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 } 363 --I2; 364 } 365 ++I2; 366 } 367 if (I2 == MBB2->begin() && I1 != MBB1->begin()) { 368 --I1; 369 while (I1->isDebugValue()) { 370 if (I1 == MBB1->begin()) 371 return TailLen; 372 --I1; 373 } 374 ++I1; 375 } 376 return TailLen; 377} 378 379void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB, 380 MachineBasicBlock *NewMBB) { 381 if (RS) { 382 RS->enterBasicBlock(CurMBB); 383 if (!CurMBB->empty()) 384 RS->forward(prior(CurMBB->end())); 385 BitVector RegsLiveAtExit(TRI->getNumRegs()); 386 RS->getRegsUsed(RegsLiveAtExit, false); 387 for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) 388 if (RegsLiveAtExit[i]) 389 NewMBB->addLiveIn(i); 390 } 391} 392 393/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything 394/// after it, replacing it with an unconditional branch to NewDest. 395void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, 396 MachineBasicBlock *NewDest) { 397 MachineBasicBlock *CurMBB = OldInst->getParent(); 398 399 TII->ReplaceTailWithBranchTo(OldInst, NewDest); 400 401 // For targets that use the register scavenger, we must maintain LiveIns. 402 MaintainLiveIns(CurMBB, NewDest); 403 404 ++NumTailMerge; 405} 406 407/// SplitMBBAt - Given a machine basic block and an iterator into it, split the 408/// MBB so that the part before the iterator falls into the part starting at the 409/// iterator. This returns the new MBB. 410MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, 411 MachineBasicBlock::iterator BBI1) { 412 if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1)) 413 return 0; 414 415 MachineFunction &MF = *CurMBB.getParent(); 416 417 // Create the fall-through block. 418 MachineFunction::iterator MBBI = &CurMBB; 419 MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); 420 CurMBB.getParent()->insert(++MBBI, NewMBB); 421 422 // Move all the successors of this block to the specified block. 423 NewMBB->transferSuccessors(&CurMBB); 424 425 // Add an edge from CurMBB to NewMBB for the fall-through. 426 CurMBB.addSuccessor(NewMBB); 427 428 // Splice the code over. 429 NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); 430 431 // For targets that use the register scavenger, we must maintain LiveIns. 432 MaintainLiveIns(&CurMBB, NewMBB); 433 434 return NewMBB; 435} 436 437/// EstimateRuntime - Make a rough estimate for how long it will take to run 438/// the specified code. 439static unsigned EstimateRuntime(MachineBasicBlock::iterator I, 440 MachineBasicBlock::iterator E) { 441 unsigned Time = 0; 442 for (; I != E; ++I) { 443 if (I->isDebugValue()) 444 continue; 445 if (I->isCall()) 446 Time += 10; 447 else if (I->mayLoad() || I->mayStore()) 448 Time += 2; 449 else 450 ++Time; 451 } 452 return Time; 453} 454 455// CurMBB needs to add an unconditional branch to SuccMBB (we removed these 456// branches temporarily for tail merging). In the case where CurMBB ends 457// with a conditional branch to the next block, optimize by reversing the 458// test and conditionally branching to SuccMBB instead. 459static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, 460 const TargetInstrInfo *TII) { 461 MachineFunction *MF = CurMBB->getParent(); 462 MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB)); 463 MachineBasicBlock *TBB = 0, *FBB = 0; 464 SmallVector<MachineOperand, 4> Cond; 465 DebugLoc dl; // FIXME: this is nowhere 466 if (I != MF->end() && 467 !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { 468 MachineBasicBlock *NextBB = I; 469 if (TBB == NextBB && !Cond.empty() && !FBB) { 470 if (!TII->ReverseBranchCondition(Cond)) { 471 TII->RemoveBranch(*CurMBB); 472 TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl); 473 return; 474 } 475 } 476 } 477 TII->InsertBranch(*CurMBB, SuccBB, NULL, 478 SmallVector<MachineOperand, 0>(), dl); 479} 480 481bool 482BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { 483 if (getHash() < o.getHash()) 484 return true; 485 else if (getHash() > o.getHash()) 486 return false; 487 else if (getBlock()->getNumber() < o.getBlock()->getNumber()) 488 return true; 489 else if (getBlock()->getNumber() > o.getBlock()->getNumber()) 490 return false; 491 else { 492 // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing 493 // an object with itself. 494#ifndef _GLIBCXX_DEBUG 495 llvm_unreachable("Predecessor appears twice"); 496#else 497 return false; 498#endif 499 } 500} 501 502/// CountTerminators - Count the number of terminators in the given 503/// block and set I to the position of the first non-terminator, if there 504/// is one, or MBB->end() otherwise. 505static unsigned CountTerminators(MachineBasicBlock *MBB, 506 MachineBasicBlock::iterator &I) { 507 I = MBB->end(); 508 unsigned NumTerms = 0; 509 for (;;) { 510 if (I == MBB->begin()) { 511 I = MBB->end(); 512 break; 513 } 514 --I; 515 if (!I->isTerminator()) break; 516 ++NumTerms; 517 } 518 return NumTerms; 519} 520 521/// ProfitableToMerge - Check if two machine basic blocks have a common tail 522/// and decide if it would be profitable to merge those tails. Return the 523/// length of the common tail and iterators to the first common instruction 524/// in each block. 525static bool ProfitableToMerge(MachineBasicBlock *MBB1, 526 MachineBasicBlock *MBB2, 527 unsigned minCommonTailLength, 528 unsigned &CommonTailLen, 529 MachineBasicBlock::iterator &I1, 530 MachineBasicBlock::iterator &I2, 531 MachineBasicBlock *SuccBB, 532 MachineBasicBlock *PredBB) { 533 CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); 534 if (CommonTailLen == 0) 535 return false; 536 DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber() 537 << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen 538 << '\n'); 539 540 // It's almost always profitable to merge any number of non-terminator 541 // instructions with the block that falls through into the common successor. 542 if (MBB1 == PredBB || MBB2 == PredBB) { 543 MachineBasicBlock::iterator I; 544 unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); 545 if (CommonTailLen > NumTerms) 546 return true; 547 } 548 549 // If one of the blocks can be completely merged and happens to be in 550 // a position where the other could fall through into it, merge any number 551 // of instructions, because it can be done without a branch. 552 // TODO: If the blocks are not adjacent, move one of them so that they are? 553 if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) 554 return true; 555 if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) 556 return true; 557 558 // If both blocks have an unconditional branch temporarily stripped out, 559 // count that as an additional common instruction for the following 560 // heuristics. 561 unsigned EffectiveTailLen = CommonTailLen; 562 if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && 563 !MBB1->back().isBarrier() && 564 !MBB2->back().isBarrier()) 565 ++EffectiveTailLen; 566 567 // Check if the common tail is long enough to be worthwhile. 568 if (EffectiveTailLen >= minCommonTailLength) 569 return true; 570 571 // If we are optimizing for code size, 2 instructions in common is enough if 572 // we don't have to split a block. At worst we will be introducing 1 new 573 // branch instruction, which is likely to be smaller than the 2 574 // instructions that would be deleted in the merge. 575 MachineFunction *MF = MBB1->getParent(); 576 if (EffectiveTailLen >= 2 && 577 MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && 578 (I1 == MBB1->begin() || I2 == MBB2->begin())) 579 return true; 580 581 return false; 582} 583 584/// ComputeSameTails - Look through all the blocks in MergePotentials that have 585/// hash CurHash (guaranteed to match the last element). Build the vector 586/// SameTails of all those that have the (same) largest number of instructions 587/// in common of any pair of these blocks. SameTails entries contain an 588/// iterator into MergePotentials (from which the MachineBasicBlock can be 589/// found) and a MachineBasicBlock::iterator into that MBB indicating the 590/// instruction where the matching code sequence begins. 591/// Order of elements in SameTails is the reverse of the order in which 592/// those blocks appear in MergePotentials (where they are not necessarily 593/// consecutive). 594unsigned BranchFolder::ComputeSameTails(unsigned CurHash, 595 unsigned minCommonTailLength, 596 MachineBasicBlock *SuccBB, 597 MachineBasicBlock *PredBB) { 598 unsigned maxCommonTailLength = 0U; 599 SameTails.clear(); 600 MachineBasicBlock::iterator TrialBBI1, TrialBBI2; 601 MPIterator HighestMPIter = prior(MergePotentials.end()); 602 for (MPIterator CurMPIter = prior(MergePotentials.end()), 603 B = MergePotentials.begin(); 604 CurMPIter != B && CurMPIter->getHash() == CurHash; 605 --CurMPIter) { 606 for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { 607 unsigned CommonTailLen; 608 if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), 609 minCommonTailLength, 610 CommonTailLen, TrialBBI1, TrialBBI2, 611 SuccBB, PredBB)) { 612 if (CommonTailLen > maxCommonTailLength) { 613 SameTails.clear(); 614 maxCommonTailLength = CommonTailLen; 615 HighestMPIter = CurMPIter; 616 SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); 617 } 618 if (HighestMPIter == CurMPIter && 619 CommonTailLen == maxCommonTailLength) 620 SameTails.push_back(SameTailElt(I, TrialBBI2)); 621 } 622 if (I == B) 623 break; 624 } 625 } 626 return maxCommonTailLength; 627} 628 629/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from 630/// MergePotentials, restoring branches at ends of blocks as appropriate. 631void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, 632 MachineBasicBlock *SuccBB, 633 MachineBasicBlock *PredBB) { 634 MPIterator CurMPIter, B; 635 for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); 636 CurMPIter->getHash() == CurHash; 637 --CurMPIter) { 638 // Put the unconditional branch back, if we need one. 639 MachineBasicBlock *CurMBB = CurMPIter->getBlock(); 640 if (SuccBB && CurMBB != PredBB) 641 FixTail(CurMBB, SuccBB, TII); 642 if (CurMPIter == B) 643 break; 644 } 645 if (CurMPIter->getHash() != CurHash) 646 CurMPIter++; 647 MergePotentials.erase(CurMPIter, MergePotentials.end()); 648} 649 650/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist 651/// only of the common tail. Create a block that does by splitting one. 652bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, 653 unsigned maxCommonTailLength, 654 unsigned &commonTailIndex) { 655 commonTailIndex = 0; 656 unsigned TimeEstimate = ~0U; 657 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 658 // Use PredBB if possible; that doesn't require a new branch. 659 if (SameTails[i].getBlock() == PredBB) { 660 commonTailIndex = i; 661 break; 662 } 663 // Otherwise, make a (fairly bogus) choice based on estimate of 664 // how long it will take the various blocks to execute. 665 unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), 666 SameTails[i].getTailStartPos()); 667 if (t <= TimeEstimate) { 668 TimeEstimate = t; 669 commonTailIndex = i; 670 } 671 } 672 673 MachineBasicBlock::iterator BBI = 674 SameTails[commonTailIndex].getTailStartPos(); 675 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 676 677 // If the common tail includes any debug info we will take it pretty 678 // randomly from one of the inputs. Might be better to remove it? 679 DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " 680 << maxCommonTailLength); 681 682 MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); 683 if (!newMBB) { 684 DEBUG(dbgs() << "... failed!"); 685 return false; 686 } 687 688 SameTails[commonTailIndex].setBlock(newMBB); 689 SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); 690 691 // If we split PredBB, newMBB is the new predecessor. 692 if (PredBB == MBB) 693 PredBB = newMBB; 694 695 return true; 696} 697 698// See if any of the blocks in MergePotentials (which all have a common single 699// successor, or all have no successor) can be tail-merged. If there is a 700// successor, any blocks in MergePotentials that are not tail-merged and 701// are not immediately before Succ must have an unconditional branch to 702// Succ added (but the predecessor/successor lists need no adjustment). 703// The lone predecessor of Succ that falls through into Succ, 704// if any, is given in PredBB. 705 706bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, 707 MachineBasicBlock *PredBB) { 708 bool MadeChange = false; 709 710 // Except for the special cases below, tail-merge if there are at least 711 // this many instructions in common. 712 unsigned minCommonTailLength = TailMergeSize; 713 714 DEBUG(dbgs() << "\nTryTailMergeBlocks: "; 715 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 716 dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() 717 << (i == e-1 ? "" : ", "); 718 dbgs() << "\n"; 719 if (SuccBB) { 720 dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; 721 if (PredBB) 722 dbgs() << " which has fall-through from BB#" 723 << PredBB->getNumber() << "\n"; 724 } 725 dbgs() << "Looking for common tails of at least " 726 << minCommonTailLength << " instruction" 727 << (minCommonTailLength == 1 ? "" : "s") << '\n'; 728 ); 729 730 // Sort by hash value so that blocks with identical end sequences sort 731 // together. 732 std::stable_sort(MergePotentials.begin(), MergePotentials.end()); 733 734 // Walk through equivalence sets looking for actual exact matches. 735 while (MergePotentials.size() > 1) { 736 unsigned CurHash = MergePotentials.back().getHash(); 737 738 // Build SameTails, identifying the set of blocks with this hash code 739 // and with the maximum number of instructions in common. 740 unsigned maxCommonTailLength = ComputeSameTails(CurHash, 741 minCommonTailLength, 742 SuccBB, PredBB); 743 744 // If we didn't find any pair that has at least minCommonTailLength 745 // instructions in common, remove all blocks with this hash code and retry. 746 if (SameTails.empty()) { 747 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 748 continue; 749 } 750 751 // If one of the blocks is the entire common tail (and not the entry 752 // block, which we can't jump to), we can treat all blocks with this same 753 // tail at once. Use PredBB if that is one of the possibilities, as that 754 // will not introduce any extra branches. 755 MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> 756 getParent()->begin(); 757 unsigned commonTailIndex = SameTails.size(); 758 // If there are two blocks, check to see if one can be made to fall through 759 // into the other. 760 if (SameTails.size() == 2 && 761 SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && 762 SameTails[1].tailIsWholeBlock()) 763 commonTailIndex = 1; 764 else if (SameTails.size() == 2 && 765 SameTails[1].getBlock()->isLayoutSuccessor( 766 SameTails[0].getBlock()) && 767 SameTails[0].tailIsWholeBlock()) 768 commonTailIndex = 0; 769 else { 770 // Otherwise just pick one, favoring the fall-through predecessor if 771 // there is one. 772 for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { 773 MachineBasicBlock *MBB = SameTails[i].getBlock(); 774 if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) 775 continue; 776 if (MBB == PredBB) { 777 commonTailIndex = i; 778 break; 779 } 780 if (SameTails[i].tailIsWholeBlock()) 781 commonTailIndex = i; 782 } 783 } 784 785 if (commonTailIndex == SameTails.size() || 786 (SameTails[commonTailIndex].getBlock() == PredBB && 787 !SameTails[commonTailIndex].tailIsWholeBlock())) { 788 // None of the blocks consist entirely of the common tail. 789 // Split a block so that one does. 790 if (!CreateCommonTailOnlyBlock(PredBB, 791 maxCommonTailLength, commonTailIndex)) { 792 RemoveBlocksWithHash(CurHash, SuccBB, PredBB); 793 continue; 794 } 795 } 796 797 MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); 798 // MBB is common tail. Adjust all other BB's to jump to this one. 799 // Traversal must be forwards so erases work. 800 DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber() 801 << " for "); 802 for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { 803 if (commonTailIndex == i) 804 continue; 805 DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() 806 << (i == e-1 ? "" : ", ")); 807 // Hack the end off BB i, making it jump to BB commonTailIndex instead. 808 ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); 809 // BB i is no longer a predecessor of SuccBB; remove it from the worklist. 810 MergePotentials.erase(SameTails[i].getMPIter()); 811 } 812 DEBUG(dbgs() << "\n"); 813 // We leave commonTailIndex in the worklist in case there are other blocks 814 // that match it with a smaller number of instructions. 815 MadeChange = true; 816 } 817 return MadeChange; 818} 819 820bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 821 bool MadeChange = false; 822 if (!EnableTailMerge) return MadeChange; 823 824 // First find blocks with no successors. 825 MergePotentials.clear(); 826 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); 827 I != E && MergePotentials.size() < TailMergeThreshold; ++I) { 828 if (TriedMerging.count(I)) 829 continue; 830 if (I->succ_empty()) 831 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I)); 832 } 833 834 // If this is a large problem, avoid visiting the same basic blocks 835 // multiple times. 836 if (MergePotentials.size() == TailMergeThreshold) 837 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 838 TriedMerging.insert(MergePotentials[i].getBlock()); 839 840 // See if we can do any tail merging on those. 841 if (MergePotentials.size() >= 2) 842 MadeChange |= TryTailMergeBlocks(NULL, NULL); 843 844 // Look at blocks (IBB) with multiple predecessors (PBB). 845 // We change each predecessor to a canonical form, by 846 // (1) temporarily removing any unconditional branch from the predecessor 847 // to IBB, and 848 // (2) alter conditional branches so they branch to the other block 849 // not IBB; this may require adding back an unconditional branch to IBB 850 // later, where there wasn't one coming in. E.g. 851 // Bcc IBB 852 // fallthrough to QBB 853 // here becomes 854 // Bncc QBB 855 // with a conceptual B to IBB after that, which never actually exists. 856 // With those changes, we see whether the predecessors' tails match, 857 // and merge them if so. We change things out of canonical form and 858 // back to the way they were later in the process. (OptimizeBranches 859 // would undo some of this, but we can't use it, because we'd get into 860 // a compile-time infinite loop repeatedly doing and undoing the same 861 // transformations.) 862 863 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 864 I != E; ++I) { 865 if (I->pred_size() < 2) continue; 866 SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; 867 MachineBasicBlock *IBB = I; 868 MachineBasicBlock *PredBB = prior(I); 869 MergePotentials.clear(); 870 for (MachineBasicBlock::pred_iterator P = I->pred_begin(), 871 E2 = I->pred_end(); 872 P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) { 873 MachineBasicBlock *PBB = *P; 874 if (TriedMerging.count(PBB)) 875 continue; 876 877 // Skip blocks that loop to themselves, can't tail merge these. 878 if (PBB == IBB) 879 continue; 880 881 // Visit each predecessor only once. 882 if (!UniquePreds.insert(PBB)) 883 continue; 884 885 // Skip blocks which may jump to a landing pad. Can't tail merge these. 886 if (PBB->getLandingPadSuccessor()) 887 continue; 888 889 MachineBasicBlock *TBB = 0, *FBB = 0; 890 SmallVector<MachineOperand, 4> Cond; 891 if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { 892 // Failing case: IBB is the target of a cbr, and we cannot reverse the 893 // branch. 894 SmallVector<MachineOperand, 4> NewCond(Cond); 895 if (!Cond.empty() && TBB == IBB) { 896 if (TII->ReverseBranchCondition(NewCond)) 897 continue; 898 // This is the QBB case described above 899 if (!FBB) 900 FBB = llvm::next(MachineFunction::iterator(PBB)); 901 } 902 903 // Failing case: the only way IBB can be reached from PBB is via 904 // exception handling. Happens for landing pads. Would be nice to have 905 // a bit in the edge so we didn't have to do all this. 906 if (IBB->isLandingPad()) { 907 MachineFunction::iterator IP = PBB; IP++; 908 MachineBasicBlock *PredNextBB = NULL; 909 if (IP != MF.end()) 910 PredNextBB = IP; 911 if (TBB == NULL) { 912 if (IBB != PredNextBB) // fallthrough 913 continue; 914 } else if (FBB) { 915 if (TBB != IBB && FBB != IBB) // cbr then ubr 916 continue; 917 } else if (Cond.empty()) { 918 if (TBB != IBB) // ubr 919 continue; 920 } else { 921 if (TBB != IBB && IBB != PredNextBB) // cbr 922 continue; 923 } 924 } 925 926 // Remove the unconditional branch at the end, if any. 927 if (TBB && (Cond.empty() || FBB)) { 928 DebugLoc dl; // FIXME: this is nowhere 929 TII->RemoveBranch(*PBB); 930 if (!Cond.empty()) 931 // reinsert conditional branch only, for now 932 TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl); 933 } 934 935 MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P)); 936 } 937 } 938 939 // If this is a large problem, avoid visiting the same basic blocks multiple 940 // times. 941 if (MergePotentials.size() == TailMergeThreshold) 942 for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) 943 TriedMerging.insert(MergePotentials[i].getBlock()); 944 945 if (MergePotentials.size() >= 2) 946 MadeChange |= TryTailMergeBlocks(IBB, PredBB); 947 948 // Reinsert an unconditional branch if needed. The 1 below can occur as a 949 // result of removing blocks in TryTailMergeBlocks. 950 PredBB = prior(I); // this may have been changed in TryTailMergeBlocks 951 if (MergePotentials.size() == 1 && 952 MergePotentials.begin()->getBlock() != PredBB) 953 FixTail(MergePotentials.begin()->getBlock(), IBB, TII); 954 } 955 956 return MadeChange; 957} 958 959//===----------------------------------------------------------------------===// 960// Branch Optimization 961//===----------------------------------------------------------------------===// 962 963bool BranchFolder::OptimizeBranches(MachineFunction &MF) { 964 bool MadeChange = false; 965 966 // Make sure blocks are numbered in order 967 MF.RenumberBlocks(); 968 969 for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); 970 I != E; ) { 971 MachineBasicBlock *MBB = I++; 972 MadeChange |= OptimizeBlock(MBB); 973 974 // If it is dead, remove it. 975 if (MBB->pred_empty()) { 976 RemoveDeadBlock(MBB); 977 MadeChange = true; 978 ++NumDeadBlocks; 979 } 980 } 981 return MadeChange; 982} 983 984// Blocks should be considered empty if they contain only debug info; 985// else the debug info would affect codegen. 986static bool IsEmptyBlock(MachineBasicBlock *MBB) { 987 if (MBB->empty()) 988 return true; 989 for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end(); 990 MBBI!=MBBE; ++MBBI) { 991 if (!MBBI->isDebugValue()) 992 return false; 993 } 994 return true; 995} 996 997// Blocks with only debug info and branches should be considered the same 998// as blocks with only branches. 999static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { 1000 MachineBasicBlock::iterator MBBI, MBBE; 1001 for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) { 1002 if (!MBBI->isDebugValue()) 1003 break; 1004 } 1005 return (MBBI->isBranch()); 1006} 1007 1008/// IsBetterFallthrough - Return true if it would be clearly better to 1009/// fall-through to MBB1 than to fall through into MBB2. This has to return 1010/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will 1011/// result in infinite loops. 1012static bool IsBetterFallthrough(MachineBasicBlock *MBB1, 1013 MachineBasicBlock *MBB2) { 1014 // Right now, we use a simple heuristic. If MBB2 ends with a call, and 1015 // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to 1016 // optimize branches that branch to either a return block or an assert block 1017 // into a fallthrough to the return. 1018 if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false; 1019 1020 // If there is a clear successor ordering we make sure that one block 1021 // will fall through to the next 1022 if (MBB1->isSuccessor(MBB2)) return true; 1023 if (MBB2->isSuccessor(MBB1)) return false; 1024 1025 // Neither block consists entirely of debug info (per IsEmptyBlock check), 1026 // so we needn't test for falling off the beginning here. 1027 MachineBasicBlock::iterator MBB1I = --MBB1->end(); 1028 while (MBB1I->isDebugValue()) 1029 --MBB1I; 1030 MachineBasicBlock::iterator MBB2I = --MBB2->end(); 1031 while (MBB2I->isDebugValue()) 1032 --MBB2I; 1033 return MBB2I->isCall() && !MBB1I->isCall(); 1034} 1035 1036/// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch 1037/// instructions on the block. Always use the DebugLoc of the first 1038/// branching instruction found unless its absent, in which case use the 1039/// DebugLoc of the second if present. 1040static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) { 1041 MachineBasicBlock::iterator I = MBB.end(); 1042 if (I == MBB.begin()) 1043 return DebugLoc(); 1044 --I; 1045 while (I->isDebugValue() && I != MBB.begin()) 1046 --I; 1047 if (I->isBranch()) 1048 return I->getDebugLoc(); 1049 return DebugLoc(); 1050} 1051 1052/// OptimizeBlock - Analyze and optimize control flow related to the specified 1053/// block. This is never called on the entry block. 1054bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { 1055 bool MadeChange = false; 1056 MachineFunction &MF = *MBB->getParent(); 1057ReoptimizeBlock: 1058 1059 MachineFunction::iterator FallThrough = MBB; 1060 ++FallThrough; 1061 1062 // If this block is empty, make everyone use its fall-through, not the block 1063 // explicitly. Landing pads should not do this since the landing-pad table 1064 // points to this block. Blocks with their addresses taken shouldn't be 1065 // optimized away. 1066 if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { 1067 // Dead block? Leave for cleanup later. 1068 if (MBB->pred_empty()) return MadeChange; 1069 1070 if (FallThrough == MF.end()) { 1071 // TODO: Simplify preds to not branch here if possible! 1072 } else { 1073 // Rewrite all predecessors of the old block to go to the fallthrough 1074 // instead. 1075 while (!MBB->pred_empty()) { 1076 MachineBasicBlock *Pred = *(MBB->pred_end()-1); 1077 Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); 1078 } 1079 // If MBB was the target of a jump table, update jump tables to go to the 1080 // fallthrough instead. 1081 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1082 MJTI->ReplaceMBBInJumpTables(MBB, FallThrough); 1083 MadeChange = true; 1084 } 1085 return MadeChange; 1086 } 1087 1088 // Check to see if we can simplify the terminator of the block before this 1089 // one. 1090 MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); 1091 1092 MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; 1093 SmallVector<MachineOperand, 4> PriorCond; 1094 bool PriorUnAnalyzable = 1095 TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); 1096 if (!PriorUnAnalyzable) { 1097 // If the CFG for the prior block has extra edges, remove them. 1098 MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, 1099 !PriorCond.empty()); 1100 1101 // If the previous branch is conditional and both conditions go to the same 1102 // destination, remove the branch, replacing it with an unconditional one or 1103 // a fall-through. 1104 if (PriorTBB && PriorTBB == PriorFBB) { 1105 DebugLoc dl = getBranchDebugLoc(PrevBB); 1106 TII->RemoveBranch(PrevBB); 1107 PriorCond.clear(); 1108 if (PriorTBB != MBB) 1109 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1110 MadeChange = true; 1111 ++NumBranchOpts; 1112 goto ReoptimizeBlock; 1113 } 1114 1115 // If the previous block unconditionally falls through to this block and 1116 // this block has no other predecessors, move the contents of this block 1117 // into the prior block. This doesn't usually happen when SimplifyCFG 1118 // has been used, but it can happen if tail merging splits a fall-through 1119 // predecessor of a block. 1120 // This has to check PrevBB->succ_size() because EH edges are ignored by 1121 // AnalyzeBranch. 1122 if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && 1123 PrevBB.succ_size() == 1 && 1124 !MBB->hasAddressTaken() && !MBB->isLandingPad()) { 1125 DEBUG(dbgs() << "\nMerging into block: " << PrevBB 1126 << "From MBB: " << *MBB); 1127 // Remove redundant DBG_VALUEs first. 1128 if (PrevBB.begin() != PrevBB.end()) { 1129 MachineBasicBlock::iterator PrevBBIter = PrevBB.end(); 1130 --PrevBBIter; 1131 MachineBasicBlock::iterator MBBIter = MBB->begin(); 1132 // Check if DBG_VALUE at the end of PrevBB is identical to the 1133 // DBG_VALUE at the beginning of MBB. 1134 while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end() 1135 && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) { 1136 if (!MBBIter->isIdenticalTo(PrevBBIter)) 1137 break; 1138 MachineInstr *DuplicateDbg = MBBIter; 1139 ++MBBIter; -- PrevBBIter; 1140 DuplicateDbg->eraseFromParent(); 1141 } 1142 } 1143 PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); 1144 PrevBB.removeSuccessor(PrevBB.succ_begin()); 1145 assert(PrevBB.succ_empty()); 1146 PrevBB.transferSuccessors(MBB); 1147 MadeChange = true; 1148 return MadeChange; 1149 } 1150 1151 // If the previous branch *only* branches to *this* block (conditional or 1152 // not) remove the branch. 1153 if (PriorTBB == MBB && PriorFBB == 0) { 1154 TII->RemoveBranch(PrevBB); 1155 MadeChange = true; 1156 ++NumBranchOpts; 1157 goto ReoptimizeBlock; 1158 } 1159 1160 // If the prior block branches somewhere else on the condition and here if 1161 // the condition is false, remove the uncond second branch. 1162 if (PriorFBB == MBB) { 1163 DebugLoc dl = getBranchDebugLoc(PrevBB); 1164 TII->RemoveBranch(PrevBB); 1165 TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl); 1166 MadeChange = true; 1167 ++NumBranchOpts; 1168 goto ReoptimizeBlock; 1169 } 1170 1171 // If the prior block branches here on true and somewhere else on false, and 1172 // if the branch condition is reversible, reverse the branch to create a 1173 // fall-through. 1174 if (PriorTBB == MBB) { 1175 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1176 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1177 DebugLoc dl = getBranchDebugLoc(PrevBB); 1178 TII->RemoveBranch(PrevBB); 1179 TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl); 1180 MadeChange = true; 1181 ++NumBranchOpts; 1182 goto ReoptimizeBlock; 1183 } 1184 } 1185 1186 // If this block has no successors (e.g. it is a return block or ends with 1187 // a call to a no-return function like abort or __cxa_throw) and if the pred 1188 // falls through into this block, and if it would otherwise fall through 1189 // into the block after this, move this block to the end of the function. 1190 // 1191 // We consider it more likely that execution will stay in the function (e.g. 1192 // due to loops) than it is to exit it. This asserts in loops etc, moving 1193 // the assert condition out of the loop body. 1194 if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && 1195 MachineFunction::iterator(PriorTBB) == FallThrough && 1196 !MBB->canFallThrough()) { 1197 bool DoTransform = true; 1198 1199 // We have to be careful that the succs of PredBB aren't both no-successor 1200 // blocks. If neither have successors and if PredBB is the second from 1201 // last block in the function, we'd just keep swapping the two blocks for 1202 // last. Only do the swap if one is clearly better to fall through than 1203 // the other. 1204 if (FallThrough == --MF.end() && 1205 !IsBetterFallthrough(PriorTBB, MBB)) 1206 DoTransform = false; 1207 1208 if (DoTransform) { 1209 // Reverse the branch so we will fall through on the previous true cond. 1210 SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); 1211 if (!TII->ReverseBranchCondition(NewPriorCond)) { 1212 DEBUG(dbgs() << "\nMoving MBB: " << *MBB 1213 << "To make fallthrough to: " << *PriorTBB << "\n"); 1214 1215 DebugLoc dl = getBranchDebugLoc(PrevBB); 1216 TII->RemoveBranch(PrevBB); 1217 TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl); 1218 1219 // Move this block to the end of the function. 1220 MBB->moveAfter(--MF.end()); 1221 MadeChange = true; 1222 ++NumBranchOpts; 1223 return MadeChange; 1224 } 1225 } 1226 } 1227 } 1228 1229 // Analyze the branch in the current block. 1230 MachineBasicBlock *CurTBB = 0, *CurFBB = 0; 1231 SmallVector<MachineOperand, 4> CurCond; 1232 bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); 1233 if (!CurUnAnalyzable) { 1234 // If the CFG for the prior block has extra edges, remove them. 1235 MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); 1236 1237 // If this is a two-way branch, and the FBB branches to this block, reverse 1238 // the condition so the single-basic-block loop is faster. Instead of: 1239 // Loop: xxx; jcc Out; jmp Loop 1240 // we want: 1241 // Loop: xxx; jncc Loop; jmp Out 1242 if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { 1243 SmallVector<MachineOperand, 4> NewCond(CurCond); 1244 if (!TII->ReverseBranchCondition(NewCond)) { 1245 DebugLoc dl = getBranchDebugLoc(*MBB); 1246 TII->RemoveBranch(*MBB); 1247 TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); 1248 MadeChange = true; 1249 ++NumBranchOpts; 1250 goto ReoptimizeBlock; 1251 } 1252 } 1253 1254 // If this branch is the only thing in its block, see if we can forward 1255 // other blocks across it. 1256 if (CurTBB && CurCond.empty() && CurFBB == 0 && 1257 IsBranchOnlyBlock(MBB) && CurTBB != MBB && 1258 !MBB->hasAddressTaken()) { 1259 DebugLoc dl = getBranchDebugLoc(*MBB); 1260 // This block may contain just an unconditional branch. Because there can 1261 // be 'non-branch terminators' in the block, try removing the branch and 1262 // then seeing if the block is empty. 1263 TII->RemoveBranch(*MBB); 1264 // If the only things remaining in the block are debug info, remove these 1265 // as well, so this will behave the same as an empty block in non-debug 1266 // mode. 1267 if (!MBB->empty()) { 1268 bool NonDebugInfoFound = false; 1269 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 1270 I != E; ++I) { 1271 if (!I->isDebugValue()) { 1272 NonDebugInfoFound = true; 1273 break; 1274 } 1275 } 1276 if (!NonDebugInfoFound) 1277 // Make the block empty, losing the debug info (we could probably 1278 // improve this in some cases.) 1279 MBB->erase(MBB->begin(), MBB->end()); 1280 } 1281 // If this block is just an unconditional branch to CurTBB, we can 1282 // usually completely eliminate the block. The only case we cannot 1283 // completely eliminate the block is when the block before this one 1284 // falls through into MBB and we can't understand the prior block's branch 1285 // condition. 1286 if (MBB->empty()) { 1287 bool PredHasNoFallThrough = !PrevBB.canFallThrough(); 1288 if (PredHasNoFallThrough || !PriorUnAnalyzable || 1289 !PrevBB.isSuccessor(MBB)) { 1290 // If the prior block falls through into us, turn it into an 1291 // explicit branch to us to make updates simpler. 1292 if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && 1293 PriorTBB != MBB && PriorFBB != MBB) { 1294 if (PriorTBB == 0) { 1295 assert(PriorCond.empty() && PriorFBB == 0 && 1296 "Bad branch analysis"); 1297 PriorTBB = MBB; 1298 } else { 1299 assert(PriorFBB == 0 && "Machine CFG out of date!"); 1300 PriorFBB = MBB; 1301 } 1302 DebugLoc pdl = getBranchDebugLoc(PrevBB); 1303 TII->RemoveBranch(PrevBB); 1304 TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl); 1305 } 1306 1307 // Iterate through all the predecessors, revectoring each in-turn. 1308 size_t PI = 0; 1309 bool DidChange = false; 1310 bool HasBranchToSelf = false; 1311 while(PI != MBB->pred_size()) { 1312 MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); 1313 if (PMBB == MBB) { 1314 // If this block has an uncond branch to itself, leave it. 1315 ++PI; 1316 HasBranchToSelf = true; 1317 } else { 1318 DidChange = true; 1319 PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); 1320 // If this change resulted in PMBB ending in a conditional 1321 // branch where both conditions go to the same destination, 1322 // change this to an unconditional branch (and fix the CFG). 1323 MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; 1324 SmallVector<MachineOperand, 4> NewCurCond; 1325 bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, 1326 NewCurFBB, NewCurCond, true); 1327 if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { 1328 DebugLoc pdl = getBranchDebugLoc(*PMBB); 1329 TII->RemoveBranch(*PMBB); 1330 NewCurCond.clear(); 1331 TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, pdl); 1332 MadeChange = true; 1333 ++NumBranchOpts; 1334 PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); 1335 } 1336 } 1337 } 1338 1339 // Change any jumptables to go to the new MBB. 1340 if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) 1341 MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); 1342 if (DidChange) { 1343 ++NumBranchOpts; 1344 MadeChange = true; 1345 if (!HasBranchToSelf) return MadeChange; 1346 } 1347 } 1348 } 1349 1350 // Add the branch back if the block is more than just an uncond branch. 1351 TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl); 1352 } 1353 } 1354 1355 // If the prior block doesn't fall through into this block, and if this 1356 // block doesn't fall through into some other block, see if we can find a 1357 // place to move this block where a fall-through will happen. 1358 if (!PrevBB.canFallThrough()) { 1359 1360 // Now we know that there was no fall-through into this block, check to 1361 // see if it has a fall-through into its successor. 1362 bool CurFallsThru = MBB->canFallThrough(); 1363 1364 if (!MBB->isLandingPad()) { 1365 // Check all the predecessors of this block. If one of them has no fall 1366 // throughs, move this block right after it. 1367 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), 1368 E = MBB->pred_end(); PI != E; ++PI) { 1369 // Analyze the branch at the end of the pred. 1370 MachineBasicBlock *PredBB = *PI; 1371 MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; 1372 MachineBasicBlock *PredTBB = 0, *PredFBB = 0; 1373 SmallVector<MachineOperand, 4> PredCond; 1374 if (PredBB != MBB && !PredBB->canFallThrough() && 1375 !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) 1376 && (!CurFallsThru || !CurTBB || !CurFBB) 1377 && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { 1378 // If the current block doesn't fall through, just move it. 1379 // If the current block can fall through and does not end with a 1380 // conditional branch, we need to append an unconditional jump to 1381 // the (current) next block. To avoid a possible compile-time 1382 // infinite loop, move blocks only backward in this case. 1383 // Also, if there are already 2 branches here, we cannot add a third; 1384 // this means we have the case 1385 // Bcc next 1386 // B elsewhere 1387 // next: 1388 if (CurFallsThru) { 1389 MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); 1390 CurCond.clear(); 1391 TII->InsertBranch(*MBB, NextBB, 0, CurCond, DebugLoc()); 1392 } 1393 MBB->moveAfter(PredBB); 1394 MadeChange = true; 1395 goto ReoptimizeBlock; 1396 } 1397 } 1398 } 1399 1400 if (!CurFallsThru) { 1401 // Check all successors to see if we can move this block before it. 1402 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1403 E = MBB->succ_end(); SI != E; ++SI) { 1404 // Analyze the branch at the end of the block before the succ. 1405 MachineBasicBlock *SuccBB = *SI; 1406 MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; 1407 1408 // If this block doesn't already fall-through to that successor, and if 1409 // the succ doesn't already have a block that can fall through into it, 1410 // and if the successor isn't an EH destination, we can arrange for the 1411 // fallthrough to happen. 1412 if (SuccBB != MBB && &*SuccPrev != MBB && 1413 !SuccPrev->canFallThrough() && !CurUnAnalyzable && 1414 !SuccBB->isLandingPad()) { 1415 MBB->moveBefore(SuccBB); 1416 MadeChange = true; 1417 goto ReoptimizeBlock; 1418 } 1419 } 1420 1421 // Okay, there is no really great place to put this block. If, however, 1422 // the block before this one would be a fall-through if this block were 1423 // removed, move this block to the end of the function. 1424 MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0; 1425 SmallVector<MachineOperand, 4> PrevCond; 1426 if (FallThrough != MF.end() && 1427 !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && 1428 PrevBB.isSuccessor(FallThrough)) { 1429 MBB->moveAfter(--MF.end()); 1430 MadeChange = true; 1431 return MadeChange; 1432 } 1433 } 1434 } 1435 1436 return MadeChange; 1437} 1438 1439//===----------------------------------------------------------------------===// 1440// Hoist Common Code 1441//===----------------------------------------------------------------------===// 1442 1443/// HoistCommonCode - Hoist common instruction sequences at the start of basic 1444/// blocks to their common predecessor. 1445bool BranchFolder::HoistCommonCode(MachineFunction &MF) { 1446 bool MadeChange = false; 1447 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) { 1448 MachineBasicBlock *MBB = I++; 1449 MadeChange |= HoistCommonCodeInSuccs(MBB); 1450 } 1451 1452 return MadeChange; 1453} 1454 1455/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 1456/// its 'true' successor. 1457static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 1458 MachineBasicBlock *TrueBB) { 1459 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 1460 E = BB->succ_end(); SI != E; ++SI) { 1461 MachineBasicBlock *SuccBB = *SI; 1462 if (SuccBB != TrueBB) 1463 return SuccBB; 1464 } 1465 return NULL; 1466} 1467 1468/// findHoistingInsertPosAndDeps - Find the location to move common instructions 1469/// in successors to. The location is usually just before the terminator, 1470/// however if the terminator is a conditional branch and its previous 1471/// instruction is the flag setting instruction, the previous instruction is 1472/// the preferred location. This function also gathers uses and defs of the 1473/// instructions from the insertion point to the end of the block. The data is 1474/// used by HoistCommonCodeInSuccs to ensure safety. 1475static 1476MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB, 1477 const TargetInstrInfo *TII, 1478 const TargetRegisterInfo *TRI, 1479 SmallSet<unsigned,4> &Uses, 1480 SmallSet<unsigned,4> &Defs) { 1481 MachineBasicBlock::iterator Loc = MBB->getFirstTerminator(); 1482 if (!TII->isUnpredicatedTerminator(Loc)) 1483 return MBB->end(); 1484 1485 for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) { 1486 const MachineOperand &MO = Loc->getOperand(i); 1487 if (!MO.isReg()) 1488 continue; 1489 unsigned Reg = MO.getReg(); 1490 if (!Reg) 1491 continue; 1492 if (MO.isUse()) { 1493 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1494 Uses.insert(*AI); 1495 } else if (!MO.isDead()) 1496 // Don't try to hoist code in the rare case the terminator defines a 1497 // register that is later used. 1498 return MBB->end(); 1499 } 1500 1501 if (Uses.empty()) 1502 return Loc; 1503 if (Loc == MBB->begin()) 1504 return MBB->end(); 1505 1506 // The terminator is probably a conditional branch, try not to separate the 1507 // branch from condition setting instruction. 1508 MachineBasicBlock::iterator PI = Loc; 1509 --PI; 1510 while (PI != MBB->begin() && Loc->isDebugValue()) 1511 --PI; 1512 1513 bool IsDef = false; 1514 for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) { 1515 const MachineOperand &MO = PI->getOperand(i); 1516 // If PI has a regmask operand, it is probably a call. Separate away. 1517 if (MO.isRegMask()) 1518 return Loc; 1519 if (!MO.isReg() || MO.isUse()) 1520 continue; 1521 unsigned Reg = MO.getReg(); 1522 if (!Reg) 1523 continue; 1524 if (Uses.count(Reg)) 1525 IsDef = true; 1526 } 1527 if (!IsDef) 1528 // The condition setting instruction is not just before the conditional 1529 // branch. 1530 return Loc; 1531 1532 // Be conservative, don't insert instruction above something that may have 1533 // side-effects. And since it's potentially bad to separate flag setting 1534 // instruction from the conditional branch, just abort the optimization 1535 // completely. 1536 // Also avoid moving code above predicated instruction since it's hard to 1537 // reason about register liveness with predicated instruction. 1538 bool DontMoveAcrossStore = true; 1539 if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) || 1540 TII->isPredicated(PI)) 1541 return MBB->end(); 1542 1543 1544 // Find out what registers are live. Note this routine is ignoring other live 1545 // registers which are only used by instructions in successor blocks. 1546 for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) { 1547 const MachineOperand &MO = PI->getOperand(i); 1548 if (!MO.isReg()) 1549 continue; 1550 unsigned Reg = MO.getReg(); 1551 if (!Reg) 1552 continue; 1553 if (MO.isUse()) { 1554 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1555 Uses.insert(*AI); 1556 } else { 1557 if (Uses.erase(Reg)) { 1558 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) 1559 Uses.erase(*SubRegs); // Use sub-registers to be conservative 1560 } 1561 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1562 Defs.insert(*AI); 1563 } 1564 } 1565 1566 return PI; 1567} 1568 1569/// HoistCommonCodeInSuccs - If the successors of MBB has common instruction 1570/// sequence at the start of the function, move the instructions before MBB 1571/// terminator if it's legal. 1572bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) { 1573 MachineBasicBlock *TBB = 0, *FBB = 0; 1574 SmallVector<MachineOperand, 4> Cond; 1575 if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty()) 1576 return false; 1577 1578 if (!FBB) FBB = findFalseBlock(MBB, TBB); 1579 if (!FBB) 1580 // Malformed bcc? True and false blocks are the same? 1581 return false; 1582 1583 // Restrict the optimization to cases where MBB is the only predecessor, 1584 // it is an obvious win. 1585 if (TBB->pred_size() > 1 || FBB->pred_size() > 1) 1586 return false; 1587 1588 // Find a suitable position to hoist the common instructions to. Also figure 1589 // out which registers are used or defined by instructions from the insertion 1590 // point to the end of the block. 1591 SmallSet<unsigned, 4> Uses, Defs; 1592 MachineBasicBlock::iterator Loc = 1593 findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs); 1594 if (Loc == MBB->end()) 1595 return false; 1596 1597 bool HasDups = false; 1598 SmallVector<unsigned, 4> LocalDefs; 1599 SmallSet<unsigned, 4> LocalDefsSet; 1600 MachineBasicBlock::iterator TIB = TBB->begin(); 1601 MachineBasicBlock::iterator FIB = FBB->begin(); 1602 MachineBasicBlock::iterator TIE = TBB->end(); 1603 MachineBasicBlock::iterator FIE = FBB->end(); 1604 while (TIB != TIE && FIB != FIE) { 1605 // Skip dbg_value instructions. These do not count. 1606 if (TIB->isDebugValue()) { 1607 while (TIB != TIE && TIB->isDebugValue()) 1608 ++TIB; 1609 if (TIB == TIE) 1610 break; 1611 } 1612 if (FIB->isDebugValue()) { 1613 while (FIB != FIE && FIB->isDebugValue()) 1614 ++FIB; 1615 if (FIB == FIE) 1616 break; 1617 } 1618 if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead)) 1619 break; 1620 1621 if (TII->isPredicated(TIB)) 1622 // Hard to reason about register liveness with predicated instruction. 1623 break; 1624 1625 bool IsSafe = true; 1626 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1627 MachineOperand &MO = TIB->getOperand(i); 1628 // Don't attempt to hoist instructions with register masks. 1629 if (MO.isRegMask()) { 1630 IsSafe = false; 1631 break; 1632 } 1633 if (!MO.isReg()) 1634 continue; 1635 unsigned Reg = MO.getReg(); 1636 if (!Reg) 1637 continue; 1638 if (MO.isDef()) { 1639 if (Uses.count(Reg)) { 1640 // Avoid clobbering a register that's used by the instruction at 1641 // the point of insertion. 1642 IsSafe = false; 1643 break; 1644 } 1645 1646 if (Defs.count(Reg) && !MO.isDead()) { 1647 // Don't hoist the instruction if the def would be clobber by the 1648 // instruction at the point insertion. FIXME: This is overly 1649 // conservative. It should be possible to hoist the instructions 1650 // in BB2 in the following example: 1651 // BB1: 1652 // r1, eflag = op1 r2, r3 1653 // brcc eflag 1654 // 1655 // BB2: 1656 // r1 = op2, ... 1657 // = op3, r1<kill> 1658 IsSafe = false; 1659 break; 1660 } 1661 } else if (!LocalDefsSet.count(Reg)) { 1662 if (Defs.count(Reg)) { 1663 // Use is defined by the instruction at the point of insertion. 1664 IsSafe = false; 1665 break; 1666 } 1667 1668 if (MO.isKill() && Uses.count(Reg)) 1669 // Kills a register that's read by the instruction at the point of 1670 // insertion. Remove the kill marker. 1671 MO.setIsKill(false); 1672 } 1673 } 1674 if (!IsSafe) 1675 break; 1676 1677 bool DontMoveAcrossStore = true; 1678 if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore)) 1679 break; 1680 1681 // Remove kills from LocalDefsSet, these registers had short live ranges. 1682 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1683 MachineOperand &MO = TIB->getOperand(i); 1684 if (!MO.isReg() || !MO.isUse() || !MO.isKill()) 1685 continue; 1686 unsigned Reg = MO.getReg(); 1687 if (!Reg || !LocalDefsSet.count(Reg)) 1688 continue; 1689 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1690 LocalDefsSet.erase(*AI); 1691 } 1692 1693 // Track local defs so we can update liveins. 1694 for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) { 1695 MachineOperand &MO = TIB->getOperand(i); 1696 if (!MO.isReg() || !MO.isDef() || MO.isDead()) 1697 continue; 1698 unsigned Reg = MO.getReg(); 1699 if (!Reg) 1700 continue; 1701 LocalDefs.push_back(Reg); 1702 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 1703 LocalDefsSet.insert(*AI); 1704 } 1705 1706 HasDups = true; 1707 ++TIB; 1708 ++FIB; 1709 } 1710 1711 if (!HasDups) 1712 return false; 1713 1714 MBB->splice(Loc, TBB, TBB->begin(), TIB); 1715 FBB->erase(FBB->begin(), FIB); 1716 1717 // Update livein's. 1718 for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) { 1719 unsigned Def = LocalDefs[i]; 1720 if (LocalDefsSet.count(Def)) { 1721 TBB->addLiveIn(Def); 1722 FBB->addLiveIn(Def); 1723 } 1724 } 1725 1726 ++NumHoist; 1727 return true; 1728} 1729