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/DenseSet.h" 18 #include "llvm/ADT/Hashing.h" 19 #include "llvm/ADT/STLExtras.h" 20 #include "llvm/ADT/SetVector.h" 21 #include "llvm/ADT/SmallPtrSet.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/Analysis/EHPersonalities.h" 24 #include "llvm/Analysis/InstructionSimplify.h" 25 #include "llvm/Analysis/MemoryBuiltins.h" 26 #include "llvm/Analysis/ValueTracking.h" 27 #include "llvm/IR/CFG.h" 28 #include "llvm/IR/Constants.h" 29 #include "llvm/IR/DIBuilder.h" 30 #include "llvm/IR/DataLayout.h" 31 #include "llvm/IR/DebugInfo.h" 32 #include "llvm/IR/DerivedTypes.h" 33 #include "llvm/IR/Dominators.h" 34 #include "llvm/IR/GetElementPtrTypeIterator.h" 35 #include "llvm/IR/GlobalAlias.h" 36 #include "llvm/IR/GlobalVariable.h" 37 #include "llvm/IR/IRBuilder.h" 38 #include "llvm/IR/Instructions.h" 39 #include "llvm/IR/IntrinsicInst.h" 40 #include "llvm/IR/Intrinsics.h" 41 #include "llvm/IR/MDBuilder.h" 42 #include "llvm/IR/Metadata.h" 43 #include "llvm/IR/Operator.h" 44 #include "llvm/IR/ValueHandle.h" 45 #include "llvm/Support/Debug.h" 46 #include "llvm/Support/MathExtras.h" 47 #include "llvm/Support/raw_ostream.h" 48 using namespace llvm; 49 50 #define DEBUG_TYPE "local" 51 52 STATISTIC(NumRemoved, "Number of unreachable basic blocks removed"); 53 54 //===----------------------------------------------------------------------===// 55 // Local constant propagation. 56 // 57 58 /// ConstantFoldTerminator - If a terminator instruction is predicated on a 59 /// constant value, convert it into an unconditional branch to the constant 60 /// destination. This is a nontrivial operation because the successors of this 61 /// basic block must have their PHI nodes updated. 62 /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch 63 /// conditions and indirectbr addresses this might make dead if 64 /// DeleteDeadConditions is true. 65 bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions, 66 const TargetLibraryInfo *TLI) { 67 TerminatorInst *T = BB->getTerminator(); 68 IRBuilder<> Builder(T); 69 70 // Branch - See if we are conditional jumping on constant 71 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 72 if (BI->isUnconditional()) return false; // Can't optimize uncond branch 73 BasicBlock *Dest1 = BI->getSuccessor(0); 74 BasicBlock *Dest2 = BI->getSuccessor(1); 75 76 if (ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition())) { 77 // Are we branching on constant? 78 // YES. Change to unconditional branch... 79 BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2; 80 BasicBlock *OldDest = Cond->getZExtValue() ? Dest2 : Dest1; 81 82 //cerr << "Function: " << T->getParent()->getParent() 83 // << "\nRemoving branch from " << T->getParent() 84 // << "\n\nTo: " << OldDest << endl; 85 86 // Let the basic block know that we are letting go of it. Based on this, 87 // it will adjust it's PHI nodes. 88 OldDest->removePredecessor(BB); 89 90 // Replace the conditional branch with an unconditional one. 91 Builder.CreateBr(Destination); 92 BI->eraseFromParent(); 93 return true; 94 } 95 96 if (Dest2 == Dest1) { // Conditional branch to same location? 97 // This branch matches something like this: 98 // br bool %cond, label %Dest, label %Dest 99 // and changes it into: br label %Dest 100 101 // Let the basic block know that we are letting go of one copy of it. 102 assert(BI->getParent() && "Terminator not inserted in block!"); 103 Dest1->removePredecessor(BI->getParent()); 104 105 // Replace the conditional branch with an unconditional one. 106 Builder.CreateBr(Dest1); 107 Value *Cond = BI->getCondition(); 108 BI->eraseFromParent(); 109 if (DeleteDeadConditions) 110 RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI); 111 return true; 112 } 113 return false; 114 } 115 116 if (SwitchInst *SI = dyn_cast<SwitchInst>(T)) { 117 // If we are switching on a constant, we can convert the switch to an 118 // unconditional branch. 119 ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition()); 120 BasicBlock *DefaultDest = SI->getDefaultDest(); 121 BasicBlock *TheOnlyDest = DefaultDest; 122 123 // If the default is unreachable, ignore it when searching for TheOnlyDest. 124 if (isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg()) && 125 SI->getNumCases() > 0) { 126 TheOnlyDest = SI->case_begin().getCaseSuccessor(); 127 } 128 129 // Figure out which case it goes to. 130 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); 131 i != e; ++i) { 132 // Found case matching a constant operand? 133 if (i.getCaseValue() == CI) { 134 TheOnlyDest = i.getCaseSuccessor(); 135 break; 136 } 137 138 // Check to see if this branch is going to the same place as the default 139 // dest. If so, eliminate it as an explicit compare. 140 if (i.getCaseSuccessor() == DefaultDest) { 141 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof); 142 unsigned NCases = SI->getNumCases(); 143 // Fold the case metadata into the default if there will be any branches 144 // left, unless the metadata doesn't match the switch. 145 if (NCases > 1 && MD && MD->getNumOperands() == 2 + NCases) { 146 // Collect branch weights into a vector. 147 SmallVector<uint32_t, 8> Weights; 148 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; 149 ++MD_i) { 150 ConstantInt *CI = 151 mdconst::dyn_extract<ConstantInt>(MD->getOperand(MD_i)); 152 assert(CI); 153 Weights.push_back(CI->getValue().getZExtValue()); 154 } 155 // Merge weight of this case to the default weight. 156 unsigned idx = i.getCaseIndex(); 157 Weights[0] += Weights[idx+1]; 158 // Remove weight for this case. 159 std::swap(Weights[idx+1], Weights.back()); 160 Weights.pop_back(); 161 SI->setMetadata(LLVMContext::MD_prof, 162 MDBuilder(BB->getContext()). 163 createBranchWeights(Weights)); 164 } 165 // Remove this entry. 166 DefaultDest->removePredecessor(SI->getParent()); 167 SI->removeCase(i); 168 --i; --e; 169 continue; 170 } 171 172 // Otherwise, check to see if the switch only branches to one destination. 173 // We do this by reseting "TheOnlyDest" to null when we find two non-equal 174 // destinations. 175 if (i.getCaseSuccessor() != TheOnlyDest) TheOnlyDest = nullptr; 176 } 177 178 if (CI && !TheOnlyDest) { 179 // Branching on a constant, but not any of the cases, go to the default 180 // successor. 181 TheOnlyDest = SI->getDefaultDest(); 182 } 183 184 // If we found a single destination that we can fold the switch into, do so 185 // now. 186 if (TheOnlyDest) { 187 // Insert the new branch. 188 Builder.CreateBr(TheOnlyDest); 189 BasicBlock *BB = SI->getParent(); 190 191 // Remove entries from PHI nodes which we no longer branch to... 192 for (BasicBlock *Succ : SI->successors()) { 193 // Found case matching a constant operand? 194 if (Succ == TheOnlyDest) 195 TheOnlyDest = nullptr; // Don't modify the first branch to TheOnlyDest 196 else 197 Succ->removePredecessor(BB); 198 } 199 200 // Delete the old switch. 201 Value *Cond = SI->getCondition(); 202 SI->eraseFromParent(); 203 if (DeleteDeadConditions) 204 RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI); 205 return true; 206 } 207 208 if (SI->getNumCases() == 1) { 209 // Otherwise, we can fold this switch into a conditional branch 210 // instruction if it has only one non-default destination. 211 SwitchInst::CaseIt FirstCase = SI->case_begin(); 212 Value *Cond = Builder.CreateICmpEQ(SI->getCondition(), 213 FirstCase.getCaseValue(), "cond"); 214 215 // Insert the new branch. 216 BranchInst *NewBr = Builder.CreateCondBr(Cond, 217 FirstCase.getCaseSuccessor(), 218 SI->getDefaultDest()); 219 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof); 220 if (MD && MD->getNumOperands() == 3) { 221 ConstantInt *SICase = 222 mdconst::dyn_extract<ConstantInt>(MD->getOperand(2)); 223 ConstantInt *SIDef = 224 mdconst::dyn_extract<ConstantInt>(MD->getOperand(1)); 225 assert(SICase && SIDef); 226 // The TrueWeight should be the weight for the single case of SI. 227 NewBr->setMetadata(LLVMContext::MD_prof, 228 MDBuilder(BB->getContext()). 229 createBranchWeights(SICase->getValue().getZExtValue(), 230 SIDef->getValue().getZExtValue())); 231 } 232 233 // Update make.implicit metadata to the newly-created conditional branch. 234 MDNode *MakeImplicitMD = SI->getMetadata(LLVMContext::MD_make_implicit); 235 if (MakeImplicitMD) 236 NewBr->setMetadata(LLVMContext::MD_make_implicit, MakeImplicitMD); 237 238 // Delete the old switch. 239 SI->eraseFromParent(); 240 return true; 241 } 242 return false; 243 } 244 245 if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(T)) { 246 // indirectbr blockaddress(@F, @BB) -> br label @BB 247 if (BlockAddress *BA = 248 dyn_cast<BlockAddress>(IBI->getAddress()->stripPointerCasts())) { 249 BasicBlock *TheOnlyDest = BA->getBasicBlock(); 250 // Insert the new branch. 251 Builder.CreateBr(TheOnlyDest); 252 253 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 254 if (IBI->getDestination(i) == TheOnlyDest) 255 TheOnlyDest = nullptr; 256 else 257 IBI->getDestination(i)->removePredecessor(IBI->getParent()); 258 } 259 Value *Address = IBI->getAddress(); 260 IBI->eraseFromParent(); 261 if (DeleteDeadConditions) 262 RecursivelyDeleteTriviallyDeadInstructions(Address, TLI); 263 264 // If we didn't find our destination in the IBI successor list, then we 265 // have undefined behavior. Replace the unconditional branch with an 266 // 'unreachable' instruction. 267 if (TheOnlyDest) { 268 BB->getTerminator()->eraseFromParent(); 269 new UnreachableInst(BB->getContext(), BB); 270 } 271 272 return true; 273 } 274 } 275 276 return false; 277 } 278 279 280 //===----------------------------------------------------------------------===// 281 // Local dead code elimination. 282 // 283 284 /// isInstructionTriviallyDead - Return true if the result produced by the 285 /// instruction is not used, and the instruction has no side effects. 286 /// 287 bool llvm::isInstructionTriviallyDead(Instruction *I, 288 const TargetLibraryInfo *TLI) { 289 if (!I->use_empty() || isa<TerminatorInst>(I)) return false; 290 291 // We don't want the landingpad-like instructions removed by anything this 292 // general. 293 if (I->isEHPad()) 294 return false; 295 296 // We don't want debug info removed by anything this general, unless 297 // debug info is empty. 298 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(I)) { 299 if (DDI->getAddress()) 300 return false; 301 return true; 302 } 303 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(I)) { 304 if (DVI->getValue()) 305 return false; 306 return true; 307 } 308 309 if (!I->mayHaveSideEffects()) return true; 310 311 // Special case intrinsics that "may have side effects" but can be deleted 312 // when dead. 313 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 314 // Safe to delete llvm.stacksave if dead. 315 if (II->getIntrinsicID() == Intrinsic::stacksave) 316 return true; 317 318 // Lifetime intrinsics are dead when their right-hand is undef. 319 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 320 II->getIntrinsicID() == Intrinsic::lifetime_end) 321 return isa<UndefValue>(II->getArgOperand(1)); 322 323 // Assumptions are dead if their condition is trivially true. 324 if (II->getIntrinsicID() == Intrinsic::assume) { 325 if (ConstantInt *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0))) 326 return !Cond->isZero(); 327 328 return false; 329 } 330 } 331 332 if (isAllocLikeFn(I, TLI)) return true; 333 334 if (CallInst *CI = isFreeCall(I, TLI)) 335 if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0))) 336 return C->isNullValue() || isa<UndefValue>(C); 337 338 return false; 339 } 340 341 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a 342 /// trivially dead instruction, delete it. If that makes any of its operands 343 /// trivially dead, delete them too, recursively. Return true if any 344 /// instructions were deleted. 345 bool 346 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V, 347 const TargetLibraryInfo *TLI) { 348 Instruction *I = dyn_cast<Instruction>(V); 349 if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI)) 350 return false; 351 352 SmallVector<Instruction*, 16> DeadInsts; 353 DeadInsts.push_back(I); 354 355 do { 356 I = DeadInsts.pop_back_val(); 357 358 // Null out all of the instruction's operands to see if any operand becomes 359 // dead as we go. 360 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 361 Value *OpV = I->getOperand(i); 362 I->setOperand(i, nullptr); 363 364 if (!OpV->use_empty()) continue; 365 366 // If the operand is an instruction that became dead as we nulled out the 367 // operand, and if it is 'trivially' dead, delete it in a future loop 368 // iteration. 369 if (Instruction *OpI = dyn_cast<Instruction>(OpV)) 370 if (isInstructionTriviallyDead(OpI, TLI)) 371 DeadInsts.push_back(OpI); 372 } 373 374 I->eraseFromParent(); 375 } while (!DeadInsts.empty()); 376 377 return true; 378 } 379 380 /// areAllUsesEqual - Check whether the uses of a value are all the same. 381 /// This is similar to Instruction::hasOneUse() except this will also return 382 /// true when there are no uses or multiple uses that all refer to the same 383 /// value. 384 static bool areAllUsesEqual(Instruction *I) { 385 Value::user_iterator UI = I->user_begin(); 386 Value::user_iterator UE = I->user_end(); 387 if (UI == UE) 388 return true; 389 390 User *TheUse = *UI; 391 for (++UI; UI != UE; ++UI) { 392 if (*UI != TheUse) 393 return false; 394 } 395 return true; 396 } 397 398 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively 399 /// dead PHI node, due to being a def-use chain of single-use nodes that 400 /// either forms a cycle or is terminated by a trivially dead instruction, 401 /// delete it. If that makes any of its operands trivially dead, delete them 402 /// too, recursively. Return true if a change was made. 403 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN, 404 const TargetLibraryInfo *TLI) { 405 SmallPtrSet<Instruction*, 4> Visited; 406 for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects(); 407 I = cast<Instruction>(*I->user_begin())) { 408 if (I->use_empty()) 409 return RecursivelyDeleteTriviallyDeadInstructions(I, TLI); 410 411 // If we find an instruction more than once, we're on a cycle that 412 // won't prove fruitful. 413 if (!Visited.insert(I).second) { 414 // Break the cycle and delete the instruction and its operands. 415 I->replaceAllUsesWith(UndefValue::get(I->getType())); 416 (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI); 417 return true; 418 } 419 } 420 return false; 421 } 422 423 static bool 424 simplifyAndDCEInstruction(Instruction *I, 425 SmallSetVector<Instruction *, 16> &WorkList, 426 const DataLayout &DL, 427 const TargetLibraryInfo *TLI) { 428 if (isInstructionTriviallyDead(I, TLI)) { 429 // Null out all of the instruction's operands to see if any operand becomes 430 // dead as we go. 431 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 432 Value *OpV = I->getOperand(i); 433 I->setOperand(i, nullptr); 434 435 if (!OpV->use_empty() || I == OpV) 436 continue; 437 438 // If the operand is an instruction that became dead as we nulled out the 439 // operand, and if it is 'trivially' dead, delete it in a future loop 440 // iteration. 441 if (Instruction *OpI = dyn_cast<Instruction>(OpV)) 442 if (isInstructionTriviallyDead(OpI, TLI)) 443 WorkList.insert(OpI); 444 } 445 446 I->eraseFromParent(); 447 448 return true; 449 } 450 451 if (Value *SimpleV = SimplifyInstruction(I, DL)) { 452 // Add the users to the worklist. CAREFUL: an instruction can use itself, 453 // in the case of a phi node. 454 for (User *U : I->users()) 455 if (U != I) 456 WorkList.insert(cast<Instruction>(U)); 457 458 // Replace the instruction with its simplified value. 459 I->replaceAllUsesWith(SimpleV); 460 I->eraseFromParent(); 461 return true; 462 } 463 return false; 464 } 465 466 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to 467 /// simplify any instructions in it and recursively delete dead instructions. 468 /// 469 /// This returns true if it changed the code, note that it can delete 470 /// instructions in other blocks as well in this block. 471 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB, 472 const TargetLibraryInfo *TLI) { 473 bool MadeChange = false; 474 const DataLayout &DL = BB->getModule()->getDataLayout(); 475 476 #ifndef NDEBUG 477 // In debug builds, ensure that the terminator of the block is never replaced 478 // or deleted by these simplifications. The idea of simplification is that it 479 // cannot introduce new instructions, and there is no way to replace the 480 // terminator of a block without introducing a new instruction. 481 AssertingVH<Instruction> TerminatorVH(&BB->back()); 482 #endif 483 484 SmallSetVector<Instruction *, 16> WorkList; 485 // Iterate over the original function, only adding insts to the worklist 486 // if they actually need to be revisited. This avoids having to pre-init 487 // the worklist with the entire function's worth of instructions. 488 for (BasicBlock::iterator BI = BB->begin(), E = std::prev(BB->end()); BI != E;) { 489 assert(!BI->isTerminator()); 490 Instruction *I = &*BI; 491 ++BI; 492 493 // We're visiting this instruction now, so make sure it's not in the 494 // worklist from an earlier visit. 495 if (!WorkList.count(I)) 496 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI); 497 } 498 499 while (!WorkList.empty()) { 500 Instruction *I = WorkList.pop_back_val(); 501 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI); 502 } 503 return MadeChange; 504 } 505 506 //===----------------------------------------------------------------------===// 507 // Control Flow Graph Restructuring. 508 // 509 510 511 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this 512 /// method is called when we're about to delete Pred as a predecessor of BB. If 513 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred. 514 /// 515 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI 516 /// nodes that collapse into identity values. For example, if we have: 517 /// x = phi(1, 0, 0, 0) 518 /// y = and x, z 519 /// 520 /// .. and delete the predecessor corresponding to the '1', this will attempt to 521 /// recursively fold the and to 0. 522 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred) { 523 // This only adjusts blocks with PHI nodes. 524 if (!isa<PHINode>(BB->begin())) 525 return; 526 527 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify 528 // them down. This will leave us with single entry phi nodes and other phis 529 // that can be removed. 530 BB->removePredecessor(Pred, true); 531 532 WeakVH PhiIt = &BB->front(); 533 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) { 534 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt)); 535 Value *OldPhiIt = PhiIt; 536 537 if (!recursivelySimplifyInstruction(PN)) 538 continue; 539 540 // If recursive simplification ended up deleting the next PHI node we would 541 // iterate to, then our iterator is invalid, restart scanning from the top 542 // of the block. 543 if (PhiIt != OldPhiIt) PhiIt = &BB->front(); 544 } 545 } 546 547 548 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its 549 /// predecessor is known to have one successor (DestBB!). Eliminate the edge 550 /// between them, moving the instructions in the predecessor into DestBB and 551 /// deleting the predecessor block. 552 /// 553 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, DominatorTree *DT) { 554 // If BB has single-entry PHI nodes, fold them. 555 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 556 Value *NewVal = PN->getIncomingValue(0); 557 // Replace self referencing PHI with undef, it must be dead. 558 if (NewVal == PN) NewVal = UndefValue::get(PN->getType()); 559 PN->replaceAllUsesWith(NewVal); 560 PN->eraseFromParent(); 561 } 562 563 BasicBlock *PredBB = DestBB->getSinglePredecessor(); 564 assert(PredBB && "Block doesn't have a single predecessor!"); 565 566 // Zap anything that took the address of DestBB. Not doing this will give the 567 // address an invalid value. 568 if (DestBB->hasAddressTaken()) { 569 BlockAddress *BA = BlockAddress::get(DestBB); 570 Constant *Replacement = 571 ConstantInt::get(llvm::Type::getInt32Ty(BA->getContext()), 1); 572 BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement, 573 BA->getType())); 574 BA->destroyConstant(); 575 } 576 577 // Anything that branched to PredBB now branches to DestBB. 578 PredBB->replaceAllUsesWith(DestBB); 579 580 // Splice all the instructions from PredBB to DestBB. 581 PredBB->getTerminator()->eraseFromParent(); 582 DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList()); 583 584 // If the PredBB is the entry block of the function, move DestBB up to 585 // become the entry block after we erase PredBB. 586 if (PredBB == &DestBB->getParent()->getEntryBlock()) 587 DestBB->moveAfter(PredBB); 588 589 if (DT) { 590 BasicBlock *PredBBIDom = DT->getNode(PredBB)->getIDom()->getBlock(); 591 DT->changeImmediateDominator(DestBB, PredBBIDom); 592 DT->eraseNode(PredBB); 593 } 594 // Nuke BB. 595 PredBB->eraseFromParent(); 596 } 597 598 /// CanMergeValues - Return true if we can choose one of these values to use 599 /// in place of the other. Note that we will always choose the non-undef 600 /// value to keep. 601 static bool CanMergeValues(Value *First, Value *Second) { 602 return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second); 603 } 604 605 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an 606 /// almost-empty BB ending in an unconditional branch to Succ, into Succ. 607 /// 608 /// Assumption: Succ is the single successor for BB. 609 /// 610 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { 611 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!"); 612 613 DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into " 614 << Succ->getName() << "\n"); 615 // Shortcut, if there is only a single predecessor it must be BB and merging 616 // is always safe 617 if (Succ->getSinglePredecessor()) return true; 618 619 // Make a list of the predecessors of BB 620 SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB)); 621 622 // Look at all the phi nodes in Succ, to see if they present a conflict when 623 // merging these blocks 624 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 625 PHINode *PN = cast<PHINode>(I); 626 627 // If the incoming value from BB is again a PHINode in 628 // BB which has the same incoming value for *PI as PN does, we can 629 // merge the phi nodes and then the blocks can still be merged 630 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB)); 631 if (BBPN && BBPN->getParent() == BB) { 632 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) { 633 BasicBlock *IBB = PN->getIncomingBlock(PI); 634 if (BBPreds.count(IBB) && 635 !CanMergeValues(BBPN->getIncomingValueForBlock(IBB), 636 PN->getIncomingValue(PI))) { 637 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in " 638 << Succ->getName() << " is conflicting with " 639 << BBPN->getName() << " with regard to common predecessor " 640 << IBB->getName() << "\n"); 641 return false; 642 } 643 } 644 } else { 645 Value* Val = PN->getIncomingValueForBlock(BB); 646 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) { 647 // See if the incoming value for the common predecessor is equal to the 648 // one for BB, in which case this phi node will not prevent the merging 649 // of the block. 650 BasicBlock *IBB = PN->getIncomingBlock(PI); 651 if (BBPreds.count(IBB) && 652 !CanMergeValues(Val, PN->getIncomingValue(PI))) { 653 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in " 654 << Succ->getName() << " is conflicting with regard to common " 655 << "predecessor " << IBB->getName() << "\n"); 656 return false; 657 } 658 } 659 } 660 } 661 662 return true; 663 } 664 665 typedef SmallVector<BasicBlock *, 16> PredBlockVector; 666 typedef DenseMap<BasicBlock *, Value *> IncomingValueMap; 667 668 /// \brief Determines the value to use as the phi node input for a block. 669 /// 670 /// Select between \p OldVal any value that we know flows from \p BB 671 /// to a particular phi on the basis of which one (if either) is not 672 /// undef. Update IncomingValues based on the selected value. 673 /// 674 /// \param OldVal The value we are considering selecting. 675 /// \param BB The block that the value flows in from. 676 /// \param IncomingValues A map from block-to-value for other phi inputs 677 /// that we have examined. 678 /// 679 /// \returns the selected value. 680 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB, 681 IncomingValueMap &IncomingValues) { 682 if (!isa<UndefValue>(OldVal)) { 683 assert((!IncomingValues.count(BB) || 684 IncomingValues.find(BB)->second == OldVal) && 685 "Expected OldVal to match incoming value from BB!"); 686 687 IncomingValues.insert(std::make_pair(BB, OldVal)); 688 return OldVal; 689 } 690 691 IncomingValueMap::const_iterator It = IncomingValues.find(BB); 692 if (It != IncomingValues.end()) return It->second; 693 694 return OldVal; 695 } 696 697 /// \brief Create a map from block to value for the operands of a 698 /// given phi. 699 /// 700 /// Create a map from block to value for each non-undef value flowing 701 /// into \p PN. 702 /// 703 /// \param PN The phi we are collecting the map for. 704 /// \param IncomingValues [out] The map from block to value for this phi. 705 static void gatherIncomingValuesToPhi(PHINode *PN, 706 IncomingValueMap &IncomingValues) { 707 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 708 BasicBlock *BB = PN->getIncomingBlock(i); 709 Value *V = PN->getIncomingValue(i); 710 711 if (!isa<UndefValue>(V)) 712 IncomingValues.insert(std::make_pair(BB, V)); 713 } 714 } 715 716 /// \brief Replace the incoming undef values to a phi with the values 717 /// from a block-to-value map. 718 /// 719 /// \param PN The phi we are replacing the undefs in. 720 /// \param IncomingValues A map from block to value. 721 static void replaceUndefValuesInPhi(PHINode *PN, 722 const IncomingValueMap &IncomingValues) { 723 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 724 Value *V = PN->getIncomingValue(i); 725 726 if (!isa<UndefValue>(V)) continue; 727 728 BasicBlock *BB = PN->getIncomingBlock(i); 729 IncomingValueMap::const_iterator It = IncomingValues.find(BB); 730 if (It == IncomingValues.end()) continue; 731 732 PN->setIncomingValue(i, It->second); 733 } 734 } 735 736 /// \brief Replace a value flowing from a block to a phi with 737 /// potentially multiple instances of that value flowing from the 738 /// block's predecessors to the phi. 739 /// 740 /// \param BB The block with the value flowing into the phi. 741 /// \param BBPreds The predecessors of BB. 742 /// \param PN The phi that we are updating. 743 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB, 744 const PredBlockVector &BBPreds, 745 PHINode *PN) { 746 Value *OldVal = PN->removeIncomingValue(BB, false); 747 assert(OldVal && "No entry in PHI for Pred BB!"); 748 749 IncomingValueMap IncomingValues; 750 751 // We are merging two blocks - BB, and the block containing PN - and 752 // as a result we need to redirect edges from the predecessors of BB 753 // to go to the block containing PN, and update PN 754 // accordingly. Since we allow merging blocks in the case where the 755 // predecessor and successor blocks both share some predecessors, 756 // and where some of those common predecessors might have undef 757 // values flowing into PN, we want to rewrite those values to be 758 // consistent with the non-undef values. 759 760 gatherIncomingValuesToPhi(PN, IncomingValues); 761 762 // If this incoming value is one of the PHI nodes in BB, the new entries 763 // in the PHI node are the entries from the old PHI. 764 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) { 765 PHINode *OldValPN = cast<PHINode>(OldVal); 766 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) { 767 // Note that, since we are merging phi nodes and BB and Succ might 768 // have common predecessors, we could end up with a phi node with 769 // identical incoming branches. This will be cleaned up later (and 770 // will trigger asserts if we try to clean it up now, without also 771 // simplifying the corresponding conditional branch). 772 BasicBlock *PredBB = OldValPN->getIncomingBlock(i); 773 Value *PredVal = OldValPN->getIncomingValue(i); 774 Value *Selected = selectIncomingValueForBlock(PredVal, PredBB, 775 IncomingValues); 776 777 // And add a new incoming value for this predecessor for the 778 // newly retargeted branch. 779 PN->addIncoming(Selected, PredBB); 780 } 781 } else { 782 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) { 783 // Update existing incoming values in PN for this 784 // predecessor of BB. 785 BasicBlock *PredBB = BBPreds[i]; 786 Value *Selected = selectIncomingValueForBlock(OldVal, PredBB, 787 IncomingValues); 788 789 // And add a new incoming value for this predecessor for the 790 // newly retargeted branch. 791 PN->addIncoming(Selected, PredBB); 792 } 793 } 794 795 replaceUndefValuesInPhi(PN, IncomingValues); 796 } 797 798 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an 799 /// unconditional branch, and contains no instructions other than PHI nodes, 800 /// potential side-effect free intrinsics and the branch. If possible, 801 /// eliminate BB by rewriting all the predecessors to branch to the successor 802 /// block and return true. If we can't transform, return false. 803 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB) { 804 assert(BB != &BB->getParent()->getEntryBlock() && 805 "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!"); 806 807 // We can't eliminate infinite loops. 808 BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0); 809 if (BB == Succ) return false; 810 811 // Check to see if merging these blocks would cause conflicts for any of the 812 // phi nodes in BB or Succ. If not, we can safely merge. 813 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false; 814 815 // Check for cases where Succ has multiple predecessors and a PHI node in BB 816 // has uses which will not disappear when the PHI nodes are merged. It is 817 // possible to handle such cases, but difficult: it requires checking whether 818 // BB dominates Succ, which is non-trivial to calculate in the case where 819 // Succ has multiple predecessors. Also, it requires checking whether 820 // constructing the necessary self-referential PHI node doesn't introduce any 821 // conflicts; this isn't too difficult, but the previous code for doing this 822 // was incorrect. 823 // 824 // Note that if this check finds a live use, BB dominates Succ, so BB is 825 // something like a loop pre-header (or rarely, a part of an irreducible CFG); 826 // folding the branch isn't profitable in that case anyway. 827 if (!Succ->getSinglePredecessor()) { 828 BasicBlock::iterator BBI = BB->begin(); 829 while (isa<PHINode>(*BBI)) { 830 for (Use &U : BBI->uses()) { 831 if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) { 832 if (PN->getIncomingBlock(U) != BB) 833 return false; 834 } else { 835 return false; 836 } 837 } 838 ++BBI; 839 } 840 } 841 842 DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB); 843 844 if (isa<PHINode>(Succ->begin())) { 845 // If there is more than one pred of succ, and there are PHI nodes in 846 // the successor, then we need to add incoming edges for the PHI nodes 847 // 848 const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB)); 849 850 // Loop over all of the PHI nodes in the successor of BB. 851 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 852 PHINode *PN = cast<PHINode>(I); 853 854 redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN); 855 } 856 } 857 858 if (Succ->getSinglePredecessor()) { 859 // BB is the only predecessor of Succ, so Succ will end up with exactly 860 // the same predecessors BB had. 861 862 // Copy over any phi, debug or lifetime instruction. 863 BB->getTerminator()->eraseFromParent(); 864 Succ->getInstList().splice(Succ->getFirstNonPHI()->getIterator(), 865 BB->getInstList()); 866 } else { 867 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) { 868 // We explicitly check for such uses in CanPropagatePredecessorsForPHIs. 869 assert(PN->use_empty() && "There shouldn't be any uses here!"); 870 PN->eraseFromParent(); 871 } 872 } 873 874 // Everything that jumped to BB now goes to Succ. 875 BB->replaceAllUsesWith(Succ); 876 if (!Succ->hasName()) Succ->takeName(BB); 877 BB->eraseFromParent(); // Delete the old basic block. 878 return true; 879 } 880 881 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI 882 /// nodes in this block. This doesn't try to be clever about PHI nodes 883 /// which differ only in the order of the incoming values, but instcombine 884 /// orders them so it usually won't matter. 885 /// 886 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) { 887 // This implementation doesn't currently consider undef operands 888 // specially. Theoretically, two phis which are identical except for 889 // one having an undef where the other doesn't could be collapsed. 890 891 struct PHIDenseMapInfo { 892 static PHINode *getEmptyKey() { 893 return DenseMapInfo<PHINode *>::getEmptyKey(); 894 } 895 static PHINode *getTombstoneKey() { 896 return DenseMapInfo<PHINode *>::getTombstoneKey(); 897 } 898 static unsigned getHashValue(PHINode *PN) { 899 // Compute a hash value on the operands. Instcombine will likely have 900 // sorted them, which helps expose duplicates, but we have to check all 901 // the operands to be safe in case instcombine hasn't run. 902 return static_cast<unsigned>(hash_combine( 903 hash_combine_range(PN->value_op_begin(), PN->value_op_end()), 904 hash_combine_range(PN->block_begin(), PN->block_end()))); 905 } 906 static bool isEqual(PHINode *LHS, PHINode *RHS) { 907 if (LHS == getEmptyKey() || LHS == getTombstoneKey() || 908 RHS == getEmptyKey() || RHS == getTombstoneKey()) 909 return LHS == RHS; 910 return LHS->isIdenticalTo(RHS); 911 } 912 }; 913 914 // Set of unique PHINodes. 915 DenseSet<PHINode *, PHIDenseMapInfo> PHISet; 916 917 // Examine each PHI. 918 bool Changed = false; 919 for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I++);) { 920 auto Inserted = PHISet.insert(PN); 921 if (!Inserted.second) { 922 // A duplicate. Replace this PHI with its duplicate. 923 PN->replaceAllUsesWith(*Inserted.first); 924 PN->eraseFromParent(); 925 Changed = true; 926 927 // The RAUW can change PHIs that we already visited. Start over from the 928 // beginning. 929 PHISet.clear(); 930 I = BB->begin(); 931 } 932 } 933 934 return Changed; 935 } 936 937 /// enforceKnownAlignment - If the specified pointer points to an object that 938 /// we control, modify the object's alignment to PrefAlign. This isn't 939 /// often possible though. If alignment is important, a more reliable approach 940 /// is to simply align all global variables and allocation instructions to 941 /// their preferred alignment from the beginning. 942 /// 943 static unsigned enforceKnownAlignment(Value *V, unsigned Align, 944 unsigned PrefAlign, 945 const DataLayout &DL) { 946 V = V->stripPointerCasts(); 947 948 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 949 // If the preferred alignment is greater than the natural stack alignment 950 // then don't round up. This avoids dynamic stack realignment. 951 if (DL.exceedsNaturalStackAlignment(PrefAlign)) 952 return Align; 953 // If there is a requested alignment and if this is an alloca, round up. 954 if (AI->getAlignment() >= PrefAlign) 955 return AI->getAlignment(); 956 AI->setAlignment(PrefAlign); 957 return PrefAlign; 958 } 959 960 if (auto *GO = dyn_cast<GlobalObject>(V)) { 961 // If there is a large requested alignment and we can, bump up the alignment 962 // of the global. If the memory we set aside for the global may not be the 963 // memory used by the final program then it is impossible for us to reliably 964 // enforce the preferred alignment. 965 if (!GO->isStrongDefinitionForLinker()) 966 return Align; 967 968 if (GO->getAlignment() >= PrefAlign) 969 return GO->getAlignment(); 970 // We can only increase the alignment of the global if it has no alignment 971 // specified or if it is not assigned a section. If it is assigned a 972 // section, the global could be densely packed with other objects in the 973 // section, increasing the alignment could cause padding issues. 974 if (!GO->hasSection() || GO->getAlignment() == 0) 975 GO->setAlignment(PrefAlign); 976 return GO->getAlignment(); 977 } 978 979 return Align; 980 } 981 982 /// getOrEnforceKnownAlignment - If the specified pointer has an alignment that 983 /// we can determine, return it, otherwise return 0. If PrefAlign is specified, 984 /// and it is more than the alignment of the ultimate object, see if we can 985 /// increase the alignment of the ultimate object, making this check succeed. 986 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign, 987 const DataLayout &DL, 988 const Instruction *CxtI, 989 AssumptionCache *AC, 990 const DominatorTree *DT) { 991 assert(V->getType()->isPointerTy() && 992 "getOrEnforceKnownAlignment expects a pointer!"); 993 unsigned BitWidth = DL.getPointerTypeSizeInBits(V->getType()); 994 995 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 996 computeKnownBits(V, KnownZero, KnownOne, DL, 0, AC, CxtI, DT); 997 unsigned TrailZ = KnownZero.countTrailingOnes(); 998 999 // Avoid trouble with ridiculously large TrailZ values, such as 1000 // those computed from a null pointer. 1001 TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1)); 1002 1003 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ); 1004 1005 // LLVM doesn't support alignments larger than this currently. 1006 Align = std::min(Align, +Value::MaximumAlignment); 1007 1008 if (PrefAlign > Align) 1009 Align = enforceKnownAlignment(V, Align, PrefAlign, DL); 1010 1011 // We don't need to make any adjustment. 1012 return Align; 1013 } 1014 1015 ///===---------------------------------------------------------------------===// 1016 /// Dbg Intrinsic utilities 1017 /// 1018 1019 /// See if there is a dbg.value intrinsic for DIVar before I. 1020 static bool LdStHasDebugValue(const DILocalVariable *DIVar, Instruction *I) { 1021 // Since we can't guarantee that the original dbg.declare instrinsic 1022 // is removed by LowerDbgDeclare(), we need to make sure that we are 1023 // not inserting the same dbg.value intrinsic over and over. 1024 llvm::BasicBlock::InstListType::iterator PrevI(I); 1025 if (PrevI != I->getParent()->getInstList().begin()) { 1026 --PrevI; 1027 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI)) 1028 if (DVI->getValue() == I->getOperand(0) && 1029 DVI->getOffset() == 0 && 1030 DVI->getVariable() == DIVar) 1031 return true; 1032 } 1033 return false; 1034 } 1035 1036 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value 1037 /// that has an associated llvm.dbg.decl intrinsic. 1038 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, 1039 StoreInst *SI, DIBuilder &Builder) { 1040 auto *DIVar = DDI->getVariable(); 1041 auto *DIExpr = DDI->getExpression(); 1042 assert(DIVar && "Missing variable"); 1043 1044 if (LdStHasDebugValue(DIVar, SI)) 1045 return true; 1046 1047 // If an argument is zero extended then use argument directly. The ZExt 1048 // may be zapped by an optimization pass in future. 1049 Argument *ExtendedArg = nullptr; 1050 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 1051 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0)); 1052 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 1053 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0)); 1054 if (ExtendedArg) 1055 Builder.insertDbgValueIntrinsic(ExtendedArg, 0, DIVar, DIExpr, 1056 DDI->getDebugLoc(), SI); 1057 else 1058 Builder.insertDbgValueIntrinsic(SI->getOperand(0), 0, DIVar, DIExpr, 1059 DDI->getDebugLoc(), SI); 1060 return true; 1061 } 1062 1063 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value 1064 /// that has an associated llvm.dbg.decl intrinsic. 1065 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, 1066 LoadInst *LI, DIBuilder &Builder) { 1067 auto *DIVar = DDI->getVariable(); 1068 auto *DIExpr = DDI->getExpression(); 1069 assert(DIVar && "Missing variable"); 1070 1071 if (LdStHasDebugValue(DIVar, LI)) 1072 return true; 1073 1074 // We are now tracking the loaded value instead of the address. In the 1075 // future if multi-location support is added to the IR, it might be 1076 // preferable to keep tracking both the loaded value and the original 1077 // address in case the alloca can not be elided. 1078 Instruction *DbgValue = Builder.insertDbgValueIntrinsic( 1079 LI, 0, DIVar, DIExpr, DDI->getDebugLoc(), (Instruction *)nullptr); 1080 DbgValue->insertAfter(LI); 1081 return true; 1082 } 1083 1084 /// Determine whether this alloca is either a VLA or an array. 1085 static bool isArray(AllocaInst *AI) { 1086 return AI->isArrayAllocation() || 1087 AI->getType()->getElementType()->isArrayTy(); 1088 } 1089 1090 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set 1091 /// of llvm.dbg.value intrinsics. 1092 bool llvm::LowerDbgDeclare(Function &F) { 1093 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false); 1094 SmallVector<DbgDeclareInst *, 4> Dbgs; 1095 for (auto &FI : F) 1096 for (Instruction &BI : FI) 1097 if (auto DDI = dyn_cast<DbgDeclareInst>(&BI)) 1098 Dbgs.push_back(DDI); 1099 1100 if (Dbgs.empty()) 1101 return false; 1102 1103 for (auto &I : Dbgs) { 1104 DbgDeclareInst *DDI = I; 1105 AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress()); 1106 // If this is an alloca for a scalar variable, insert a dbg.value 1107 // at each load and store to the alloca and erase the dbg.declare. 1108 // The dbg.values allow tracking a variable even if it is not 1109 // stored on the stack, while the dbg.declare can only describe 1110 // the stack slot (and at a lexical-scope granularity). Later 1111 // passes will attempt to elide the stack slot. 1112 if (AI && !isArray(AI)) { 1113 for (User *U : AI->users()) 1114 if (StoreInst *SI = dyn_cast<StoreInst>(U)) 1115 ConvertDebugDeclareToDebugValue(DDI, SI, DIB); 1116 else if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1117 ConvertDebugDeclareToDebugValue(DDI, LI, DIB); 1118 else if (CallInst *CI = dyn_cast<CallInst>(U)) { 1119 // This is a call by-value or some other instruction that 1120 // takes a pointer to the variable. Insert a *value* 1121 // intrinsic that describes the alloca. 1122 SmallVector<uint64_t, 1> NewDIExpr; 1123 auto *DIExpr = DDI->getExpression(); 1124 NewDIExpr.push_back(dwarf::DW_OP_deref); 1125 NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end()); 1126 DIB.insertDbgValueIntrinsic(AI, 0, DDI->getVariable(), 1127 DIB.createExpression(NewDIExpr), 1128 DDI->getDebugLoc(), CI); 1129 } 1130 DDI->eraseFromParent(); 1131 } 1132 } 1133 return true; 1134 } 1135 1136 /// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the 1137 /// alloca 'V', if any. 1138 DbgDeclareInst *llvm::FindAllocaDbgDeclare(Value *V) { 1139 if (auto *L = LocalAsMetadata::getIfExists(V)) 1140 if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L)) 1141 for (User *U : MDV->users()) 1142 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U)) 1143 return DDI; 1144 1145 return nullptr; 1146 } 1147 1148 bool llvm::replaceDbgDeclare(Value *Address, Value *NewAddress, 1149 Instruction *InsertBefore, DIBuilder &Builder, 1150 bool Deref, int Offset) { 1151 DbgDeclareInst *DDI = FindAllocaDbgDeclare(Address); 1152 if (!DDI) 1153 return false; 1154 DebugLoc Loc = DDI->getDebugLoc(); 1155 auto *DIVar = DDI->getVariable(); 1156 auto *DIExpr = DDI->getExpression(); 1157 assert(DIVar && "Missing variable"); 1158 1159 if (Deref || Offset) { 1160 // Create a copy of the original DIDescriptor for user variable, prepending 1161 // "deref" operation to a list of address elements, as new llvm.dbg.declare 1162 // will take a value storing address of the memory for variable, not 1163 // alloca itself. 1164 SmallVector<uint64_t, 4> NewDIExpr; 1165 if (Deref) 1166 NewDIExpr.push_back(dwarf::DW_OP_deref); 1167 if (Offset > 0) { 1168 NewDIExpr.push_back(dwarf::DW_OP_plus); 1169 NewDIExpr.push_back(Offset); 1170 } else if (Offset < 0) { 1171 NewDIExpr.push_back(dwarf::DW_OP_minus); 1172 NewDIExpr.push_back(-Offset); 1173 } 1174 if (DIExpr) 1175 NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end()); 1176 DIExpr = Builder.createExpression(NewDIExpr); 1177 } 1178 1179 // Insert llvm.dbg.declare immediately after the original alloca, and remove 1180 // old llvm.dbg.declare. 1181 Builder.insertDeclare(NewAddress, DIVar, DIExpr, Loc, InsertBefore); 1182 DDI->eraseFromParent(); 1183 return true; 1184 } 1185 1186 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress, 1187 DIBuilder &Builder, bool Deref, int Offset) { 1188 return replaceDbgDeclare(AI, NewAllocaAddress, AI->getNextNode(), Builder, 1189 Deref, Offset); 1190 } 1191 1192 void llvm::changeToUnreachable(Instruction *I, bool UseLLVMTrap) { 1193 BasicBlock *BB = I->getParent(); 1194 // Loop over all of the successors, removing BB's entry from any PHI 1195 // nodes. 1196 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 1197 (*SI)->removePredecessor(BB); 1198 1199 // Insert a call to llvm.trap right before this. This turns the undefined 1200 // behavior into a hard fail instead of falling through into random code. 1201 if (UseLLVMTrap) { 1202 Function *TrapFn = 1203 Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap); 1204 CallInst *CallTrap = CallInst::Create(TrapFn, "", I); 1205 CallTrap->setDebugLoc(I->getDebugLoc()); 1206 } 1207 new UnreachableInst(I->getContext(), I); 1208 1209 // All instructions after this are dead. 1210 BasicBlock::iterator BBI = I->getIterator(), BBE = BB->end(); 1211 while (BBI != BBE) { 1212 if (!BBI->use_empty()) 1213 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 1214 BB->getInstList().erase(BBI++); 1215 } 1216 } 1217 1218 /// changeToCall - Convert the specified invoke into a normal call. 1219 static void changeToCall(InvokeInst *II) { 1220 SmallVector<Value*, 8> Args(II->arg_begin(), II->arg_end()); 1221 SmallVector<OperandBundleDef, 1> OpBundles; 1222 II->getOperandBundlesAsDefs(OpBundles); 1223 CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, OpBundles, 1224 "", II); 1225 NewCall->takeName(II); 1226 NewCall->setCallingConv(II->getCallingConv()); 1227 NewCall->setAttributes(II->getAttributes()); 1228 NewCall->setDebugLoc(II->getDebugLoc()); 1229 II->replaceAllUsesWith(NewCall); 1230 1231 // Follow the call by a branch to the normal destination. 1232 BranchInst::Create(II->getNormalDest(), II); 1233 1234 // Update PHI nodes in the unwind destination 1235 II->getUnwindDest()->removePredecessor(II->getParent()); 1236 II->eraseFromParent(); 1237 } 1238 1239 static bool markAliveBlocks(Function &F, 1240 SmallPtrSetImpl<BasicBlock*> &Reachable) { 1241 1242 SmallVector<BasicBlock*, 128> Worklist; 1243 BasicBlock *BB = &F.front(); 1244 Worklist.push_back(BB); 1245 Reachable.insert(BB); 1246 bool Changed = false; 1247 do { 1248 BB = Worklist.pop_back_val(); 1249 1250 // Do a quick scan of the basic block, turning any obviously unreachable 1251 // instructions into LLVM unreachable insts. The instruction combining pass 1252 // canonicalizes unreachable insts into stores to null or undef. 1253 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E;++BBI){ 1254 // Assumptions that are known to be false are equivalent to unreachable. 1255 // Also, if the condition is undefined, then we make the choice most 1256 // beneficial to the optimizer, and choose that to also be unreachable. 1257 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI)) 1258 if (II->getIntrinsicID() == Intrinsic::assume) { 1259 bool MakeUnreachable = false; 1260 if (isa<UndefValue>(II->getArgOperand(0))) 1261 MakeUnreachable = true; 1262 else if (ConstantInt *Cond = 1263 dyn_cast<ConstantInt>(II->getArgOperand(0))) 1264 MakeUnreachable = Cond->isZero(); 1265 1266 if (MakeUnreachable) { 1267 // Don't insert a call to llvm.trap right before the unreachable. 1268 changeToUnreachable(&*BBI, false); 1269 Changed = true; 1270 break; 1271 } 1272 } 1273 1274 if (CallInst *CI = dyn_cast<CallInst>(BBI)) { 1275 if (CI->doesNotReturn()) { 1276 // If we found a call to a no-return function, insert an unreachable 1277 // instruction after it. Make sure there isn't *already* one there 1278 // though. 1279 ++BBI; 1280 if (!isa<UnreachableInst>(BBI)) { 1281 // Don't insert a call to llvm.trap right before the unreachable. 1282 changeToUnreachable(&*BBI, false); 1283 Changed = true; 1284 } 1285 break; 1286 } 1287 } 1288 1289 // Store to undef and store to null are undefined and used to signal that 1290 // they should be changed to unreachable by passes that can't modify the 1291 // CFG. 1292 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 1293 // Don't touch volatile stores. 1294 if (SI->isVolatile()) continue; 1295 1296 Value *Ptr = SI->getOperand(1); 1297 1298 if (isa<UndefValue>(Ptr) || 1299 (isa<ConstantPointerNull>(Ptr) && 1300 SI->getPointerAddressSpace() == 0)) { 1301 changeToUnreachable(SI, true); 1302 Changed = true; 1303 break; 1304 } 1305 } 1306 } 1307 1308 // Turn invokes that call 'nounwind' functions into ordinary calls. 1309 if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator())) { 1310 Value *Callee = II->getCalledValue(); 1311 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 1312 changeToUnreachable(II, true); 1313 Changed = true; 1314 } else if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(&F)) { 1315 if (II->use_empty() && II->onlyReadsMemory()) { 1316 // jump to the normal destination branch. 1317 BranchInst::Create(II->getNormalDest(), II); 1318 II->getUnwindDest()->removePredecessor(II->getParent()); 1319 II->eraseFromParent(); 1320 } else 1321 changeToCall(II); 1322 Changed = true; 1323 } 1324 } 1325 1326 Changed |= ConstantFoldTerminator(BB, true); 1327 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) 1328 if (Reachable.insert(*SI).second) 1329 Worklist.push_back(*SI); 1330 } while (!Worklist.empty()); 1331 return Changed; 1332 } 1333 1334 void llvm::removeUnwindEdge(BasicBlock *BB) { 1335 TerminatorInst *TI = BB->getTerminator(); 1336 1337 if (auto *II = dyn_cast<InvokeInst>(TI)) { 1338 changeToCall(II); 1339 return; 1340 } 1341 1342 TerminatorInst *NewTI; 1343 BasicBlock *UnwindDest; 1344 1345 if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) { 1346 NewTI = CleanupReturnInst::Create(CRI->getCleanupPad(), nullptr, CRI); 1347 UnwindDest = CRI->getUnwindDest(); 1348 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(TI)) { 1349 auto *NewCatchSwitch = CatchSwitchInst::Create( 1350 CatchSwitch->getParentPad(), nullptr, CatchSwitch->getNumHandlers(), 1351 CatchSwitch->getName(), CatchSwitch); 1352 for (BasicBlock *PadBB : CatchSwitch->handlers()) 1353 NewCatchSwitch->addHandler(PadBB); 1354 1355 NewTI = NewCatchSwitch; 1356 UnwindDest = CatchSwitch->getUnwindDest(); 1357 } else { 1358 llvm_unreachable("Could not find unwind successor"); 1359 } 1360 1361 NewTI->takeName(TI); 1362 NewTI->setDebugLoc(TI->getDebugLoc()); 1363 UnwindDest->removePredecessor(BB); 1364 TI->replaceAllUsesWith(NewTI); 1365 TI->eraseFromParent(); 1366 } 1367 1368 /// removeUnreachableBlocksFromFn - Remove blocks that are not reachable, even 1369 /// if they are in a dead cycle. Return true if a change was made, false 1370 /// otherwise. 1371 bool llvm::removeUnreachableBlocks(Function &F) { 1372 SmallPtrSet<BasicBlock*, 128> Reachable; 1373 bool Changed = markAliveBlocks(F, Reachable); 1374 1375 // If there are unreachable blocks in the CFG... 1376 if (Reachable.size() == F.size()) 1377 return Changed; 1378 1379 assert(Reachable.size() < F.size()); 1380 NumRemoved += F.size()-Reachable.size(); 1381 1382 // Loop over all of the basic blocks that are not reachable, dropping all of 1383 // their internal references... 1384 for (Function::iterator BB = ++F.begin(), E = F.end(); BB != E; ++BB) { 1385 if (Reachable.count(&*BB)) 1386 continue; 1387 1388 for (succ_iterator SI = succ_begin(&*BB), SE = succ_end(&*BB); SI != SE; 1389 ++SI) 1390 if (Reachable.count(*SI)) 1391 (*SI)->removePredecessor(&*BB); 1392 BB->dropAllReferences(); 1393 } 1394 1395 for (Function::iterator I = ++F.begin(); I != F.end();) 1396 if (!Reachable.count(&*I)) 1397 I = F.getBasicBlockList().erase(I); 1398 else 1399 ++I; 1400 1401 return true; 1402 } 1403 1404 void llvm::combineMetadata(Instruction *K, const Instruction *J, 1405 ArrayRef<unsigned> KnownIDs) { 1406 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata; 1407 K->dropUnknownNonDebugMetadata(KnownIDs); 1408 K->getAllMetadataOtherThanDebugLoc(Metadata); 1409 for (unsigned i = 0, n = Metadata.size(); i < n; ++i) { 1410 unsigned Kind = Metadata[i].first; 1411 MDNode *JMD = J->getMetadata(Kind); 1412 MDNode *KMD = Metadata[i].second; 1413 1414 switch (Kind) { 1415 default: 1416 K->setMetadata(Kind, nullptr); // Remove unknown metadata 1417 break; 1418 case LLVMContext::MD_dbg: 1419 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg"); 1420 case LLVMContext::MD_tbaa: 1421 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 1422 break; 1423 case LLVMContext::MD_alias_scope: 1424 K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD)); 1425 break; 1426 case LLVMContext::MD_noalias: 1427 K->setMetadata(Kind, MDNode::intersect(JMD, KMD)); 1428 break; 1429 case LLVMContext::MD_range: 1430 K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD)); 1431 break; 1432 case LLVMContext::MD_fpmath: 1433 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 1434 break; 1435 case LLVMContext::MD_invariant_load: 1436 // Only set the !invariant.load if it is present in both instructions. 1437 K->setMetadata(Kind, JMD); 1438 break; 1439 case LLVMContext::MD_nonnull: 1440 // Only set the !nonnull if it is present in both instructions. 1441 K->setMetadata(Kind, JMD); 1442 break; 1443 case LLVMContext::MD_invariant_group: 1444 // Preserve !invariant.group in K. 1445 break; 1446 case LLVMContext::MD_align: 1447 K->setMetadata(Kind, 1448 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD)); 1449 break; 1450 case LLVMContext::MD_dereferenceable: 1451 case LLVMContext::MD_dereferenceable_or_null: 1452 K->setMetadata(Kind, 1453 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD)); 1454 break; 1455 } 1456 } 1457 // Set !invariant.group from J if J has it. If both instructions have it 1458 // then we will just pick it from J - even when they are different. 1459 // Also make sure that K is load or store - f.e. combining bitcast with load 1460 // could produce bitcast with invariant.group metadata, which is invalid. 1461 // FIXME: we should try to preserve both invariant.group md if they are 1462 // different, but right now instruction can only have one invariant.group. 1463 if (auto *JMD = J->getMetadata(LLVMContext::MD_invariant_group)) 1464 if (isa<LoadInst>(K) || isa<StoreInst>(K)) 1465 K->setMetadata(LLVMContext::MD_invariant_group, JMD); 1466 } 1467 1468 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To, 1469 DominatorTree &DT, 1470 const BasicBlockEdge &Root) { 1471 assert(From->getType() == To->getType()); 1472 1473 unsigned Count = 0; 1474 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 1475 UI != UE; ) { 1476 Use &U = *UI++; 1477 if (DT.dominates(Root, U)) { 1478 U.set(To); 1479 DEBUG(dbgs() << "Replace dominated use of '" 1480 << From->getName() << "' as " 1481 << *To << " in " << *U << "\n"); 1482 ++Count; 1483 } 1484 } 1485 return Count; 1486 } 1487 1488 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To, 1489 DominatorTree &DT, 1490 const BasicBlock *BB) { 1491 assert(From->getType() == To->getType()); 1492 1493 unsigned Count = 0; 1494 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 1495 UI != UE;) { 1496 Use &U = *UI++; 1497 auto *I = cast<Instruction>(U.getUser()); 1498 if (DT.dominates(BB, I->getParent())) { 1499 U.set(To); 1500 DEBUG(dbgs() << "Replace dominated use of '" << From->getName() << "' as " 1501 << *To << " in " << *U << "\n"); 1502 ++Count; 1503 } 1504 } 1505 return Count; 1506 } 1507 1508 bool llvm::callsGCLeafFunction(ImmutableCallSite CS) { 1509 if (isa<IntrinsicInst>(CS.getInstruction())) 1510 // Most LLVM intrinsics are things which can never take a safepoint. 1511 // As a result, we don't need to have the stack parsable at the 1512 // callsite. This is a highly useful optimization since intrinsic 1513 // calls are fairly prevalent, particularly in debug builds. 1514 return true; 1515 1516 // Check if the function is specifically marked as a gc leaf function. 1517 // 1518 // TODO: we should be checking the attributes on the call site as well. 1519 if (const Function *F = CS.getCalledFunction()) 1520 return F->hasFnAttribute("gc-leaf-function"); 1521 1522 return false; 1523 } 1524