BranchFolding.cpp revision 95ba669e09c30f9aa1a7d754199548b8e6a227ce
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 // Remove the block. 112 MF->erase(MBB); 113} 114 115/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def 116/// followed by terminators, and if the implicitly defined registers are not 117/// used by the terminators, remove those implicit_def's. e.g. 118/// BB1: 119/// r0 = implicit_def 120/// r1 = implicit_def 121/// br 122/// This block can be optimized away later if the implicit instructions are 123/// removed. 124bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { 125 SmallSet<unsigned, 4> ImpDefRegs; 126 MachineBasicBlock::iterator I = MBB->begin(); 127 while (I != MBB->end()) { 128 if (!I->isImplicitDef()) 129 break; 130 unsigned Reg = I->getOperand(0).getReg(); 131 ImpDefRegs.insert(Reg); 132 for (const unsigned *SubRegs = TRI->getSubRegisters(Reg); 133 unsigned SubReg = *SubRegs; ++SubRegs) 134 ImpDefRegs.insert(SubReg); 135 ++I; 136 } 137 if (ImpDefRegs.empty()) 138 return false; 139 140 MachineBasicBlock::iterator FirstTerm = I; 141 while (I != MBB->end()) { 142 if (!TII->isUnpredicatedTerminator(I)) 143 return false; 144 // See if it uses any of the implicitly defined registers. 145 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 146 MachineOperand &MO = I->getOperand(i); 147 if (!MO.isReg() || !MO.isUse()) 148 continue; 149 unsigned Reg = MO.getReg(); 150 if (ImpDefRegs.count(Reg)) 151 return false; 152 } 153 ++I; 154 } 155 156 I = MBB->begin(); 157 while (I != FirstTerm) { 158 MachineInstr *ImpDefMI = &*I; 159 ++I; 160 MBB->erase(ImpDefMI); 161 } 162 163 return true; 164} 165 166/// OptimizeFunction - Perhaps branch folding, tail merging and other 167/// CFG optimizations on the given function. 168bool BranchFolder::OptimizeFunction(MachineFunction &MF, 169 const TargetInstrInfo *tii, 170 const TargetRegisterInfo *tri, 171 MachineModuleInfo *mmi) { 172 if (!tii) return false; 173 174 TII = tii; 175 TRI = tri; 176 MMI = mmi; 177 178 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; 179 180 // Fix CFG. The later algorithms expect it to be right. 181 bool MadeChange = false; 182 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { 183 MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; 184 SmallVector<MachineOperand, 4> Cond; 185 if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) 186 MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 187 MadeChange |= OptimizeImpDefsBlock(MBB); 188 } 189 190 bool MadeChangeThisIteration = true; 191 while (MadeChangeThisIteration) { 192 MadeChangeThisIteration = TailMergeBlocks(MF); 193 MadeChangeThisIteration |= OptimizeBranches(MF); 194 if (EnableHoistCommonCode) 195 MadeChangeThisIteration |= HoistCommonCode(MF); 196 MadeChange |= MadeChangeThisIteration; 197 } 198 199 // See if any jump tables have become dead as the code generator 200 // did its thing. 201 MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); 202 if (JTI == 0) { 203 delete RS; 204 return MadeChange; 205 } 206 207 // Walk the function to find jump tables that are live. 208 BitVector JTIsLive(JTI->getJumpTables().size()); 209 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); 210 BB != E; ++BB) { 211 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); 212 I != E; ++I) 213 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { 214 MachineOperand &Op = I->getOperand(op); 215 if (!Op.isJTI()) continue; 216 217 // Remember that this JT is live. 218 JTIsLive.set(Op.getIndex()); 219 } 220 } 221 222 // Finally, remove dead jump tables. This happens when the 223 // indirect jump was unreachable (and thus deleted). 224 for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) 225 if (!JTIsLive.test(i)) { 226 JTI->RemoveJumpTable(i); 227 MadeChange = true; 228 } 229 230 delete RS; 231 return MadeChange; 232} 233 234//===----------------------------------------------------------------------===// 235// Tail Merging of Blocks 236//===----------------------------------------------------------------------===// 237 238/// HashMachineInstr - Compute a hash value for MI and its operands. 239static unsigned HashMachineInstr(const MachineInstr *MI) { 240 unsigned Hash = MI->getOpcode(); 241 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 242 const MachineOperand &Op = MI->getOperand(i); 243 244 // Merge in bits from the operand if easy. 245 unsigned OperandHash = 0; 246 switch (Op.getType()) { 247 case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; 248 case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; 249 case MachineOperand::MO_MachineBasicBlock: 250 OperandHash = Op.getMBB()->getNumber(); 251 break; 252 case MachineOperand::MO_FrameIndex: 253 case MachineOperand::MO_ConstantPoolIndex: 254 case MachineOperand::MO_JumpTableIndex: 255 OperandHash = Op.getIndex(); 256 break; 257 case MachineOperand::MO_GlobalAddress: 258 case MachineOperand::MO_ExternalSymbol: 259 // Global address / external symbol are too hard, don't bother, but do 260 // pull in the offset. 261 OperandHash = Op.getOffset(); 262 break; 263 default: break; 264 } 265 266 Hash += ((OperandHash << 3) | Op.getType()) << (i&31); 267 } 268 return Hash; 269} 270 271/// HashEndOfMBB - Hash the last instruction in the MBB. 272static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) { 273 MachineBasicBlock::const_iterator I = MBB->end(); 274 if (I == MBB->begin()) 275 return 0; // Empty MBB. 276 277 --I; 278 // Skip debug info so it will not affect codegen. 279 while (I->isDebugValue()) { 280 if (I==MBB->begin()) 281 return 0; // MBB empty except for debug info. 282 --I; 283 } 284 285 return HashMachineInstr(I); 286} 287 288/// ComputeCommonTailLength - Given two machine basic blocks, compute the number 289/// of instructions they actually have in common together at their end. Return 290/// iterators for the first shared instruction in each block. 291static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, 292 MachineBasicBlock *MBB2, 293 MachineBasicBlock::iterator &I1, 294 MachineBasicBlock::iterator &I2) { 295 I1 = MBB1->end(); 296 I2 = MBB2->end(); 297 298 unsigned TailLen = 0; 299 while (I1 != MBB1->begin() && I2 != MBB2->begin()) { 300 --I1; --I2; 301 // Skip debugging pseudos; necessary to avoid changing the code. 302 while (I1->isDebugValue()) { 303 if (I1==MBB1->begin()) { 304 while (I2->isDebugValue()) { 305 if (I2==MBB2->begin()) 306 // I1==DBG at begin; I2==DBG at begin 307 return TailLen; 308 --I2; 309 } 310 ++I2; 311 // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin 312 return TailLen; 313 } 314 --I1; 315 } 316 // I1==first (untested) non-DBG preceding known match 317 while (I2->isDebugValue()) { 318 if (I2==MBB2->begin()) { 319 ++I1; 320 // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin 321 return TailLen; 322 } 323 --I2; 324 } 325 // I1, I2==first (untested) non-DBGs preceding known match 326 if (!I1->isIdenticalTo(I2) || 327 // FIXME: This check is dubious. It's used to get around a problem where 328 // people incorrectly expect inline asm directives to remain in the same 329 // relative order. This is untenable because normal compiler 330 // optimizations (like this one) may reorder and/or merge these 331 // directives. 332 I1->isInlineAsm()) { 333 ++I1; ++I2; 334 break; 335 } 336 ++TailLen; 337 } 338 // Back past possible debugging pseudos at beginning of block. This matters 339 // when one block differs from the other only by whether debugging pseudos 340 // are present at the beginning. (This way, the various checks later for 341 // I1==MBB1->begin() work as expected.) 342 if (I1 == MBB1->begin() && I2 != MBB2->begin()) { 343 --I2; 344 while (I2->isDebugValue()) { 345 if (I2 == MBB2->begin()) { 346 return TailLen; 347 } 348 --I2; 349 } 350 ++I2; 351 } 352 if (I2 == MBB2->begin() && I1 != MBB1->begin()) { 353 --I1; 354 while (I1->isDebugValue()) { 355 if (I1 == MBB1->begin()) 356 return TailLen; 357 --I1; 358 } 359 ++I1; 360 } 361 return TailLen; 362} 363 364/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 365/// after it, replacing it with an unconditional branch to NewDest. 366void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 367 MachineBasicBlock *NewDest) { 368 TII->ReplaceTailWithBranchTo(OldInst, NewDest); 369 ++NumTailMerge; 370} 371 372/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 373/// MBB so that the part before the iterator falls into the part starting at the 374/// iterator. This returns the new MBB. 375MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 376 MachineBasicBlock::iterator BBI1) { 377 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1)) 378 return 0; 379 380 MachineFunction &MF = *CurMBB.getParent(); 381 382 // Create the fall-through block. 383 MachineFunction::iterator MBBI = &CurMBB; 384 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 385 CurMBB.getParent()->insert(++MBBI, NewMBB); 386 387 // Move all the successors of this block to the specified block. 388 NewMBB->transferSuccessors(&CurMBB); 389 390 // Add an edge from CurMBB to NewMBB for the fall-through. 391 CurMBB.addSuccessor(NewMBB); 392 393 // Splice the code over. 394 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 395 396 // For targets that use the register scavenger, we must maintain LiveIns. 397 if (RS) { 398 RS->enterBasicBlock(&CurMBB); 399 if (!CurMBB.empty()) 400 RS->forward(prior(CurMBB.end())); 401 BitVector RegsLiveAtExit(TRI->getNumRegs()); 402 RS->getRegsUsed(RegsLiveAtExit, false); 403 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 404 if (RegsLiveAtExit[i]) 405 NewMBB->addLiveIn(i); 406 } 407 408 return NewMBB; 409} 410 411/// EstimateRuntime - Make a rough estimate for how long it will take to run 412/// the specified code. 413static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 414 MachineBasicBlock::iterator E) { 415 unsigned Time = 0; 416 for (; I != E; ++I) { 417 if (I->isDebugValue()) 418 continue; 419 const TargetInstrDesc &TID = I->getDesc(); 420 if (TID.isCall()) 421 Time += 10; 422 else if (TID.mayLoad() || TID.mayStore()) 423 Time += 2; 424 else 425 ++Time; 426 } 427 return Time; 428} 429 430// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 431// branches temporarily for tail merging). In the case where CurMBB ends 432// with a conditional branch to the next block, optimize by reversing the 433// test and conditionally branching to SuccMBB instead. 434static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 435 const TargetInstrInfo *TII) { 436 MachineFunction *MF = CurMBB->getParent(); 437 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB)); 438 MachineBasicBlock *TBB = 0, *FBB = 0; 439 SmallVector<MachineOperand, 4> Cond; 440 DebugLoc dl; // FIXME: this is nowhere 441 if (I != MF->end() && 442 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 443 MachineBasicBlock *NextBB = I; 444 if (TBB == NextBB && !Cond.empty() && !FBB) { 445 if (!TII->ReverseBranchCondition(Cond)) { 446 TII->RemoveBranch(*CurMBB); 447 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl); 448 return; 449 } 450 } 451 } 452 TII->InsertBranch(*CurMBB, SuccBB, NULL, 453 SmallVector<MachineOperand, 0>(), dl); 454} 455 456bool 457BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 458 if (getHash() < o.getHash()) 459 return true; 460 else if (getHash() > o.getHash()) 461 return false; 462 else if (getBlock()->getNumber() < o.getBlock()->getNumber()) 463 return true; 464 else if (getBlock()->getNumber() > o.getBlock()->getNumber()) 465 return false; 466 else { 467 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 468 // an object with itself. 469#ifndef _GLIBCXX_DEBUG 470 llvm_unreachable("Predecessor appears twice"); 471#endif 472 return false; 473 } 474} 475 476/// CountTerminators - Count the number of terminators in the given 477/// block and set I to the position of the first non-terminator, if there 478/// is one, or MBB->end() otherwise. 479static unsigned CountTerminators(MachineBasicBlock *MBB, 480 MachineBasicBlock::iterator &I) { 481 I = MBB->end(); 482 unsigned NumTerms = 0; 483 for (;;) { 484 if (I == MBB->begin()) { 485 I = MBB->end(); 486 break; 487 } 488 --I; 489 if (!I->getDesc().isTerminator()) break; 490 ++NumTerms; 491 } 492 return NumTerms; 493} 494 495/// ProfitableToMerge - Check if two machine basic blocks have a common tail 496/// and decide if it would be profitable to merge those tails. Return the 497/// length of the common tail and iterators to the first common instruction 498/// in each block. 499static bool ProfitableToMerge(MachineBasicBlock *MBB1, 500 MachineBasicBlock *MBB2, 501 unsigned minCommonTailLength, 502 unsigned &CommonTailLen, 503 MachineBasicBlock::iterator &I1, 504 MachineBasicBlock::iterator &I2, 505 MachineBasicBlock *SuccBB, 506 MachineBasicBlock *PredBB) { 507 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 508 if (CommonTailLen == 0) 509 return false; 510 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber() 511 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen 512 << '\n'); 513 514 // It's almost always profitable to merge any number of non-terminator 515 // instructions with the block that falls through into the common successor. 516 if (MBB1 == PredBB || MBB2 == PredBB) { 517 MachineBasicBlock::iterator I; 518 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 519 if (CommonTailLen > NumTerms) 520 return true; 521 } 522 523 // If one of the blocks can be completely merged and happens to be in 524 // a position where the other could fall through into it, merge any number 525 // of instructions, because it can be done without a branch. 526 // TODO: If the blocks are not adjacent, move one of them so that they are? 527 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 528 return true; 529 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 530 return true; 531 532 // If both blocks have an unconditional branch temporarily stripped out, 533 // count that as an additional common instruction for the following 534 // heuristics. 535 unsigned EffectiveTailLen = CommonTailLen; 536 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 537 !MBB1->back().getDesc().isBarrier() && 538 !MBB2->back().getDesc().isBarrier()) 539 ++EffectiveTailLen; 540 541 // Check if the common tail is long enough to be worthwhile. 542 if (EffectiveTailLen >= minCommonTailLength) 543 return true; 544 545 // If we are optimizing for code size, 2 instructions in common is enough if 546 // we don't have to split a block. At worst we will be introducing 1 new 547 // branch instruction, which is likely to be smaller than the 2 548 // instructions that would be deleted in the merge. 549 MachineFunction *MF = MBB1->getParent(); 550 if (EffectiveTailLen >= 2 && 551 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 552 (I1 == MBB1->begin() || I2 == MBB2->begin())) 553 return true; 554 555 return false; 556} 557 558/// ComputeSameTails - Look through all the blocks in MergePotentials that have 559/// hash CurHash (guaranteed to match the last element). Build the vector 560/// SameTails of all those that have the (same) largest number of instructions 561/// in common of any pair of these blocks. SameTails entries contain an 562/// iterator into MergePotentials (from which the MachineBasicBlock can be 563/// found) and a MachineBasicBlock::iterator into that MBB indicating the 564/// instruction where the matching code sequence begins. 565/// Order of elements in SameTails is the reverse of the order in which 566/// those blocks appear in MergePotentials (where they are not necessarily 567/// consecutive). 568unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 569 unsigned minCommonTailLength, 570 MachineBasicBlock *SuccBB, 571 MachineBasicBlock *PredBB) { 572 unsigned maxCommonTailLength = 0U; 573 SameTails.clear(); 574 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 575 MPIterator HighestMPIter = prior(MergePotentials.end()); 576 for (MPIterator CurMPIter = prior(MergePotentials.end()), 577 B = MergePotentials.begin(); 578 CurMPIter != B && CurMPIter->getHash() == CurHash; 579 --CurMPIter) { 580 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 581 unsigned CommonTailLen; 582 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 583 minCommonTailLength, 584 CommonTailLen, TrialBBI1, TrialBBI2, 585 SuccBB, PredBB)) { 586 if (CommonTailLen > maxCommonTailLength) { 587 SameTails.clear(); 588 maxCommonTailLength = CommonTailLen; 589 HighestMPIter = CurMPIter; 590 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 591 } 592 if (HighestMPIter == CurMPIter && 593 CommonTailLen == maxCommonTailLength) 594 SameTails.push_back(SameTailElt(I, TrialBBI2)); 595 } 596 if (I == B) 597 break; 598 } 599 } 600 return maxCommonTailLength; 601} 602 603/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 604/// MergePotentials, restoring branches at ends of blocks as appropriate. 605void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 606 MachineBasicBlock *SuccBB, 607 MachineBasicBlock *PredBB) { 608 MPIterator CurMPIter, B; 609 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 610 CurMPIter->getHash() == CurHash; 611 --CurMPIter) { 612 // Put the unconditional branch back, if we need one. 613 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 614 if (SuccBB && CurMBB != PredBB) 615 FixTail(CurMBB, SuccBB, TII); 616 if (CurMPIter == B) 617 break; 618 } 619 if (CurMPIter->getHash() != CurHash) 620 CurMPIter++; 621 MergePotentials.erase(CurMPIter, MergePotentials.end()); 622} 623 624/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 625/// only of the common tail. Create a block that does by splitting one. 626bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 627 unsigned maxCommonTailLength, 628 unsigned &commonTailIndex) { 629 commonTailIndex = 0; 630 unsigned TimeEstimate = ~0U; 631 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 632 // Use PredBB if possible; that doesn't require a new branch. 633 if (SameTails[i].getBlock() == PredBB) { 634 commonTailIndex = i; 635 break; 636 } 637 // Otherwise, make a (fairly bogus) choice based on estimate of 638 // how long it will take the various blocks to execute. 639 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 640 SameTails[i].getTailStartPos()); 641 if (t <= TimeEstimate) { 642 TimeEstimate = t; 643 commonTailIndex = i; 644 } 645 } 646 647 MachineBasicBlock::iterator BBI = 648 SameTails[commonTailIndex].getTailStartPos(); 649 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 650 651 // If the common tail includes any debug info we will take it pretty 652 // randomly from one of the inputs. Might be better to remove it? 653 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 654 << maxCommonTailLength); 655 656 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 657 if (!newMBB) { 658 DEBUG(dbgs() << "... failed!"); 659 return false; 660 } 661 662 SameTails[commonTailIndex].setBlock(newMBB); 663 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 664 665 // If we split PredBB, newMBB is the new predecessor. 666 if (PredBB == MBB) 667 PredBB = newMBB; 668 669 return true; 670} 671 672// See if any of the blocks in MergePotentials (which all have a common single 673// successor, or all have no successor) can be tail-merged. If there is a 674// successor, any blocks in MergePotentials that are not tail-merged and 675// are not immediately before Succ must have an unconditional branch to 676// Succ added (but the predecessor/successor lists need no adjustment). 677// The lone predecessor of Succ that falls through into Succ, 678// if any, is given in PredBB. 679 680bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 681 MachineBasicBlock *PredBB) { 682 bool MadeChange = false; 683 684 // Except for the special cases below, tail-merge if there are at least 685 // this many instructions in common. 686 unsigned minCommonTailLength = TailMergeSize; 687 688 DEBUG(dbgs() << "\nTryTailMergeBlocks: "; 689 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 690 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 691 << (i == e-1 ? "" : ", "); 692 dbgs() << "\n"; 693 if (SuccBB) { 694 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 695 if (PredBB) 696 dbgs() << " which has fall-through from BB#" 697 << PredBB->getNumber() << "\n"; 698 } 699 dbgs() << "Looking for common tails of at least " 700 << minCommonTailLength << " instruction" 701 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 702 ); 703 704 // Sort by hash value so that blocks with identical end sequences sort 705 // together. 706 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 707 708 // Walk through equivalence sets looking for actual exact matches. 709 while (MergePotentials.size() > 1) { 710 unsigned CurHash = MergePotentials.back().getHash(); 711 712 // Build SameTails, identifying the set of blocks with this hash code 713 // and with the maximum number of instructions in common. 714 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 715 minCommonTailLength, 716 SuccBB, PredBB); 717 718 // If we didn't find any pair that has at least minCommonTailLength 719 // instructions in common, remove all blocks with this hash code and retry. 720 if (SameTails.empty()) { 721 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 722 continue; 723 } 724 725 // If one of the blocks is the entire common tail (and not the entry 726 // block, which we can't jump to), we can treat all blocks with this same 727 // tail at once. Use PredBB if that is one of the possibilities, as that 728 // will not introduce any extra branches. 729 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 730 getParent()->begin(); 731 unsigned commonTailIndex = SameTails.size(); 732 // If there are two blocks, check to see if one can be made to fall through 733 // into the other. 734 if (SameTails.size() == 2 && 735 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 736 SameTails[1].tailIsWholeBlock()) 737 commonTailIndex = 1; 738 else if (SameTails.size() == 2 && 739 SameTails[1].getBlock()->isLayoutSuccessor( 740 SameTails[0].getBlock()) && 741 SameTails[0].tailIsWholeBlock()) 742 commonTailIndex = 0; 743 else { 744 // Otherwise just pick one, favoring the fall-through predecessor if 745 // there is one. 746 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 747 MachineBasicBlock *MBB = SameTails[i].getBlock(); 748 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 749 continue; 750 if (MBB == PredBB) { 751 commonTailIndex = i; 752 break; 753 } 754 if (SameTails[i].tailIsWholeBlock()) 755 commonTailIndex = i; 756 } 757 } 758 759 if (commonTailIndex == SameTails.size() || 760 (SameTails[commonTailIndex].getBlock() == PredBB && 761 !SameTails[commonTailIndex].tailIsWholeBlock())) { 762 // None of the blocks consist entirely of the common tail. 763 // Split a block so that one does. 764 if (!CreateCommonTailOnlyBlock(PredBB, 765 maxCommonTailLength, commonTailIndex)) { 766 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 767 continue; 768 } 769 } 770 771 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 772 // MBB is common tail. Adjust all other BB's to jump to this one. 773 // Traversal must be forwards so erases work. 774 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber() 775 << " for "); 776 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 777 if (commonTailIndex == i) 778 continue; 779 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() 780 << (i == e-1 ? "" : ", ")); 781 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 782 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 783 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 784 MergePotentials.erase(SameTails[i].getMPIter()); 785 } 786 DEBUG(dbgs() << "\n"); 787 // We leave commonTailIndex in the worklist in case there are other blocks 788 // that match it with a smaller number of instructions. 789 MadeChange = true; 790 } 791 return MadeChange; 792} 793 794bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 795 796 if (!EnableTailMerge) return false; 797 798 bool MadeChange = false; 799 800 // First find blocks with no successors. 801 MergePotentials.clear(); 802 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 803 if (I->succ_empty()) 804 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I)); 805 } 806 807 // See if we can do any tail merging on those. 808 if (MergePotentials.size() < TailMergeThreshold && 809 MergePotentials.size() >= 2) 810 MadeChange |= TryTailMergeBlocks(NULL, NULL); 811 812 // Look at blocks (IBB) with multiple predecessors (PBB). 813 // We change each predecessor to a canonical form, by 814 // (1) temporarily removing any unconditional branch from the predecessor 815 // to IBB, and 816 // (2) alter conditional branches so they branch to the other block 817 // not IBB; this may require adding back an unconditional branch to IBB 818 // later, where there wasn't one coming in. E.g. 819 // Bcc IBB 820 // fallthrough to QBB 821 // here becomes 822 // Bncc QBB 823 // with a conceptual B to IBB after that, which never actually exists. 824 // With those changes, we see whether the predecessors' tails match, 825 // and merge them if so. We change things out of canonical form and 826 // back to the way they were later in the process. (OptimizeBranches 827 // would undo some of this, but we can't use it, because we'd get into 828 // a compile-time infinite loop repeatedly doing and undoing the same 829 // transformations.) 830 831 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 832 I != E; ++I) { 833 if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { 834 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 835 MachineBasicBlock *IBB = I; 836 MachineBasicBlock *PredBB = prior(I); 837 MergePotentials.clear(); 838 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 839 E2 = I->pred_end(); 840 P != E2; ++P) { 841 MachineBasicBlock *PBB = *P; 842 // Skip blocks that loop to themselves, can't tail merge these. 843 if (PBB == IBB) 844 continue; 845 // Visit each predecessor only once. 846 if (!UniquePreds.insert(PBB)) 847 continue; 848 MachineBasicBlock *TBB = 0, *FBB = 0; 849 SmallVector<MachineOperand, 4> Cond; 850 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 851 // Failing case: IBB is the target of a cbr, and 852 // we cannot reverse the branch. 853 SmallVector<MachineOperand, 4> NewCond(Cond); 854 if (!Cond.empty() && TBB == IBB) { 855 if (TII->ReverseBranchCondition(NewCond)) 856 continue; 857 // This is the QBB case described above 858 if (!FBB) 859 FBB = llvm::next(MachineFunction::iterator(PBB)); 860 } 861 // Failing case: the only way IBB can be reached from PBB is via 862 // exception handling. Happens for landing pads. Would be nice 863 // to have a bit in the edge so we didn't have to do all this. 864 if (IBB->isLandingPad()) { 865 MachineFunction::iterator IP = PBB; IP++; 866 MachineBasicBlock *PredNextBB = NULL; 867 if (IP != MF.end()) 868 PredNextBB = IP; 869 if (TBB == NULL) { 870 if (IBB != PredNextBB) // fallthrough 871 continue; 872 } else if (FBB) { 873 if (TBB != IBB && FBB != IBB) // cbr then ubr 874 continue; 875 } else if (Cond.empty()) { 876 if (TBB != IBB) // ubr 877 continue; 878 } else { 879 if (TBB != IBB && IBB != PredNextBB) // cbr 880 continue; 881 } 882 } 883 // Remove the unconditional branch at the end, if any. 884 if (TBB && (Cond.empty() || FBB)) { 885 DebugLoc dl; // FIXME: this is nowhere 886 TII->RemoveBranch(*PBB); 887 if (!Cond.empty()) 888 // reinsert conditional branch only, for now 889 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl); 890 } 891 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P)); 892 } 893 } 894 if (MergePotentials.size() >= 2) 895 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 896 // Reinsert an unconditional branch if needed. 897 // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks. 898 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 899 if (MergePotentials.size() == 1 && 900 MergePotentials.begin()->getBlock() != PredBB) 901 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 902 } 903 } 904 return MadeChange; 905} 906 907//===----------------------------------------------------------------------===// 908// Branch Optimization 909//===----------------------------------------------------------------------===// 910 911bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 912 bool MadeChange = false; 913 914 // Make sure blocks are numbered in order 915 MF.RenumberBlocks(); 916 917 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 918 I != E; ) { 919 MachineBasicBlock *MBB = I++; 920 MadeChange |= OptimizeBlock(MBB); 921 922 // If it is dead, remove it. 923 if (MBB->pred_empty()) { 924 RemoveDeadBlock(MBB); 925 MadeChange = true; 926 ++NumDeadBlocks; 927 } 928 } 929 return MadeChange; 930} 931 932// Blocks should be considered empty if they contain only debug info; 933// else the debug info would affect codegen. 934static bool IsEmptyBlock(MachineBasicBlock *MBB) { 935 if (MBB->empty()) 936 return true; 937 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end(); 938 MBBI!=MBBE; ++MBBI) { 939 if (!MBBI->isDebugValue()) 940 return false; 941 } 942 return true; 943} 944 945// Blocks with only debug info and branches should be considered the same 946// as blocks with only branches. 947static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { 948 MachineBasicBlock::iterator MBBI, MBBE; 949 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) { 950 if (!MBBI->isDebugValue()) 951 break; 952 } 953 return (MBBI->getDesc().isBranch()); 954} 955 956/// IsBetterFallthrough - Return true if it would be clearly better to 957/// fall-through to MBB1 than to fall through into MBB2. This has to return 958/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 959/// result in infinite loops. 960static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 961 MachineBasicBlock *MBB2) { 962 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 963 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 964 // optimize branches that branch to either a return block or an assert block 965 // into a fallthrough to the return. 966 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false; 967 968 // If there is a clear successor ordering we make sure that one block 969 // will fall through to the next 970 if (MBB1->isSuccessor(MBB2)) return true; 971 if (MBB2->isSuccessor(MBB1)) return false; 972 973 // Neither block consists entirely of debug info (per IsEmptyBlock check), 974 // so we needn't test for falling off the beginning here. 975 MachineBasicBlock::iterator MBB1I = --MBB1->end(); 976 while (MBB1I->isDebugValue()) 977 --MBB1I; 978 MachineBasicBlock::iterator MBB2I = --MBB2->end(); 979 while (MBB2I->isDebugValue()) 980 --MBB2I; 981 return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); 982} 983 984/// OptimizeBlock - Analyze and optimize control flow related to the specified 985/// block. This is never called on the entry block. 986bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 987 bool MadeChange = false; 988 MachineFunction &MF = *MBB->getParent(); 989 DebugLoc dl; // FIXME: this is nowhere 990ReoptimizeBlock: 991 992 MachineFunction::iterator FallThrough = MBB; 993 ++FallThrough; 994 995 // If this block is empty, make everyone use its fall-through, not the block 996 // explicitly. Landing pads should not do this since the landing-pad table 997 // points to this block. Blocks with their addresses taken shouldn't be 998 // optimized away. 999 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 1000 // Dead block? Leave for cleanup later. 1001 if (MBB->pred_empty()) return MadeChange; 1002 1003 if (FallThrough == MF.end()) { 1004 // TODO: Simplify preds to not branch here if possible! 1005 } else { 1006 // Rewrite all predecessors of the old block to go to the fallthrough 1007 // instead. 1008 while (!MBB->pred_empty()) { 1009 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1010 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1011 } 1012 // If MBB was the target of a jump table, update jump tables to go to the 1013 // fallthrough instead. 1014 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1015 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough); 1016 MadeChange = true; 1017 } 1018 return MadeChange; 1019 } 1020 1021 // Check to see if we can simplify the terminator of the block before this 1022 // one. 1023 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1024 1025 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1026 SmallVector<MachineOperand, 4> PriorCond; 1027 bool PriorUnAnalyzable = 1028 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1029 if (!PriorUnAnalyzable) { 1030 // If the CFG for the prior block has extra edges, remove them. 1031 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1032 !PriorCond.empty()); 1033 1034 // If the previous branch is conditional and both conditions go to the same 1035 // destination, remove the branch, replacing it with an unconditional one or 1036 // a fall-through. 1037 if (PriorTBB && PriorTBB == PriorFBB) { 1038 TII->RemoveBranch(PrevBB); 1039 PriorCond.clear(); 1040 if (PriorTBB != MBB) 1041 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1042 MadeChange = true; 1043 ++NumBranchOpts; 1044 goto ReoptimizeBlock; 1045 } 1046 1047 // If the previous block unconditionally falls through to this block and 1048 // this block has no other predecessors, move the contents of this block 1049 // into the prior block. This doesn't usually happen when SimplifyCFG 1050 // has been used, but it can happen if tail merging splits a fall-through 1051 // predecessor of a block. 1052 // This has to check PrevBB->succ_size() because EH edges are ignored by 1053 // AnalyzeBranch. 1054 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1055 PrevBB.succ_size() == 1 && 1056 !MBB->hasAddressTaken() && !MBB->isLandingPad()) { 1057 DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1058 << "From MBB: " << *MBB); 1059 // Remove redundant DBG_VALUEs first. 1060 if (PrevBB.begin() != PrevBB.end()) { 1061 MachineBasicBlock::iterator PrevBBIter = PrevBB.end(); 1062 --PrevBBIter; 1063 MachineBasicBlock::iterator MBBIter = MBB->begin(); 1064 // Check if DBG_VALUE at the end of PrevBB is identical to the 1065 // DBG_VALUE at the beginning of MBB. 1066 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end() 1067 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) { 1068 if (!MBBIter->isIdenticalTo(PrevBBIter)) 1069 break; 1070 MachineInstr *DuplicateDbg = MBBIter; 1071 ++MBBIter; -- PrevBBIter; 1072 DuplicateDbg->eraseFromParent(); 1073 } 1074 } 1075 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1076 PrevBB.removeSuccessor(PrevBB.succ_begin());; 1077 assert(PrevBB.succ_empty()); 1078 PrevBB.transferSuccessors(MBB); 1079 MadeChange = true; 1080 return MadeChange; 1081 } 1082 1083 // If the previous branch *only* branches to *this* block (conditional or 1084 // not) remove the branch. 1085 if (PriorTBB == MBB && PriorFBB == 0) { 1086 TII->RemoveBranch(PrevBB); 1087 MadeChange = true; 1088 ++NumBranchOpts; 1089 goto ReoptimizeBlock; 1090 } 1091 1092 // If the prior block branches somewhere else on the condition and here if 1093 // the condition is false, remove the uncond second branch. 1094 if (PriorFBB == MBB) { 1095 TII->RemoveBranch(PrevBB); 1096 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1097 MadeChange = true; 1098 ++NumBranchOpts; 1099 goto ReoptimizeBlock; 1100 } 1101 1102 // If the prior block branches here on true and somewhere else on false, and 1103 // if the branch condition is reversible, reverse the branch to create a 1104 // fall-through. 1105 if (PriorTBB == MBB) { 1106 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1107 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1108 TII->RemoveBranch(PrevBB); 1109 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl); 1110 MadeChange = true; 1111 ++NumBranchOpts; 1112 goto ReoptimizeBlock; 1113 } 1114 } 1115 1116 // If this block has no successors (e.g. it is a return block or ends with 1117 // a call to a no-return function like abort or __cxa_throw) and if the pred 1118 // falls through into this block, and if it would otherwise fall through 1119 // into the block after this, move this block to the end of the function. 1120 // 1121 // We consider it more likely that execution will stay in the function (e.g. 1122 // due to loops) than it is to exit it. This asserts in loops etc, moving 1123 // the assert condition out of the loop body. 1124 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1125 MachineFunction::iterator(PriorTBB) == FallThrough && 1126 !MBB->canFallThrough()) { 1127 bool DoTransform = true; 1128 1129 // We have to be careful that the succs of PredBB aren't both no-successor 1130 // blocks. If neither have successors and if PredBB is the second from 1131 // last block in the function, we'd just keep swapping the two blocks for 1132 // last. Only do the swap if one is clearly better to fall through than 1133 // the other. 1134 if (FallThrough == --MF.end() && 1135 !IsBetterFallthrough(PriorTBB, MBB)) 1136 DoTransform = false; 1137 1138 if (DoTransform) { 1139 // Reverse the branch so we will fall through on the previous true cond. 1140 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1141 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1142 DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1143 << "To make fallthrough to: " << *PriorTBB << "\n"); 1144 1145 TII->RemoveBranch(PrevBB); 1146 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl); 1147 1148 // Move this block to the end of the function. 1149 MBB->moveAfter(--MF.end()); 1150 MadeChange = true; 1151 ++NumBranchOpts; 1152 return MadeChange; 1153 } 1154 } 1155 } 1156 } 1157 1158 // Analyze the branch in the current block. 1159 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1160 SmallVector<MachineOperand, 4> CurCond; 1161 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1162 if (!CurUnAnalyzable) { 1163 // If the CFG for the prior block has extra edges, remove them. 1164 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1165 1166 // If this is a two-way branch, and the FBB branches to this block, reverse 1167 // the condition so the single-basic-block loop is faster. Instead of: 1168 // Loop: xxx; jcc Out; jmp Loop 1169 // we want: 1170 // Loop: xxx; jncc Loop; jmp Out 1171 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1172 SmallVector<MachineOperand, 4> NewCond(CurCond); 1173 if (!TII->ReverseBranchCondition(NewCond)) { 1174 TII->RemoveBranch(*MBB); 1175 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); 1176 MadeChange = true; 1177 ++NumBranchOpts; 1178 goto ReoptimizeBlock; 1179 } 1180 } 1181 1182 // If this branch is the only thing in its block, see if we can forward 1183 // other blocks across it. 1184 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1185 IsBranchOnlyBlock(MBB) && CurTBB != MBB && 1186 !MBB->hasAddressTaken()) { 1187 // This block may contain just an unconditional branch. Because there can 1188 // be 'non-branch terminators' in the block, try removing the branch and 1189 // then seeing if the block is empty. 1190 TII->RemoveBranch(*MBB); 1191 // If the only things remaining in the block are debug info, remove these 1192 // as well, so this will behave the same as an empty block in non-debug 1193 // mode. 1194 if (!MBB->empty()) { 1195 bool NonDebugInfoFound = false; 1196 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 1197 I != E; ++I) { 1198 if (!I->isDebugValue()) { 1199 NonDebugInfoFound = true; 1200 break; 1201 } 1202 } 1203 if (!NonDebugInfoFound) 1204 // Make the block empty, losing the debug info (we could probably 1205 // improve this in some cases.) 1206 MBB->erase(MBB->begin(), MBB->end()); 1207 } 1208 // If this block is just an unconditional branch to CurTBB, we can 1209 // usually completely eliminate the block. The only case we cannot 1210 // completely eliminate the block is when the block before this one 1211 // falls through into MBB and we can't understand the prior block's branch 1212 // condition. 1213 if (MBB->empty()) { 1214 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1215 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1216 !PrevBB.isSuccessor(MBB)) { 1217 // If the prior block falls through into us, turn it into an 1218 // explicit branch to us to make updates simpler. 1219 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1220 PriorTBB != MBB && PriorFBB != MBB) { 1221 if (PriorTBB == 0) { 1222 assert(PriorCond.empty() && PriorFBB == 0 && 1223 "Bad branch analysis"); 1224 PriorTBB = MBB; 1225 } else { 1226 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1227 PriorFBB = MBB; 1228 } 1229 TII->RemoveBranch(PrevBB); 1230 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, dl); 1231 } 1232 1233 // Iterate through all the predecessors, revectoring each in-turn. 1234 size_t PI = 0; 1235 bool DidChange = false; 1236 bool HasBranchToSelf = false; 1237 while(PI != MBB->pred_size()) { 1238 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1239 if (PMBB == MBB) { 1240 // If this block has an uncond branch to itself, leave it. 1241 ++PI; 1242 HasBranchToSelf = true; 1243 } else { 1244 DidChange = true; 1245 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1246 // If this change resulted in PMBB ending in a conditional 1247 // branch where both conditions go to the same destination, 1248 // change this to an unconditional branch (and fix the CFG). 1249 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1250 SmallVector<MachineOperand, 4> NewCurCond; 1251 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1252 NewCurFBB, NewCurCond, true); 1253 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1254 TII->RemoveBranch(*PMBB); 1255 NewCurCond.clear(); 1256 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, dl); 1257 MadeChange = true; 1258 ++NumBranchOpts; 1259 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1260 } 1261 } 1262 } 1263 1264 // Change any jumptables to go to the new MBB. 1265 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1266 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); 1267 if (DidChange) { 1268 ++NumBranchOpts; 1269 MadeChange = true; 1270 if (!HasBranchToSelf) return MadeChange; 1271 } 1272 } 1273 } 1274 1275 // Add the branch back if the block is more than just an uncond branch. 1276 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl); 1277 } 1278 } 1279 1280 // If the prior block doesn't fall through into this block, and if this 1281 // block doesn't fall through into some other block, see if we can find a 1282 // place to move this block where a fall-through will happen. 1283 if (!PrevBB.canFallThrough()) { 1284 1285 // Now we know that there was no fall-through into this block, check to 1286 // see if it has a fall-through into its successor. 1287 bool CurFallsThru = MBB->canFallThrough(); 1288 1289 if (!MBB->isLandingPad()) { 1290 // Check all the predecessors of this block. If one of them has no fall 1291 // throughs, move this block right after it. 1292 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1293 E = MBB->pred_end(); PI != E; ++PI) { 1294 // Analyze the branch at the end of the pred. 1295 MachineBasicBlock *PredBB = *PI; 1296 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1297 MachineBasicBlock *PredTBB = 0, *PredFBB = 0; 1298 SmallVector<MachineOperand, 4> PredCond; 1299 if (PredBB != MBB && !PredBB->canFallThrough() && 1300 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1301 && (!CurFallsThru || !CurTBB || !CurFBB) 1302 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1303 // If the current block doesn't fall through, just move it. 1304 // If the current block can fall through and does not end with a 1305 // conditional branch, we need to append an unconditional jump to 1306 // the (current) next block. To avoid a possible compile-time 1307 // infinite loop, move blocks only backward in this case. 1308 // Also, if there are already 2 branches here, we cannot add a third; 1309 // this means we have the case 1310 // Bcc next 1311 // B elsewhere 1312 // next: 1313 if (CurFallsThru) { 1314 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); 1315 CurCond.clear(); 1316 TII->InsertBranch(*MBB, NextBB, 0, CurCond, dl); 1317 } 1318 MBB->moveAfter(PredBB); 1319 MadeChange = true; 1320 goto ReoptimizeBlock; 1321 } 1322 } 1323 } 1324 1325 if (!CurFallsThru) { 1326 // Check all successors to see if we can move this block before it. 1327 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1328 E = MBB->succ_end(); SI != E; ++SI) { 1329 // Analyze the branch at the end of the block before the succ. 1330 MachineBasicBlock *SuccBB = *SI; 1331 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1332 1333 // If this block doesn't already fall-through to that successor, and if 1334 // the succ doesn't already have a block that can fall through into it, 1335 // and if the successor isn't an EH destination, we can arrange for the 1336 // fallthrough to happen. 1337 if (SuccBB != MBB && &*SuccPrev != MBB && 1338 !SuccPrev->canFallThrough() && !CurUnAnalyzable && 1339 !SuccBB->isLandingPad()) { 1340 MBB->moveBefore(SuccBB); 1341 MadeChange = true; 1342 goto ReoptimizeBlock; 1343 } 1344 } 1345 1346 // Okay, there is no really great place to put this block. If, however, 1347 // the block before this one would be a fall-through if this block were 1348 // removed, move this block to the end of the function. 1349 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0; 1350 SmallVector<MachineOperand, 4> PrevCond; 1351 if (FallThrough != MF.end() && 1352 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1353 PrevBB.isSuccessor(FallThrough)) { 1354 MBB->moveAfter(--MF.end()); 1355 MadeChange = true; 1356 return MadeChange; 1357 } 1358 } 1359 } 1360 1361 return MadeChange; 1362} 1363 1364//===----------------------------------------------------------------------===// 1365// Hoist Common Code 1366//===----------------------------------------------------------------------===// 1367 1368/// HoistCommonCode - Hoist common instruction sequences at the start of basic 1369/// blocks to their common predecessor. 1370bool BranchFolder::HoistCommonCode(MachineFunction &MF) { 1371 bool MadeChange = false; 1372 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) { 1373 MachineBasicBlock *MBB = I++; 1374 MadeChange |= HoistCommonCodeInSuccs(MBB); 1375 } 1376 1377 return MadeChange; 1378} 1379 1380/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 1381/// its 'true' successor. 1382static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 1383 MachineBasicBlock *TrueBB) { 1384 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 1385 E = BB->succ_end(); SI != E; ++SI) { 1386 MachineBasicBlock *SuccBB = *SI; 1387 if (SuccBB != TrueBB) 1388 return SuccBB; 1389 } 1390 return NULL; 1391} 1392 1393/// findHoistingInsertPosAndDeps - Find the location to move common instructions 1394/// in successors to. The location is ususally just before the terminator, 1395/// however if the terminator is a conditional branch and its previous 1396/// instruction is the flag setting instruction, the previous instruction is 1397/// the preferred location. This function also gathers uses and defs of the 1398/// instructions from the insertion point to the end of the block. The data is 1399/// used by HoistCommonCodeInSuccs to ensure safety. 1400static 1401MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB, 1402 const TargetInstrInfo *TII, 1403 const TargetRegisterInfo *TRI, 1404 SmallSet<unsigned,4> &Uses, 1405 SmallSet<unsigned,4> &Defs) { 1406 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator(); 1407 if (!TII->isUnpredicatedTerminator(Loc)) 1408 return MBB->end(); 1409 1410 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) { 1411 const MachineOperand &MO = Loc->getOperand(i); 1412 if (!MO.isReg()) 1413 continue; 1414 unsigned Reg = MO.getReg(); 1415 if (!Reg) 1416 continue; 1417 if (MO.isUse()) { 1418 Uses.insert(Reg); 1419 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1420 Uses.insert(*AS); 1421 } else if (!MO.isDead()) 1422 // Don't try to hoist code in the rare case the terminator defines a 1423 // register that is later used. 1424 return MBB->end(); 1425 } 1426 1427 if (Uses.empty()) 1428 return Loc; 1429 if (Loc == MBB->begin()) 1430 return MBB->end(); 1431 1432 // The terminator is probably a conditional branch, try not to separate the 1433 // branch from condition setting instruction. 1434 MachineBasicBlock::iterator PI = Loc; 1435 --PI; 1436 while (PI != MBB->begin() && Loc->isDebugValue()) 1437 --PI; 1438 1439 bool IsDef = false; 1440 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) { 1441 const MachineOperand &MO = PI->getOperand(i); 1442 if (!MO.isReg() || MO.isUse()) 1443 continue; 1444 unsigned Reg = MO.getReg(); 1445 if (!Reg) 1446 continue; 1447 if (Uses.count(Reg)) 1448 IsDef = true; 1449 } 1450 if (!IsDef) 1451 // The condition setting instruction is not just before the conditional 1452 // branch. 1453 return Loc; 1454 1455 // Be conservative, don't insert instruction above something that may have 1456 // side-effects. And since it's potentially bad to separate flag setting 1457 // instruction from the conditional branch, just abort the optimization 1458 // completely. 1459 // Also avoid moving code above predicated instruction since it's hard to 1460 // reason about register liveness with predicated instruction. 1461 bool DontMoveAcrossStore = true; 1462 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) || 1463 TII->isPredicated(PI)) 1464 return MBB->end(); 1465 1466 1467 // Find out what registers are live. Note this routine is ignoring other live 1468 // registers which are only used by instructions in successor blocks. 1469 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) { 1470 const MachineOperand &MO = PI->getOperand(i); 1471 if (!MO.isReg()) 1472 continue; 1473 unsigned Reg = MO.getReg(); 1474 if (!Reg) 1475 continue; 1476 if (MO.isUse()) { 1477 Uses.insert(Reg); 1478 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1479 Uses.insert(*AS); 1480 } else { 1481 if (Uses.count(Reg)) { 1482 Uses.erase(Reg); 1483 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1484 Uses.erase(*SR); // Use getSubRegisters to be conservative 1485 } 1486 Defs.insert(Reg); 1487 for (const unsigned *AS = TRI->getAliasSet(Reg); *AS; ++AS) 1488 Defs.insert(*AS); 1489 } 1490 } 1491 1492 return PI; 1493} 1494 1495/// HoistCommonCodeInSuccs - If the successors of MBB has common instruction 1496/// sequence at the start of the function, move the instructions before MBB 1497/// terminator if it's legal. 1498bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) { 1499 MachineBasicBlock *TBB = 0, *FBB = 0; 1500 SmallVector<MachineOperand, 4> Cond; 1501 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty()) 1502 return false; 1503 1504 if (!FBB) FBB = findFalseBlock(MBB, TBB); 1505 if (!FBB) 1506 // Malformed bcc? True and false blocks are the same? 1507 return false; 1508 1509 // Restrict the optimization to cases where MBB is the only predecessor, 1510 // it is an obvious win. 1511 if (TBB->pred_size() > 1 || FBB->pred_size() > 1) 1512 return false; 1513 1514 // Find a suitable position to hoist the common instructions to. Also figure 1515 // out which registers are used or defined by instructions from the insertion 1516 // point to the end of the block. 1517 SmallSet<unsigned, 4> Uses, Defs; 1518 MachineBasicBlock::iterator Loc = 1519 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs); 1520 if (Loc == MBB->end()) 1521 return false; 1522 1523 bool HasDups = false; 1524 SmallVector<unsigned, 4> LocalDefs; 1525 SmallSet<unsigned, 4> LocalDefsSet; 1526 MachineBasicBlock::iterator TIB = TBB->begin(); 1527 MachineBasicBlock::iterator FIB = FBB->begin(); 1528 MachineBasicBlock::iterator TIE = TBB->end(); 1529 MachineBasicBlock::iterator FIE = FBB->end(); 1530 while (TIB != TIE && FIB != FIE) { 1531 // Skip dbg_value instructions. These do not count. 1532 if (TIB->isDebugValue()) { 1533 while (TIB != TIE && TIB->isDebugValue()) 1534 ++TIB; 1535 if (TIB == TIE) 1536 break; 1537 } 1538 if (FIB->isDebugValue()) { 1539 while (FIB != FIE && FIB->isDebugValue()) 1540 ++FIB; 1541 if (FIB == FIE) 1542 break; 1543 } 1544 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead)) 1545 break; 1546 1547 if (TII->isPredicated(TIB)) 1548 // Hard to reason about register liveness with predicated instruction. 1549 break; 1550 1551 bool IsSafe = true; 1552 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1553 MachineOperand &MO = TIB->getOperand(i); 1554 if (!MO.isReg()) 1555 continue; 1556 unsigned Reg = MO.getReg(); 1557 if (!Reg) 1558 continue; 1559 if (MO.isDef()) { 1560 if (Uses.count(Reg)) { 1561 // Avoid clobbering a register that's used by the instruction at 1562 // the point of insertion. 1563 IsSafe = false; 1564 break; 1565 } 1566 1567 if (Defs.count(Reg) && !MO.isDead()) { 1568 // Don't hoist the instruction if the def would be clobber by the 1569 // instruction at the point insertion. FIXME: This is overly 1570 // conservative. It should be possible to hoist the instructions 1571 // in BB2 in the following example: 1572 // BB1: 1573 // r1, eflag = op1 r2, r3 1574 // brcc eflag 1575 // 1576 // BB2: 1577 // r1 = op2, ... 1578 // = op3, r1<kill> 1579 IsSafe = false; 1580 break; 1581 } 1582 } else if (!LocalDefsSet.count(Reg)) { 1583 if (Defs.count(Reg)) { 1584 // Use is defined by the instruction at the point of insertion. 1585 IsSafe = false; 1586 break; 1587 } 1588 } 1589 } 1590 if (!IsSafe) 1591 break; 1592 1593 bool DontMoveAcrossStore = true; 1594 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore)) 1595 break; 1596 1597 // Track local defs so we can update liveins. 1598 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1599 MachineOperand &MO = TIB->getOperand(i); 1600 if (!MO.isReg()) 1601 continue; 1602 unsigned Reg = MO.getReg(); 1603 if (!Reg) 1604 continue; 1605 if (MO.isDef()) { 1606 if (!MO.isDead()) { 1607 LocalDefs.push_back(Reg); 1608 LocalDefsSet.insert(Reg); 1609 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1610 LocalDefsSet.insert(*SR); 1611 } 1612 } else if (MO.isKill() && LocalDefsSet.count(Reg)) { 1613 LocalDefsSet.erase(Reg); 1614 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1615 LocalDefsSet.erase(*SR); 1616 } 1617 } 1618 1619 HasDups = true;; 1620 ++TIB; 1621 ++FIB; 1622 } 1623 1624 if (!HasDups) 1625 return false; 1626 1627 MBB->splice(Loc, TBB, TBB->begin(), TIB); 1628 FBB->erase(FBB->begin(), FIB); 1629 1630 // Update livein's. 1631 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) { 1632 unsigned Def = LocalDefs[i]; 1633 if (LocalDefsSet.count(Def)) { 1634 TBB->addLiveIn(Def); 1635 FBB->addLiveIn(Def); 1636 } 1637 } 1638 1639 ++NumHoist; 1640 return true; 1641} 1642