool IsEmptyBlock(MachineBasicBlock *MBB) { 929 if (MBB->empty()) 930 return true; 931 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end(); 932 MBBI!=MBBE; ++MBBI) { 933 if (!MBBI->isDebugValue()) 934 return false; 935 } 936 return true; 937} 938 939// Blocks with only debug info and branches should be considered the same 940// as blocks with only branches. 941static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { 942 MachineBasicBlock::iterator MBBI, MBBE; 943 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) { 944 if (!MBBI->isDebugValue()) 945 break; 946 } 947 return (MBBI->getDesc().isBranch()); 948} 949 950/// IsBetterFallthrough - Return true if it would be clearly better to 951/// fall-through to MBB1 than to fall through into MBB2. This has to return 952/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 953/// result in infinite loops. 954static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 955 MachineBasicBlock *MBB2) { 956 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 957 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 958 // optimize branches that branch to either a return block or an assert block 959 // into a fallthrough to the return. 960 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false; 961 962 // If there is a clear successor ordering we make sure that one block 963 // will fall through to the next 964 if (MBB1->isSuccessor(MBB2)) return true; 965 if (MBB2->isSuccessor(MBB1)) return false; 966 967 // Neither block consists entirely of debug info (per IsEmptyBlock check), 968 // so we needn't test for falling off the beginning here. 969 MachineBasicBlock::iterator MBB1I = --MBB1->end(); 970 while (MBB1I->isDebugValue()) 971 --MBB1I; 972 MachineBasicBlock::iterator MBB2I = --MBB2->end(); 973 while (MBB2I->isDebugValue()) 974 --MBB2I; 975 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 976} 977 978/// OptimizeBlock - Analyze and optimize control flow related to the specified 979/// block. This is never called on the entry block. 980bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 981 bool MadeChange = false; 982 MachineFunction &MF = *MBB->getParent(); 983 DebugLoc dl; // FIXME: this is nowhere 984ReoptimizeBlock: 985 986 MachineFunction::iterator FallThrough = MBB; 987 ++FallThrough; 988 989 // If this block is empty, make everyone use its fall-through, not the block 990 // explicitly. Landing pads should not do this since the landing-pad table 991 // points to this block. Blocks with their addresses taken shouldn't be 992 // optimized away. 993 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 994 // Dead block? Leave for cleanup later. 995 if (MBB->pred_empty()) return MadeChange; 996 997 if (FallThrough == MF.end()) { 998 // TODO: Simplify preds to not branch here if possible! 999 } else { 1000 // Rewrite all predecessors of the old block to go to the fallthrough 1001 // instead. 1002 while (!MBB->pred_empty()) { 1003 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1004 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1005 } 1006 // If MBB was the target of a jump table, update jump tables to go to the 1007 // fallthrough instead. 1008 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1009 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough); 1010 MadeChange = true; 1011 } 1012 return MadeChange; 1013 } 1014 1015 // Check to see if we can simplify the terminator of the block before this 1016 // one. 1017 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1018 1019 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1020 SmallVector<MachineOperand, 4> PriorCond; 1021 bool PriorUnAnalyzable = 1022 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1023 if (!PriorUnAnalyzable) { 1024 // If the CFG for the prior block has extra edges, remove them. 1025 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1026 !PriorCond.empty()); 1027 1028 // If the previous branch is conditional and both conditions go to the same 1029 // destination, remove the branch, replacing it with an unconditional one or 1030 // a fall-through. 1031 if (PriorTBB && PriorTBB == PriorFBB) { 1032 TII->RemoveBranch(PrevBB); 1033 PriorCond.clear(); 1034 if (PriorTBB != MBB) 1035 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1036 MadeChange = true; 1037 ++NumBranchOpts; 1038 goto ReoptimizeBlock; 1039 } 1040 1041 // If the previous block unconditionally falls through to this block and 1042 // this block has no other predecessors, move the contents of this block 1043 // into the prior block. This doesn't usually happen when SimplifyCFG 1044 // has been used, but it can happen if tail merging splits a fall-through 1045 // predecessor of a block. 1046 // This has to check PrevBB->succ_size() because EH edges are ignored by 1047 // AnalyzeBranch. 1048 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1049 PrevBB.succ_size() == 1 && 1050 !MBB->hasAddressTaken()) { 1051 DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1052 << "From MBB: " << *MBB); 1053 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1054 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1055 assert(PrevBB.succ_empty()); 1056 PrevBB.transferSuccessors(MBB); 1057 MadeChange = true; 1058 return MadeChange; 1059 } 1060 1061 // If the previous branch *only* branches to *this* block (conditional or 1062 // not) remove the branch. 1063 if (PriorTBB == MBB && PriorFBB == 0) { 1064 TII->RemoveBranch(PrevBB); 1065 MadeChange = true; 1066 ++NumBranchOpts; 1067 goto ReoptimizeBlock; 1068 } 1069 1070 // If the prior block branches somewhere else on the condition and here if 1071 // the condition is false, remove the uncond second branch. 1072 if (PriorFBB == MBB) { 1073 TII->RemoveBranch(PrevBB); 1074 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1075 MadeChange = true; 1076 ++NumBranchOpts; 1077 goto ReoptimizeBlock; 1078 } 1079 1080 // If the prior block branches here on true and somewhere else on false, and 1081 // if the branch condition is reversible, reverse the branch to create a 1082 // fall-through. 1083 if (PriorTBB == MBB) { 1084 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1085 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1086 TII->RemoveBranch(PrevBB); 1087 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl); 1088 MadeChange = true; 1089 ++NumBranchOpts; 1090 goto ReoptimizeBlock; 1091 } 1092 } 1093 1094 // If this block has no successors (e.g. it is a return block or ends with 1095 // a call to a no-return function like abort or __cxa_throw) and if the pred 1096 // falls through into this block, and if it would otherwise fall through 1097 // into the block after this, move this block to the end of the function. 1098 // 1099 // We consider it more likely that execution will stay in the function (e.g. 1100 // due to loops) than it is to exit it. This asserts in loops etc, moving 1101 // the assert condition out of the loop body. 1102 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1103 MachineFunction::iterator(PriorTBB) == FallThrough && 1104 !MBB->canFallThrough()) { 1105 bool DoTransform = true; 1106 1107 // We have to be careful that the succs of PredBB aren't both no-successor 1108 // blocks. If neither have successors and if PredBB is the second from 1109 // last block in the function, we'd just keep swapping the two blocks for 1110 // last. Only do the swap if one is clearly better to fall through than 1111 // the other. 1112 if (FallThrough == --MF.end() && 1113 !IsBetterFallthrough(PriorTBB, MBB)) 1114 DoTransform = false; 1115 1116 if (DoTransform) { 1117 // Reverse the branch so we will fall through on the previous true cond. 1118 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1119 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1120 DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1121 << "To make fallthrough to: " << *PriorTBB << "\n"); 1122 1123 TII->RemoveBranch(PrevBB); 1124 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl); 1125 1126 // Move this block to the end of the function. 1127 MBB->moveAfter(--MF.end()); 1128 MadeChange = true; 1129 ++NumBranchOpts; 1130 return MadeChange; 1131 } 1132 } 1133 } 1134 } 1135 1136 // Analyze the branch in the current block. 1137 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1138 SmallVector<MachineOperand, 4> CurCond; 1139 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1140 if (!CurUnAnalyzable) { 1141 // If the CFG for the prior block has extra edges, remove them. 1142 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1143 1144 // If this is a two-way branch, and the FBB branches to this block, reverse 1145 // the condition so the single-basic-block loop is faster. Instead of: 1146 // Loop: xxx; jcc Out; jmp Loop 1147 // we want: 1148 // Loop: xxx; jncc Loop; jmp Out 1149 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1150 SmallVector<MachineOperand, 4> NewCond(CurCond); 1151 if (!TII->ReverseBranchCondition(NewCond)) { 1152 TII->RemoveBranch(*MBB); 1153 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); 1154 MadeChange = true; 1155 ++NumBranchOpts; 1156 goto ReoptimizeBlock; 1157 } 1158 } 1159 1160 // If this branch is the only thing in its block, see if we can forward 1161 // other blocks across it. 1162 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1163 IsBranchOnlyBlock(MBB) && CurTBB != MBB && 1164 !MBB->hasAddressTaken()) { 1165 // This block may contain just an unconditional branch. Because there can 1166 // be 'non-branch terminators' in the block, try removing the branch and 1167 // then seeing if the block is empty. 1168 TII->RemoveBranch(*MBB); 1169 // If the only things remaining in the block are debug info, remove these 1170 // as well, so this will behave the same as an empty block in non-debug 1171 // mode. 1172 if (!MBB->empty()) { 1173 bool NonDebugInfoFound = false; 1174 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 1175 I != E; ++I) { 1176 if (!I->isDebugValue()) { 1177 NonDebugInfoFound = true; 1178 break; 1179 } 1180 } 1181 if (!NonDebugInfoFound) 1182 // Make the block empty, losing the debug info (we could probably 1183 // improve this in some cases.) 1184 MBB->erase(MBB->begin(), MBB->end()); 1185 } 1186 // If this block is just an unconditional branch to CurTBB, we can 1187 // usually completely eliminate the block. The only case we cannot 1188 // completely eliminate the block is when the block before this one 1189 // falls through into MBB and we can't understand the prior block's branch 1190 // condition. 1191 if (MBB->empty()) { 1192 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1193 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1194 !PrevBB.isSuccessor(MBB)) { 1195 // If the prior block falls through into us, turn it into an 1196 // explicit branch to us to make updates simpler. 1197 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1198 PriorTBB != MBB && PriorFBB != MBB) { 1199 if (PriorTBB == 0) { 1200 assert(PriorCond.empty() && PriorFBB == 0 && 1201 "Bad branch analysis"); 1202 PriorTBB = MBB; 1203 } else { 1204 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1205 PriorFBB = MBB; 1206 } 1207 TII->RemoveBranch(PrevBB); 1208 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, dl); 1209 } 1210 1211 // Iterate through all the predecessors, revectoring each in-turn. 1212 size_t PI = 0; 1213 bool DidChange = false; 1214 bool HasBranchToSelf = false; 1215 while(PI != MBB->pred_size()) { 1216 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1217 if (PMBB == MBB) { 1218 // If this block has an uncond branch to itself, leave it. 1219 ++PI; 1220 HasBranchToSelf = true; 1221 } else { 1222 DidChange = true; 1223 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1224 // If this change resulted in PMBB ending in a conditional 1225 // branch where both conditions go to the same destination, 1226 // change this to an unconditional branch (and fix the CFG). 1227 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1228 SmallVector<MachineOperand, 4> NewCurCond; 1229 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1230 NewCurFBB, NewCurCond, true); 1231 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1232 TII->RemoveBranch(*PMBB); 1233 NewCurCond.clear(); 1234 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, dl); 1235 MadeChange = true; 1236 ++NumBranchOpts; 1237 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1238 } 1239 } 1240 } 1241 1242 // Change any jumptables to go to the new MBB. 1243 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1244 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); 1245 if (DidChange) { 1246 ++NumBranchOpts; 1247 MadeChange = true; 1248 if (!HasBranchToSelf) return MadeChange; 1249 } 1250 } 1251 } 1252 1253 // Add the branch back if the block is more than just an uncond branch. 1254 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl); 1255 } 1256 } 1257 1258 // If the prior block doesn't fall through into this block, and if this 1259 // block doesn't fall through into some other block, see if we can find a 1260 // place to move this block where a fall-through will happen. 1261 if (!PrevBB.canFallThrough()) { 1262 1263 // Now we know that there was no fall-through into this block, check to 1264 // see if it has a fall-through into its successor. 1265 bool CurFallsThru = MBB->canFallThrough(); 1266 1267 if (!MBB->isLandingPad()) { 1268 // Check all the predecessors of this block. If one of them has no fall 1269 // throughs, move this block right after it. 1270 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1271 E = MBB->pred_end(); PI != E; ++PI) { 1272 // Analyze the branch at the end of the pred. 1273 MachineBasicBlock *PredBB = *PI; 1274 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1275 MachineBasicBlock *PredTBB = 0, *PredFBB = 0; 1276 SmallVector<MachineOperand, 4> PredCond; 1277 if (PredBB != MBB && !PredBB->canFallThrough() && 1278 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1279 && (!CurFallsThru || !CurTBB || !CurFBB) 1280 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1281 // If the current block doesn't fall through, just move it. 1282 // If the current block can fall through and does not end with a 1283 // conditional branch, we need to append an unconditional jump to 1284 // the (current) next block. To avoid a possible compile-time 1285 // infinite loop, move blocks only backward in this case. 1286 // Also, if there are already 2 branches here, we cannot add a third; 1287 // this means we have the case 1288 // Bcc next 1289 // B elsewhere 1290 // next: 1291 if (CurFallsThru) { 1292 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); 1293 CurCond.clear(); 1294 TII->InsertBranch(*MBB, NextBB, 0, CurCond, dl); 1295 } 1296 MBB->moveAfter(PredBB); 1297 MadeChange = true; 1298 goto ReoptimizeBlock; 1299 } 1300 } 1301 } 1302 1303 if (!CurFallsThru) { 1304 // Check all successors to see if we can move this block before it. 1305 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1306 E = MBB->succ_end(); SI != E; ++SI) { 1307 // Analyze the branch at the end of the block before the succ. 1308 MachineBasicBlock *SuccBB = *SI; 1309 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1310 1311 // If this block doesn't already fall-through to that successor, and if 1312 // the succ doesn't already have a block that can fall through into it, 1313 // and if the successor isn't an EH destination, we can arrange for the 1314 // fallthrough to happen. 1315 if (SuccBB != MBB && &*SuccPrev != MBB && 1316 !SuccPrev->canFallThrough() && !CurUnAnalyzable && 1317 !SuccBB->isLandingPad()) { 1318 MBB->moveBefore(SuccBB); 1319 MadeChange = true; 1320 goto ReoptimizeBlock; 1321 } 1322 } 1323 1324 // Okay, there is no really great place to put this block. If, however, 1325 // the block before this one would be a fall-through if this block were 1326 // removed, move this block to the end of the function. 1327 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0; 1328 SmallVector<MachineOperand, 4> PrevCond; 1329 if (FallThrough != MF.end() && 1330 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1331 PrevBB.isSuccessor(FallThrough)) { 1332 MBB->moveAfter(--MF.end()); 1333 MadeChange = true; 1334 return MadeChange; 1335 } 1336 } 1337 } 1338 1339 return MadeChange; 1340} 1341