1 //===-- Local.cpp - Functions to perform local transformations ------------===// 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 family of functions perform various local transformations to the 11 // program. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Transforms/Utils/Local.h" 16 #include "llvm/ADT/DenseMap.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SmallPtrSet.h" 19 #include "llvm/ADT/Statistic.h" 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/Analysis/MemoryBuiltins.h" 22 #include "llvm/Analysis/ValueTracking.h" 23 #include "llvm/IR/CFG.h" 24 #include "llvm/IR/Constants.h" 25 #include "llvm/IR/DIBuilder.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/DebugInfo.h" 28 #include "llvm/IR/DerivedTypes.h" 29 #include "llvm/IR/Dominators.h" 30 #include "llvm/IR/GetElementPtrTypeIterator.h" 31 #include "llvm/IR/GlobalAlias.h" 32 #include "llvm/IR/GlobalVariable.h" 33 #include "llvm/IR/IRBuilder.h" 34 #include "llvm/IR/Instructions.h" 35 #include "llvm/IR/IntrinsicInst.h" 36 #include "llvm/IR/Intrinsics.h" 37 #include "llvm/IR/MDBuilder.h" 38 #include "llvm/IR/Metadata.h" 39 #include "llvm/IR/Operator.h" 40 #include "llvm/IR/ValueHandle.h" 41 #include "llvm/Support/Debug.h" 42 #include "llvm/Support/MathExtras.h" 43 #include "llvm/Support/raw_ostream.h" 44 using namespace llvm; 45 46 #define DEBUG_TYPE "local" 47 48 STATISTIC(NumRemoved, "Number of unreachable basic blocks removed"); 49 50 //===----------------------------------------------------------------------===// 51 // Local constant propagation. 52 // 53 54 /// ConstantFoldTerminator - If a terminator instruction is predicated on a 55 /// constant value, convert it into an unconditional branch to the constant 56 /// destination. This is a nontrivial operation because the successors of this 57 /// basic block must have their PHI nodes updated. 58 /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch 59 /// conditions and indirectbr addresses this might make dead if 60 /// DeleteDeadConditions is true. 61 bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions, 62 const TargetLibraryInfo *TLI) { 63 TerminatorInst *T = BB->getTerminator(); 64 IRBuilder<> Builder(T); 65 66 // Branch - See if we are conditional jumping on constant 67 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 68 if (BI->isUnconditional()) return false; // Can't optimize uncond branch 69 BasicBlock *Dest1 = BI->getSuccessor(0); 70 BasicBlock *Dest2 = BI->getSuccessor(1); 71 72 if (ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition())) { 73 // Are we branching on constant? 74 // YES. Change to unconditional branch... 75 BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2; 76 BasicBlock *OldDest = Cond->getZExtValue() ? Dest2 : Dest1; 77 78 //cerr << "Function: " << T->getParent()->getParent() 79 // << "\nRemoving branch from " << T->getParent() 80 // << "\n\nTo: " << OldDest << endl; 81 82 // Let the basic block know that we are letting go of it. Based on this, 83 // it will adjust it's PHI nodes. 84 OldDest->removePredecessor(BB); 85 86 // Replace the conditional branch with an unconditional one. 87 Builder.CreateBr(Destination); 88 BI->eraseFromParent(); 89 return true; 90 } 91 92 if (Dest2 == Dest1) { // Conditional branch to same location? 93 // This branch matches something like this: 94 // br bool %cond, label %Dest, label %Dest 95 // and changes it into: br label %Dest 96 97 // Let the basic block know that we are letting go of one copy of it. 98 assert(BI->getParent() && "Terminator not inserted in block!"); 99 Dest1->removePredecessor(BI->getParent()); 100 101 // Replace the conditional branch with an unconditional one. 102 Builder.CreateBr(Dest1); 103 Value *Cond = BI->getCondition(); 104 BI->eraseFromParent(); 105 if (DeleteDeadConditions) 106 RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI); 107 return true; 108 } 109 return false; 110 } 111 112 if (SwitchInst *SI = dyn_cast<SwitchInst>(T)) { 113 // If we are switching on a constant, we can convert the switch into a 114 // single branch instruction! 115 ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition()); 116 BasicBlock *TheOnlyDest = SI->getDefaultDest(); 117 BasicBlock *DefaultDest = TheOnlyDest; 118 119 // Figure out which case it goes to. 120 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 121 i != e; ++i) { 122 // Found case matching a constant operand? 123 if (i.getCaseValue() == CI) { 124 TheOnlyDest = i.getCaseSuccessor(); 125 break; 126 } 127 128 // Check to see if this branch is going to the same place as the default 129 // dest. If so, eliminate it as an explicit compare. 130 if (i.getCaseSuccessor() == DefaultDest) { 131 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); 132 unsigned NCases = SI->getNumCases(); 133 // Fold the case metadata into the default if there will be any branches 134 // left, unless the metadata doesn't match the switch. 135 if (NCases > 1 && MD && MD->getNumOperands() == 2 + NCases) { 136 // Collect branch weights into a vector. 137 SmallVector<uint32_t, 8> Weights; 138 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; 139 ++MD_i) { 140 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i)); 141 assert(CI); 142 Weights.push_back(CI->getValue().getZExtValue()); 143 } 144 // Merge weight of this case to the default weight. 145 unsigned idx = i.getCaseIndex(); 146 Weights[0] += Weights[idx+1]; 147 // Remove weight for this case. 148 std::swap(Weights[idx+1], Weights.back()); 149 Weights.pop_back(); 150 SI->setMetadata(LLVMContext::MD_prof, 151 MDBuilder(BB->getContext()). 152 createBranchWeights(Weights)); 153 } 154 // Remove this entry. 155 DefaultDest->removePredecessor(SI->getParent()); 156 SI->removeCase(i); 157 --i; --e; 158 continue; 159 } 160 161 // Otherwise, check to see if the switch only branches to one destination. 162 // We do this by reseting "TheOnlyDest" to null when we find two non-equal 163 // destinations. 164 if (i.getCaseSuccessor() != TheOnlyDest) TheOnlyDest = nullptr; 165 } 166 167 if (CI && !TheOnlyDest) { 168 // Branching on a constant, but not any of the cases, go to the default 169 // successor. 170 TheOnlyDest = SI->getDefaultDest(); 171 } 172 173 // If we found a single destination that we can fold the switch into, do so 174 // now. 175 if (TheOnlyDest) { 176 // Insert the new branch. 177 Builder.CreateBr(TheOnlyDest); 178 BasicBlock *BB = SI->getParent(); 179 180 // Remove entries from PHI nodes which we no longer branch to... 181 for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) { 182 // Found case matching a constant operand? 183 BasicBlock *Succ = SI->getSuccessor(i); 184 if (Succ == TheOnlyDest) 185 TheOnlyDest = nullptr; // Don't modify the first branch to TheOnlyDest 186 else 187 Succ->removePredecessor(BB); 188 } 189 190 // Delete the old switch. 191 Value *Cond = SI->getCondition(); 192 SI->eraseFromParent(); 193 if (DeleteDeadConditions) 194 RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI); 195 return true; 196 } 197 198 if (SI->getNumCases() == 1) { 199 // Otherwise, we can fold this switch into a conditional branch 200 // instruction if it has only one non-default destination. 201 SwitchInst::CaseIt FirstCase = SI->case_begin(); 202 Value *Cond = Builder.CreateICmpEQ(SI->getCondition(), 203 FirstCase.getCaseValue(), "cond"); 204 205 // Insert the new branch. 206 BranchInst *NewBr = Builder.CreateCondBr(Cond, 207 FirstCase.getCaseSuccessor(), 208 SI->getDefaultDest()); 209 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof); 210 if (MD && MD->getNumOperands() == 3) { 211 ConstantInt *SICase = dyn_cast<ConstantInt>(MD->getOperand(2)); 212 ConstantInt *SIDef = dyn_cast<ConstantInt>(MD->getOperand(1)); 213 assert(SICase && SIDef); 214 // The TrueWeight should be the weight for the single case of SI. 215 NewBr->setMetadata(LLVMContext::MD_prof, 216 MDBuilder(BB->getContext()). 217 createBranchWeights(SICase->getValue().getZExtValue(), 218 SIDef->getValue().getZExtValue())); 219 } 220 221 // Delete the old switch. 222 SI->eraseFromParent(); 223 return true; 224 } 225 return false; 226 } 227 228 if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(T)) { 229 // indirectbr blockaddress(@F, @BB) -> br label @BB 230 if (BlockAddress *BA = 231 dyn_cast<BlockAddress>(IBI->getAddress()->stripPointerCasts())) { 232 BasicBlock *TheOnlyDest = BA->getBasicBlock(); 233 // Insert the new branch. 234 Builder.CreateBr(TheOnlyDest); 235 236 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 237 if (IBI->getDestination(i) == TheOnlyDest) 238 TheOnlyDest = nullptr; 239 else 240 IBI->getDestination(i)->removePredecessor(IBI->getParent()); 241 } 242 Value *Address = IBI->getAddress(); 243 IBI->eraseFromParent(); 244 if (DeleteDeadConditions) 245 RecursivelyDeleteTriviallyDeadInstructions(Address, TLI); 246 247 // If we didn't find our destination in the IBI successor list, then we 248 // have undefined behavior. Replace the unconditional branch with an 249 // 'unreachable' instruction. 250 if (TheOnlyDest) { 251 BB->getTerminator()->eraseFromParent(); 252 new UnreachableInst(BB->getContext(), BB); 253 } 254 255 return true; 256 } 257 } 258 259 return false; 260 } 261 262 263 //===----------------------------------------------------------------------===// 264 // Local dead code elimination. 265 // 266 267 /// isInstructionTriviallyDead - Return true if the result produced by the 268 /// instruction is not used, and the instruction has no side effects. 269 /// 270 bool llvm::isInstructionTriviallyDead(Instruction *I, 271 const TargetLibraryInfo *TLI) { 272 if (!I->use_empty() || isa<TerminatorInst>(I)) return false; 273 274 // We don't want the landingpad instruction removed by anything this general. 275 if (isa<LandingPadInst>(I)) 276 return false; 277 278 // We don't want debug info removed by anything this general, unless 279 // debug info is empty. 280 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(I)) { 281 if (DDI->getAddress()) 282 return false; 283 return true; 284 } 285 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(I)) { 286 if (DVI->getValue()) 287 return false; 288 return true; 289 } 290 291 if (!I->mayHaveSideEffects()) return true; 292 293 // Special case intrinsics that "may have side effects" but can be deleted 294 // when dead. 295 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 296 // Safe to delete llvm.stacksave if dead. 297 if (II->getIntrinsicID() == Intrinsic::stacksave) 298 return true; 299 300 // Lifetime intrinsics are dead when their right-hand is undef. 301 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 302 II->getIntrinsicID() == Intrinsic::lifetime_end) 303 return isa<UndefValue>(II->getArgOperand(1)); 304 } 305 306 if (isAllocLikeFn(I, TLI)) return true; 307 308 if (CallInst *CI = isFreeCall(I, TLI)) 309 if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0))) 310 return C->isNullValue() || isa<UndefValue>(C); 311 312 return false; 313 } 314 315 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a 316 /// trivially dead instruction, delete it. If that makes any of its operands 317 /// trivially dead, delete them too, recursively. Return true if any 318 /// instructions were deleted. 319 bool 320 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V, 321 const TargetLibraryInfo *TLI) { 322 Instruction *I = dyn_cast<Instruction>(V); 323 if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI)) 324 return false; 325 326 SmallVector<Instruction*, 16> DeadInsts; 327 DeadInsts.push_back(I); 328 329 do { 330 I = DeadInsts.pop_back_val(); 331 332 // Null out all of the instruction's operands to see if any operand becomes 333 // dead as we go. 334 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 335 Value *OpV = I->getOperand(i); 336 I->setOperand(i, nullptr); 337 338 if (!OpV->use_empty()) continue; 339 340 // If the operand is an instruction that became dead as we nulled out the 341 // operand, and if it is 'trivially' dead, delete it in a future loop 342 // iteration. 343 if (Instruction *OpI = dyn_cast<Instruction>(OpV)) 344 if (isInstructionTriviallyDead(OpI, TLI)) 345 DeadInsts.push_back(OpI); 346 } 347 348 I->eraseFromParent(); 349 } while (!DeadInsts.empty()); 350 351 return true; 352 } 353 354 /// areAllUsesEqual - Check whether the uses of a value are all the same. 355 /// This is similar to Instruction::hasOneUse() except this will also return 356 /// true when there are no uses or multiple uses that all refer to the same 357 /// value. 358 static bool areAllUsesEqual(Instruction *I) { 359 Value::user_iterator UI = I->user_begin(); 360 Value::user_iterator UE = I->user_end(); 361 if (UI == UE) 362 return true; 363 364 User *TheUse = *UI; 365 for (++UI; UI != UE; ++UI) { 366 if (*UI != TheUse) 367 return false; 368 } 369 return true; 370 } 371 372 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively 373 /// dead PHI node, due to being a def-use chain of single-use nodes that 374 /// either forms a cycle or is terminated by a trivially dead instruction, 375 /// delete it. If that makes any of its operands trivially dead, delete them 376 /// too, recursively. Return true if a change was made. 377 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN, 378 const TargetLibraryInfo *TLI) { 379 SmallPtrSet<Instruction*, 4> Visited; 380 for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects(); 381 I = cast<Instruction>(*I->user_begin())) { 382 if (I->use_empty()) 383 return RecursivelyDeleteTriviallyDeadInstructions(I, TLI); 384 385 // If we find an instruction more than once, we're on a cycle that 386 // won't prove fruitful. 387 if (!Visited.insert(I)) { 388 // Break the cycle and delete the instruction and its operands. 389 I->replaceAllUsesWith(UndefValue::get(I->getType())); 390 (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI); 391 return true; 392 } 393 } 394 return false; 395 } 396 397 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to 398 /// simplify any instructions in it and recursively delete dead instructions. 399 /// 400 /// This returns true if it changed the code, note that it can delete 401 /// instructions in other blocks as well in this block. 402 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB, const DataLayout *TD, 403 const TargetLibraryInfo *TLI) { 404 bool MadeChange = false; 405 406 #ifndef NDEBUG 407 // In debug builds, ensure that the terminator of the block is never replaced 408 // or deleted by these simplifications. The idea of simplification is that it 409 // cannot introduce new instructions, and there is no way to replace the 410 // terminator of a block without introducing a new instruction. 411 AssertingVH<Instruction> TerminatorVH(--BB->end()); 412 #endif 413 414 for (BasicBlock::iterator BI = BB->begin(), E = --BB->end(); BI != E; ) { 415 assert(!BI->isTerminator()); 416 Instruction *Inst = BI++; 417 418 WeakVH BIHandle(BI); 419 if (recursivelySimplifyInstruction(Inst, TD, TLI)) { 420 MadeChange = true; 421 if (BIHandle != BI) 422 BI = BB->begin(); 423 continue; 424 } 425 426 MadeChange |= RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI); 427 if (BIHandle != BI) 428 BI = BB->begin(); 429 } 430 return MadeChange; 431 } 432 433 //===----------------------------------------------------------------------===// 434 // Control Flow Graph Restructuring. 435 // 436 437 438 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this 439 /// method is called when we're about to delete Pred as a predecessor of BB. If 440 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred. 441 /// 442 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI 443 /// nodes that collapse into identity values. For example, if we have: 444 /// x = phi(1, 0, 0, 0) 445 /// y = and x, z 446 /// 447 /// .. and delete the predecessor corresponding to the '1', this will attempt to 448 /// recursively fold the and to 0. 449 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred, 450 DataLayout *TD) { 451 // This only adjusts blocks with PHI nodes. 452 if (!isa<PHINode>(BB->begin())) 453 return; 454 455 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify 456 // them down. This will leave us with single entry phi nodes and other phis 457 // that can be removed. 458 BB->removePredecessor(Pred, true); 459 460 WeakVH PhiIt = &BB->front(); 461 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) { 462 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt)); 463 Value *OldPhiIt = PhiIt; 464 465 if (!recursivelySimplifyInstruction(PN, TD)) 466 continue; 467 468 // If recursive simplification ended up deleting the next PHI node we would 469 // iterate to, then our iterator is invalid, restart scanning from the top 470 // of the block. 471 if (PhiIt != OldPhiIt) PhiIt = &BB->front(); 472 } 473 } 474 475 476 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its 477 /// predecessor is known to have one successor (DestBB!). Eliminate the edge 478 /// between them, moving the instructions in the predecessor into DestBB and 479 /// deleting the predecessor block. 480 /// 481 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, Pass *P) { 482 // If BB has single-entry PHI nodes, fold them. 483 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 484 Value *NewVal = PN->getIncomingValue(0); 485 // Replace self referencing PHI with undef, it must be dead. 486 if (NewVal == PN) NewVal = UndefValue::get(PN->getType()); 487 PN->replaceAllUsesWith(NewVal); 488 PN->eraseFromParent(); 489 } 490 491 BasicBlock *PredBB = DestBB->getSinglePredecessor(); 492 assert(PredBB && "Block doesn't have a single predecessor!"); 493 494 // Zap anything that took the address of DestBB. Not doing this will give the 495 // address an invalid value. 496 if (DestBB->hasAddressTaken()) { 497 BlockAddress *BA = BlockAddress::get(DestBB); 498 Constant *Replacement = 499 ConstantInt::get(llvm::Type::getInt32Ty(BA->getContext()), 1); 500 BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement, 501 BA->getType())); 502 BA->destroyConstant(); 503 } 504 505 // Anything that branched to PredBB now branches to DestBB. 506 PredBB->replaceAllUsesWith(DestBB); 507 508 // Splice all the instructions from PredBB to DestBB. 509 PredBB->getTerminator()->eraseFromParent(); 510 DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList()); 511 512 if (P) { 513 if (DominatorTreeWrapperPass *DTWP = 514 P->getAnalysisIfAvailable<DominatorTreeWrapperPass>()) { 515 DominatorTree &DT = DTWP->getDomTree(); 516 BasicBlock *PredBBIDom = DT.getNode(PredBB)->getIDom()->getBlock(); 517 DT.changeImmediateDominator(DestBB, PredBBIDom); 518 DT.eraseNode(PredBB); 519 } 520 } 521 // Nuke BB. 522 PredBB->eraseFromParent(); 523 } 524 525 /// CanMergeValues - Return true if we can choose one of these values to use 526 /// in place of the other. Note that we will always choose the non-undef 527 /// value to keep. 528 static bool CanMergeValues(Value *First, Value *Second) { 529 return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second); 530 } 531 532 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an 533 /// almost-empty BB ending in an unconditional branch to Succ, into Succ. 534 /// 535 /// Assumption: Succ is the single successor for BB. 536 /// 537 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { 538 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!"); 539 540 DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into " 541 << Succ->getName() << "\n"); 542 // Shortcut, if there is only a single predecessor it must be BB and merging 543 // is always safe 544 if (Succ->getSinglePredecessor()) return true; 545 546 // Make a list of the predecessors of BB 547 SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB)); 548 549 // Look at all the phi nodes in Succ, to see if they present a conflict when 550 // merging these blocks 551 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 552 PHINode *PN = cast<PHINode>(I); 553 554 // If the incoming value from BB is again a PHINode in 555 // BB which has the same incoming value for *PI as PN does, we can 556 // merge the phi nodes and then the blocks can still be merged 557 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB)); 558 if (BBPN && BBPN->getParent() == BB) { 559 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) { 560 BasicBlock *IBB = PN->getIncomingBlock(PI); 561 if (BBPreds.count(IBB) && 562 !CanMergeValues(BBPN->getIncomingValueForBlock(IBB), 563 PN->getIncomingValue(PI))) { 564 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in " 565 << Succ->getName() << " is conflicting with " 566 << BBPN->getName() << " with regard to common predecessor " 567 << IBB->getName() << "\n"); 568 return false; 569 } 570 } 571 } else { 572 Value* Val = PN->getIncomingValueForBlock(BB); 573 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) { 574 // See if the incoming value for the common predecessor is equal to the 575 // one for BB, in which case this phi node will not prevent the merging 576 // of the block. 577 BasicBlock *IBB = PN->getIncomingBlock(PI); 578 if (BBPreds.count(IBB) && 579 !CanMergeValues(Val, PN->getIncomingValue(PI))) { 580 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in " 581 << Succ->getName() << " is conflicting with regard to common " 582 << "predecessor " << IBB->getName() << "\n"); 583 return false; 584 } 585 } 586 } 587 } 588 589 return true; 590 } 591 592 typedef SmallVector<BasicBlock *, 16> PredBlockVector; 593 typedef DenseMap<BasicBlock *, Value *> IncomingValueMap; 594 595 /// \brief Determines the value to use as the phi node input for a block. 596 /// 597 /// Select between \p OldVal any value that we know flows from \p BB 598 /// to a particular phi on the basis of which one (if either) is not 599 /// undef. Update IncomingValues based on the selected value. 600 /// 601 /// \param OldVal The value we are considering selecting. 602 /// \param BB The block that the value flows in from. 603 /// \param IncomingValues A map from block-to-value for other phi inputs 604 /// that we have examined. 605 /// 606 /// \returns the selected value. 607 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB, 608 IncomingValueMap &IncomingValues) { 609 if (!isa<UndefValue>(OldVal)) { 610 assert((!IncomingValues.count(BB) || 611 IncomingValues.find(BB)->second == OldVal) && 612 "Expected OldVal to match incoming value from BB!"); 613 614 IncomingValues.insert(std::make_pair(BB, OldVal)); 615 return OldVal; 616 } 617 618 IncomingValueMap::const_iterator It = IncomingValues.find(BB); 619 if (It != IncomingValues.end()) return It->second; 620 621 return OldVal; 622 } 623 624 /// \brief Create a map from block to value for the operands of a 625 /// given phi. 626 /// 627 /// Create a map from block to value for each non-undef value flowing 628 /// into \p PN. 629 /// 630 /// \param PN The phi we are collecting the map for. 631 /// \param IncomingValues [out] The map from block to value for this phi. 632 static void gatherIncomingValuesToPhi(PHINode *PN, 633 IncomingValueMap &IncomingValues) { 634 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 635 BasicBlock *BB = PN->getIncomingBlock(i); 636 Value *V = PN->getIncomingValue(i); 637 638 if (!isa<UndefValue>(V)) 639 IncomingValues.insert(std::make_pair(BB, V)); 640 } 641 } 642 643 /// \brief Replace the incoming undef values to a phi with the values 644 /// from a block-to-value map. 645 /// 646 /// \param PN The phi we are replacing the undefs in. 647 /// \param IncomingValues A map from block to value. 648 static void replaceUndefValuesInPhi(PHINode *PN, 649 const IncomingValueMap &IncomingValues) { 650 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 651 Value *V = PN->getIncomingValue(i); 652 653 if (!isa<UndefValue>(V)) continue; 654 655 BasicBlock *BB = PN->getIncomingBlock(i); 656 IncomingValueMap::const_iterator It = IncomingValues.find(BB); 657 if (It == IncomingValues.end()) continue; 658 659 PN->setIncomingValue(i, It->second); 660 } 661 } 662 663 /// \brief Replace a value flowing from a block to a phi with 664 /// potentially multiple instances of that value flowing from the 665 /// block's predecessors to the phi. 666 /// 667 /// \param BB The block with the value flowing into the phi. 668 /// \param BBPreds The predecessors of BB. 669 /// \param PN The phi that we are updating. 670 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB, 671 const PredBlockVector &BBPreds, 672 PHINode *PN) { 673 Value *OldVal = PN->removeIncomingValue(BB, false); 674 assert(OldVal && "No entry in PHI for Pred BB!"); 675 676 IncomingValueMap IncomingValues; 677 678 // We are merging two blocks - BB, and the block containing PN - and 679 // as a result we need to redirect edges from the predecessors of BB 680 // to go to the block containing PN, and update PN 681 // accordingly. Since we allow merging blocks in the case where the 682 // predecessor and successor blocks both share some predecessors, 683 // and where some of those common predecessors might have undef 684 // values flowing into PN, we want to rewrite those values to be 685 // consistent with the non-undef values. 686 687 gatherIncomingValuesToPhi(PN, IncomingValues); 688 689 // If this incoming value is one of the PHI nodes in BB, the new entries 690 // in the PHI node are the entries from the old PHI. 691 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) { 692 PHINode *OldValPN = cast<PHINode>(OldVal); 693 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) { 694 // Note that, since we are merging phi nodes and BB and Succ might 695 // have common predecessors, we could end up with a phi node with 696 // identical incoming branches. This will be cleaned up later (and 697 // will trigger asserts if we try to clean it up now, without also 698 // simplifying the corresponding conditional branch). 699 BasicBlock *PredBB = OldValPN->getIncomingBlock(i); 700 Value *PredVal = OldValPN->getIncomingValue(i); 701 Value *Selected = selectIncomingValueForBlock(PredVal, PredBB, 702 IncomingValues); 703 704 // And add a new incoming value for this predecessor for the 705 // newly retargeted branch. 706 PN->addIncoming(Selected, PredBB); 707 } 708 } else { 709 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) { 710 // Update existing incoming values in PN for this 711 // predecessor of BB. 712 BasicBlock *PredBB = BBPreds[i]; 713 Value *Selected = selectIncomingValueForBlock(OldVal, PredBB, 714 IncomingValues); 715 716 // And add a new incoming value for this predecessor for the 717 // newly retargeted branch. 718 PN->addIncoming(Selected, PredBB); 719 } 720 } 721 722 replaceUndefValuesInPhi(PN, IncomingValues); 723 } 724 725 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an 726 /// unconditional branch, and contains no instructions other than PHI nodes, 727 /// potential side-effect free intrinsics and the branch. If possible, 728 /// eliminate BB by rewriting all the predecessors to branch to the successor 729 /// block and return true. If we can't transform, return false. 730 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB) { 731 assert(BB != &BB->getParent()->getEntryBlock() && 732 "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!"); 733 734 // We can't eliminate infinite loops. 735 BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0); 736 if (BB == Succ) return false; 737 738 // Check to see if merging these blocks would cause conflicts for any of the 739 // phi nodes in BB or Succ. If not, we can safely merge. 740 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false; 741 742 // Check for cases where Succ has multiple predecessors and a PHI node in BB 743 // has uses which will not disappear when the PHI nodes are merged. It is 744 // possible to handle such cases, but difficult: it requires checking whether 745 // BB dominates Succ, which is non-trivial to calculate in the case where 746 // Succ has multiple predecessors. Also, it requires checking whether 747 // constructing the necessary self-referential PHI node doesn't introduce any 748 // conflicts; this isn't too difficult, but the previous code for doing this 749 // was incorrect. 750 // 751 // Note that if this check finds a live use, BB dominates Succ, so BB is 752 // something like a loop pre-header (or rarely, a part of an irreducible CFG); 753 // folding the branch isn't profitable in that case anyway. 754 if (!Succ->getSinglePredecessor()) { 755 BasicBlock::iterator BBI = BB->begin(); 756 while (isa<PHINode>(*BBI)) { 757 for (Use &U : BBI->uses()) { 758 if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) { 759 if (PN->getIncomingBlock(U) != BB) 760 return false; 761 } else { 762 return false; 763 } 764 } 765 ++BBI; 766 } 767 } 768 769 DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB); 770 771 if (isa<PHINode>(Succ->begin())) { 772 // If there is more than one pred of succ, and there are PHI nodes in 773 // the successor, then we need to add incoming edges for the PHI nodes 774 // 775 const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB)); 776 777 // Loop over all of the PHI nodes in the successor of BB. 778 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 779 PHINode *PN = cast<PHINode>(I); 780 781 redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN); 782 } 783 } 784 785 if (Succ->getSinglePredecessor()) { 786 // BB is the only predecessor of Succ, so Succ will end up with exactly 787 // the same predecessors BB had. 788 789 // Copy over any phi, debug or lifetime instruction. 790 BB->getTerminator()->eraseFromParent(); 791 Succ->getInstList().splice(Succ->getFirstNonPHI(), BB->getInstList()); 792 } else { 793 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) { 794 // We explicitly check for such uses in CanPropagatePredecessorsForPHIs. 795 assert(PN->use_empty() && "There shouldn't be any uses here!"); 796 PN->eraseFromParent(); 797 } 798 } 799 800 // Everything that jumped to BB now goes to Succ. 801 BB->replaceAllUsesWith(Succ); 802 if (!Succ->hasName()) Succ->takeName(BB); 803 BB->eraseFromParent(); // Delete the old basic block. 804 return true; 805 } 806 807 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI 808 /// nodes in this block. This doesn't try to be clever about PHI nodes 809 /// which differ only in the order of the incoming values, but instcombine 810 /// orders them so it usually won't matter. 811 /// 812 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) { 813 bool Changed = false; 814 815 // This implementation doesn't currently consider undef operands 816 // specially. Theoretically, two phis which are identical except for 817 // one having an undef where the other doesn't could be collapsed. 818 819 // Map from PHI hash values to PHI nodes. If multiple PHIs have 820 // the same hash value, the element is the first PHI in the 821 // linked list in CollisionMap. 822 DenseMap<uintptr_t, PHINode *> HashMap; 823 824 // Maintain linked lists of PHI nodes with common hash values. 825 DenseMap<PHINode *, PHINode *> CollisionMap; 826 827 // Examine each PHI. 828 for (BasicBlock::iterator I = BB->begin(); 829 PHINode *PN = dyn_cast<PHINode>(I++); ) { 830 // Compute a hash value on the operands. Instcombine will likely have sorted 831 // them, which helps expose duplicates, but we have to check all the 832 // operands to be safe in case instcombine hasn't run. 833 uintptr_t Hash = 0; 834 // This hash algorithm is quite weak as hash functions go, but it seems 835 // to do a good enough job for this particular purpose, and is very quick. 836 for (User::op_iterator I = PN->op_begin(), E = PN->op_end(); I != E; ++I) { 837 Hash ^= reinterpret_cast<uintptr_t>(static_cast<Value *>(*I)); 838 Hash = (Hash << 7) | (Hash >> (sizeof(uintptr_t) * CHAR_BIT - 7)); 839 } 840 for (PHINode::block_iterator I = PN->block_begin(), E = PN->block_end(); 841 I != E; ++I) { 842 Hash ^= reinterpret_cast<uintptr_t>(static_cast<BasicBlock *>(*I)); 843 Hash = (Hash << 7) | (Hash >> (sizeof(uintptr_t) * CHAR_BIT - 7)); 844 } 845 // Avoid colliding with the DenseMap sentinels ~0 and ~0-1. 846 Hash >>= 1; 847 // If we've never seen this hash value before, it's a unique PHI. 848 std::pair<DenseMap<uintptr_t, PHINode *>::iterator, bool> Pair = 849 HashMap.insert(std::make_pair(Hash, PN)); 850 if (Pair.second) continue; 851 // Otherwise it's either a duplicate or a hash collision. 852 for (PHINode *OtherPN = Pair.first->second; ; ) { 853 if (OtherPN->isIdenticalTo(PN)) { 854 // A duplicate. Replace this PHI with its duplicate. 855 PN->replaceAllUsesWith(OtherPN); 856 PN->eraseFromParent(); 857 Changed = true; 858 break; 859 } 860 // A non-duplicate hash collision. 861 DenseMap<PHINode *, PHINode *>::iterator I = CollisionMap.find(OtherPN); 862 if (I == CollisionMap.end()) { 863 // Set this PHI to be the head of the linked list of colliding PHIs. 864 PHINode *Old = Pair.first->second; 865 Pair.first->second = PN; 866 CollisionMap[PN] = Old; 867 break; 868 } 869 // Proceed to the next PHI in the list. 870 OtherPN = I->second; 871 } 872 } 873 874 return Changed; 875 } 876 877 /// enforceKnownAlignment - If the specified pointer points to an object that 878 /// we control, modify the object's alignment to PrefAlign. This isn't 879 /// often possible though. If alignment is important, a more reliable approach 880 /// is to simply align all global variables and allocation instructions to 881 /// their preferred alignment from the beginning. 882 /// 883 static unsigned enforceKnownAlignment(Value *V, unsigned Align, 884 unsigned PrefAlign, const DataLayout *TD) { 885 V = V->stripPointerCasts(); 886 887 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 888 // If the preferred alignment is greater than the natural stack alignment 889 // then don't round up. This avoids dynamic stack realignment. 890 if (TD && TD->exceedsNaturalStackAlignment(PrefAlign)) 891 return Align; 892 // If there is a requested alignment and if this is an alloca, round up. 893 if (AI->getAlignment() >= PrefAlign) 894 return AI->getAlignment(); 895 AI->setAlignment(PrefAlign); 896 return PrefAlign; 897 } 898 899 if (auto *GO = dyn_cast<GlobalObject>(V)) { 900 // If there is a large requested alignment and we can, bump up the alignment 901 // of the global. 902 if (GO->isDeclaration()) 903 return Align; 904 // If the memory we set aside for the global may not be the memory used by 905 // the final program then it is impossible for us to reliably enforce the 906 // preferred alignment. 907 if (GO->isWeakForLinker()) 908 return Align; 909 910 if (GO->getAlignment() >= PrefAlign) 911 return GO->getAlignment(); 912 // We can only increase the alignment of the global if it has no alignment 913 // specified or if it is not assigned a section. If it is assigned a 914 // section, the global could be densely packed with other objects in the 915 // section, increasing the alignment could cause padding issues. 916 if (!GO->hasSection() || GO->getAlignment() == 0) 917 GO->setAlignment(PrefAlign); 918 return GO->getAlignment(); 919 } 920 921 return Align; 922 } 923 924 /// getOrEnforceKnownAlignment - If the specified pointer has an alignment that 925 /// we can determine, return it, otherwise return 0. If PrefAlign is specified, 926 /// and it is more than the alignment of the ultimate object, see if we can 927 /// increase the alignment of the ultimate object, making this check succeed. 928 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign, 929 const DataLayout *DL) { 930 assert(V->getType()->isPointerTy() && 931 "getOrEnforceKnownAlignment expects a pointer!"); 932 unsigned BitWidth = DL ? DL->getPointerTypeSizeInBits(V->getType()) : 64; 933 934 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 935 computeKnownBits(V, KnownZero, KnownOne, DL); 936 unsigned TrailZ = KnownZero.countTrailingOnes(); 937 938 // Avoid trouble with ridiculously large TrailZ values, such as 939 // those computed from a null pointer. 940 TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1)); 941 942 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ); 943 944 // LLVM doesn't support alignments larger than this currently. 945 Align = std::min(Align, +Value::MaximumAlignment); 946 947 if (PrefAlign > Align) 948 Align = enforceKnownAlignment(V, Align, PrefAlign, DL); 949 950 // We don't need to make any adjustment. 951 return Align; 952 } 953 954 ///===---------------------------------------------------------------------===// 955 /// Dbg Intrinsic utilities 956 /// 957 958 /// See if there is a dbg.value intrinsic for DIVar before I. 959 static bool LdStHasDebugValue(DIVariable &DIVar, Instruction *I) { 960 // Since we can't guarantee that the original dbg.declare instrinsic 961 // is removed by LowerDbgDeclare(), we need to make sure that we are 962 // not inserting the same dbg.value intrinsic over and over. 963 llvm::BasicBlock::InstListType::iterator PrevI(I); 964 if (PrevI != I->getParent()->getInstList().begin()) { 965 --PrevI; 966 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI)) 967 if (DVI->getValue() == I->getOperand(0) && 968 DVI->getOffset() == 0 && 969 DVI->getVariable() == DIVar) 970 return true; 971 } 972 return false; 973 } 974 975 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value 976 /// that has an associated llvm.dbg.decl intrinsic. 977 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, 978 StoreInst *SI, DIBuilder &Builder) { 979 DIVariable DIVar(DDI->getVariable()); 980 assert((!DIVar || DIVar.isVariable()) && 981 "Variable in DbgDeclareInst should be either null or a DIVariable."); 982 if (!DIVar) 983 return false; 984 985 if (LdStHasDebugValue(DIVar, SI)) 986 return true; 987 988 Instruction *DbgVal = nullptr; 989 // If an argument is zero extended then use argument directly. The ZExt 990 // may be zapped by an optimization pass in future. 991 Argument *ExtendedArg = nullptr; 992 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 993 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0)); 994 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 995 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0)); 996 if (ExtendedArg) 997 DbgVal = Builder.insertDbgValueIntrinsic(ExtendedArg, 0, DIVar, SI); 998 else 999 DbgVal = Builder.insertDbgValueIntrinsic(SI->getOperand(0), 0, DIVar, SI); 1000 DbgVal->setDebugLoc(DDI->getDebugLoc()); 1001 return true; 1002 } 1003 1004 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value 1005 /// that has an associated llvm.dbg.decl intrinsic. 1006 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, 1007 LoadInst *LI, DIBuilder &Builder) { 1008 DIVariable DIVar(DDI->getVariable()); 1009 assert((!DIVar || DIVar.isVariable()) && 1010 "Variable in DbgDeclareInst should be either null or a DIVariable."); 1011 if (!DIVar) 1012 return false; 1013 1014 if (LdStHasDebugValue(DIVar, LI)) 1015 return true; 1016 1017 Instruction *DbgVal = 1018 Builder.insertDbgValueIntrinsic(LI->getOperand(0), 0, 1019 DIVar, LI); 1020 DbgVal->setDebugLoc(DDI->getDebugLoc()); 1021 return true; 1022 } 1023 1024 /// Determine whether this alloca is either a VLA or an array. 1025 static bool isArray(AllocaInst *AI) { 1026 return AI->isArrayAllocation() || 1027 AI->getType()->getElementType()->isArrayTy(); 1028 } 1029 1030 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set 1031 /// of llvm.dbg.value intrinsics. 1032 bool llvm::LowerDbgDeclare(Function &F) { 1033 DIBuilder DIB(*F.getParent()); 1034 SmallVector<DbgDeclareInst *, 4> Dbgs; 1035 for (auto &FI : F) 1036 for (BasicBlock::iterator BI : FI) 1037 if (auto DDI = dyn_cast<DbgDeclareInst>(BI)) 1038 Dbgs.push_back(DDI); 1039 1040 if (Dbgs.empty()) 1041 return false; 1042 1043 for (auto &I : Dbgs) { 1044 DbgDeclareInst *DDI = I; 1045 AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress()); 1046 // If this is an alloca for a scalar variable, insert a dbg.value 1047 // at each load and store to the alloca and erase the dbg.declare. 1048 // The dbg.values allow tracking a variable even if it is not 1049 // stored on the stack, while the dbg.declare can only describe 1050 // the stack slot (and at a lexical-scope granularity). Later 1051 // passes will attempt to elide the stack slot. 1052 if (AI && !isArray(AI)) { 1053 for (User *U : AI->users()) 1054 if (StoreInst *SI = dyn_cast<StoreInst>(U)) 1055 ConvertDebugDeclareToDebugValue(DDI, SI, DIB); 1056 else if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1057 ConvertDebugDeclareToDebugValue(DDI, LI, DIB); 1058 else if (CallInst *CI = dyn_cast<CallInst>(U)) { 1059 // This is a call by-value or some other instruction that 1060 // takes a pointer to the variable. Insert a *value* 1061 // intrinsic that describes the alloca. 1062 auto DbgVal = 1063 DIB.insertDbgValueIntrinsic(AI, 0, 1064 DIVariable(DDI->getVariable()), CI); 1065 DbgVal->setDebugLoc(DDI->getDebugLoc()); 1066 } 1067 DDI->eraseFromParent(); 1068 } 1069 } 1070 return true; 1071 } 1072 1073 /// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the 1074 /// alloca 'V', if any. 1075 DbgDeclareInst *llvm::FindAllocaDbgDeclare(Value *V) { 1076 if (MDNode *DebugNode = MDNode::getIfExists(V->getContext(), V)) 1077 for (User *U : DebugNode->users()) 1078 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U)) 1079 return DDI; 1080 1081 return nullptr; 1082 } 1083 1084 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress, 1085 DIBuilder &Builder) { 1086 DbgDeclareInst *DDI = FindAllocaDbgDeclare(AI); 1087 if (!DDI) 1088 return false; 1089 DIVariable DIVar(DDI->getVariable()); 1090 assert((!DIVar || DIVar.isVariable()) && 1091 "Variable in DbgDeclareInst should be either null or a DIVariable."); 1092 if (!DIVar) 1093 return false; 1094 1095 // Create a copy of the original DIDescriptor for user variable, appending 1096 // "deref" operation to a list of address elements, as new llvm.dbg.declare 1097 // will take a value storing address of the memory for variable, not 1098 // alloca itself. 1099 Type *Int64Ty = Type::getInt64Ty(AI->getContext()); 1100 SmallVector<Value*, 4> NewDIVarAddress; 1101 if (DIVar.hasComplexAddress()) { 1102 for (unsigned i = 0, n = DIVar.getNumAddrElements(); i < n; ++i) { 1103 NewDIVarAddress.push_back( 1104 ConstantInt::get(Int64Ty, DIVar.getAddrElement(i))); 1105 } 1106 } 1107 NewDIVarAddress.push_back(ConstantInt::get(Int64Ty, DIBuilder::OpDeref)); 1108 DIVariable NewDIVar = Builder.createComplexVariable( 1109 DIVar.getTag(), DIVar.getContext(), DIVar.getName(), 1110 DIVar.getFile(), DIVar.getLineNumber(), DIVar.getType(), 1111 NewDIVarAddress, DIVar.getArgNumber()); 1112 1113 // Insert llvm.dbg.declare in the same basic block as the original alloca, 1114 // and remove old llvm.dbg.declare. 1115 BasicBlock *BB = AI->getParent(); 1116 Builder.insertDeclare(NewAllocaAddress, NewDIVar, BB); 1117 DDI->eraseFromParent(); 1118 return true; 1119 } 1120 1121 /// changeToUnreachable - Insert an unreachable instruction before the specified 1122 /// instruction, making it and the rest of the code in the block dead. 1123 static void changeToUnreachable(Instruction *I, bool UseLLVMTrap) { 1124 BasicBlock *BB = I->getParent(); 1125 // Loop over all of the successors, removing BB's entry from any PHI 1126 // nodes. 1127 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 1128 (*SI)->removePredecessor(BB); 1129 1130 // Insert a call to llvm.trap right before this. This turns the undefined 1131 // behavior into a hard fail instead of falling through into random code. 1132 if (UseLLVMTrap) { 1133 Function *TrapFn = 1134 Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap); 1135 CallInst *CallTrap = CallInst::Create(TrapFn, "", I); 1136 CallTrap->setDebugLoc(I->getDebugLoc()); 1137 } 1138 new UnreachableInst(I->getContext(), I); 1139 1140 // All instructions after this are dead. 1141 BasicBlock::iterator BBI = I, BBE = BB->end(); 1142 while (BBI != BBE) { 1143 if (!BBI->use_empty()) 1144 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 1145 BB->getInstList().erase(BBI++); 1146 } 1147 } 1148 1149 /// changeToCall - Convert the specified invoke into a normal call. 1150 static void changeToCall(InvokeInst *II) { 1151 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3); 1152 CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, "", II); 1153 NewCall->takeName(II); 1154 NewCall->setCallingConv(II->getCallingConv()); 1155 NewCall->setAttributes(II->getAttributes()); 1156 NewCall->setDebugLoc(II->getDebugLoc()); 1157 II->replaceAllUsesWith(NewCall); 1158 1159 // Follow the call by a branch to the normal destination. 1160 BranchInst::Create(II->getNormalDest(), II); 1161 1162 // Update PHI nodes in the unwind destination 1163 II->getUnwindDest()->removePredecessor(II->getParent()); 1164 II->eraseFromParent(); 1165 } 1166 1167 static bool markAliveBlocks(BasicBlock *BB, 1168 SmallPtrSet<BasicBlock*, 128> &Reachable) { 1169 1170 SmallVector<BasicBlock*, 128> Worklist; 1171 Worklist.push_back(BB); 1172 Reachable.insert(BB); 1173 bool Changed = false; 1174 do { 1175 BB = Worklist.pop_back_val(); 1176 1177 // Do a quick scan of the basic block, turning any obviously unreachable 1178 // instructions into LLVM unreachable insts. The instruction combining pass 1179 // canonicalizes unreachable insts into stores to null or undef. 1180 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E;++BBI){ 1181 if (CallInst *CI = dyn_cast<CallInst>(BBI)) { 1182 if (CI->doesNotReturn()) { 1183 // If we found a call to a no-return function, insert an unreachable 1184 // instruction after it. Make sure there isn't *already* one there 1185 // though. 1186 ++BBI; 1187 if (!isa<UnreachableInst>(BBI)) { 1188 // Don't insert a call to llvm.trap right before the unreachable. 1189 changeToUnreachable(BBI, false); 1190 Changed = true; 1191 } 1192 break; 1193 } 1194 } 1195 1196 // Store to undef and store to null are undefined and used to signal that 1197 // they should be changed to unreachable by passes that can't modify the 1198 // CFG. 1199 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 1200 // Don't touch volatile stores. 1201 if (SI->isVolatile()) continue; 1202 1203 Value *Ptr = SI->getOperand(1); 1204 1205 if (isa<UndefValue>(Ptr) || 1206 (isa<ConstantPointerNull>(Ptr) && 1207 SI->getPointerAddressSpace() == 0)) { 1208 changeToUnreachable(SI, true); 1209 Changed = true; 1210 break; 1211 } 1212 } 1213 } 1214 1215 // Turn invokes that call 'nounwind' functions into ordinary calls. 1216 if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator())) { 1217 Value *Callee = II->getCalledValue(); 1218 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 1219 changeToUnreachable(II, true); 1220 Changed = true; 1221 } else if (II->doesNotThrow()) { 1222 if (II->use_empty() && II->onlyReadsMemory()) { 1223 // jump to the normal destination branch. 1224 BranchInst::Create(II->getNormalDest(), II); 1225 II->getUnwindDest()->removePredecessor(II->getParent()); 1226 II->eraseFromParent(); 1227 } else 1228 changeToCall(II); 1229 Changed = true; 1230 } 1231 } 1232 1233 Changed |= ConstantFoldTerminator(BB, true); 1234 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 1235 if (Reachable.insert(*SI)) 1236 Worklist.push_back(*SI); 1237 } while (!Worklist.empty()); 1238 return Changed; 1239 } 1240 1241 /// removeUnreachableBlocksFromFn - Remove blocks that are not reachable, even 1242 /// if they are in a dead cycle. Return true if a change was made, false 1243 /// otherwise. 1244 bool llvm::removeUnreachableBlocks(Function &F) { 1245 SmallPtrSet<BasicBlock*, 128> Reachable; 1246 bool Changed = markAliveBlocks(F.begin(), Reachable); 1247 1248 // If there are unreachable blocks in the CFG... 1249 if (Reachable.size() == F.size()) 1250 return Changed; 1251 1252 assert(Reachable.size() < F.size()); 1253 NumRemoved += F.size()-Reachable.size(); 1254 1255 // Loop over all of the basic blocks that are not reachable, dropping all of 1256 // their internal references... 1257 for (Function::iterator BB = ++F.begin(), E = F.end(); BB != E; ++BB) { 1258 if (Reachable.count(BB)) 1259 continue; 1260 1261 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 1262 if (Reachable.count(*SI)) 1263 (*SI)->removePredecessor(BB); 1264 BB->dropAllReferences(); 1265 } 1266 1267 for (Function::iterator I = ++F.begin(); I != F.end();) 1268 if (!Reachable.count(I)) 1269 I = F.getBasicBlockList().erase(I); 1270 else 1271 ++I; 1272 1273 return true; 1274 } 1275