BranchFolding.cpp revision 14ce91cc2720da9547bff1ac199697b6d157d348
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/MachineRegisterInfo.h" 27#include "llvm/CodeGen/RegisterScavenging.h" 28#include "llvm/Target/TargetInstrInfo.h" 29#include "llvm/Target/TargetMachine.h" 30#include "llvm/Target/TargetRegisterInfo.h" 31#include "llvm/Support/CommandLine.h" 32#include "llvm/Support/Debug.h" 33#include "llvm/Support/ErrorHandling.h" 34#include "llvm/Support/raw_ostream.h" 35#include "llvm/ADT/SmallSet.h" 36#include "llvm/ADT/SetVector.h" 37#include "llvm/ADT/Statistic.h" 38#include "llvm/ADT/STLExtras.h" 39#include <algorithm> 40using namespace llvm; 41 42STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); 43STATISTIC(NumBranchOpts, "Number of branches optimized"); 44STATISTIC(NumTailMerge , "Number of block tails merged"); 45STATISTIC(NumHoist , "Number of times common instructions are hoisted"); 46 47static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", 48 cl::init(cl::BOU_UNSET), cl::Hidden); 49 50// Throttle for huge numbers of predecessors (compile speed problems) 51static cl::opt<unsigned> 52TailMergeThreshold("tail-merge-threshold", 53 cl::desc("Max number of predecessors to consider tail merging"), 54 cl::init(150), cl::Hidden); 55 56// Heuristic for tail merging (and, inversely, tail duplication). 57// TODO: This should be replaced with a target query. 58static cl::opt<unsigned> 59TailMergeSize("tail-merge-size", 60 cl::desc("Min number of instructions to consider tail merging"), 61 cl::init(3), cl::Hidden); 62 63namespace { 64 /// BranchFolderPass - Wrap branch folder in a machine function pass. 65 class BranchFolderPass : public MachineFunctionPass { 66 public: 67 static char ID; 68 explicit BranchFolderPass(): MachineFunctionPass(ID) {} 69 70 virtual bool runOnMachineFunction(MachineFunction &MF); 71 72 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 73 AU.addRequired<TargetPassConfig>(); 74 MachineFunctionPass::getAnalysisUsage(AU); 75 } 76 }; 77} 78 79char BranchFolderPass::ID = 0; 80char &llvm::BranchFolderPassID = BranchFolderPass::ID; 81 82INITIALIZE_PASS(BranchFolderPass, "branch-folder", 83 "Control Flow Optimizer", false, false) 84 85bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { 86 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>(); 87 BranchFolder Folder(PassConfig->getEnableTailMerge(), /*CommonHoist=*/true); 88 return Folder.OptimizeFunction(MF, 89 MF.getTarget().getInstrInfo(), 90 MF.getTarget().getRegisterInfo(), 91 getAnalysisIfAvailable<MachineModuleInfo>()); 92} 93 94 95BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist) { 96 switch (FlagEnableTailMerge) { 97 case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; 98 case cl::BOU_TRUE: EnableTailMerge = true; break; 99 case cl::BOU_FALSE: EnableTailMerge = false; break; 100 } 101 102 EnableHoistCommonCode = CommonHoist; 103} 104 105/// RemoveDeadBlock - Remove the specified dead machine basic block from the 106/// function, updating the CFG. 107void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { 108 assert(MBB->pred_empty() && "MBB must be dead!"); 109 DEBUG(dbgs() << "\nRemoving MBB: " << *MBB); 110 111 MachineFunction *MF = MBB->getParent(); 112 // drop all successors. 113 while (!MBB->succ_empty()) 114 MBB->removeSuccessor(MBB->succ_end()-1); 115 116 // Avoid matching if this pointer gets reused. 117 TriedMerging.erase(MBB); 118 119 // Remove the block. 120 MF->erase(MBB); 121} 122 123/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 124/// followed by terminators, and if the implicitly defined registers are not 125/// used by the terminators, remove those implicit_def's. e.g. 126/// BB1: 127/// r0 = implicit_def 128/// r1 = implicit_def 129/// br 130/// This block can be optimized away later if the implicit instructions are 131/// removed. 132bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 133 SmallSet<unsigned, 4> ImpDefRegs; 134 MachineBasicBlock::iterator I = MBB->begin(); 135 while (I != MBB->end()) { 136 if (!I->isImplicitDef()) 137 break; 138 unsigned Reg = I->getOperand(0).getReg(); 139 ImpDefRegs.insert(Reg); 140 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) 141 ImpDefRegs.insert(*SubRegs); 142 ++I; 143 } 144 if (ImpDefRegs.empty()) 145 return false; 146 147 MachineBasicBlock::iterator FirstTerm = I; 148 while (I != MBB->end()) { 149 if (!TII->isUnpredicatedTerminator(I)) 150 return false; 151 // See if it uses any of the implicitly defined registers. 152 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 153 MachineOperand &MO = I->getOperand(i); 154 if (!MO.isReg() || !MO.isUse()) 155 continue; 156 unsigned Reg = MO.getReg(); 157 if (ImpDefRegs.count(Reg)) 158 return false; 159 } 160 ++I; 161 } 162 163 I = MBB->begin(); 164 while (I != FirstTerm) { 165 MachineInstr *ImpDefMI = &*I; 166 ++I; 167 MBB->erase(ImpDefMI); 168 } 169 170 return true; 171} 172 173/// OptimizeFunction - Perhaps branch folding, tail merging and other 174/// CFG optimizations on the given function. 175bool BranchFolder::OptimizeFunction(MachineFunction &MF, 176 const TargetInstrInfo *tii, 177 const TargetRegisterInfo *tri, 178 MachineModuleInfo *mmi) { 179 if (!tii) return false; 180 181 TriedMerging.clear(); 182 183 TII = tii; 184 TRI = tri; 185 MMI = mmi; 186 RS = NULL; 187 188 // Use a RegScavenger to help update liveness when required. 189 MachineRegisterInfo &MRI = MF.getRegInfo(); 190 if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF)) 191 RS = new RegScavenger(); 192 else 193 MRI.invalidateLiveness(); 194 195 // Fix CFG. The later algorithms expect it to be right. 196 bool MadeChange = false; 197 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 198 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 199 SmallVector<MachineOperand, 4> Cond; 200 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 201 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 202 MadeChange |= OptimizeImpDefsBlock(MBB); 203 } 204 205 bool MadeChangeThisIteration = true; 206 while (MadeChangeThisIteration) { 207 MadeChangeThisIteration = TailMergeBlocks(MF); 208 MadeChangeThisIteration |= OptimizeBranches(MF); 209 if (EnableHoistCommonCode) 210 MadeChangeThisIteration |= HoistCommonCode(MF); 211 MadeChange |= MadeChangeThisIteration; 212 } 213 214 // See if any jump tables have become dead as the code generator 215 // did its thing. 216 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 217 if (JTI == 0) { 218 delete RS; 219 return MadeChange; 220 } 221 222 // Walk the function to find jump tables that are live. 223 BitVector JTIsLive(JTI->getJumpTables().size()); 224 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 225 BB != E; ++BB) { 226 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 227 I != E; ++I) 228 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 229 MachineOperand &Op = I->getOperand(op); 230 if (!Op.isJTI()) continue; 231 232 // Remember that this JT is live. 233 JTIsLive.set(Op.getIndex()); 234 } 235 } 236 237 // Finally, remove dead jump tables. This happens when the 238 // indirect jump was unreachable (and thus deleted). 239 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 240 if (!JTIsLive.test(i)) { 241 JTI->RemoveJumpTable(i); 242 MadeChange = true; 243 } 244 245 delete RS; 246 return MadeChange; 247} 248 249//===----------------------------------------------------------------------===// 250// Tail Merging of Blocks 251//===----------------------------------------------------------------------===// 252 253/// HashMachineInstr - Compute a hash value for MI and its operands. 254static unsigned HashMachineInstr(const MachineInstr *MI) { 255 unsigned Hash = MI->getOpcode(); 256 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 257 const MachineOperand &Op = MI->getOperand(i); 258 259 // Merge in bits from the operand if easy. 260 unsigned OperandHash = 0; 261 switch (Op.getType()) { 262 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 263 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 264 case MachineOperand::MO_MachineBasicBlock: 265 OperandHash = Op.getMBB()->getNumber(); 266 break; 267 case MachineOperand::MO_FrameIndex: 268 case MachineOperand::MO_ConstantPoolIndex: 269 case MachineOperand::MO_JumpTableIndex: 270 OperandHash = Op.getIndex(); 271 break; 272 case MachineOperand::MO_GlobalAddress: 273 case MachineOperand::MO_ExternalSymbol: 274 // Global address / external symbol are too hard, don't bother, but do 275 // pull in the offset. 276 OperandHash = Op.getOffset(); 277 break; 278 default: break; 279 } 280 281 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 282 } 283 return Hash; 284} 285 286/// HashEndOfMBB - Hash the last instruction in the MBB. 287static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) { 288 MachineBasicBlock::const_iterator I = MBB->end(); 289 if (I == MBB->begin()) 290 return 0; // Empty MBB. 291 292 --I; 293 // Skip debug info so it will not affect codegen. 294 while (I->isDebugValue()) { 295 if (I==MBB->begin()) 296 return 0; // MBB empty except for debug info. 297 --I; 298 } 299 300 return HashMachineInstr(I); 301} 302 303/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 304/// of instructions they actually have in common together at their end. Return 305/// iterators for the first shared instruction in each block. 306static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 307 MachineBasicBlock *MBB2, 308 MachineBasicBlock::iterator &I1, 309 MachineBasicBlock::iterator &I2) { 310 I1 = MBB1->end(); 311 I2 = MBB2->end(); 312 313 unsigned TailLen = 0; 314 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 315 --I1; --I2; 316 // Skip debugging pseudos; necessary to avoid changing the code. 317 while (I1->isDebugValue()) { 318 if (I1==MBB1->begin()) { 319 while (I2->isDebugValue()) { 320 if (I2==MBB2->begin()) 321 // I1==DBG at begin; I2==DBG at begin 322 return TailLen; 323 --I2; 324 } 325 ++I2; 326 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin 327 return TailLen; 328 } 329 --I1; 330 } 331 // I1==first (untested) non-DBG preceding known match 332 while (I2->isDebugValue()) { 333 if (I2==MBB2->begin()) { 334 ++I1; 335 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin 336 return TailLen; 337 } 338 --I2; 339 } 340 // I1, I2==first (untested) non-DBGs preceding known match 341 if (!I1->isIdenticalTo(I2) || 342 // FIXME: This check is dubious. It's used to get around a problem where 343 // people incorrectly expect inline asm directives to remain in the same 344 // relative order. This is untenable because normal compiler 345 // optimizations (like this one) may reorder and/or merge these 346 // directives. 347 I1->isInlineAsm()) { 348 ++I1; ++I2; 349 break; 350 } 351 ++TailLen; 352 } 353 // Back past possible debugging pseudos at beginning of block. This matters 354 // when one block differs from the other only by whether debugging pseudos 355 // are present at the beginning. (This way, the various checks later for 356 // I1==MBB1->begin() work as expected.) 357 if (I1 == MBB1->begin() && I2 != MBB2->begin()) { 358 --I2; 359 while (I2->isDebugValue()) { 360 if (I2 == MBB2->begin()) 361 return TailLen; 362 --I2; 363 } 364 ++I2; 365 } 366 if (I2 == MBB2->begin() && I1 != MBB1->begin()) { 367 --I1; 368 while (I1->isDebugValue()) { 369 if (I1 == MBB1->begin()) 370 return TailLen; 371 --I1; 372 } 373 ++I1; 374 } 375 return TailLen; 376} 377 378void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB, 379 MachineBasicBlock *NewMBB) { 380 if (RS) { 381 RS->enterBasicBlock(CurMBB); 382 if (!CurMBB->empty()) 383 RS->forward(prior(CurMBB->end())); 384 BitVector RegsLiveAtExit(TRI->getNumRegs()); 385 RS->getRegsUsed(RegsLiveAtExit, false); 386 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 387 if (RegsLiveAtExit[i]) 388 NewMBB->addLiveIn(i); 389 } 390} 391 392/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 393/// after it, replacing it with an unconditional branch to NewDest. 394void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 395 MachineBasicBlock *NewDest) { 396 MachineBasicBlock *CurMBB = OldInst->getParent(); 397 398 TII->ReplaceTailWithBranchTo(OldInst, NewDest); 399 400 // For targets that use the register scavenger, we must maintain LiveIns. 401 MaintainLiveIns(CurMBB, NewDest); 402 403 ++NumTailMerge; 404} 405 406/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 407/// MBB so that the part before the iterator falls into the part starting at the 408/// iterator. This returns the new MBB. 409MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 410 MachineBasicBlock::iterator BBI1) { 411 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1)) 412 return 0; 413 414 MachineFunction &MF = *CurMBB.getParent(); 415 416 // Create the fall-through block. 417 MachineFunction::iterator MBBI = &CurMBB; 418 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 419 CurMBB.getParent()->insert(++MBBI, NewMBB); 420 421 // Move all the successors of this block to the specified block. 422 NewMBB->transferSuccessors(&CurMBB); 423 424 // Add an edge from CurMBB to NewMBB for the fall-through. 425 CurMBB.addSuccessor(NewMBB); 426 427 // Splice the code over. 428 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 429 430 // For targets that use the register scavenger, we must maintain LiveIns. 431 MaintainLiveIns(&CurMBB, NewMBB); 432 433 return NewMBB; 434} 435 436/// EstimateRuntime - Make a rough estimate for how long it will take to run 437/// the specified code. 438static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 439 MachineBasicBlock::iterator E) { 440 unsigned Time = 0; 441 for (; I != E; ++I) { 442 if (I->isDebugValue()) 443 continue; 444 if (I->isCall()) 445 Time += 10; 446 else if (I->mayLoad() || I->mayStore()) 447 Time += 2; 448 else 449 ++Time; 450 } 451 return Time; 452} 453 454// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 455// branches temporarily for tail merging). In the case where CurMBB ends 456// with a conditional branch to the next block, optimize by reversing the 457// test and conditionally branching to SuccMBB instead. 458static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 459 const TargetInstrInfo *TII) { 460 MachineFunction *MF = CurMBB->getParent(); 461 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB)); 462 MachineBasicBlock *TBB = 0, *FBB = 0; 463 SmallVector<MachineOperand, 4> Cond; 464 DebugLoc dl; // FIXME: this is nowhere 465 if (I != MF->end() && 466 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 467 MachineBasicBlock *NextBB = I; 468 if (TBB == NextBB && !Cond.empty() && !FBB) { 469 if (!TII->ReverseBranchCondition(Cond)) { 470 TII->RemoveBranch(*CurMBB); 471 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl); 472 return; 473 } 474 } 475 } 476 TII->InsertBranch(*CurMBB, SuccBB, NULL, 477 SmallVector<MachineOperand, 0>(), dl); 478} 479 480bool 481BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 482 if (getHash() < o.getHash()) 483 return true; 484 if (getHash() > o.getHash()) 485 return false; 486 if (getBlock()->getNumber() < o.getBlock()->getNumber()) 487 return true; 488 if (getBlock()->getNumber() > o.getBlock()->getNumber()) 489 return false; 490 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 491 // an object with itself. 492#ifndef _GLIBCXX_DEBUG 493 llvm_unreachable("Predecessor appears twice"); 494#else 495 return false; 496#endif 497} 498 499/// CountTerminators - Count the number of terminators in the given 500/// block and set I to the position of the first non-terminator, if there 501/// is one, or MBB->end() otherwise. 502static unsigned CountTerminators(MachineBasicBlock *MBB, 503 MachineBasicBlock::iterator &I) { 504 I = MBB->end(); 505 unsigned NumTerms = 0; 506 for (;;) { 507 if (I == MBB->begin()) { 508 I = MBB->end(); 509 break; 510 } 511 --I; 512 if (!I->isTerminator()) break; 513 ++NumTerms; 514 } 515 return NumTerms; 516} 517 518/// ProfitableToMerge - Check if two machine basic blocks have a common tail 519/// and decide if it would be profitable to merge those tails. Return the 520/// length of the common tail and iterators to the first common instruction 521/// in each block. 522static bool ProfitableToMerge(MachineBasicBlock *MBB1, 523 MachineBasicBlock *MBB2, 524 unsigned minCommonTailLength, 525 unsigned &CommonTailLen, 526 MachineBasicBlock::iterator &I1, 527 MachineBasicBlock::iterator &I2, 528 MachineBasicBlock *SuccBB, 529 MachineBasicBlock *PredBB) { 530 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 531 if (CommonTailLen == 0) 532 return false; 533 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber() 534 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen 535 << '\n'); 536 537 // It's almost always profitable to merge any number of non-terminator 538 // instructions with the block that falls through into the common successor. 539 if (MBB1 == PredBB || MBB2 == PredBB) { 540 MachineBasicBlock::iterator I; 541 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 542 if (CommonTailLen > NumTerms) 543 return true; 544 } 545 546 // If one of the blocks can be completely merged and happens to be in 547 // a position where the other could fall through into it, merge any number 548 // of instructions, because it can be done without a branch. 549 // TODO: If the blocks are not adjacent, move one of them so that they are? 550 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 551 return true; 552 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 553 return true; 554 555 // If both blocks have an unconditional branch temporarily stripped out, 556 // count that as an additional common instruction for the following 557 // heuristics. 558 unsigned EffectiveTailLen = CommonTailLen; 559 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 560 !MBB1->back().isBarrier() && 561 !MBB2->back().isBarrier()) 562 ++EffectiveTailLen; 563 564 // Check if the common tail is long enough to be worthwhile. 565 if (EffectiveTailLen >= minCommonTailLength) 566 return true; 567 568 // If we are optimizing for code size, 2 instructions in common is enough if 569 // we don't have to split a block. At worst we will be introducing 1 new 570 // branch instruction, which is likely to be smaller than the 2 571 // instructions that would be deleted in the merge. 572 MachineFunction *MF = MBB1->getParent(); 573 if (EffectiveTailLen >= 2 && 574 MF->getFunction()->getFnAttributes().hasOptimizeForSizeAttr() && 575 (I1 == MBB1->begin() || I2 == MBB2->begin())) 576 return true; 577 578 return false; 579} 580 581/// ComputeSameTails - Look through all the blocks in MergePotentials that have 582/// hash CurHash (guaranteed to match the last element). Build the vector 583/// SameTails of all those that have the (same) largest number of instructions 584/// in common of any pair of these blocks. SameTails entries contain an 585/// iterator into MergePotentials (from which the MachineBasicBlock can be 586/// found) and a MachineBasicBlock::iterator into that MBB indicating the 587/// instruction where the matching code sequence begins. 588/// Order of elements in SameTails is the reverse of the order in which 589/// those blocks appear in MergePotentials (where they are not necessarily 590/// consecutive). 591unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 592 unsigned minCommonTailLength, 593 MachineBasicBlock *SuccBB, 594 MachineBasicBlock *PredBB) { 595 unsigned maxCommonTailLength = 0U; 596 SameTails.clear(); 597 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 598 MPIterator HighestMPIter = prior(MergePotentials.end()); 599 for (MPIterator CurMPIter = prior(MergePotentials.end()), 600 B = MergePotentials.begin(); 601 CurMPIter != B && CurMPIter->getHash() == CurHash; 602 --CurMPIter) { 603 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 604 unsigned CommonTailLen; 605 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 606 minCommonTailLength, 607 CommonTailLen, TrialBBI1, TrialBBI2, 608 SuccBB, PredBB)) { 609 if (CommonTailLen > maxCommonTailLength) { 610 SameTails.clear(); 611 maxCommonTailLength = CommonTailLen; 612 HighestMPIter = CurMPIter; 613 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 614 } 615 if (HighestMPIter == CurMPIter && 616 CommonTailLen == maxCommonTailLength) 617 SameTails.push_back(SameTailElt(I, TrialBBI2)); 618 } 619 if (I == B) 620 break; 621 } 622 } 623 return maxCommonTailLength; 624} 625 626/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 627/// MergePotentials, restoring branches at ends of blocks as appropriate. 628void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 629 MachineBasicBlock *SuccBB, 630 MachineBasicBlock *PredBB) { 631 MPIterator CurMPIter, B; 632 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 633 CurMPIter->getHash() == CurHash; 634 --CurMPIter) { 635 // Put the unconditional branch back, if we need one. 636 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 637 if (SuccBB && CurMBB != PredBB) 638 FixTail(CurMBB, SuccBB, TII); 639 if (CurMPIter == B) 640 break; 641 } 642 if (CurMPIter->getHash() != CurHash) 643 CurMPIter++; 644 MergePotentials.erase(CurMPIter, MergePotentials.end()); 645} 646 647/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 648/// only of the common tail. Create a block that does by splitting one. 649bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 650 unsigned maxCommonTailLength, 651 unsigned &commonTailIndex) { 652 commonTailIndex = 0; 653 unsigned TimeEstimate = ~0U; 654 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 655 // Use PredBB if possible; that doesn't require a new branch. 656 if (SameTails[i].getBlock() == PredBB) { 657 commonTailIndex = i; 658 break; 659 } 660 // Otherwise, make a (fairly bogus) choice based on estimate of 661 // how long it will take the various blocks to execute. 662 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 663 SameTails[i].getTailStartPos()); 664 if (t <= TimeEstimate) { 665 TimeEstimate = t; 666 commonTailIndex = i; 667 } 668 } 669 670 MachineBasicBlock::iterator BBI = 671 SameTails[commonTailIndex].getTailStartPos(); 672 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 673 674 // If the common tail includes any debug info we will take it pretty 675 // randomly from one of the inputs. Might be better to remove it? 676 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 677 << maxCommonTailLength); 678 679 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 680 if (!newMBB) { 681 DEBUG(dbgs() << "... failed!"); 682 return false; 683 } 684 685 SameTails[commonTailIndex].setBlock(newMBB); 686 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 687 688 // If we split PredBB, newMBB is the new predecessor. 689 if (PredBB == MBB) 690 PredBB = newMBB; 691 692 return true; 693} 694 695// See if any of the blocks in MergePotentials (which all have a common single 696// successor, or all have no successor) can be tail-merged. If there is a 697// successor, any blocks in MergePotentials that are not tail-merged and 698// are not immediately before Succ must have an unconditional branch to 699// Succ added (but the predecessor/successor lists need no adjustment). 700// The lone predecessor of Succ that falls through into Succ, 701// if any, is given in PredBB. 702 703bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 704 MachineBasicBlock *PredBB) { 705 bool MadeChange = false; 706 707 // Except for the special cases below, tail-merge if there are at least 708 // this many instructions in common. 709 unsigned minCommonTailLength = TailMergeSize; 710 711 DEBUG(dbgs() << "\nTryTailMergeBlocks: "; 712 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 713 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 714 << (i == e-1 ? "" : ", "); 715 dbgs() << "\n"; 716 if (SuccBB) { 717 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 718 if (PredBB) 719 dbgs() << " which has fall-through from BB#" 720 << PredBB->getNumber() << "\n"; 721 } 722 dbgs() << "Looking for common tails of at least " 723 << minCommonTailLength << " instruction" 724 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 725 ); 726 727 // Sort by hash value so that blocks with identical end sequences sort 728 // together. 729 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 730 731 // Walk through equivalence sets looking for actual exact matches. 732 while (MergePotentials.size() > 1) { 733 unsigned CurHash = MergePotentials.back().getHash(); 734 735 // Build SameTails, identifying the set of blocks with this hash code 736 // and with the maximum number of instructions in common. 737 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 738 minCommonTailLength, 739 SuccBB, PredBB); 740 741 // If we didn't find any pair that has at least minCommonTailLength 742 // instructions in common, remove all blocks with this hash code and retry. 743 if (SameTails.empty()) { 744 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 745 continue; 746 } 747 748 // If one of the blocks is the entire common tail (and not the entry 749 // block, which we can't jump to), we can treat all blocks with this same 750 // tail at once. Use PredBB if that is one of the possibilities, as that 751 // will not introduce any extra branches. 752 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 753 getParent()->begin(); 754 unsigned commonTailIndex = SameTails.size(); 755 // If there are two blocks, check to see if one can be made to fall through 756 // into the other. 757 if (SameTails.size() == 2 && 758 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 759 SameTails[1].tailIsWholeBlock()) 760 commonTailIndex = 1; 761 else if (SameTails.size() == 2 && 762 SameTails[1].getBlock()->isLayoutSuccessor( 763 SameTails[0].getBlock()) && 764 SameTails[0].tailIsWholeBlock()) 765 commonTailIndex = 0; 766 else { 767 // Otherwise just pick one, favoring the fall-through predecessor if 768 // there is one. 769 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 770 MachineBasicBlock *MBB = SameTails[i].getBlock(); 771 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 772 continue; 773 if (MBB == PredBB) { 774 commonTailIndex = i; 775 break; 776 } 777 if (SameTails[i].tailIsWholeBlock()) 778 commonTailIndex = i; 779 } 780 } 781 782 if (commonTailIndex == SameTails.size() || 783 (SameTails[commonTailIndex].getBlock() == PredBB && 784 !SameTails[commonTailIndex].tailIsWholeBlock())) { 785 // None of the blocks consist entirely of the common tail. 786 // Split a block so that one does. 787 if (!CreateCommonTailOnlyBlock(PredBB, 788 maxCommonTailLength, commonTailIndex)) { 789 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 790 continue; 791 } 792 } 793 794 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 795 // MBB is common tail. Adjust all other BB's to jump to this one. 796 // Traversal must be forwards so erases work. 797 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber() 798 << " for "); 799 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 800 if (commonTailIndex == i) 801 continue; 802 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() 803 << (i == e-1 ? "" : ", ")); 804 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 805 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 806 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 807 MergePotentials.erase(SameTails[i].getMPIter()); 808 } 809 DEBUG(dbgs() << "\n"); 810 // We leave commonTailIndex in the worklist in case there are other blocks 811 // that match it with a smaller number of instructions. 812 MadeChange = true; 813 } 814 return MadeChange; 815} 816 817bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 818 bool MadeChange = false; 819 if (!EnableTailMerge) return MadeChange; 820 821 // First find blocks with no successors. 822 MergePotentials.clear(); 823 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); 824 I != E && MergePotentials.size() < TailMergeThreshold; ++I) { 825 if (TriedMerging.count(I)) 826 continue; 827 if (I->succ_empty()) 828 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I)); 829 } 830 831 // If this is a large problem, avoid visiting the same basic blocks 832 // multiple times. 833 if (MergePotentials.size() == TailMergeThreshold) 834 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 835 TriedMerging.insert(MergePotentials[i].getBlock()); 836 837 // See if we can do any tail merging on those. 838 if (MergePotentials.size() >= 2) 839 MadeChange |= TryTailMergeBlocks(NULL, NULL); 840 841 // Look at blocks (IBB) with multiple predecessors (PBB). 842 // We change each predecessor to a canonical form, by 843 // (1) temporarily removing any unconditional branch from the predecessor 844 // to IBB, and 845 // (2) alter conditional branches so they branch to the other block 846 // not IBB; this may require adding back an unconditional branch to IBB 847 // later, where there wasn't one coming in. E.g. 848 // Bcc IBB 849 // fallthrough to QBB 850 // here becomes 851 // Bncc QBB 852 // with a conceptual B to IBB after that, which never actually exists. 853 // With those changes, we see whether the predecessors' tails match, 854 // and merge them if so. We change things out of canonical form and 855 // back to the way they were later in the process. (OptimizeBranches 856 // would undo some of this, but we can't use it, because we'd get into 857 // a compile-time infinite loop repeatedly doing and undoing the same 858 // transformations.) 859 860 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 861 I != E; ++I) { 862 if (I->pred_size() < 2) continue; 863 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 864 MachineBasicBlock *IBB = I; 865 MachineBasicBlock *PredBB = prior(I); 866 MergePotentials.clear(); 867 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 868 E2 = I->pred_end(); 869 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) { 870 MachineBasicBlock *PBB = *P; 871 if (TriedMerging.count(PBB)) 872 continue; 873 874 // Skip blocks that loop to themselves, can't tail merge these. 875 if (PBB == IBB) 876 continue; 877 878 // Visit each predecessor only once. 879 if (!UniquePreds.insert(PBB)) 880 continue; 881 882 // Skip blocks which may jump to a landing pad. Can't tail merge these. 883 if (PBB->getLandingPadSuccessor()) 884 continue; 885 886 MachineBasicBlock *TBB = 0, *FBB = 0; 887 SmallVector<MachineOperand, 4> Cond; 888 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 889 // Failing case: IBB is the target of a cbr, and we cannot reverse the 890 // branch. 891 SmallVector<MachineOperand, 4> NewCond(Cond); 892 if (!Cond.empty() && TBB == IBB) { 893 if (TII->ReverseBranchCondition(NewCond)) 894 continue; 895 // This is the QBB case described above 896 if (!FBB) 897 FBB = llvm::next(MachineFunction::iterator(PBB)); 898 } 899 900 // Failing case: the only way IBB can be reached from PBB is via 901 // exception handling. Happens for landing pads. Would be nice to have 902 // a bit in the edge so we didn't have to do all this. 903 if (IBB->isLandingPad()) { 904 MachineFunction::iterator IP = PBB; IP++; 905 MachineBasicBlock *PredNextBB = NULL; 906 if (IP != MF.end()) 907 PredNextBB = IP; 908 if (TBB == NULL) { 909 if (IBB != PredNextBB) // fallthrough 910 continue; 911 } else if (FBB) { 912 if (TBB != IBB && FBB != IBB) // cbr then ubr 913 continue; 914 } else if (Cond.empty()) { 915 if (TBB != IBB) // ubr 916 continue; 917 } else { 918 if (TBB != IBB && IBB != PredNextBB) // cbr 919 continue; 920 } 921 } 922 923 // Remove the unconditional branch at the end, if any. 924 if (TBB && (Cond.empty() || FBB)) { 925 DebugLoc dl; // FIXME: this is nowhere 926 TII->RemoveBranch(*PBB); 927 if (!Cond.empty()) 928 // reinsert conditional branch only, for now 929 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl); 930 } 931 932 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P)); 933 } 934 } 935 936 // If this is a large problem, avoid visiting the same basic blocks multiple 937 // times. 938 if (MergePotentials.size() == TailMergeThreshold) 939 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 940 TriedMerging.insert(MergePotentials[i].getBlock()); 941 942 if (MergePotentials.size() >= 2) 943 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 944 945 // Reinsert an unconditional branch if needed. The 1 below can occur as a 946 // result of removing blocks in TryTailMergeBlocks. 947 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 948 if (MergePotentials.size() == 1 && 949 MergePotentials.begin()->getBlock() != PredBB) 950 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 951 } 952 953 return MadeChange; 954} 955 956//===----------------------------------------------------------------------===// 957// Branch Optimization 958//===----------------------------------------------------------------------===// 959 960bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 961 bool MadeChange = false; 962 963 // Make sure blocks are numbered in order 964 MF.RenumberBlocks(); 965 966 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 967 I != E; ) { 968 MachineBasicBlock *MBB = I++; 969 MadeChange |= OptimizeBlock(MBB); 970 971 // If it is dead, remove it. 972 if (MBB->pred_empty()) { 973 RemoveDeadBlock(MBB); 974 MadeChange = true; 975 ++NumDeadBlocks; 976 } 977 } 978 return MadeChange; 979} 980 981// Blocks should be considered empty if they contain only debug info; 982// else the debug info would affect codegen. 983static bool IsEmptyBlock(MachineBasicBlock *MBB) { 984 if (MBB->empty()) 985 return true; 986 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end(); 987 MBBI!=MBBE; ++MBBI) { 988 if (!MBBI->isDebugValue()) 989 return false; 990 } 991 return true; 992} 993 994// Blocks with only debug info and branches should be considered the same 995// as blocks with only branches. 996static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { 997 MachineBasicBlock::iterator MBBI, MBBE; 998 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) { 999 if (!MBBI->isDebugValue()) 1000 break; 1001 } 1002 return (MBBI->isBranch()); 1003} 1004 1005/// IsBetterFallthrough - Return true if it would be clearly better to 1006/// fall-through to MBB1 than to fall through into MBB2. This has to return 1007/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 1008/// result in infinite loops. 1009static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 1010 MachineBasicBlock *MBB2) { 1011 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 1012 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 1013 // optimize branches that branch to either a return block or an assert block 1014 // into a fallthrough to the return. 1015 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false; 1016 1017 // If there is a clear successor ordering we make sure that one block 1018 // will fall through to the next 1019 if (MBB1->isSuccessor(MBB2)) return true; 1020 if (MBB2->isSuccessor(MBB1)) return false; 1021 1022 // Neither block consists entirely of debug info (per IsEmptyBlock check), 1023 // so we needn't test for falling off the beginning here. 1024 MachineBasicBlock::iterator MBB1I = --MBB1->end(); 1025 while (MBB1I->isDebugValue()) 1026 --MBB1I; 1027 MachineBasicBlock::iterator MBB2I = --MBB2->end(); 1028 while (MBB2I->isDebugValue()) 1029 --MBB2I; 1030 return MBB2I->isCall() && !MBB1I->isCall(); 1031} 1032 1033/// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch 1034/// instructions on the block. Always use the DebugLoc of the first 1035/// branching instruction found unless its absent, in which case use the 1036/// DebugLoc of the second if present. 1037static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) { 1038 MachineBasicBlock::iterator I = MBB.end(); 1039 if (I == MBB.begin()) 1040 return DebugLoc(); 1041 --I; 1042 while (I->isDebugValue() && I != MBB.begin()) 1043 --I; 1044 if (I->isBranch()) 1045 return I->getDebugLoc(); 1046 return DebugLoc(); 1047} 1048 1049/// OptimizeBlock - Analyze and optimize control flow related to the specified 1050/// block. This is never called on the entry block. 1051bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1052 bool MadeChange = false; 1053 MachineFunction &MF = *MBB->getParent(); 1054ReoptimizeBlock: 1055 1056 MachineFunction::iterator FallThrough = MBB; 1057 ++FallThrough; 1058 1059 // If this block is empty, make everyone use its fall-through, not the block 1060 // explicitly. Landing pads should not do this since the landing-pad table 1061 // points to this block. Blocks with their addresses taken shouldn't be 1062 // optimized away. 1063 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 1064 // Dead block? Leave for cleanup later. 1065 if (MBB->pred_empty()) return MadeChange; 1066 1067 if (FallThrough == MF.end()) { 1068 // TODO: Simplify preds to not branch here if possible! 1069 } else { 1070 // Rewrite all predecessors of the old block to go to the fallthrough 1071 // instead. 1072 while (!MBB->pred_empty()) { 1073 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1074 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1075 } 1076 // If MBB was the target of a jump table, update jump tables to go to the 1077 // fallthrough instead. 1078 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1079 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough); 1080 MadeChange = true; 1081 } 1082 return MadeChange; 1083 } 1084 1085 // Check to see if we can simplify the terminator of the block before this 1086 // one. 1087 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1088 1089 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1090 SmallVector<MachineOperand, 4> PriorCond; 1091 bool PriorUnAnalyzable = 1092 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1093 if (!PriorUnAnalyzable) { 1094 // If the CFG for the prior block has extra edges, remove them. 1095 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1096 !PriorCond.empty()); 1097 1098 // If the previous branch is conditional and both conditions go to the same 1099 // destination, remove the branch, replacing it with an unconditional one or 1100 // a fall-through. 1101 if (PriorTBB && PriorTBB == PriorFBB) { 1102 DebugLoc dl = getBranchDebugLoc(PrevBB); 1103 TII->RemoveBranch(PrevBB); 1104 PriorCond.clear(); 1105 if (PriorTBB != MBB) 1106 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1107 MadeChange = true; 1108 ++NumBranchOpts; 1109 goto ReoptimizeBlock; 1110 } 1111 1112 // If the previous block unconditionally falls through to this block and 1113 // this block has no other predecessors, move the contents of this block 1114 // into the prior block. This doesn't usually happen when SimplifyCFG 1115 // has been used, but it can happen if tail merging splits a fall-through 1116 // predecessor of a block. 1117 // This has to check PrevBB->succ_size() because EH edges are ignored by 1118 // AnalyzeBranch. 1119 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1120 PrevBB.succ_size() == 1 && 1121 !MBB->hasAddressTaken() && !MBB->isLandingPad()) { 1122 DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1123 << "From MBB: " << *MBB); 1124 // Remove redundant DBG_VALUEs first. 1125 if (PrevBB.begin() != PrevBB.end()) { 1126 MachineBasicBlock::iterator PrevBBIter = PrevBB.end(); 1127 --PrevBBIter; 1128 MachineBasicBlock::iterator MBBIter = MBB->begin(); 1129 // Check if DBG_VALUE at the end of PrevBB is identical to the 1130 // DBG_VALUE at the beginning of MBB. 1131 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end() 1132 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) { 1133 if (!MBBIter->isIdenticalTo(PrevBBIter)) 1134 break; 1135 MachineInstr *DuplicateDbg = MBBIter; 1136 ++MBBIter; -- PrevBBIter; 1137 DuplicateDbg->eraseFromParent(); 1138 } 1139 } 1140 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1141 PrevBB.removeSuccessor(PrevBB.succ_begin()); 1142 assert(PrevBB.succ_empty()); 1143 PrevBB.transferSuccessors(MBB); 1144 MadeChange = true; 1145 return MadeChange; 1146 } 1147 1148 // If the previous branch *only* branches to *this* block (conditional or 1149 // not) remove the branch. 1150 if (PriorTBB == MBB && PriorFBB == 0) { 1151 TII->RemoveBranch(PrevBB); 1152 MadeChange = true; 1153 ++NumBranchOpts; 1154 goto ReoptimizeBlock; 1155 } 1156 1157 // If the prior block branches somewhere else on the condition and here if 1158 // the condition is false, remove the uncond second branch. 1159 if (PriorFBB == MBB) { 1160 DebugLoc dl = getBranchDebugLoc(PrevBB); 1161 TII->RemoveBranch(PrevBB); 1162 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1163 MadeChange = true; 1164 ++NumBranchOpts; 1165 goto ReoptimizeBlock; 1166 } 1167 1168 // If the prior block branches here on true and somewhere else on false, and 1169 // if the branch condition is reversible, reverse the branch to create a 1170 // fall-through. 1171 if (PriorTBB == MBB) { 1172 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1173 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1174 DebugLoc dl = getBranchDebugLoc(PrevBB); 1175 TII->RemoveBranch(PrevBB); 1176 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl); 1177 MadeChange = true; 1178 ++NumBranchOpts; 1179 goto ReoptimizeBlock; 1180 } 1181 } 1182 1183 // If this block has no successors (e.g. it is a return block or ends with 1184 // a call to a no-return function like abort or __cxa_throw) and if the pred 1185 // falls through into this block, and if it would otherwise fall through 1186 // into the block after this, move this block to the end of the function. 1187 // 1188 // We consider it more likely that execution will stay in the function (e.g. 1189 // due to loops) than it is to exit it. This asserts in loops etc, moving 1190 // the assert condition out of the loop body. 1191 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1192 MachineFunction::iterator(PriorTBB) == FallThrough && 1193 !MBB->canFallThrough()) { 1194 bool DoTransform = true; 1195 1196 // We have to be careful that the succs of PredBB aren't both no-successor 1197 // blocks. If neither have successors and if PredBB is the second from 1198 // last block in the function, we'd just keep swapping the two blocks for 1199 // last. Only do the swap if one is clearly better to fall through than 1200 // the other. 1201 if (FallThrough == --MF.end() && 1202 !IsBetterFallthrough(PriorTBB, MBB)) 1203 DoTransform = false; 1204 1205 if (DoTransform) { 1206 // Reverse the branch so we will fall through on the previous true cond. 1207 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1208 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1209 DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1210 << "To make fallthrough to: " << *PriorTBB << "\n"); 1211 1212 DebugLoc dl = getBranchDebugLoc(PrevBB); 1213 TII->RemoveBranch(PrevBB); 1214 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl); 1215 1216 // Move this block to the end of the function. 1217 MBB->moveAfter(--MF.end()); 1218 MadeChange = true; 1219 ++NumBranchOpts; 1220 return MadeChange; 1221 } 1222 } 1223 } 1224 } 1225 1226 // Analyze the branch in the current block. 1227 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1228 SmallVector<MachineOperand, 4> CurCond; 1229 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1230 if (!CurUnAnalyzable) { 1231 // If the CFG for the prior block has extra edges, remove them. 1232 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1233 1234 // If this is a two-way branch, and the FBB branches to this block, reverse 1235 // the condition so the single-basic-block loop is faster. Instead of: 1236 // Loop: xxx; jcc Out; jmp Loop 1237 // we want: 1238 // Loop: xxx; jncc Loop; jmp Out 1239 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1240 SmallVector<MachineOperand, 4> NewCond(CurCond); 1241 if (!TII->ReverseBranchCondition(NewCond)) { 1242 DebugLoc dl = getBranchDebugLoc(*MBB); 1243 TII->RemoveBranch(*MBB); 1244 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); 1245 MadeChange = true; 1246 ++NumBranchOpts; 1247 goto ReoptimizeBlock; 1248 } 1249 } 1250 1251 // If this branch is the only thing in its block, see if we can forward 1252 // other blocks across it. 1253 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1254 IsBranchOnlyBlock(MBB) && CurTBB != MBB && 1255 !MBB->hasAddressTaken()) { 1256 DebugLoc dl = getBranchDebugLoc(*MBB); 1257 // This block may contain just an unconditional branch. Because there can 1258 // be 'non-branch terminators' in the block, try removing the branch and 1259 // then seeing if the block is empty. 1260 TII->RemoveBranch(*MBB); 1261 // If the only things remaining in the block are debug info, remove these 1262 // as well, so this will behave the same as an empty block in non-debug 1263 // mode. 1264 if (!MBB->empty()) { 1265 bool NonDebugInfoFound = false; 1266 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 1267 I != E; ++I) { 1268 if (!I->isDebugValue()) { 1269 NonDebugInfoFound = true; 1270 break; 1271 } 1272 } 1273 if (!NonDebugInfoFound) 1274 // Make the block empty, losing the debug info (we could probably 1275 // improve this in some cases.) 1276 MBB->erase(MBB->begin(), MBB->end()); 1277 } 1278 // If this block is just an unconditional branch to CurTBB, we can 1279 // usually completely eliminate the block. The only case we cannot 1280 // completely eliminate the block is when the block before this one 1281 // falls through into MBB and we can't understand the prior block's branch 1282 // condition. 1283 if (MBB->empty()) { 1284 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1285 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1286 !PrevBB.isSuccessor(MBB)) { 1287 // If the prior block falls through into us, turn it into an 1288 // explicit branch to us to make updates simpler. 1289 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1290 PriorTBB != MBB && PriorFBB != MBB) { 1291 if (PriorTBB == 0) { 1292 assert(PriorCond.empty() && PriorFBB == 0 && 1293 "Bad branch analysis"); 1294 PriorTBB = MBB; 1295 } else { 1296 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1297 PriorFBB = MBB; 1298 } 1299 DebugLoc pdl = getBranchDebugLoc(PrevBB); 1300 TII->RemoveBranch(PrevBB); 1301 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl); 1302 } 1303 1304 // Iterate through all the predecessors, revectoring each in-turn. 1305 size_t PI = 0; 1306 bool DidChange = false; 1307 bool HasBranchToSelf = false; 1308 while(PI != MBB->pred_size()) { 1309 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1310 if (PMBB == MBB) { 1311 // If this block has an uncond branch to itself, leave it. 1312 ++PI; 1313 HasBranchToSelf = true; 1314 } else { 1315 DidChange = true; 1316 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1317 // If this change resulted in PMBB ending in a conditional 1318 // branch where both conditions go to the same destination, 1319 // change this to an unconditional branch (and fix the CFG). 1320 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1321 SmallVector<MachineOperand, 4> NewCurCond; 1322 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1323 NewCurFBB, NewCurCond, true); 1324 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1325 DebugLoc pdl = getBranchDebugLoc(*PMBB); 1326 TII->RemoveBranch(*PMBB); 1327 NewCurCond.clear(); 1328 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, pdl); 1329 MadeChange = true; 1330 ++NumBranchOpts; 1331 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1332 } 1333 } 1334 } 1335 1336 // Change any jumptables to go to the new MBB. 1337 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1338 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); 1339 if (DidChange) { 1340 ++NumBranchOpts; 1341 MadeChange = true; 1342 if (!HasBranchToSelf) return MadeChange; 1343 } 1344 } 1345 } 1346 1347 // Add the branch back if the block is more than just an uncond branch. 1348 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl); 1349 } 1350 } 1351 1352 // If the prior block doesn't fall through into this block, and if this 1353 // block doesn't fall through into some other block, see if we can find a 1354 // place to move this block where a fall-through will happen. 1355 if (!PrevBB.canFallThrough()) { 1356 1357 // Now we know that there was no fall-through into this block, check to 1358 // see if it has a fall-through into its successor. 1359 bool CurFallsThru = MBB->canFallThrough(); 1360 1361 if (!MBB->isLandingPad()) { 1362 // Check all the predecessors of this block. If one of them has no fall 1363 // throughs, move this block right after it. 1364 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1365 E = MBB->pred_end(); PI != E; ++PI) { 1366 // Analyze the branch at the end of the pred. 1367 MachineBasicBlock *PredBB = *PI; 1368 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1369 MachineBasicBlock *PredTBB = 0, *PredFBB = 0; 1370 SmallVector<MachineOperand, 4> PredCond; 1371 if (PredBB != MBB && !PredBB->canFallThrough() && 1372 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1373 && (!CurFallsThru || !CurTBB || !CurFBB) 1374 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1375 // If the current block doesn't fall through, just move it. 1376 // If the current block can fall through and does not end with a 1377 // conditional branch, we need to append an unconditional jump to 1378 // the (current) next block. To avoid a possible compile-time 1379 // infinite loop, move blocks only backward in this case. 1380 // Also, if there are already 2 branches here, we cannot add a third; 1381 // this means we have the case 1382 // Bcc next 1383 // B elsewhere 1384 // next: 1385 if (CurFallsThru) { 1386 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); 1387 CurCond.clear(); 1388 TII->InsertBranch(*MBB, NextBB, 0, CurCond, DebugLoc()); 1389 } 1390 MBB->moveAfter(PredBB); 1391 MadeChange = true; 1392 goto ReoptimizeBlock; 1393 } 1394 } 1395 } 1396 1397 if (!CurFallsThru) { 1398 // Check all successors to see if we can move this block before it. 1399 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1400 E = MBB->succ_end(); SI != E; ++SI) { 1401 // Analyze the branch at the end of the block before the succ. 1402 MachineBasicBlock *SuccBB = *SI; 1403 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1404 1405 // If this block doesn't already fall-through to that successor, and if 1406 // the succ doesn't already have a block that can fall through into it, 1407 // and if the successor isn't an EH destination, we can arrange for the 1408 // fallthrough to happen. 1409 if (SuccBB != MBB && &*SuccPrev != MBB && 1410 !SuccPrev->canFallThrough() && !CurUnAnalyzable && 1411 !SuccBB->isLandingPad()) { 1412 MBB->moveBefore(SuccBB); 1413 MadeChange = true; 1414 goto ReoptimizeBlock; 1415 } 1416 } 1417 1418 // Okay, there is no really great place to put this block. If, however, 1419 // the block before this one would be a fall-through if this block were 1420 // removed, move this block to the end of the function. 1421 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0; 1422 SmallVector<MachineOperand, 4> PrevCond; 1423 if (FallThrough != MF.end() && 1424 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1425 PrevBB.isSuccessor(FallThrough)) { 1426 MBB->moveAfter(--MF.end()); 1427 MadeChange = true; 1428 return MadeChange; 1429 } 1430 } 1431 } 1432 1433 return MadeChange; 1434} 1435 1436//===----------------------------------------------------------------------===// 1437// Hoist Common Code 1438//===----------------------------------------------------------------------===// 1439 1440/// HoistCommonCode - Hoist common instruction sequences at the start of basic 1441/// blocks to their common predecessor. 1442bool BranchFolder::HoistCommonCode(MachineFunction &MF) { 1443 bool MadeChange = false; 1444 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) { 1445 MachineBasicBlock *MBB = I++; 1446 MadeChange |= HoistCommonCodeInSuccs(MBB); 1447 } 1448 1449 return MadeChange; 1450} 1451 1452/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 1453/// its 'true' successor. 1454static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 1455 MachineBasicBlock *TrueBB) { 1456 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 1457 E = BB->succ_end(); SI != E; ++SI) { 1458 MachineBasicBlock *SuccBB = *SI; 1459 if (SuccBB != TrueBB) 1460 return SuccBB; 1461 } 1462 return NULL; 1463} 1464 1465/// findHoistingInsertPosAndDeps - Find the location to move common instructions 1466/// in successors to. The location is usually just before the terminator, 1467/// however if the terminator is a conditional branch and its previous 1468/// instruction is the flag setting instruction, the previous instruction is 1469/// the preferred location. This function also gathers uses and defs of the 1470/// instructions from the insertion point to the end of the block. The data is 1471/// used by HoistCommonCodeInSuccs to ensure safety. 1472static 1473MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB, 1474 const TargetInstrInfo *TII, 1475 const TargetRegisterInfo *TRI, 1476 SmallSet<unsigned,4> &Uses, 1477 SmallSet<unsigned,4> &Defs) { 1478 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator(); 1479 if (!TII->isUnpredicatedTerminator(Loc)) 1480 return MBB->end(); 1481 1482 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) { 1483 const MachineOperand &MO = Loc->getOperand(i); 1484 if (!MO.isReg()) 1485 continue; 1486 unsigned Reg = MO.getReg(); 1487 if (!Reg) 1488 continue; 1489 if (MO.isUse()) { 1490 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1491 Uses.insert(*AI); 1492 } else if (!MO.isDead()) 1493 // Don't try to hoist code in the rare case the terminator defines a 1494 // register that is later used. 1495 return MBB->end(); 1496 } 1497 1498 if (Uses.empty()) 1499 return Loc; 1500 if (Loc == MBB->begin()) 1501 return MBB->end(); 1502 1503 // The terminator is probably a conditional branch, try not to separate the 1504 // branch from condition setting instruction. 1505 MachineBasicBlock::iterator PI = Loc; 1506 --PI; 1507 while (PI != MBB->begin() && Loc->isDebugValue()) 1508 --PI; 1509 1510 bool IsDef = false; 1511 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) { 1512 const MachineOperand &MO = PI->getOperand(i); 1513 // If PI has a regmask operand, it is probably a call. Separate away. 1514 if (MO.isRegMask()) 1515 return Loc; 1516 if (!MO.isReg() || MO.isUse()) 1517 continue; 1518 unsigned Reg = MO.getReg(); 1519 if (!Reg) 1520 continue; 1521 if (Uses.count(Reg)) 1522 IsDef = true; 1523 } 1524 if (!IsDef) 1525 // The condition setting instruction is not just before the conditional 1526 // branch. 1527 return Loc; 1528 1529 // Be conservative, don't insert instruction above something that may have 1530 // side-effects. And since it's potentially bad to separate flag setting 1531 // instruction from the conditional branch, just abort the optimization 1532 // completely. 1533 // Also avoid moving code above predicated instruction since it's hard to 1534 // reason about register liveness with predicated instruction. 1535 bool DontMoveAcrossStore = true; 1536 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) || 1537 TII->isPredicated(PI)) 1538 return MBB->end(); 1539 1540 1541 // Find out what registers are live. Note this routine is ignoring other live 1542 // registers which are only used by instructions in successor blocks. 1543 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) { 1544 const MachineOperand &MO = PI->getOperand(i); 1545 if (!MO.isReg()) 1546 continue; 1547 unsigned Reg = MO.getReg(); 1548 if (!Reg) 1549 continue; 1550 if (MO.isUse()) { 1551 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1552 Uses.insert(*AI); 1553 } else { 1554 if (Uses.erase(Reg)) { 1555 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) 1556 Uses.erase(*SubRegs); // Use sub-registers to be conservative 1557 } 1558 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1559 Defs.insert(*AI); 1560 } 1561 } 1562 1563 return PI; 1564} 1565 1566/// HoistCommonCodeInSuccs - If the successors of MBB has common instruction 1567/// sequence at the start of the function, move the instructions before MBB 1568/// terminator if it's legal. 1569bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) { 1570 MachineBasicBlock *TBB = 0, *FBB = 0; 1571 SmallVector<MachineOperand, 4> Cond; 1572 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty()) 1573 return false; 1574 1575 if (!FBB) FBB = findFalseBlock(MBB, TBB); 1576 if (!FBB) 1577 // Malformed bcc? True and false blocks are the same? 1578 return false; 1579 1580 // Restrict the optimization to cases where MBB is the only predecessor, 1581 // it is an obvious win. 1582 if (TBB->pred_size() > 1 || FBB->pred_size() > 1) 1583 return false; 1584 1585 // Find a suitable position to hoist the common instructions to. Also figure 1586 // out which registers are used or defined by instructions from the insertion 1587 // point to the end of the block. 1588 SmallSet<unsigned, 4> Uses, Defs; 1589 MachineBasicBlock::iterator Loc = 1590 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs); 1591 if (Loc == MBB->end()) 1592 return false; 1593 1594 bool HasDups = false; 1595 SmallVector<unsigned, 4> LocalDefs; 1596 SmallSet<unsigned, 4> LocalDefsSet; 1597 MachineBasicBlock::iterator TIB = TBB->begin(); 1598 MachineBasicBlock::iterator FIB = FBB->begin(); 1599 MachineBasicBlock::iterator TIE = TBB->end(); 1600 MachineBasicBlock::iterator FIE = FBB->end(); 1601 while (TIB != TIE && FIB != FIE) { 1602 // Skip dbg_value instructions. These do not count. 1603 if (TIB->isDebugValue()) { 1604 while (TIB != TIE && TIB->isDebugValue()) 1605 ++TIB; 1606 if (TIB == TIE) 1607 break; 1608 } 1609 if (FIB->isDebugValue()) { 1610 while (FIB != FIE && FIB->isDebugValue()) 1611 ++FIB; 1612 if (FIB == FIE) 1613 break; 1614 } 1615 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead)) 1616 break; 1617 1618 if (TII->isPredicated(TIB)) 1619 // Hard to reason about register liveness with predicated instruction. 1620 break; 1621 1622 bool IsSafe = true; 1623 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1624 MachineOperand &MO = TIB->getOperand(i); 1625 // Don't attempt to hoist instructions with register masks. 1626 if (MO.isRegMask()) { 1627 IsSafe = false; 1628 break; 1629 } 1630 if (!MO.isReg()) 1631 continue; 1632 unsigned Reg = MO.getReg(); 1633 if (!Reg) 1634 continue; 1635 if (MO.isDef()) { 1636 if (Uses.count(Reg)) { 1637 // Avoid clobbering a register that's used by the instruction at 1638 // the point of insertion. 1639 IsSafe = false; 1640 break; 1641 } 1642 1643 if (Defs.count(Reg) && !MO.isDead()) { 1644 // Don't hoist the instruction if the def would be clobber by the 1645 // instruction at the point insertion. FIXME: This is overly 1646 // conservative. It should be possible to hoist the instructions 1647 // in BB2 in the following example: 1648 // BB1: 1649 // r1, eflag = op1 r2, r3 1650 // brcc eflag 1651 // 1652 // BB2: 1653 // r1 = op2, ... 1654 // = op3, r1<kill> 1655 IsSafe = false; 1656 break; 1657 } 1658 } else if (!LocalDefsSet.count(Reg)) { 1659 if (Defs.count(Reg)) { 1660 // Use is defined by the instruction at the point of insertion. 1661 IsSafe = false; 1662 break; 1663 } 1664 1665 if (MO.isKill() && Uses.count(Reg)) 1666 // Kills a register that's read by the instruction at the point of 1667 // insertion. Remove the kill marker. 1668 MO.setIsKill(false); 1669 } 1670 } 1671 if (!IsSafe) 1672 break; 1673 1674 bool DontMoveAcrossStore = true; 1675 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore)) 1676 break; 1677 1678 // Remove kills from LocalDefsSet, these registers had short live ranges. 1679 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1680 MachineOperand &MO = TIB->getOperand(i); 1681 if (!MO.isReg() || !MO.isUse() || !MO.isKill()) 1682 continue; 1683 unsigned Reg = MO.getReg(); 1684 if (!Reg || !LocalDefsSet.count(Reg)) 1685 continue; 1686 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1687 LocalDefsSet.erase(*AI); 1688 } 1689 1690 // Track local defs so we can update liveins. 1691 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1692 MachineOperand &MO = TIB->getOperand(i); 1693 if (!MO.isReg() || !MO.isDef() || MO.isDead()) 1694 continue; 1695 unsigned Reg = MO.getReg(); 1696 if (!Reg) 1697 continue; 1698 LocalDefs.push_back(Reg); 1699 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1700 LocalDefsSet.insert(*AI); 1701 } 1702 1703 HasDups = true; 1704 ++TIB; 1705 ++FIB; 1706 } 1707 1708 if (!HasDups) 1709 return false; 1710 1711 MBB->splice(Loc, TBB, TBB->begin(), TIB); 1712 FBB->erase(FBB->begin(), FIB); 1713 1714 // Update livein's. 1715 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) { 1716 unsigned Def = LocalDefs[i]; 1717 if (LocalDefsSet.count(Def)) { 1718 TBB->addLiveIn(Def); 1719 FBB->addLiveIn(Def); 1720 } 1721 } 1722 1723 ++NumHoist; 1724 return true; 1725} 1726