BranchFolding.cpp revision cd6a8e23207f276f915f20cad9f3905a862df5d9
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"); 44static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 45 cl::init(cl::BOU_UNSET), cl::Hidden); 46// Throttle for huge numbers of predecessors (compile speed problems) 47static cl::opt<unsigned> 48TailMergeThreshold("tail-merge-threshold", 49 cl::desc("Max number of predecessors to consider tail merging"), 50 cl::init(150), cl::Hidden); 51 52// Heuristic for tail merging (and, inversely, tail duplication). 53// TODO: This should be replaced with a target query. 54static cl::opt<unsigned> 55TailMergeSize("tail-merge-size", 56 cl::desc("Min number of instructions to consider tail merging"), 57 cl::init(3), cl::Hidden); 58 59namespace { 60 /// BranchFolderPass - Wrap branch folder in a machine function pass. 61 class BranchFolderPass : public MachineFunctionPass, 62 public BranchFolder { 63 public: 64 static char ID; 65 explicit BranchFolderPass(bool defaultEnableTailMerge) 66 : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {} 67 68 virtual bool runOnMachineFunction(MachineFunction &MF); 69 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 70 }; 71} 72 73char BranchFolderPass::ID = 0; 74 75FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 76 return new BranchFolderPass(DefaultEnableTailMerge); 77} 78 79bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { 80 return OptimizeFunction(MF, 81 MF.getTarget().getInstrInfo(), 82 MF.getTarget().getRegisterInfo(), 83 getAnalysisIfAvailable<MachineModuleInfo>()); 84} 85 86 87BranchFolder::BranchFolder(bool defaultEnableTailMerge) { 88 switch (FlagEnableTailMerge) { 89 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 90 case cl::BOU_TRUE: EnableTailMerge = true; break; 91 case cl::BOU_FALSE: EnableTailMerge = false; break; 92 } 93} 94 95/// RemoveDeadBlock - Remove the specified dead machine basic block from the 96/// function, updating the CFG. 97void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 98 assert(MBB->pred_empty() && "MBB must be dead!"); 99 DEBUG(errs() << "\nRemoving MBB: " << *MBB); 100 101 MachineFunction *MF = MBB->getParent(); 102 // drop all successors. 103 while (!MBB->succ_empty()) 104 MBB->removeSuccessor(MBB->succ_end()-1); 105 106 // If there are any labels in the basic block, unregister them from 107 // MachineModuleInfo. 108 if (MMI && !MBB->empty()) { 109 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 110 I != E; ++I) { 111 if (I->isLabel()) 112 // The label ID # is always operand #0, an immediate. 113 MMI->InvalidateLabel(I->getOperand(0).getImm()); 114 } 115 } 116 117 // Remove the block. 118 MF->erase(MBB); 119} 120 121/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 122/// followed by terminators, and if the implicitly defined registers are not 123/// used by the terminators, remove those implicit_def's. e.g. 124/// BB1: 125/// r0 = implicit_def 126/// r1 = implicit_def 127/// br 128/// This block can be optimized away later if the implicit instructions are 129/// removed. 130bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 131 SmallSet<unsigned, 4> ImpDefRegs; 132 MachineBasicBlock::iterator I = MBB->begin(); 133 while (I != MBB->end()) { 134 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) 135 break; 136 unsigned Reg = I->getOperand(0).getReg(); 137 ImpDefRegs.insert(Reg); 138 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg); 139 unsigned SubReg = *SubRegs; ++SubRegs) 140 ImpDefRegs.insert(SubReg); 141 ++I; 142 } 143 if (ImpDefRegs.empty()) 144 return false; 145 146 MachineBasicBlock::iterator FirstTerm = I; 147 while (I != MBB->end()) { 148 if (!TII->isUnpredicatedTerminator(I)) 149 return false; 150 // See if it uses any of the implicitly defined registers. 151 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 152 MachineOperand &MO = I->getOperand(i); 153 if (!MO.isReg() || !MO.isUse()) 154 continue; 155 unsigned Reg = MO.getReg(); 156 if (ImpDefRegs.count(Reg)) 157 return false; 158 } 159 ++I; 160 } 161 162 I = MBB->begin(); 163 while (I != FirstTerm) { 164 MachineInstr *ImpDefMI = &*I; 165 ++I; 166 MBB->erase(ImpDefMI); 167 } 168 169 return true; 170} 171 172/// OptimizeFunction - Perhaps branch folding, tail merging and other 173/// CFG optimizations on the given function. 174bool BranchFolder::OptimizeFunction(MachineFunction &MF, 175 const TargetInstrInfo *tii, 176 const TargetRegisterInfo *tri, 177 MachineModuleInfo *mmi) { 178 if (!tii) return false; 179 180 TII = tii; 181 TRI = tri; 182 MMI = mmi; 183 184 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 185 186 // Fix CFG. The later algorithms expect it to be right. 187 bool MadeChange = false; 188 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 189 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 190 SmallVector<MachineOperand, 4> Cond; 191 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 192 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 193 MadeChange |= OptimizeImpDefsBlock(MBB); 194 } 195 196 197 bool MadeChangeThisIteration = true; 198 while (MadeChangeThisIteration) { 199 MadeChangeThisIteration = false; 200 MadeChangeThisIteration |= TailMergeBlocks(MF); 201 MadeChangeThisIteration |= OptimizeBranches(MF); 202 MadeChange |= MadeChangeThisIteration; 203 } 204 205 // See if any jump tables have become mergable or dead as the code generator 206 // did its thing. 207 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 208 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 209 if (!JTs.empty()) { 210 // Figure out how these jump tables should be merged. 211 std::vector<unsigned> JTMapping; 212 JTMapping.reserve(JTs.size()); 213 214 // We always keep the 0th jump table. 215 JTMapping.push_back(0); 216 217 // Scan the jump tables, seeing if there are any duplicates. Note that this 218 // is N^2, which should be fixed someday. 219 for (unsigned i = 1, e = JTs.size(); i != e; ++i) { 220 if (JTs[i].MBBs.empty()) 221 JTMapping.push_back(i); 222 else 223 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 224 } 225 226 // If a jump table was merge with another one, walk the function rewriting 227 // references to jump tables to reference the new JT ID's. Keep track of 228 // whether we see a jump table idx, if not, we can delete the JT. 229 BitVector JTIsLive(JTs.size()); 230 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 231 BB != E; ++BB) { 232 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 233 I != E; ++I) 234 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 235 MachineOperand &Op = I->getOperand(op); 236 if (!Op.isJTI()) continue; 237 unsigned NewIdx = JTMapping[Op.getIndex()]; 238 Op.setIndex(NewIdx); 239 240 // Remember that this JT is live. 241 JTIsLive.set(NewIdx); 242 } 243 } 244 245 // Finally, remove dead jump tables. This happens either because the 246 // indirect jump was unreachable (and thus deleted) or because the jump 247 // table was merged with some other one. 248 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 249 if (!JTIsLive.test(i)) { 250 JTI->RemoveJumpTable(i); 251 MadeChange = true; 252 } 253 } 254 255 delete RS; 256 return MadeChange; 257} 258 259//===----------------------------------------------------------------------===// 260// Tail Merging of Blocks 261//===----------------------------------------------------------------------===// 262 263/// HashMachineInstr - Compute a hash value for MI and its operands. 264static unsigned HashMachineInstr(const MachineInstr *MI) { 265 unsigned Hash = MI->getOpcode(); 266 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 267 const MachineOperand &Op = MI->getOperand(i); 268 269 // Merge in bits from the operand if easy. 270 unsigned OperandHash = 0; 271 switch (Op.getType()) { 272 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 273 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 274 case MachineOperand::MO_MachineBasicBlock: 275 OperandHash = Op.getMBB()->getNumber(); 276 break; 277 case MachineOperand::MO_FrameIndex: 278 case MachineOperand::MO_ConstantPoolIndex: 279 case MachineOperand::MO_JumpTableIndex: 280 OperandHash = Op.getIndex(); 281 break; 282 case MachineOperand::MO_GlobalAddress: 283 case MachineOperand::MO_ExternalSymbol: 284 // Global address / external symbol are too hard, don't bother, but do 285 // pull in the offset. 286 OperandHash = Op.getOffset(); 287 break; 288 default: break; 289 } 290 291 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 292 } 293 return Hash; 294} 295 296/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 297/// with no successors, we hash two instructions, because cross-jumping 298/// only saves code when at least two instructions are removed (since a 299/// branch must be inserted). For blocks with a successor, one of the 300/// two blocks to be tail-merged will end with a branch already, so 301/// it gains to cross-jump even for one instruction. 302static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 303 unsigned minCommonTailLength) { 304 MachineBasicBlock::const_iterator I = MBB->end(); 305 if (I == MBB->begin()) 306 return 0; // Empty MBB. 307 308 --I; 309 unsigned Hash = HashMachineInstr(I); 310 311 if (I == MBB->begin() || minCommonTailLength == 1) 312 return Hash; // Single instr MBB. 313 314 --I; 315 // Hash in the second-to-last instruction. 316 Hash ^= HashMachineInstr(I) << 2; 317 return Hash; 318} 319 320/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 321/// of instructions they actually have in common together at their end. Return 322/// iterators for the first shared instruction in each block. 323static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 324 MachineBasicBlock *MBB2, 325 MachineBasicBlock::iterator &I1, 326 MachineBasicBlock::iterator &I2) { 327 I1 = MBB1->end(); 328 I2 = MBB2->end(); 329 330 unsigned TailLen = 0; 331 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 332 --I1; --I2; 333 if (!I1->isIdenticalTo(I2) || 334 // FIXME: This check is dubious. It's used to get around a problem where 335 // people incorrectly expect inline asm directives to remain in the same 336 // relative order. This is untenable because normal compiler 337 // optimizations (like this one) may reorder and/or merge these 338 // directives. 339 I1->getOpcode() == TargetInstrInfo::INLINEASM) { 340 ++I1; ++I2; 341 break; 342 } 343 ++TailLen; 344 } 345 return TailLen; 346} 347 348/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 349/// after it, replacing it with an unconditional branch to NewDest. This 350/// returns true if OldInst's block is modified, false if NewDest is modified. 351void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 352 MachineBasicBlock *NewDest) { 353 MachineBasicBlock *OldBB = OldInst->getParent(); 354 355 // Remove all the old successors of OldBB from the CFG. 356 while (!OldBB->succ_empty()) 357 OldBB->removeSuccessor(OldBB->succ_begin()); 358 359 // Remove all the dead instructions from the end of OldBB. 360 OldBB->erase(OldInst, OldBB->end()); 361 362 // If OldBB isn't immediately before OldBB, insert a branch to it. 363 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 364 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>()); 365 OldBB->addSuccessor(NewDest); 366 ++NumTailMerge; 367} 368 369/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 370/// MBB so that the part before the iterator falls into the part starting at the 371/// iterator. This returns the new MBB. 372MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 373 MachineBasicBlock::iterator BBI1) { 374 MachineFunction &MF = *CurMBB.getParent(); 375 376 // Create the fall-through block. 377 MachineFunction::iterator MBBI = &CurMBB; 378 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 379 CurMBB.getParent()->insert(++MBBI, NewMBB); 380 381 // Move all the successors of this block to the specified block. 382 NewMBB->transferSuccessors(&CurMBB); 383 384 // Add an edge from CurMBB to NewMBB for the fall-through. 385 CurMBB.addSuccessor(NewMBB); 386 387 // Splice the code over. 388 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 389 390 // For targets that use the register scavenger, we must maintain LiveIns. 391 if (RS) { 392 RS->enterBasicBlock(&CurMBB); 393 if (!CurMBB.empty()) 394 RS->forward(prior(CurMBB.end())); 395 BitVector RegsLiveAtExit(TRI->getNumRegs()); 396 RS->getRegsUsed(RegsLiveAtExit, false); 397 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 398 if (RegsLiveAtExit[i]) 399 NewMBB->addLiveIn(i); 400 } 401 402 return NewMBB; 403} 404 405/// EstimateRuntime - Make a rough estimate for how long it will take to run 406/// the specified code. 407static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 408 MachineBasicBlock::iterator E) { 409 unsigned Time = 0; 410 for (; I != E; ++I) { 411 const TargetInstrDesc &TID = I->getDesc(); 412 if (TID.isCall()) 413 Time += 10; 414 else if (TID.mayLoad() || TID.mayStore()) 415 Time += 2; 416 else 417 ++Time; 418 } 419 return Time; 420} 421 422// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 423// branches temporarily for tail merging). In the case where CurMBB ends 424// with a conditional branch to the next block, optimize by reversing the 425// test and conditionally branching to SuccMBB instead. 426static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, 427 const TargetInstrInfo *TII) { 428 MachineFunction *MF = CurMBB->getParent(); 429 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 430 MachineBasicBlock *TBB = 0, *FBB = 0; 431 SmallVector<MachineOperand, 4> Cond; 432 if (I != MF->end() && 433 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 434 MachineBasicBlock *NextBB = I; 435 if (TBB == NextBB && !Cond.empty() && !FBB) { 436 if (!TII->ReverseBranchCondition(Cond)) { 437 TII->RemoveBranch(*CurMBB); 438 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 439 return; 440 } 441 } 442 } 443 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>()); 444} 445 446bool 447BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 448 if (getHash() < o.getHash()) 449 return true; 450 else if (getHash() > o.getHash()) 451 return false; 452 else if (getBlock()->getNumber() < o.getBlock()->getNumber()) 453 return true; 454 else if (getBlock()->getNumber() > o.getBlock()->getNumber()) 455 return false; 456 else { 457 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 458 // an object with itself. 459#ifndef _GLIBCXX_DEBUG 460 llvm_unreachable("Predecessor appears twice"); 461#endif 462 return false; 463 } 464} 465 466/// CountTerminators - Count the number of terminators in the given 467/// block and set I to the position of the first non-terminator, if there 468/// is one, or MBB->end() otherwise. 469static unsigned CountTerminators(MachineBasicBlock *MBB, 470 MachineBasicBlock::iterator &I) { 471 I = MBB->end(); 472 unsigned NumTerms = 0; 473 for (;;) { 474 if (I == MBB->begin()) { 475 I = MBB->end(); 476 break; 477 } 478 --I; 479 if (!I->getDesc().isTerminator()) break; 480 ++NumTerms; 481 } 482 return NumTerms; 483} 484 485/// ProfitableToMerge - Check if two machine basic blocks have a common tail 486/// and decide if it would be profitable to merge those tails. Return the 487/// length of the common tail and iterators to the first common instruction 488/// in each block. 489static bool ProfitableToMerge(MachineBasicBlock *MBB1, 490 MachineBasicBlock *MBB2, 491 unsigned minCommonTailLength, 492 unsigned &CommonTailLen, 493 MachineBasicBlock::iterator &I1, 494 MachineBasicBlock::iterator &I2, 495 MachineBasicBlock *SuccBB, 496 MachineBasicBlock *PredBB) { 497 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 498 MachineFunction *MF = MBB1->getParent(); 499 500 if (CommonTailLen == 0) 501 return false; 502 503 // It's almost always profitable to merge any number of non-terminator 504 // instructions with the block that falls through into the common successor. 505 if (MBB1 == PredBB || MBB2 == PredBB) { 506 MachineBasicBlock::iterator I; 507 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 508 if (CommonTailLen > NumTerms) 509 return true; 510 } 511 512 // If one of the blocks can be completely merged and happens to be in 513 // a position where the other could fall through into it, merge any number 514 // of instructions, because it can be done without a branch. 515 // TODO: If the blocks are not adjacent, move one of them so that they are? 516 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 517 return true; 518 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 519 return true; 520 521 // If both blocks have an unconditional branch temporarily stripped out, 522 // count that as an additional common instruction for the following 523 // heuristics. 524 unsigned EffectiveTailLen = CommonTailLen; 525 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 526 !MBB1->back().getDesc().isBarrier() && 527 !MBB2->back().getDesc().isBarrier()) 528 ++EffectiveTailLen; 529 530 // Check if the common tail is long enough to be worthwhile. 531 if (EffectiveTailLen >= minCommonTailLength) 532 return true; 533 534 // If we are optimizing for code size, 2 instructions in common is enough if 535 // we don't have to split a block. At worst we will be introducing 1 new 536 // branch instruction, which is likely to be smaller than the 2 537 // instructions that would be deleted in the merge. 538 if (EffectiveTailLen >= 2 && 539 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 540 (I1 == MBB1->begin() || I2 == MBB2->begin())) 541 return true; 542 543 return false; 544} 545 546/// ComputeSameTails - Look through all the blocks in MergePotentials that have 547/// hash CurHash (guaranteed to match the last element). Build the vector 548/// SameTails of all those that have the (same) largest number of instructions 549/// in common of any pair of these blocks. SameTails entries contain an 550/// iterator into MergePotentials (from which the MachineBasicBlock can be 551/// found) and a MachineBasicBlock::iterator into that MBB indicating the 552/// instruction where the matching code sequence begins. 553/// Order of elements in SameTails is the reverse of the order in which 554/// those blocks appear in MergePotentials (where they are not necessarily 555/// consecutive). 556unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 557 unsigned minCommonTailLength, 558 MachineBasicBlock *SuccBB, 559 MachineBasicBlock *PredBB) { 560 unsigned maxCommonTailLength = 0U; 561 SameTails.clear(); 562 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 563 MPIterator HighestMPIter = prior(MergePotentials.end()); 564 for (MPIterator CurMPIter = prior(MergePotentials.end()), 565 B = MergePotentials.begin(); 566 CurMPIter != B && CurMPIter->getHash() == CurHash; 567 --CurMPIter) { 568 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 569 unsigned CommonTailLen; 570 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 571 minCommonTailLength, 572 CommonTailLen, TrialBBI1, TrialBBI2, 573 SuccBB, PredBB)) { 574 if (CommonTailLen > maxCommonTailLength) { 575 SameTails.clear(); 576 maxCommonTailLength = CommonTailLen; 577 HighestMPIter = CurMPIter; 578 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 579 } 580 if (HighestMPIter == CurMPIter && 581 CommonTailLen == maxCommonTailLength) 582 SameTails.push_back(SameTailElt(I, TrialBBI2)); 583 } 584 if (I == B) 585 break; 586 } 587 } 588 return maxCommonTailLength; 589} 590 591/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 592/// MergePotentials, restoring branches at ends of blocks as appropriate. 593void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 594 MachineBasicBlock* SuccBB, 595 MachineBasicBlock* PredBB) { 596 MPIterator CurMPIter, B; 597 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 598 CurMPIter->getHash() == CurHash; 599 --CurMPIter) { 600 // Put the unconditional branch back, if we need one. 601 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 602 if (SuccBB && CurMBB != PredBB) 603 FixTail(CurMBB, SuccBB, TII); 604 if (CurMPIter == B) 605 break; 606 } 607 if (CurMPIter->getHash() != CurHash) 608 CurMPIter++; 609 MergePotentials.erase(CurMPIter, MergePotentials.end()); 610} 611 612/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 613/// only of the common tail. Create a block that does by splitting one. 614unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 615 unsigned maxCommonTailLength) { 616 unsigned commonTailIndex = 0; 617 unsigned TimeEstimate = ~0U; 618 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 619 // Use PredBB if possible; that doesn't require a new branch. 620 if (SameTails[i].getBlock() == PredBB) { 621 commonTailIndex = i; 622 break; 623 } 624 // Otherwise, make a (fairly bogus) choice based on estimate of 625 // how long it will take the various blocks to execute. 626 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 627 SameTails[i].getTailStartPos()); 628 if (t <= TimeEstimate) { 629 TimeEstimate = t; 630 commonTailIndex = i; 631 } 632 } 633 634 MachineBasicBlock::iterator BBI = 635 SameTails[commonTailIndex].getTailStartPos(); 636 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 637 638 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 639 << maxCommonTailLength); 640 641 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 642 SameTails[commonTailIndex].setBlock(newMBB); 643 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 644 645 // If we split PredBB, newMBB is the new predecessor. 646 if (PredBB == MBB) 647 PredBB = newMBB; 648 649 return commonTailIndex; 650} 651 652// See if any of the blocks in MergePotentials (which all have a common single 653// successor, or all have no successor) can be tail-merged. If there is a 654// successor, any blocks in MergePotentials that are not tail-merged and 655// are not immediately before Succ must have an unconditional branch to 656// Succ added (but the predecessor/successor lists need no adjustment). 657// The lone predecessor of Succ that falls through into Succ, 658// if any, is given in PredBB. 659 660bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 661 MachineBasicBlock* PredBB) { 662 bool MadeChange = false; 663 664 // Except for the special cases below, tail-merge if there are at least 665 // this many instructions in common. 666 unsigned minCommonTailLength = TailMergeSize; 667 668 // If there's a successor block, there are some cases which don't require 669 // new branching and as such are very likely to be profitable. 670 if (SuccBB) { 671 if (SuccBB->pred_size() == MergePotentials.size() && 672 !MergePotentials[0].getBlock()->empty()) { 673 // If all the predecessors have at least one tail instruction in common, 674 // merging is very likely to be a win since it won't require an increase 675 // in static branches, and it will decrease the static instruction count. 676 bool AllPredsMatch = true; 677 MachineBasicBlock::iterator FirstNonTerm; 678 unsigned MinNumTerms = CountTerminators(MergePotentials[0].getBlock(), 679 FirstNonTerm); 680 if (FirstNonTerm != MergePotentials[0].getBlock()->end()) { 681 for (unsigned i = 1, e = MergePotentials.size(); i != e; ++i) { 682 MachineBasicBlock::iterator OtherFirstNonTerm; 683 unsigned NumTerms = CountTerminators(MergePotentials[0].getBlock(), 684 OtherFirstNonTerm); 685 if (NumTerms < MinNumTerms) 686 MinNumTerms = NumTerms; 687 if (OtherFirstNonTerm == MergePotentials[i].getBlock()->end() || 688 OtherFirstNonTerm->isIdenticalTo(FirstNonTerm)) { 689 AllPredsMatch = false; 690 break; 691 } 692 } 693 694 // If they all have an instruction in common, do any amount of merging. 695 if (AllPredsMatch) 696 minCommonTailLength = MinNumTerms + 1; 697 } 698 } 699 } 700 701 DEBUG(errs() << "\nTryTailMergeBlocks: "; 702 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 703 errs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 704 << (i == e-1 ? "" : ", "); 705 errs() << "\n"; 706 if (SuccBB) { 707 errs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 708 if (PredBB) 709 errs() << " which has fall-through from BB#" 710 << PredBB->getNumber() << "\n"; 711 } 712 errs() << "Looking for common tails of at least " 713 << minCommonTailLength << " instruction" 714 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 715 ); 716 717 // Sort by hash value so that blocks with identical end sequences sort 718 // together. 719 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 720 721 // Walk through equivalence sets looking for actual exact matches. 722 while (MergePotentials.size() > 1) { 723 unsigned CurHash = MergePotentials.back().getHash(); 724 725 // Build SameTails, identifying the set of blocks with this hash code 726 // and with the maximum number of instructions in common. 727 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 728 minCommonTailLength, 729 SuccBB, PredBB); 730 731 // If we didn't find any pair that has at least minCommonTailLength 732 // instructions in common, remove all blocks with this hash code and retry. 733 if (SameTails.empty()) { 734 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 735 continue; 736 } 737 738 // If one of the blocks is the entire common tail (and not the entry 739 // block, which we can't jump to), we can treat all blocks with this same 740 // tail at once. Use PredBB if that is one of the possibilities, as that 741 // will not introduce any extra branches. 742 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 743 getParent()->begin(); 744 unsigned commonTailIndex = SameTails.size(); 745 // If there are two blocks, check to see if one can be made to fall through 746 // into the other. 747 if (SameTails.size() == 2 && 748 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 749 SameTails[1].tailIsWholeBlock()) 750 commonTailIndex = 1; 751 else if (SameTails.size() == 2 && 752 SameTails[1].getBlock()->isLayoutSuccessor( 753 SameTails[0].getBlock()) && 754 SameTails[0].tailIsWholeBlock()) 755 commonTailIndex = 0; 756 else { 757 // Otherwise just pick one, favoring the fall-through predecessor if 758 // there is one. 759 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 760 MachineBasicBlock *MBB = SameTails[i].getBlock(); 761 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 762 continue; 763 if (MBB == PredBB) { 764 commonTailIndex = i; 765 break; 766 } 767 if (SameTails[i].tailIsWholeBlock()) 768 commonTailIndex = i; 769 } 770 } 771 772 if (commonTailIndex == SameTails.size() || 773 (SameTails[commonTailIndex].getBlock() == PredBB && 774 !SameTails[commonTailIndex].tailIsWholeBlock())) { 775 // None of the blocks consist entirely of the common tail. 776 // Split a block so that one does. 777 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); 778 } 779 780 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 781 // MBB is common tail. Adjust all other BB's to jump to this one. 782 // Traversal must be forwards so erases work. 783 DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber() 784 << " for "); 785 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 786 if (commonTailIndex == i) 787 continue; 788 DEBUG(errs() << "BB#" << SameTails[i].getBlock()->getNumber() 789 << (i == e-1 ? "" : ", ")); 790 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 791 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 792 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 793 MergePotentials.erase(SameTails[i].getMPIter()); 794 } 795 DEBUG(errs() << "\n"); 796 // We leave commonTailIndex in the worklist in case there are other blocks 797 // that match it with a smaller number of instructions. 798 MadeChange = true; 799 } 800 return MadeChange; 801} 802 803bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 804 805 if (!EnableTailMerge) return false; 806 807 bool MadeChange = false; 808 809 // First find blocks with no successors. 810 MergePotentials.clear(); 811 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 812 if (I->succ_empty()) 813 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I)); 814 } 815 816 // See if we can do any tail merging on those. 817 if (MergePotentials.size() < TailMergeThreshold && 818 MergePotentials.size() >= 2) 819 MadeChange |= TryTailMergeBlocks(NULL, NULL); 820 821 // Look at blocks (IBB) with multiple predecessors (PBB). 822 // We change each predecessor to a canonical form, by 823 // (1) temporarily removing any unconditional branch from the predecessor 824 // to IBB, and 825 // (2) alter conditional branches so they branch to the other block 826 // not IBB; this may require adding back an unconditional branch to IBB 827 // later, where there wasn't one coming in. E.g. 828 // Bcc IBB 829 // fallthrough to QBB 830 // here becomes 831 // Bncc QBB 832 // with a conceptual B to IBB after that, which never actually exists. 833 // With those changes, we see whether the predecessors' tails match, 834 // and merge them if so. We change things out of canonical form and 835 // back to the way they were later in the process. (OptimizeBranches 836 // would undo some of this, but we can't use it, because we'd get into 837 // a compile-time infinite loop repeatedly doing and undoing the same 838 // transformations.) 839 840 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end(); 841 I != E; ++I) { 842 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 843 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 844 MachineBasicBlock *IBB = I; 845 MachineBasicBlock *PredBB = prior(I); 846 MergePotentials.clear(); 847 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 848 E2 = I->pred_end(); 849 P != E2; ++P) { 850 MachineBasicBlock* PBB = *P; 851 // Skip blocks that loop to themselves, can't tail merge these. 852 if (PBB == IBB) 853 continue; 854 // Visit each predecessor only once. 855 if (!UniquePreds.insert(PBB)) 856 continue; 857 MachineBasicBlock *TBB = 0, *FBB = 0; 858 SmallVector<MachineOperand, 4> Cond; 859 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 860 // Failing case: IBB is the target of a cbr, and 861 // we cannot reverse the branch. 862 SmallVector<MachineOperand, 4> NewCond(Cond); 863 if (!Cond.empty() && TBB == IBB) { 864 if (TII->ReverseBranchCondition(NewCond)) 865 continue; 866 // This is the QBB case described above 867 if (!FBB) 868 FBB = next(MachineFunction::iterator(PBB)); 869 } 870 // Failing case: the only way IBB can be reached from PBB is via 871 // exception handling. Happens for landing pads. Would be nice 872 // to have a bit in the edge so we didn't have to do all this. 873 if (IBB->isLandingPad()) { 874 MachineFunction::iterator IP = PBB; IP++; 875 MachineBasicBlock* PredNextBB = NULL; 876 if (IP != MF.end()) 877 PredNextBB = IP; 878 if (TBB == NULL) { 879 if (IBB != PredNextBB) // fallthrough 880 continue; 881 } else if (FBB) { 882 if (TBB != IBB && FBB != IBB) // cbr then ubr 883 continue; 884 } else if (Cond.empty()) { 885 if (TBB != IBB) // ubr 886 continue; 887 } else { 888 if (TBB != IBB && IBB != PredNextBB) // cbr 889 continue; 890 } 891 } 892 // Remove the unconditional branch at the end, if any. 893 if (TBB && (Cond.empty() || FBB)) { 894 TII->RemoveBranch(*PBB); 895 if (!Cond.empty()) 896 // reinsert conditional branch only, for now 897 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond); 898 } 899 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U), 900 *P)); 901 } 902 } 903 if (MergePotentials.size() >= 2) 904 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 905 // Reinsert an unconditional branch if needed. 906 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks. 907 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 908 if (MergePotentials.size() == 1 && 909 MergePotentials.begin()->getBlock() != PredBB) 910 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 911 } 912 } 913 return MadeChange; 914} 915 916//===----------------------------------------------------------------------===// 917// Branch Optimization 918//===----------------------------------------------------------------------===// 919 920bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 921 bool MadeChange = false; 922 923 // Make sure blocks are numbered in order 924 MF.RenumberBlocks(); 925 926 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 927 MachineBasicBlock *MBB = I++; 928 MadeChange |= OptimizeBlock(MBB); 929 930 // If it is dead, remove it. 931 if (MBB->pred_empty()) { 932 RemoveDeadBlock(MBB); 933 MadeChange = true; 934 ++NumDeadBlocks; 935 } 936 } 937 return MadeChange; 938} 939 940 941/// CanFallThrough - Return true if the specified block (with the specified 942/// branch condition) can implicitly transfer control to the block after it by 943/// falling off the end of it. This should return false if it can reach the 944/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 945/// 946/// True is a conservative answer. 947/// 948bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 949 bool BranchUnAnalyzable, 950 MachineBasicBlock *TBB, 951 MachineBasicBlock *FBB, 952 const SmallVectorImpl<MachineOperand> &Cond) { 953 MachineFunction::iterator Fallthrough = CurBB; 954 ++Fallthrough; 955 // If FallthroughBlock is off the end of the function, it can't fall through. 956 if (Fallthrough == CurBB->getParent()->end()) 957 return false; 958 959 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 960 if (!CurBB->isSuccessor(Fallthrough)) 961 return false; 962 963 // If we couldn't analyze the branch, examine the last instruction. 964 // If the block doesn't end in a known control barrier, assume fallthrough 965 // is possible. The isPredicable check is needed because this code can be 966 // called during IfConversion, where an instruction which is normally a 967 // Barrier is predicated and thus no longer an actual control barrier. This 968 // is over-conservative though, because if an instruction isn't actually 969 // predicated we could still treat it like a barrier. 970 if (BranchUnAnalyzable) 971 return CurBB->empty() || !CurBB->back().getDesc().isBarrier() || 972 CurBB->back().getDesc().isPredicable(); 973 974 // If there is no branch, control always falls through. 975 if (TBB == 0) return true; 976 977 // If there is some explicit branch to the fallthrough block, it can obviously 978 // reach, even though the branch should get folded to fall through implicitly. 979 if (MachineFunction::iterator(TBB) == Fallthrough || 980 MachineFunction::iterator(FBB) == Fallthrough) 981 return true; 982 983 // If it's an unconditional branch to some block not the fall through, it 984 // doesn't fall through. 985 if (Cond.empty()) return false; 986 987 // Otherwise, if it is conditional and has no explicit false block, it falls 988 // through. 989 return FBB == 0; 990} 991 992/// CanFallThrough - Return true if the specified can implicitly transfer 993/// control to the block after it by falling off the end of it. This should 994/// return false if it can reach the block after it, but it uses an explicit 995/// branch to do so (e.g. a table jump). 996/// 997/// True is a conservative answer. 998/// 999bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 1000 MachineBasicBlock *TBB = 0, *FBB = 0; 1001 SmallVector<MachineOperand, 4> Cond; 1002 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true); 1003 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 1004} 1005 1006/// IsBetterFallthrough - Return true if it would be clearly better to 1007/// fall-through to MBB1 than to fall through into MBB2. This has to return 1008/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 1009/// result in infinite loops. 1010static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 1011 MachineBasicBlock *MBB2) { 1012 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 1013 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 1014 // optimize branches that branch to either a return block or an assert block 1015 // into a fallthrough to the return. 1016 if (MBB1->empty() || MBB2->empty()) return false; 1017 1018 // If there is a clear successor ordering we make sure that one block 1019 // will fall through to the next 1020 if (MBB1->isSuccessor(MBB2)) return true; 1021 if (MBB2->isSuccessor(MBB1)) return false; 1022 1023 MachineInstr *MBB1I = --MBB1->end(); 1024 MachineInstr *MBB2I = --MBB2->end(); 1025 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 1026} 1027 1028/// TailDuplicate - MBB unconditionally branches to SuccBB. If it is profitable, 1029/// duplicate SuccBB's contents in MBB to eliminate the branch. 1030bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB, 1031 bool PrevFallsThrough, 1032 MachineFunction &MF) { 1033 // Don't try to tail-duplicate single-block loops. 1034 if (TailBB->isSuccessor(TailBB)) 1035 return false; 1036 1037 // Don't tail-duplicate a block which will soon be folded into its successor. 1038 if (TailBB->succ_size() == 1 && 1039 TailBB->succ_begin()[0]->pred_size() == 1) 1040 return false; 1041 1042 // Duplicate up to one less that the tail-merge threshold, so that we don't 1043 // get into an infinite loop between duplicating and merging. When optimizing 1044 // for size, duplicate only one, because one branch instruction can be 1045 // eliminated to compensate for the duplication. 1046 unsigned MaxDuplicateCount = 1047 MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ? 1048 1 : (TailMergeSize - 1); 1049 1050 // Check the instructions in the block to determine whether tail-duplication 1051 // is invalid or unlikely to be unprofitable. 1052 unsigned i = 0; 1053 bool HasCall = false; 1054 for (MachineBasicBlock::iterator I = TailBB->begin(); 1055 I != TailBB->end(); ++I, ++i) { 1056 // Non-duplicable things shouldn't be tail-duplicated. 1057 if (I->getDesc().isNotDuplicable()) return false; 1058 // Don't duplicate more than the threshold. 1059 if (i == MaxDuplicateCount) return false; 1060 // Remember if we saw a call. 1061 if (I->getDesc().isCall()) HasCall = true; 1062 } 1063 // Heuristically, don't tail-duplicate calls if it would expand code size, 1064 // as it's less likely to be worth the extra cost. 1065 if (i > 1 && HasCall) 1066 return false; 1067 1068 // Iterate through all the unique predecessors and tail-duplicate this 1069 // block into them, if possible. Copying the list ahead of time also 1070 // avoids trouble with the predecessor list reallocating. 1071 bool Changed = false; 1072 SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(), 1073 TailBB->pred_end()); 1074 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(), 1075 PE = Preds.end(); PI != PE; ++PI) { 1076 MachineBasicBlock *PredBB = *PI; 1077 1078 assert(TailBB != PredBB && 1079 "Single-block loop should have been rejected earlier!"); 1080 if (PredBB->succ_size() > 1) continue; 1081 1082 MachineBasicBlock *PredTBB, *PredFBB; 1083 SmallVector<MachineOperand, 4> PredCond; 1084 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)) 1085 continue; 1086 if (!PredCond.empty()) 1087 continue; 1088 // EH edges are ignored by AnalyzeBranch. 1089 if (PredBB->succ_size() != 1) 1090 continue; 1091 // Don't duplicate into a fall-through predecessor unless its the 1092 // only predecessor. 1093 if (PredBB->isLayoutSuccessor(TailBB) && 1094 PrevFallsThrough && 1095 TailBB->pred_size() != 1) 1096 continue; 1097 1098 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB 1099 << "From Succ: " << *TailBB); 1100 1101 // Remove PredBB's unconditional branch. 1102 TII->RemoveBranch(*PredBB); 1103 // Clone the contents of TailBB into PredBB. 1104 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end(); 1105 I != E; ++I) { 1106 MachineInstr *NewMI = MF.CloneMachineInstr(I); 1107 PredBB->insert(PredBB->end(), NewMI); 1108 } 1109 1110 // Update the CFG. 1111 PredBB->removeSuccessor(PredBB->succ_begin()); 1112 assert(PredBB->succ_empty() && 1113 "TailDuplicate called on block with multiple successors!"); 1114 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(), 1115 E = TailBB->succ_end(); I != E; ++I) 1116 PredBB->addSuccessor(*I); 1117 1118 Changed = true; 1119 } 1120 1121 return Changed; 1122} 1123 1124/// OptimizeBlock - Analyze and optimize control flow related to the specified 1125/// block. This is never called on the entry block. 1126bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1127 bool MadeChange = false; 1128 MachineFunction &MF = *MBB->getParent(); 1129ReoptimizeBlock: 1130 1131 MachineFunction::iterator FallThrough = MBB; 1132 ++FallThrough; 1133 1134 // If this block is empty, make everyone use its fall-through, not the block 1135 // explicitly. Landing pads should not do this since the landing-pad table 1136 // points to this block. Blocks with their addresses taken shouldn't be 1137 // optimized away. 1138 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 1139 // Dead block? Leave for cleanup later. 1140 if (MBB->pred_empty()) return MadeChange; 1141 1142 if (FallThrough == MF.end()) { 1143 // TODO: Simplify preds to not branch here if possible! 1144 } else { 1145 // Rewrite all predecessors of the old block to go to the fallthrough 1146 // instead. 1147 while (!MBB->pred_empty()) { 1148 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1149 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1150 } 1151 // If MBB was the target of a jump table, update jump tables to go to the 1152 // fallthrough instead. 1153 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough); 1154 MadeChange = true; 1155 } 1156 return MadeChange; 1157 } 1158 1159 // Check to see if we can simplify the terminator of the block before this 1160 // one. 1161 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1162 1163 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1164 SmallVector<MachineOperand, 4> PriorCond; 1165 bool PriorUnAnalyzable = 1166 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1167 if (!PriorUnAnalyzable) { 1168 // If the CFG for the prior block has extra edges, remove them. 1169 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1170 !PriorCond.empty()); 1171 1172 // If the previous branch is conditional and both conditions go to the same 1173 // destination, remove the branch, replacing it with an unconditional one or 1174 // a fall-through. 1175 if (PriorTBB && PriorTBB == PriorFBB) { 1176 TII->RemoveBranch(PrevBB); 1177 PriorCond.clear(); 1178 if (PriorTBB != MBB) 1179 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 1180 MadeChange = true; 1181 ++NumBranchOpts; 1182 goto ReoptimizeBlock; 1183 } 1184 1185 // If the previous block unconditionally falls through to this block and 1186 // this block has no other predecessors, move the contents of this block 1187 // into the prior block. This doesn't usually happen when SimplifyCFG 1188 // has been used, but it can happen tail duplication eliminates all the 1189 // non-branch predecessors of a block leaving only the fall-through edge. 1190 // This has to check PrevBB->succ_size() because EH edges are ignored by 1191 // AnalyzeBranch. 1192 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1193 PrevBB.succ_size() == 1 && 1194 !MBB->hasAddressTaken()) { 1195 DEBUG(errs() << "\nMerging into block: " << PrevBB 1196 << "From MBB: " << *MBB); 1197 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1198 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1199 assert(PrevBB.succ_empty()); 1200 PrevBB.transferSuccessors(MBB); 1201 MadeChange = true; 1202 return MadeChange; 1203 } 1204 1205 // If the previous branch *only* branches to *this* block (conditional or 1206 // not) remove the branch. 1207 if (PriorTBB == MBB && PriorFBB == 0) { 1208 TII->RemoveBranch(PrevBB); 1209 MadeChange = true; 1210 ++NumBranchOpts; 1211 goto ReoptimizeBlock; 1212 } 1213 1214 // If the prior block branches somewhere else on the condition and here if 1215 // the condition is false, remove the uncond second branch. 1216 if (PriorFBB == MBB) { 1217 TII->RemoveBranch(PrevBB); 1218 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 1219 MadeChange = true; 1220 ++NumBranchOpts; 1221 goto ReoptimizeBlock; 1222 } 1223 1224 // If the prior block branches here on true and somewhere else on false, and 1225 // if the branch condition is reversible, reverse the branch to create a 1226 // fall-through. 1227 if (PriorTBB == MBB) { 1228 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1229 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1230 TII->RemoveBranch(PrevBB); 1231 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 1232 MadeChange = true; 1233 ++NumBranchOpts; 1234 goto ReoptimizeBlock; 1235 } 1236 } 1237 1238 // If this block has no successors (e.g. it is a return block or ends with 1239 // a call to a no-return function like abort or __cxa_throw) and if the pred 1240 // falls through into this block, and if it would otherwise fall through 1241 // into the block after this, move this block to the end of the function. 1242 // 1243 // We consider it more likely that execution will stay in the function (e.g. 1244 // due to loops) than it is to exit it. This asserts in loops etc, moving 1245 // the assert condition out of the loop body. 1246 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1247 MachineFunction::iterator(PriorTBB) == FallThrough && 1248 !CanFallThrough(MBB)) { 1249 bool DoTransform = true; 1250 1251 // We have to be careful that the succs of PredBB aren't both no-successor 1252 // blocks. If neither have successors and if PredBB is the second from 1253 // last block in the function, we'd just keep swapping the two blocks for 1254 // last. Only do the swap if one is clearly better to fall through than 1255 // the other. 1256 if (FallThrough == --MF.end() && 1257 !IsBetterFallthrough(PriorTBB, MBB)) 1258 DoTransform = false; 1259 1260 // We don't want to do this transformation if we have control flow like: 1261 // br cond BB2 1262 // BB1: 1263 // .. 1264 // jmp BBX 1265 // BB2: 1266 // .. 1267 // ret 1268 // 1269 // In this case, we could actually be moving the return block *into* a 1270 // loop! 1271 if (DoTransform && !MBB->succ_empty() && 1272 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 1273 DoTransform = false; 1274 1275 1276 if (DoTransform) { 1277 // Reverse the branch so we will fall through on the previous true cond. 1278 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1279 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1280 DEBUG(errs() << "\nMoving MBB: " << *MBB 1281 << "To make fallthrough to: " << *PriorTBB << "\n"); 1282 1283 TII->RemoveBranch(PrevBB); 1284 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 1285 1286 // Move this block to the end of the function. 1287 MBB->moveAfter(--MF.end()); 1288 MadeChange = true; 1289 ++NumBranchOpts; 1290 return MadeChange; 1291 } 1292 } 1293 } 1294 } 1295 1296 // Analyze the branch in the current block. 1297 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1298 SmallVector<MachineOperand, 4> CurCond; 1299 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1300 if (!CurUnAnalyzable) { 1301 // If the CFG for the prior block has extra edges, remove them. 1302 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1303 1304 // If this is a two-way branch, and the FBB branches to this block, reverse 1305 // the condition so the single-basic-block loop is faster. Instead of: 1306 // Loop: xxx; jcc Out; jmp Loop 1307 // we want: 1308 // Loop: xxx; jncc Loop; jmp Out 1309 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1310 SmallVector<MachineOperand, 4> NewCond(CurCond); 1311 if (!TII->ReverseBranchCondition(NewCond)) { 1312 TII->RemoveBranch(*MBB); 1313 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 1314 MadeChange = true; 1315 ++NumBranchOpts; 1316 goto ReoptimizeBlock; 1317 } 1318 } 1319 1320 1321 // If this branch is the only thing in its block, see if we can forward 1322 // other blocks across it. 1323 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1324 MBB->begin()->getDesc().isBranch() && CurTBB != MBB && 1325 !MBB->hasAddressTaken()) { 1326 // This block may contain just an unconditional branch. Because there can 1327 // be 'non-branch terminators' in the block, try removing the branch and 1328 // then seeing if the block is empty. 1329 TII->RemoveBranch(*MBB); 1330 1331 // If this block is just an unconditional branch to CurTBB, we can 1332 // usually completely eliminate the block. The only case we cannot 1333 // completely eliminate the block is when the block before this one 1334 // falls through into MBB and we can't understand the prior block's branch 1335 // condition. 1336 if (MBB->empty()) { 1337 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 1338 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1339 !PrevBB.isSuccessor(MBB)) { 1340 // If the prior block falls through into us, turn it into an 1341 // explicit branch to us to make updates simpler. 1342 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1343 PriorTBB != MBB && PriorFBB != MBB) { 1344 if (PriorTBB == 0) { 1345 assert(PriorCond.empty() && PriorFBB == 0 && 1346 "Bad branch analysis"); 1347 PriorTBB = MBB; 1348 } else { 1349 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1350 PriorFBB = MBB; 1351 } 1352 TII->RemoveBranch(PrevBB); 1353 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1354 } 1355 1356 // Iterate through all the predecessors, revectoring each in-turn. 1357 size_t PI = 0; 1358 bool DidChange = false; 1359 bool HasBranchToSelf = false; 1360 while(PI != MBB->pred_size()) { 1361 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1362 if (PMBB == MBB) { 1363 // If this block has an uncond branch to itself, leave it. 1364 ++PI; 1365 HasBranchToSelf = true; 1366 } else { 1367 DidChange = true; 1368 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1369 // If this change resulted in PMBB ending in a conditional 1370 // branch where both conditions go to the same destination, 1371 // change this to an unconditional branch (and fix the CFG). 1372 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1373 SmallVector<MachineOperand, 4> NewCurCond; 1374 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1375 NewCurFBB, NewCurCond, true); 1376 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1377 TII->RemoveBranch(*PMBB); 1378 NewCurCond.clear(); 1379 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond); 1380 MadeChange = true; 1381 ++NumBranchOpts; 1382 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1383 } 1384 } 1385 } 1386 1387 // Change any jumptables to go to the new MBB. 1388 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB); 1389 if (DidChange) { 1390 ++NumBranchOpts; 1391 MadeChange = true; 1392 if (!HasBranchToSelf) return MadeChange; 1393 } 1394 } 1395 } 1396 1397 // Add the branch back if the block is more than just an uncond branch. 1398 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1399 } 1400 } 1401 1402 // Now we know that there was no fall-through into this block, check to 1403 // see if it has a fall-through into its successor. 1404 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1405 CurCond); 1406 bool PrevFallsThru = CanFallThrough(&PrevBB, PriorUnAnalyzable, 1407 PriorTBB, PriorFBB, PriorCond); 1408 1409 // If this block is small, unconditionally branched to, and does not 1410 // fall through, tail-duplicate its instructions into its predecessors 1411 // to eliminate a (dynamic) branch. 1412 if (!CurFallsThru) 1413 if (TailDuplicate(MBB, PrevFallsThru, MF)) { 1414 MadeChange = true; 1415 return MadeChange; 1416 } 1417 1418 // If the prior block doesn't fall through into this block, and if this 1419 // block doesn't fall through into some other block, see if we can find a 1420 // place to move this block where a fall-through will happen. 1421 if (!PrevFallsThru) { 1422 if (!MBB->isLandingPad()) { 1423 // Check all the predecessors of this block. If one of them has no fall 1424 // throughs, move this block right after it. 1425 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1426 E = MBB->pred_end(); PI != E; ++PI) { 1427 // Analyze the branch at the end of the pred. 1428 MachineBasicBlock *PredBB = *PI; 1429 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1430 MachineBasicBlock *PredTBB, *PredFBB; 1431 SmallVector<MachineOperand, 4> PredCond; 1432 if (PredBB != MBB && !CanFallThrough(PredBB) && 1433 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1434 && (!CurFallsThru || !CurTBB || !CurFBB) 1435 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1436 // If the current block doesn't fall through, just move it. 1437 // If the current block can fall through and does not end with a 1438 // conditional branch, we need to append an unconditional jump to 1439 // the (current) next block. To avoid a possible compile-time 1440 // infinite loop, move blocks only backward in this case. 1441 // Also, if there are already 2 branches here, we cannot add a third; 1442 // this means we have the case 1443 // Bcc next 1444 // B elsewhere 1445 // next: 1446 if (CurFallsThru) { 1447 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1448 CurCond.clear(); 1449 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1450 } 1451 MBB->moveAfter(PredBB); 1452 MadeChange = true; 1453 goto ReoptimizeBlock; 1454 } 1455 } 1456 } 1457 1458 if (!CurFallsThru) { 1459 // Check all successors to see if we can move this block before it. 1460 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1461 E = MBB->succ_end(); SI != E; ++SI) { 1462 // Analyze the branch at the end of the block before the succ. 1463 MachineBasicBlock *SuccBB = *SI; 1464 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1465 1466 // If this block doesn't already fall-through to that successor, and if 1467 // the succ doesn't already have a block that can fall through into it, 1468 // and if the successor isn't an EH destination, we can arrange for the 1469 // fallthrough to happen. 1470 if (SuccBB != MBB && &*SuccPrev != MBB && 1471 !CanFallThrough(SuccPrev) && !CurUnAnalyzable && 1472 !SuccBB->isLandingPad()) { 1473 MBB->moveBefore(SuccBB); 1474 MadeChange = true; 1475 goto ReoptimizeBlock; 1476 } 1477 } 1478 1479 // Okay, there is no really great place to put this block. If, however, 1480 // the block before this one would be a fall-through if this block were 1481 // removed, move this block to the end of the function. 1482 MachineBasicBlock *PrevTBB, *PrevFBB; 1483 SmallVector<MachineOperand, 4> PrevCond; 1484 if (FallThrough != MF.end() && 1485 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1486 PrevBB.isSuccessor(FallThrough)) { 1487 MBB->moveAfter(--MF.end()); 1488 MadeChange = true; 1489 return MadeChange; 1490 } 1491 } 1492 } 1493 1494 return MadeChange; 1495} 1496