BranchFolding.cpp revision 4d6ccb5f68cd7c6418a209f1fa4dbade569e4493
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"); 44STATISTIC(NumHoist , "Number of times common instructions are hoisted"); 45 46static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 47 cl::init(cl::BOU_UNSET), cl::Hidden); 48 49// Throttle for huge numbers of predecessors (compile speed problems) 50static cl::opt<unsigned> 51TailMergeThreshold("tail-merge-threshold", 52 cl::desc("Max number of predecessors to consider tail merging"), 53 cl::init(150), cl::Hidden); 54 55// Heuristic for tail merging (and, inversely, tail duplication). 56// TODO: This should be replaced with a target query. 57static cl::opt<unsigned> 58TailMergeSize("tail-merge-size", 59 cl::desc("Min number of instructions to consider tail merging"), 60 cl::init(3), cl::Hidden); 61 62namespace { 63 /// BranchFolderPass - Wrap branch folder in a machine function pass. 64 class BranchFolderPass : public MachineFunctionPass, 65 public BranchFolder { 66 public: 67 static char ID; 68 explicit BranchFolderPass(bool defaultEnableTailMerge) 69 : MachineFunctionPass(ID), BranchFolder(defaultEnableTailMerge, true) {} 70 71 virtual bool runOnMachineFunction(MachineFunction &MF); 72 virtual const char *getPassName() const { return "Control Flow Optimizer"; } 73 }; 74} 75 76char BranchFolderPass::ID = 0; 77 78FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { 79 return new BranchFolderPass(DefaultEnableTailMerge); 80} 81 82bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { 83 return OptimizeFunction(MF, 84 MF.getTarget().getInstrInfo(), 85 MF.getTarget().getRegisterInfo(), 86 getAnalysisIfAvailable<MachineModuleInfo>()); 87} 88 89 90BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist) { 91 switch (FlagEnableTailMerge) { 92 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 93 case cl::BOU_TRUE: EnableTailMerge = true; break; 94 case cl::BOU_FALSE: EnableTailMerge = false; break; 95 } 96 97 EnableHoistCommonCode = CommonHoist; 98} 99 100/// RemoveDeadBlock - Remove the specified dead machine basic block from the 101/// function, updating the CFG. 102void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 103 assert(MBB->pred_empty() && "MBB must be dead!"); 104 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB); 105 106 MachineFunction *MF = MBB->getParent(); 107 // drop all successors. 108 while (!MBB->succ_empty()) 109 MBB->removeSuccessor(MBB->succ_end()-1); 110 111 // Avoid matching if this pointer gets reused. 112 TriedMerging.erase(MBB); 113 114 // Remove the block. 115 MF->erase(MBB); 116} 117 118/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 119/// followed by terminators, and if the implicitly defined registers are not 120/// used by the terminators, remove those implicit_def's. e.g. 121/// BB1: 122/// r0 = implicit_def 123/// r1 = implicit_def 124/// br 125/// This block can be optimized away later if the implicit instructions are 126/// removed. 127bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 128 SmallSet<unsigned, 4> ImpDefRegs; 129 MachineBasicBlock::iterator I = MBB->begin(); 130 while (I != MBB->end()) { 131 if (!I->isImplicitDef()) 132 break; 133 unsigned Reg = I->getOperand(0).getReg(); 134 ImpDefRegs.insert(Reg); 135 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg); 136 unsigned SubReg = *SubRegs; ++SubRegs) 137 ImpDefRegs.insert(SubReg); 138 ++I; 139 } 140 if (ImpDefRegs.empty()) 141 return false; 142 143 MachineBasicBlock::iterator FirstTerm = I; 144 while (I != MBB->end()) { 145 if (!TII->isUnpredicatedTerminator(I)) 146 return false; 147 // See if it uses any of the implicitly defined registers. 148 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 149 MachineOperand &MO = I->getOperand(i); 150 if (!MO.isReg() || !MO.isUse()) 151 continue; 152 unsigned Reg = MO.getReg(); 153 if (ImpDefRegs.count(Reg)) 154 return false; 155 } 156 ++I; 157 } 158 159 I = MBB->begin(); 160 while (I != FirstTerm) { 161 MachineInstr *ImpDefMI = &*I; 162 ++I; 163 MBB->erase(ImpDefMI); 164 } 165 166 return true; 167} 168 169/// OptimizeFunction - Perhaps branch folding, tail merging and other 170/// CFG optimizations on the given function. 171bool BranchFolder::OptimizeFunction(MachineFunction &MF, 172 const TargetInstrInfo *tii, 173 const TargetRegisterInfo *tri, 174 MachineModuleInfo *mmi) { 175 if (!tii) return false; 176 177 TriedMerging.clear(); 178 179 TII = tii; 180 TRI = tri; 181 MMI = mmi; 182 183 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 184 185 // Fix CFG. The later algorithms expect it to be right. 186 bool MadeChange = false; 187 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 188 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 189 SmallVector<MachineOperand, 4> Cond; 190 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 191 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 192 MadeChange |= OptimizeImpDefsBlock(MBB); 193 } 194 195 bool MadeChangeThisIteration = true; 196 while (MadeChangeThisIteration) { 197 MadeChangeThisIteration = TailMergeBlocks(MF); 198 MadeChangeThisIteration |= OptimizeBranches(MF); 199 if (EnableHoistCommonCode) 200 MadeChangeThisIteration |= HoistCommonCode(MF); 201 MadeChange |= MadeChangeThisIteration; 202 } 203 204 // See if any jump tables have become dead as the code generator 205 // did its thing. 206 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 207 if (JTI == 0) { 208 delete RS; 209 return MadeChange; 210 } 211 212 // Walk the function to find jump tables that are live. 213 BitVector JTIsLive(JTI->getJumpTables().size()); 214 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 215 BB != E; ++BB) { 216 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 217 I != E; ++I) 218 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 219 MachineOperand &Op = I->getOperand(op); 220 if (!Op.isJTI()) continue; 221 222 // Remember that this JT is live. 223 JTIsLive.set(Op.getIndex()); 224 } 225 } 226 227 // Finally, remove dead jump tables. This happens when the 228 // indirect jump was unreachable (and thus deleted). 229 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 230 if (!JTIsLive.test(i)) { 231 JTI->RemoveJumpTable(i); 232 MadeChange = true; 233 } 234 235 delete RS; 236 return MadeChange; 237} 238 239//===----------------------------------------------------------------------===// 240// Tail Merging of Blocks 241//===----------------------------------------------------------------------===// 242 243/// HashMachineInstr - Compute a hash value for MI and its operands. 244static unsigned HashMachineInstr(const MachineInstr *MI) { 245 unsigned Hash = MI->getOpcode(); 246 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 247 const MachineOperand &Op = MI->getOperand(i); 248 249 // Merge in bits from the operand if easy. 250 unsigned OperandHash = 0; 251 switch (Op.getType()) { 252 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 253 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 254 case MachineOperand::MO_MachineBasicBlock: 255 OperandHash = Op.getMBB()->getNumber(); 256 break; 257 case MachineOperand::MO_FrameIndex: 258 case MachineOperand::MO_ConstantPoolIndex: 259 case MachineOperand::MO_JumpTableIndex: 260 OperandHash = Op.getIndex(); 261 break; 262 case MachineOperand::MO_GlobalAddress: 263 case MachineOperand::MO_ExternalSymbol: 264 // Global address / external symbol are too hard, don't bother, but do 265 // pull in the offset. 266 OperandHash = Op.getOffset(); 267 break; 268 default: break; 269 } 270 271 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 272 } 273 return Hash; 274} 275 276/// HashEndOfMBB - Hash the last instruction in the MBB. 277static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) { 278 MachineBasicBlock::const_iterator I = MBB->end(); 279 if (I == MBB->begin()) 280 return 0; // Empty MBB. 281 282 --I; 283 // Skip debug info so it will not affect codegen. 284 while (I->isDebugValue()) { 285 if (I==MBB->begin()) 286 return 0; // MBB empty except for debug info. 287 --I; 288 } 289 290 return HashMachineInstr(I); 291} 292 293/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 294/// of instructions they actually have in common together at their end. Return 295/// iterators for the first shared instruction in each block. 296static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 297 MachineBasicBlock *MBB2, 298 MachineBasicBlock::iterator &I1, 299 MachineBasicBlock::iterator &I2) { 300 I1 = MBB1->end(); 301 I2 = MBB2->end(); 302 303 unsigned TailLen = 0; 304 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 305 --I1; --I2; 306 // Skip debugging pseudos; necessary to avoid changing the code. 307 while (I1->isDebugValue()) { 308 if (I1==MBB1->begin()) { 309 while (I2->isDebugValue()) { 310 if (I2==MBB2->begin()) 311 // I1==DBG at begin; I2==DBG at begin 312 return TailLen; 313 --I2; 314 } 315 ++I2; 316 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin 317 return TailLen; 318 } 319 --I1; 320 } 321 // I1==first (untested) non-DBG preceding known match 322 while (I2->isDebugValue()) { 323 if (I2==MBB2->begin()) { 324 ++I1; 325 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin 326 return TailLen; 327 } 328 --I2; 329 } 330 // I1, I2==first (untested) non-DBGs preceding known match 331 if (!I1->isIdenticalTo(I2) || 332 // FIXME: This check is dubious. It's used to get around a problem where 333 // people incorrectly expect inline asm directives to remain in the same 334 // relative order. This is untenable because normal compiler 335 // optimizations (like this one) may reorder and/or merge these 336 // directives. 337 I1->isInlineAsm()) { 338 ++I1; ++I2; 339 break; 340 } 341 ++TailLen; 342 } 343 // Back past possible debugging pseudos at beginning of block. This matters 344 // when one block differs from the other only by whether debugging pseudos 345 // are present at the beginning. (This way, the various checks later for 346 // I1==MBB1->begin() work as expected.) 347 if (I1 == MBB1->begin() && I2 != MBB2->begin()) { 348 --I2; 349 while (I2->isDebugValue()) { 350 if (I2 == MBB2->begin()) { 351 return TailLen; 352 } 353 --I2; 354 } 355 ++I2; 356 } 357 if (I2 == MBB2->begin() && I1 != MBB1->begin()) { 358 --I1; 359 while (I1->isDebugValue()) { 360 if (I1 == MBB1->begin()) 361 return TailLen; 362 --I1; 363 } 364 ++I1; 365 } 366 return TailLen; 367} 368 369void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB, 370 MachineBasicBlock *NewMBB) { 371 if (RS) { 372 RS->enterBasicBlock(CurMBB); 373 if (!CurMBB->empty()) 374 RS->forward(prior(CurMBB->end())); 375 BitVector RegsLiveAtExit(TRI->getNumRegs()); 376 RS->getRegsUsed(RegsLiveAtExit, false); 377 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 378 if (RegsLiveAtExit[i]) 379 NewMBB->addLiveIn(i); 380 } 381} 382 383/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 384/// after it, replacing it with an unconditional branch to NewDest. 385void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 386 MachineBasicBlock *NewDest) { 387 MachineBasicBlock *CurMBB = OldInst->getParent(); 388 389 TII->ReplaceTailWithBranchTo(OldInst, NewDest); 390 391 // For targets that use the register scavenger, we must maintain LiveIns. 392 MaintainLiveIns(CurMBB, NewDest); 393 394 ++NumTailMerge; 395} 396 397/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 398/// MBB so that the part before the iterator falls into the part starting at the 399/// iterator. This returns the new MBB. 400MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 401 MachineBasicBlock::iterator BBI1) { 402 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1)) 403 return 0; 404 405 MachineFunction &MF = *CurMBB.getParent(); 406 407 // Create the fall-through block. 408 MachineFunction::iterator MBBI = &CurMBB; 409 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 410 CurMBB.getParent()->insert(++MBBI, NewMBB); 411 412 // Move all the successors of this block to the specified block. 413 NewMBB->transferSuccessors(&CurMBB); 414 415 // Add an edge from CurMBB to NewMBB for the fall-through. 416 CurMBB.addSuccessor(NewMBB); 417 418 // Splice the code over. 419 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 420 421 // For targets that use the register scavenger, we must maintain LiveIns. 422 MaintainLiveIns(&CurMBB, NewMBB); 423 424 return NewMBB; 425} 426 427/// EstimateRuntime - Make a rough estimate for how long it will take to run 428/// the specified code. 429static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 430 MachineBasicBlock::iterator E) { 431 unsigned Time = 0; 432 for (; I != E; ++I) { 433 if (I->isDebugValue()) 434 continue; 435 if (I->isCall()) 436 Time += 10; 437 else if (I->mayLoad() || I->mayStore()) 438 Time += 2; 439 else 440 ++Time; 441 } 442 return Time; 443} 444 445// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 446// branches temporarily for tail merging). In the case where CurMBB ends 447// with a conditional branch to the next block, optimize by reversing the 448// test and conditionally branching to SuccMBB instead. 449static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 450 const TargetInstrInfo *TII) { 451 MachineFunction *MF = CurMBB->getParent(); 452 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB)); 453 MachineBasicBlock *TBB = 0, *FBB = 0; 454 SmallVector<MachineOperand, 4> Cond; 455 DebugLoc dl; // FIXME: this is nowhere 456 if (I != MF->end() && 457 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 458 MachineBasicBlock *NextBB = I; 459 if (TBB == NextBB && !Cond.empty() && !FBB) { 460 if (!TII->ReverseBranchCondition(Cond)) { 461 TII->RemoveBranch(*CurMBB); 462 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl); 463 return; 464 } 465 } 466 } 467 TII->InsertBranch(*CurMBB, SuccBB, NULL, 468 SmallVector<MachineOperand, 0>(), dl); 469} 470 471bool 472BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 473 if (getHash() < o.getHash()) 474 return true; 475 else if (getHash() > o.getHash()) 476 return false; 477 else if (getBlock()->getNumber() < o.getBlock()->getNumber()) 478 return true; 479 else if (getBlock()->getNumber() > o.getBlock()->getNumber()) 480 return false; 481 else { 482 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 483 // an object with itself. 484#ifndef _GLIBCXX_DEBUG 485 llvm_unreachable("Predecessor appears twice"); 486#else 487 return false; 488#endif 489 } 490} 491 492/// CountTerminators - Count the number of terminators in the given 493/// block and set I to the position of the first non-terminator, if there 494/// is one, or MBB->end() otherwise. 495static unsigned CountTerminators(MachineBasicBlock *MBB, 496 MachineBasicBlock::iterator &I) { 497 I = MBB->end(); 498 unsigned NumTerms = 0; 499 for (;;) { 500 if (I == MBB->begin()) { 501 I = MBB->end(); 502 break; 503 } 504 --I; 505 if (!I->isTerminator()) break; 506 ++NumTerms; 507 } 508 return NumTerms; 509} 510 511/// ProfitableToMerge - Check if two machine basic blocks have a common tail 512/// and decide if it would be profitable to merge those tails. Return the 513/// length of the common tail and iterators to the first common instruction 514/// in each block. 515static bool ProfitableToMerge(MachineBasicBlock *MBB1, 516 MachineBasicBlock *MBB2, 517 unsigned minCommonTailLength, 518 unsigned &CommonTailLen, 519 MachineBasicBlock::iterator &I1, 520 MachineBasicBlock::iterator &I2, 521 MachineBasicBlock *SuccBB, 522 MachineBasicBlock *PredBB) { 523 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 524 if (CommonTailLen == 0) 525 return false; 526 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber() 527 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen 528 << '\n'); 529 530 // It's almost always profitable to merge any number of non-terminator 531 // instructions with the block that falls through into the common successor. 532 if (MBB1 == PredBB || MBB2 == PredBB) { 533 MachineBasicBlock::iterator I; 534 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 535 if (CommonTailLen > NumTerms) 536 return true; 537 } 538 539 // If one of the blocks can be completely merged and happens to be in 540 // a position where the other could fall through into it, merge any number 541 // of instructions, because it can be done without a branch. 542 // TODO: If the blocks are not adjacent, move one of them so that they are? 543 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 544 return true; 545 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 546 return true; 547 548 // If both blocks have an unconditional branch temporarily stripped out, 549 // count that as an additional common instruction for the following 550 // heuristics. 551 unsigned EffectiveTailLen = CommonTailLen; 552 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 553 !MBB1->back().isBarrier() && 554 !MBB2->back().isBarrier()) 555 ++EffectiveTailLen; 556 557 // Check if the common tail is long enough to be worthwhile. 558 if (EffectiveTailLen >= minCommonTailLength) 559 return true; 560 561 // If we are optimizing for code size, 2 instructions in common is enough if 562 // we don't have to split a block. At worst we will be introducing 1 new 563 // branch instruction, which is likely to be smaller than the 2 564 // instructions that would be deleted in the merge. 565 MachineFunction *MF = MBB1->getParent(); 566 if (EffectiveTailLen >= 2 && 567 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 568 (I1 == MBB1->begin() || I2 == MBB2->begin())) 569 return true; 570 571 return false; 572} 573 574/// ComputeSameTails - Look through all the blocks in MergePotentials that have 575/// hash CurHash (guaranteed to match the last element). Build the vector 576/// SameTails of all those that have the (same) largest number of instructions 577/// in common of any pair of these blocks. SameTails entries contain an 578/// iterator into MergePotentials (from which the MachineBasicBlock can be 579/// found) and a MachineBasicBlock::iterator into that MBB indicating the 580/// instruction where the matching code sequence begins. 581/// Order of elements in SameTails is the reverse of the order in which 582/// those blocks appear in MergePotentials (where they are not necessarily 583/// consecutive). 584unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 585 unsigned minCommonTailLength, 586 MachineBasicBlock *SuccBB, 587 MachineBasicBlock *PredBB) { 588 unsigned maxCommonTailLength = 0U; 589 SameTails.clear(); 590 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 591 MPIterator HighestMPIter = prior(MergePotentials.end()); 592 for (MPIterator CurMPIter = prior(MergePotentials.end()), 593 B = MergePotentials.begin(); 594 CurMPIter != B && CurMPIter->getHash() == CurHash; 595 --CurMPIter) { 596 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 597 unsigned CommonTailLen; 598 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 599 minCommonTailLength, 600 CommonTailLen, TrialBBI1, TrialBBI2, 601 SuccBB, PredBB)) { 602 if (CommonTailLen > maxCommonTailLength) { 603 SameTails.clear(); 604 maxCommonTailLength = CommonTailLen; 605 HighestMPIter = CurMPIter; 606 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 607 } 608 if (HighestMPIter == CurMPIter && 609 CommonTailLen == maxCommonTailLength) 610 SameTails.push_back(SameTailElt(I, TrialBBI2)); 611 } 612 if (I == B) 613 break; 614 } 615 } 616 return maxCommonTailLength; 617} 618 619/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 620/// MergePotentials, restoring branches at ends of blocks as appropriate. 621void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 622 MachineBasicBlock *SuccBB, 623 MachineBasicBlock *PredBB) { 624 MPIterator CurMPIter, B; 625 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 626 CurMPIter->getHash() == CurHash; 627 --CurMPIter) { 628 // Put the unconditional branch back, if we need one. 629 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 630 if (SuccBB && CurMBB != PredBB) 631 FixTail(CurMBB, SuccBB, TII); 632 if (CurMPIter == B) 633 break; 634 } 635 if (CurMPIter->getHash() != CurHash) 636 CurMPIter++; 637 MergePotentials.erase(CurMPIter, MergePotentials.end()); 638} 639 640/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 641/// only of the common tail. Create a block that does by splitting one. 642bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 643 unsigned maxCommonTailLength, 644 unsigned &commonTailIndex) { 645 commonTailIndex = 0; 646 unsigned TimeEstimate = ~0U; 647 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 648 // Use PredBB if possible; that doesn't require a new branch. 649 if (SameTails[i].getBlock() == PredBB) { 650 commonTailIndex = i; 651 break; 652 } 653 // Otherwise, make a (fairly bogus) choice based on estimate of 654 // how long it will take the various blocks to execute. 655 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 656 SameTails[i].getTailStartPos()); 657 if (t <= TimeEstimate) { 658 TimeEstimate = t; 659 commonTailIndex = i; 660 } 661 } 662 663 MachineBasicBlock::iterator BBI = 664 SameTails[commonTailIndex].getTailStartPos(); 665 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 666 667 // If the common tail includes any debug info we will take it pretty 668 // randomly from one of the inputs. Might be better to remove it? 669 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 670 << maxCommonTailLength); 671 672 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 673 if (!newMBB) { 674 DEBUG(dbgs() << "... failed!"); 675 return false; 676 } 677 678 SameTails[commonTailIndex].setBlock(newMBB); 679 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 680 681 // If we split PredBB, newMBB is the new predecessor. 682 if (PredBB == MBB) 683 PredBB = newMBB; 684 685 return true; 686} 687 688// See if any of the blocks in MergePotentials (which all have a common single 689// successor, or all have no successor) can be tail-merged. If there is a 690// successor, any blocks in MergePotentials that are not tail-merged and 691// are not immediately before Succ must have an unconditional branch to 692// Succ added (but the predecessor/successor lists need no adjustment). 693// The lone predecessor of Succ that falls through into Succ, 694// if any, is given in PredBB. 695 696bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 697 MachineBasicBlock *PredBB) { 698 bool MadeChange = false; 699 700 // Except for the special cases below, tail-merge if there are at least 701 // this many instructions in common. 702 unsigned minCommonTailLength = TailMergeSize; 703 704 DEBUG(dbgs() << "\nTryTailMergeBlocks: "; 705 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 706 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 707 << (i == e-1 ? "" : ", "); 708 dbgs() << "\n"; 709 if (SuccBB) { 710 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 711 if (PredBB) 712 dbgs() << " which has fall-through from BB#" 713 << PredBB->getNumber() << "\n"; 714 } 715 dbgs() << "Looking for common tails of at least " 716 << minCommonTailLength << " instruction" 717 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 718 ); 719 720 // Sort by hash value so that blocks with identical end sequences sort 721 // together. 722 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 723 724 // Walk through equivalence sets looking for actual exact matches. 725 while (MergePotentials.size() > 1) { 726 unsigned CurHash = MergePotentials.back().getHash(); 727 728 // Build SameTails, identifying the set of blocks with this hash code 729 // and with the maximum number of instructions in common. 730 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 731 minCommonTailLength, 732 SuccBB, PredBB); 733 734 // If we didn't find any pair that has at least minCommonTailLength 735 // instructions in common, remove all blocks with this hash code and retry. 736 if (SameTails.empty()) { 737 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 738 continue; 739 } 740 741 // If one of the blocks is the entire common tail (and not the entry 742 // block, which we can't jump to), we can treat all blocks with this same 743 // tail at once. Use PredBB if that is one of the possibilities, as that 744 // will not introduce any extra branches. 745 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 746 getParent()->begin(); 747 unsigned commonTailIndex = SameTails.size(); 748 // If there are two blocks, check to see if one can be made to fall through 749 // into the other. 750 if (SameTails.size() == 2 && 751 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 752 SameTails[1].tailIsWholeBlock()) 753 commonTailIndex = 1; 754 else if (SameTails.size() == 2 && 755 SameTails[1].getBlock()->isLayoutSuccessor( 756 SameTails[0].getBlock()) && 757 SameTails[0].tailIsWholeBlock()) 758 commonTailIndex = 0; 759 else { 760 // Otherwise just pick one, favoring the fall-through predecessor if 761 // there is one. 762 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 763 MachineBasicBlock *MBB = SameTails[i].getBlock(); 764 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 765 continue; 766 if (MBB == PredBB) { 767 commonTailIndex = i; 768 break; 769 } 770 if (SameTails[i].tailIsWholeBlock()) 771 commonTailIndex = i; 772 } 773 } 774 775 if (commonTailIndex == SameTails.size() || 776 (SameTails[commonTailIndex].getBlock() == PredBB && 777 !SameTails[commonTailIndex].tailIsWholeBlock())) { 778 // None of the blocks consist entirely of the common tail. 779 // Split a block so that one does. 780 if (!CreateCommonTailOnlyBlock(PredBB, 781 maxCommonTailLength, commonTailIndex)) { 782 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 783 continue; 784 } 785 } 786 787 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 788 // MBB is common tail. Adjust all other BB's to jump to this one. 789 // Traversal must be forwards so erases work. 790 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber() 791 << " for "); 792 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 793 if (commonTailIndex == i) 794 continue; 795 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() 796 << (i == e-1 ? "" : ", ")); 797 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 798 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 799 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 800 MergePotentials.erase(SameTails[i].getMPIter()); 801 } 802 DEBUG(dbgs() << "\n"); 803 // We leave commonTailIndex in the worklist in case there are other blocks 804 // that match it with a smaller number of instructions. 805 MadeChange = true; 806 } 807 return MadeChange; 808} 809 810bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 811 812 if (!EnableTailMerge) return false; 813 814 bool MadeChange = false; 815 816 // First find blocks with no successors. 817 MergePotentials.clear(); 818 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); 819 I != E && MergePotentials.size() < TailMergeThreshold; ++I) { 820 if (TriedMerging.count(I)) 821 continue; 822 if (I->succ_empty()) 823 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I)); 824 } 825 826 // If this is a large problem, avoid visiting the same basic blocks 827 // multiple times. 828 if (MergePotentials.size() == TailMergeThreshold) 829 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 830 TriedMerging.insert(MergePotentials[i].getBlock()); 831 // See if we can do any tail merging on those. 832 if (MergePotentials.size() >= 2) 833 MadeChange |= TryTailMergeBlocks(NULL, NULL); 834 835 // Look at blocks (IBB) with multiple predecessors (PBB). 836 // We change each predecessor to a canonical form, by 837 // (1) temporarily removing any unconditional branch from the predecessor 838 // to IBB, and 839 // (2) alter conditional branches so they branch to the other block 840 // not IBB; this may require adding back an unconditional branch to IBB 841 // later, where there wasn't one coming in. E.g. 842 // Bcc IBB 843 // fallthrough to QBB 844 // here becomes 845 // Bncc QBB 846 // with a conceptual B to IBB after that, which never actually exists. 847 // With those changes, we see whether the predecessors' tails match, 848 // and merge them if so. We change things out of canonical form and 849 // back to the way they were later in the process. (OptimizeBranches 850 // would undo some of this, but we can't use it, because we'd get into 851 // a compile-time infinite loop repeatedly doing and undoing the same 852 // transformations.) 853 854 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 855 I != E; ++I) { 856 if (I->pred_size() >= 2) { 857 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 858 MachineBasicBlock *IBB = I; 859 MachineBasicBlock *PredBB = prior(I); 860 MergePotentials.clear(); 861 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 862 E2 = I->pred_end(); 863 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) { 864 MachineBasicBlock *PBB = *P; 865 if (TriedMerging.count(PBB)) 866 continue; 867 // Skip blocks that loop to themselves, can't tail merge these. 868 if (PBB == IBB) 869 continue; 870 // Visit each predecessor only once. 871 if (!UniquePreds.insert(PBB)) 872 continue; 873 // Skip blocks which may jump to a landing pad. Can't tail merge these. 874 if (PBB->getLandingPadSuccessor()) 875 continue; 876 MachineBasicBlock *TBB = 0, *FBB = 0; 877 SmallVector<MachineOperand, 4> Cond; 878 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 879 // Failing case: IBB is the target of a cbr, and 880 // we cannot reverse the branch. 881 SmallVector<MachineOperand, 4> NewCond(Cond); 882 if (!Cond.empty() && TBB == IBB) { 883 if (TII->ReverseBranchCondition(NewCond)) 884 continue; 885 // This is the QBB case described above 886 if (!FBB) 887 FBB = llvm::next(MachineFunction::iterator(PBB)); 888 } 889 // Failing case: the only way IBB can be reached from PBB is via 890 // exception handling. Happens for landing pads. Would be nice 891 // to have a bit in the edge so we didn't have to do all this. 892 if (IBB->isLandingPad()) { 893 MachineFunction::iterator IP = PBB; IP++; 894 MachineBasicBlock *PredNextBB = NULL; 895 if (IP != MF.end()) 896 PredNextBB = IP; 897 if (TBB == NULL) { 898 if (IBB != PredNextBB) // fallthrough 899 continue; 900 } else if (FBB) { 901 if (TBB != IBB && FBB != IBB) // cbr then ubr 902 continue; 903 } else if (Cond.empty()) { 904 if (TBB != IBB) // ubr 905 continue; 906 } else { 907 if (TBB != IBB && IBB != PredNextBB) // cbr 908 continue; 909 } 910 } 911 // Remove the unconditional branch at the end, if any. 912 if (TBB && (Cond.empty() || FBB)) { 913 DebugLoc dl; // FIXME: this is nowhere 914 TII->RemoveBranch(*PBB); 915 if (!Cond.empty()) 916 // reinsert conditional branch only, for now 917 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl); 918 } 919 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P)); 920 } 921 } 922 // If this is a large problem, avoid visiting the same basic blocks 923 // multiple times. 924 if (MergePotentials.size() == TailMergeThreshold) 925 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 926 TriedMerging.insert(MergePotentials[i].getBlock()); 927 if (MergePotentials.size() >= 2) 928 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 929 // Reinsert an unconditional branch if needed. 930 // The 1 below can occur as a result of removing blocks in 931 // TryTailMergeBlocks. 932 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 933 if (MergePotentials.size() == 1 && 934 MergePotentials.begin()->getBlock() != PredBB) 935 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 936 } 937 } 938 return MadeChange; 939} 940 941//===----------------------------------------------------------------------===// 942// Branch Optimization 943//===----------------------------------------------------------------------===// 944 945bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 946 bool MadeChange = false; 947 948 // Make sure blocks are numbered in order 949 MF.RenumberBlocks(); 950 951 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 952 I != E; ) { 953 MachineBasicBlock *MBB = I++; 954 MadeChange |= OptimizeBlock(MBB); 955 956 // If it is dead, remove it. 957 if (MBB->pred_empty()) { 958 RemoveDeadBlock(MBB); 959 MadeChange = true; 960 ++NumDeadBlocks; 961 } 962 } 963 return MadeChange; 964} 965 966// Blocks should be considered empty if they contain only debug info; 967// else the debug info would affect codegen. 968static bool IsEmptyBlock(MachineBasicBlock *MBB) { 969 if (MBB->empty()) 970 return true; 971 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end(); 972 MBBI!=MBBE; ++MBBI) { 973 if (!MBBI->isDebugValue()) 974 return false; 975 } 976 return true; 977} 978 979// Blocks with only debug info and branches should be considered the same 980// as blocks with only branches. 981static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { 982 MachineBasicBlock::iterator MBBI, MBBE; 983 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) { 984 if (!MBBI->isDebugValue()) 985 break; 986 } 987 return (MBBI->isBranch()); 988} 989 990/// IsBetterFallthrough - Return true if it would be clearly better to 991/// fall-through to MBB1 than to fall through into MBB2. This has to return 992/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 993/// result in infinite loops. 994static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 995 MachineBasicBlock *MBB2) { 996 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 997 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 998 // optimize branches that branch to either a return block or an assert block 999 // into a fallthrough to the return. 1000 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false; 1001 1002 // If there is a clear successor ordering we make sure that one block 1003 // will fall through to the next 1004 if (MBB1->isSuccessor(MBB2)) return true; 1005 if (MBB2->isSuccessor(MBB1)) return false; 1006 1007 // Neither block consists entirely of debug info (per IsEmptyBlock check), 1008 // so we needn't test for falling off the beginning here. 1009 MachineBasicBlock::iterator MBB1I = --MBB1->end(); 1010 while (MBB1I->isDebugValue()) 1011 --MBB1I; 1012 MachineBasicBlock::iterator MBB2I = --MBB2->end(); 1013 while (MBB2I->isDebugValue()) 1014 --MBB2I; 1015 return MBB2I->isCall() && !MBB1I->isCall(); 1016} 1017 1018/// OptimizeBlock - Analyze and optimize control flow related to the specified 1019/// block. This is never called on the entry block. 1020bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1021 bool MadeChange = false; 1022 MachineFunction &MF = *MBB->getParent(); 1023 DebugLoc dl; // FIXME: this is nowhere 1024ReoptimizeBlock: 1025 1026 MachineFunction::iterator FallThrough = MBB; 1027 ++FallThrough; 1028 1029 // If this block is empty, make everyone use its fall-through, not the block 1030 // explicitly. Landing pads should not do this since the landing-pad table 1031 // points to this block. Blocks with their addresses taken shouldn't be 1032 // optimized away. 1033 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 1034 // Dead block? Leave for cleanup later. 1035 if (MBB->pred_empty()) return MadeChange; 1036 1037 if (FallThrough == MF.end()) { 1038 // TODO: Simplify preds to not branch here if possible! 1039 } else { 1040 // Rewrite all predecessors of the old block to go to the fallthrough 1041 // instead. 1042 while (!MBB->pred_empty()) { 1043 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1044 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1045 } 1046 // If MBB was the target of a jump table, update jump tables to go to the 1047 // fallthrough instead. 1048 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1049 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough); 1050 MadeChange = true; 1051 } 1052 return MadeChange; 1053 } 1054 1055 // Check to see if we can simplify the terminator of the block before this 1056 // one. 1057 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1058 1059 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1060 SmallVector<MachineOperand, 4> PriorCond; 1061 bool PriorUnAnalyzable = 1062 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1063 if (!PriorUnAnalyzable) { 1064 // If the CFG for the prior block has extra edges, remove them. 1065 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1066 !PriorCond.empty()); 1067 1068 // If the previous branch is conditional and both conditions go to the same 1069 // destination, remove the branch, replacing it with an unconditional one or 1070 // a fall-through. 1071 if (PriorTBB && PriorTBB == PriorFBB) { 1072 TII->RemoveBranch(PrevBB); 1073 PriorCond.clear(); 1074 if (PriorTBB != MBB) 1075 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1076 MadeChange = true; 1077 ++NumBranchOpts; 1078 goto ReoptimizeBlock; 1079 } 1080 1081 // If the previous block unconditionally falls through to this block and 1082 // this block has no other predecessors, move the contents of this block 1083 // into the prior block. This doesn't usually happen when SimplifyCFG 1084 // has been used, but it can happen if tail merging splits a fall-through 1085 // predecessor of a block. 1086 // This has to check PrevBB->succ_size() because EH edges are ignored by 1087 // AnalyzeBranch. 1088 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1089 PrevBB.succ_size() == 1 && 1090 !MBB->hasAddressTaken() && !MBB->isLandingPad()) { 1091 DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1092 << "From MBB: " << *MBB); 1093 // Remove redundant DBG_VALUEs first. 1094 if (PrevBB.begin() != PrevBB.end()) { 1095 MachineBasicBlock::iterator PrevBBIter = PrevBB.end(); 1096 --PrevBBIter; 1097 MachineBasicBlock::iterator MBBIter = MBB->begin(); 1098 // Check if DBG_VALUE at the end of PrevBB is identical to the 1099 // DBG_VALUE at the beginning of MBB. 1100 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end() 1101 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) { 1102 if (!MBBIter->isIdenticalTo(PrevBBIter)) 1103 break; 1104 MachineInstr *DuplicateDbg = MBBIter; 1105 ++MBBIter; -- PrevBBIter; 1106 DuplicateDbg->eraseFromParent(); 1107 } 1108 } 1109 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1110 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1111 assert(PrevBB.succ_empty()); 1112 PrevBB.transferSuccessors(MBB); 1113 MadeChange = true; 1114 return MadeChange; 1115 } 1116 1117 // If the previous branch *only* branches to *this* block (conditional or 1118 // not) remove the branch. 1119 if (PriorTBB == MBB && PriorFBB == 0) { 1120 TII->RemoveBranch(PrevBB); 1121 MadeChange = true; 1122 ++NumBranchOpts; 1123 goto ReoptimizeBlock; 1124 } 1125 1126 // If the prior block branches somewhere else on the condition and here if 1127 // the condition is false, remove the uncond second branch. 1128 if (PriorFBB == MBB) { 1129 TII->RemoveBranch(PrevBB); 1130 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1131 MadeChange = true; 1132 ++NumBranchOpts; 1133 goto ReoptimizeBlock; 1134 } 1135 1136 // If the prior block branches here on true and somewhere else on false, and 1137 // if the branch condition is reversible, reverse the branch to create a 1138 // fall-through. 1139 if (PriorTBB == MBB) { 1140 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1141 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1142 TII->RemoveBranch(PrevBB); 1143 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl); 1144 MadeChange = true; 1145 ++NumBranchOpts; 1146 goto ReoptimizeBlock; 1147 } 1148 } 1149 1150 // If this block has no successors (e.g. it is a return block or ends with 1151 // a call to a no-return function like abort or __cxa_throw) and if the pred 1152 // falls through into this block, and if it would otherwise fall through 1153 // into the block after this, move this block to the end of the function. 1154 // 1155 // We consider it more likely that execution will stay in the function (e.g. 1156 // due to loops) than it is to exit it. This asserts in loops etc, moving 1157 // the assert condition out of the loop body. 1158 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1159 MachineFunction::iterator(PriorTBB) == FallThrough && 1160 !MBB->canFallThrough()) { 1161 bool DoTransform = true; 1162 1163 // We have to be careful that the succs of PredBB aren't both no-successor 1164 // blocks. If neither have successors and if PredBB is the second from 1165 // last block in the function, we'd just keep swapping the two blocks for 1166 // last. Only do the swap if one is clearly better to fall through than 1167 // the other. 1168 if (FallThrough == --MF.end() && 1169 !IsBetterFallthrough(PriorTBB, MBB)) 1170 DoTransform = false; 1171 1172 if (DoTransform) { 1173 // Reverse the branch so we will fall through on the previous true cond. 1174 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1175 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1176 DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1177 << "To make fallthrough to: " << *PriorTBB << "\n"); 1178 1179 TII->RemoveBranch(PrevBB); 1180 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl); 1181 1182 // Move this block to the end of the function. 1183 MBB->moveAfter(--MF.end()); 1184 MadeChange = true; 1185 ++NumBranchOpts; 1186 return MadeChange; 1187 } 1188 } 1189 } 1190 } 1191 1192 // Analyze the branch in the current block. 1193 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1194 SmallVector<MachineOperand, 4> CurCond; 1195 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1196 if (!CurUnAnalyzable) { 1197 // If the CFG for the prior block has extra edges, remove them. 1198 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1199 1200 // If this is a two-way branch, and the FBB branches to this block, reverse 1201 // the condition so the single-basic-block loop is faster. Instead of: 1202 // Loop: xxx; jcc Out; jmp Loop 1203 // we want: 1204 // Loop: xxx; jncc Loop; jmp Out 1205 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1206 SmallVector<MachineOperand, 4> NewCond(CurCond); 1207 if (!TII->ReverseBranchCondition(NewCond)) { 1208 TII->RemoveBranch(*MBB); 1209 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); 1210 MadeChange = true; 1211 ++NumBranchOpts; 1212 goto ReoptimizeBlock; 1213 } 1214 } 1215 1216 // If this branch is the only thing in its block, see if we can forward 1217 // other blocks across it. 1218 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1219 IsBranchOnlyBlock(MBB) && CurTBB != MBB && 1220 !MBB->hasAddressTaken()) { 1221 // This block may contain just an unconditional branch. Because there can 1222 // be 'non-branch terminators' in the block, try removing the branch and 1223 // then seeing if the block is empty. 1224 TII->RemoveBranch(*MBB); 1225 // If the only things remaining in the block are debug info, remove these 1226 // as well, so this will behave the same as an empty block in non-debug 1227 // mode. 1228 if (!MBB->empty()) { 1229 bool NonDebugInfoFound = false; 1230 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 1231 I != E; ++I) { 1232 if (!I->isDebugValue()) { 1233 NonDebugInfoFound = true; 1234 break; 1235 } 1236 } 1237 if (!NonDebugInfoFound) 1238 // Make the block empty, losing the debug info (we could probably 1239 // improve this in some cases.) 1240 MBB->erase(MBB->begin(), MBB->end()); 1241 } 1242 // If this block is just an unconditional branch to CurTBB, we can 1243 // usually completely eliminate the block. The only case we cannot 1244 // completely eliminate the block is when the block before this one 1245 // falls through into MBB and we can't understand the prior block's branch 1246 // condition. 1247 if (MBB->empty()) { 1248 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1249 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1250 !PrevBB.isSuccessor(MBB)) { 1251 // If the prior block falls through into us, turn it into an 1252 // explicit branch to us to make updates simpler. 1253 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1254 PriorTBB != MBB && PriorFBB != MBB) { 1255 if (PriorTBB == 0) { 1256 assert(PriorCond.empty() && PriorFBB == 0 && 1257 "Bad branch analysis"); 1258 PriorTBB = MBB; 1259 } else { 1260 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1261 PriorFBB = MBB; 1262 } 1263 TII->RemoveBranch(PrevBB); 1264 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, dl); 1265 } 1266 1267 // Iterate through all the predecessors, revectoring each in-turn. 1268 size_t PI = 0; 1269 bool DidChange = false; 1270 bool HasBranchToSelf = false; 1271 while(PI != MBB->pred_size()) { 1272 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1273 if (PMBB == MBB) { 1274 // If this block has an uncond branch to itself, leave it. 1275 ++PI; 1276 HasBranchToSelf = true; 1277 } else { 1278 DidChange = true; 1279 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1280 // If this change resulted in PMBB ending in a conditional 1281 // branch where both conditions go to the same destination, 1282 // change this to an unconditional branch (and fix the CFG). 1283 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1284 SmallVector<MachineOperand, 4> NewCurCond; 1285 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1286 NewCurFBB, NewCurCond, true); 1287 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1288 TII->RemoveBranch(*PMBB); 1289 NewCurCond.clear(); 1290 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, dl); 1291 MadeChange = true; 1292 ++NumBranchOpts; 1293 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1294 } 1295 } 1296 } 1297 1298 // Change any jumptables to go to the new MBB. 1299 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1300 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); 1301 if (DidChange) { 1302 ++NumBranchOpts; 1303 MadeChange = true; 1304 if (!HasBranchToSelf) return MadeChange; 1305 } 1306 } 1307 } 1308 1309 // Add the branch back if the block is more than just an uncond branch. 1310 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl); 1311 } 1312 } 1313 1314 // If the prior block doesn't fall through into this block, and if this 1315 // block doesn't fall through into some other block, see if we can find a 1316 // place to move this block where a fall-through will happen. 1317 if (!PrevBB.canFallThrough()) { 1318 1319 // Now we know that there was no fall-through into this block, check to 1320 // see if it has a fall-through into its successor. 1321 bool CurFallsThru = MBB->canFallThrough(); 1322 1323 if (!MBB->isLandingPad()) { 1324 // Check all the predecessors of this block. If one of them has no fall 1325 // throughs, move this block right after it. 1326 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1327 E = MBB->pred_end(); PI != E; ++PI) { 1328 // Analyze the branch at the end of the pred. 1329 MachineBasicBlock *PredBB = *PI; 1330 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1331 MachineBasicBlock *PredTBB = 0, *PredFBB = 0; 1332 SmallVector<MachineOperand, 4> PredCond; 1333 if (PredBB != MBB && !PredBB->canFallThrough() && 1334 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1335 && (!CurFallsThru || !CurTBB || !CurFBB) 1336 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1337 // If the current block doesn't fall through, just move it. 1338 // If the current block can fall through and does not end with a 1339 // conditional branch, we need to append an unconditional jump to 1340 // the (current) next block. To avoid a possible compile-time 1341 // infinite loop, move blocks only backward in this case. 1342 // Also, if there are already 2 branches here, we cannot add a third; 1343 // this means we have the case 1344 // Bcc next 1345 // B elsewhere 1346 // next: 1347 if (CurFallsThru) { 1348 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); 1349 CurCond.clear(); 1350 TII->InsertBranch(*MBB, NextBB, 0, CurCond, dl); 1351 } 1352 MBB->moveAfter(PredBB); 1353 MadeChange = true; 1354 goto ReoptimizeBlock; 1355 } 1356 } 1357 } 1358 1359 if (!CurFallsThru) { 1360 // Check all successors to see if we can move this block before it. 1361 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1362 E = MBB->succ_end(); SI != E; ++SI) { 1363 // Analyze the branch at the end of the block before the succ. 1364 MachineBasicBlock *SuccBB = *SI; 1365 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1366 1367 // If this block doesn't already fall-through to that successor, and if 1368 // the succ doesn't already have a block that can fall through into it, 1369 // and if the successor isn't an EH destination, we can arrange for the 1370 // fallthrough to happen. 1371 if (SuccBB != MBB && &*SuccPrev != MBB && 1372 !SuccPrev->canFallThrough() && !CurUnAnalyzable && 1373 !SuccBB->isLandingPad()) { 1374 MBB->moveBefore(SuccBB); 1375 MadeChange = true; 1376 goto ReoptimizeBlock; 1377 } 1378 } 1379 1380 // Okay, there is no really great place to put this block. If, however, 1381 // the block before this one would be a fall-through if this block were 1382 // removed, move this block to the end of the function. 1383 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0; 1384 SmallVector<MachineOperand, 4> PrevCond; 1385 if (FallThrough != MF.end() && 1386 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1387 PrevBB.isSuccessor(FallThrough)) { 1388 MBB->moveAfter(--MF.end()); 1389 MadeChange = true; 1390 return MadeChange; 1391 } 1392 } 1393 } 1394 1395 return MadeChange; 1396} 1397 1398//===----------------------------------------------------------------------===// 1399// Hoist Common Code 1400//===----------------------------------------------------------------------===// 1401 1402/// HoistCommonCode - Hoist common instruction sequences at the start of basic 1403/// blocks to their common predecessor. 1404bool BranchFolder::HoistCommonCode(MachineFunction &MF) { 1405 bool MadeChange = false; 1406 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) { 1407 MachineBasicBlock *MBB = I++; 1408 MadeChange |= HoistCommonCodeInSuccs(MBB); 1409 } 1410 1411 return MadeChange; 1412} 1413 1414/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 1415/// its 'true' successor. 1416static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 1417 MachineBasicBlock *TrueBB) { 1418 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 1419 E = BB->succ_end(); SI != E; ++SI) { 1420 MachineBasicBlock *SuccBB = *SI; 1421 if (SuccBB != TrueBB) 1422 return SuccBB; 1423 } 1424 return NULL; 1425} 1426 1427/// findHoistingInsertPosAndDeps - Find the location to move common instructions 1428/// in successors to. The location is ususally just before the terminator, 1429/// however if the terminator is a conditional branch and its previous 1430/// instruction is the flag setting instruction, the previous instruction is 1431/// the preferred location. This function also gathers uses and defs of the 1432/// instructions from the insertion point to the end of the block. The data is 1433/// used by HoistCommonCodeInSuccs to ensure safety. 1434static 1435MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB, 1436 const TargetInstrInfo *TII, 1437 const TargetRegisterInfo *TRI, 1438 SmallSet<unsigned,4> &Uses, 1439 SmallSet<unsigned,4> &Defs) { 1440 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator(); 1441 if (!TII->isUnpredicatedTerminator(Loc)) 1442 return MBB->end(); 1443 1444 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) { 1445 const MachineOperand &MO = Loc->getOperand(i); 1446 if (!MO.isReg()) 1447 continue; 1448 unsigned Reg = MO.getReg(); 1449 if (!Reg) 1450 continue; 1451 if (MO.isUse()) { 1452 Uses.insert(Reg); 1453 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1454 Uses.insert(*AS); 1455 } else if (!MO.isDead()) 1456 // Don't try to hoist code in the rare case the terminator defines a 1457 // register that is later used. 1458 return MBB->end(); 1459 } 1460 1461 if (Uses.empty()) 1462 return Loc; 1463 if (Loc == MBB->begin()) 1464 return MBB->end(); 1465 1466 // The terminator is probably a conditional branch, try not to separate the 1467 // branch from condition setting instruction. 1468 MachineBasicBlock::iterator PI = Loc; 1469 --PI; 1470 while (PI != MBB->begin() && Loc->isDebugValue()) 1471 --PI; 1472 1473 bool IsDef = false; 1474 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) { 1475 const MachineOperand &MO = PI->getOperand(i); 1476 if (!MO.isReg() || MO.isUse()) 1477 continue; 1478 unsigned Reg = MO.getReg(); 1479 if (!Reg) 1480 continue; 1481 if (Uses.count(Reg)) 1482 IsDef = true; 1483 } 1484 if (!IsDef) 1485 // The condition setting instruction is not just before the conditional 1486 // branch. 1487 return Loc; 1488 1489 // Be conservative, don't insert instruction above something that may have 1490 // side-effects. And since it's potentially bad to separate flag setting 1491 // instruction from the conditional branch, just abort the optimization 1492 // completely. 1493 // Also avoid moving code above predicated instruction since it's hard to 1494 // reason about register liveness with predicated instruction. 1495 bool DontMoveAcrossStore = true; 1496 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) || 1497 TII->isPredicated(PI)) 1498 return MBB->end(); 1499 1500 1501 // Find out what registers are live. Note this routine is ignoring other live 1502 // registers which are only used by instructions in successor blocks. 1503 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) { 1504 const MachineOperand &MO = PI->getOperand(i); 1505 if (!MO.isReg()) 1506 continue; 1507 unsigned Reg = MO.getReg(); 1508 if (!Reg) 1509 continue; 1510 if (MO.isUse()) { 1511 Uses.insert(Reg); 1512 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1513 Uses.insert(*AS); 1514 } else { 1515 if (Uses.count(Reg)) { 1516 Uses.erase(Reg); 1517 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1518 Uses.erase(*SR); // Use getSubRegisters to be conservative 1519 } 1520 Defs.insert(Reg); 1521 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1522 Defs.insert(*AS); 1523 } 1524 } 1525 1526 return PI; 1527} 1528 1529/// HoistCommonCodeInSuccs - If the successors of MBB has common instruction 1530/// sequence at the start of the function, move the instructions before MBB 1531/// terminator if it's legal. 1532bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) { 1533 MachineBasicBlock *TBB = 0, *FBB = 0; 1534 SmallVector<MachineOperand, 4> Cond; 1535 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty()) 1536 return false; 1537 1538 if (!FBB) FBB = findFalseBlock(MBB, TBB); 1539 if (!FBB) 1540 // Malformed bcc? True and false blocks are the same? 1541 return false; 1542 1543 // Restrict the optimization to cases where MBB is the only predecessor, 1544 // it is an obvious win. 1545 if (TBB->pred_size() > 1 || FBB->pred_size() > 1) 1546 return false; 1547 1548 // Find a suitable position to hoist the common instructions to. Also figure 1549 // out which registers are used or defined by instructions from the insertion 1550 // point to the end of the block. 1551 SmallSet<unsigned, 4> Uses, Defs; 1552 MachineBasicBlock::iterator Loc = 1553 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs); 1554 if (Loc == MBB->end()) 1555 return false; 1556 1557 bool HasDups = false; 1558 SmallVector<unsigned, 4> LocalDefs; 1559 SmallSet<unsigned, 4> LocalDefsSet; 1560 MachineBasicBlock::iterator TIB = TBB->begin(); 1561 MachineBasicBlock::iterator FIB = FBB->begin(); 1562 MachineBasicBlock::iterator TIE = TBB->end(); 1563 MachineBasicBlock::iterator FIE = FBB->end(); 1564 while (TIB != TIE && FIB != FIE) { 1565 // Skip dbg_value instructions. These do not count. 1566 if (TIB->isDebugValue()) { 1567 while (TIB != TIE && TIB->isDebugValue()) 1568 ++TIB; 1569 if (TIB == TIE) 1570 break; 1571 } 1572 if (FIB->isDebugValue()) { 1573 while (FIB != FIE && FIB->isDebugValue()) 1574 ++FIB; 1575 if (FIB == FIE) 1576 break; 1577 } 1578 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead)) 1579 break; 1580 1581 if (TII->isPredicated(TIB)) 1582 // Hard to reason about register liveness with predicated instruction. 1583 break; 1584 1585 bool IsSafe = true; 1586 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1587 MachineOperand &MO = TIB->getOperand(i); 1588 if (!MO.isReg()) 1589 continue; 1590 unsigned Reg = MO.getReg(); 1591 if (!Reg) 1592 continue; 1593 if (MO.isDef()) { 1594 if (Uses.count(Reg)) { 1595 // Avoid clobbering a register that's used by the instruction at 1596 // the point of insertion. 1597 IsSafe = false; 1598 break; 1599 } 1600 1601 if (Defs.count(Reg) && !MO.isDead()) { 1602 // Don't hoist the instruction if the def would be clobber by the 1603 // instruction at the point insertion. FIXME: This is overly 1604 // conservative. It should be possible to hoist the instructions 1605 // in BB2 in the following example: 1606 // BB1: 1607 // r1, eflag = op1 r2, r3 1608 // brcc eflag 1609 // 1610 // BB2: 1611 // r1 = op2, ... 1612 // = op3, r1<kill> 1613 IsSafe = false; 1614 break; 1615 } 1616 } else if (!LocalDefsSet.count(Reg)) { 1617 if (Defs.count(Reg)) { 1618 // Use is defined by the instruction at the point of insertion. 1619 IsSafe = false; 1620 break; 1621 } 1622 1623 if (MO.isKill() && Uses.count(Reg)) 1624 // Kills a register that's read by the instruction at the point of 1625 // insertion. Remove the kill marker. 1626 MO.setIsKill(false); 1627 } 1628 } 1629 if (!IsSafe) 1630 break; 1631 1632 bool DontMoveAcrossStore = true; 1633 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore)) 1634 break; 1635 1636 // Remove kills from LocalDefsSet, these registers had short live ranges. 1637 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1638 MachineOperand &MO = TIB->getOperand(i); 1639 if (!MO.isReg() || !MO.isUse() || !MO.isKill()) 1640 continue; 1641 unsigned Reg = MO.getReg(); 1642 if (!Reg || !LocalDefsSet.count(Reg)) 1643 continue; 1644 for (const unsigned *OR = TRI->getOverlaps(Reg); *OR; ++OR) 1645 LocalDefsSet.erase(*OR); 1646 } 1647 1648 // Track local defs so we can update liveins. 1649 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1650 MachineOperand &MO = TIB->getOperand(i); 1651 if (!MO.isReg() || !MO.isDef() || MO.isDead()) 1652 continue; 1653 unsigned Reg = MO.getReg(); 1654 if (!Reg) 1655 continue; 1656 LocalDefs.push_back(Reg); 1657 for (const unsigned *OR = TRI->getOverlaps(Reg); *OR; ++OR) 1658 LocalDefsSet.insert(*OR); 1659 } 1660 1661 HasDups = true;; 1662 ++TIB; 1663 ++FIB; 1664 } 1665 1666 if (!HasDups) 1667 return false; 1668 1669 MBB->splice(Loc, TBB, TBB->begin(), TIB); 1670 FBB->erase(FBB->begin(), FIB); 1671 1672 // Update livein's. 1673 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) { 1674 unsigned Def = LocalDefs[i]; 1675 if (LocalDefsSet.count(Def)) { 1676 TBB->addLiveIn(Def); 1677 FBB->addLiveIn(Def); 1678 } 1679 } 1680 1681 ++NumHoist; 1682 return true; 1683} 1684