BranchFolding.cpp revision c16c25fbc3b53da99dcaf27685a6116249f79b30
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#endif 487 return false; 488 } 489} 490 491/// CountTerminators - Count the number of terminators in the given 492/// block and set I to the position of the first non-terminator, if there 493/// is one, or MBB->end() otherwise. 494static unsigned CountTerminators(MachineBasicBlock *MBB, 495 MachineBasicBlock::iterator &I) { 496 I = MBB->end(); 497 unsigned NumTerms = 0; 498 for (;;) { 499 if (I == MBB->begin()) { 500 I = MBB->end(); 501 break; 502 } 503 --I; 504 if (!I->isTerminator()) break; 505 ++NumTerms; 506 } 507 return NumTerms; 508} 509 510/// ProfitableToMerge - Check if two machine basic blocks have a common tail 511/// and decide if it would be profitable to merge those tails. Return the 512/// length of the common tail and iterators to the first common instruction 513/// in each block. 514static bool ProfitableToMerge(MachineBasicBlock *MBB1, 515 MachineBasicBlock *MBB2, 516 unsigned minCommonTailLength, 517 unsigned &CommonTailLen, 518 MachineBasicBlock::iterator &I1, 519 MachineBasicBlock::iterator &I2, 520 MachineBasicBlock *SuccBB, 521 MachineBasicBlock *PredBB) { 522 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 523 if (CommonTailLen == 0) 524 return false; 525 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber() 526 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen 527 << '\n'); 528 529 // It's almost always profitable to merge any number of non-terminator 530 // instructions with the block that falls through into the common successor. 531 if (MBB1 == PredBB || MBB2 == PredBB) { 532 MachineBasicBlock::iterator I; 533 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 534 if (CommonTailLen > NumTerms) 535 return true; 536 } 537 538 // If one of the blocks can be completely merged and happens to be in 539 // a position where the other could fall through into it, merge any number 540 // of instructions, because it can be done without a branch. 541 // TODO: If the blocks are not adjacent, move one of them so that they are? 542 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 543 return true; 544 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 545 return true; 546 547 // If both blocks have an unconditional branch temporarily stripped out, 548 // count that as an additional common instruction for the following 549 // heuristics. 550 unsigned EffectiveTailLen = CommonTailLen; 551 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 552 !MBB1->back().isBarrier() && 553 !MBB2->back().isBarrier()) 554 ++EffectiveTailLen; 555 556 // Check if the common tail is long enough to be worthwhile. 557 if (EffectiveTailLen >= minCommonTailLength) 558 return true; 559 560 // If we are optimizing for code size, 2 instructions in common is enough if 561 // we don't have to split a block. At worst we will be introducing 1 new 562 // branch instruction, which is likely to be smaller than the 2 563 // instructions that would be deleted in the merge. 564 MachineFunction *MF = MBB1->getParent(); 565 if (EffectiveTailLen >= 2 && 566 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 567 (I1 == MBB1->begin() || I2 == MBB2->begin())) 568 return true; 569 570 return false; 571} 572 573/// ComputeSameTails - Look through all the blocks in MergePotentials that have 574/// hash CurHash (guaranteed to match the last element). Build the vector 575/// SameTails of all those that have the (same) largest number of instructions 576/// in common of any pair of these blocks. SameTails entries contain an 577/// iterator into MergePotentials (from which the MachineBasicBlock can be 578/// found) and a MachineBasicBlock::iterator into that MBB indicating the 579/// instruction where the matching code sequence begins. 580/// Order of elements in SameTails is the reverse of the order in which 581/// those blocks appear in MergePotentials (where they are not necessarily 582/// consecutive). 583unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 584 unsigned minCommonTailLength, 585 MachineBasicBlock *SuccBB, 586 MachineBasicBlock *PredBB) { 587 unsigned maxCommonTailLength = 0U; 588 SameTails.clear(); 589 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 590 MPIterator HighestMPIter = prior(MergePotentials.end()); 591 for (MPIterator CurMPIter = prior(MergePotentials.end()), 592 B = MergePotentials.begin(); 593 CurMPIter != B && CurMPIter->getHash() == CurHash; 594 --CurMPIter) { 595 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 596 unsigned CommonTailLen; 597 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 598 minCommonTailLength, 599 CommonTailLen, TrialBBI1, TrialBBI2, 600 SuccBB, PredBB)) { 601 if (CommonTailLen > maxCommonTailLength) { 602 SameTails.clear(); 603 maxCommonTailLength = CommonTailLen; 604 HighestMPIter = CurMPIter; 605 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 606 } 607 if (HighestMPIter == CurMPIter && 608 CommonTailLen == maxCommonTailLength) 609 SameTails.push_back(SameTailElt(I, TrialBBI2)); 610 } 611 if (I == B) 612 break; 613 } 614 } 615 return maxCommonTailLength; 616} 617 618/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 619/// MergePotentials, restoring branches at ends of blocks as appropriate. 620void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 621 MachineBasicBlock *SuccBB, 622 MachineBasicBlock *PredBB) { 623 MPIterator CurMPIter, B; 624 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 625 CurMPIter->getHash() == CurHash; 626 --CurMPIter) { 627 // Put the unconditional branch back, if we need one. 628 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 629 if (SuccBB && CurMBB != PredBB) 630 FixTail(CurMBB, SuccBB, TII); 631 if (CurMPIter == B) 632 break; 633 } 634 if (CurMPIter->getHash() != CurHash) 635 CurMPIter++; 636 MergePotentials.erase(CurMPIter, MergePotentials.end()); 637} 638 639/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 640/// only of the common tail. Create a block that does by splitting one. 641bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 642 unsigned maxCommonTailLength, 643 unsigned &commonTailIndex) { 644 commonTailIndex = 0; 645 unsigned TimeEstimate = ~0U; 646 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 647 // Use PredBB if possible; that doesn't require a new branch. 648 if (SameTails[i].getBlock() == PredBB) { 649 commonTailIndex = i; 650 break; 651 } 652 // Otherwise, make a (fairly bogus) choice based on estimate of 653 // how long it will take the various blocks to execute. 654 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 655 SameTails[i].getTailStartPos()); 656 if (t <= TimeEstimate) { 657 TimeEstimate = t; 658 commonTailIndex = i; 659 } 660 } 661 662 MachineBasicBlock::iterator BBI = 663 SameTails[commonTailIndex].getTailStartPos(); 664 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 665 666 // If the common tail includes any debug info we will take it pretty 667 // randomly from one of the inputs. Might be better to remove it? 668 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 669 << maxCommonTailLength); 670 671 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 672 if (!newMBB) { 673 DEBUG(dbgs() << "... failed!"); 674 return false; 675 } 676 677 SameTails[commonTailIndex].setBlock(newMBB); 678 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 679 680 // If we split PredBB, newMBB is the new predecessor. 681 if (PredBB == MBB) 682 PredBB = newMBB; 683 684 return true; 685} 686 687// See if any of the blocks in MergePotentials (which all have a common single 688// successor, or all have no successor) can be tail-merged. If there is a 689// successor, any blocks in MergePotentials that are not tail-merged and 690// are not immediately before Succ must have an unconditional branch to 691// Succ added (but the predecessor/successor lists need no adjustment). 692// The lone predecessor of Succ that falls through into Succ, 693// if any, is given in PredBB. 694 695bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 696 MachineBasicBlock *PredBB) { 697 bool MadeChange = false; 698 699 // Except for the special cases below, tail-merge if there are at least 700 // this many instructions in common. 701 unsigned minCommonTailLength = TailMergeSize; 702 703 DEBUG(dbgs() << "\nTryTailMergeBlocks: "; 704 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 705 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 706 << (i == e-1 ? "" : ", "); 707 dbgs() << "\n"; 708 if (SuccBB) { 709 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 710 if (PredBB) 711 dbgs() << " which has fall-through from BB#" 712 << PredBB->getNumber() << "\n"; 713 } 714 dbgs() << "Looking for common tails of at least " 715 << minCommonTailLength << " instruction" 716 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 717 ); 718 719 // Sort by hash value so that blocks with identical end sequences sort 720 // together. 721 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 722 723 // Walk through equivalence sets looking for actual exact matches. 724 while (MergePotentials.size() > 1) { 725 unsigned CurHash = MergePotentials.back().getHash(); 726 727 // Build SameTails, identifying the set of blocks with this hash code 728 // and with the maximum number of instructions in common. 729 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 730 minCommonTailLength, 731 SuccBB, PredBB); 732 733 // If we didn't find any pair that has at least minCommonTailLength 734 // instructions in common, remove all blocks with this hash code and retry. 735 if (SameTails.empty()) { 736 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 737 continue; 738 } 739 740 // If one of the blocks is the entire common tail (and not the entry 741 // block, which we can't jump to), we can treat all blocks with this same 742 // tail at once. Use PredBB if that is one of the possibilities, as that 743 // will not introduce any extra branches. 744 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 745 getParent()->begin(); 746 unsigned commonTailIndex = SameTails.size(); 747 // If there are two blocks, check to see if one can be made to fall through 748 // into the other. 749 if (SameTails.size() == 2 && 750 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 751 SameTails[1].tailIsWholeBlock()) 752 commonTailIndex = 1; 753 else if (SameTails.size() == 2 && 754 SameTails[1].getBlock()->isLayoutSuccessor( 755 SameTails[0].getBlock()) && 756 SameTails[0].tailIsWholeBlock()) 757 commonTailIndex = 0; 758 else { 759 // Otherwise just pick one, favoring the fall-through predecessor if 760 // there is one. 761 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 762 MachineBasicBlock *MBB = SameTails[i].getBlock(); 763 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 764 continue; 765 if (MBB == PredBB) { 766 commonTailIndex = i; 767 break; 768 } 769 if (SameTails[i].tailIsWholeBlock()) 770 commonTailIndex = i; 771 } 772 } 773 774 if (commonTailIndex == SameTails.size() || 775 (SameTails[commonTailIndex].getBlock() == PredBB && 776 !SameTails[commonTailIndex].tailIsWholeBlock())) { 777 // None of the blocks consist entirely of the common tail. 778 // Split a block so that one does. 779 if (!CreateCommonTailOnlyBlock(PredBB, 780 maxCommonTailLength, commonTailIndex)) { 781 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 782 continue; 783 } 784 } 785 786 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 787 // MBB is common tail. Adjust all other BB's to jump to this one. 788 // Traversal must be forwards so erases work. 789 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber() 790 << " for "); 791 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 792 if (commonTailIndex == i) 793 continue; 794 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() 795 << (i == e-1 ? "" : ", ")); 796 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 797 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 798 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 799 MergePotentials.erase(SameTails[i].getMPIter()); 800 } 801 DEBUG(dbgs() << "\n"); 802 // We leave commonTailIndex in the worklist in case there are other blocks 803 // that match it with a smaller number of instructions. 804 MadeChange = true; 805 } 806 return MadeChange; 807} 808 809bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 810 811 if (!EnableTailMerge) return false; 812 813 bool MadeChange = false; 814 815 // First find blocks with no successors. 816 MergePotentials.clear(); 817 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); 818 I != E && MergePotentials.size() < TailMergeThreshold; ++I) { 819 if (TriedMerging.count(I)) 820 continue; 821 if (I->succ_empty()) 822 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I)); 823 } 824 825 // If this is a large problem, avoid visiting the same basic blocks 826 // multiple times. 827 if (MergePotentials.size() == TailMergeThreshold) 828 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 829 TriedMerging.insert(MergePotentials[i].getBlock()); 830 // See if we can do any tail merging on those. 831 if (MergePotentials.size() >= 2) 832 MadeChange |= TryTailMergeBlocks(NULL, NULL); 833 834 // Look at blocks (IBB) with multiple predecessors (PBB). 835 // We change each predecessor to a canonical form, by 836 // (1) temporarily removing any unconditional branch from the predecessor 837 // to IBB, and 838 // (2) alter conditional branches so they branch to the other block 839 // not IBB; this may require adding back an unconditional branch to IBB 840 // later, where there wasn't one coming in. E.g. 841 // Bcc IBB 842 // fallthrough to QBB 843 // here becomes 844 // Bncc QBB 845 // with a conceptual B to IBB after that, which never actually exists. 846 // With those changes, we see whether the predecessors' tails match, 847 // and merge them if so. We change things out of canonical form and 848 // back to the way they were later in the process. (OptimizeBranches 849 // would undo some of this, but we can't use it, because we'd get into 850 // a compile-time infinite loop repeatedly doing and undoing the same 851 // transformations.) 852 853 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 854 I != E; ++I) { 855 if (I->pred_size() >= 2) { 856 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 857 MachineBasicBlock *IBB = I; 858 MachineBasicBlock *PredBB = prior(I); 859 MergePotentials.clear(); 860 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 861 E2 = I->pred_end(); 862 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) { 863 MachineBasicBlock *PBB = *P; 864 if (TriedMerging.count(PBB)) 865 continue; 866 // Skip blocks that loop to themselves, can't tail merge these. 867 if (PBB == IBB) 868 continue; 869 // Visit each predecessor only once. 870 if (!UniquePreds.insert(PBB)) 871 continue; 872 // Skip blocks which may jump to a landing pad. Can't tail merge these. 873 if (PBB->getLandingPadSuccessor()) 874 continue; 875 MachineBasicBlock *TBB = 0, *FBB = 0; 876 SmallVector<MachineOperand, 4> Cond; 877 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 878 // Failing case: IBB is the target of a cbr, and 879 // we cannot reverse the branch. 880 SmallVector<MachineOperand, 4> NewCond(Cond); 881 if (!Cond.empty() && TBB == IBB) { 882 if (TII->ReverseBranchCondition(NewCond)) 883 continue; 884 // This is the QBB case described above 885 if (!FBB) 886 FBB = llvm::next(MachineFunction::iterator(PBB)); 887 } 888 // Failing case: the only way IBB can be reached from PBB is via 889 // exception handling. Happens for landing pads. Would be nice 890 // to have a bit in the edge so we didn't have to do all this. 891 if (IBB->isLandingPad()) { 892 MachineFunction::iterator IP = PBB; IP++; 893 MachineBasicBlock *PredNextBB = NULL; 894 if (IP != MF.end()) 895 PredNextBB = IP; 896 if (TBB == NULL) { 897 if (IBB != PredNextBB) // fallthrough 898 continue; 899 } else if (FBB) { 900 if (TBB != IBB && FBB != IBB) // cbr then ubr 901 continue; 902 } else if (Cond.empty()) { 903 if (TBB != IBB) // ubr 904 continue; 905 } else { 906 if (TBB != IBB && IBB != PredNextBB) // cbr 907 continue; 908 } 909 } 910 // Remove the unconditional branch at the end, if any. 911 if (TBB && (Cond.empty() || FBB)) { 912 DebugLoc dl; // FIXME: this is nowhere 913 TII->RemoveBranch(*PBB); 914 if (!Cond.empty()) 915 // reinsert conditional branch only, for now 916 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl); 917 } 918 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P)); 919 } 920 } 921 // If this is a large problem, avoid visiting the same basic blocks 922 // multiple times. 923 if (MergePotentials.size() == TailMergeThreshold) 924 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 925 TriedMerging.insert(MergePotentials[i].getBlock()); 926 if (MergePotentials.size() >= 2) 927 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 928 // Reinsert an unconditional branch if needed. 929 // The 1 below can occur as a result of removing blocks in 930 // TryTailMergeBlocks. 931 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 932 if (MergePotentials.size() == 1 && 933 MergePotentials.begin()->getBlock() != PredBB) 934 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 935 } 936 } 937 return MadeChange; 938} 939 940//===----------------------------------------------------------------------===// 941// Branch Optimization 942//===----------------------------------------------------------------------===// 943 944bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 945 bool MadeChange = false; 946 947 // Make sure blocks are numbered in order 948 MF.RenumberBlocks(); 949 950 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 951 I != E; ) { 952 MachineBasicBlock *MBB = I++; 953 MadeChange |= OptimizeBlock(MBB); 954 955 // If it is dead, remove it. 956 if (MBB->pred_empty()) { 957 RemoveDeadBlock(MBB); 958 MadeChange = true; 959 ++NumDeadBlocks; 960 } 961 } 962 return MadeChange; 963} 964 965// Blocks should be considered empty if they contain only debug info; 966// else the debug info would affect codegen. 967static bool IsEmptyBlock(MachineBasicBlock *MBB) { 968 if (MBB->empty()) 969 return true; 970 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end(); 971 MBBI!=MBBE; ++MBBI) { 972 if (!MBBI->isDebugValue()) 973 return false; 974 } 975 return true; 976} 977 978// Blocks with only debug info and branches should be considered the same 979// as blocks with only branches. 980static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { 981 MachineBasicBlock::iterator MBBI, MBBE; 982 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) { 983 if (!MBBI->isDebugValue()) 984 break; 985 } 986 return (MBBI->isBranch()); 987} 988 989/// IsBetterFallthrough - Return true if it would be clearly better to 990/// fall-through to MBB1 than to fall through into MBB2. This has to return 991/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 992/// result in infinite loops. 993static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 994 MachineBasicBlock *MBB2) { 995 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 996 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 997 // optimize branches that branch to either a return block or an assert block 998 // into a fallthrough to the return. 999 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false; 1000 1001 // If there is a clear successor ordering we make sure that one block 1002 // will fall through to the next 1003 if (MBB1->isSuccessor(MBB2)) return true; 1004 if (MBB2->isSuccessor(MBB1)) return false; 1005 1006 // Neither block consists entirely of debug info (per IsEmptyBlock check), 1007 // so we needn't test for falling off the beginning here. 1008 MachineBasicBlock::iterator MBB1I = --MBB1->end(); 1009 while (MBB1I->isDebugValue()) 1010 --MBB1I; 1011 MachineBasicBlock::iterator MBB2I = --MBB2->end(); 1012 while (MBB2I->isDebugValue()) 1013 --MBB2I; 1014 return MBB2I->isCall() && !MBB1I->isCall(); 1015} 1016 1017/// OptimizeBlock - Analyze and optimize control flow related to the specified 1018/// block. This is never called on the entry block. 1019bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1020 bool MadeChange = false; 1021 MachineFunction &MF = *MBB->getParent(); 1022 DebugLoc dl; // FIXME: this is nowhere 1023ReoptimizeBlock: 1024 1025 MachineFunction::iterator FallThrough = MBB; 1026 ++FallThrough; 1027 1028 // If this block is empty, make everyone use its fall-through, not the block 1029 // explicitly. Landing pads should not do this since the landing-pad table 1030 // points to this block. Blocks with their addresses taken shouldn't be 1031 // optimized away. 1032 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 1033 // Dead block? Leave for cleanup later. 1034 if (MBB->pred_empty()) return MadeChange; 1035 1036 if (FallThrough == MF.end()) { 1037 // TODO: Simplify preds to not branch here if possible! 1038 } else { 1039 // Rewrite all predecessors of the old block to go to the fallthrough 1040 // instead. 1041 while (!MBB->pred_empty()) { 1042 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1043 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1044 } 1045 // If MBB was the target of a jump table, update jump tables to go to the 1046 // fallthrough instead. 1047 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1048 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough); 1049 MadeChange = true; 1050 } 1051 return MadeChange; 1052 } 1053 1054 // Check to see if we can simplify the terminator of the block before this 1055 // one. 1056 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1057 1058 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1059 SmallVector<MachineOperand, 4> PriorCond; 1060 bool PriorUnAnalyzable = 1061 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1062 if (!PriorUnAnalyzable) { 1063 // If the CFG for the prior block has extra edges, remove them. 1064 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1065 !PriorCond.empty()); 1066 1067 // If the previous branch is conditional and both conditions go to the same 1068 // destination, remove the branch, replacing it with an unconditional one or 1069 // a fall-through. 1070 if (PriorTBB && PriorTBB == PriorFBB) { 1071 TII->RemoveBranch(PrevBB); 1072 PriorCond.clear(); 1073 if (PriorTBB != MBB) 1074 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1075 MadeChange = true; 1076 ++NumBranchOpts; 1077 goto ReoptimizeBlock; 1078 } 1079 1080 // If the previous block unconditionally falls through to this block and 1081 // this block has no other predecessors, move the contents of this block 1082 // into the prior block. This doesn't usually happen when SimplifyCFG 1083 // has been used, but it can happen if tail merging splits a fall-through 1084 // predecessor of a block. 1085 // This has to check PrevBB->succ_size() because EH edges are ignored by 1086 // AnalyzeBranch. 1087 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1088 PrevBB.succ_size() == 1 && 1089 !MBB->hasAddressTaken() && !MBB->isLandingPad()) { 1090 DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1091 << "From MBB: " << *MBB); 1092 // Remove redundant DBG_VALUEs first. 1093 if (PrevBB.begin() != PrevBB.end()) { 1094 MachineBasicBlock::iterator PrevBBIter = PrevBB.end(); 1095 --PrevBBIter; 1096 MachineBasicBlock::iterator MBBIter = MBB->begin(); 1097 // Check if DBG_VALUE at the end of PrevBB is identical to the 1098 // DBG_VALUE at the beginning of MBB. 1099 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end() 1100 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) { 1101 if (!MBBIter->isIdenticalTo(PrevBBIter)) 1102 break; 1103 MachineInstr *DuplicateDbg = MBBIter; 1104 ++MBBIter; -- PrevBBIter; 1105 DuplicateDbg->eraseFromParent(); 1106 } 1107 } 1108 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1109 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1110 assert(PrevBB.succ_empty()); 1111 PrevBB.transferSuccessors(MBB); 1112 MadeChange = true; 1113 return MadeChange; 1114 } 1115 1116 // If the previous branch *only* branches to *this* block (conditional or 1117 // not) remove the branch. 1118 if (PriorTBB == MBB && PriorFBB == 0) { 1119 TII->RemoveBranch(PrevBB); 1120 MadeChange = true; 1121 ++NumBranchOpts; 1122 goto ReoptimizeBlock; 1123 } 1124 1125 // If the prior block branches somewhere else on the condition and here if 1126 // the condition is false, remove the uncond second branch. 1127 if (PriorFBB == MBB) { 1128 TII->RemoveBranch(PrevBB); 1129 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1130 MadeChange = true; 1131 ++NumBranchOpts; 1132 goto ReoptimizeBlock; 1133 } 1134 1135 // If the prior block branches here on true and somewhere else on false, and 1136 // if the branch condition is reversible, reverse the branch to create a 1137 // fall-through. 1138 if (PriorTBB == MBB) { 1139 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1140 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1141 TII->RemoveBranch(PrevBB); 1142 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl); 1143 MadeChange = true; 1144 ++NumBranchOpts; 1145 goto ReoptimizeBlock; 1146 } 1147 } 1148 1149 // If this block has no successors (e.g. it is a return block or ends with 1150 // a call to a no-return function like abort or __cxa_throw) and if the pred 1151 // falls through into this block, and if it would otherwise fall through 1152 // into the block after this, move this block to the end of the function. 1153 // 1154 // We consider it more likely that execution will stay in the function (e.g. 1155 // due to loops) than it is to exit it. This asserts in loops etc, moving 1156 // the assert condition out of the loop body. 1157 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1158 MachineFunction::iterator(PriorTBB) == FallThrough && 1159 !MBB->canFallThrough()) { 1160 bool DoTransform = true; 1161 1162 // We have to be careful that the succs of PredBB aren't both no-successor 1163 // blocks. If neither have successors and if PredBB is the second from 1164 // last block in the function, we'd just keep swapping the two blocks for 1165 // last. Only do the swap if one is clearly better to fall through than 1166 // the other. 1167 if (FallThrough == --MF.end() && 1168 !IsBetterFallthrough(PriorTBB, MBB)) 1169 DoTransform = false; 1170 1171 if (DoTransform) { 1172 // Reverse the branch so we will fall through on the previous true cond. 1173 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1174 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1175 DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1176 << "To make fallthrough to: " << *PriorTBB << "\n"); 1177 1178 TII->RemoveBranch(PrevBB); 1179 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl); 1180 1181 // Move this block to the end of the function. 1182 MBB->moveAfter(--MF.end()); 1183 MadeChange = true; 1184 ++NumBranchOpts; 1185 return MadeChange; 1186 } 1187 } 1188 } 1189 } 1190 1191 // Analyze the branch in the current block. 1192 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1193 SmallVector<MachineOperand, 4> CurCond; 1194 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1195 if (!CurUnAnalyzable) { 1196 // If the CFG for the prior block has extra edges, remove them. 1197 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1198 1199 // If this is a two-way branch, and the FBB branches to this block, reverse 1200 // the condition so the single-basic-block loop is faster. Instead of: 1201 // Loop: xxx; jcc Out; jmp Loop 1202 // we want: 1203 // Loop: xxx; jncc Loop; jmp Out 1204 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1205 SmallVector<MachineOperand, 4> NewCond(CurCond); 1206 if (!TII->ReverseBranchCondition(NewCond)) { 1207 TII->RemoveBranch(*MBB); 1208 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); 1209 MadeChange = true; 1210 ++NumBranchOpts; 1211 goto ReoptimizeBlock; 1212 } 1213 } 1214 1215 // If this branch is the only thing in its block, see if we can forward 1216 // other blocks across it. 1217 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1218 IsBranchOnlyBlock(MBB) && CurTBB != MBB && 1219 !MBB->hasAddressTaken()) { 1220 // This block may contain just an unconditional branch. Because there can 1221 // be 'non-branch terminators' in the block, try removing the branch and 1222 // then seeing if the block is empty. 1223 TII->RemoveBranch(*MBB); 1224 // If the only things remaining in the block are debug info, remove these 1225 // as well, so this will behave the same as an empty block in non-debug 1226 // mode. 1227 if (!MBB->empty()) { 1228 bool NonDebugInfoFound = false; 1229 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 1230 I != E; ++I) { 1231 if (!I->isDebugValue()) { 1232 NonDebugInfoFound = true; 1233 break; 1234 } 1235 } 1236 if (!NonDebugInfoFound) 1237 // Make the block empty, losing the debug info (we could probably 1238 // improve this in some cases.) 1239 MBB->erase(MBB->begin(), MBB->end()); 1240 } 1241 // If this block is just an unconditional branch to CurTBB, we can 1242 // usually completely eliminate the block. The only case we cannot 1243 // completely eliminate the block is when the block before this one 1244 // falls through into MBB and we can't understand the prior block's branch 1245 // condition. 1246 if (MBB->empty()) { 1247 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1248 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1249 !PrevBB.isSuccessor(MBB)) { 1250 // If the prior block falls through into us, turn it into an 1251 // explicit branch to us to make updates simpler. 1252 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1253 PriorTBB != MBB && PriorFBB != MBB) { 1254 if (PriorTBB == 0) { 1255 assert(PriorCond.empty() && PriorFBB == 0 && 1256 "Bad branch analysis"); 1257 PriorTBB = MBB; 1258 } else { 1259 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1260 PriorFBB = MBB; 1261 } 1262 TII->RemoveBranch(PrevBB); 1263 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, dl); 1264 } 1265 1266 // Iterate through all the predecessors, revectoring each in-turn. 1267 size_t PI = 0; 1268 bool DidChange = false; 1269 bool HasBranchToSelf = false; 1270 while(PI != MBB->pred_size()) { 1271 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1272 if (PMBB == MBB) { 1273 // If this block has an uncond branch to itself, leave it. 1274 ++PI; 1275 HasBranchToSelf = true; 1276 } else { 1277 DidChange = true; 1278 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1279 // If this change resulted in PMBB ending in a conditional 1280 // branch where both conditions go to the same destination, 1281 // change this to an unconditional branch (and fix the CFG). 1282 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1283 SmallVector<MachineOperand, 4> NewCurCond; 1284 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1285 NewCurFBB, NewCurCond, true); 1286 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1287 TII->RemoveBranch(*PMBB); 1288 NewCurCond.clear(); 1289 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, dl); 1290 MadeChange = true; 1291 ++NumBranchOpts; 1292 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1293 } 1294 } 1295 } 1296 1297 // Change any jumptables to go to the new MBB. 1298 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1299 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); 1300 if (DidChange) { 1301 ++NumBranchOpts; 1302 MadeChange = true; 1303 if (!HasBranchToSelf) return MadeChange; 1304 } 1305 } 1306 } 1307 1308 // Add the branch back if the block is more than just an uncond branch. 1309 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl); 1310 } 1311 } 1312 1313 // If the prior block doesn't fall through into this block, and if this 1314 // block doesn't fall through into some other block, see if we can find a 1315 // place to move this block where a fall-through will happen. 1316 if (!PrevBB.canFallThrough()) { 1317 1318 // Now we know that there was no fall-through into this block, check to 1319 // see if it has a fall-through into its successor. 1320 bool CurFallsThru = MBB->canFallThrough(); 1321 1322 if (!MBB->isLandingPad()) { 1323 // Check all the predecessors of this block. If one of them has no fall 1324 // throughs, move this block right after it. 1325 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1326 E = MBB->pred_end(); PI != E; ++PI) { 1327 // Analyze the branch at the end of the pred. 1328 MachineBasicBlock *PredBB = *PI; 1329 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1330 MachineBasicBlock *PredTBB = 0, *PredFBB = 0; 1331 SmallVector<MachineOperand, 4> PredCond; 1332 if (PredBB != MBB && !PredBB->canFallThrough() && 1333 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1334 && (!CurFallsThru || !CurTBB || !CurFBB) 1335 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1336 // If the current block doesn't fall through, just move it. 1337 // If the current block can fall through and does not end with a 1338 // conditional branch, we need to append an unconditional jump to 1339 // the (current) next block. To avoid a possible compile-time 1340 // infinite loop, move blocks only backward in this case. 1341 // Also, if there are already 2 branches here, we cannot add a third; 1342 // this means we have the case 1343 // Bcc next 1344 // B elsewhere 1345 // next: 1346 if (CurFallsThru) { 1347 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); 1348 CurCond.clear(); 1349 TII->InsertBranch(*MBB, NextBB, 0, CurCond, dl); 1350 } 1351 MBB->moveAfter(PredBB); 1352 MadeChange = true; 1353 goto ReoptimizeBlock; 1354 } 1355 } 1356 } 1357 1358 if (!CurFallsThru) { 1359 // Check all successors to see if we can move this block before it. 1360 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1361 E = MBB->succ_end(); SI != E; ++SI) { 1362 // Analyze the branch at the end of the block before the succ. 1363 MachineBasicBlock *SuccBB = *SI; 1364 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1365 1366 // If this block doesn't already fall-through to that successor, and if 1367 // the succ doesn't already have a block that can fall through into it, 1368 // and if the successor isn't an EH destination, we can arrange for the 1369 // fallthrough to happen. 1370 if (SuccBB != MBB && &*SuccPrev != MBB && 1371 !SuccPrev->canFallThrough() && !CurUnAnalyzable && 1372 !SuccBB->isLandingPad()) { 1373 MBB->moveBefore(SuccBB); 1374 MadeChange = true; 1375 goto ReoptimizeBlock; 1376 } 1377 } 1378 1379 // Okay, there is no really great place to put this block. If, however, 1380 // the block before this one would be a fall-through if this block were 1381 // removed, move this block to the end of the function. 1382 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0; 1383 SmallVector<MachineOperand, 4> PrevCond; 1384 if (FallThrough != MF.end() && 1385 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1386 PrevBB.isSuccessor(FallThrough)) { 1387 MBB->moveAfter(--MF.end()); 1388 MadeChange = true; 1389 return MadeChange; 1390 } 1391 } 1392 } 1393 1394 return MadeChange; 1395} 1396 1397//===----------------------------------------------------------------------===// 1398// Hoist Common Code 1399//===----------------------------------------------------------------------===// 1400 1401/// HoistCommonCode - Hoist common instruction sequences at the start of basic 1402/// blocks to their common predecessor. 1403bool BranchFolder::HoistCommonCode(MachineFunction &MF) { 1404 bool MadeChange = false; 1405 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) { 1406 MachineBasicBlock *MBB = I++; 1407 MadeChange |= HoistCommonCodeInSuccs(MBB); 1408 } 1409 1410 return MadeChange; 1411} 1412 1413/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 1414/// its 'true' successor. 1415static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 1416 MachineBasicBlock *TrueBB) { 1417 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 1418 E = BB->succ_end(); SI != E; ++SI) { 1419 MachineBasicBlock *SuccBB = *SI; 1420 if (SuccBB != TrueBB) 1421 return SuccBB; 1422 } 1423 return NULL; 1424} 1425 1426/// findHoistingInsertPosAndDeps - Find the location to move common instructions 1427/// in successors to. The location is ususally just before the terminator, 1428/// however if the terminator is a conditional branch and its previous 1429/// instruction is the flag setting instruction, the previous instruction is 1430/// the preferred location. This function also gathers uses and defs of the 1431/// instructions from the insertion point to the end of the block. The data is 1432/// used by HoistCommonCodeInSuccs to ensure safety. 1433static 1434MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB, 1435 const TargetInstrInfo *TII, 1436 const TargetRegisterInfo *TRI, 1437 SmallSet<unsigned,4> &Uses, 1438 SmallSet<unsigned,4> &Defs) { 1439 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator(); 1440 if (!TII->isUnpredicatedTerminator(Loc)) 1441 return MBB->end(); 1442 1443 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) { 1444 const MachineOperand &MO = Loc->getOperand(i); 1445 if (!MO.isReg()) 1446 continue; 1447 unsigned Reg = MO.getReg(); 1448 if (!Reg) 1449 continue; 1450 if (MO.isUse()) { 1451 Uses.insert(Reg); 1452 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1453 Uses.insert(*AS); 1454 } else if (!MO.isDead()) 1455 // Don't try to hoist code in the rare case the terminator defines a 1456 // register that is later used. 1457 return MBB->end(); 1458 } 1459 1460 if (Uses.empty()) 1461 return Loc; 1462 if (Loc == MBB->begin()) 1463 return MBB->end(); 1464 1465 // The terminator is probably a conditional branch, try not to separate the 1466 // branch from condition setting instruction. 1467 MachineBasicBlock::iterator PI = Loc; 1468 --PI; 1469 while (PI != MBB->begin() && Loc->isDebugValue()) 1470 --PI; 1471 1472 bool IsDef = false; 1473 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) { 1474 const MachineOperand &MO = PI->getOperand(i); 1475 if (!MO.isReg() || MO.isUse()) 1476 continue; 1477 unsigned Reg = MO.getReg(); 1478 if (!Reg) 1479 continue; 1480 if (Uses.count(Reg)) 1481 IsDef = true; 1482 } 1483 if (!IsDef) 1484 // The condition setting instruction is not just before the conditional 1485 // branch. 1486 return Loc; 1487 1488 // Be conservative, don't insert instruction above something that may have 1489 // side-effects. And since it's potentially bad to separate flag setting 1490 // instruction from the conditional branch, just abort the optimization 1491 // completely. 1492 // Also avoid moving code above predicated instruction since it's hard to 1493 // reason about register liveness with predicated instruction. 1494 bool DontMoveAcrossStore = true; 1495 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) || 1496 TII->isPredicated(PI)) 1497 return MBB->end(); 1498 1499 1500 // Find out what registers are live. Note this routine is ignoring other live 1501 // registers which are only used by instructions in successor blocks. 1502 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) { 1503 const MachineOperand &MO = PI->getOperand(i); 1504 if (!MO.isReg()) 1505 continue; 1506 unsigned Reg = MO.getReg(); 1507 if (!Reg) 1508 continue; 1509 if (MO.isUse()) { 1510 Uses.insert(Reg); 1511 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1512 Uses.insert(*AS); 1513 } else { 1514 if (Uses.count(Reg)) { 1515 Uses.erase(Reg); 1516 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1517 Uses.erase(*SR); // Use getSubRegisters to be conservative 1518 } 1519 Defs.insert(Reg); 1520 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1521 Defs.insert(*AS); 1522 } 1523 } 1524 1525 return PI; 1526} 1527 1528/// HoistCommonCodeInSuccs - If the successors of MBB has common instruction 1529/// sequence at the start of the function, move the instructions before MBB 1530/// terminator if it's legal. 1531bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) { 1532 MachineBasicBlock *TBB = 0, *FBB = 0; 1533 SmallVector<MachineOperand, 4> Cond; 1534 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty()) 1535 return false; 1536 1537 if (!FBB) FBB = findFalseBlock(MBB, TBB); 1538 if (!FBB) 1539 // Malformed bcc? True and false blocks are the same? 1540 return false; 1541 1542 // Restrict the optimization to cases where MBB is the only predecessor, 1543 // it is an obvious win. 1544 if (TBB->pred_size() > 1 || FBB->pred_size() > 1) 1545 return false; 1546 1547 // Find a suitable position to hoist the common instructions to. Also figure 1548 // out which registers are used or defined by instructions from the insertion 1549 // point to the end of the block. 1550 SmallSet<unsigned, 4> Uses, Defs; 1551 MachineBasicBlock::iterator Loc = 1552 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs); 1553 if (Loc == MBB->end()) 1554 return false; 1555 1556 bool HasDups = false; 1557 SmallVector<unsigned, 4> LocalDefs; 1558 SmallSet<unsigned, 4> LocalDefsSet; 1559 MachineBasicBlock::iterator TIB = TBB->begin(); 1560 MachineBasicBlock::iterator FIB = FBB->begin(); 1561 MachineBasicBlock::iterator TIE = TBB->end(); 1562 MachineBasicBlock::iterator FIE = FBB->end(); 1563 while (TIB != TIE && FIB != FIE) { 1564 // Skip dbg_value instructions. These do not count. 1565 if (TIB->isDebugValue()) { 1566 while (TIB != TIE && TIB->isDebugValue()) 1567 ++TIB; 1568 if (TIB == TIE) 1569 break; 1570 } 1571 if (FIB->isDebugValue()) { 1572 while (FIB != FIE && FIB->isDebugValue()) 1573 ++FIB; 1574 if (FIB == FIE) 1575 break; 1576 } 1577 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead)) 1578 break; 1579 1580 if (TII->isPredicated(TIB)) 1581 // Hard to reason about register liveness with predicated instruction. 1582 break; 1583 1584 bool IsSafe = true; 1585 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1586 MachineOperand &MO = TIB->getOperand(i); 1587 if (!MO.isReg()) 1588 continue; 1589 unsigned Reg = MO.getReg(); 1590 if (!Reg) 1591 continue; 1592 if (MO.isDef()) { 1593 if (Uses.count(Reg)) { 1594 // Avoid clobbering a register that's used by the instruction at 1595 // the point of insertion. 1596 IsSafe = false; 1597 break; 1598 } 1599 1600 if (Defs.count(Reg) && !MO.isDead()) { 1601 // Don't hoist the instruction if the def would be clobber by the 1602 // instruction at the point insertion. FIXME: This is overly 1603 // conservative. It should be possible to hoist the instructions 1604 // in BB2 in the following example: 1605 // BB1: 1606 // r1, eflag = op1 r2, r3 1607 // brcc eflag 1608 // 1609 // BB2: 1610 // r1 = op2, ... 1611 // = op3, r1<kill> 1612 IsSafe = false; 1613 break; 1614 } 1615 } else if (!LocalDefsSet.count(Reg)) { 1616 if (Defs.count(Reg)) { 1617 // Use is defined by the instruction at the point of insertion. 1618 IsSafe = false; 1619 break; 1620 } 1621 1622 if (MO.isKill() && Uses.count(Reg)) 1623 // Kills a register that's read by the instruction at the point of 1624 // insertion. Remove the kill marker. 1625 MO.setIsKill(false); 1626 } 1627 } 1628 if (!IsSafe) 1629 break; 1630 1631 bool DontMoveAcrossStore = true; 1632 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore)) 1633 break; 1634 1635 // Remove kills from LocalDefsSet, these registers had short live ranges. 1636 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1637 MachineOperand &MO = TIB->getOperand(i); 1638 if (!MO.isReg() || !MO.isUse() || !MO.isKill()) 1639 continue; 1640 unsigned Reg = MO.getReg(); 1641 if (!Reg || !LocalDefsSet.count(Reg)) 1642 continue; 1643 for (const unsigned *OR = TRI->getOverlaps(Reg); *OR; ++OR) 1644 LocalDefsSet.erase(*OR); 1645 } 1646 1647 // Track local defs so we can update liveins. 1648 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1649 MachineOperand &MO = TIB->getOperand(i); 1650 if (!MO.isReg() || !MO.isDef() || MO.isDead()) 1651 continue; 1652 unsigned Reg = MO.getReg(); 1653 if (!Reg) 1654 continue; 1655 LocalDefs.push_back(Reg); 1656 for (const unsigned *OR = TRI->getOverlaps(Reg); *OR; ++OR) 1657 LocalDefsSet.insert(*OR); 1658 } 1659 1660 HasDups = true;; 1661 ++TIB; 1662 ++FIB; 1663 } 1664 1665 if (!HasDups) 1666 return false; 1667 1668 MBB->splice(Loc, TBB, TBB->begin(), TIB); 1669 FBB->erase(FBB->begin(), FIB); 1670 1671 // Update livein's. 1672 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) { 1673 unsigned Def = LocalDefs[i]; 1674 if (LocalDefsSet.count(Def)) { 1675 TBB->addLiveIn(Def); 1676 FBB->addLiveIn(Def); 1677 } 1678 } 1679 1680 ++NumHoist; 1681 return true; 1682} 1683