BranchFolding.cpp revision 2210c0bea83aa8a8585d793a1f63e8c01b65be38
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 59char BranchFolderPass::ID = 0; 60 61Pass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 62 return new BranchFolderPass(DefaultEnableTailMerge); 63} 64 65bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { 66 return OptimizeFunction(MF, 67 MF.getTarget().getInstrInfo(), 68 MF.getTarget().getRegisterInfo(), 69 getAnalysisIfAvailable<MachineModuleInfo>()); 70} 71 72 73BranchFolder::BranchFolder(bool defaultEnableTailMerge) { 74 switch (FlagEnableTailMerge) { 75 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 76 case cl::BOU_TRUE: EnableTailMerge = true; break; 77 case cl::BOU_FALSE: EnableTailMerge = false; break; 78 } 79} 80 81/// RemoveDeadBlock - Remove the specified dead machine basic block from the 82/// function, updating the CFG. 83void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 84 assert(MBB->pred_empty() && "MBB must be dead!"); 85 DEBUG(errs() << "\nRemoving MBB: " << *MBB); 86 87 MachineFunction *MF = MBB->getParent(); 88 // drop all successors. 89 while (!MBB->succ_empty()) 90 MBB->removeSuccessor(MBB->succ_end()-1); 91 92 // If there are any labels in the basic block, unregister them from 93 // MachineModuleInfo. 94 if (MMI && !MBB->empty()) { 95 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 96 I != E; ++I) { 97 if (I->isLabel()) 98 // The label ID # is always operand #0, an immediate. 99 MMI->InvalidateLabel(I->getOperand(0).getImm()); 100 } 101 } 102 103 // Remove the block. 104 MF->erase(MBB); 105} 106 107/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 108/// followed by terminators, and if the implicitly defined registers are not 109/// used by the terminators, remove those implicit_def's. e.g. 110/// BB1: 111/// r0 = implicit_def 112/// r1 = implicit_def 113/// br 114/// This block can be optimized away later if the implicit instructions are 115/// removed. 116bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 117 SmallSet<unsigned, 4> ImpDefRegs; 118 MachineBasicBlock::iterator I = MBB->begin(); 119 while (I != MBB->end()) { 120 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) 121 break; 122 unsigned Reg = I->getOperand(0).getReg(); 123 ImpDefRegs.insert(Reg); 124 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg); 125 unsigned SubReg = *SubRegs; ++SubRegs) 126 ImpDefRegs.insert(SubReg); 127 ++I; 128 } 129 if (ImpDefRegs.empty()) 130 return false; 131 132 MachineBasicBlock::iterator FirstTerm = I; 133 while (I != MBB->end()) { 134 if (!TII->isUnpredicatedTerminator(I)) 135 return false; 136 // See if it uses any of the implicitly defined registers. 137 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 138 MachineOperand &MO = I->getOperand(i); 139 if (!MO.isReg() || !MO.isUse()) 140 continue; 141 unsigned Reg = MO.getReg(); 142 if (ImpDefRegs.count(Reg)) 143 return false; 144 } 145 ++I; 146 } 147 148 I = MBB->begin(); 149 while (I != FirstTerm) { 150 MachineInstr *ImpDefMI = &*I; 151 ++I; 152 MBB->erase(ImpDefMI); 153 } 154 155 return true; 156} 157 158/// OptimizeFunction - Perhaps branch folding, tail merging and other 159/// CFG optimizations on the given function. 160bool BranchFolder::OptimizeFunction(MachineFunction &MF, 161 const TargetInstrInfo *tii, 162 const TargetRegisterInfo *tri, 163 MachineModuleInfo *mmi) { 164 if (!tii) return false; 165 166 TII = tii; 167 TRI = tri; 168 MMI = mmi; 169 170 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 171 172 // Fix CFG. The later algorithms expect it to be right. 173 bool MadeChange = false; 174 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 175 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 176 SmallVector<MachineOperand, 4> Cond; 177 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 178 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 179 MadeChange |= OptimizeImpDefsBlock(MBB); 180 } 181 182 183 bool MadeChangeThisIteration = true; 184 while (MadeChangeThisIteration) { 185 MadeChangeThisIteration = false; 186 MadeChangeThisIteration |= TailMergeBlocks(MF); 187 MadeChangeThisIteration |= OptimizeBranches(MF); 188 MadeChange |= MadeChangeThisIteration; 189 } 190 191 // See if any jump tables have become mergable or dead as the code generator 192 // did its thing. 193 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 194 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 195 if (!JTs.empty()) { 196 // Figure out how these jump tables should be merged. 197 std::vector<unsigned> JTMapping; 198 JTMapping.reserve(JTs.size()); 199 200 // We always keep the 0th jump table. 201 JTMapping.push_back(0); 202 203 // Scan the jump tables, seeing if there are any duplicates. Note that this 204 // is N^2, which should be fixed someday. 205 for (unsigned i = 1, e = JTs.size(); i != e; ++i) { 206 if (JTs[i].MBBs.empty()) 207 JTMapping.push_back(i); 208 else 209 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 210 } 211 212 // If a jump table was merge with another one, walk the function rewriting 213 // references to jump tables to reference the new JT ID's. Keep track of 214 // whether we see a jump table idx, if not, we can delete the JT. 215 BitVector JTIsLive(JTs.size()); 216 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 217 BB != E; ++BB) { 218 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 219 I != E; ++I) 220 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 221 MachineOperand &Op = I->getOperand(op); 222 if (!Op.isJTI()) continue; 223 unsigned NewIdx = JTMapping[Op.getIndex()]; 224 Op.setIndex(NewIdx); 225 226 // Remember that this JT is live. 227 JTIsLive.set(NewIdx); 228 } 229 } 230 231 // Finally, remove dead jump tables. This happens either because the 232 // indirect jump was unreachable (and thus deleted) or because the jump 233 // table was merged with some other one. 234 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 235 if (!JTIsLive.test(i)) { 236 JTI->RemoveJumpTable(i); 237 MadeChange = true; 238 } 239 } 240 241 delete RS; 242 return MadeChange; 243} 244 245//===----------------------------------------------------------------------===// 246// Tail Merging of Blocks 247//===----------------------------------------------------------------------===// 248 249/// HashMachineInstr - Compute a hash value for MI and its operands. 250static unsigned HashMachineInstr(const MachineInstr *MI) { 251 unsigned Hash = MI->getOpcode(); 252 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 253 const MachineOperand &Op = MI->getOperand(i); 254 255 // Merge in bits from the operand if easy. 256 unsigned OperandHash = 0; 257 switch (Op.getType()) { 258 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 259 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 260 case MachineOperand::MO_MachineBasicBlock: 261 OperandHash = Op.getMBB()->getNumber(); 262 break; 263 case MachineOperand::MO_FrameIndex: 264 case MachineOperand::MO_ConstantPoolIndex: 265 case MachineOperand::MO_JumpTableIndex: 266 OperandHash = Op.getIndex(); 267 break; 268 case MachineOperand::MO_GlobalAddress: 269 case MachineOperand::MO_ExternalSymbol: 270 // Global address / external symbol are too hard, don't bother, but do 271 // pull in the offset. 272 OperandHash = Op.getOffset(); 273 break; 274 default: break; 275 } 276 277 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 278 } 279 return Hash; 280} 281 282/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 283/// with no successors, we hash two instructions, because cross-jumping 284/// only saves code when at least two instructions are removed (since a 285/// branch must be inserted). For blocks with a successor, one of the 286/// two blocks to be tail-merged will end with a branch already, so 287/// it gains to cross-jump even for one instruction. 288static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 289 unsigned minCommonTailLength) { 290 MachineBasicBlock::const_iterator I = MBB->end(); 291 if (I == MBB->begin()) 292 return 0; // Empty MBB. 293 294 --I; 295 unsigned Hash = HashMachineInstr(I); 296 297 if (I == MBB->begin() || minCommonTailLength == 1) 298 return Hash; // Single instr MBB. 299 300 --I; 301 // Hash in the second-to-last instruction. 302 Hash ^= HashMachineInstr(I) << 2; 303 return Hash; 304} 305 306/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 307/// of instructions they actually have in common together at their end. Return 308/// iterators for the first shared instruction in each block. 309static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 310 MachineBasicBlock *MBB2, 311 MachineBasicBlock::iterator &I1, 312 MachineBasicBlock::iterator &I2) { 313 I1 = MBB1->end(); 314 I2 = MBB2->end(); 315 316 unsigned TailLen = 0; 317 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 318 --I1; --I2; 319 if (!I1->isIdenticalTo(I2) || 320 // FIXME: This check is dubious. It's used to get around a problem where 321 // people incorrectly expect inline asm directives to remain in the same 322 // relative order. This is untenable because normal compiler 323 // optimizations (like this one) may reorder and/or merge these 324 // directives. 325 I1->getOpcode() == TargetInstrInfo::INLINEASM) { 326 ++I1; ++I2; 327 break; 328 } 329 ++TailLen; 330 } 331 return TailLen; 332} 333 334/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 335/// after it, replacing it with an unconditional branch to NewDest. This 336/// returns true if OldInst's block is modified, false if NewDest is modified. 337void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 338 MachineBasicBlock *NewDest) { 339 MachineBasicBlock *OldBB = OldInst->getParent(); 340 341 // Remove all the old successors of OldBB from the CFG. 342 while (!OldBB->succ_empty()) 343 OldBB->removeSuccessor(OldBB->succ_begin()); 344 345 // Remove all the dead instructions from the end of OldBB. 346 OldBB->erase(OldInst, OldBB->end()); 347 348 // If OldBB isn't immediately before OldBB, insert a branch to it. 349 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 350 TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>()); 351 OldBB->addSuccessor(NewDest); 352 ++NumTailMerge; 353} 354 355/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 356/// MBB so that the part before the iterator falls into the part starting at the 357/// iterator. This returns the new MBB. 358MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 359 MachineBasicBlock::iterator BBI1) { 360 MachineFunction &MF = *CurMBB.getParent(); 361 362 // Create the fall-through block. 363 MachineFunction::iterator MBBI = &CurMBB; 364 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 365 CurMBB.getParent()->insert(++MBBI, NewMBB); 366 367 // Move all the successors of this block to the specified block. 368 NewMBB->transferSuccessors(&CurMBB); 369 370 // Add an edge from CurMBB to NewMBB for the fall-through. 371 CurMBB.addSuccessor(NewMBB); 372 373 // Splice the code over. 374 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 375 376 // For targets that use the register scavenger, we must maintain LiveIns. 377 if (RS) { 378 RS->enterBasicBlock(&CurMBB); 379 if (!CurMBB.empty()) 380 RS->forward(prior(CurMBB.end())); 381 BitVector RegsLiveAtExit(TRI->getNumRegs()); 382 RS->getRegsUsed(RegsLiveAtExit, false); 383 for (unsigned int i=0, e=TRI->getNumRegs(); i!=e; i++) 384 if (RegsLiveAtExit[i]) 385 NewMBB->addLiveIn(i); 386 } 387 388 return NewMBB; 389} 390 391/// EstimateRuntime - Make a rough estimate for how long it will take to run 392/// the specified code. 393static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 394 MachineBasicBlock::iterator E) { 395 unsigned Time = 0; 396 for (; I != E; ++I) { 397 const TargetInstrDesc &TID = I->getDesc(); 398 if (TID.isCall()) 399 Time += 10; 400 else if (TID.mayLoad() || TID.mayStore()) 401 Time += 2; 402 else 403 ++Time; 404 } 405 return Time; 406} 407 408// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 409// branches temporarily for tail merging). In the case where CurMBB ends 410// with a conditional branch to the next block, optimize by reversing the 411// test and conditionally branching to SuccMBB instead. 412static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, 413 const TargetInstrInfo *TII) { 414 MachineFunction *MF = CurMBB->getParent(); 415 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 416 MachineBasicBlock *TBB = 0, *FBB = 0; 417 SmallVector<MachineOperand, 4> Cond; 418 if (I != MF->end() && 419 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 420 MachineBasicBlock *NextBB = I; 421 if (TBB == NextBB && !Cond.empty() && !FBB) { 422 if (!TII->ReverseBranchCondition(Cond)) { 423 TII->RemoveBranch(*CurMBB); 424 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 425 return; 426 } 427 } 428 } 429 TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>()); 430} 431 432static bool MergeCompare(const std::pair<unsigned,MachineBasicBlock*> &p, 433 const std::pair<unsigned,MachineBasicBlock*> &q) { 434 if (p.first < q.first) 435 return true; 436 else if (p.first > q.first) 437 return false; 438 else if (p.second->getNumber() < q.second->getNumber()) 439 return true; 440 else if (p.second->getNumber() > q.second->getNumber()) 441 return false; 442 else { 443 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 444 // an object with itself. 445#ifndef _GLIBCXX_DEBUG 446 llvm_unreachable("Predecessor appears twice"); 447#endif 448 return false; 449 } 450} 451 452/// CountTerminators - Count the number of terminators in the given 453/// block and set I to the position of the first non-terminator, if there 454/// is one, or MBB->end() otherwise. 455static unsigned CountTerminators(MachineBasicBlock *MBB, 456 MachineBasicBlock::iterator &I) { 457 I = MBB->end(); 458 unsigned NumTerms = 0; 459 for (;;) { 460 if (I == MBB->begin()) { 461 I = MBB->end(); 462 break; 463 } 464 --I; 465 if (!I->getDesc().isTerminator()) break; 466 ++NumTerms; 467 } 468 return NumTerms; 469} 470 471/// ProfitableToMerge - Check if two machine basic blocks have a common tail 472/// and decide if it would be profitable to merge those tails. Return the 473/// length of the common tail and iterators to the first common instruction 474/// in each block. 475static bool ProfitableToMerge(MachineBasicBlock *MBB1, 476 MachineBasicBlock *MBB2, 477 unsigned minCommonTailLength, 478 unsigned &CommonTailLen, 479 MachineBasicBlock::iterator &I1, 480 MachineBasicBlock::iterator &I2, 481 MachineBasicBlock *SuccBB, 482 MachineBasicBlock *PredBB) { 483 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 484 MachineFunction *MF = MBB1->getParent(); 485 486 if (CommonTailLen == 0) 487 return false; 488 489 // It's almost always profitable to merge any number of non-terminator 490 // instructions with the block that falls through into the common successor. 491 if (MBB1 == PredBB || MBB2 == PredBB) { 492 MachineBasicBlock::iterator I; 493 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 494 if (CommonTailLen > NumTerms) 495 return true; 496 } 497 498 // If both blocks have an unconditional branch temporarily stripped out, 499 // treat that as an additional common instruction. 500 if (MBB1 != PredBB && MBB2 != PredBB && 501 !MBB1->back().getDesc().isBarrier() && 502 !MBB2->back().getDesc().isBarrier()) 503 --minCommonTailLength; 504 505 // Check if the common tail is long enough to be worthwhile. 506 if (CommonTailLen >= minCommonTailLength) 507 return true; 508 509 // If we are optimizing for code size, 1 instruction in common is enough if 510 // we don't have to split a block. At worst we will be replacing a 511 // fallthrough into the common tail with a branch, which at worst breaks 512 // even with falling through into the duplicated common tail. 513 if (MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 514 (I1 == MBB1->begin() || I2 == MBB2->begin())) 515 return true; 516 517 return false; 518} 519 520/// ComputeSameTails - Look through all the blocks in MergePotentials that have 521/// hash CurHash (guaranteed to match the last element). Build the vector 522/// SameTails of all those that have the (same) largest number of instructions 523/// in common of any pair of these blocks. SameTails entries contain an 524/// iterator into MergePotentials (from which the MachineBasicBlock can be 525/// found) and a MachineBasicBlock::iterator into that MBB indicating the 526/// instruction where the matching code sequence begins. 527/// Order of elements in SameTails is the reverse of the order in which 528/// those blocks appear in MergePotentials (where they are not necessarily 529/// consecutive). 530unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 531 unsigned minCommonTailLength, 532 MachineBasicBlock *SuccBB, 533 MachineBasicBlock *PredBB) { 534 unsigned maxCommonTailLength = 0U; 535 SameTails.clear(); 536 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 537 MPIterator HighestMPIter = prior(MergePotentials.end()); 538 for (MPIterator CurMPIter = prior(MergePotentials.end()), 539 B = MergePotentials.begin(); 540 CurMPIter!=B && CurMPIter->first == CurHash; 541 --CurMPIter) { 542 for (MPIterator I = prior(CurMPIter); I->first == CurHash ; --I) { 543 unsigned CommonTailLen; 544 if (ProfitableToMerge(CurMPIter->second, I->second, minCommonTailLength, 545 CommonTailLen, TrialBBI1, TrialBBI2, 546 SuccBB, PredBB)) { 547 if (CommonTailLen > maxCommonTailLength) { 548 SameTails.clear(); 549 maxCommonTailLength = CommonTailLen; 550 HighestMPIter = CurMPIter; 551 SameTails.push_back(std::make_pair(CurMPIter, TrialBBI1)); 552 } 553 if (HighestMPIter == CurMPIter && 554 CommonTailLen == maxCommonTailLength) 555 SameTails.push_back(std::make_pair(I, TrialBBI2)); 556 } 557 if (I == B) 558 break; 559 } 560 } 561 return maxCommonTailLength; 562} 563 564/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 565/// MergePotentials, restoring branches at ends of blocks as appropriate. 566void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 567 MachineBasicBlock* SuccBB, 568 MachineBasicBlock* PredBB) { 569 MPIterator CurMPIter, B; 570 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 571 CurMPIter->first == CurHash; 572 --CurMPIter) { 573 // Put the unconditional branch back, if we need one. 574 MachineBasicBlock *CurMBB = CurMPIter->second; 575 if (SuccBB && CurMBB != PredBB) 576 FixTail(CurMBB, SuccBB, TII); 577 if (CurMPIter == B) 578 break; 579 } 580 if (CurMPIter->first!=CurHash) 581 CurMPIter++; 582 MergePotentials.erase(CurMPIter, MergePotentials.end()); 583} 584 585/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 586/// only of the common tail. Create a block that does by splitting one. 587unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 588 unsigned maxCommonTailLength) { 589 unsigned i, commonTailIndex; 590 unsigned TimeEstimate = ~0U; 591 for (i=0, commonTailIndex=0; i<SameTails.size(); i++) { 592 // Use PredBB if possible; that doesn't require a new branch. 593 if (SameTails[i].first->second == PredBB) { 594 commonTailIndex = i; 595 break; 596 } 597 // Otherwise, make a (fairly bogus) choice based on estimate of 598 // how long it will take the various blocks to execute. 599 unsigned t = EstimateRuntime(SameTails[i].first->second->begin(), 600 SameTails[i].second); 601 if (t <= TimeEstimate) { 602 TimeEstimate = t; 603 commonTailIndex = i; 604 } 605 } 606 607 MachineBasicBlock::iterator BBI = SameTails[commonTailIndex].second; 608 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; 609 610 DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 611 << maxCommonTailLength); 612 613 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 614 SameTails[commonTailIndex].first->second = newMBB; 615 SameTails[commonTailIndex].second = newMBB->begin(); 616 617 // If we split PredBB, newMBB is the new predecessor. 618 if (PredBB == MBB) 619 PredBB = newMBB; 620 621 return commonTailIndex; 622} 623 624// See if any of the blocks in MergePotentials (which all have a common single 625// successor, or all have no successor) can be tail-merged. If there is a 626// successor, any blocks in MergePotentials that are not tail-merged and 627// are not immediately before Succ must have an unconditional branch to 628// Succ added (but the predecessor/successor lists need no adjustment). 629// The lone predecessor of Succ that falls through into Succ, 630// if any, is given in PredBB. 631 632bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 633 MachineBasicBlock* PredBB) { 634 bool MadeChange = false; 635 636 // Except for the special cases below, tail-merge if there are at least 637 // this many instructions in common. 638 unsigned minCommonTailLength = TailMergeSize; 639 640 // If there's a successor block, there are some cases which don't require 641 // new branching and as such are very likely to be profitable. 642 if (SuccBB) { 643 if (SuccBB->pred_size() == MergePotentials.size() && 644 !MergePotentials[0].second->empty()) { 645 // If all the predecessors have at least one tail instruction in common, 646 // merging is very likely to be a win since it won't require an increase 647 // in static branches, and it will decrease the static instruction count. 648 bool AllPredsMatch = true; 649 MachineBasicBlock::iterator FirstNonTerm; 650 unsigned MinNumTerms = CountTerminators(MergePotentials[0].second, 651 FirstNonTerm); 652 if (FirstNonTerm != MergePotentials[0].second->end()) { 653 for (unsigned i = 1, e = MergePotentials.size(); i != e; ++i) { 654 MachineBasicBlock::iterator OtherFirstNonTerm; 655 unsigned NumTerms = CountTerminators(MergePotentials[0].second, 656 OtherFirstNonTerm); 657 if (NumTerms < MinNumTerms) 658 MinNumTerms = NumTerms; 659 if (OtherFirstNonTerm == MergePotentials[i].second->end() || 660 OtherFirstNonTerm->isIdenticalTo(FirstNonTerm)) { 661 AllPredsMatch = false; 662 break; 663 } 664 } 665 666 // If they all have an instruction in common, do any amount of merging. 667 if (AllPredsMatch) 668 minCommonTailLength = MinNumTerms + 1; 669 } 670 } 671 } 672 673 DEBUG(errs() << "\nTryTailMergeBlocks: "; 674 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 675 errs() << "BB#" << MergePotentials[i].second->getNumber() 676 << (i == e-1 ? "" : ", "); 677 errs() << "\n"; 678 if (SuccBB) { 679 errs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 680 if (PredBB) 681 errs() << " which has fall-through from BB#" 682 << PredBB->getNumber() << "\n"; 683 } 684 errs() << "Looking for common tails of at least " 685 << minCommonTailLength << " instruction" 686 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 687 ); 688 689 // Sort by hash value so that blocks with identical end sequences sort 690 // together. 691 std::stable_sort(MergePotentials.begin(), MergePotentials.end(),MergeCompare); 692 693 // Walk through equivalence sets looking for actual exact matches. 694 while (MergePotentials.size() > 1) { 695 unsigned CurHash = MergePotentials.back().first; 696 697 // Build SameTails, identifying the set of blocks with this hash code 698 // and with the maximum number of instructions in common. 699 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 700 minCommonTailLength, 701 SuccBB, PredBB); 702 703 // If we didn't find any pair that has at least minCommonTailLength 704 // instructions in common, remove all blocks with this hash code and retry. 705 if (SameTails.empty()) { 706 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 707 continue; 708 } 709 710 // If one of the blocks is the entire common tail (and not the entry 711 // block, which we can't jump to), we can treat all blocks with this same 712 // tail at once. Use PredBB if that is one of the possibilities, as that 713 // will not introduce any extra branches. 714 MachineBasicBlock *EntryBB = MergePotentials.begin()->second-> 715 getParent()->begin(); 716 unsigned int commonTailIndex, i; 717 for (commonTailIndex=SameTails.size(), i=0; i<SameTails.size(); i++) { 718 MachineBasicBlock *MBB = SameTails[i].first->second; 719 if (MBB == EntryBB) 720 continue; 721 if (MBB == PredBB) { 722 commonTailIndex = i; 723 break; 724 } 725 if (MBB->begin() == SameTails[i].second) 726 commonTailIndex = i; 727 } 728 729 if (commonTailIndex == SameTails.size() || 730 (SameTails[commonTailIndex].first->second == PredBB && 731 SameTails[commonTailIndex].first->second->begin() != 732 SameTails[i].second)) { 733 // None of the blocks consist entirely of the common tail. 734 // Split a block so that one does. 735 commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); 736 } 737 738 MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; 739 // MBB is common tail. Adjust all other BB's to jump to this one. 740 // Traversal must be forwards so erases work. 741 DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber() 742 << " for "); 743 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 744 if (commonTailIndex == i) 745 continue; 746 DEBUG(errs() << "BB#" << SameTails[i].first->second->getNumber() 747 << (i == e-1 ? "" : ", ")); 748 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 749 ReplaceTailWithBranchTo(SameTails[i].second, MBB); 750 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 751 MergePotentials.erase(SameTails[i].first); 752 } 753 DEBUG(errs() << "\n"); 754 // We leave commonTailIndex in the worklist in case there are other blocks 755 // that match it with a smaller number of instructions. 756 MadeChange = true; 757 } 758 return MadeChange; 759} 760 761bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 762 763 if (!EnableTailMerge) return false; 764 765 bool MadeChange = false; 766 767 // First find blocks with no successors. 768 MergePotentials.clear(); 769 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 770 if (I->succ_empty()) 771 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I)); 772 } 773 774 // See if we can do any tail merging on those. 775 if (MergePotentials.size() < TailMergeThreshold && 776 MergePotentials.size() >= 2) 777 MadeChange |= TryTailMergeBlocks(NULL, NULL); 778 779 // Look at blocks (IBB) with multiple predecessors (PBB). 780 // We change each predecessor to a canonical form, by 781 // (1) temporarily removing any unconditional branch from the predecessor 782 // to IBB, and 783 // (2) alter conditional branches so they branch to the other block 784 // not IBB; this may require adding back an unconditional branch to IBB 785 // later, where there wasn't one coming in. E.g. 786 // Bcc IBB 787 // fallthrough to QBB 788 // here becomes 789 // Bncc QBB 790 // with a conceptual B to IBB after that, which never actually exists. 791 // With those changes, we see whether the predecessors' tails match, 792 // and merge them if so. We change things out of canonical form and 793 // back to the way they were later in the process. (OptimizeBranches 794 // would undo some of this, but we can't use it, because we'd get into 795 // a compile-time infinite loop repeatedly doing and undoing the same 796 // transformations.) 797 798 for (MachineFunction::iterator I = next(MF.begin()), E = MF.end(); 799 I != E; ++I) { 800 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 801 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 802 MachineBasicBlock *IBB = I; 803 MachineBasicBlock *PredBB = prior(I); 804 MergePotentials.clear(); 805 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 806 E2 = I->pred_end(); 807 P != E2; ++P) { 808 MachineBasicBlock* PBB = *P; 809 // Skip blocks that loop to themselves, can't tail merge these. 810 if (PBB == IBB) 811 continue; 812 // Visit each predecessor only once. 813 if (!UniquePreds.insert(PBB)) 814 continue; 815 MachineBasicBlock *TBB = 0, *FBB = 0; 816 SmallVector<MachineOperand, 4> Cond; 817 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 818 // Failing case: IBB is the target of a cbr, and 819 // we cannot reverse the branch. 820 SmallVector<MachineOperand, 4> NewCond(Cond); 821 if (!Cond.empty() && TBB == IBB) { 822 if (TII->ReverseBranchCondition(NewCond)) 823 continue; 824 // This is the QBB case described above 825 if (!FBB) 826 FBB = next(MachineFunction::iterator(PBB)); 827 } 828 // Failing case: the only way IBB can be reached from PBB is via 829 // exception handling. Happens for landing pads. Would be nice 830 // to have a bit in the edge so we didn't have to do all this. 831 if (IBB->isLandingPad()) { 832 MachineFunction::iterator IP = PBB; IP++; 833 MachineBasicBlock* PredNextBB = NULL; 834 if (IP!=MF.end()) 835 PredNextBB = IP; 836 if (TBB == NULL) { 837 if (IBB!=PredNextBB) // fallthrough 838 continue; 839 } else if (FBB) { 840 if (TBB!=IBB && FBB!=IBB) // cbr then ubr 841 continue; 842 } else if (Cond.empty()) { 843 if (TBB!=IBB) // ubr 844 continue; 845 } else { 846 if (TBB!=IBB && IBB!=PredNextBB) // cbr 847 continue; 848 } 849 } 850 // Remove the unconditional branch at the end, if any. 851 if (TBB && (Cond.empty() || FBB)) { 852 TII->RemoveBranch(*PBB); 853 if (!Cond.empty()) 854 // reinsert conditional branch only, for now 855 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond); 856 } 857 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P)); 858 } 859 } 860 if (MergePotentials.size() >= 2) 861 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 862 // Reinsert an unconditional branch if needed. 863 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks. 864 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 865 if (MergePotentials.size() == 1 && 866 MergePotentials.begin()->second != PredBB) 867 FixTail(MergePotentials.begin()->second, IBB, TII); 868 } 869 } 870 return MadeChange; 871} 872 873//===----------------------------------------------------------------------===// 874// Branch Optimization 875//===----------------------------------------------------------------------===// 876 877bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 878 bool MadeChange = false; 879 880 // Make sure blocks are numbered in order 881 MF.RenumberBlocks(); 882 883 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 884 MachineBasicBlock *MBB = I++; 885 MadeChange |= OptimizeBlock(MBB); 886 887 // If it is dead, remove it. 888 if (MBB->pred_empty()) { 889 RemoveDeadBlock(MBB); 890 MadeChange = true; 891 ++NumDeadBlocks; 892 } 893 } 894 return MadeChange; 895} 896 897 898/// CanFallThrough - Return true if the specified block (with the specified 899/// branch condition) can implicitly transfer control to the block after it by 900/// falling off the end of it. This should return false if it can reach the 901/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 902/// 903/// True is a conservative answer. 904/// 905bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 906 bool BranchUnAnalyzable, 907 MachineBasicBlock *TBB, 908 MachineBasicBlock *FBB, 909 const SmallVectorImpl<MachineOperand> &Cond) { 910 MachineFunction::iterator Fallthrough = CurBB; 911 ++Fallthrough; 912 // If FallthroughBlock is off the end of the function, it can't fall through. 913 if (Fallthrough == CurBB->getParent()->end()) 914 return false; 915 916 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 917 if (!CurBB->isSuccessor(Fallthrough)) 918 return false; 919 920 // If we couldn't analyze the branch, examine the last instruction. 921 // If the block doesn't end in a known control barrier, assume fallthrough 922 // is possible. The isPredicable check is needed because this code can be 923 // called during IfConversion, where an instruction which is normally a 924 // Barrier is predicated and thus no longer an actual control barrier. This 925 // is over-conservative though, because if an instruction isn't actually 926 // predicated we could still treat it like a barrier. 927 if (BranchUnAnalyzable) 928 return CurBB->empty() || !CurBB->back().getDesc().isBarrier() || 929 CurBB->back().getDesc().isPredicable(); 930 931 // If there is no branch, control always falls through. 932 if (TBB == 0) return true; 933 934 // If there is some explicit branch to the fallthrough block, it can obviously 935 // reach, even though the branch should get folded to fall through implicitly. 936 if (MachineFunction::iterator(TBB) == Fallthrough || 937 MachineFunction::iterator(FBB) == Fallthrough) 938 return true; 939 940 // If it's an unconditional branch to some block not the fall through, it 941 // doesn't fall through. 942 if (Cond.empty()) return false; 943 944 // Otherwise, if it is conditional and has no explicit false block, it falls 945 // through. 946 return FBB == 0; 947} 948 949/// CanFallThrough - Return true if the specified can implicitly transfer 950/// control to the block after it by falling off the end of it. This should 951/// return false if it can reach the block after it, but it uses an explicit 952/// branch to do so (e.g. a table jump). 953/// 954/// True is a conservative answer. 955/// 956bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 957 MachineBasicBlock *TBB = 0, *FBB = 0; 958 SmallVector<MachineOperand, 4> Cond; 959 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true); 960 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 961} 962 963/// IsBetterFallthrough - Return true if it would be clearly better to 964/// fall-through to MBB1 than to fall through into MBB2. This has to return 965/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 966/// result in infinite loops. 967static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 968 MachineBasicBlock *MBB2) { 969 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 970 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 971 // optimize branches that branch to either a return block or an assert block 972 // into a fallthrough to the return. 973 if (MBB1->empty() || MBB2->empty()) return false; 974 975 // If there is a clear successor ordering we make sure that one block 976 // will fall through to the next 977 if (MBB1->isSuccessor(MBB2)) return true; 978 if (MBB2->isSuccessor(MBB1)) return false; 979 980 MachineInstr *MBB1I = --MBB1->end(); 981 MachineInstr *MBB2I = --MBB2->end(); 982 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 983} 984 985/// TailDuplicate - MBB unconditionally branches to SuccBB. If it is profitable, 986/// duplicate SuccBB's contents in MBB to eliminate the branch. 987bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB, 988 bool PrevFallsThrough, 989 MachineFunction &MF) { 990 // Don't try to tail-duplicate single-block loops. 991 if (TailBB->isSuccessor(TailBB)) 992 return false; 993 994 // Don't tail-duplicate a block which will soon be folded into its successor. 995 if (TailBB->succ_size() == 1 && 996 TailBB->succ_begin()[0]->pred_size() == 1) 997 return false; 998 999 // Duplicate up to one less that the tail-merge threshold, so that we don't 1000 // get into an infinite loop between duplicating and merging. When optimizing 1001 // for size, duplicate only one, because one branch instruction can be 1002 // eliminated to compensate for the duplication. 1003 unsigned MaxDuplicateCount = 1004 MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ? 1005 1 : (TailMergeSize - 1); 1006 1007 // Check the instructions in the block to determine whether tail-duplication 1008 // is invalid or unlikely to be unprofitable. 1009 unsigned i = 0; 1010 bool HasCall = false; 1011 for (MachineBasicBlock::iterator I = TailBB->begin(); 1012 I != TailBB->end(); ++I, ++i) { 1013 // Non-duplicable things shouldn't be tail-duplicated. 1014 if (I->getDesc().isNotDuplicable()) return false; 1015 // Don't duplicate more than the threshold. 1016 if (i == MaxDuplicateCount) return false; 1017 // Remember if we saw a call. 1018 if (I->getDesc().isCall()) HasCall = true; 1019 } 1020 // Heuristically, don't tail-duplicate calls if it would expand code size, 1021 // as it's less likely to be worth the extra cost. 1022 if (i > 1 && HasCall) 1023 return false; 1024 1025 // Iterate through all the unique predecessors and tail-duplicate this 1026 // block into them, if possible. Copying the list ahead of time also 1027 // avoids trouble with the predecessor list reallocating. 1028 bool Changed = false; 1029 SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(), 1030 TailBB->pred_end()); 1031 for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(), 1032 PE = Preds.end(); PI != PE; ++PI) { 1033 MachineBasicBlock *PredBB = *PI; 1034 1035 assert(TailBB != PredBB && 1036 "Single-block loop should have been rejected earlier!"); 1037 if (PredBB->succ_size() > 1) continue; 1038 1039 MachineBasicBlock *PredTBB, *PredFBB; 1040 SmallVector<MachineOperand, 4> PredCond; 1041 if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)) 1042 continue; 1043 if (!PredCond.empty()) 1044 continue; 1045 // EH edges are ignored by AnalyzeBranch. 1046 if (PredBB->succ_size() != 1) 1047 continue; 1048 // Don't duplicate into a fall-through predecessor unless its the 1049 // only predecessor. 1050 if (&*next(MachineFunction::iterator(PredBB)) == TailBB && 1051 PrevFallsThrough && 1052 TailBB->pred_size() != 1) 1053 continue; 1054 1055 DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB 1056 << "From Succ: " << *TailBB); 1057 1058 // Remove PredBB's unconditional branch. 1059 TII->RemoveBranch(*PredBB); 1060 // Clone the contents of TailBB into PredBB. 1061 for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end(); 1062 I != E; ++I) { 1063 MachineInstr *NewMI = MF.CloneMachineInstr(I); 1064 PredBB->insert(PredBB->end(), NewMI); 1065 } 1066 1067 // Update the CFG. 1068 PredBB->removeSuccessor(PredBB->succ_begin()); 1069 assert(PredBB->succ_empty() && 1070 "TailDuplicate called on block with multiple successors!"); 1071 for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(), 1072 E = TailBB->succ_end(); I != E; ++I) 1073 PredBB->addSuccessor(*I); 1074 1075 Changed = true; 1076 } 1077 1078 return Changed; 1079} 1080 1081/// OptimizeBlock - Analyze and optimize control flow related to the specified 1082/// block. This is never called on the entry block. 1083bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1084 bool MadeChange = false; 1085 MachineFunction &MF = *MBB->getParent(); 1086ReoptimizeBlock: 1087 1088 MachineFunction::iterator FallThrough = MBB; 1089 ++FallThrough; 1090 1091 // If this block is empty, make everyone use its fall-through, not the block 1092 // explicitly. Landing pads should not do this since the landing-pad table 1093 // points to this block. Blocks with their addresses taken shouldn't be 1094 // optimized away. 1095 if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 1096 // Dead block? Leave for cleanup later. 1097 if (MBB->pred_empty()) return MadeChange; 1098 1099 if (FallThrough == MF.end()) { 1100 // TODO: Simplify preds to not branch here if possible! 1101 } else { 1102 // Rewrite all predecessors of the old block to go to the fallthrough 1103 // instead. 1104 while (!MBB->pred_empty()) { 1105 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1106 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1107 } 1108 // If MBB was the target of a jump table, update jump tables to go to the 1109 // fallthrough instead. 1110 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough); 1111 MadeChange = true; 1112 } 1113 return MadeChange; 1114 } 1115 1116 // Check to see if we can simplify the terminator of the block before this 1117 // one. 1118 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1119 1120 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1121 SmallVector<MachineOperand, 4> PriorCond; 1122 bool PriorUnAnalyzable = 1123 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1124 if (!PriorUnAnalyzable) { 1125 // If the CFG for the prior block has extra edges, remove them. 1126 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1127 !PriorCond.empty()); 1128 1129 // If the previous branch is conditional and both conditions go to the same 1130 // destination, remove the branch, replacing it with an unconditional one or 1131 // a fall-through. 1132 if (PriorTBB && PriorTBB == PriorFBB) { 1133 TII->RemoveBranch(PrevBB); 1134 PriorCond.clear(); 1135 if (PriorTBB != MBB) 1136 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 1137 MadeChange = true; 1138 ++NumBranchOpts; 1139 goto ReoptimizeBlock; 1140 } 1141 1142 // If the previous block unconditionally falls through to this block and 1143 // this block has no other predecessors, move the contents of this block 1144 // into the prior block. This doesn't usually happen when SimplifyCFG 1145 // has been used, but it can happen tail duplication eliminates all the 1146 // non-branch predecessors of a block leaving only the fall-through edge. 1147 // This has to check PrevBB->succ_size() because EH edges are ignored by 1148 // AnalyzeBranch. 1149 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1150 PrevBB.succ_size() == 1 && 1151 !MBB->hasAddressTaken()) { 1152 DEBUG(errs() << "\nMerging into block: " << PrevBB 1153 << "From MBB: " << *MBB); 1154 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1155 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1156 assert(PrevBB.succ_empty()); 1157 PrevBB.transferSuccessors(MBB); 1158 MadeChange = true; 1159 return MadeChange; 1160 } 1161 1162 // If the previous branch *only* branches to *this* block (conditional or 1163 // not) remove the branch. 1164 if (PriorTBB == MBB && PriorFBB == 0) { 1165 TII->RemoveBranch(PrevBB); 1166 MadeChange = true; 1167 ++NumBranchOpts; 1168 goto ReoptimizeBlock; 1169 } 1170 1171 // If the prior block branches somewhere else on the condition and here if 1172 // the condition is false, remove the uncond second branch. 1173 if (PriorFBB == MBB) { 1174 TII->RemoveBranch(PrevBB); 1175 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 1176 MadeChange = true; 1177 ++NumBranchOpts; 1178 goto ReoptimizeBlock; 1179 } 1180 1181 // If the prior block branches here on true and somewhere else on false, and 1182 // if the branch condition is reversible, reverse the branch to create a 1183 // fall-through. 1184 if (PriorTBB == MBB) { 1185 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1186 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1187 TII->RemoveBranch(PrevBB); 1188 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 1189 MadeChange = true; 1190 ++NumBranchOpts; 1191 goto ReoptimizeBlock; 1192 } 1193 } 1194 1195 // If this block has no successors (e.g. it is a return block or ends with 1196 // a call to a no-return function like abort or __cxa_throw) and if the pred 1197 // falls through into this block, and if it would otherwise fall through 1198 // into the block after this, move this block to the end of the function. 1199 // 1200 // We consider it more likely that execution will stay in the function (e.g. 1201 // due to loops) than it is to exit it. This asserts in loops etc, moving 1202 // the assert condition out of the loop body. 1203 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1204 MachineFunction::iterator(PriorTBB) == FallThrough && 1205 !CanFallThrough(MBB)) { 1206 bool DoTransform = true; 1207 1208 // We have to be careful that the succs of PredBB aren't both no-successor 1209 // blocks. If neither have successors and if PredBB is the second from 1210 // last block in the function, we'd just keep swapping the two blocks for 1211 // last. Only do the swap if one is clearly better to fall through than 1212 // the other. 1213 if (FallThrough == --MF.end() && 1214 !IsBetterFallthrough(PriorTBB, MBB)) 1215 DoTransform = false; 1216 1217 // We don't want to do this transformation if we have control flow like: 1218 // br cond BB2 1219 // BB1: 1220 // .. 1221 // jmp BBX 1222 // BB2: 1223 // .. 1224 // ret 1225 // 1226 // In this case, we could actually be moving the return block *into* a 1227 // loop! 1228 if (DoTransform && !MBB->succ_empty() && 1229 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 1230 DoTransform = false; 1231 1232 1233 if (DoTransform) { 1234 // Reverse the branch so we will fall through on the previous true cond. 1235 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1236 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1237 DEBUG(errs() << "\nMoving MBB: " << *MBB 1238 << "To make fallthrough to: " << *PriorTBB << "\n"); 1239 1240 TII->RemoveBranch(PrevBB); 1241 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 1242 1243 // Move this block to the end of the function. 1244 MBB->moveAfter(--MF.end()); 1245 MadeChange = true; 1246 ++NumBranchOpts; 1247 return MadeChange; 1248 } 1249 } 1250 } 1251 } 1252 1253 // Analyze the branch in the current block. 1254 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1255 SmallVector<MachineOperand, 4> CurCond; 1256 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1257 if (!CurUnAnalyzable) { 1258 // If the CFG for the prior block has extra edges, remove them. 1259 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1260 1261 // If this is a two-way branch, and the FBB branches to this block, reverse 1262 // the condition so the single-basic-block loop is faster. Instead of: 1263 // Loop: xxx; jcc Out; jmp Loop 1264 // we want: 1265 // Loop: xxx; jncc Loop; jmp Out 1266 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1267 SmallVector<MachineOperand, 4> NewCond(CurCond); 1268 if (!TII->ReverseBranchCondition(NewCond)) { 1269 TII->RemoveBranch(*MBB); 1270 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 1271 MadeChange = true; 1272 ++NumBranchOpts; 1273 goto ReoptimizeBlock; 1274 } 1275 } 1276 1277 1278 // If this branch is the only thing in its block, see if we can forward 1279 // other blocks across it. 1280 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1281 MBB->begin()->getDesc().isBranch() && CurTBB != MBB && 1282 !MBB->hasAddressTaken()) { 1283 // This block may contain just an unconditional branch. Because there can 1284 // be 'non-branch terminators' in the block, try removing the branch and 1285 // then seeing if the block is empty. 1286 TII->RemoveBranch(*MBB); 1287 1288 // If this block is just an unconditional branch to CurTBB, we can 1289 // usually completely eliminate the block. The only case we cannot 1290 // completely eliminate the block is when the block before this one 1291 // falls through into MBB and we can't understand the prior block's branch 1292 // condition. 1293 if (MBB->empty()) { 1294 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 1295 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1296 !PrevBB.isSuccessor(MBB)) { 1297 // If the prior block falls through into us, turn it into an 1298 // explicit branch to us to make updates simpler. 1299 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1300 PriorTBB != MBB && PriorFBB != MBB) { 1301 if (PriorTBB == 0) { 1302 assert(PriorCond.empty() && PriorFBB == 0 && 1303 "Bad branch analysis"); 1304 PriorTBB = MBB; 1305 } else { 1306 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1307 PriorFBB = MBB; 1308 } 1309 TII->RemoveBranch(PrevBB); 1310 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1311 } 1312 1313 // Iterate through all the predecessors, revectoring each in-turn. 1314 size_t PI = 0; 1315 bool DidChange = false; 1316 bool HasBranchToSelf = false; 1317 while(PI != MBB->pred_size()) { 1318 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1319 if (PMBB == MBB) { 1320 // If this block has an uncond branch to itself, leave it. 1321 ++PI; 1322 HasBranchToSelf = true; 1323 } else { 1324 DidChange = true; 1325 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1326 // If this change resulted in PMBB ending in a conditional 1327 // branch where both conditions go to the same destination, 1328 // change this to an unconditional branch (and fix the CFG). 1329 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1330 SmallVector<MachineOperand, 4> NewCurCond; 1331 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1332 NewCurFBB, NewCurCond, true); 1333 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1334 TII->RemoveBranch(*PMBB); 1335 NewCurCond.clear(); 1336 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond); 1337 MadeChange = true; 1338 ++NumBranchOpts; 1339 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1340 } 1341 } 1342 } 1343 1344 // Change any jumptables to go to the new MBB. 1345 MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB); 1346 if (DidChange) { 1347 ++NumBranchOpts; 1348 MadeChange = true; 1349 if (!HasBranchToSelf) return MadeChange; 1350 } 1351 } 1352 } 1353 1354 // Add the branch back if the block is more than just an uncond branch. 1355 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1356 } 1357 } 1358 1359 // Now we know that there was no fall-through into this block, check to 1360 // see if it has a fall-through into its successor. 1361 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1362 CurCond); 1363 bool PrevFallsThru = CanFallThrough(&PrevBB, PriorUnAnalyzable, 1364 PriorTBB, PriorFBB, PriorCond); 1365 1366 // If this block is small, unconditionally branched to, and does not 1367 // fall through, tail-duplicate its instructions into its predecessors 1368 // to eliminate a (dynamic) branch. 1369 if (!CurFallsThru) 1370 if (TailDuplicate(MBB, PrevFallsThru, MF)) { 1371 MadeChange = true; 1372 return MadeChange; 1373 } 1374 1375 // If the prior block doesn't fall through into this block, and if this 1376 // block doesn't fall through into some other block, see if we can find a 1377 // place to move this block where a fall-through will happen. 1378 if (!PrevFallsThru) { 1379 if (!MBB->isLandingPad()) { 1380 // Check all the predecessors of this block. If one of them has no fall 1381 // throughs, move this block right after it. 1382 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1383 E = MBB->pred_end(); PI != E; ++PI) { 1384 // Analyze the branch at the end of the pred. 1385 MachineBasicBlock *PredBB = *PI; 1386 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1387 MachineBasicBlock *PredTBB, *PredFBB; 1388 SmallVector<MachineOperand, 4> PredCond; 1389 if (PredBB != MBB && !CanFallThrough(PredBB) && 1390 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1391 && (!CurFallsThru || !CurTBB || !CurFBB) 1392 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1393 // If the current block doesn't fall through, just move it. 1394 // If the current block can fall through and does not end with a 1395 // conditional branch, we need to append an unconditional jump to 1396 // the (current) next block. To avoid a possible compile-time 1397 // infinite loop, move blocks only backward in this case. 1398 // Also, if there are already 2 branches here, we cannot add a third; 1399 // this means we have the case 1400 // Bcc next 1401 // B elsewhere 1402 // next: 1403 if (CurFallsThru) { 1404 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1405 CurCond.clear(); 1406 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1407 } 1408 MBB->moveAfter(PredBB); 1409 MadeChange = true; 1410 goto ReoptimizeBlock; 1411 } 1412 } 1413 } 1414 1415 if (!CurFallsThru) { 1416 // Check all successors to see if we can move this block before it. 1417 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1418 E = MBB->succ_end(); SI != E; ++SI) { 1419 // Analyze the branch at the end of the block before the succ. 1420 MachineBasicBlock *SuccBB = *SI; 1421 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1422 1423 // If this block doesn't already fall-through to that successor, and if 1424 // the succ doesn't already have a block that can fall through into it, 1425 // and if the successor isn't an EH destination, we can arrange for the 1426 // fallthrough to happen. 1427 if (SuccBB != MBB && &*SuccPrev != MBB && 1428 !CanFallThrough(SuccPrev) && !CurUnAnalyzable && 1429 !SuccBB->isLandingPad()) { 1430 MBB->moveBefore(SuccBB); 1431 MadeChange = true; 1432 goto ReoptimizeBlock; 1433 } 1434 } 1435 1436 // Okay, there is no really great place to put this block. If, however, 1437 // the block before this one would be a fall-through if this block were 1438 // removed, move this block to the end of the function. 1439 MachineBasicBlock *PrevTBB, *PrevFBB; 1440 SmallVector<MachineOperand, 4> PrevCond; 1441 if (FallThrough != MF.end() && 1442 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1443 PrevBB.isSuccessor(FallThrough)) { 1444 MBB->moveAfter(--MF.end()); 1445 MadeChange = true; 1446 return MadeChange; 1447 } 1448 } 1449 } 1450 1451 return MadeChange; 1452} 1453