BranchFolding.cpp revision 6b8583cbf1f8a3df5ae859d3da2ca690ff57f91c
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 "llvm/CodeGen/Passes.h" 21#include "llvm/CodeGen/MachineModuleInfo.h" 22#include "llvm/CodeGen/MachineFunctionPass.h" 23#include "llvm/CodeGen/MachineJumpTableInfo.h" 24#include "llvm/CodeGen/RegisterScavenging.h" 25#include "llvm/Target/TargetInstrInfo.h" 26#include "llvm/Target/TargetMachine.h" 27#include "llvm/Target/TargetRegisterInfo.h" 28#include "llvm/Support/CommandLine.h" 29#include "llvm/Support/Debug.h" 30#include "llvm/ADT/SmallSet.h" 31#include "llvm/ADT/Statistic.h" 32#include "llvm/ADT/STLExtras.h" 33#include <algorithm> 34using namespace llvm; 35 36STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 37STATISTIC(NumBranchOpts, "Number of branches optimized"); 38STATISTIC(NumTailMerge , "Number of block tails merged"); 39static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 40 cl::init(cl::BOU_UNSET), cl::Hidden); 41namespace { 42 // Throttle for huge numbers of predecessors (compile speed problems) 43 static cl::opt<unsigned> 44 TailMergeThreshold("tail-merge-threshold", 45 cl::desc("Max number of predecessors to consider tail merging"), 46 cl::init(100), cl::Hidden); 47 48 struct VISIBILITY_HIDDEN BranchFolder : public MachineFunctionPass { 49 static char ID; 50 explicit BranchFolder(bool defaultEnableTailMerge) : 51 MachineFunctionPass((intptr_t)&ID) { 52 switch (FlagEnableTailMerge) { 53 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 54 case cl::BOU_TRUE: EnableTailMerge = true; break; 55 case cl::BOU_FALSE: EnableTailMerge = false; break; 56 } 57 } 58 59 virtual bool runOnMachineFunction(MachineFunction &MF); 60 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 61 const TargetInstrInfo *TII; 62 MachineModuleInfo *MMI; 63 bool MadeChange; 64 private: 65 // Tail Merging. 66 bool EnableTailMerge; 67 bool TailMergeBlocks(MachineFunction &MF); 68 bool TryMergeBlocks(MachineBasicBlock* SuccBB, 69 MachineBasicBlock* PredBB); 70 void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 71 MachineBasicBlock *NewDest); 72 MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB, 73 MachineBasicBlock::iterator BBI1); 74 unsigned ComputeSameTails(unsigned CurHash, unsigned minCommonTailLength); 75 void RemoveBlocksWithHash(unsigned CurHash, MachineBasicBlock* SuccBB, 76 MachineBasicBlock* PredBB); 77 78 typedef std::pair<unsigned,MachineBasicBlock*> MergePotentialsElt; 79 typedef std::vector<MergePotentialsElt>::iterator MPIterator; 80 std::vector<MergePotentialsElt> MergePotentials; 81 typedef std::pair<MPIterator, MachineBasicBlock::iterator> SameTailElt; 82 std::vector<SameTailElt> SameTails; 83 84 const TargetRegisterInfo *RegInfo; 85 RegScavenger *RS; 86 // Branch optzn. 87 bool OptimizeBranches(MachineFunction &MF); 88 void OptimizeBlock(MachineBasicBlock *MBB); 89 void RemoveDeadBlock(MachineBasicBlock *MBB); 90 bool OptimizeImpDefsBlock(MachineBasicBlock *MBB); 91 92 bool CanFallThrough(MachineBasicBlock *CurBB); 93 bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable, 94 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 95 const std::vector<MachineOperand> &Cond); 96 }; 97 char BranchFolder::ID = 0; 98} 99 100FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 101 return new BranchFolder(DefaultEnableTailMerge); } 102 103/// RemoveDeadBlock - Remove the specified dead machine basic block from the 104/// function, updating the CFG. 105void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 106 assert(MBB->pred_empty() && "MBB must be dead!"); 107 DOUT << "\nRemoving MBB: " << *MBB; 108 109 MachineFunction *MF = MBB->getParent(); 110 // drop all successors. 111 while (!MBB->succ_empty()) 112 MBB->removeSuccessor(MBB->succ_end()-1); 113 114 // If there is DWARF info to active, check to see if there are any LABEL 115 // records in the basic block. If so, unregister them from MachineModuleInfo. 116 if (MMI && !MBB->empty()) { 117 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 118 I != E; ++I) { 119 if ((unsigned)I->getOpcode() == TargetInstrInfo::LABEL) { 120 // The label ID # is always operand #0, an immediate. 121 MMI->InvalidateLabel(I->getOperand(0).getImm()); 122 } 123 } 124 } 125 126 // Remove the block. 127 MF->getBasicBlockList().erase(MBB); 128} 129 130/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 131/// followed by terminators, and if the implicitly defined registers are not 132/// used by the terminators, remove those implicit_def's. e.g. 133/// BB1: 134/// r0 = implicit_def 135/// r1 = implicit_def 136/// br 137/// This block can be optimized away later if the implicit instructions are 138/// removed. 139bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 140 SmallSet<unsigned, 4> ImpDefRegs; 141 MachineBasicBlock::iterator I = MBB->begin(); 142 while (I != MBB->end()) { 143 if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) 144 break; 145 unsigned Reg = I->getOperand(0).getReg(); 146 ImpDefRegs.insert(Reg); 147 for (const unsigned *SubRegs = RegInfo->getSubRegisters(Reg); 148 unsigned SubReg = *SubRegs; ++SubRegs) 149 ImpDefRegs.insert(SubReg); 150 ++I; 151 } 152 if (ImpDefRegs.empty()) 153 return false; 154 155 MachineBasicBlock::iterator FirstTerm = I; 156 while (I != MBB->end()) { 157 if (!TII->isUnpredicatedTerminator(I)) 158 return false; 159 // See if it uses any of the implicitly defined registers. 160 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 161 MachineOperand &MO = I->getOperand(i); 162 if (!MO.isReg() || !MO.isUse()) 163 continue; 164 unsigned Reg = MO.getReg(); 165 if (ImpDefRegs.count(Reg)) 166 return false; 167 } 168 ++I; 169 } 170 171 I = MBB->begin(); 172 while (I != FirstTerm) { 173 MachineInstr *ImpDefMI = &*I; 174 ++I; 175 MBB->erase(ImpDefMI); 176 } 177 178 return true; 179} 180 181bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { 182 TII = MF.getTarget().getInstrInfo(); 183 if (!TII) return false; 184 185 RegInfo = MF.getTarget().getRegisterInfo(); 186 187 // Fix CFG. The later algorithms expect it to be right. 188 bool EverMadeChange = false; 189 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 190 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 191 std::vector<MachineOperand> Cond; 192 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond)) 193 EverMadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 194 EverMadeChange |= OptimizeImpDefsBlock(MBB); 195 } 196 197 RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 198 199 MMI = getAnalysisToUpdate<MachineModuleInfo>(); 200 201 bool MadeChangeThisIteration = true; 202 while (MadeChangeThisIteration) { 203 MadeChangeThisIteration = false; 204 MadeChangeThisIteration |= TailMergeBlocks(MF); 205 MadeChangeThisIteration |= OptimizeBranches(MF); 206 EverMadeChange |= MadeChangeThisIteration; 207 } 208 209 // See if any jump tables have become mergable or dead as the code generator 210 // did its thing. 211 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 212 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 213 if (!JTs.empty()) { 214 // Figure out how these jump tables should be merged. 215 std::vector<unsigned> JTMapping; 216 JTMapping.reserve(JTs.size()); 217 218 // We always keep the 0th jump table. 219 JTMapping.push_back(0); 220 221 // Scan the jump tables, seeing if there are any duplicates. Note that this 222 // is N^2, which should be fixed someday. 223 for (unsigned i = 1, e = JTs.size(); i != e; ++i) 224 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 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.isJumpTableIndex()) 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 EverMadeChange = true; 252 } 253 } 254 255 delete RS; 256 return EverMadeChange; 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. 302 303static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 304 unsigned minCommonTailLength) { 305 MachineBasicBlock::const_iterator I = MBB->end(); 306 if (I == MBB->begin()) 307 return 0; // Empty MBB. 308 309 --I; 310 unsigned Hash = HashMachineInstr(I); 311 312 if (I == MBB->begin() || minCommonTailLength == 1) 313 return Hash; // Single instr MBB. 314 315 --I; 316 // Hash in the second-to-last instruction. 317 Hash ^= HashMachineInstr(I) << 2; 318 return Hash; 319} 320 321/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 322/// of instructions they actually have in common together at their end. Return 323/// iterators for the first shared instruction in each block. 324static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 325 MachineBasicBlock *MBB2, 326 MachineBasicBlock::iterator &I1, 327 MachineBasicBlock::iterator &I2) { 328 I1 = MBB1->end(); 329 I2 = MBB2->end(); 330 331 unsigned TailLen = 0; 332 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 333 --I1; --I2; 334 if (!I1->isIdenticalTo(I2) || 335 // FIXME: This check is dubious. It's used to get around a problem where 336 // people incorrectly expect inline asm directives to remain in the same 337 // relative order. This is untenable because normal compiler 338 // optimizations (like this one) may reorder and/or merge these 339 // directives. 340 I1->getOpcode() == TargetInstrInfo::INLINEASM) { 341 ++I1; ++I2; 342 break; 343 } 344 ++TailLen; 345 } 346 return TailLen; 347} 348 349/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 350/// after it, replacing it with an unconditional branch to NewDest. This 351/// returns true if OldInst's block is modified, false if NewDest is modified. 352void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 353 MachineBasicBlock *NewDest) { 354 MachineBasicBlock *OldBB = OldInst->getParent(); 355 356 // Remove all the old successors of OldBB from the CFG. 357 while (!OldBB->succ_empty()) 358 OldBB->removeSuccessor(OldBB->succ_begin()); 359 360 // Remove all the dead instructions from the end of OldBB. 361 OldBB->erase(OldInst, OldBB->end()); 362 363 // If OldBB isn't immediately before OldBB, insert a branch to it. 364 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 365 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>()); 366 OldBB->addSuccessor(NewDest); 367 ++NumTailMerge; 368} 369 370/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 371/// MBB so that the part before the iterator falls into the part starting at the 372/// iterator. This returns the new MBB. 373MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 374 MachineBasicBlock::iterator BBI1) { 375 // Create the fall-through block. 376 MachineFunction::iterator MBBI = &CurMBB; 377 MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock()); 378 CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB); 379 380 // Move all the successors of this block to the specified block. 381 while (!CurMBB.succ_empty()) { 382 MachineBasicBlock *S = *(CurMBB.succ_end()-1); 383 NewMBB->addSuccessor(S); 384 CurMBB.removeSuccessor(S); 385 } 386 387 // Add an edge from CurMBB to NewMBB for the fall-through. 388 CurMBB.addSuccessor(NewMBB); 389 390 // Splice the code over. 391 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 392 393 // For targets that use the register scavenger, we must maintain LiveIns. 394 if (RS) { 395 RS->enterBasicBlock(&CurMBB); 396 if (!CurMBB.empty()) 397 RS->forward(prior(CurMBB.end())); 398 BitVector RegsLiveAtExit(RegInfo->getNumRegs()); 399 RS->getRegsUsed(RegsLiveAtExit, false); 400 for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++) 401 if (RegsLiveAtExit[i]) 402 NewMBB->addLiveIn(i); 403 } 404 405 return NewMBB; 406} 407 408/// EstimateRuntime - Make a rough estimate for how long it will take to run 409/// the specified code. 410static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 411 MachineBasicBlock::iterator E) { 412 unsigned Time = 0; 413 for (; I != E; ++I) { 414 const TargetInstrDesc &TID = I->getDesc(); 415 if (TID.isCall()) 416 Time += 10; 417 else if (TID.isSimpleLoad() || TID.mayStore()) 418 Time += 2; 419 else 420 ++Time; 421 } 422 return Time; 423} 424 425/// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at 426/// MBB2I and then insert an unconditional branch in the other block. Determine 427/// which is the best to split 428static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1, 429 MachineBasicBlock::iterator MBB1I, 430 MachineBasicBlock *MBB2, 431 MachineBasicBlock::iterator MBB2I, 432 MachineBasicBlock *PredBB) { 433 // If one block is the entry block, split the other one; we can't generate 434 // a branch to the entry block, as its label is not emitted. 435 MachineBasicBlock *Entry = MBB1->getParent()->begin(); 436 if (MBB1 == Entry) 437 return false; 438 if (MBB2 == Entry) 439 return true; 440 441 // If one block falls through into the common successor, choose that 442 // one to split; it is one instruction less to do that. 443 if (PredBB) { 444 if (MBB1 == PredBB) 445 return true; 446 else if (MBB2 == PredBB) 447 return false; 448 } 449 // TODO: if we had some notion of which block was hotter, we could split 450 // the hot block, so it is the fall-through. Since we don't have profile info 451 // make a decision based on which will hurt most to split. 452 unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I); 453 unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I); 454 455 // If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the 456 // MBB1 block so it falls through. This will penalize the MBB2 path, but will 457 // have a lower overall impact on the program execution. 458 return MBB1Time < MBB2Time; 459} 460 461// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 462// branches temporarily for tail merging). In the case where CurMBB ends 463// with a conditional branch to the next block, optimize by reversing the 464// test and conditionally branching to SuccMBB instead. 465 466static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, 467 const TargetInstrInfo *TII) { 468 MachineFunction *MF = CurMBB->getParent(); 469 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 470 MachineBasicBlock *TBB = 0, *FBB = 0; 471 std::vector<MachineOperand> Cond; 472 if (I != MF->end() && 473 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) { 474 MachineBasicBlock *NextBB = I; 475 if (TBB == NextBB && !Cond.empty() && !FBB) { 476 if (!TII->ReverseBranchCondition(Cond)) { 477 TII->RemoveBranch(*CurMBB); 478 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 479 return; 480 } 481 } 482 } 483 TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector<MachineOperand>()); 484} 485 486static bool MergeCompare(const std::pair<unsigned,MachineBasicBlock*> &p, 487 const std::pair<unsigned,MachineBasicBlock*> &q) { 488 if (p.first < q.first) 489 return true; 490 else if (p.first > q.first) 491 return false; 492 else if (p.second->getNumber() < q.second->getNumber()) 493 return true; 494 else if (p.second->getNumber() > q.second->getNumber()) 495 return false; 496 else { 497 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 498 // an object with itself. 499#ifndef _GLIBCXX_DEBUG 500 assert(0 && "Predecessor appears twice"); 501#endif 502 return(false); 503 } 504} 505 506/// ComputeSameTails - Look through all the blocks in MergePotentials that have 507/// hash CurHash (guaranteed to match the last element). Build the vector 508/// SameTails of all those that have the (same) largest number of instructions 509/// in common of any pair of these blocks. SameTails entries contain an 510/// iterator into MergePotentials (from which the MachineBasicBlock can be 511/// found) and a MachineBasicBlock::iterator into that MBB indicating the 512/// instruction where the matching code sequence begins. 513/// Order of elements in SameTails is the reverse of the order in which 514/// those blocks appear in MergePotentials (where they are not necessarily 515/// consecutive). 516unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 517 unsigned minCommonTailLength) { 518 unsigned maxCommonTailLength = 0U; 519 SameTails.clear(); 520 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 521 MPIterator HighestMPIter = prior(MergePotentials.end()); 522 for (MPIterator CurMPIter = prior(MergePotentials.end()), 523 B = MergePotentials.begin(); 524 CurMPIter!=B && CurMPIter->first==CurHash; 525 --CurMPIter) { 526 for (MPIterator I = prior(CurMPIter); I->first==CurHash ; --I) { 527 unsigned CommonTailLen = ComputeCommonTailLength( 528 CurMPIter->second, 529 I->second, 530 TrialBBI1, TrialBBI2); 531 if (CommonTailLen >= minCommonTailLength) { 532 if (CommonTailLen > maxCommonTailLength) { 533 SameTails.clear(); 534 maxCommonTailLength = CommonTailLen; 535 HighestMPIter = CurMPIter; 536 SameTails.push_back(std::make_pair(CurMPIter, TrialBBI1)); 537 } 538 if (HighestMPIter == CurMPIter && 539 CommonTailLen == maxCommonTailLength) 540 SameTails.push_back(std::make_pair(I, TrialBBI2)); 541 } 542 if (I==B) 543 break; 544 } 545 } 546 return maxCommonTailLength; 547} 548 549/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 550/// MergePotentials, restoring branches at ends of blocks as appropriate. 551void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 552 MachineBasicBlock* SuccBB, 553 MachineBasicBlock* PredBB) { 554 for (MPIterator CurMPIter = prior(MergePotentials.end()), 555 B = MergePotentials.begin(); 556 CurMPIter->first==CurHash; 557 --CurMPIter) { 558 // Put the unconditional branch back, if we need one. 559 MachineBasicBlock *CurMBB = CurMPIter->second; 560 if (SuccBB && CurMBB != PredBB) 561 FixTail(CurMBB, SuccBB, TII); 562 MergePotentials.erase(CurMPIter); 563 if (CurMPIter==B) 564 break; 565 } 566} 567 568// See if any of the blocks in MergePotentials (which all have a common single 569// successor, or all have no successor) can be tail-merged. If there is a 570// successor, any blocks in MergePotentials that are not tail-merged and 571// are not immediately before Succ must have an unconditional branch to 572// Succ added (but the predecessor/successor lists need no adjustment). 573// The lone predecessor of Succ that falls through into Succ, 574// if any, is given in PredBB. 575 576bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, 577 MachineBasicBlock* PredBB) { 578 // We cannot jump to the entry block, which affects various choices below. 579 MachineBasicBlock *Entry = MergePotentials.begin()->second-> 580 getParent()->begin(); 581 582 // It doesn't make sense to save a single instruction since tail merging 583 // will add a jump. 584 // FIXME: Ask the target to provide the threshold? 585 unsigned minCommonTailLength = (SuccBB ? 1 : 2) + 1; 586 MadeChange = false; 587 588 DOUT << "\nTryMergeBlocks " << MergePotentials.size(); 589 590 // Sort by hash value so that blocks with identical end sequences sort 591 // together. 592 std::stable_sort(MergePotentials.begin(), MergePotentials.end(),MergeCompare); 593 594 // Walk through equivalence sets looking for actual exact matches. 595 while (MergePotentials.size() > 1) { 596 unsigned CurHash = prior(MergePotentials.end())->first; 597 598 // Build SameTails, identifying the set of blocks with this hash code 599 // and with the maximum number of instructions in common. 600 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 601 minCommonTailLength); 602 603 // If we didn't find any pair that has at least minCommonTailLength 604 // instructions in common, remove all blocks with this hash code and retry. 605 if (SameTails.empty()) { 606 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 607 continue; 608 } 609 610 // If one of the blocks is the entire common tail (and not the entry 611 // block, which we can't jump to), treat all blocks with this same 612 // tail at once. 613 unsigned int i; 614 for (i=0; i<SameTails.size(); i++) { 615 MachineBasicBlock *MBB = SameTails[i].first->second; 616 if (MBB->begin() == SameTails[i].second && MBB != Entry) 617 break; 618 } 619 if (i!=SameTails.size()) { 620 MachineBasicBlock *MBB = SameTails[i].first->second; 621 // MBB is common tail. Adjust all other BB's to jump to this one. 622 // Traversal must be forwards so erases work. 623 DOUT << "\nUsing common tail " << MBB->getNumber() << " for "; 624 for (unsigned int j=0; j<SameTails.size(); ++j) { 625 if (i==j) 626 continue; 627 DOUT << SameTails[j].first->second->getNumber() << ","; 628 // Hack the end off BB j, making it jump to BB i instead. 629 ReplaceTailWithBranchTo(SameTails[j].second, MBB); 630 // This modifies BB j, so remove it from the worklist. 631 MergePotentials.erase(SameTails[j].first); 632 } 633 DOUT << "\n"; 634 // We leave i in the worklist in case there are other blocks that 635 // match it with a smaller number of instructions. 636 MadeChange = true; 637 continue; 638 } 639 640 // Otherwise, merge the 2 blocks in SameTails that are latest in 641 // MergePotentials; these are at indices 0 and 1 in SameTails. 642 MachineBasicBlock::iterator BBI1 = (SameTails[0]).second; 643 MachineBasicBlock::iterator BBI2 = (SameTails[1]).second; 644 MachineBasicBlock *MBB1 = (SameTails[0]).first->second; 645 MachineBasicBlock *MBB2 = (SameTails[1]).first->second; 646 647 DOUT << "\nMerging " << MBB1->getNumber() << "," << 648 MBB2->getNumber() << ", size " << maxCommonTailLength; 649 650 // Neither block is the entire common tail; split the tail of one block 651 // to make it redundant with the other tail. We cannot jump to the 652 // entry block, so if one block is the entry block, split the other one. 653 654 // The second half of the split block will remain in SameTails, and will 655 // consist entirely of common code. Thus in the case where there are 656 // multiple blocks that would all need to be split, the next iteration of 657 // the outer loop will handle all the rest of them. 658 659 // Decide whether we want to split MBB1 or MBB2. 660 if (ShouldSplitFirstBlock(MBB1, BBI1, MBB2, BBI2, PredBB)) { 661 MBB1 = SplitMBBAt(*MBB1, BBI1); 662 BBI1 = MBB1->begin(); 663 SameTails[0].first->second = MBB1; 664 } else { 665 MBB2 = SplitMBBAt(*MBB2, BBI2); 666 BBI2 = MBB2->begin(); 667 SameTails[1].first->second = MBB2; 668 } 669 670 if (MBB2->begin() == BBI2 && MBB2 != Entry) { 671 // Hack the end off MBB1, making it jump to MBB2 instead. 672 ReplaceTailWithBranchTo(BBI1, MBB2); 673 // This modifies MBB1, so remove it from the worklist. 674 MergePotentials.erase(SameTails[0].first); 675 } else { 676 assert(MBB1->begin() == BBI1 && MBB1 != Entry && 677 "Didn't split block correctly?"); 678 // Hack the end off MBB2, making it jump to MBB1 instead. 679 ReplaceTailWithBranchTo(BBI2, MBB1); 680 // This modifies MBB2, so remove it from the worklist. 681 MergePotentials.erase(SameTails[1].first); 682 } 683 MadeChange = true; 684 } 685 return MadeChange; 686} 687 688bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 689 690 if (!EnableTailMerge) return false; 691 692 MadeChange = false; 693 694 // First find blocks with no successors. 695 MergePotentials.clear(); 696 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 697 if (I->succ_empty()) 698 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I)); 699 } 700 // See if we can do any tail merging on those. 701 if (MergePotentials.size() < TailMergeThreshold && 702 MergePotentials.size() >= 2) 703 MadeChange |= TryMergeBlocks(NULL, NULL); 704 705 // Look at blocks (IBB) with multiple predecessors (PBB). 706 // We change each predecessor to a canonical form, by 707 // (1) temporarily removing any unconditional branch from the predecessor 708 // to IBB, and 709 // (2) alter conditional branches so they branch to the other block 710 // not IBB; this may require adding back an unconditional branch to IBB 711 // later, where there wasn't one coming in. E.g. 712 // Bcc IBB 713 // fallthrough to QBB 714 // here becomes 715 // Bncc QBB 716 // with a conceptual B to IBB after that, which never actually exists. 717 // With those changes, we see whether the predecessors' tails match, 718 // and merge them if so. We change things out of canonical form and 719 // back to the way they were later in the process. (OptimizeBranches 720 // would undo some of this, but we can't use it, because we'd get into 721 // a compile-time infinite loop repeatedly doing and undoing the same 722 // transformations.) 723 724 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 725 if (!I->succ_empty() && I->pred_size() >= 2 && 726 I->pred_size() < TailMergeThreshold) { 727 MachineBasicBlock *IBB = I; 728 MachineBasicBlock *PredBB = prior(I); 729 MergePotentials.clear(); 730 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 731 E2 = I->pred_end(); 732 P != E2; ++P) { 733 MachineBasicBlock* PBB = *P; 734 // Skip blocks that loop to themselves, can't tail merge these. 735 if (PBB==IBB) 736 continue; 737 MachineBasicBlock *TBB = 0, *FBB = 0; 738 std::vector<MachineOperand> Cond; 739 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond)) { 740 // Failing case: IBB is the target of a cbr, and 741 // we cannot reverse the branch. 742 std::vector<MachineOperand> NewCond(Cond); 743 if (!Cond.empty() && TBB==IBB) { 744 if (TII->ReverseBranchCondition(NewCond)) 745 continue; 746 // This is the QBB case described above 747 if (!FBB) 748 FBB = next(MachineFunction::iterator(PBB)); 749 } 750 // Failing case: the only way IBB can be reached from PBB is via 751 // exception handling. Happens for landing pads. Would be nice 752 // to have a bit in the edge so we didn't have to do all this. 753 if (IBB->isLandingPad()) { 754 MachineFunction::iterator IP = PBB; IP++; 755 MachineBasicBlock* PredNextBB = NULL; 756 if (IP!=MF.end()) 757 PredNextBB = IP; 758 if (TBB==NULL) { 759 if (IBB!=PredNextBB) // fallthrough 760 continue; 761 } else if (FBB) { 762 if (TBB!=IBB && FBB!=IBB) // cbr then ubr 763 continue; 764 } else if (Cond.empty()) { 765 if (TBB!=IBB) // ubr 766 continue; 767 } else { 768 if (TBB!=IBB && IBB!=PredNextBB) // cbr 769 continue; 770 } 771 } 772 // Remove the unconditional branch at the end, if any. 773 if (TBB && (Cond.empty() || FBB)) { 774 TII->RemoveBranch(*PBB); 775 if (!Cond.empty()) 776 // reinsert conditional branch only, for now 777 TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond); 778 } 779 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P)); 780 } 781 } 782 if (MergePotentials.size() >= 2) 783 MadeChange |= TryMergeBlocks(I, PredBB); 784 // Reinsert an unconditional branch if needed. 785 // The 1 below can be either an original single predecessor, or a result 786 // of removing blocks in TryMergeBlocks. 787 PredBB = prior(I); // this may have been changed in TryMergeBlocks 788 if (MergePotentials.size()==1 && 789 (MergePotentials.begin())->second != PredBB) 790 FixTail((MergePotentials.begin())->second, I, TII); 791 } 792 } 793 return MadeChange; 794} 795 796//===----------------------------------------------------------------------===// 797// Branch Optimization 798//===----------------------------------------------------------------------===// 799 800bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 801 MadeChange = false; 802 803 // Make sure blocks are numbered in order 804 MF.RenumberBlocks(); 805 806 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 807 MachineBasicBlock *MBB = I++; 808 OptimizeBlock(MBB); 809 810 // If it is dead, remove it. 811 if (MBB->pred_empty()) { 812 RemoveDeadBlock(MBB); 813 MadeChange = true; 814 ++NumDeadBlocks; 815 } 816 } 817 return MadeChange; 818} 819 820 821/// CanFallThrough - Return true if the specified block (with the specified 822/// branch condition) can implicitly transfer control to the block after it by 823/// falling off the end of it. This should return false if it can reach the 824/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 825/// 826/// True is a conservative answer. 827/// 828bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 829 bool BranchUnAnalyzable, 830 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 831 const std::vector<MachineOperand> &Cond) { 832 MachineFunction::iterator Fallthrough = CurBB; 833 ++Fallthrough; 834 // If FallthroughBlock is off the end of the function, it can't fall through. 835 if (Fallthrough == CurBB->getParent()->end()) 836 return false; 837 838 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 839 if (!CurBB->isSuccessor(Fallthrough)) 840 return false; 841 842 // If we couldn't analyze the branch, assume it could fall through. 843 if (BranchUnAnalyzable) return true; 844 845 // If there is no branch, control always falls through. 846 if (TBB == 0) return true; 847 848 // If there is some explicit branch to the fallthrough block, it can obviously 849 // reach, even though the branch should get folded to fall through implicitly. 850 if (MachineFunction::iterator(TBB) == Fallthrough || 851 MachineFunction::iterator(FBB) == Fallthrough) 852 return true; 853 854 // If it's an unconditional branch to some block not the fall through, it 855 // doesn't fall through. 856 if (Cond.empty()) return false; 857 858 // Otherwise, if it is conditional and has no explicit false block, it falls 859 // through. 860 return FBB == 0; 861} 862 863/// CanFallThrough - Return true if the specified can implicitly transfer 864/// control to the block after it by falling off the end of it. This should 865/// return false if it can reach the block after it, but it uses an explicit 866/// branch to do so (e.g. a table jump). 867/// 868/// True is a conservative answer. 869/// 870bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 871 MachineBasicBlock *TBB = 0, *FBB = 0; 872 std::vector<MachineOperand> Cond; 873 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond); 874 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 875} 876 877/// IsBetterFallthrough - Return true if it would be clearly better to 878/// fall-through to MBB1 than to fall through into MBB2. This has to return 879/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 880/// result in infinite loops. 881static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 882 MachineBasicBlock *MBB2) { 883 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 884 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 885 // optimize branches that branch to either a return block or an assert block 886 // into a fallthrough to the return. 887 if (MBB1->empty() || MBB2->empty()) return false; 888 889 // If there is a clear successor ordering we make sure that one block 890 // will fall through to the next 891 if (MBB1->isSuccessor(MBB2)) return true; 892 if (MBB2->isSuccessor(MBB1)) return false; 893 894 MachineInstr *MBB1I = --MBB1->end(); 895 MachineInstr *MBB2I = --MBB2->end(); 896 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 897} 898 899/// OptimizeBlock - Analyze and optimize control flow related to the specified 900/// block. This is never called on the entry block. 901void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 902 MachineFunction::iterator FallThrough = MBB; 903 ++FallThrough; 904 905 // If this block is empty, make everyone use its fall-through, not the block 906 // explicitly. Landing pads should not do this since the landing-pad table 907 // points to this block. 908 if (MBB->empty() && !MBB->isLandingPad()) { 909 // Dead block? Leave for cleanup later. 910 if (MBB->pred_empty()) return; 911 912 if (FallThrough == MBB->getParent()->end()) { 913 // TODO: Simplify preds to not branch here if possible! 914 } else { 915 // Rewrite all predecessors of the old block to go to the fallthrough 916 // instead. 917 while (!MBB->pred_empty()) { 918 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 919 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 920 } 921 922 // If MBB was the target of a jump table, update jump tables to go to the 923 // fallthrough instead. 924 MBB->getParent()->getJumpTableInfo()-> 925 ReplaceMBBInJumpTables(MBB, FallThrough); 926 MadeChange = true; 927 } 928 return; 929 } 930 931 // Check to see if we can simplify the terminator of the block before this 932 // one. 933 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 934 935 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 936 std::vector<MachineOperand> PriorCond; 937 bool PriorUnAnalyzable = 938 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 939 if (!PriorUnAnalyzable) { 940 // If the CFG for the prior block has extra edges, remove them. 941 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 942 !PriorCond.empty()); 943 944 // If the previous branch is conditional and both conditions go to the same 945 // destination, remove the branch, replacing it with an unconditional one or 946 // a fall-through. 947 if (PriorTBB && PriorTBB == PriorFBB) { 948 TII->RemoveBranch(PrevBB); 949 PriorCond.clear(); 950 if (PriorTBB != MBB) 951 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 952 MadeChange = true; 953 ++NumBranchOpts; 954 return OptimizeBlock(MBB); 955 } 956 957 // If the previous branch *only* branches to *this* block (conditional or 958 // not) remove the branch. 959 if (PriorTBB == MBB && PriorFBB == 0) { 960 TII->RemoveBranch(PrevBB); 961 MadeChange = true; 962 ++NumBranchOpts; 963 return OptimizeBlock(MBB); 964 } 965 966 // If the prior block branches somewhere else on the condition and here if 967 // the condition is false, remove the uncond second branch. 968 if (PriorFBB == MBB) { 969 TII->RemoveBranch(PrevBB); 970 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 971 MadeChange = true; 972 ++NumBranchOpts; 973 return OptimizeBlock(MBB); 974 } 975 976 // If the prior block branches here on true and somewhere else on false, and 977 // if the branch condition is reversible, reverse the branch to create a 978 // fall-through. 979 if (PriorTBB == MBB) { 980 std::vector<MachineOperand> NewPriorCond(PriorCond); 981 if (!TII->ReverseBranchCondition(NewPriorCond)) { 982 TII->RemoveBranch(PrevBB); 983 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 984 MadeChange = true; 985 ++NumBranchOpts; 986 return OptimizeBlock(MBB); 987 } 988 } 989 990 // If this block doesn't fall through (e.g. it ends with an uncond branch or 991 // has no successors) and if the pred falls through into this block, and if 992 // it would otherwise fall through into the block after this, move this 993 // block to the end of the function. 994 // 995 // We consider it more likely that execution will stay in the function (e.g. 996 // due to loops) than it is to exit it. This asserts in loops etc, moving 997 // the assert condition out of the loop body. 998 if (!PriorCond.empty() && PriorFBB == 0 && 999 MachineFunction::iterator(PriorTBB) == FallThrough && 1000 !CanFallThrough(MBB)) { 1001 bool DoTransform = true; 1002 1003 // We have to be careful that the succs of PredBB aren't both no-successor 1004 // blocks. If neither have successors and if PredBB is the second from 1005 // last block in the function, we'd just keep swapping the two blocks for 1006 // last. Only do the swap if one is clearly better to fall through than 1007 // the other. 1008 if (FallThrough == --MBB->getParent()->end() && 1009 !IsBetterFallthrough(PriorTBB, MBB)) 1010 DoTransform = false; 1011 1012 // We don't want to do this transformation if we have control flow like: 1013 // br cond BB2 1014 // BB1: 1015 // .. 1016 // jmp BBX 1017 // BB2: 1018 // .. 1019 // ret 1020 // 1021 // In this case, we could actually be moving the return block *into* a 1022 // loop! 1023 if (DoTransform && !MBB->succ_empty() && 1024 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 1025 DoTransform = false; 1026 1027 1028 if (DoTransform) { 1029 // Reverse the branch so we will fall through on the previous true cond. 1030 std::vector<MachineOperand> NewPriorCond(PriorCond); 1031 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1032 DOUT << "\nMoving MBB: " << *MBB; 1033 DOUT << "To make fallthrough to: " << *PriorTBB << "\n"; 1034 1035 TII->RemoveBranch(PrevBB); 1036 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 1037 1038 // Move this block to the end of the function. 1039 MBB->moveAfter(--MBB->getParent()->end()); 1040 MadeChange = true; 1041 ++NumBranchOpts; 1042 return; 1043 } 1044 } 1045 } 1046 } 1047 1048 // Analyze the branch in the current block. 1049 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1050 std::vector<MachineOperand> CurCond; 1051 bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond); 1052 if (!CurUnAnalyzable) { 1053 // If the CFG for the prior block has extra edges, remove them. 1054 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1055 1056 // If this is a two-way branch, and the FBB branches to this block, reverse 1057 // the condition so the single-basic-block loop is faster. Instead of: 1058 // Loop: xxx; jcc Out; jmp Loop 1059 // we want: 1060 // Loop: xxx; jncc Loop; jmp Out 1061 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1062 std::vector<MachineOperand> NewCond(CurCond); 1063 if (!TII->ReverseBranchCondition(NewCond)) { 1064 TII->RemoveBranch(*MBB); 1065 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 1066 MadeChange = true; 1067 ++NumBranchOpts; 1068 return OptimizeBlock(MBB); 1069 } 1070 } 1071 1072 1073 // If this branch is the only thing in its block, see if we can forward 1074 // other blocks across it. 1075 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1076 MBB->begin()->getDesc().isBranch() && CurTBB != MBB) { 1077 // This block may contain just an unconditional branch. Because there can 1078 // be 'non-branch terminators' in the block, try removing the branch and 1079 // then seeing if the block is empty. 1080 TII->RemoveBranch(*MBB); 1081 1082 // If this block is just an unconditional branch to CurTBB, we can 1083 // usually completely eliminate the block. The only case we cannot 1084 // completely eliminate the block is when the block before this one 1085 // falls through into MBB and we can't understand the prior block's branch 1086 // condition. 1087 if (MBB->empty()) { 1088 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 1089 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1090 !PrevBB.isSuccessor(MBB)) { 1091 // If the prior block falls through into us, turn it into an 1092 // explicit branch to us to make updates simpler. 1093 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1094 PriorTBB != MBB && PriorFBB != MBB) { 1095 if (PriorTBB == 0) { 1096 assert(PriorCond.empty() && PriorFBB == 0 && 1097 "Bad branch analysis"); 1098 PriorTBB = MBB; 1099 } else { 1100 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1101 PriorFBB = MBB; 1102 } 1103 TII->RemoveBranch(PrevBB); 1104 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 1105 } 1106 1107 // Iterate through all the predecessors, revectoring each in-turn. 1108 size_t PI = 0; 1109 bool DidChange = false; 1110 bool HasBranchToSelf = false; 1111 while(PI != MBB->pred_size()) { 1112 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1113 if (PMBB == MBB) { 1114 // If this block has an uncond branch to itself, leave it. 1115 ++PI; 1116 HasBranchToSelf = true; 1117 } else { 1118 DidChange = true; 1119 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1120 } 1121 } 1122 1123 // Change any jumptables to go to the new MBB. 1124 MBB->getParent()->getJumpTableInfo()-> 1125 ReplaceMBBInJumpTables(MBB, CurTBB); 1126 if (DidChange) { 1127 ++NumBranchOpts; 1128 MadeChange = true; 1129 if (!HasBranchToSelf) return; 1130 } 1131 } 1132 } 1133 1134 // Add the branch back if the block is more than just an uncond branch. 1135 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1136 } 1137 } 1138 1139 // If the prior block doesn't fall through into this block, and if this 1140 // block doesn't fall through into some other block, see if we can find a 1141 // place to move this block where a fall-through will happen. 1142 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 1143 PriorTBB, PriorFBB, PriorCond)) { 1144 // Now we know that there was no fall-through into this block, check to 1145 // see if it has a fall-through into its successor. 1146 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1147 CurCond); 1148 1149 if (!MBB->isLandingPad()) { 1150 // Check all the predecessors of this block. If one of them has no fall 1151 // throughs, move this block right after it. 1152 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1153 E = MBB->pred_end(); PI != E; ++PI) { 1154 // Analyze the branch at the end of the pred. 1155 MachineBasicBlock *PredBB = *PI; 1156 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1157 if (PredBB != MBB && !CanFallThrough(PredBB) 1158 && (!CurFallsThru || !CurTBB || !CurFBB) 1159 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1160 // If the current block doesn't fall through, just move it. 1161 // If the current block can fall through and does not end with a 1162 // conditional branch, we need to append an unconditional jump to 1163 // the (current) next block. To avoid a possible compile-time 1164 // infinite loop, move blocks only backward in this case. 1165 // Also, if there are already 2 branches here, we cannot add a third; 1166 // this means we have the case 1167 // Bcc next 1168 // B elsewhere 1169 // next: 1170 if (CurFallsThru) { 1171 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1172 CurCond.clear(); 1173 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1174 } 1175 MBB->moveAfter(PredBB); 1176 MadeChange = true; 1177 return OptimizeBlock(MBB); 1178 } 1179 } 1180 } 1181 1182 if (!CurFallsThru) { 1183 // Check all successors to see if we can move this block before it. 1184 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1185 E = MBB->succ_end(); SI != E; ++SI) { 1186 // Analyze the branch at the end of the block before the succ. 1187 MachineBasicBlock *SuccBB = *SI; 1188 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1189 std::vector<MachineOperand> SuccPrevCond; 1190 1191 // If this block doesn't already fall-through to that successor, and if 1192 // the succ doesn't already have a block that can fall through into it, 1193 // and if the successor isn't an EH destination, we can arrange for the 1194 // fallthrough to happen. 1195 if (SuccBB != MBB && !CanFallThrough(SuccPrev) && 1196 !SuccBB->isLandingPad()) { 1197 MBB->moveBefore(SuccBB); 1198 MadeChange = true; 1199 return OptimizeBlock(MBB); 1200 } 1201 } 1202 1203 // Okay, there is no really great place to put this block. If, however, 1204 // the block before this one would be a fall-through if this block were 1205 // removed, move this block to the end of the function. 1206 if (FallThrough != MBB->getParent()->end() && 1207 PrevBB.isSuccessor(FallThrough)) { 1208 MBB->moveAfter(--MBB->getParent()->end()); 1209 MadeChange = true; 1210 return; 1211 } 1212 } 1213 } 1214} 1215