1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===// 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 // Peephole optimize the CFG. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #define DEBUG_TYPE "simplifycfg" 15 #include "llvm/Transforms/Utils/Local.h" 16 #include "llvm/Constants.h" 17 #include "llvm/Instructions.h" 18 #include "llvm/IntrinsicInst.h" 19 #include "llvm/LLVMContext.h" 20 #include "llvm/Type.h" 21 #include "llvm/DerivedTypes.h" 22 #include "llvm/GlobalVariable.h" 23 #include "llvm/Support/CFG.h" 24 #include "llvm/Support/Debug.h" 25 #include "llvm/Support/raw_ostream.h" 26 #include "llvm/Analysis/ConstantFolding.h" 27 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 28 #include "llvm/ADT/SmallVector.h" 29 #include "llvm/ADT/SmallPtrSet.h" 30 #include "llvm/ADT/Statistic.h" 31 #include <algorithm> 32 #include <functional> 33 #include <set> 34 #include <map> 35 using namespace llvm; 36 37 STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 38 39 /// SafeToMergeTerminators - Return true if it is safe to merge these two 40 /// terminator instructions together. 41 /// 42 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 43 if (SI1 == SI2) return false; // Can't merge with self! 44 45 // It is not safe to merge these two switch instructions if they have a common 46 // successor, and if that successor has a PHI node, and if *that* PHI node has 47 // conflicting incoming values from the two switch blocks. 48 BasicBlock *SI1BB = SI1->getParent(); 49 BasicBlock *SI2BB = SI2->getParent(); 50 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 51 52 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 53 if (SI1Succs.count(*I)) 54 for (BasicBlock::iterator BBI = (*I)->begin(); 55 isa<PHINode>(BBI); ++BBI) { 56 PHINode *PN = cast<PHINode>(BBI); 57 if (PN->getIncomingValueForBlock(SI1BB) != 58 PN->getIncomingValueForBlock(SI2BB)) 59 return false; 60 } 61 62 return true; 63 } 64 65 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 66 /// now be entries in it from the 'NewPred' block. The values that will be 67 /// flowing into the PHI nodes will be the same as those coming in from 68 /// ExistPred, an existing predecessor of Succ. 69 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 70 BasicBlock *ExistPred) { 71 assert(std::find(succ_begin(ExistPred), succ_end(ExistPred), Succ) != 72 succ_end(ExistPred) && "ExistPred is not a predecessor of Succ!"); 73 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 74 75 PHINode *PN; 76 for (BasicBlock::iterator I = Succ->begin(); 77 (PN = dyn_cast<PHINode>(I)); ++I) 78 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 79 } 80 81 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an 82 /// almost-empty BB ending in an unconditional branch to Succ, into succ. 83 /// 84 /// Assumption: Succ is the single successor for BB. 85 /// 86 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { 87 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!"); 88 89 DOUT << "Looking to fold " << BB->getNameStart() << " into " 90 << Succ->getNameStart() << "\n"; 91 // Shortcut, if there is only a single predecessor it must be BB and merging 92 // is always safe 93 if (Succ->getSinglePredecessor()) return true; 94 95 typedef SmallPtrSet<Instruction*, 16> InstrSet; 96 InstrSet BBPHIs; 97 98 // Make a list of all phi nodes in BB 99 BasicBlock::iterator BBI = BB->begin(); 100 while (isa<PHINode>(*BBI)) BBPHIs.insert(BBI++); 101 102 // Make a list of the predecessors of BB 103 typedef SmallPtrSet<BasicBlock*, 16> BlockSet; 104 BlockSet BBPreds(pred_begin(BB), pred_end(BB)); 105 106 // Use that list to make another list of common predecessors of BB and Succ 107 BlockSet CommonPreds; 108 for (pred_iterator PI = pred_begin(Succ), PE = pred_end(Succ); 109 PI != PE; ++PI) 110 if (BBPreds.count(*PI)) 111 CommonPreds.insert(*PI); 112 113 // Shortcut, if there are no common predecessors, merging is always safe 114 if (CommonPreds.empty()) 115 return true; 116 117 // Look at all the phi nodes in Succ, to see if they present a conflict when 118 // merging these blocks 119 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 120 PHINode *PN = cast<PHINode>(I); 121 122 // If the incoming value from BB is again a PHINode in 123 // BB which has the same incoming value for *PI as PN does, we can 124 // merge the phi nodes and then the blocks can still be merged 125 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB)); 126 if (BBPN && BBPN->getParent() == BB) { 127 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 128 PI != PE; PI++) { 129 if (BBPN->getIncomingValueForBlock(*PI) 130 != PN->getIncomingValueForBlock(*PI)) { 131 DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " 132 << Succ->getNameStart() << " is conflicting with " 133 << BBPN->getNameStart() << " with regard to common predecessor " 134 << (*PI)->getNameStart() << "\n"; 135 return false; 136 } 137 } 138 // Remove this phinode from the list of phis in BB, since it has been 139 // handled. 140 BBPHIs.erase(BBPN); 141 } else { 142 Value* Val = PN->getIncomingValueForBlock(BB); 143 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 144 PI != PE; PI++) { 145 // See if the incoming value for the common predecessor is equal to the 146 // one for BB, in which case this phi node will not prevent the merging 147 // of the block. 148 if (Val != PN->getIncomingValueForBlock(*PI)) { 149 DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " 150 << Succ->getNameStart() << " is conflicting with regard to common " 151 << "predecessor " << (*PI)->getNameStart() << "\n"; 152 return false; 153 } 154 } 155 } 156 } 157 158 // If there are any other phi nodes in BB that don't have a phi node in Succ 159 // to merge with, they must be moved to Succ completely. However, for any 160 // predecessors of Succ, branches will be added to the phi node that just 161 // point to itself. So, for any common predecessors, this must not cause 162 // conflicts. 163 for (InstrSet::iterator I = BBPHIs.begin(), E = BBPHIs.end(); 164 I != E; I++) { 165 PHINode *PN = cast<PHINode>(*I); 166 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 167 PI != PE; PI++) 168 if (PN->getIncomingValueForBlock(*PI) != PN) { 169 DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " 170 << BB->getNameStart() << " is conflicting with regard to common " 171 << "predecessor " << (*PI)->getNameStart() << "\n"; 172 return false; 173 } 174 } 175 176 return true; 177 } 178 179 /// TryToSimplifyUncondBranchFromEmptyBlock - BB contains an unconditional 180 /// branch to Succ, and contains no instructions other than PHI nodes and the 181 /// branch. If possible, eliminate BB. 182 static bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, 183 BasicBlock *Succ) { 184 // Check to see if merging these blocks would cause conflicts for any of the 185 // phi nodes in BB or Succ. If not, we can safely merge. 186 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false; 187 188 DOUT << "Killing Trivial BB: \n" << *BB; 189 190 if (isa<PHINode>(Succ->begin())) { 191 // If there is more than one pred of succ, and there are PHI nodes in 192 // the successor, then we need to add incoming edges for the PHI nodes 193 // 194 const SmallVector<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB)); 195 196 // Loop over all of the PHI nodes in the successor of BB. 197 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 198 PHINode *PN = cast<PHINode>(I); 199 Value *OldVal = PN->removeIncomingValue(BB, false); 200 assert(OldVal && "No entry in PHI for Pred BB!"); 201 202 // If this incoming value is one of the PHI nodes in BB, the new entries 203 // in the PHI node are the entries from the old PHI. 204 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) { 205 PHINode *OldValPN = cast<PHINode>(OldVal); 206 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) 207 // Note that, since we are merging phi nodes and BB and Succ might 208 // have common predecessors, we could end up with a phi node with 209 // identical incoming branches. This will be cleaned up later (and 210 // will trigger asserts if we try to clean it up now, without also 211 // simplifying the corresponding conditional branch). 212 PN->addIncoming(OldValPN->getIncomingValue(i), 213 OldValPN->getIncomingBlock(i)); 214 } else { 215 // Add an incoming value for each of the new incoming values. 216 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) 217 PN->addIncoming(OldVal, BBPreds[i]); 218 } 219 } 220 } 221 222 if (isa<PHINode>(&BB->front())) { 223 SmallVector<BasicBlock*, 16> 224 OldSuccPreds(pred_begin(Succ), pred_end(Succ)); 225 226 // Move all PHI nodes in BB to Succ if they are alive, otherwise 227 // delete them. 228 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) { 229 if (PN->use_empty()) { 230 // Just remove the dead phi. This happens if Succ's PHIs were the only 231 // users of the PHI nodes. 232 PN->eraseFromParent(); 233 continue; 234 } 235 236 // The instruction is alive, so this means that BB must dominate all 237 // predecessors of Succ (Since all uses of the PN are after its 238 // definition, so in Succ or a block dominated by Succ. If a predecessor 239 // of Succ would not be dominated by BB, PN would violate the def before 240 // use SSA demand). Therefore, we can simply move the phi node to the 241 // next block. 242 Succ->getInstList().splice(Succ->begin(), 243 BB->getInstList(), BB->begin()); 244 245 // We need to add new entries for the PHI node to account for 246 // predecessors of Succ that the PHI node does not take into 247 // account. At this point, since we know that BB dominated succ and all 248 // of its predecessors, this means that we should any newly added 249 // incoming edges should use the PHI node itself as the value for these 250 // edges, because they are loop back edges. 251 for (unsigned i = 0, e = OldSuccPreds.size(); i != e; ++i) 252 if (OldSuccPreds[i] != BB) 253 PN->addIncoming(PN, OldSuccPreds[i]); 254 } 255 } 256 257 // Everything that jumped to BB now goes to Succ. 258 BB->replaceAllUsesWith(Succ); 259 if (!Succ->hasName()) Succ->takeName(BB); 260 BB->eraseFromParent(); // Delete the old basic block. 261 return true; 262 } 263 264 /// GetIfCondition - Given a basic block (BB) with two predecessors (and 265 /// presumably PHI nodes in it), check to see if the merge at this block is due 266 /// to an "if condition". If so, return the boolean condition that determines 267 /// which entry into BB will be taken. Also, return by references the block 268 /// that will be entered from if the condition is true, and the block that will 269 /// be entered if the condition is false. 270 /// 271 /// 272 static Value *GetIfCondition(BasicBlock *BB, 273 BasicBlock *&IfTrue, BasicBlock *&IfFalse) { 274 assert(std::distance(pred_begin(BB), pred_end(BB)) == 2 && 275 "Function can only handle blocks with 2 predecessors!"); 276 BasicBlock *Pred1 = *pred_begin(BB); 277 BasicBlock *Pred2 = *++pred_begin(BB); 278 279 // We can only handle branches. Other control flow will be lowered to 280 // branches if possible anyway. 281 if (!isa<BranchInst>(Pred1->getTerminator()) || 282 !isa<BranchInst>(Pred2->getTerminator())) 283 return 0; 284 BranchInst *Pred1Br = cast<BranchInst>(Pred1->getTerminator()); 285 BranchInst *Pred2Br = cast<BranchInst>(Pred2->getTerminator()); 286 287 // Eliminate code duplication by ensuring that Pred1Br is conditional if 288 // either are. 289 if (Pred2Br->isConditional()) { 290 // If both branches are conditional, we don't have an "if statement". In 291 // reality, we could transform this case, but since the condition will be 292 // required anyway, we stand no chance of eliminating it, so the xform is 293 // probably not profitable. 294 if (Pred1Br->isConditional()) 295 return 0; 296 297 std::swap(Pred1, Pred2); 298 std::swap(Pred1Br, Pred2Br); 299 } 300 301 if (Pred1Br->isConditional()) { 302 // If we found a conditional branch predecessor, make sure that it branches 303 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 304 if (Pred1Br->getSuccessor(0) == BB && 305 Pred1Br->getSuccessor(1) == Pred2) { 306 IfTrue = Pred1; 307 IfFalse = Pred2; 308 } else if (Pred1Br->getSuccessor(0) == Pred2 && 309 Pred1Br->getSuccessor(1) == BB) { 310 IfTrue = Pred2; 311 IfFalse = Pred1; 312 } else { 313 // We know that one arm of the conditional goes to BB, so the other must 314 // go somewhere unrelated, and this must not be an "if statement". 315 return 0; 316 } 317 318 // The only thing we have to watch out for here is to make sure that Pred2 319 // doesn't have incoming edges from other blocks. If it does, the condition 320 // doesn't dominate BB. 321 if (++pred_begin(Pred2) != pred_end(Pred2)) 322 return 0; 323 324 return Pred1Br->getCondition(); 325 } 326 327 // Ok, if we got here, both predecessors end with an unconditional branch to 328 // BB. Don't panic! If both blocks only have a single (identical) 329 // predecessor, and THAT is a conditional branch, then we're all ok! 330 if (pred_begin(Pred1) == pred_end(Pred1) || 331 ++pred_begin(Pred1) != pred_end(Pred1) || 332 pred_begin(Pred2) == pred_end(Pred2) || 333 ++pred_begin(Pred2) != pred_end(Pred2) || 334 *pred_begin(Pred1) != *pred_begin(Pred2)) 335 return 0; 336 337 // Otherwise, if this is a conditional branch, then we can use it! 338 BasicBlock *CommonPred = *pred_begin(Pred1); 339 if (BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator())) { 340 assert(BI->isConditional() && "Two successors but not conditional?"); 341 if (BI->getSuccessor(0) == Pred1) { 342 IfTrue = Pred1; 343 IfFalse = Pred2; 344 } else { 345 IfTrue = Pred2; 346 IfFalse = Pred1; 347 } 348 return BI->getCondition(); 349 } 350 return 0; 351 } 352 353 /// DominatesMergePoint - If we have a merge point of an "if condition" as 354 /// accepted above, return true if the specified value dominates the block. We 355 /// don't handle the true generality of domination here, just a special case 356 /// which works well enough for us. 357 /// 358 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 359 /// see if V (which must be an instruction) is cheap to compute and is 360 /// non-trapping. If both are true, the instruction is inserted into the set 361 /// and true is returned. 362 static bool DominatesMergePoint(Value *V, BasicBlock *BB, 363 std::set<Instruction*> *AggressiveInsts) { 364 Instruction *I = dyn_cast<Instruction>(V); 365 if (!I) { 366 // Non-instructions all dominate instructions, but not all constantexprs 367 // can be executed unconditionally. 368 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 369 if (C->canTrap()) 370 return false; 371 return true; 372 } 373 BasicBlock *PBB = I->getParent(); 374 375 // We don't want to allow weird loops that might have the "if condition" in 376 // the bottom of this block. 377 if (PBB == BB) return false; 378 379 // If this instruction is defined in a block that contains an unconditional 380 // branch to BB, then it must be in the 'conditional' part of the "if 381 // statement". 382 if (BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator())) 383 if (BI->isUnconditional() && BI->getSuccessor(0) == BB) { 384 if (!AggressiveInsts) return false; 385 // Okay, it looks like the instruction IS in the "condition". Check to 386 // see if its a cheap instruction to unconditionally compute, and if it 387 // only uses stuff defined outside of the condition. If so, hoist it out. 388 if (!I->isSafeToSpeculativelyExecute()) 389 return false; 390 391 switch (I->getOpcode()) { 392 default: return false; // Cannot hoist this out safely. 393 case Instruction::Load: { 394 // We have to check to make sure there are no instructions before the 395 // load in its basic block, as we are going to hoist the loop out to 396 // its predecessor. 397 BasicBlock::iterator IP = PBB->begin(); 398 while (isa<DbgInfoIntrinsic>(IP)) 399 IP++; 400 if (IP != BasicBlock::iterator(I)) 401 return false; 402 break; 403 } 404 case Instruction::Add: 405 case Instruction::Sub: 406 case Instruction::And: 407 case Instruction::Or: 408 case Instruction::Xor: 409 case Instruction::Shl: 410 case Instruction::LShr: 411 case Instruction::AShr: 412 case Instruction::ICmp: 413 break; // These are all cheap and non-trapping instructions. 414 } 415 416 // Okay, we can only really hoist these out if their operands are not 417 // defined in the conditional region. 418 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 419 if (!DominatesMergePoint(*i, BB, 0)) 420 return false; 421 // Okay, it's safe to do this! Remember this instruction. 422 AggressiveInsts->insert(I); 423 } 424 425 return true; 426 } 427 428 /// GatherConstantSetEQs - Given a potentially 'or'd together collection of 429 /// icmp_eq instructions that compare a value against a constant, return the 430 /// value being compared, and stick the constant into the Values vector. 431 static Value *GatherConstantSetEQs(Value *V, std::vector<ConstantInt*> &Values){ 432 if (Instruction *Inst = dyn_cast<Instruction>(V)) { 433 if (Inst->getOpcode() == Instruction::ICmp && 434 cast<ICmpInst>(Inst)->getPredicate() == ICmpInst::ICMP_EQ) { 435 if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(1))) { 436 Values.push_back(C); 437 return Inst->getOperand(0); 438 } else if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(0))) { 439 Values.push_back(C); 440 return Inst->getOperand(1); 441 } 442 } else if (Inst->getOpcode() == Instruction::Or) { 443 if (Value *LHS = GatherConstantSetEQs(Inst->getOperand(0), Values)) 444 if (Value *RHS = GatherConstantSetEQs(Inst->getOperand(1), Values)) 445 if (LHS == RHS) 446 return LHS; 447 } 448 } 449 return 0; 450 } 451 452 /// GatherConstantSetNEs - Given a potentially 'and'd together collection of 453 /// setne instructions that compare a value against a constant, return the value 454 /// being compared, and stick the constant into the Values vector. 455 static Value *GatherConstantSetNEs(Value *V, std::vector<ConstantInt*> &Values){ 456 if (Instruction *Inst = dyn_cast<Instruction>(V)) { 457 if (Inst->getOpcode() == Instruction::ICmp && 458 cast<ICmpInst>(Inst)->getPredicate() == ICmpInst::ICMP_NE) { 459 if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(1))) { 460 Values.push_back(C); 461 return Inst->getOperand(0); 462 } else if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(0))) { 463 Values.push_back(C); 464 return Inst->getOperand(1); 465 } 466 } else if (Inst->getOpcode() == Instruction::And) { 467 if (Value *LHS = GatherConstantSetNEs(Inst->getOperand(0), Values)) 468 if (Value *RHS = GatherConstantSetNEs(Inst->getOperand(1), Values)) 469 if (LHS == RHS) 470 return LHS; 471 } 472 } 473 return 0; 474 } 475 476 /// GatherValueComparisons - If the specified Cond is an 'and' or 'or' of a 477 /// bunch of comparisons of one value against constants, return the value and 478 /// the constants being compared. 479 static bool GatherValueComparisons(Instruction *Cond, Value *&CompVal, 480 std::vector<ConstantInt*> &Values) { 481 if (Cond->getOpcode() == Instruction::Or) { 482 CompVal = GatherConstantSetEQs(Cond, Values); 483 484 // Return true to indicate that the condition is true if the CompVal is 485 // equal to one of the constants. 486 return true; 487 } else if (Cond->getOpcode() == Instruction::And) { 488 CompVal = GatherConstantSetNEs(Cond, Values); 489 490 // Return false to indicate that the condition is false if the CompVal is 491 // equal to one of the constants. 492 return false; 493 } 494 return false; 495 } 496 497 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { 498 Instruction* Cond = 0; 499 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 500 Cond = dyn_cast<Instruction>(SI->getCondition()); 501 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 502 if (BI->isConditional()) 503 Cond = dyn_cast<Instruction>(BI->getCondition()); 504 } 505 506 TI->eraseFromParent(); 507 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); 508 } 509 510 /// isValueEqualityComparison - Return true if the specified terminator checks 511 /// to see if a value is equal to constant integer value. 512 static Value *isValueEqualityComparison(TerminatorInst *TI) { 513 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 514 // Do not permit merging of large switch instructions into their 515 // predecessors unless there is only one predecessor. 516 if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()), 517 pred_end(SI->getParent())) > 128) 518 return 0; 519 520 return SI->getCondition(); 521 } 522 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 523 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 524 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 525 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 526 ICI->getPredicate() == ICmpInst::ICMP_NE) && 527 isa<ConstantInt>(ICI->getOperand(1))) 528 return ICI->getOperand(0); 529 return 0; 530 } 531 532 /// GetValueEqualityComparisonCases - Given a value comparison instruction, 533 /// decode all of the 'cases' that it represents and return the 'default' block. 534 static BasicBlock * 535 GetValueEqualityComparisonCases(TerminatorInst *TI, 536 std::vector<std::pair<ConstantInt*, 537 BasicBlock*> > &Cases) { 538 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 539 Cases.reserve(SI->getNumCases()); 540 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 541 Cases.push_back(std::make_pair(SI->getCaseValue(i), SI->getSuccessor(i))); 542 return SI->getDefaultDest(); 543 } 544 545 BranchInst *BI = cast<BranchInst>(TI); 546 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 547 Cases.push_back(std::make_pair(cast<ConstantInt>(ICI->getOperand(1)), 548 BI->getSuccessor(ICI->getPredicate() == 549 ICmpInst::ICMP_NE))); 550 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 551 } 552 553 554 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 555 /// in the list that match the specified block. 556 static void EliminateBlockCases(BasicBlock *BB, 557 std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases) { 558 for (unsigned i = 0, e = Cases.size(); i != e; ++i) 559 if (Cases[i].second == BB) { 560 Cases.erase(Cases.begin()+i); 561 --i; --e; 562 } 563 } 564 565 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 566 /// well. 567 static bool 568 ValuesOverlap(std::vector<std::pair<ConstantInt*, BasicBlock*> > &C1, 569 std::vector<std::pair<ConstantInt*, BasicBlock*> > &C2) { 570 std::vector<std::pair<ConstantInt*, BasicBlock*> > *V1 = &C1, *V2 = &C2; 571 572 // Make V1 be smaller than V2. 573 if (V1->size() > V2->size()) 574 std::swap(V1, V2); 575 576 if (V1->size() == 0) return false; 577 if (V1->size() == 1) { 578 // Just scan V2. 579 ConstantInt *TheVal = (*V1)[0].first; 580 for (unsigned i = 0, e = V2->size(); i != e; ++i) 581 if (TheVal == (*V2)[i].first) 582 return true; 583 } 584 585 // Otherwise, just sort both lists and compare element by element. 586 std::sort(V1->begin(), V1->end()); 587 std::sort(V2->begin(), V2->end()); 588 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 589 while (i1 != e1 && i2 != e2) { 590 if ((*V1)[i1].first == (*V2)[i2].first) 591 return true; 592 if ((*V1)[i1].first < (*V2)[i2].first) 593 ++i1; 594 else 595 ++i2; 596 } 597 return false; 598 } 599 600 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 601 /// terminator instruction and its block is known to only have a single 602 /// predecessor block, check to see if that predecessor is also a value 603 /// comparison with the same value, and if that comparison determines the 604 /// outcome of this comparison. If so, simplify TI. This does a very limited 605 /// form of jump threading. 606 static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 607 BasicBlock *Pred) { 608 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 609 if (!PredVal) return false; // Not a value comparison in predecessor. 610 611 Value *ThisVal = isValueEqualityComparison(TI); 612 assert(ThisVal && "This isn't a value comparison!!"); 613 if (ThisVal != PredVal) return false; // Different predicates. 614 615 // Find out information about when control will move from Pred to TI's block. 616 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 617 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 618 PredCases); 619 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 620 621 // Find information about how control leaves this block. 622 std::vector<std::pair<ConstantInt*, BasicBlock*> > ThisCases; 623 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 624 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 625 626 // If TI's block is the default block from Pred's comparison, potentially 627 // simplify TI based on this knowledge. 628 if (PredDef == TI->getParent()) { 629 // If we are here, we know that the value is none of those cases listed in 630 // PredCases. If there are any cases in ThisCases that are in PredCases, we 631 // can simplify TI. 632 if (ValuesOverlap(PredCases, ThisCases)) { 633 if (isa<BranchInst>(TI)) { 634 // Okay, one of the successors of this condbr is dead. Convert it to a 635 // uncond br. 636 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 637 // Insert the new branch. 638 Instruction *NI = BranchInst::Create(ThisDef, TI); 639 640 // Remove PHI node entries for the dead edge. 641 ThisCases[0].second->removePredecessor(TI->getParent()); 642 643 DOUT << "Threading pred instr: " << *Pred->getTerminator() 644 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"; 645 646 EraseTerminatorInstAndDCECond(TI); 647 return true; 648 649 } else { 650 SwitchInst *SI = cast<SwitchInst>(TI); 651 // Okay, TI has cases that are statically dead, prune them away. 652 SmallPtrSet<Constant*, 16> DeadCases; 653 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 654 DeadCases.insert(PredCases[i].first); 655 656 DOUT << "Threading pred instr: " << *Pred->getTerminator() 657 << "Through successor TI: " << *TI; 658 659 for (unsigned i = SI->getNumCases()-1; i != 0; --i) 660 if (DeadCases.count(SI->getCaseValue(i))) { 661 SI->getSuccessor(i)->removePredecessor(TI->getParent()); 662 SI->removeCase(i); 663 } 664 665 DOUT << "Leaving: " << *TI << "\n"; 666 return true; 667 } 668 } 669 670 } else { 671 // Otherwise, TI's block must correspond to some matched value. Find out 672 // which value (or set of values) this is. 673 ConstantInt *TIV = 0; 674 BasicBlock *TIBB = TI->getParent(); 675 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 676 if (PredCases[i].second == TIBB) { 677 if (TIV == 0) 678 TIV = PredCases[i].first; 679 else 680 return false; // Cannot handle multiple values coming to this block. 681 } 682 assert(TIV && "No edge from pred to succ?"); 683 684 // Okay, we found the one constant that our value can be if we get into TI's 685 // BB. Find out which successor will unconditionally be branched to. 686 BasicBlock *TheRealDest = 0; 687 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 688 if (ThisCases[i].first == TIV) { 689 TheRealDest = ThisCases[i].second; 690 break; 691 } 692 693 // If not handled by any explicit cases, it is handled by the default case. 694 if (TheRealDest == 0) TheRealDest = ThisDef; 695 696 // Remove PHI node entries for dead edges. 697 BasicBlock *CheckEdge = TheRealDest; 698 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 699 if (*SI != CheckEdge) 700 (*SI)->removePredecessor(TIBB); 701 else 702 CheckEdge = 0; 703 704 // Insert the new branch. 705 Instruction *NI = BranchInst::Create(TheRealDest, TI); 706 707 DOUT << "Threading pred instr: " << *Pred->getTerminator() 708 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"; 709 710 EraseTerminatorInstAndDCECond(TI); 711 return true; 712 } 713 return false; 714 } 715 716 namespace { 717 /// ConstantIntOrdering - This class implements a stable ordering of constant 718 /// integers that does not depend on their address. This is important for 719 /// applications that sort ConstantInt's to ensure uniqueness. 720 struct ConstantIntOrdering { 721 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 722 return LHS->getValue().ult(RHS->getValue()); 723 } 724 }; 725 } 726 727 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value 728 /// equality comparison instruction (either a switch or a branch on "X == c"). 729 /// See if any of the predecessors of the terminator block are value comparisons 730 /// on the same value. If so, and if safe to do so, fold them together. 731 static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { 732 BasicBlock *BB = TI->getParent(); 733 Value *CV = isValueEqualityComparison(TI); // CondVal 734 assert(CV && "Not a comparison?"); 735 bool Changed = false; 736 737 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 738 while (!Preds.empty()) { 739 BasicBlock *Pred = Preds.pop_back_val(); 740 741 // See if the predecessor is a comparison with the same value. 742 TerminatorInst *PTI = Pred->getTerminator(); 743 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 744 745 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 746 // Figure out which 'cases' to copy from SI to PSI. 747 std::vector<std::pair<ConstantInt*, BasicBlock*> > BBCases; 748 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 749 750 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 751 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 752 753 // Based on whether the default edge from PTI goes to BB or not, fill in 754 // PredCases and PredDefault with the new switch cases we would like to 755 // build. 756 SmallVector<BasicBlock*, 8> NewSuccessors; 757 758 if (PredDefault == BB) { 759 // If this is the default destination from PTI, only the edges in TI 760 // that don't occur in PTI, or that branch to BB will be activated. 761 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 762 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 763 if (PredCases[i].second != BB) 764 PTIHandled.insert(PredCases[i].first); 765 else { 766 // The default destination is BB, we don't need explicit targets. 767 std::swap(PredCases[i], PredCases.back()); 768 PredCases.pop_back(); 769 --i; --e; 770 } 771 772 // Reconstruct the new switch statement we will be building. 773 if (PredDefault != BBDefault) { 774 PredDefault->removePredecessor(Pred); 775 PredDefault = BBDefault; 776 NewSuccessors.push_back(BBDefault); 777 } 778 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 779 if (!PTIHandled.count(BBCases[i].first) && 780 BBCases[i].second != BBDefault) { 781 PredCases.push_back(BBCases[i]); 782 NewSuccessors.push_back(BBCases[i].second); 783 } 784 785 } else { 786 // If this is not the default destination from PSI, only the edges 787 // in SI that occur in PSI with a destination of BB will be 788 // activated. 789 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled; 790 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 791 if (PredCases[i].second == BB) { 792 PTIHandled.insert(PredCases[i].first); 793 std::swap(PredCases[i], PredCases.back()); 794 PredCases.pop_back(); 795 --i; --e; 796 } 797 798 // Okay, now we know which constants were sent to BB from the 799 // predecessor. Figure out where they will all go now. 800 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 801 if (PTIHandled.count(BBCases[i].first)) { 802 // If this is one we are capable of getting... 803 PredCases.push_back(BBCases[i]); 804 NewSuccessors.push_back(BBCases[i].second); 805 PTIHandled.erase(BBCases[i].first);// This constant is taken care of 806 } 807 808 // If there are any constants vectored to BB that TI doesn't handle, 809 // they must go to the default destination of TI. 810 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I = 811 PTIHandled.begin(), 812 E = PTIHandled.end(); I != E; ++I) { 813 PredCases.push_back(std::make_pair(*I, BBDefault)); 814 NewSuccessors.push_back(BBDefault); 815 } 816 } 817 818 // Okay, at this point, we know which new successor Pred will get. Make 819 // sure we update the number of entries in the PHI nodes for these 820 // successors. 821 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 822 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 823 824 // Now that the successors are updated, create the new Switch instruction. 825 SwitchInst *NewSI = SwitchInst::Create(CV, PredDefault, 826 PredCases.size(), PTI); 827 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 828 NewSI->addCase(PredCases[i].first, PredCases[i].second); 829 830 EraseTerminatorInstAndDCECond(PTI); 831 832 // Okay, last check. If BB is still a successor of PSI, then we must 833 // have an infinite loop case. If so, add an infinitely looping block 834 // to handle the case to preserve the behavior of the code. 835 BasicBlock *InfLoopBlock = 0; 836 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 837 if (NewSI->getSuccessor(i) == BB) { 838 if (InfLoopBlock == 0) { 839 // Insert it at the end of the function, because it's either code, 840 // or it won't matter if it's hot. :) 841 InfLoopBlock = BasicBlock::Create("infloop", BB->getParent()); 842 BranchInst::Create(InfLoopBlock, InfLoopBlock); 843 } 844 NewSI->setSuccessor(i, InfLoopBlock); 845 } 846 847 Changed = true; 848 } 849 } 850 return Changed; 851 } 852 853 // isSafeToHoistInvoke - If we would need to insert a select that uses the 854 // value of this invoke (comments in HoistThenElseCodeToIf explain why we 855 // would need to do this), we can't hoist the invoke, as there is nowhere 856 // to put the select in this case. 857 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2, 858 Instruction *I1, Instruction *I2) { 859 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 860 PHINode *PN; 861 for (BasicBlock::iterator BBI = SI->begin(); 862 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 863 Value *BB1V = PN->getIncomingValueForBlock(BB1); 864 Value *BB2V = PN->getIncomingValueForBlock(BB2); 865 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) { 866 return false; 867 } 868 } 869 } 870 return true; 871 } 872 873 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 874 /// BB2, hoist any common code in the two blocks up into the branch block. The 875 /// caller of this function guarantees that BI's block dominates BB1 and BB2. 876 static bool HoistThenElseCodeToIf(BranchInst *BI) { 877 // This does very trivial matching, with limited scanning, to find identical 878 // instructions in the two blocks. In particular, we don't want to get into 879 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 880 // such, we currently just scan for obviously identical instructions in an 881 // identical order. 882 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 883 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 884 885 BasicBlock::iterator BB1_Itr = BB1->begin(); 886 BasicBlock::iterator BB2_Itr = BB2->begin(); 887 888 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; 889 while (isa<DbgInfoIntrinsic>(I1)) 890 I1 = BB1_Itr++; 891 while (isa<DbgInfoIntrinsic>(I2)) 892 I2 = BB2_Itr++; 893 if (I1->getOpcode() != I2->getOpcode() || isa<PHINode>(I1) || 894 !I1->isIdenticalTo(I2) || 895 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) 896 return false; 897 898 // If we get here, we can hoist at least one instruction. 899 BasicBlock *BIParent = BI->getParent(); 900 901 do { 902 // If we are hoisting the terminator instruction, don't move one (making a 903 // broken BB), instead clone it, and remove BI. 904 if (isa<TerminatorInst>(I1)) 905 goto HoistTerminator; 906 907 // For a normal instruction, we just move one to right before the branch, 908 // then replace all uses of the other with the first. Finally, we remove 909 // the now redundant second instruction. 910 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 911 if (!I2->use_empty()) 912 I2->replaceAllUsesWith(I1); 913 BB2->getInstList().erase(I2); 914 915 I1 = BB1_Itr++; 916 while (isa<DbgInfoIntrinsic>(I1)) 917 I1 = BB1_Itr++; 918 I2 = BB2_Itr++; 919 while (isa<DbgInfoIntrinsic>(I2)) 920 I2 = BB2_Itr++; 921 } while (I1->getOpcode() == I2->getOpcode() && I1->isIdenticalTo(I2)); 922 923 return true; 924 925 HoistTerminator: 926 // It may not be possible to hoist an invoke. 927 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) 928 return true; 929 930 // Okay, it is safe to hoist the terminator. 931 Instruction *NT = I1->clone(BB1->getContext()); 932 BIParent->getInstList().insert(BI, NT); 933 if (NT->getType() != Type::VoidTy) { 934 I1->replaceAllUsesWith(NT); 935 I2->replaceAllUsesWith(NT); 936 NT->takeName(I1); 937 } 938 939 // Hoisting one of the terminators from our successor is a great thing. 940 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 941 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 942 // nodes, so we insert select instruction to compute the final result. 943 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 944 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 945 PHINode *PN; 946 for (BasicBlock::iterator BBI = SI->begin(); 947 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 948 Value *BB1V = PN->getIncomingValueForBlock(BB1); 949 Value *BB2V = PN->getIncomingValueForBlock(BB2); 950 if (BB1V != BB2V) { 951 // These values do not agree. Insert a select instruction before NT 952 // that determines the right value. 953 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 954 if (SI == 0) 955 SI = SelectInst::Create(BI->getCondition(), BB1V, BB2V, 956 BB1V->getName()+"."+BB2V->getName(), NT); 957 // Make the PHI node use the select for all incoming values for BB1/BB2 958 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 959 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 960 PN->setIncomingValue(i, SI); 961 } 962 } 963 } 964 965 // Update any PHI nodes in our new successors. 966 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 967 AddPredecessorToBlock(*SI, BIParent, BB1); 968 969 EraseTerminatorInstAndDCECond(BI); 970 return true; 971 } 972 973 /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 974 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code 975 /// (for now, restricted to a single instruction that's side effect free) from 976 /// the BB1 into the branch block to speculatively execute it. 977 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 978 // Only speculatively execution a single instruction (not counting the 979 // terminator) for now. 980 Instruction *HInst = NULL; 981 Instruction *Term = BB1->getTerminator(); 982 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end(); 983 BBI != BBE; ++BBI) { 984 Instruction *I = BBI; 985 // Skip debug info. 986 if (isa<DbgInfoIntrinsic>(I)) continue; 987 if (I == Term) break; 988 989 if (!HInst) 990 HInst = I; 991 else 992 return false; 993 } 994 if (!HInst) 995 return false; 996 997 // Be conservative for now. FP select instruction can often be expensive. 998 Value *BrCond = BI->getCondition(); 999 if (isa<Instruction>(BrCond) && 1000 cast<Instruction>(BrCond)->getOpcode() == Instruction::FCmp) 1001 return false; 1002 1003 // If BB1 is actually on the false edge of the conditional branch, remember 1004 // to swap the select operands later. 1005 bool Invert = false; 1006 if (BB1 != BI->getSuccessor(0)) { 1007 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 1008 Invert = true; 1009 } 1010 1011 // Turn 1012 // BB: 1013 // %t1 = icmp 1014 // br i1 %t1, label %BB1, label %BB2 1015 // BB1: 1016 // %t3 = add %t2, c 1017 // br label BB2 1018 // BB2: 1019 // => 1020 // BB: 1021 // %t1 = icmp 1022 // %t4 = add %t2, c 1023 // %t3 = select i1 %t1, %t2, %t3 1024 switch (HInst->getOpcode()) { 1025 default: return false; // Not safe / profitable to hoist. 1026 case Instruction::Add: 1027 case Instruction::Sub: 1028 // Not worth doing for vector ops. 1029 if (isa<VectorType>(HInst->getType())) 1030 return false; 1031 break; 1032 case Instruction::And: 1033 case Instruction::Or: 1034 case Instruction::Xor: 1035 case Instruction::Shl: 1036 case Instruction::LShr: 1037 case Instruction::AShr: 1038 // Don't mess with vector operations. 1039 if (isa<VectorType>(HInst->getType())) 1040 return false; 1041 break; // These are all cheap and non-trapping instructions. 1042 } 1043 1044 // If the instruction is obviously dead, don't try to predicate it. 1045 if (HInst->use_empty()) { 1046 HInst->eraseFromParent(); 1047 return true; 1048 } 1049 1050 // Can we speculatively execute the instruction? And what is the value 1051 // if the condition is false? Consider the phi uses, if the incoming value 1052 // from the "if" block are all the same V, then V is the value of the 1053 // select if the condition is false. 1054 BasicBlock *BIParent = BI->getParent(); 1055 SmallVector<PHINode*, 4> PHIUses; 1056 Value *FalseV = NULL; 1057 1058 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); 1059 for (Value::use_iterator UI = HInst->use_begin(), E = HInst->use_end(); 1060 UI != E; ++UI) { 1061 // Ignore any user that is not a PHI node in BB2. These can only occur in 1062 // unreachable blocks, because they would not be dominated by the instr. 1063 PHINode *PN = dyn_cast<PHINode>(UI); 1064 if (!PN || PN->getParent() != BB2) 1065 return false; 1066 PHIUses.push_back(PN); 1067 1068 Value *PHIV = PN->getIncomingValueForBlock(BIParent); 1069 if (!FalseV) 1070 FalseV = PHIV; 1071 else if (FalseV != PHIV) 1072 return false; // Inconsistent value when condition is false. 1073 } 1074 1075 assert(FalseV && "Must have at least one user, and it must be a PHI"); 1076 1077 // Do not hoist the instruction if any of its operands are defined but not 1078 // used in this BB. The transformation will prevent the operand from 1079 // being sunk into the use block. 1080 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end(); 1081 i != e; ++i) { 1082 Instruction *OpI = dyn_cast<Instruction>(*i); 1083 if (OpI && OpI->getParent() == BIParent && 1084 !OpI->isUsedInBasicBlock(BIParent)) 1085 return false; 1086 } 1087 1088 // If we get here, we can hoist the instruction. Try to place it 1089 // before the icmp instruction preceding the conditional branch. 1090 BasicBlock::iterator InsertPos = BI; 1091 if (InsertPos != BIParent->begin()) 1092 --InsertPos; 1093 // Skip debug info between condition and branch. 1094 while (InsertPos != BIParent->begin() && isa<DbgInfoIntrinsic>(InsertPos)) 1095 --InsertPos; 1096 if (InsertPos == BrCond && !isa<PHINode>(BrCond)) { 1097 SmallPtrSet<Instruction *, 4> BB1Insns; 1098 for(BasicBlock::iterator BB1I = BB1->begin(), BB1E = BB1->end(); 1099 BB1I != BB1E; ++BB1I) 1100 BB1Insns.insert(BB1I); 1101 for(Value::use_iterator UI = BrCond->use_begin(), UE = BrCond->use_end(); 1102 UI != UE; ++UI) { 1103 Instruction *Use = cast<Instruction>(*UI); 1104 if (BB1Insns.count(Use)) { 1105 // If BrCond uses the instruction that place it just before 1106 // branch instruction. 1107 InsertPos = BI; 1108 break; 1109 } 1110 } 1111 } else 1112 InsertPos = BI; 1113 BIParent->getInstList().splice(InsertPos, BB1->getInstList(), HInst); 1114 1115 // Create a select whose true value is the speculatively executed value and 1116 // false value is the previously determined FalseV. 1117 SelectInst *SI; 1118 if (Invert) 1119 SI = SelectInst::Create(BrCond, FalseV, HInst, 1120 FalseV->getName() + "." + HInst->getName(), BI); 1121 else 1122 SI = SelectInst::Create(BrCond, HInst, FalseV, 1123 HInst->getName() + "." + FalseV->getName(), BI); 1124 1125 // Make the PHI node use the select for all incoming values for "then" and 1126 // "if" blocks. 1127 for (unsigned i = 0, e = PHIUses.size(); i != e; ++i) { 1128 PHINode *PN = PHIUses[i]; 1129 for (unsigned j = 0, ee = PN->getNumIncomingValues(); j != ee; ++j) 1130 if (PN->getIncomingBlock(j) == BB1 || 1131 PN->getIncomingBlock(j) == BIParent) 1132 PN->setIncomingValue(j, SI); 1133 } 1134 1135 ++NumSpeculations; 1136 return true; 1137 } 1138 1139 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1140 /// across this block. 1141 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1142 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1143 unsigned Size = 0; 1144 1145 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1146 if (isa<DbgInfoIntrinsic>(BBI)) 1147 continue; 1148 if (Size > 10) return false; // Don't clone large BB's. 1149 ++Size; 1150 1151 // We can only support instructions that do not define values that are 1152 // live outside of the current basic block. 1153 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1154 UI != E; ++UI) { 1155 Instruction *U = cast<Instruction>(*UI); 1156 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1157 } 1158 1159 // Looks ok, continue checking. 1160 } 1161 1162 return true; 1163 } 1164 1165 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1166 /// that is defined in the same block as the branch and if any PHI entries are 1167 /// constants, thread edges corresponding to that entry to be branches to their 1168 /// ultimate destination. 1169 static bool FoldCondBranchOnPHI(BranchInst *BI) { 1170 BasicBlock *BB = BI->getParent(); 1171 LLVMContext &Context = BB->getContext(); 1172 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1173 // NOTE: we currently cannot transform this case if the PHI node is used 1174 // outside of the block. 1175 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1176 return false; 1177 1178 // Degenerate case of a single entry PHI. 1179 if (PN->getNumIncomingValues() == 1) { 1180 FoldSingleEntryPHINodes(PN->getParent()); 1181 return true; 1182 } 1183 1184 // Now we know that this block has multiple preds and two succs. 1185 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1186 1187 // Okay, this is a simple enough basic block. See if any phi values are 1188 // constants. 1189 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1190 ConstantInt *CB; 1191 if ((CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i))) && 1192 CB->getType() == Type::Int1Ty) { 1193 // Okay, we now know that all edges from PredBB should be revectored to 1194 // branch to RealDest. 1195 BasicBlock *PredBB = PN->getIncomingBlock(i); 1196 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1197 1198 if (RealDest == BB) continue; // Skip self loops. 1199 1200 // The dest block might have PHI nodes, other predecessors and other 1201 // difficult cases. Instead of being smart about this, just insert a new 1202 // block that jumps to the destination block, effectively splitting 1203 // the edge we are about to create. 1204 BasicBlock *EdgeBB = BasicBlock::Create(RealDest->getName()+".critedge", 1205 RealDest->getParent(), RealDest); 1206 BranchInst::Create(RealDest, EdgeBB); 1207 PHINode *PN; 1208 for (BasicBlock::iterator BBI = RealDest->begin(); 1209 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1210 Value *V = PN->getIncomingValueForBlock(BB); 1211 PN->addIncoming(V, EdgeBB); 1212 } 1213 1214 // BB may have instructions that are being threaded over. Clone these 1215 // instructions into EdgeBB. We know that there will be no uses of the 1216 // cloned instructions outside of EdgeBB. 1217 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1218 std::map<Value*, Value*> TranslateMap; // Track translated values. 1219 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1220 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1221 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1222 } else { 1223 // Clone the instruction. 1224 Instruction *N = BBI->clone(Context); 1225 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1226 1227 // Update operands due to translation. 1228 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1229 i != e; ++i) { 1230 std::map<Value*, Value*>::iterator PI = 1231 TranslateMap.find(*i); 1232 if (PI != TranslateMap.end()) 1233 *i = PI->second; 1234 } 1235 1236 // Check for trivial simplification. 1237 if (Constant *C = ConstantFoldInstruction(N, Context)) { 1238 TranslateMap[BBI] = C; 1239 delete N; // Constant folded away, don't need actual inst 1240 } else { 1241 // Insert the new instruction into its new home. 1242 EdgeBB->getInstList().insert(InsertPt, N); 1243 if (!BBI->use_empty()) 1244 TranslateMap[BBI] = N; 1245 } 1246 } 1247 } 1248 1249 // Loop over all of the edges from PredBB to BB, changing them to branch 1250 // to EdgeBB instead. 1251 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1252 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1253 if (PredBBTI->getSuccessor(i) == BB) { 1254 BB->removePredecessor(PredBB); 1255 PredBBTI->setSuccessor(i, EdgeBB); 1256 } 1257 1258 // Recurse, simplifying any other constants. 1259 return FoldCondBranchOnPHI(BI) | true; 1260 } 1261 } 1262 1263 return false; 1264 } 1265 1266 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1267 /// PHI node, see if we can eliminate it. 1268 static bool FoldTwoEntryPHINode(PHINode *PN) { 1269 LLVMContext &Context = PN->getParent()->getContext(); 1270 1271 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1272 // statement", which has a very simple dominance structure. Basically, we 1273 // are trying to find the condition that is being branched on, which 1274 // subsequently causes this merge to happen. We really want control 1275 // dependence information for this check, but simplifycfg can't keep it up 1276 // to date, and this catches most of the cases we care about anyway. 1277 // 1278 BasicBlock *BB = PN->getParent(); 1279 BasicBlock *IfTrue, *IfFalse; 1280 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1281 if (!IfCond) return false; 1282 1283 // Okay, we found that we can merge this two-entry phi node into a select. 1284 // Doing so would require us to fold *all* two entry phi nodes in this block. 1285 // At some point this becomes non-profitable (particularly if the target 1286 // doesn't support cmov's). Only do this transformation if there are two or 1287 // fewer PHI nodes in this block. 1288 unsigned NumPhis = 0; 1289 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1290 if (NumPhis > 2) 1291 return false; 1292 1293 DEBUG(errs() << "FOUND IF CONDITION! " << *IfCond << " T: " 1294 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); 1295 1296 // Loop over the PHI's seeing if we can promote them all to select 1297 // instructions. While we are at it, keep track of the instructions 1298 // that need to be moved to the dominating block. 1299 std::set<Instruction*> AggressiveInsts; 1300 1301 BasicBlock::iterator AfterPHIIt = BB->begin(); 1302 while (isa<PHINode>(AfterPHIIt)) { 1303 PHINode *PN = cast<PHINode>(AfterPHIIt++); 1304 if (PN->getIncomingValue(0) == PN->getIncomingValue(1)) { 1305 if (PN->getIncomingValue(0) != PN) 1306 PN->replaceAllUsesWith(PN->getIncomingValue(0)); 1307 else 1308 PN->replaceAllUsesWith(Context.getUndef(PN->getType())); 1309 } else if (!DominatesMergePoint(PN->getIncomingValue(0), BB, 1310 &AggressiveInsts) || 1311 !DominatesMergePoint(PN->getIncomingValue(1), BB, 1312 &AggressiveInsts)) { 1313 return false; 1314 } 1315 } 1316 1317 // If we all PHI nodes are promotable, check to make sure that all 1318 // instructions in the predecessor blocks can be promoted as well. If 1319 // not, we won't be able to get rid of the control flow, so it's not 1320 // worth promoting to select instructions. 1321 BasicBlock *DomBlock = 0, *IfBlock1 = 0, *IfBlock2 = 0; 1322 PN = cast<PHINode>(BB->begin()); 1323 BasicBlock *Pred = PN->getIncomingBlock(0); 1324 if (cast<BranchInst>(Pred->getTerminator())->isUnconditional()) { 1325 IfBlock1 = Pred; 1326 DomBlock = *pred_begin(Pred); 1327 for (BasicBlock::iterator I = Pred->begin(); 1328 !isa<TerminatorInst>(I); ++I) 1329 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1330 // This is not an aggressive instruction that we can promote. 1331 // Because of this, we won't be able to get rid of the control 1332 // flow, so the xform is not worth it. 1333 return false; 1334 } 1335 } 1336 1337 Pred = PN->getIncomingBlock(1); 1338 if (cast<BranchInst>(Pred->getTerminator())->isUnconditional()) { 1339 IfBlock2 = Pred; 1340 DomBlock = *pred_begin(Pred); 1341 for (BasicBlock::iterator I = Pred->begin(); 1342 !isa<TerminatorInst>(I); ++I) 1343 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) { 1344 // This is not an aggressive instruction that we can promote. 1345 // Because of this, we won't be able to get rid of the control 1346 // flow, so the xform is not worth it. 1347 return false; 1348 } 1349 } 1350 1351 // If we can still promote the PHI nodes after this gauntlet of tests, 1352 // do all of the PHI's now. 1353 1354 // Move all 'aggressive' instructions, which are defined in the 1355 // conditional parts of the if's up to the dominating block. 1356 if (IfBlock1) { 1357 DomBlock->getInstList().splice(DomBlock->getTerminator(), 1358 IfBlock1->getInstList(), 1359 IfBlock1->begin(), 1360 IfBlock1->getTerminator()); 1361 } 1362 if (IfBlock2) { 1363 DomBlock->getInstList().splice(DomBlock->getTerminator(), 1364 IfBlock2->getInstList(), 1365 IfBlock2->begin(), 1366 IfBlock2->getTerminator()); 1367 } 1368 1369 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1370 // Change the PHI node into a select instruction. 1371 Value *TrueVal = 1372 PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1373 Value *FalseVal = 1374 PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1375 1376 Value *NV = SelectInst::Create(IfCond, TrueVal, FalseVal, "", AfterPHIIt); 1377 PN->replaceAllUsesWith(NV); 1378 NV->takeName(PN); 1379 1380 BB->getInstList().erase(PN); 1381 } 1382 return true; 1383 } 1384 1385 /// isTerminatorFirstRelevantInsn - Return true if Term is very first 1386 /// instruction ignoring Phi nodes and dbg intrinsics. 1387 static bool isTerminatorFirstRelevantInsn(BasicBlock *BB, Instruction *Term) { 1388 BasicBlock::iterator BBI = Term; 1389 while (BBI != BB->begin()) { 1390 --BBI; 1391 if (!isa<DbgInfoIntrinsic>(BBI)) 1392 break; 1393 } 1394 1395 if (isa<PHINode>(BBI) || &*BBI == Term || isa<DbgInfoIntrinsic>(BBI)) 1396 return true; 1397 return false; 1398 } 1399 1400 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1401 /// to two returning blocks, try to merge them together into one return, 1402 /// introducing a select if the return values disagree. 1403 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI) { 1404 assert(BI->isConditional() && "Must be a conditional branch"); 1405 BasicBlock *TrueSucc = BI->getSuccessor(0); 1406 BasicBlock *FalseSucc = BI->getSuccessor(1); 1407 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1408 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1409 1410 // Check to ensure both blocks are empty (just a return) or optionally empty 1411 // with PHI nodes. If there are other instructions, merging would cause extra 1412 // computation on one path or the other. 1413 if (!isTerminatorFirstRelevantInsn(TrueSucc, TrueRet)) 1414 return false; 1415 if (!isTerminatorFirstRelevantInsn(FalseSucc, FalseRet)) 1416 return false; 1417 1418 // Okay, we found a branch that is going to two return nodes. If 1419 // there is no return value for this function, just change the 1420 // branch into a return. 1421 if (FalseRet->getNumOperands() == 0) { 1422 TrueSucc->removePredecessor(BI->getParent()); 1423 FalseSucc->removePredecessor(BI->getParent()); 1424 ReturnInst::Create(0, BI); 1425 EraseTerminatorInstAndDCECond(BI); 1426 return true; 1427 } 1428 1429 // Otherwise, figure out what the true and false return values are 1430 // so we can insert a new select instruction. 1431 Value *TrueValue = TrueRet->getReturnValue(); 1432 Value *FalseValue = FalseRet->getReturnValue(); 1433 1434 // Unwrap any PHI nodes in the return blocks. 1435 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1436 if (TVPN->getParent() == TrueSucc) 1437 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1438 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1439 if (FVPN->getParent() == FalseSucc) 1440 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1441 1442 // In order for this transformation to be safe, we must be able to 1443 // unconditionally execute both operands to the return. This is 1444 // normally the case, but we could have a potentially-trapping 1445 // constant expression that prevents this transformation from being 1446 // safe. 1447 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1448 if (TCV->canTrap()) 1449 return false; 1450 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1451 if (FCV->canTrap()) 1452 return false; 1453 1454 // Okay, we collected all the mapped values and checked them for sanity, and 1455 // defined to really do this transformation. First, update the CFG. 1456 TrueSucc->removePredecessor(BI->getParent()); 1457 FalseSucc->removePredecessor(BI->getParent()); 1458 1459 // Insert select instructions where needed. 1460 Value *BrCond = BI->getCondition(); 1461 if (TrueValue) { 1462 // Insert a select if the results differ. 1463 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1464 } else if (isa<UndefValue>(TrueValue)) { 1465 TrueValue = FalseValue; 1466 } else { 1467 TrueValue = SelectInst::Create(BrCond, TrueValue, 1468 FalseValue, "retval", BI); 1469 } 1470 } 1471 1472 Value *RI = !TrueValue ? 1473 ReturnInst::Create(BI) : 1474 ReturnInst::Create(TrueValue, BI); 1475 1476 DOUT << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1477 << "\n " << *BI << "NewRet = " << *RI 1478 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc; 1479 1480 EraseTerminatorInstAndDCECond(BI); 1481 1482 return true; 1483 } 1484 1485 /// FoldBranchToCommonDest - If this basic block is ONLY a setcc and a branch, 1486 /// and if a predecessor branches to us and one of our successors, fold the 1487 /// setcc into the predecessor and use logical operations to pick the right 1488 /// destination. 1489 bool llvm::FoldBranchToCommonDest(BranchInst *BI) { 1490 BasicBlock *BB = BI->getParent(); 1491 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 1492 if (Cond == 0) return false; 1493 1494 1495 // Only allow this if the condition is a simple instruction that can be 1496 // executed unconditionally. It must be in the same block as the branch, and 1497 // must be at the front of the block. 1498 BasicBlock::iterator FrontIt = BB->front(); 1499 // Ignore dbg intrinsics. 1500 while(isa<DbgInfoIntrinsic>(FrontIt)) 1501 ++FrontIt; 1502 if ((!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1503 Cond->getParent() != BB || &*FrontIt != Cond || !Cond->hasOneUse()) { 1504 return false; 1505 } 1506 1507 // Make sure the instruction after the condition is the cond branch. 1508 BasicBlock::iterator CondIt = Cond; ++CondIt; 1509 // Ingore dbg intrinsics. 1510 while(isa<DbgInfoIntrinsic>(CondIt)) 1511 ++CondIt; 1512 if (&*CondIt != BI) { 1513 assert (!isa<DbgInfoIntrinsic>(CondIt) && "Hey do not forget debug info!"); 1514 return false; 1515 } 1516 1517 // Cond is known to be a compare or binary operator. Check to make sure that 1518 // neither operand is a potentially-trapping constant expression. 1519 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0))) 1520 if (CE->canTrap()) 1521 return false; 1522 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1))) 1523 if (CE->canTrap()) 1524 return false; 1525 1526 1527 // Finally, don't infinitely unroll conditional loops. 1528 BasicBlock *TrueDest = BI->getSuccessor(0); 1529 BasicBlock *FalseDest = BI->getSuccessor(1); 1530 if (TrueDest == BB || FalseDest == BB) 1531 return false; 1532 1533 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1534 BasicBlock *PredBlock = *PI; 1535 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1536 1537 // Check that we have two conditional branches. If there is a PHI node in 1538 // the common successor, verify that the same value flows in from both 1539 // blocks. 1540 if (PBI == 0 || PBI->isUnconditional() || 1541 !SafeToMergeTerminators(BI, PBI)) 1542 continue; 1543 1544 Instruction::BinaryOps Opc; 1545 bool InvertPredCond = false; 1546 1547 if (PBI->getSuccessor(0) == TrueDest) 1548 Opc = Instruction::Or; 1549 else if (PBI->getSuccessor(1) == FalseDest) 1550 Opc = Instruction::And; 1551 else if (PBI->getSuccessor(0) == FalseDest) 1552 Opc = Instruction::And, InvertPredCond = true; 1553 else if (PBI->getSuccessor(1) == TrueDest) 1554 Opc = Instruction::Or, InvertPredCond = true; 1555 else 1556 continue; 1557 1558 DOUT << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB; 1559 1560 // If we need to invert the condition in the pred block to match, do so now. 1561 if (InvertPredCond) { 1562 Value *NewCond = 1563 BinaryOperator::CreateNot(BI->getParent()->getContext(), 1564 PBI->getCondition(), 1565 PBI->getCondition()->getName()+".not", PBI); 1566 PBI->setCondition(NewCond); 1567 BasicBlock *OldTrue = PBI->getSuccessor(0); 1568 BasicBlock *OldFalse = PBI->getSuccessor(1); 1569 PBI->setSuccessor(0, OldFalse); 1570 PBI->setSuccessor(1, OldTrue); 1571 } 1572 1573 // Clone Cond into the predecessor basic block, and or/and the 1574 // two conditions together. 1575 Instruction *New = Cond->clone(BB->getContext()); 1576 PredBlock->getInstList().insert(PBI, New); 1577 New->takeName(Cond); 1578 Cond->setName(New->getName()+".old"); 1579 1580 Value *NewCond = BinaryOperator::Create(Opc, PBI->getCondition(), 1581 New, "or.cond", PBI); 1582 PBI->setCondition(NewCond); 1583 if (PBI->getSuccessor(0) == BB) { 1584 AddPredecessorToBlock(TrueDest, PredBlock, BB); 1585 PBI->setSuccessor(0, TrueDest); 1586 } 1587 if (PBI->getSuccessor(1) == BB) { 1588 AddPredecessorToBlock(FalseDest, PredBlock, BB); 1589 PBI->setSuccessor(1, FalseDest); 1590 } 1591 return true; 1592 } 1593 return false; 1594 } 1595 1596 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 1597 /// predecessor of another block, this function tries to simplify it. We know 1598 /// that PBI and BI are both conditional branches, and BI is in one of the 1599 /// successor blocks of PBI - PBI branches to BI. 1600 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 1601 assert(PBI->isConditional() && BI->isConditional()); 1602 BasicBlock *BB = BI->getParent(); 1603 LLVMContext &Context = BB->getContext(); 1604 1605 // If this block ends with a branch instruction, and if there is a 1606 // predecessor that ends on a branch of the same condition, make 1607 // this conditional branch redundant. 1608 if (PBI->getCondition() == BI->getCondition() && 1609 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1610 // Okay, the outcome of this conditional branch is statically 1611 // knowable. If this block had a single pred, handle specially. 1612 if (BB->getSinglePredecessor()) { 1613 // Turn this into a branch on constant. 1614 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1615 BI->setCondition(ConstantInt::get(Type::Int1Ty, CondIsTrue)); 1616 return true; // Nuke the branch on constant. 1617 } 1618 1619 // Otherwise, if there are multiple predecessors, insert a PHI that merges 1620 // in the constant and simplify the block result. Subsequent passes of 1621 // simplifycfg will thread the block. 1622 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 1623 PHINode *NewPN = PHINode::Create(Type::Int1Ty, 1624 BI->getCondition()->getName() + ".pr", 1625 BB->begin()); 1626 // Okay, we're going to insert the PHI node. Since PBI is not the only 1627 // predecessor, compute the PHI'd conditional value for all of the preds. 1628 // Any predecessor where the condition is not computable we keep symbolic. 1629 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 1630 if ((PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) && 1631 PBI != BI && PBI->isConditional() && 1632 PBI->getCondition() == BI->getCondition() && 1633 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1634 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1635 NewPN->addIncoming(ConstantInt::get(Type::Int1Ty, 1636 CondIsTrue), *PI); 1637 } else { 1638 NewPN->addIncoming(BI->getCondition(), *PI); 1639 } 1640 1641 BI->setCondition(NewPN); 1642 return true; 1643 } 1644 } 1645 1646 // If this is a conditional branch in an empty block, and if any 1647 // predecessors is a conditional branch to one of our destinations, 1648 // fold the conditions into logical ops and one cond br. 1649 BasicBlock::iterator BBI = BB->begin(); 1650 // Ignore dbg intrinsics. 1651 while (isa<DbgInfoIntrinsic>(BBI)) 1652 ++BBI; 1653 if (&*BBI != BI) 1654 return false; 1655 1656 1657 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition())) 1658 if (CE->canTrap()) 1659 return false; 1660 1661 int PBIOp, BIOp; 1662 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 1663 PBIOp = BIOp = 0; 1664 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 1665 PBIOp = 0, BIOp = 1; 1666 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 1667 PBIOp = 1, BIOp = 0; 1668 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 1669 PBIOp = BIOp = 1; 1670 else 1671 return false; 1672 1673 // Check to make sure that the other destination of this branch 1674 // isn't BB itself. If so, this is an infinite loop that will 1675 // keep getting unwound. 1676 if (PBI->getSuccessor(PBIOp) == BB) 1677 return false; 1678 1679 // Do not perform this transformation if it would require 1680 // insertion of a large number of select instructions. For targets 1681 // without predication/cmovs, this is a big pessimization. 1682 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 1683 1684 unsigned NumPhis = 0; 1685 for (BasicBlock::iterator II = CommonDest->begin(); 1686 isa<PHINode>(II); ++II, ++NumPhis) 1687 if (NumPhis > 2) // Disable this xform. 1688 return false; 1689 1690 // Finally, if everything is ok, fold the branches to logical ops. 1691 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 1692 1693 DOUT << "FOLDING BRs:" << *PBI->getParent() 1694 << "AND: " << *BI->getParent(); 1695 1696 1697 // If OtherDest *is* BB, then BB is a basic block with a single conditional 1698 // branch in it, where one edge (OtherDest) goes back to itself but the other 1699 // exits. We don't *know* that the program avoids the infinite loop 1700 // (even though that seems likely). If we do this xform naively, we'll end up 1701 // recursively unpeeling the loop. Since we know that (after the xform is 1702 // done) that the block *is* infinite if reached, we just make it an obviously 1703 // infinite loop with no cond branch. 1704 if (OtherDest == BB) { 1705 // Insert it at the end of the function, because it's either code, 1706 // or it won't matter if it's hot. :) 1707 BasicBlock *InfLoopBlock = BasicBlock::Create("infloop", BB->getParent()); 1708 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1709 OtherDest = InfLoopBlock; 1710 } 1711 1712 DOUT << *PBI->getParent()->getParent(); 1713 1714 // BI may have other predecessors. Because of this, we leave 1715 // it alone, but modify PBI. 1716 1717 // Make sure we get to CommonDest on True&True directions. 1718 Value *PBICond = PBI->getCondition(); 1719 if (PBIOp) 1720 PBICond = BinaryOperator::CreateNot(Context, PBICond, 1721 PBICond->getName()+".not", 1722 PBI); 1723 Value *BICond = BI->getCondition(); 1724 if (BIOp) 1725 BICond = BinaryOperator::CreateNot(Context, BICond, 1726 BICond->getName()+".not", 1727 PBI); 1728 // Merge the conditions. 1729 Value *Cond = BinaryOperator::CreateOr(PBICond, BICond, "brmerge", PBI); 1730 1731 // Modify PBI to branch on the new condition to the new dests. 1732 PBI->setCondition(Cond); 1733 PBI->setSuccessor(0, CommonDest); 1734 PBI->setSuccessor(1, OtherDest); 1735 1736 // OtherDest may have phi nodes. If so, add an entry from PBI's 1737 // block that are identical to the entries for BI's block. 1738 PHINode *PN; 1739 for (BasicBlock::iterator II = OtherDest->begin(); 1740 (PN = dyn_cast<PHINode>(II)); ++II) { 1741 Value *V = PN->getIncomingValueForBlock(BB); 1742 PN->addIncoming(V, PBI->getParent()); 1743 } 1744 1745 // We know that the CommonDest already had an edge from PBI to 1746 // it. If it has PHIs though, the PHIs may have different 1747 // entries for BB and PBI's BB. If so, insert a select to make 1748 // them agree. 1749 for (BasicBlock::iterator II = CommonDest->begin(); 1750 (PN = dyn_cast<PHINode>(II)); ++II) { 1751 Value *BIV = PN->getIncomingValueForBlock(BB); 1752 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 1753 Value *PBIV = PN->getIncomingValue(PBBIdx); 1754 if (BIV != PBIV) { 1755 // Insert a select in PBI to pick the right value. 1756 Value *NV = SelectInst::Create(PBICond, PBIV, BIV, 1757 PBIV->getName()+".mux", PBI); 1758 PN->setIncomingValue(PBBIdx, NV); 1759 } 1760 } 1761 1762 DOUT << "INTO: " << *PBI->getParent(); 1763 1764 DOUT << *PBI->getParent()->getParent(); 1765 1766 // This basic block is probably dead. We know it has at least 1767 // one fewer predecessor. 1768 return true; 1769 } 1770 1771 1772 /// SimplifyCFG - This function is used to do simplification of a CFG. For 1773 /// example, it adjusts branches to branches to eliminate the extra hop, it 1774 /// eliminates unreachable basic blocks, and does other "peephole" optimization 1775 /// of the CFG. It returns true if a modification was made. 1776 /// 1777 /// WARNING: The entry node of a function may not be simplified. 1778 /// 1779 bool llvm::SimplifyCFG(BasicBlock *BB) { 1780 bool Changed = false; 1781 Function *M = BB->getParent(); 1782 1783 assert(BB && BB->getParent() && "Block not embedded in function!"); 1784 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 1785 assert(&BB->getParent()->getEntryBlock() != BB && 1786 "Can't Simplify entry block!"); 1787 1788 // Remove basic blocks that have no predecessors... or that just have themself 1789 // as a predecessor. These are unreachable. 1790 if (pred_begin(BB) == pred_end(BB) || BB->getSinglePredecessor() == BB) { 1791 DOUT << "Removing BB: \n" << *BB; 1792 DeleteDeadBlock(BB); 1793 return true; 1794 } 1795 1796 // Check to see if we can constant propagate this terminator instruction 1797 // away... 1798 Changed |= ConstantFoldTerminator(BB); 1799 1800 // If there is a trivial two-entry PHI node in this basic block, and we can 1801 // eliminate it, do so now. 1802 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 1803 if (PN->getNumIncomingValues() == 2) 1804 Changed |= FoldTwoEntryPHINode(PN); 1805 1806 // If this is a returning block with only PHI nodes in it, fold the return 1807 // instruction into any unconditional branch predecessors. 1808 // 1809 // If any predecessor is a conditional branch that just selects among 1810 // different return values, fold the replace the branch/return with a select 1811 // and return. 1812 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 1813 if (isTerminatorFirstRelevantInsn(BB, BB->getTerminator())) { 1814 // Find predecessors that end with branches. 1815 SmallVector<BasicBlock*, 8> UncondBranchPreds; 1816 SmallVector<BranchInst*, 8> CondBranchPreds; 1817 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1818 TerminatorInst *PTI = (*PI)->getTerminator(); 1819 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 1820 if (BI->isUnconditional()) 1821 UncondBranchPreds.push_back(*PI); 1822 else 1823 CondBranchPreds.push_back(BI); 1824 } 1825 } 1826 1827 // If we found some, do the transformation! 1828 if (!UncondBranchPreds.empty()) { 1829 while (!UncondBranchPreds.empty()) { 1830 BasicBlock *Pred = UncondBranchPreds.pop_back_val(); 1831 DOUT << "FOLDING: " << *BB 1832 << "INTO UNCOND BRANCH PRED: " << *Pred; 1833 Instruction *UncondBranch = Pred->getTerminator(); 1834 // Clone the return and add it to the end of the predecessor. 1835 Instruction *NewRet = RI->clone(BB->getContext()); 1836 Pred->getInstList().push_back(NewRet); 1837 1838 BasicBlock::iterator BBI = RI; 1839 if (BBI != BB->begin()) { 1840 // Move region end info into the predecessor. 1841 if (DbgRegionEndInst *DREI = dyn_cast<DbgRegionEndInst>(--BBI)) 1842 DREI->moveBefore(NewRet); 1843 } 1844 1845 // If the return instruction returns a value, and if the value was a 1846 // PHI node in "BB", propagate the right value into the return. 1847 for (User::op_iterator i = NewRet->op_begin(), e = NewRet->op_end(); 1848 i != e; ++i) 1849 if (PHINode *PN = dyn_cast<PHINode>(*i)) 1850 if (PN->getParent() == BB) 1851 *i = PN->getIncomingValueForBlock(Pred); 1852 1853 // Update any PHI nodes in the returning block to realize that we no 1854 // longer branch to them. 1855 BB->removePredecessor(Pred); 1856 Pred->getInstList().erase(UncondBranch); 1857 } 1858 1859 // If we eliminated all predecessors of the block, delete the block now. 1860 if (pred_begin(BB) == pred_end(BB)) 1861 // We know there are no successors, so just nuke the block. 1862 M->getBasicBlockList().erase(BB); 1863 1864 return true; 1865 } 1866 1867 // Check out all of the conditional branches going to this return 1868 // instruction. If any of them just select between returns, change the 1869 // branch itself into a select/return pair. 1870 while (!CondBranchPreds.empty()) { 1871 BranchInst *BI = CondBranchPreds.pop_back_val(); 1872 1873 // Check to see if the non-BB successor is also a return block. 1874 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 1875 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 1876 SimplifyCondBranchToTwoReturns(BI)) 1877 return true; 1878 } 1879 } 1880 } else if (isa<UnwindInst>(BB->begin())) { 1881 // Check to see if the first instruction in this block is just an unwind. 1882 // If so, replace any invoke instructions which use this as an exception 1883 // destination with call instructions, and any unconditional branch 1884 // predecessor with an unwind. 1885 // 1886 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 1887 while (!Preds.empty()) { 1888 BasicBlock *Pred = Preds.back(); 1889 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) { 1890 if (BI->isUnconditional()) { 1891 Pred->getInstList().pop_back(); // nuke uncond branch 1892 new UnwindInst(Pred); // Use unwind. 1893 Changed = true; 1894 } 1895 } else if (InvokeInst *II = dyn_cast<InvokeInst>(Pred->getTerminator())) 1896 if (II->getUnwindDest() == BB) { 1897 // Insert a new branch instruction before the invoke, because this 1898 // is now a fall through... 1899 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 1900 Pred->getInstList().remove(II); // Take out of symbol table 1901 1902 // Insert the call now... 1903 SmallVector<Value*,8> Args(II->op_begin()+3, II->op_end()); 1904 CallInst *CI = CallInst::Create(II->getCalledValue(), 1905 Args.begin(), Args.end(), 1906 II->getName(), BI); 1907 CI->setCallingConv(II->getCallingConv()); 1908 CI->setAttributes(II->getAttributes()); 1909 // If the invoke produced a value, the Call now does instead 1910 II->replaceAllUsesWith(CI); 1911 delete II; 1912 Changed = true; 1913 } 1914 1915 Preds.pop_back(); 1916 } 1917 1918 // If this block is now dead, remove it. 1919 if (pred_begin(BB) == pred_end(BB)) { 1920 // We know there are no successors, so just nuke the block. 1921 M->getBasicBlockList().erase(BB); 1922 return true; 1923 } 1924 1925 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 1926 if (isValueEqualityComparison(SI)) { 1927 // If we only have one predecessor, and if it is a branch on this value, 1928 // see if that predecessor totally determines the outcome of this switch. 1929 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 1930 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred)) 1931 return SimplifyCFG(BB) || 1; 1932 1933 // If the block only contains the switch, see if we can fold the block 1934 // away into any preds. 1935 BasicBlock::iterator BBI = BB->begin(); 1936 // Ignore dbg intrinsics. 1937 while (isa<DbgInfoIntrinsic>(BBI)) 1938 ++BBI; 1939 if (SI == &*BBI) 1940 if (FoldValueComparisonIntoPredecessors(SI)) 1941 return SimplifyCFG(BB) || 1; 1942 } 1943 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 1944 if (BI->isUnconditional()) { 1945 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 1946 1947 BasicBlock *Succ = BI->getSuccessor(0); 1948 // Ignore dbg intrinsics. 1949 while (isa<DbgInfoIntrinsic>(BBI)) 1950 ++BBI; 1951 if (BBI->isTerminator() && // Terminator is the only non-phi instruction! 1952 Succ != BB) // Don't hurt infinite loops! 1953 if (TryToSimplifyUncondBranchFromEmptyBlock(BB, Succ)) 1954 return true; 1955 1956 } else { // Conditional branch 1957 if (isValueEqualityComparison(BI)) { 1958 // If we only have one predecessor, and if it is a branch on this value, 1959 // see if that predecessor totally determines the outcome of this 1960 // switch. 1961 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 1962 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred)) 1963 return SimplifyCFG(BB) || 1; 1964 1965 // This block must be empty, except for the setcond inst, if it exists. 1966 // Ignore dbg intrinsics. 1967 BasicBlock::iterator I = BB->begin(); 1968 // Ignore dbg intrinsics. 1969 while (isa<DbgInfoIntrinsic>(I)) 1970 ++I; 1971 if (&*I == BI) { 1972 if (FoldValueComparisonIntoPredecessors(BI)) 1973 return SimplifyCFG(BB) | true; 1974 } else if (&*I == cast<Instruction>(BI->getCondition())){ 1975 ++I; 1976 // Ignore dbg intrinsics. 1977 while (isa<DbgInfoIntrinsic>(I)) 1978 ++I; 1979 if(&*I == BI) { 1980 if (FoldValueComparisonIntoPredecessors(BI)) 1981 return SimplifyCFG(BB) | true; 1982 } 1983 } 1984 } 1985 1986 // If this is a branch on a phi node in the current block, thread control 1987 // through this block if any PHI node entries are constants. 1988 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 1989 if (PN->getParent() == BI->getParent()) 1990 if (FoldCondBranchOnPHI(BI)) 1991 return SimplifyCFG(BB) | true; 1992 1993 // If this basic block is ONLY a setcc and a branch, and if a predecessor 1994 // branches to us and one of our successors, fold the setcc into the 1995 // predecessor and use logical operations to pick the right destination. 1996 if (FoldBranchToCommonDest(BI)) 1997 return SimplifyCFG(BB) | 1; 1998 1999 2000 // Scan predecessor blocks for conditional branches. 2001 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 2002 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 2003 if (PBI != BI && PBI->isConditional()) 2004 if (SimplifyCondBranchToCondBranch(PBI, BI)) 2005 return SimplifyCFG(BB) | true; 2006 } 2007 } else if (isa<UnreachableInst>(BB->getTerminator())) { 2008 // If there are any instructions immediately before the unreachable that can 2009 // be removed, do so. 2010 Instruction *Unreachable = BB->getTerminator(); 2011 while (Unreachable != BB->begin()) { 2012 BasicBlock::iterator BBI = Unreachable; 2013 --BBI; 2014 // Do not delete instructions that can have side effects, like calls 2015 // (which may never return) and volatile loads and stores. 2016 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break; 2017 2018 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) 2019 if (SI->isVolatile()) 2020 break; 2021 2022 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) 2023 if (LI->isVolatile()) 2024 break; 2025 2026 // Delete this instruction 2027 BB->getInstList().erase(BBI); 2028 Changed = true; 2029 } 2030 2031 // If the unreachable instruction is the first in the block, take a gander 2032 // at all of the predecessors of this instruction, and simplify them. 2033 if (&BB->front() == Unreachable) { 2034 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 2035 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 2036 TerminatorInst *TI = Preds[i]->getTerminator(); 2037 2038 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 2039 if (BI->isUnconditional()) { 2040 if (BI->getSuccessor(0) == BB) { 2041 new UnreachableInst(TI); 2042 TI->eraseFromParent(); 2043 Changed = true; 2044 } 2045 } else { 2046 if (BI->getSuccessor(0) == BB) { 2047 BranchInst::Create(BI->getSuccessor(1), BI); 2048 EraseTerminatorInstAndDCECond(BI); 2049 } else if (BI->getSuccessor(1) == BB) { 2050 BranchInst::Create(BI->getSuccessor(0), BI); 2051 EraseTerminatorInstAndDCECond(BI); 2052 Changed = true; 2053 } 2054 } 2055 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 2056 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2057 if (SI->getSuccessor(i) == BB) { 2058 BB->removePredecessor(SI->getParent()); 2059 SI->removeCase(i); 2060 --i; --e; 2061 Changed = true; 2062 } 2063 // If the default value is unreachable, figure out the most popular 2064 // destination and make it the default. 2065 if (SI->getSuccessor(0) == BB) { 2066 std::map<BasicBlock*, unsigned> Popularity; 2067 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2068 Popularity[SI->getSuccessor(i)]++; 2069 2070 // Find the most popular block. 2071 unsigned MaxPop = 0; 2072 BasicBlock *MaxBlock = 0; 2073 for (std::map<BasicBlock*, unsigned>::iterator 2074 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 2075 if (I->second > MaxPop) { 2076 MaxPop = I->second; 2077 MaxBlock = I->first; 2078 } 2079 } 2080 if (MaxBlock) { 2081 // Make this the new default, allowing us to delete any explicit 2082 // edges to it. 2083 SI->setSuccessor(0, MaxBlock); 2084 Changed = true; 2085 2086 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 2087 // it. 2088 if (isa<PHINode>(MaxBlock->begin())) 2089 for (unsigned i = 0; i != MaxPop-1; ++i) 2090 MaxBlock->removePredecessor(SI->getParent()); 2091 2092 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2093 if (SI->getSuccessor(i) == MaxBlock) { 2094 SI->removeCase(i); 2095 --i; --e; 2096 } 2097 } 2098 } 2099 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2100 if (II->getUnwindDest() == BB) { 2101 // Convert the invoke to a call instruction. This would be a good 2102 // place to note that the call does not throw though. 2103 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 2104 II->removeFromParent(); // Take out of symbol table 2105 2106 // Insert the call now... 2107 SmallVector<Value*, 8> Args(II->op_begin()+3, II->op_end()); 2108 CallInst *CI = CallInst::Create(II->getCalledValue(), 2109 Args.begin(), Args.end(), 2110 II->getName(), BI); 2111 CI->setCallingConv(II->getCallingConv()); 2112 CI->setAttributes(II->getAttributes()); 2113 // If the invoke produced a value, the Call does now instead. 2114 II->replaceAllUsesWith(CI); 2115 delete II; 2116 Changed = true; 2117 } 2118 } 2119 } 2120 2121 // If this block is now dead, remove it. 2122 if (pred_begin(BB) == pred_end(BB)) { 2123 // We know there are no successors, so just nuke the block. 2124 M->getBasicBlockList().erase(BB); 2125 return true; 2126 } 2127 } 2128 } 2129 2130 // Merge basic blocks into their predecessor if there is only one distinct 2131 // pred, and if there is only one distinct successor of the predecessor, and 2132 // if there are no PHI nodes. 2133 // 2134 if (MergeBlockIntoPredecessor(BB)) 2135 return true; 2136 2137 // Otherwise, if this block only has a single predecessor, and if that block 2138 // is a conditional branch, see if we can hoist any code from this block up 2139 // into our predecessor. 2140 pred_iterator PI(pred_begin(BB)), PE(pred_end(BB)); 2141 BasicBlock *OnlyPred = *PI++; 2142 for (; PI != PE; ++PI) // Search all predecessors, see if they are all same 2143 if (*PI != OnlyPred) { 2144 OnlyPred = 0; // There are multiple different predecessors... 2145 break; 2146 } 2147 2148 if (OnlyPred) 2149 if (BranchInst *BI = dyn_cast<BranchInst>(OnlyPred->getTerminator())) 2150 if (BI->isConditional()) { 2151 // Get the other block. 2152 BasicBlock *OtherBB = BI->getSuccessor(BI->getSuccessor(0) == BB); 2153 PI = pred_begin(OtherBB); 2154 ++PI; 2155 2156 if (PI == pred_end(OtherBB)) { 2157 // We have a conditional branch to two blocks that are only reachable 2158 // from the condbr. We know that the condbr dominates the two blocks, 2159 // so see if there is any identical code in the "then" and "else" 2160 // blocks. If so, we can hoist it up to the branching block. 2161 Changed |= HoistThenElseCodeToIf(BI); 2162 } else { 2163 BasicBlock* OnlySucc = NULL; 2164 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); 2165 SI != SE; ++SI) { 2166 if (!OnlySucc) 2167 OnlySucc = *SI; 2168 else if (*SI != OnlySucc) { 2169 OnlySucc = 0; // There are multiple distinct successors! 2170 break; 2171 } 2172 } 2173 2174 if (OnlySucc == OtherBB) { 2175 // If BB's only successor is the other successor of the predecessor, 2176 // i.e. a triangle, see if we can hoist any code from this block up 2177 // to the "if" block. 2178 Changed |= SpeculativelyExecuteBB(BI, BB); 2179 } 2180 } 2181 } 2182 2183 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 2184 if (BranchInst *BI = dyn_cast<BranchInst>((*PI)->getTerminator())) 2185 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2186 if (BI->isConditional() && isa<Instruction>(BI->getCondition())) { 2187 Instruction *Cond = cast<Instruction>(BI->getCondition()); 2188 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2189 // 'setne's and'ed together, collect them. 2190 Value *CompVal = 0; 2191 std::vector<ConstantInt*> Values; 2192 bool TrueWhenEqual = GatherValueComparisons(Cond, CompVal, Values); 2193 if (CompVal && CompVal->getType()->isInteger()) { 2194 // There might be duplicate constants in the list, which the switch 2195 // instruction can't handle, remove them now. 2196 std::sort(Values.begin(), Values.end(), ConstantIntOrdering()); 2197 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2198 2199 // Figure out which block is which destination. 2200 BasicBlock *DefaultBB = BI->getSuccessor(1); 2201 BasicBlock *EdgeBB = BI->getSuccessor(0); 2202 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2203 2204 // Create the new switch instruction now. 2205 SwitchInst *New = SwitchInst::Create(CompVal, DefaultBB, 2206 Values.size(), BI); 2207 2208 // Add all of the 'cases' to the switch instruction. 2209 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2210 New->addCase(Values[i], EdgeBB); 2211 2212 // We added edges from PI to the EdgeBB. As such, if there were any 2213 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2214 // the number of edges added. 2215 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2216 isa<PHINode>(BBI); ++BBI) { 2217 PHINode *PN = cast<PHINode>(BBI); 2218 Value *InVal = PN->getIncomingValueForBlock(*PI); 2219 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2220 PN->addIncoming(InVal, *PI); 2221 } 2222 2223 // Erase the old branch instruction. 2224 EraseTerminatorInstAndDCECond(BI); 2225 return true; 2226 } 2227 } 2228 2229 return Changed; 2230 } 2231