BranchFolding.cpp revision e837dead3c8dc3445ef6a0e2322179c57e264a13
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} 368 369/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 370/// after it, replacing it with an unconditional branch to NewDest. 371void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 372 MachineBasicBlock *NewDest) { 373 TII->ReplaceTailWithBranchTo(OldInst, NewDest); 374 ++NumTailMerge; 375} 376 377/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 378/// MBB so that the part before the iterator falls into the part starting at the 379/// iterator. This returns the new MBB. 380MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 381 MachineBasicBlock::iterator BBI1) { 382 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1)) 383 return 0; 384 385 MachineFunction &MF = *CurMBB.getParent(); 386 387 // Create the fall-through block. 388 MachineFunction::iterator MBBI = &CurMBB; 389 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 390 CurMBB.getParent()->insert(++MBBI, NewMBB); 391 392 // Move all the successors of this block to the specified block. 393 NewMBB->transferSuccessors(&CurMBB); 394 395 // Add an edge from CurMBB to NewMBB for the fall-through. 396 CurMBB.addSuccessor(NewMBB); 397 398 // Splice the code over. 399 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 400 401 // For targets that use the register scavenger, we must maintain LiveIns. 402 if (RS) { 403 RS->enterBasicBlock(&CurMBB); 404 if (!CurMBB.empty()) 405 RS->forward(prior(CurMBB.end())); 406 BitVector RegsLiveAtExit(TRI->getNumRegs()); 407 RS->getRegsUsed(RegsLiveAtExit, false); 408 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 409 if (RegsLiveAtExit[i]) 410 NewMBB->addLiveIn(i); 411 } 412 413 return NewMBB; 414} 415 416/// EstimateRuntime - Make a rough estimate for how long it will take to run 417/// the specified code. 418static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 419 MachineBasicBlock::iterator E) { 420 unsigned Time = 0; 421 for (; I != E; ++I) { 422 if (I->isDebugValue()) 423 continue; 424 const MCInstrDesc &MCID = I->getDesc(); 425 if (MCID.isCall()) 426 Time += 10; 427 else if (MCID.mayLoad() || MCID.mayStore()) 428 Time += 2; 429 else 430 ++Time; 431 } 432 return Time; 433} 434 435// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 436// branches temporarily for tail merging). In the case where CurMBB ends 437// with a conditional branch to the next block, optimize by reversing the 438// test and conditionally branching to SuccMBB instead. 439static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 440 const TargetInstrInfo *TII) { 441 MachineFunction *MF = CurMBB->getParent(); 442 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB)); 443 MachineBasicBlock *TBB = 0, *FBB = 0; 444 SmallVector<MachineOperand, 4> Cond; 445 DebugLoc dl; // FIXME: this is nowhere 446 if (I != MF->end() && 447 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 448 MachineBasicBlock *NextBB = I; 449 if (TBB == NextBB && !Cond.empty() && !FBB) { 450 if (!TII->ReverseBranchCondition(Cond)) { 451 TII->RemoveBranch(*CurMBB); 452 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl); 453 return; 454 } 455 } 456 } 457 TII->InsertBranch(*CurMBB, SuccBB, NULL, 458 SmallVector<MachineOperand, 0>(), dl); 459} 460 461bool 462BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 463 if (getHash() < o.getHash()) 464 return true; 465 else if (getHash() > o.getHash()) 466 return false; 467 else if (getBlock()->getNumber() < o.getBlock()->getNumber()) 468 return true; 469 else if (getBlock()->getNumber() > o.getBlock()->getNumber()) 470 return false; 471 else { 472 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 473 // an object with itself. 474#ifndef _GLIBCXX_DEBUG 475 llvm_unreachable("Predecessor appears twice"); 476#endif 477 return false; 478 } 479} 480 481/// CountTerminators - Count the number of terminators in the given 482/// block and set I to the position of the first non-terminator, if there 483/// is one, or MBB->end() otherwise. 484static unsigned CountTerminators(MachineBasicBlock *MBB, 485 MachineBasicBlock::iterator &I) { 486 I = MBB->end(); 487 unsigned NumTerms = 0; 488 for (;;) { 489 if (I == MBB->begin()) { 490 I = MBB->end(); 491 break; 492 } 493 --I; 494 if (!I->getDesc().isTerminator()) break; 495 ++NumTerms; 496 } 497 return NumTerms; 498} 499 500/// ProfitableToMerge - Check if two machine basic blocks have a common tail 501/// and decide if it would be profitable to merge those tails. Return the 502/// length of the common tail and iterators to the first common instruction 503/// in each block. 504static bool ProfitableToMerge(MachineBasicBlock *MBB1, 505 MachineBasicBlock *MBB2, 506 unsigned minCommonTailLength, 507 unsigned &CommonTailLen, 508 MachineBasicBlock::iterator &I1, 509 MachineBasicBlock::iterator &I2, 510 MachineBasicBlock *SuccBB, 511 MachineBasicBlock *PredBB) { 512 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 513 if (CommonTailLen == 0) 514 return false; 515 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber() 516 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen 517 << '\n'); 518 519 // It's almost always profitable to merge any number of non-terminator 520 // instructions with the block that falls through into the common successor. 521 if (MBB1 == PredBB || MBB2 == PredBB) { 522 MachineBasicBlock::iterator I; 523 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 524 if (CommonTailLen > NumTerms) 525 return true; 526 } 527 528 // If one of the blocks can be completely merged and happens to be in 529 // a position where the other could fall through into it, merge any number 530 // of instructions, because it can be done without a branch. 531 // TODO: If the blocks are not adjacent, move one of them so that they are? 532 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 533 return true; 534 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 535 return true; 536 537 // If both blocks have an unconditional branch temporarily stripped out, 538 // count that as an additional common instruction for the following 539 // heuristics. 540 unsigned EffectiveTailLen = CommonTailLen; 541 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 542 !MBB1->back().getDesc().isBarrier() && 543 !MBB2->back().getDesc().isBarrier()) 544 ++EffectiveTailLen; 545 546 // Check if the common tail is long enough to be worthwhile. 547 if (EffectiveTailLen >= minCommonTailLength) 548 return true; 549 550 // If we are optimizing for code size, 2 instructions in common is enough if 551 // we don't have to split a block. At worst we will be introducing 1 new 552 // branch instruction, which is likely to be smaller than the 2 553 // instructions that would be deleted in the merge. 554 MachineFunction *MF = MBB1->getParent(); 555 if (EffectiveTailLen >= 2 && 556 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 557 (I1 == MBB1->begin() || I2 == MBB2->begin())) 558 return true; 559 560 return false; 561} 562 563/// ComputeSameTails - Look through all the blocks in MergePotentials that have 564/// hash CurHash (guaranteed to match the last element). Build the vector 565/// SameTails of all those that have the (same) largest number of instructions 566/// in common of any pair of these blocks. SameTails entries contain an 567/// iterator into MergePotentials (from which the MachineBasicBlock can be 568/// found) and a MachineBasicBlock::iterator into that MBB indicating the 569/// instruction where the matching code sequence begins. 570/// Order of elements in SameTails is the reverse of the order in which 571/// those blocks appear in MergePotentials (where they are not necessarily 572/// consecutive). 573unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 574 unsigned minCommonTailLength, 575 MachineBasicBlock *SuccBB, 576 MachineBasicBlock *PredBB) { 577 unsigned maxCommonTailLength = 0U; 578 SameTails.clear(); 579 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 580 MPIterator HighestMPIter = prior(MergePotentials.end()); 581 for (MPIterator CurMPIter = prior(MergePotentials.end()), 582 B = MergePotentials.begin(); 583 CurMPIter != B && CurMPIter->getHash() == CurHash; 584 --CurMPIter) { 585 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 586 unsigned CommonTailLen; 587 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 588 minCommonTailLength, 589 CommonTailLen, TrialBBI1, TrialBBI2, 590 SuccBB, PredBB)) { 591 if (CommonTailLen > maxCommonTailLength) { 592 SameTails.clear(); 593 maxCommonTailLength = CommonTailLen; 594 HighestMPIter = CurMPIter; 595 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 596 } 597 if (HighestMPIter == CurMPIter && 598 CommonTailLen == maxCommonTailLength) 599 SameTails.push_back(SameTailElt(I, TrialBBI2)); 600 } 601 if (I == B) 602 break; 603 } 604 } 605 return maxCommonTailLength; 606} 607 608/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 609/// MergePotentials, restoring branches at ends of blocks as appropriate. 610void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 611 MachineBasicBlock *SuccBB, 612 MachineBasicBlock *PredBB) { 613 MPIterator CurMPIter, B; 614 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 615 CurMPIter->getHash() == CurHash; 616 --CurMPIter) { 617 // Put the unconditional branch back, if we need one. 618 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 619 if (SuccBB && CurMBB != PredBB) 620 FixTail(CurMBB, SuccBB, TII); 621 if (CurMPIter == B) 622 break; 623 } 624 if (CurMPIter->getHash() != CurHash) 625 CurMPIter++; 626 MergePotentials.erase(CurMPIter, MergePotentials.end()); 627} 628 629/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 630/// only of the common tail. Create a block that does by splitting one. 631bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 632 unsigned maxCommonTailLength, 633 unsigned &commonTailIndex) { 634 commonTailIndex = 0; 635 unsigned TimeEstimate = ~0U; 636 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 637 // Use PredBB if possible; that doesn't require a new branch. 638 if (SameTails[i].getBlock() == PredBB) { 639 commonTailIndex = i; 640 break; 641 } 642 // Otherwise, make a (fairly bogus) choice based on estimate of 643 // how long it will take the various blocks to execute. 644 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 645 SameTails[i].getTailStartPos()); 646 if (t <= TimeEstimate) { 647 TimeEstimate = t; 648 commonTailIndex = i; 649 } 650 } 651 652 MachineBasicBlock::iterator BBI = 653 SameTails[commonTailIndex].getTailStartPos(); 654 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 655 656 // If the common tail includes any debug info we will take it pretty 657 // randomly from one of the inputs. Might be better to remove it? 658 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 659 << maxCommonTailLength); 660 661 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 662 if (!newMBB) { 663 DEBUG(dbgs() << "... failed!"); 664 return false; 665 } 666 667 SameTails[commonTailIndex].setBlock(newMBB); 668 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 669 670 // If we split PredBB, newMBB is the new predecessor. 671 if (PredBB == MBB) 672 PredBB = newMBB; 673 674 return true; 675} 676 677// See if any of the blocks in MergePotentials (which all have a common single 678// successor, or all have no successor) can be tail-merged. If there is a 679// successor, any blocks in MergePotentials that are not tail-merged and 680// are not immediately before Succ must have an unconditional branch to 681// Succ added (but the predecessor/successor lists need no adjustment). 682// The lone predecessor of Succ that falls through into Succ, 683// if any, is given in PredBB. 684 685bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 686 MachineBasicBlock *PredBB) { 687 bool MadeChange = false; 688 689 // Except for the special cases below, tail-merge if there are at least 690 // this many instructions in common. 691 unsigned minCommonTailLength = TailMergeSize; 692 693 DEBUG(dbgs() << "\nTryTailMergeBlocks: "; 694 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 695 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 696 << (i == e-1 ? "" : ", "); 697 dbgs() << "\n"; 698 if (SuccBB) { 699 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 700 if (PredBB) 701 dbgs() << " which has fall-through from BB#" 702 << PredBB->getNumber() << "\n"; 703 } 704 dbgs() << "Looking for common tails of at least " 705 << minCommonTailLength << " instruction" 706 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 707 ); 708 709 // Sort by hash value so that blocks with identical end sequences sort 710 // together. 711 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 712 713 // Walk through equivalence sets looking for actual exact matches. 714 while (MergePotentials.size() > 1) { 715 unsigned CurHash = MergePotentials.back().getHash(); 716 717 // Build SameTails, identifying the set of blocks with this hash code 718 // and with the maximum number of instructions in common. 719 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 720 minCommonTailLength, 721 SuccBB, PredBB); 722 723 // If we didn't find any pair that has at least minCommonTailLength 724 // instructions in common, remove all blocks with this hash code and retry. 725 if (SameTails.empty()) { 726 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 727 continue; 728 } 729 730 // If one of the blocks is the entire common tail (and not the entry 731 // block, which we can't jump to), we can treat all blocks with this same 732 // tail at once. Use PredBB if that is one of the possibilities, as that 733 // will not introduce any extra branches. 734 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 735 getParent()->begin(); 736 unsigned commonTailIndex = SameTails.size(); 737 // If there are two blocks, check to see if one can be made to fall through 738 // into the other. 739 if (SameTails.size() == 2 && 740 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 741 SameTails[1].tailIsWholeBlock()) 742 commonTailIndex = 1; 743 else if (SameTails.size() == 2 && 744 SameTails[1].getBlock()->isLayoutSuccessor( 745 SameTails[0].getBlock()) && 746 SameTails[0].tailIsWholeBlock()) 747 commonTailIndex = 0; 748 else { 749 // Otherwise just pick one, favoring the fall-through predecessor if 750 // there is one. 751 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 752 MachineBasicBlock *MBB = SameTails[i].getBlock(); 753 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 754 continue; 755 if (MBB == PredBB) { 756 commonTailIndex = i; 757 break; 758 } 759 if (SameTails[i].tailIsWholeBlock()) 760 commonTailIndex = i; 761 } 762 } 763 764 if (commonTailIndex == SameTails.size() || 765 (SameTails[commonTailIndex].getBlock() == PredBB && 766 !SameTails[commonTailIndex].tailIsWholeBlock())) { 767 // None of the blocks consist entirely of the common tail. 768 // Split a block so that one does. 769 if (!CreateCommonTailOnlyBlock(PredBB, 770 maxCommonTailLength, commonTailIndex)) { 771 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 772 continue; 773 } 774 } 775 776 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 777 // MBB is common tail. Adjust all other BB's to jump to this one. 778 // Traversal must be forwards so erases work. 779 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber() 780 << " for "); 781 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 782 if (commonTailIndex == i) 783 continue; 784 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() 785 << (i == e-1 ? "" : ", ")); 786 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 787 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 788 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 789 MergePotentials.erase(SameTails[i].getMPIter()); 790 } 791 DEBUG(dbgs() << "\n"); 792 // We leave commonTailIndex in the worklist in case there are other blocks 793 // that match it with a smaller number of instructions. 794 MadeChange = true; 795 } 796 return MadeChange; 797} 798 799bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 800 801 if (!EnableTailMerge) return false; 802 803 bool MadeChange = false; 804 805 // First find blocks with no successors. 806 MergePotentials.clear(); 807 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); 808 I != E && MergePotentials.size() < TailMergeThreshold; ++I) { 809 if (TriedMerging.count(I)) 810 continue; 811 if (I->succ_empty()) 812 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I)); 813 } 814 815 // If this is a large problem, avoid visiting the same basic blocks 816 // multiple times. 817 if (MergePotentials.size() == TailMergeThreshold) 818 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 819 TriedMerging.insert(MergePotentials[i].getBlock()); 820 // See if we can do any tail merging on those. 821 if (MergePotentials.size() >= 2) 822 MadeChange |= TryTailMergeBlocks(NULL, NULL); 823 824 // Look at blocks (IBB) with multiple predecessors (PBB). 825 // We change each predecessor to a canonical form, by 826 // (1) temporarily removing any unconditional branch from the predecessor 827 // to IBB, and 828 // (2) alter conditional branches so they branch to the other block 829 // not IBB; this may require adding back an unconditional branch to IBB 830 // later, where there wasn't one coming in. E.g. 831 // Bcc IBB 832 // fallthrough to QBB 833 // here becomes 834 // Bncc QBB 835 // with a conceptual B to IBB after that, which never actually exists. 836 // With those changes, we see whether the predecessors' tails match, 837 // and merge them if so. We change things out of canonical form and 838 // back to the way they were later in the process. (OptimizeBranches 839 // would undo some of this, but we can't use it, because we'd get into 840 // a compile-time infinite loop repeatedly doing and undoing the same 841 // transformations.) 842 843 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 844 I != E; ++I) { 845 if (I->pred_size() >= 2) { 846 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 847 MachineBasicBlock *IBB = I; 848 MachineBasicBlock *PredBB = prior(I); 849 MergePotentials.clear(); 850 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 851 E2 = I->pred_end(); 852 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) { 853 MachineBasicBlock *PBB = *P; 854 if (TriedMerging.count(PBB)) 855 continue; 856 // Skip blocks that loop to themselves, can't tail merge these. 857 if (PBB == IBB) 858 continue; 859 // Visit each predecessor only once. 860 if (!UniquePreds.insert(PBB)) 861 continue; 862 MachineBasicBlock *TBB = 0, *FBB = 0; 863 SmallVector<MachineOperand, 4> Cond; 864 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 865 // Failing case: IBB is the target of a cbr, and 866 // we cannot reverse the branch. 867 SmallVector<MachineOperand, 4> NewCond(Cond); 868 if (!Cond.empty() && TBB == IBB) { 869 if (TII->ReverseBranchCondition(NewCond)) 870 continue; 871 // This is the QBB case described above 872 if (!FBB) 873 FBB = llvm::next(MachineFunction::iterator(PBB)); 874 } 875 // Failing case: the only way IBB can be reached from PBB is via 876 // exception handling. Happens for landing pads. Would be nice 877 // to have a bit in the edge so we didn't have to do all this. 878 if (IBB->isLandingPad()) { 879 MachineFunction::iterator IP = PBB; IP++; 880 MachineBasicBlock *PredNextBB = NULL; 881 if (IP != MF.end()) 882 PredNextBB = IP; 883 if (TBB == NULL) { 884 if (IBB != PredNextBB) // fallthrough 885 continue; 886 } else if (FBB) { 887 if (TBB != IBB && FBB != IBB) // cbr then ubr 888 continue; 889 } else if (Cond.empty()) { 890 if (TBB != IBB) // ubr 891 continue; 892 } else { 893 if (TBB != IBB && IBB != PredNextBB) // cbr 894 continue; 895 } 896 } 897 // Remove the unconditional branch at the end, if any. 898 if (TBB && (Cond.empty() || FBB)) { 899 DebugLoc dl; // FIXME: this is nowhere 900 TII->RemoveBranch(*PBB); 901 if (!Cond.empty()) 902 // reinsert conditional branch only, for now 903 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl); 904 } 905 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P)); 906 } 907 } 908 // If this is a large problem, avoid visiting the same basic blocks 909 // multiple times. 910 if (MergePotentials.size() == TailMergeThreshold) 911 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 912 TriedMerging.insert(MergePotentials[i].getBlock()); 913 if (MergePotentials.size() >= 2) 914 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 915 // Reinsert an unconditional branch if needed. 916 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks. 917 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 918 if (MergePotentials.size() == 1 && 919 MergePotentials.begin()->getBlock() != PredBB) 920 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 921 } 922 } 923 return MadeChange; 924} 925 926//===----------------------------------------------------------------------===// 927// Branch Optimization 928//===----------------------------------------------------------------------===// 929 930bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 931 bool MadeChange = false; 932 933 // Make sure blocks are numbered in order 934 MF.RenumberBlocks(); 935 936 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 937 I != E; ) { 938 MachineBasicBlock *MBB = I++; 939 MadeChange |= OptimizeBlock(MBB); 940 941 // If it is dead, remove it. 942 if (MBB->pred_empty()) { 943 RemoveDeadBlock(MBB); 944 MadeChange = true; 945 ++NumDeadBlocks; 946 } 947 } 948 return MadeChange; 949} 950 951// Blocks should be considered empty if they contain only debug info; 952// else the debug info would affect codegen. 953static bool IsEmptyBlock(MachineBasicBlock *MBB) { 954 if (MBB->empty()) 955 return true; 956 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end(); 957 MBBI!=MBBE; ++MBBI) { 958 if (!MBBI->isDebugValue()) 959 return false; 960 } 961 return true; 962} 963 964// Blocks with only debug info and branches should be considered the same 965// as blocks with only branches. 966static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { 967 MachineBasicBlock::iterator MBBI, MBBE; 968 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) { 969 if (!MBBI->isDebugValue()) 970 break; 971 } 972 return (MBBI->getDesc().isBranch()); 973} 974 975/// IsBetterFallthrough - Return true if it would be clearly better to 976/// fall-through to MBB1 than to fall through into MBB2. This has to return 977/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 978/// result in infinite loops. 979static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 980 MachineBasicBlock *MBB2) { 981 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 982 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 983 // optimize branches that branch to either a return block or an assert block 984 // into a fallthrough to the return. 985 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false; 986 987 // If there is a clear successor ordering we make sure that one block 988 // will fall through to the next 989 if (MBB1->isSuccessor(MBB2)) return true; 990 if (MBB2->isSuccessor(MBB1)) return false; 991 992 // Neither block consists entirely of debug info (per IsEmptyBlock check), 993 // so we needn't test for falling off the beginning here. 994 MachineBasicBlock::iterator MBB1I = --MBB1->end(); 995 while (MBB1I->isDebugValue()) 996 --MBB1I; 997 MachineBasicBlock::iterator MBB2I = --MBB2->end(); 998 while (MBB2I->isDebugValue()) 999 --MBB2I; 1000 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 1001} 1002 1003/// OptimizeBlock - Analyze and optimize control flow related to the specified 1004/// block. This is never called on the entry block. 1005bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1006 bool MadeChange = false; 1007 MachineFunction &MF = *MBB->getParent(); 1008 DebugLoc dl; // FIXME: this is nowhere 1009ReoptimizeBlock: 1010 1011 MachineFunction::iterator FallThrough = MBB; 1012 ++FallThrough; 1013 1014 // If this block is empty, make everyone use its fall-through, not the block 1015 // explicitly. Landing pads should not do this since the landing-pad table 1016 // points to this block. Blocks with their addresses taken shouldn't be 1017 // optimized away. 1018 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 1019 // Dead block? Leave for cleanup later. 1020 if (MBB->pred_empty()) return MadeChange; 1021 1022 if (FallThrough == MF.end()) { 1023 // TODO: Simplify preds to not branch here if possible! 1024 } else { 1025 // Rewrite all predecessors of the old block to go to the fallthrough 1026 // instead. 1027 while (!MBB->pred_empty()) { 1028 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1029 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1030 } 1031 // If MBB was the target of a jump table, update jump tables to go to the 1032 // fallthrough instead. 1033 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1034 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough); 1035 MadeChange = true; 1036 } 1037 return MadeChange; 1038 } 1039 1040 // Check to see if we can simplify the terminator of the block before this 1041 // one. 1042 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1043 1044 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1045 SmallVector<MachineOperand, 4> PriorCond; 1046 bool PriorUnAnalyzable = 1047 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1048 if (!PriorUnAnalyzable) { 1049 // If the CFG for the prior block has extra edges, remove them. 1050 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1051 !PriorCond.empty()); 1052 1053 // If the previous branch is conditional and both conditions go to the same 1054 // destination, remove the branch, replacing it with an unconditional one or 1055 // a fall-through. 1056 if (PriorTBB && PriorTBB == PriorFBB) { 1057 TII->RemoveBranch(PrevBB); 1058 PriorCond.clear(); 1059 if (PriorTBB != MBB) 1060 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1061 MadeChange = true; 1062 ++NumBranchOpts; 1063 goto ReoptimizeBlock; 1064 } 1065 1066 // If the previous block unconditionally falls through to this block and 1067 // this block has no other predecessors, move the contents of this block 1068 // into the prior block. This doesn't usually happen when SimplifyCFG 1069 // has been used, but it can happen if tail merging splits a fall-through 1070 // predecessor of a block. 1071 // This has to check PrevBB->succ_size() because EH edges are ignored by 1072 // AnalyzeBranch. 1073 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1074 PrevBB.succ_size() == 1 && 1075 !MBB->hasAddressTaken() && !MBB->isLandingPad()) { 1076 DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1077 << "From MBB: " << *MBB); 1078 // Remove redundant DBG_VALUEs first. 1079 if (PrevBB.begin() != PrevBB.end()) { 1080 MachineBasicBlock::iterator PrevBBIter = PrevBB.end(); 1081 --PrevBBIter; 1082 MachineBasicBlock::iterator MBBIter = MBB->begin(); 1083 // Check if DBG_VALUE at the end of PrevBB is identical to the 1084 // DBG_VALUE at the beginning of MBB. 1085 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end() 1086 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) { 1087 if (!MBBIter->isIdenticalTo(PrevBBIter)) 1088 break; 1089 MachineInstr *DuplicateDbg = MBBIter; 1090 ++MBBIter; -- PrevBBIter; 1091 DuplicateDbg->eraseFromParent(); 1092 } 1093 } 1094 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1095 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1096 assert(PrevBB.succ_empty()); 1097 PrevBB.transferSuccessors(MBB); 1098 MadeChange = true; 1099 return MadeChange; 1100 } 1101 1102 // If the previous branch *only* branches to *this* block (conditional or 1103 // not) remove the branch. 1104 if (PriorTBB == MBB && PriorFBB == 0) { 1105 TII->RemoveBranch(PrevBB); 1106 MadeChange = true; 1107 ++NumBranchOpts; 1108 goto ReoptimizeBlock; 1109 } 1110 1111 // If the prior block branches somewhere else on the condition and here if 1112 // the condition is false, remove the uncond second branch. 1113 if (PriorFBB == MBB) { 1114 TII->RemoveBranch(PrevBB); 1115 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1116 MadeChange = true; 1117 ++NumBranchOpts; 1118 goto ReoptimizeBlock; 1119 } 1120 1121 // If the prior block branches here on true and somewhere else on false, and 1122 // if the branch condition is reversible, reverse the branch to create a 1123 // fall-through. 1124 if (PriorTBB == MBB) { 1125 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1126 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1127 TII->RemoveBranch(PrevBB); 1128 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl); 1129 MadeChange = true; 1130 ++NumBranchOpts; 1131 goto ReoptimizeBlock; 1132 } 1133 } 1134 1135 // If this block has no successors (e.g. it is a return block or ends with 1136 // a call to a no-return function like abort or __cxa_throw) and if the pred 1137 // falls through into this block, and if it would otherwise fall through 1138 // into the block after this, move this block to the end of the function. 1139 // 1140 // We consider it more likely that execution will stay in the function (e.g. 1141 // due to loops) than it is to exit it. This asserts in loops etc, moving 1142 // the assert condition out of the loop body. 1143 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1144 MachineFunction::iterator(PriorTBB) == FallThrough && 1145 !MBB->canFallThrough()) { 1146 bool DoTransform = true; 1147 1148 // We have to be careful that the succs of PredBB aren't both no-successor 1149 // blocks. If neither have successors and if PredBB is the second from 1150 // last block in the function, we'd just keep swapping the two blocks for 1151 // last. Only do the swap if one is clearly better to fall through than 1152 // the other. 1153 if (FallThrough == --MF.end() && 1154 !IsBetterFallthrough(PriorTBB, MBB)) 1155 DoTransform = false; 1156 1157 if (DoTransform) { 1158 // Reverse the branch so we will fall through on the previous true cond. 1159 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1160 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1161 DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1162 << "To make fallthrough to: " << *PriorTBB << "\n"); 1163 1164 TII->RemoveBranch(PrevBB); 1165 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl); 1166 1167 // Move this block to the end of the function. 1168 MBB->moveAfter(--MF.end()); 1169 MadeChange = true; 1170 ++NumBranchOpts; 1171 return MadeChange; 1172 } 1173 } 1174 } 1175 } 1176 1177 // Analyze the branch in the current block. 1178 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1179 SmallVector<MachineOperand, 4> CurCond; 1180 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1181 if (!CurUnAnalyzable) { 1182 // If the CFG for the prior block has extra edges, remove them. 1183 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1184 1185 // If this is a two-way branch, and the FBB branches to this block, reverse 1186 // the condition so the single-basic-block loop is faster. Instead of: 1187 // Loop: xxx; jcc Out; jmp Loop 1188 // we want: 1189 // Loop: xxx; jncc Loop; jmp Out 1190 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1191 SmallVector<MachineOperand, 4> NewCond(CurCond); 1192 if (!TII->ReverseBranchCondition(NewCond)) { 1193 TII->RemoveBranch(*MBB); 1194 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); 1195 MadeChange = true; 1196 ++NumBranchOpts; 1197 goto ReoptimizeBlock; 1198 } 1199 } 1200 1201 // If this branch is the only thing in its block, see if we can forward 1202 // other blocks across it. 1203 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1204 IsBranchOnlyBlock(MBB) && CurTBB != MBB && 1205 !MBB->hasAddressTaken()) { 1206 // This block may contain just an unconditional branch. Because there can 1207 // be 'non-branch terminators' in the block, try removing the branch and 1208 // then seeing if the block is empty. 1209 TII->RemoveBranch(*MBB); 1210 // If the only things remaining in the block are debug info, remove these 1211 // as well, so this will behave the same as an empty block in non-debug 1212 // mode. 1213 if (!MBB->empty()) { 1214 bool NonDebugInfoFound = false; 1215 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 1216 I != E; ++I) { 1217 if (!I->isDebugValue()) { 1218 NonDebugInfoFound = true; 1219 break; 1220 } 1221 } 1222 if (!NonDebugInfoFound) 1223 // Make the block empty, losing the debug info (we could probably 1224 // improve this in some cases.) 1225 MBB->erase(MBB->begin(), MBB->end()); 1226 } 1227 // If this block is just an unconditional branch to CurTBB, we can 1228 // usually completely eliminate the block. The only case we cannot 1229 // completely eliminate the block is when the block before this one 1230 // falls through into MBB and we can't understand the prior block's branch 1231 // condition. 1232 if (MBB->empty()) { 1233 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1234 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1235 !PrevBB.isSuccessor(MBB)) { 1236 // If the prior block falls through into us, turn it into an 1237 // explicit branch to us to make updates simpler. 1238 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1239 PriorTBB != MBB && PriorFBB != MBB) { 1240 if (PriorTBB == 0) { 1241 assert(PriorCond.empty() && PriorFBB == 0 && 1242 "Bad branch analysis"); 1243 PriorTBB = MBB; 1244 } else { 1245 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1246 PriorFBB = MBB; 1247 } 1248 TII->RemoveBranch(PrevBB); 1249 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, dl); 1250 } 1251 1252 // Iterate through all the predecessors, revectoring each in-turn. 1253 size_t PI = 0; 1254 bool DidChange = false; 1255 bool HasBranchToSelf = false; 1256 while(PI != MBB->pred_size()) { 1257 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1258 if (PMBB == MBB) { 1259 // If this block has an uncond branch to itself, leave it. 1260 ++PI; 1261 HasBranchToSelf = true; 1262 } else { 1263 DidChange = true; 1264 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1265 // If this change resulted in PMBB ending in a conditional 1266 // branch where both conditions go to the same destination, 1267 // change this to an unconditional branch (and fix the CFG). 1268 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1269 SmallVector<MachineOperand, 4> NewCurCond; 1270 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1271 NewCurFBB, NewCurCond, true); 1272 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1273 TII->RemoveBranch(*PMBB); 1274 NewCurCond.clear(); 1275 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, dl); 1276 MadeChange = true; 1277 ++NumBranchOpts; 1278 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1279 } 1280 } 1281 } 1282 1283 // Change any jumptables to go to the new MBB. 1284 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1285 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); 1286 if (DidChange) { 1287 ++NumBranchOpts; 1288 MadeChange = true; 1289 if (!HasBranchToSelf) return MadeChange; 1290 } 1291 } 1292 } 1293 1294 // Add the branch back if the block is more than just an uncond branch. 1295 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl); 1296 } 1297 } 1298 1299 // If the prior block doesn't fall through into this block, and if this 1300 // block doesn't fall through into some other block, see if we can find a 1301 // place to move this block where a fall-through will happen. 1302 if (!PrevBB.canFallThrough()) { 1303 1304 // Now we know that there was no fall-through into this block, check to 1305 // see if it has a fall-through into its successor. 1306 bool CurFallsThru = MBB->canFallThrough(); 1307 1308 if (!MBB->isLandingPad()) { 1309 // Check all the predecessors of this block. If one of them has no fall 1310 // throughs, move this block right after it. 1311 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1312 E = MBB->pred_end(); PI != E; ++PI) { 1313 // Analyze the branch at the end of the pred. 1314 MachineBasicBlock *PredBB = *PI; 1315 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1316 MachineBasicBlock *PredTBB = 0, *PredFBB = 0; 1317 SmallVector<MachineOperand, 4> PredCond; 1318 if (PredBB != MBB && !PredBB->canFallThrough() && 1319 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1320 && (!CurFallsThru || !CurTBB || !CurFBB) 1321 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1322 // If the current block doesn't fall through, just move it. 1323 // If the current block can fall through and does not end with a 1324 // conditional branch, we need to append an unconditional jump to 1325 // the (current) next block. To avoid a possible compile-time 1326 // infinite loop, move blocks only backward in this case. 1327 // Also, if there are already 2 branches here, we cannot add a third; 1328 // this means we have the case 1329 // Bcc next 1330 // B elsewhere 1331 // next: 1332 if (CurFallsThru) { 1333 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); 1334 CurCond.clear(); 1335 TII->InsertBranch(*MBB, NextBB, 0, CurCond, dl); 1336 } 1337 MBB->moveAfter(PredBB); 1338 MadeChange = true; 1339 goto ReoptimizeBlock; 1340 } 1341 } 1342 } 1343 1344 if (!CurFallsThru) { 1345 // Check all successors to see if we can move this block before it. 1346 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1347 E = MBB->succ_end(); SI != E; ++SI) { 1348 // Analyze the branch at the end of the block before the succ. 1349 MachineBasicBlock *SuccBB = *SI; 1350 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1351 1352 // If this block doesn't already fall-through to that successor, and if 1353 // the succ doesn't already have a block that can fall through into it, 1354 // and if the successor isn't an EH destination, we can arrange for the 1355 // fallthrough to happen. 1356 if (SuccBB != MBB && &*SuccPrev != MBB && 1357 !SuccPrev->canFallThrough() && !CurUnAnalyzable && 1358 !SuccBB->isLandingPad()) { 1359 MBB->moveBefore(SuccBB); 1360 MadeChange = true; 1361 goto ReoptimizeBlock; 1362 } 1363 } 1364 1365 // Okay, there is no really great place to put this block. If, however, 1366 // the block before this one would be a fall-through if this block were 1367 // removed, move this block to the end of the function. 1368 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0; 1369 SmallVector<MachineOperand, 4> PrevCond; 1370 if (FallThrough != MF.end() && 1371 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1372 PrevBB.isSuccessor(FallThrough)) { 1373 MBB->moveAfter(--MF.end()); 1374 MadeChange = true; 1375 return MadeChange; 1376 } 1377 } 1378 } 1379 1380 return MadeChange; 1381} 1382 1383//===----------------------------------------------------------------------===// 1384// Hoist Common Code 1385//===----------------------------------------------------------------------===// 1386 1387/// HoistCommonCode - Hoist common instruction sequences at the start of basic 1388/// blocks to their common predecessor. 1389bool BranchFolder::HoistCommonCode(MachineFunction &MF) { 1390 bool MadeChange = false; 1391 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) { 1392 MachineBasicBlock *MBB = I++; 1393 MadeChange |= HoistCommonCodeInSuccs(MBB); 1394 } 1395 1396 return MadeChange; 1397} 1398 1399/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 1400/// its 'true' successor. 1401static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 1402 MachineBasicBlock *TrueBB) { 1403 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 1404 E = BB->succ_end(); SI != E; ++SI) { 1405 MachineBasicBlock *SuccBB = *SI; 1406 if (SuccBB != TrueBB) 1407 return SuccBB; 1408 } 1409 return NULL; 1410} 1411 1412/// findHoistingInsertPosAndDeps - Find the location to move common instructions 1413/// in successors to. The location is ususally just before the terminator, 1414/// however if the terminator is a conditional branch and its previous 1415/// instruction is the flag setting instruction, the previous instruction is 1416/// the preferred location. This function also gathers uses and defs of the 1417/// instructions from the insertion point to the end of the block. The data is 1418/// used by HoistCommonCodeInSuccs to ensure safety. 1419static 1420MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB, 1421 const TargetInstrInfo *TII, 1422 const TargetRegisterInfo *TRI, 1423 SmallSet<unsigned,4> &Uses, 1424 SmallSet<unsigned,4> &Defs) { 1425 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator(); 1426 if (!TII->isUnpredicatedTerminator(Loc)) 1427 return MBB->end(); 1428 1429 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) { 1430 const MachineOperand &MO = Loc->getOperand(i); 1431 if (!MO.isReg()) 1432 continue; 1433 unsigned Reg = MO.getReg(); 1434 if (!Reg) 1435 continue; 1436 if (MO.isUse()) { 1437 Uses.insert(Reg); 1438 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1439 Uses.insert(*AS); 1440 } else if (!MO.isDead()) 1441 // Don't try to hoist code in the rare case the terminator defines a 1442 // register that is later used. 1443 return MBB->end(); 1444 } 1445 1446 if (Uses.empty()) 1447 return Loc; 1448 if (Loc == MBB->begin()) 1449 return MBB->end(); 1450 1451 // The terminator is probably a conditional branch, try not to separate the 1452 // branch from condition setting instruction. 1453 MachineBasicBlock::iterator PI = Loc; 1454 --PI; 1455 while (PI != MBB->begin() && Loc->isDebugValue()) 1456 --PI; 1457 1458 bool IsDef = false; 1459 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) { 1460 const MachineOperand &MO = PI->getOperand(i); 1461 if (!MO.isReg() || MO.isUse()) 1462 continue; 1463 unsigned Reg = MO.getReg(); 1464 if (!Reg) 1465 continue; 1466 if (Uses.count(Reg)) 1467 IsDef = true; 1468 } 1469 if (!IsDef) 1470 // The condition setting instruction is not just before the conditional 1471 // branch. 1472 return Loc; 1473 1474 // Be conservative, don't insert instruction above something that may have 1475 // side-effects. And since it's potentially bad to separate flag setting 1476 // instruction from the conditional branch, just abort the optimization 1477 // completely. 1478 // Also avoid moving code above predicated instruction since it's hard to 1479 // reason about register liveness with predicated instruction. 1480 bool DontMoveAcrossStore = true; 1481 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) || 1482 TII->isPredicated(PI)) 1483 return MBB->end(); 1484 1485 1486 // Find out what registers are live. Note this routine is ignoring other live 1487 // registers which are only used by instructions in successor blocks. 1488 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) { 1489 const MachineOperand &MO = PI->getOperand(i); 1490 if (!MO.isReg()) 1491 continue; 1492 unsigned Reg = MO.getReg(); 1493 if (!Reg) 1494 continue; 1495 if (MO.isUse()) { 1496 Uses.insert(Reg); 1497 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1498 Uses.insert(*AS); 1499 } else { 1500 if (Uses.count(Reg)) { 1501 Uses.erase(Reg); 1502 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1503 Uses.erase(*SR); // Use getSubRegisters to be conservative 1504 } 1505 Defs.insert(Reg); 1506 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1507 Defs.insert(*AS); 1508 } 1509 } 1510 1511 return PI; 1512} 1513 1514/// HoistCommonCodeInSuccs - If the successors of MBB has common instruction 1515/// sequence at the start of the function, move the instructions before MBB 1516/// terminator if it's legal. 1517bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) { 1518 MachineBasicBlock *TBB = 0, *FBB = 0; 1519 SmallVector<MachineOperand, 4> Cond; 1520 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty()) 1521 return false; 1522 1523 if (!FBB) FBB = findFalseBlock(MBB, TBB); 1524 if (!FBB) 1525 // Malformed bcc? True and false blocks are the same? 1526 return false; 1527 1528 // Restrict the optimization to cases where MBB is the only predecessor, 1529 // it is an obvious win. 1530 if (TBB->pred_size() > 1 || FBB->pred_size() > 1) 1531 return false; 1532 1533 // Find a suitable position to hoist the common instructions to. Also figure 1534 // out which registers are used or defined by instructions from the insertion 1535 // point to the end of the block. 1536 SmallSet<unsigned, 4> Uses, Defs; 1537 MachineBasicBlock::iterator Loc = 1538 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs); 1539 if (Loc == MBB->end()) 1540 return false; 1541 1542 bool HasDups = false; 1543 SmallVector<unsigned, 4> LocalDefs; 1544 SmallSet<unsigned, 4> LocalDefsSet; 1545 MachineBasicBlock::iterator TIB = TBB->begin(); 1546 MachineBasicBlock::iterator FIB = FBB->begin(); 1547 MachineBasicBlock::iterator TIE = TBB->end(); 1548 MachineBasicBlock::iterator FIE = FBB->end(); 1549 while (TIB != TIE && FIB != FIE) { 1550 // Skip dbg_value instructions. These do not count. 1551 if (TIB->isDebugValue()) { 1552 while (TIB != TIE && TIB->isDebugValue()) 1553 ++TIB; 1554 if (TIB == TIE) 1555 break; 1556 } 1557 if (FIB->isDebugValue()) { 1558 while (FIB != FIE && FIB->isDebugValue()) 1559 ++FIB; 1560 if (FIB == FIE) 1561 break; 1562 } 1563 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead)) 1564 break; 1565 1566 if (TII->isPredicated(TIB)) 1567 // Hard to reason about register liveness with predicated instruction. 1568 break; 1569 1570 bool IsSafe = true; 1571 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1572 MachineOperand &MO = TIB->getOperand(i); 1573 if (!MO.isReg()) 1574 continue; 1575 unsigned Reg = MO.getReg(); 1576 if (!Reg) 1577 continue; 1578 if (MO.isDef()) { 1579 if (Uses.count(Reg)) { 1580 // Avoid clobbering a register that's used by the instruction at 1581 // the point of insertion. 1582 IsSafe = false; 1583 break; 1584 } 1585 1586 if (Defs.count(Reg) && !MO.isDead()) { 1587 // Don't hoist the instruction if the def would be clobber by the 1588 // instruction at the point insertion. FIXME: This is overly 1589 // conservative. It should be possible to hoist the instructions 1590 // in BB2 in the following example: 1591 // BB1: 1592 // r1, eflag = op1 r2, r3 1593 // brcc eflag 1594 // 1595 // BB2: 1596 // r1 = op2, ... 1597 // = op3, r1<kill> 1598 IsSafe = false; 1599 break; 1600 } 1601 } else if (!LocalDefsSet.count(Reg)) { 1602 if (Defs.count(Reg)) { 1603 // Use is defined by the instruction at the point of insertion. 1604 IsSafe = false; 1605 break; 1606 } 1607 } 1608 } 1609 if (!IsSafe) 1610 break; 1611 1612 bool DontMoveAcrossStore = true; 1613 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore)) 1614 break; 1615 1616 // Track local defs so we can update liveins. 1617 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1618 MachineOperand &MO = TIB->getOperand(i); 1619 if (!MO.isReg()) 1620 continue; 1621 unsigned Reg = MO.getReg(); 1622 if (!Reg) 1623 continue; 1624 if (MO.isDef()) { 1625 if (!MO.isDead()) { 1626 LocalDefs.push_back(Reg); 1627 LocalDefsSet.insert(Reg); 1628 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1629 LocalDefsSet.insert(*SR); 1630 } 1631 } else if (MO.isKill() && LocalDefsSet.count(Reg)) { 1632 LocalDefsSet.erase(Reg); 1633 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1634 LocalDefsSet.erase(*SR); 1635 } 1636 } 1637 1638 HasDups = true;; 1639 ++TIB; 1640 ++FIB; 1641 } 1642 1643 if (!HasDups) 1644 return false; 1645 1646 MBB->splice(Loc, TBB, TBB->begin(), TIB); 1647 FBB->erase(FBB->begin(), FIB); 1648 1649 // Update livein's. 1650 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) { 1651 unsigned Def = LocalDefs[i]; 1652 if (LocalDefsSet.count(Def)) { 1653 TBB->addLiveIn(Def); 1654 FBB->addLiveIn(Def); 1655 } 1656 } 1657 1658 ++NumHoist; 1659 return true; 1660} 1661