BranchFolding.cpp revision 3a78bbfeacce9988ff60210fd472cef9789abcc2
1//===-- BranchFolding.cpp - Fold machine code branch instructions ---------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source 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/MRegisterInfo.h" 28#include "llvm/Support/CommandLine.h" 29#include "llvm/Support/Debug.h" 30#include "llvm/ADT/Statistic.h" 31#include "llvm/ADT/STLExtras.h" 32#include <algorithm> 33using namespace llvm; 34 35STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 36STATISTIC(NumBranchOpts, "Number of branches optimized"); 37STATISTIC(NumTailMerge , "Number of block tails merged"); 38static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 39 cl::init(cl::BOU_UNSET), cl::Hidden); 40namespace { 41 // Throttle for huge numbers of predecessors (compile speed problems) 42 cl::opt<unsigned> 43 TailMergeThreshold("tail-merge-threshold", 44 cl::desc("Max number of predecessors to consider tail merging"), 45 cl::init(100), cl::Hidden); 46 47 struct BranchFolder : public MachineFunctionPass { 48 static char ID; 49 explicit BranchFolder(bool defaultEnableTailMerge) : 50 MachineFunctionPass((intptr_t)&ID) { 51 switch (FlagEnableTailMerge) { 52 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 53 case cl::BOU_TRUE: EnableTailMerge = true; break; 54 case cl::BOU_FALSE: EnableTailMerge = false; break; 55 } 56 } 57 58 virtual bool runOnMachineFunction(MachineFunction &MF); 59 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 60 const TargetInstrInfo *TII; 61 MachineModuleInfo *MMI; 62 bool MadeChange; 63 private: 64 // Tail Merging. 65 bool EnableTailMerge; 66 bool TailMergeBlocks(MachineFunction &MF); 67 bool TryMergeBlocks(MachineBasicBlock* SuccBB, 68 MachineBasicBlock* PredBB); 69 void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 70 MachineBasicBlock *NewDest); 71 MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB, 72 MachineBasicBlock::iterator BBI1); 73 74 std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials; 75 const MRegisterInfo *RegInfo; 76 RegScavenger *RS; 77 // Branch optzn. 78 bool OptimizeBranches(MachineFunction &MF); 79 void OptimizeBlock(MachineBasicBlock *MBB); 80 void RemoveDeadBlock(MachineBasicBlock *MBB); 81 82 bool CanFallThrough(MachineBasicBlock *CurBB); 83 bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable, 84 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 85 const std::vector<MachineOperand> &Cond); 86 }; 87 char BranchFolder::ID = 0; 88} 89 90FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 91 return new BranchFolder(DefaultEnableTailMerge); } 92 93/// RemoveDeadBlock - Remove the specified dead machine basic block from the 94/// function, updating the CFG. 95void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 96 assert(MBB->pred_empty() && "MBB must be dead!"); 97 DOUT << "\nRemoving MBB: " << *MBB; 98 99 MachineFunction *MF = MBB->getParent(); 100 // drop all successors. 101 while (!MBB->succ_empty()) 102 MBB->removeSuccessor(MBB->succ_end()-1); 103 104 // If there is DWARF info to active, check to see if there are any LABEL 105 // records in the basic block. If so, unregister them from MachineModuleInfo. 106 if (MMI && !MBB->empty()) { 107 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 108 I != E; ++I) { 109 if ((unsigned)I->getOpcode() == TargetInstrInfo::LABEL) { 110 // The label ID # is always operand #0, an immediate. 111 MMI->InvalidateLabel(I->getOperand(0).getImm()); 112 } 113 } 114 } 115 116 // Remove the block. 117 MF->getBasicBlockList().erase(MBB); 118} 119 120bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { 121 TII = MF.getTarget().getInstrInfo(); 122 if (!TII) return false; 123 124 // Fix CFG. The later algorithms expect it to be right. 125 bool EverMadeChange = false; 126 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 127 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 128 std::vector<MachineOperand> Cond; 129 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond)) 130 EverMadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 131 } 132 133 RegInfo = MF.getTarget().getRegisterInfo(); 134 RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 135 136 MMI = getAnalysisToUpdate<MachineModuleInfo>(); 137 138 bool MadeChangeThisIteration = true; 139 while (MadeChangeThisIteration) { 140 MadeChangeThisIteration = false; 141 MadeChangeThisIteration |= TailMergeBlocks(MF); 142 MadeChangeThisIteration |= OptimizeBranches(MF); 143 EverMadeChange |= MadeChangeThisIteration; 144 } 145 146 // See if any jump tables have become mergable or dead as the code generator 147 // did its thing. 148 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 149 const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables(); 150 if (!JTs.empty()) { 151 // Figure out how these jump tables should be merged. 152 std::vector<unsigned> JTMapping; 153 JTMapping.reserve(JTs.size()); 154 155 // We always keep the 0th jump table. 156 JTMapping.push_back(0); 157 158 // Scan the jump tables, seeing if there are any duplicates. Note that this 159 // is N^2, which should be fixed someday. 160 for (unsigned i = 1, e = JTs.size(); i != e; ++i) 161 JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); 162 163 // If a jump table was merge with another one, walk the function rewriting 164 // references to jump tables to reference the new JT ID's. Keep track of 165 // whether we see a jump table idx, if not, we can delete the JT. 166 std::vector<bool> JTIsLive; 167 JTIsLive.resize(JTs.size()); 168 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 169 BB != E; ++BB) { 170 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 171 I != E; ++I) 172 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 173 MachineOperand &Op = I->getOperand(op); 174 if (!Op.isJumpTableIndex()) continue; 175 unsigned NewIdx = JTMapping[Op.getJumpTableIndex()]; 176 Op.setJumpTableIndex(NewIdx); 177 178 // Remember that this JT is live. 179 JTIsLive[NewIdx] = true; 180 } 181 } 182 183 // Finally, remove dead jump tables. This happens either because the 184 // indirect jump was unreachable (and thus deleted) or because the jump 185 // table was merged with some other one. 186 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 187 if (!JTIsLive[i]) { 188 JTI->RemoveJumpTable(i); 189 EverMadeChange = true; 190 } 191 } 192 193 delete RS; 194 return EverMadeChange; 195} 196 197//===----------------------------------------------------------------------===// 198// Tail Merging of Blocks 199//===----------------------------------------------------------------------===// 200 201/// HashMachineInstr - Compute a hash value for MI and its operands. 202static unsigned HashMachineInstr(const MachineInstr *MI) { 203 unsigned Hash = MI->getOpcode(); 204 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 205 const MachineOperand &Op = MI->getOperand(i); 206 207 // Merge in bits from the operand if easy. 208 unsigned OperandHash = 0; 209 switch (Op.getType()) { 210 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 211 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 212 case MachineOperand::MO_MachineBasicBlock: 213 OperandHash = Op.getMachineBasicBlock()->getNumber(); 214 break; 215 case MachineOperand::MO_FrameIndex: OperandHash = Op.getFrameIndex(); break; 216 case MachineOperand::MO_ConstantPoolIndex: 217 OperandHash = Op.getConstantPoolIndex(); 218 break; 219 case MachineOperand::MO_JumpTableIndex: 220 OperandHash = Op.getJumpTableIndex(); 221 break; 222 case MachineOperand::MO_GlobalAddress: 223 case MachineOperand::MO_ExternalSymbol: 224 // Global address / external symbol are too hard, don't bother, but do 225 // pull in the offset. 226 OperandHash = Op.getOffset(); 227 break; 228 default: break; 229 } 230 231 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 232 } 233 return Hash; 234} 235 236/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks 237/// with no successors, we hash two instructions, because cross-jumping 238/// only saves code when at least two instructions are removed (since a 239/// branch must be inserted). For blocks with a successor, one of the 240/// two blocks to be tail-merged will end with a branch already, so 241/// it gains to cross-jump even for one instruction. 242 243static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, 244 unsigned minCommonTailLength) { 245 MachineBasicBlock::const_iterator I = MBB->end(); 246 if (I == MBB->begin()) 247 return 0; // Empty MBB. 248 249 --I; 250 unsigned Hash = HashMachineInstr(I); 251 252 if (I == MBB->begin() || minCommonTailLength == 1) 253 return Hash; // Single instr MBB. 254 255 --I; 256 // Hash in the second-to-last instruction. 257 Hash ^= HashMachineInstr(I) << 2; 258 return Hash; 259} 260 261/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 262/// of instructions they actually have in common together at their end. Return 263/// iterators for the first shared instruction in each block. 264static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 265 MachineBasicBlock *MBB2, 266 MachineBasicBlock::iterator &I1, 267 MachineBasicBlock::iterator &I2) { 268 I1 = MBB1->end(); 269 I2 = MBB2->end(); 270 271 unsigned TailLen = 0; 272 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 273 --I1; --I2; 274 if (!I1->isIdenticalTo(I2)) { 275 ++I1; ++I2; 276 break; 277 } 278 ++TailLen; 279 } 280 return TailLen; 281} 282 283/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 284/// after it, replacing it with an unconditional branch to NewDest. This 285/// returns true if OldInst's block is modified, false if NewDest is modified. 286void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 287 MachineBasicBlock *NewDest) { 288 MachineBasicBlock *OldBB = OldInst->getParent(); 289 290 // Remove all the old successors of OldBB from the CFG. 291 while (!OldBB->succ_empty()) 292 OldBB->removeSuccessor(OldBB->succ_begin()); 293 294 // Remove all the dead instructions from the end of OldBB. 295 OldBB->erase(OldInst, OldBB->end()); 296 297 // If OldBB isn't immediately before OldBB, insert a branch to it. 298 if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) 299 TII->InsertBranch(*OldBB, NewDest, 0, std::vector<MachineOperand>()); 300 OldBB->addSuccessor(NewDest); 301 ++NumTailMerge; 302} 303 304/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 305/// MBB so that the part before the iterator falls into the part starting at the 306/// iterator. This returns the new MBB. 307MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 308 MachineBasicBlock::iterator BBI1) { 309 // Create the fall-through block. 310 MachineFunction::iterator MBBI = &CurMBB; 311 MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock()); 312 CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB); 313 314 // Move all the successors of this block to the specified block. 315 while (!CurMBB.succ_empty()) { 316 MachineBasicBlock *S = *(CurMBB.succ_end()-1); 317 NewMBB->addSuccessor(S); 318 CurMBB.removeSuccessor(S); 319 } 320 321 // Add an edge from CurMBB to NewMBB for the fall-through. 322 CurMBB.addSuccessor(NewMBB); 323 324 // Splice the code over. 325 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 326 327 // For targets that use the register scavenger, we must maintain LiveIns. 328 if (RS) { 329 RS->enterBasicBlock(&CurMBB); 330 if (!CurMBB.empty()) 331 RS->forward(prior(CurMBB.end())); 332 BitVector RegsLiveAtExit(RegInfo->getNumRegs()); 333 RS->getRegsUsed(RegsLiveAtExit, false); 334 for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++) 335 if (RegsLiveAtExit[i]) 336 NewMBB->addLiveIn(i); 337 } 338 339 return NewMBB; 340} 341 342/// EstimateRuntime - Make a rough estimate for how long it will take to run 343/// the specified code. 344static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 345 MachineBasicBlock::iterator E, 346 const TargetInstrInfo *TII) { 347 unsigned Time = 0; 348 for (; I != E; ++I) { 349 const TargetInstrDescriptor &TID = TII->get(I->getOpcode()); 350 if (TID.Flags & M_CALL_FLAG) 351 Time += 10; 352 else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG)) 353 Time += 2; 354 else 355 ++Time; 356 } 357 return Time; 358} 359 360/// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at 361/// MBB2I and then insert an unconditional branch in the other block. Determine 362/// which is the best to split 363static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1, 364 MachineBasicBlock::iterator MBB1I, 365 MachineBasicBlock *MBB2, 366 MachineBasicBlock::iterator MBB2I, 367 const TargetInstrInfo *TII, 368 MachineBasicBlock *PredBB) { 369 // If one block is the entry block, split the other one; we can't generate 370 // a branch to the entry block, as its label is not emitted. 371 MachineBasicBlock *Entry = MBB1->getParent()->begin(); 372 if (MBB1 == Entry) 373 return false; 374 if (MBB2 == Entry) 375 return true; 376 377 // If one block falls through into the common successor, choose that 378 // one to split; it is one instruction less to do that. 379 if (PredBB) { 380 if (MBB1 == PredBB) 381 return true; 382 else if (MBB2 == PredBB) 383 return false; 384 } 385 // TODO: if we had some notion of which block was hotter, we could split 386 // the hot block, so it is the fall-through. Since we don't have profile info 387 // make a decision based on which will hurt most to split. 388 unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I, TII); 389 unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I, TII); 390 391 // If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the 392 // MBB1 block so it falls through. This will penalize the MBB2 path, but will 393 // have a lower overall impact on the program execution. 394 return MBB1Time < MBB2Time; 395} 396 397// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 398// branches temporarily for tail merging). In the case where CurMBB ends 399// with a conditional branch to the next block, optimize by reversing the 400// test and conditionally branching to SuccMBB instead. 401 402static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, 403 const TargetInstrInfo *TII) { 404 MachineFunction *MF = CurMBB->getParent(); 405 MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); 406 MachineBasicBlock *TBB = 0, *FBB = 0; 407 std::vector<MachineOperand> Cond; 408 if (I != MF->end() && 409 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) { 410 MachineBasicBlock *NextBB = I; 411 if (TBB == NextBB && Cond.size() && !FBB) { 412 if (!TII->ReverseBranchCondition(Cond)) { 413 TII->RemoveBranch(*CurMBB); 414 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); 415 return; 416 } 417 } 418 } 419 TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector<MachineOperand>()); 420} 421 422static bool MergeCompare(const std::pair<unsigned,MachineBasicBlock*> &p, 423 const std::pair<unsigned,MachineBasicBlock*> &q) { 424 if (p.first < q.first) 425 return true; 426 else if (p.first > q.first) 427 return false; 428 else if (p.second->getNumber() < q.second->getNumber()) 429 return true; 430 else if (p.second->getNumber() > q.second->getNumber()) 431 return false; 432 else { 433 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 434 // an object with itself. 435#ifndef _GLIBCXX_DEBUG 436 assert(0 && "Predecessor appears twice"); 437#endif 438 return(false); 439 } 440} 441 442// See if any of the blocks in MergePotentials (which all have a common single 443// successor, or all have no successor) can be tail-merged. If there is a 444// successor, any blocks in MergePotentials that are not tail-merged and 445// are not immediately before Succ must have an unconditional branch to 446// Succ added (but the predecessor/successor lists need no adjustment). 447// The lone predecessor of Succ that falls through into Succ, 448// if any, is given in PredBB. 449 450bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, 451 MachineBasicBlock* PredBB) { 452 unsigned minCommonTailLength = (SuccBB ? 1 : 2); 453 MadeChange = false; 454 455 // Sort by hash value so that blocks with identical end sequences sort 456 // together. 457 std::stable_sort(MergePotentials.begin(), MergePotentials.end(), MergeCompare); 458 459 // Walk through equivalence sets looking for actual exact matches. 460 while (MergePotentials.size() > 1) { 461 unsigned CurHash = (MergePotentials.end()-1)->first; 462 unsigned PrevHash = (MergePotentials.end()-2)->first; 463 MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second; 464 465 // If there is nothing that matches the hash of the current basic block, 466 // give up. 467 if (CurHash != PrevHash) { 468 if (SuccBB && CurMBB != PredBB) 469 FixTail(CurMBB, SuccBB, TII); 470 MergePotentials.pop_back(); 471 continue; 472 } 473 474 // Look through all the pairs of blocks that have the same hash as this 475 // one, and find the pair that has the largest number of instructions in 476 // common. 477 // Since instructions may get combined later (e.g. single stores into 478 // store multiple) this measure is not particularly accurate. 479 MachineBasicBlock::iterator BBI1, BBI2; 480 481 unsigned FoundI = ~0U, FoundJ = ~0U; 482 unsigned maxCommonTailLength = 0U; 483 for (int i = MergePotentials.size()-1; 484 i != -1 && MergePotentials[i].first == CurHash; --i) { 485 for (int j = i-1; 486 j != -1 && MergePotentials[j].first == CurHash; --j) { 487 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 488 unsigned CommonTailLen = ComputeCommonTailLength( 489 MergePotentials[i].second, 490 MergePotentials[j].second, 491 TrialBBI1, TrialBBI2); 492 if (CommonTailLen >= minCommonTailLength && 493 CommonTailLen > maxCommonTailLength) { 494 FoundI = i; 495 FoundJ = j; 496 maxCommonTailLength = CommonTailLen; 497 BBI1 = TrialBBI1; 498 BBI2 = TrialBBI2; 499 } 500 } 501 } 502 503 // If we didn't find any pair that has at least minCommonTailLength 504 // instructions in common, bail out. All entries with this 505 // hash code can go away now. 506 if (FoundI == ~0U) { 507 for (int i = MergePotentials.size()-1; 508 i != -1 && MergePotentials[i].first == CurHash; --i) { 509 // Put the unconditional branch back, if we need one. 510 CurMBB = MergePotentials[i].second; 511 if (SuccBB && CurMBB != PredBB) 512 FixTail(CurMBB, SuccBB, TII); 513 MergePotentials.pop_back(); 514 } 515 continue; 516 } 517 518 // Otherwise, move the block(s) to the right position(s). So that 519 // BBI1/2 will be valid, the last must be I and the next-to-last J. 520 if (FoundI != MergePotentials.size()-1) 521 std::swap(MergePotentials[FoundI], *(MergePotentials.end()-1)); 522 if (FoundJ != MergePotentials.size()-2) 523 std::swap(MergePotentials[FoundJ], *(MergePotentials.end()-2)); 524 525 CurMBB = (MergePotentials.end()-1)->second; 526 MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second; 527 528 // If neither block is the entire common tail, split the tail of one block 529 // to make it redundant with the other tail. Also, we cannot jump to the 530 // entry block, so if one block is the entry block, split the other one. 531 MachineBasicBlock *Entry = CurMBB->getParent()->begin(); 532 if (CurMBB->begin() == BBI1 && CurMBB != Entry) 533 ; // CurMBB is common tail 534 else if (MBB2->begin() == BBI2 && MBB2 != Entry) 535 ; // MBB2 is common tail 536 else { 537 if (0) { // Enable this to disable partial tail merges. 538 MergePotentials.pop_back(); 539 continue; 540 } 541 542 // Decide whether we want to split CurMBB or MBB2. 543 if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII, PredBB)) { 544 CurMBB = SplitMBBAt(*CurMBB, BBI1); 545 BBI1 = CurMBB->begin(); 546 MergePotentials.back().second = CurMBB; 547 } else { 548 MBB2 = SplitMBBAt(*MBB2, BBI2); 549 BBI2 = MBB2->begin(); 550 (MergePotentials.end()-2)->second = MBB2; 551 } 552 } 553 554 if (MBB2->begin() == BBI2 && MBB2 != Entry) { 555 // Hack the end off CurMBB, making it jump to MBBI@ instead. 556 ReplaceTailWithBranchTo(BBI1, MBB2); 557 // This modifies CurMBB, so remove it from the worklist. 558 MergePotentials.pop_back(); 559 } else { 560 assert(CurMBB->begin() == BBI1 && CurMBB != Entry && 561 "Didn't split block correctly?"); 562 // Hack the end off MBB2, making it jump to CurMBB instead. 563 ReplaceTailWithBranchTo(BBI2, CurMBB); 564 // This modifies MBB2, so remove it from the worklist. 565 MergePotentials.erase(MergePotentials.end()-2); 566 } 567 MadeChange = true; 568 } 569 return MadeChange; 570} 571 572bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 573 574 if (!EnableTailMerge) return false; 575 576 MadeChange = false; 577 578 // First find blocks with no successors. 579 MergePotentials.clear(); 580 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 581 if (I->succ_empty()) 582 MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I)); 583 } 584 // See if we can do any tail merging on those. 585 if (MergePotentials.size() < TailMergeThreshold) 586 MadeChange |= TryMergeBlocks(NULL, NULL); 587 588 // Look at blocks (IBB) with multiple predecessors (PBB). 589 // We change each predecessor to a canonical form, by 590 // (1) temporarily removing any unconditional branch from the predecessor 591 // to IBB, and 592 // (2) alter conditional branches so they branch to the other block 593 // not IBB; this may require adding back an unconditional branch to IBB 594 // later, where there wasn't one coming in. E.g. 595 // Bcc IBB 596 // fallthrough to QBB 597 // here becomes 598 // Bncc QBB 599 // with a conceptual B to IBB after that, which never actually exists. 600 // With those changes, we see whether the predecessors' tails match, 601 // and merge them if so. We change things out of canonical form and 602 // back to the way they were later in the process. (OptimizeBranches 603 // would undo some of this, but we can't use it, because we'd get into 604 // a compile-time infinite loop repeatedly doing and undoing the same 605 // transformations.) 606 607 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 608 if (!I->succ_empty() && I->pred_size() >= 2 && 609 I->pred_size() < TailMergeThreshold) { 610 MachineBasicBlock *IBB = I; 611 MachineBasicBlock *PredBB = prior(I); 612 MergePotentials.clear(); 613 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 614 E2 = I->pred_end(); 615 P != E2; ++P) { 616 MachineBasicBlock* PBB = *P; 617 // Skip blocks that loop to themselves, can't tail merge these. 618 if (PBB==IBB) 619 continue; 620 MachineBasicBlock *TBB = 0, *FBB = 0; 621 std::vector<MachineOperand> Cond; 622 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond)) { 623 // Failing case: IBB is the target of a cbr, and 624 // we cannot reverse the branch. 625 std::vector<MachineOperand> NewCond(Cond); 626 if (Cond.size() && TBB==IBB) { 627 if (TII->ReverseBranchCondition(NewCond)) 628 continue; 629 // This is the QBB case described above 630 if (!FBB) 631 FBB = next(MachineFunction::iterator(PBB)); 632 } 633 // Failing case: the only way IBB can be reached from PBB is via 634 // exception handling. Happens for landing pads. Would be nice 635 // to have a bit in the edge so we didn't have to do all this. 636 if (IBB->isLandingPad()) { 637 MachineFunction::iterator IP = PBB; IP++; 638 MachineBasicBlock* PredNextBB = NULL; 639 if (IP!=MF.end()) 640 PredNextBB = IP; 641 if (TBB==NULL) { 642 if (IBB!=PredNextBB) // fallthrough 643 continue; 644 } else if (FBB) { 645 if (TBB!=IBB && FBB!=IBB) // cbr then ubr 646 continue; 647 } else if (Cond.size() == 0) { 648 if (TBB!=IBB) // ubr 649 continue; 650 } else { 651 if (TBB!=IBB && IBB!=PredNextBB) // cbr 652 continue; 653 } 654 } 655 // Remove the unconditional branch at the end, if any. 656 if (TBB && (Cond.size()==0 || FBB)) { 657 TII->RemoveBranch(*PBB); 658 if (Cond.size()) 659 // reinsert conditional branch only, for now 660 TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond); 661 } 662 MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P)); 663 } 664 } 665 if (MergePotentials.size() >= 2) 666 MadeChange |= TryMergeBlocks(I, PredBB); 667 // Reinsert an unconditional branch if needed. 668 // The 1 below can be either an original single predecessor, or a result 669 // of removing blocks in TryMergeBlocks. 670 PredBB = prior(I); // this may have been changed in TryMergeBlocks 671 if (MergePotentials.size()==1 && 672 (MergePotentials.begin())->second != PredBB) 673 FixTail((MergePotentials.begin())->second, I, TII); 674 } 675 } 676 return MadeChange; 677} 678 679//===----------------------------------------------------------------------===// 680// Branch Optimization 681//===----------------------------------------------------------------------===// 682 683bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 684 MadeChange = false; 685 686 // Make sure blocks are numbered in order 687 MF.RenumberBlocks(); 688 689 for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { 690 MachineBasicBlock *MBB = I++; 691 OptimizeBlock(MBB); 692 693 // If it is dead, remove it. 694 if (MBB->pred_empty()) { 695 RemoveDeadBlock(MBB); 696 MadeChange = true; 697 ++NumDeadBlocks; 698 } 699 } 700 return MadeChange; 701} 702 703 704/// CanFallThrough - Return true if the specified block (with the specified 705/// branch condition) can implicitly transfer control to the block after it by 706/// falling off the end of it. This should return false if it can reach the 707/// block after it, but it uses an explicit branch to do so (e.g. a table jump). 708/// 709/// True is a conservative answer. 710/// 711bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, 712 bool BranchUnAnalyzable, 713 MachineBasicBlock *TBB, MachineBasicBlock *FBB, 714 const std::vector<MachineOperand> &Cond) { 715 MachineFunction::iterator Fallthrough = CurBB; 716 ++Fallthrough; 717 // If FallthroughBlock is off the end of the function, it can't fall through. 718 if (Fallthrough == CurBB->getParent()->end()) 719 return false; 720 721 // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. 722 if (!CurBB->isSuccessor(Fallthrough)) 723 return false; 724 725 // If we couldn't analyze the branch, assume it could fall through. 726 if (BranchUnAnalyzable) return true; 727 728 // If there is no branch, control always falls through. 729 if (TBB == 0) return true; 730 731 // If there is some explicit branch to the fallthrough block, it can obviously 732 // reach, even though the branch should get folded to fall through implicitly. 733 if (MachineFunction::iterator(TBB) == Fallthrough || 734 MachineFunction::iterator(FBB) == Fallthrough) 735 return true; 736 737 // If it's an unconditional branch to some block not the fall through, it 738 // doesn't fall through. 739 if (Cond.empty()) return false; 740 741 // Otherwise, if it is conditional and has no explicit false block, it falls 742 // through. 743 return FBB == 0; 744} 745 746/// CanFallThrough - Return true if the specified can implicitly transfer 747/// control to the block after it by falling off the end of it. This should 748/// return false if it can reach the block after it, but it uses an explicit 749/// branch to do so (e.g. a table jump). 750/// 751/// True is a conservative answer. 752/// 753bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { 754 MachineBasicBlock *TBB = 0, *FBB = 0; 755 std::vector<MachineOperand> Cond; 756 bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond); 757 return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); 758} 759 760/// IsBetterFallthrough - Return true if it would be clearly better to 761/// fall-through to MBB1 than to fall through into MBB2. This has to return 762/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 763/// result in infinite loops. 764static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 765 MachineBasicBlock *MBB2, 766 const TargetInstrInfo &TII) { 767 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 768 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 769 // optimize branches that branch to either a return block or an assert block 770 // into a fallthrough to the return. 771 if (MBB1->empty() || MBB2->empty()) return false; 772 773 MachineInstr *MBB1I = --MBB1->end(); 774 MachineInstr *MBB2I = --MBB2->end(); 775 return TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode()); 776} 777 778/// OptimizeBlock - Analyze and optimize control flow related to the specified 779/// block. This is never called on the entry block. 780void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 781 MachineFunction::iterator FallThrough = MBB; 782 ++FallThrough; 783 784 // If this block is empty, make everyone use its fall-through, not the block 785 // explicitly. Landing pads should not do this since the landing-pad table 786 // points to this block. 787 if (MBB->empty() && !MBB->isLandingPad()) { 788 // Dead block? Leave for cleanup later. 789 if (MBB->pred_empty()) return; 790 791 if (FallThrough == MBB->getParent()->end()) { 792 // TODO: Simplify preds to not branch here if possible! 793 } else { 794 // Rewrite all predecessors of the old block to go to the fallthrough 795 // instead. 796 while (!MBB->pred_empty()) { 797 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 798 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 799 } 800 801 // If MBB was the target of a jump table, update jump tables to go to the 802 // fallthrough instead. 803 MBB->getParent()->getJumpTableInfo()-> 804 ReplaceMBBInJumpTables(MBB, FallThrough); 805 MadeChange = true; 806 } 807 return; 808 } 809 810 // Check to see if we can simplify the terminator of the block before this 811 // one. 812 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 813 814 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 815 std::vector<MachineOperand> PriorCond; 816 bool PriorUnAnalyzable = 817 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 818 if (!PriorUnAnalyzable) { 819 // If the CFG for the prior block has extra edges, remove them. 820 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 821 !PriorCond.empty()); 822 823 // If the previous branch is conditional and both conditions go to the same 824 // destination, remove the branch, replacing it with an unconditional one or 825 // a fall-through. 826 if (PriorTBB && PriorTBB == PriorFBB) { 827 TII->RemoveBranch(PrevBB); 828 PriorCond.clear(); 829 if (PriorTBB != MBB) 830 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 831 MadeChange = true; 832 ++NumBranchOpts; 833 return OptimizeBlock(MBB); 834 } 835 836 // If the previous branch *only* branches to *this* block (conditional or 837 // not) remove the branch. 838 if (PriorTBB == MBB && PriorFBB == 0) { 839 TII->RemoveBranch(PrevBB); 840 MadeChange = true; 841 ++NumBranchOpts; 842 return OptimizeBlock(MBB); 843 } 844 845 // If the prior block branches somewhere else on the condition and here if 846 // the condition is false, remove the uncond second branch. 847 if (PriorFBB == MBB) { 848 TII->RemoveBranch(PrevBB); 849 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); 850 MadeChange = true; 851 ++NumBranchOpts; 852 return OptimizeBlock(MBB); 853 } 854 855 // If the prior block branches here on true and somewhere else on false, and 856 // if the branch condition is reversible, reverse the branch to create a 857 // fall-through. 858 if (PriorTBB == MBB) { 859 std::vector<MachineOperand> NewPriorCond(PriorCond); 860 if (!TII->ReverseBranchCondition(NewPriorCond)) { 861 TII->RemoveBranch(PrevBB); 862 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); 863 MadeChange = true; 864 ++NumBranchOpts; 865 return OptimizeBlock(MBB); 866 } 867 } 868 869 // If this block doesn't fall through (e.g. it ends with an uncond branch or 870 // has no successors) and if the pred falls through into this block, and if 871 // it would otherwise fall through into the block after this, move this 872 // block to the end of the function. 873 // 874 // We consider it more likely that execution will stay in the function (e.g. 875 // due to loops) than it is to exit it. This asserts in loops etc, moving 876 // the assert condition out of the loop body. 877 if (!PriorCond.empty() && PriorFBB == 0 && 878 MachineFunction::iterator(PriorTBB) == FallThrough && 879 !CanFallThrough(MBB)) { 880 bool DoTransform = true; 881 882 // We have to be careful that the succs of PredBB aren't both no-successor 883 // blocks. If neither have successors and if PredBB is the second from 884 // last block in the function, we'd just keep swapping the two blocks for 885 // last. Only do the swap if one is clearly better to fall through than 886 // the other. 887 if (FallThrough == --MBB->getParent()->end() && 888 !IsBetterFallthrough(PriorTBB, MBB, *TII)) 889 DoTransform = false; 890 891 // We don't want to do this transformation if we have control flow like: 892 // br cond BB2 893 // BB1: 894 // .. 895 // jmp BBX 896 // BB2: 897 // .. 898 // ret 899 // 900 // In this case, we could actually be moving the return block *into* a 901 // loop! 902 if (DoTransform && !MBB->succ_empty() && 903 (!CanFallThrough(PriorTBB) || PriorTBB->empty())) 904 DoTransform = false; 905 906 907 if (DoTransform) { 908 // Reverse the branch so we will fall through on the previous true cond. 909 std::vector<MachineOperand> NewPriorCond(PriorCond); 910 if (!TII->ReverseBranchCondition(NewPriorCond)) { 911 DOUT << "\nMoving MBB: " << *MBB; 912 DOUT << "To make fallthrough to: " << *PriorTBB << "\n"; 913 914 TII->RemoveBranch(PrevBB); 915 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); 916 917 // Move this block to the end of the function. 918 MBB->moveAfter(--MBB->getParent()->end()); 919 MadeChange = true; 920 ++NumBranchOpts; 921 return; 922 } 923 } 924 } 925 } 926 927 // Analyze the branch in the current block. 928 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 929 std::vector<MachineOperand> CurCond; 930 bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond); 931 if (!CurUnAnalyzable) { 932 // If the CFG for the prior block has extra edges, remove them. 933 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 934 935 // If this is a two-way branch, and the FBB branches to this block, reverse 936 // the condition so the single-basic-block loop is faster. Instead of: 937 // Loop: xxx; jcc Out; jmp Loop 938 // we want: 939 // Loop: xxx; jncc Loop; jmp Out 940 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 941 std::vector<MachineOperand> NewCond(CurCond); 942 if (!TII->ReverseBranchCondition(NewCond)) { 943 TII->RemoveBranch(*MBB); 944 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); 945 MadeChange = true; 946 ++NumBranchOpts; 947 return OptimizeBlock(MBB); 948 } 949 } 950 951 952 // If this branch is the only thing in its block, see if we can forward 953 // other blocks across it. 954 if (CurTBB && CurCond.empty() && CurFBB == 0 && 955 TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) { 956 // This block may contain just an unconditional branch. Because there can 957 // be 'non-branch terminators' in the block, try removing the branch and 958 // then seeing if the block is empty. 959 TII->RemoveBranch(*MBB); 960 961 // If this block is just an unconditional branch to CurTBB, we can 962 // usually completely eliminate the block. The only case we cannot 963 // completely eliminate the block is when the block before this one 964 // falls through into MBB and we can't understand the prior block's branch 965 // condition. 966 if (MBB->empty()) { 967 bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); 968 if (PredHasNoFallThrough || !PriorUnAnalyzable || 969 !PrevBB.isSuccessor(MBB)) { 970 // If the prior block falls through into us, turn it into an 971 // explicit branch to us to make updates simpler. 972 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 973 PriorTBB != MBB && PriorFBB != MBB) { 974 if (PriorTBB == 0) { 975 assert(PriorCond.empty() && PriorFBB == 0 && 976 "Bad branch analysis"); 977 PriorTBB = MBB; 978 } else { 979 assert(PriorFBB == 0 && "Machine CFG out of date!"); 980 PriorFBB = MBB; 981 } 982 TII->RemoveBranch(PrevBB); 983 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 984 } 985 986 // Iterate through all the predecessors, revectoring each in-turn. 987 size_t PI = 0; 988 bool DidChange = false; 989 bool HasBranchToSelf = false; 990 while(PI != MBB->pred_size()) { 991 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 992 if (PMBB == MBB) { 993 // If this block has an uncond branch to itself, leave it. 994 ++PI; 995 HasBranchToSelf = true; 996 } else { 997 DidChange = true; 998 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 999 } 1000 } 1001 1002 // Change any jumptables to go to the new MBB. 1003 MBB->getParent()->getJumpTableInfo()-> 1004 ReplaceMBBInJumpTables(MBB, CurTBB); 1005 if (DidChange) { 1006 ++NumBranchOpts; 1007 MadeChange = true; 1008 if (!HasBranchToSelf) return; 1009 } 1010 } 1011 } 1012 1013 // Add the branch back if the block is more than just an uncond branch. 1014 TII->InsertBranch(*MBB, CurTBB, 0, CurCond); 1015 } 1016 } 1017 1018 // If the prior block doesn't fall through into this block, and if this 1019 // block doesn't fall through into some other block, see if we can find a 1020 // place to move this block where a fall-through will happen. 1021 if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, 1022 PriorTBB, PriorFBB, PriorCond)) { 1023 // Now we know that there was no fall-through into this block, check to 1024 // see if it has a fall-through into its successor. 1025 bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, 1026 CurCond); 1027 1028 if (!MBB->isLandingPad()) { 1029 // Check all the predecessors of this block. If one of them has no fall 1030 // throughs, move this block right after it. 1031 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1032 E = MBB->pred_end(); PI != E; ++PI) { 1033 // Analyze the branch at the end of the pred. 1034 MachineBasicBlock *PredBB = *PI; 1035 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1036 if (PredBB != MBB && !CanFallThrough(PredBB) 1037 && (!CurFallsThru || !CurTBB || !CurFBB) 1038 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1039 // If the current block doesn't fall through, just move it. 1040 // If the current block can fall through and does not end with a 1041 // conditional branch, we need to append an unconditional jump to 1042 // the (current) next block. To avoid a possible compile-time 1043 // infinite loop, move blocks only backward in this case. 1044 // Also, if there are already 2 branches here, we cannot add a third; 1045 // this means we have the case 1046 // Bcc next 1047 // B elsewhere 1048 // next: 1049 if (CurFallsThru) { 1050 MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); 1051 CurCond.clear(); 1052 TII->InsertBranch(*MBB, NextBB, 0, CurCond); 1053 } 1054 MBB->moveAfter(PredBB); 1055 MadeChange = true; 1056 return OptimizeBlock(MBB); 1057 } 1058 } 1059 } 1060 1061 if (!CurFallsThru) { 1062 // Check all successors to see if we can move this block before it. 1063 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1064 E = MBB->succ_end(); SI != E; ++SI) { 1065 // Analyze the branch at the end of the block before the succ. 1066 MachineBasicBlock *SuccBB = *SI; 1067 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1068 std::vector<MachineOperand> SuccPrevCond; 1069 1070 // If this block doesn't already fall-through to that successor, and if 1071 // the succ doesn't already have a block that can fall through into it, 1072 // and if the successor isn't an EH destination, we can arrange for the 1073 // fallthrough to happen. 1074 if (SuccBB != MBB && !CanFallThrough(SuccPrev) && 1075 !SuccBB->isLandingPad()) { 1076 MBB->moveBefore(SuccBB); 1077 MadeChange = true; 1078 return OptimizeBlock(MBB); 1079 } 1080 } 1081 1082 // Okay, there is no really great place to put this block. If, however, 1083 // the block before this one would be a fall-through if this block were 1084 // removed, move this block to the end of the function. 1085 if (FallThrough != MBB->getParent()->end() && 1086 PrevBB.isSuccessor(FallThrough)) { 1087 MBB->moveAfter(--MBB->getParent()->end()); 1088 MadeChange = true; 1089 return; 1090 } 1091 } 1092 } 1093} 1094