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/APInt.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/DenseMapInfo.h" 19 #include "llvm/ADT/DenseSet.h" 20 #include "llvm/ADT/Hashing.h" 21 #include "llvm/ADT/None.h" 22 #include "llvm/ADT/Optional.h" 23 #include "llvm/ADT/STLExtras.h" 24 #include "llvm/ADT/SetVector.h" 25 #include "llvm/ADT/SmallPtrSet.h" 26 #include "llvm/ADT/SmallVector.h" 27 #include "llvm/ADT/Statistic.h" 28 #include "llvm/ADT/TinyPtrVector.h" 29 #include "llvm/Analysis/ConstantFolding.h" 30 #include "llvm/Analysis/EHPersonalities.h" 31 #include "llvm/Analysis/InstructionSimplify.h" 32 #include "llvm/Analysis/LazyValueInfo.h" 33 #include "llvm/Analysis/MemoryBuiltins.h" 34 #include "llvm/Analysis/TargetLibraryInfo.h" 35 #include "llvm/Analysis/ValueTracking.h" 36 #include "llvm/BinaryFormat/Dwarf.h" 37 #include "llvm/IR/Argument.h" 38 #include "llvm/IR/Attributes.h" 39 #include "llvm/IR/BasicBlock.h" 40 #include "llvm/IR/CFG.h" 41 #include "llvm/IR/CallSite.h" 42 #include "llvm/IR/Constant.h" 43 #include "llvm/IR/ConstantRange.h" 44 #include "llvm/IR/Constants.h" 45 #include "llvm/IR/DIBuilder.h" 46 #include "llvm/IR/DataLayout.h" 47 #include "llvm/IR/DebugInfoMetadata.h" 48 #include "llvm/IR/DebugLoc.h" 49 #include "llvm/IR/DerivedTypes.h" 50 #include "llvm/IR/Dominators.h" 51 #include "llvm/IR/Function.h" 52 #include "llvm/IR/GetElementPtrTypeIterator.h" 53 #include "llvm/IR/GlobalObject.h" 54 #include "llvm/IR/IRBuilder.h" 55 #include "llvm/IR/InstrTypes.h" 56 #include "llvm/IR/Instruction.h" 57 #include "llvm/IR/Instructions.h" 58 #include "llvm/IR/IntrinsicInst.h" 59 #include "llvm/IR/Intrinsics.h" 60 #include "llvm/IR/LLVMContext.h" 61 #include "llvm/IR/MDBuilder.h" 62 #include "llvm/IR/Metadata.h" 63 #include "llvm/IR/Module.h" 64 #include "llvm/IR/Operator.h" 65 #include "llvm/IR/PatternMatch.h" 66 #include "llvm/IR/Type.h" 67 #include "llvm/IR/Use.h" 68 #include "llvm/IR/User.h" 69 #include "llvm/IR/Value.h" 70 #include "llvm/IR/ValueHandle.h" 71 #include "llvm/Support/Casting.h" 72 #include "llvm/Support/Debug.h" 73 #include "llvm/Support/ErrorHandling.h" 74 #include "llvm/Support/KnownBits.h" 75 #include "llvm/Support/raw_ostream.h" 76 #include <algorithm> 77 #include <cassert> 78 #include <climits> 79 #include <cstdint> 80 #include <iterator> 81 #include <map> 82 #include <utility> 83 84 using namespace llvm; 85 using namespace llvm::PatternMatch; 86 87 #define DEBUG_TYPE "local" 88 89 STATISTIC(NumRemoved, "Number of unreachable basic blocks removed"); 90 91 //===----------------------------------------------------------------------===// 92 // Local constant propagation. 93 // 94 95 /// ConstantFoldTerminator - If a terminator instruction is predicated on a 96 /// constant value, convert it into an unconditional branch to the constant 97 /// destination. This is a nontrivial operation because the successors of this 98 /// basic block must have their PHI nodes updated. 99 /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch 100 /// conditions and indirectbr addresses this might make dead if 101 /// DeleteDeadConditions is true. 102 bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions, 103 const TargetLibraryInfo *TLI) { 104 TerminatorInst *T = BB->getTerminator(); 105 IRBuilder<> Builder(T); 106 107 // Branch - See if we are conditional jumping on constant 108 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 109 if (BI->isUnconditional()) return false; // Can't optimize uncond branch 110 BasicBlock *Dest1 = BI->getSuccessor(0); 111 BasicBlock *Dest2 = BI->getSuccessor(1); 112 113 if (ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition())) { 114 // Are we branching on constant? 115 // YES. Change to unconditional branch... 116 BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2; 117 BasicBlock *OldDest = Cond->getZExtValue() ? Dest2 : Dest1; 118 119 //cerr << "Function: " << T->getParent()->getParent() 120 // << "\nRemoving branch from " << T->getParent() 121 // << "\n\nTo: " << OldDest << endl; 122 123 // Let the basic block know that we are letting go of it. Based on this, 124 // it will adjust it's PHI nodes. 125 OldDest->removePredecessor(BB); 126 127 // Replace the conditional branch with an unconditional one. 128 Builder.CreateBr(Destination); 129 BI->eraseFromParent(); 130 return true; 131 } 132 133 if (Dest2 == Dest1) { // Conditional branch to same location? 134 // This branch matches something like this: 135 // br bool %cond, label %Dest, label %Dest 136 // and changes it into: br label %Dest 137 138 // Let the basic block know that we are letting go of one copy of it. 139 assert(BI->getParent() && "Terminator not inserted in block!"); 140 Dest1->removePredecessor(BI->getParent()); 141 142 // Replace the conditional branch with an unconditional one. 143 Builder.CreateBr(Dest1); 144 Value *Cond = BI->getCondition(); 145 BI->eraseFromParent(); 146 if (DeleteDeadConditions) 147 RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI); 148 return true; 149 } 150 return false; 151 } 152 153 if (SwitchInst *SI = dyn_cast<SwitchInst>(T)) { 154 // If we are switching on a constant, we can convert the switch to an 155 // unconditional branch. 156 ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition()); 157 BasicBlock *DefaultDest = SI->getDefaultDest(); 158 BasicBlock *TheOnlyDest = DefaultDest; 159 160 // If the default is unreachable, ignore it when searching for TheOnlyDest. 161 if (isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg()) && 162 SI->getNumCases() > 0) { 163 TheOnlyDest = SI->case_begin()->getCaseSuccessor(); 164 } 165 166 // Figure out which case it goes to. 167 for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) { 168 // Found case matching a constant operand? 169 if (i->getCaseValue() == CI) { 170 TheOnlyDest = i->getCaseSuccessor(); 171 break; 172 } 173 174 // Check to see if this branch is going to the same place as the default 175 // dest. If so, eliminate it as an explicit compare. 176 if (i->getCaseSuccessor() == DefaultDest) { 177 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof); 178 unsigned NCases = SI->getNumCases(); 179 // Fold the case metadata into the default if there will be any branches 180 // left, unless the metadata doesn't match the switch. 181 if (NCases > 1 && MD && MD->getNumOperands() == 2 + NCases) { 182 // Collect branch weights into a vector. 183 SmallVector<uint32_t, 8> Weights; 184 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e; 185 ++MD_i) { 186 auto *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i)); 187 Weights.push_back(CI->getValue().getZExtValue()); 188 } 189 // Merge weight of this case to the default weight. 190 unsigned idx = i->getCaseIndex(); 191 Weights[0] += Weights[idx+1]; 192 // Remove weight for this case. 193 std::swap(Weights[idx+1], Weights.back()); 194 Weights.pop_back(); 195 SI->setMetadata(LLVMContext::MD_prof, 196 MDBuilder(BB->getContext()). 197 createBranchWeights(Weights)); 198 } 199 // Remove this entry. 200 DefaultDest->removePredecessor(SI->getParent()); 201 i = SI->removeCase(i); 202 e = SI->case_end(); 203 continue; 204 } 205 206 // Otherwise, check to see if the switch only branches to one destination. 207 // We do this by reseting "TheOnlyDest" to null when we find two non-equal 208 // destinations. 209 if (i->getCaseSuccessor() != TheOnlyDest) 210 TheOnlyDest = nullptr; 211 212 // Increment this iterator as we haven't removed the case. 213 ++i; 214 } 215 216 if (CI && !TheOnlyDest) { 217 // Branching on a constant, but not any of the cases, go to the default 218 // successor. 219 TheOnlyDest = SI->getDefaultDest(); 220 } 221 222 // If we found a single destination that we can fold the switch into, do so 223 // now. 224 if (TheOnlyDest) { 225 // Insert the new branch. 226 Builder.CreateBr(TheOnlyDest); 227 BasicBlock *BB = SI->getParent(); 228 229 // Remove entries from PHI nodes which we no longer branch to... 230 for (BasicBlock *Succ : SI->successors()) { 231 // Found case matching a constant operand? 232 if (Succ == TheOnlyDest) 233 TheOnlyDest = nullptr; // Don't modify the first branch to TheOnlyDest 234 else 235 Succ->removePredecessor(BB); 236 } 237 238 // Delete the old switch. 239 Value *Cond = SI->getCondition(); 240 SI->eraseFromParent(); 241 if (DeleteDeadConditions) 242 RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI); 243 return true; 244 } 245 246 if (SI->getNumCases() == 1) { 247 // Otherwise, we can fold this switch into a conditional branch 248 // instruction if it has only one non-default destination. 249 auto FirstCase = *SI->case_begin(); 250 Value *Cond = Builder.CreateICmpEQ(SI->getCondition(), 251 FirstCase.getCaseValue(), "cond"); 252 253 // Insert the new branch. 254 BranchInst *NewBr = Builder.CreateCondBr(Cond, 255 FirstCase.getCaseSuccessor(), 256 SI->getDefaultDest()); 257 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof); 258 if (MD && MD->getNumOperands() == 3) { 259 ConstantInt *SICase = 260 mdconst::dyn_extract<ConstantInt>(MD->getOperand(2)); 261 ConstantInt *SIDef = 262 mdconst::dyn_extract<ConstantInt>(MD->getOperand(1)); 263 assert(SICase && SIDef); 264 // The TrueWeight should be the weight for the single case of SI. 265 NewBr->setMetadata(LLVMContext::MD_prof, 266 MDBuilder(BB->getContext()). 267 createBranchWeights(SICase->getValue().getZExtValue(), 268 SIDef->getValue().getZExtValue())); 269 } 270 271 // Update make.implicit metadata to the newly-created conditional branch. 272 MDNode *MakeImplicitMD = SI->getMetadata(LLVMContext::MD_make_implicit); 273 if (MakeImplicitMD) 274 NewBr->setMetadata(LLVMContext::MD_make_implicit, MakeImplicitMD); 275 276 // Delete the old switch. 277 SI->eraseFromParent(); 278 return true; 279 } 280 return false; 281 } 282 283 if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(T)) { 284 // indirectbr blockaddress(@F, @BB) -> br label @BB 285 if (BlockAddress *BA = 286 dyn_cast<BlockAddress>(IBI->getAddress()->stripPointerCasts())) { 287 BasicBlock *TheOnlyDest = BA->getBasicBlock(); 288 // Insert the new branch. 289 Builder.CreateBr(TheOnlyDest); 290 291 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 292 if (IBI->getDestination(i) == TheOnlyDest) 293 TheOnlyDest = nullptr; 294 else 295 IBI->getDestination(i)->removePredecessor(IBI->getParent()); 296 } 297 Value *Address = IBI->getAddress(); 298 IBI->eraseFromParent(); 299 if (DeleteDeadConditions) 300 RecursivelyDeleteTriviallyDeadInstructions(Address, TLI); 301 302 // If we didn't find our destination in the IBI successor list, then we 303 // have undefined behavior. Replace the unconditional branch with an 304 // 'unreachable' instruction. 305 if (TheOnlyDest) { 306 BB->getTerminator()->eraseFromParent(); 307 new UnreachableInst(BB->getContext(), BB); 308 } 309 310 return true; 311 } 312 } 313 314 return false; 315 } 316 317 //===----------------------------------------------------------------------===// 318 // Local dead code elimination. 319 // 320 321 /// isInstructionTriviallyDead - Return true if the result produced by the 322 /// instruction is not used, and the instruction has no side effects. 323 /// 324 bool llvm::isInstructionTriviallyDead(Instruction *I, 325 const TargetLibraryInfo *TLI) { 326 if (!I->use_empty()) 327 return false; 328 return wouldInstructionBeTriviallyDead(I, TLI); 329 } 330 331 bool llvm::wouldInstructionBeTriviallyDead(Instruction *I, 332 const TargetLibraryInfo *TLI) { 333 if (isa<TerminatorInst>(I)) 334 return false; 335 336 // We don't want the landingpad-like instructions removed by anything this 337 // general. 338 if (I->isEHPad()) 339 return false; 340 341 // We don't want debug info removed by anything this general, unless 342 // debug info is empty. 343 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(I)) { 344 if (DDI->getAddress()) 345 return false; 346 return true; 347 } 348 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(I)) { 349 if (DVI->getValue()) 350 return false; 351 return true; 352 } 353 354 if (!I->mayHaveSideEffects()) 355 return true; 356 357 // Special case intrinsics that "may have side effects" but can be deleted 358 // when dead. 359 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 360 // Safe to delete llvm.stacksave if dead. 361 if (II->getIntrinsicID() == Intrinsic::stacksave) 362 return true; 363 364 // Lifetime intrinsics are dead when their right-hand is undef. 365 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 366 II->getIntrinsicID() == Intrinsic::lifetime_end) 367 return isa<UndefValue>(II->getArgOperand(1)); 368 369 // Assumptions are dead if their condition is trivially true. Guards on 370 // true are operationally no-ops. In the future we can consider more 371 // sophisticated tradeoffs for guards considering potential for check 372 // widening, but for now we keep things simple. 373 if (II->getIntrinsicID() == Intrinsic::assume || 374 II->getIntrinsicID() == Intrinsic::experimental_guard) { 375 if (ConstantInt *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0))) 376 return !Cond->isZero(); 377 378 return false; 379 } 380 } 381 382 if (isAllocLikeFn(I, TLI)) 383 return true; 384 385 if (CallInst *CI = isFreeCall(I, TLI)) 386 if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0))) 387 return C->isNullValue() || isa<UndefValue>(C); 388 389 if (CallSite CS = CallSite(I)) 390 if (isMathLibCallNoop(CS, TLI)) 391 return true; 392 393 return false; 394 } 395 396 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a 397 /// trivially dead instruction, delete it. If that makes any of its operands 398 /// trivially dead, delete them too, recursively. Return true if any 399 /// instructions were deleted. 400 bool 401 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V, 402 const TargetLibraryInfo *TLI) { 403 Instruction *I = dyn_cast<Instruction>(V); 404 if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI)) 405 return false; 406 407 SmallVector<Instruction*, 16> DeadInsts; 408 DeadInsts.push_back(I); 409 410 do { 411 I = DeadInsts.pop_back_val(); 412 413 // Null out all of the instruction's operands to see if any operand becomes 414 // dead as we go. 415 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 416 Value *OpV = I->getOperand(i); 417 I->setOperand(i, nullptr); 418 419 if (!OpV->use_empty()) continue; 420 421 // If the operand is an instruction that became dead as we nulled out the 422 // operand, and if it is 'trivially' dead, delete it in a future loop 423 // iteration. 424 if (Instruction *OpI = dyn_cast<Instruction>(OpV)) 425 if (isInstructionTriviallyDead(OpI, TLI)) 426 DeadInsts.push_back(OpI); 427 } 428 429 I->eraseFromParent(); 430 } while (!DeadInsts.empty()); 431 432 return true; 433 } 434 435 /// areAllUsesEqual - Check whether the uses of a value are all the same. 436 /// This is similar to Instruction::hasOneUse() except this will also return 437 /// true when there are no uses or multiple uses that all refer to the same 438 /// value. 439 static bool areAllUsesEqual(Instruction *I) { 440 Value::user_iterator UI = I->user_begin(); 441 Value::user_iterator UE = I->user_end(); 442 if (UI == UE) 443 return true; 444 445 User *TheUse = *UI; 446 for (++UI; UI != UE; ++UI) { 447 if (*UI != TheUse) 448 return false; 449 } 450 return true; 451 } 452 453 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively 454 /// dead PHI node, due to being a def-use chain of single-use nodes that 455 /// either forms a cycle or is terminated by a trivially dead instruction, 456 /// delete it. If that makes any of its operands trivially dead, delete them 457 /// too, recursively. Return true if a change was made. 458 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN, 459 const TargetLibraryInfo *TLI) { 460 SmallPtrSet<Instruction*, 4> Visited; 461 for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects(); 462 I = cast<Instruction>(*I->user_begin())) { 463 if (I->use_empty()) 464 return RecursivelyDeleteTriviallyDeadInstructions(I, TLI); 465 466 // If we find an instruction more than once, we're on a cycle that 467 // won't prove fruitful. 468 if (!Visited.insert(I).second) { 469 // Break the cycle and delete the instruction and its operands. 470 I->replaceAllUsesWith(UndefValue::get(I->getType())); 471 (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI); 472 return true; 473 } 474 } 475 return false; 476 } 477 478 static bool 479 simplifyAndDCEInstruction(Instruction *I, 480 SmallSetVector<Instruction *, 16> &WorkList, 481 const DataLayout &DL, 482 const TargetLibraryInfo *TLI) { 483 if (isInstructionTriviallyDead(I, TLI)) { 484 // Null out all of the instruction's operands to see if any operand becomes 485 // dead as we go. 486 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 487 Value *OpV = I->getOperand(i); 488 I->setOperand(i, nullptr); 489 490 if (!OpV->use_empty() || I == OpV) 491 continue; 492 493 // If the operand is an instruction that became dead as we nulled out the 494 // operand, and if it is 'trivially' dead, delete it in a future loop 495 // iteration. 496 if (Instruction *OpI = dyn_cast<Instruction>(OpV)) 497 if (isInstructionTriviallyDead(OpI, TLI)) 498 WorkList.insert(OpI); 499 } 500 501 I->eraseFromParent(); 502 503 return true; 504 } 505 506 if (Value *SimpleV = SimplifyInstruction(I, DL)) { 507 // Add the users to the worklist. CAREFUL: an instruction can use itself, 508 // in the case of a phi node. 509 for (User *U : I->users()) { 510 if (U != I) { 511 WorkList.insert(cast<Instruction>(U)); 512 } 513 } 514 515 // Replace the instruction with its simplified value. 516 bool Changed = false; 517 if (!I->use_empty()) { 518 I->replaceAllUsesWith(SimpleV); 519 Changed = true; 520 } 521 if (isInstructionTriviallyDead(I, TLI)) { 522 I->eraseFromParent(); 523 Changed = true; 524 } 525 return Changed; 526 } 527 return false; 528 } 529 530 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to 531 /// simplify any instructions in it and recursively delete dead instructions. 532 /// 533 /// This returns true if it changed the code, note that it can delete 534 /// instructions in other blocks as well in this block. 535 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB, 536 const TargetLibraryInfo *TLI) { 537 bool MadeChange = false; 538 const DataLayout &DL = BB->getModule()->getDataLayout(); 539 540 #ifndef NDEBUG 541 // In debug builds, ensure that the terminator of the block is never replaced 542 // or deleted by these simplifications. The idea of simplification is that it 543 // cannot introduce new instructions, and there is no way to replace the 544 // terminator of a block without introducing a new instruction. 545 AssertingVH<Instruction> TerminatorVH(&BB->back()); 546 #endif 547 548 SmallSetVector<Instruction *, 16> WorkList; 549 // Iterate over the original function, only adding insts to the worklist 550 // if they actually need to be revisited. This avoids having to pre-init 551 // the worklist with the entire function's worth of instructions. 552 for (BasicBlock::iterator BI = BB->begin(), E = std::prev(BB->end()); 553 BI != E;) { 554 assert(!BI->isTerminator()); 555 Instruction *I = &*BI; 556 ++BI; 557 558 // We're visiting this instruction now, so make sure it's not in the 559 // worklist from an earlier visit. 560 if (!WorkList.count(I)) 561 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI); 562 } 563 564 while (!WorkList.empty()) { 565 Instruction *I = WorkList.pop_back_val(); 566 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI); 567 } 568 return MadeChange; 569 } 570 571 //===----------------------------------------------------------------------===// 572 // Control Flow Graph Restructuring. 573 // 574 575 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this 576 /// method is called when we're about to delete Pred as a predecessor of BB. If 577 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred. 578 /// 579 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI 580 /// nodes that collapse into identity values. For example, if we have: 581 /// x = phi(1, 0, 0, 0) 582 /// y = and x, z 583 /// 584 /// .. and delete the predecessor corresponding to the '1', this will attempt to 585 /// recursively fold the and to 0. 586 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred) { 587 // This only adjusts blocks with PHI nodes. 588 if (!isa<PHINode>(BB->begin())) 589 return; 590 591 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify 592 // them down. This will leave us with single entry phi nodes and other phis 593 // that can be removed. 594 BB->removePredecessor(Pred, true); 595 596 WeakTrackingVH PhiIt = &BB->front(); 597 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) { 598 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt)); 599 Value *OldPhiIt = PhiIt; 600 601 if (!recursivelySimplifyInstruction(PN)) 602 continue; 603 604 // If recursive simplification ended up deleting the next PHI node we would 605 // iterate to, then our iterator is invalid, restart scanning from the top 606 // of the block. 607 if (PhiIt != OldPhiIt) PhiIt = &BB->front(); 608 } 609 } 610 611 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its 612 /// predecessor is known to have one successor (DestBB!). Eliminate the edge 613 /// between them, moving the instructions in the predecessor into DestBB and 614 /// deleting the predecessor block. 615 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, DominatorTree *DT) { 616 // If BB has single-entry PHI nodes, fold them. 617 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 618 Value *NewVal = PN->getIncomingValue(0); 619 // Replace self referencing PHI with undef, it must be dead. 620 if (NewVal == PN) NewVal = UndefValue::get(PN->getType()); 621 PN->replaceAllUsesWith(NewVal); 622 PN->eraseFromParent(); 623 } 624 625 BasicBlock *PredBB = DestBB->getSinglePredecessor(); 626 assert(PredBB && "Block doesn't have a single predecessor!"); 627 628 // Zap anything that took the address of DestBB. Not doing this will give the 629 // address an invalid value. 630 if (DestBB->hasAddressTaken()) { 631 BlockAddress *BA = BlockAddress::get(DestBB); 632 Constant *Replacement = 633 ConstantInt::get(Type::getInt32Ty(BA->getContext()), 1); 634 BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement, 635 BA->getType())); 636 BA->destroyConstant(); 637 } 638 639 // Anything that branched to PredBB now branches to DestBB. 640 PredBB->replaceAllUsesWith(DestBB); 641 642 // Splice all the instructions from PredBB to DestBB. 643 PredBB->getTerminator()->eraseFromParent(); 644 DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList()); 645 646 // If the PredBB is the entry block of the function, move DestBB up to 647 // become the entry block after we erase PredBB. 648 if (PredBB == &DestBB->getParent()->getEntryBlock()) 649 DestBB->moveAfter(PredBB); 650 651 if (DT) { 652 // For some irreducible CFG we end up having forward-unreachable blocks 653 // so check if getNode returns a valid node before updating the domtree. 654 if (DomTreeNode *DTN = DT->getNode(PredBB)) { 655 BasicBlock *PredBBIDom = DTN->getIDom()->getBlock(); 656 DT->changeImmediateDominator(DestBB, PredBBIDom); 657 DT->eraseNode(PredBB); 658 } 659 } 660 // Nuke BB. 661 PredBB->eraseFromParent(); 662 } 663 664 /// CanMergeValues - Return true if we can choose one of these values to use 665 /// in place of the other. Note that we will always choose the non-undef 666 /// value to keep. 667 static bool CanMergeValues(Value *First, Value *Second) { 668 return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second); 669 } 670 671 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an 672 /// almost-empty BB ending in an unconditional branch to Succ, into Succ. 673 /// 674 /// Assumption: Succ is the single successor for BB. 675 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { 676 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!"); 677 678 DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into " 679 << Succ->getName() << "\n"); 680 // Shortcut, if there is only a single predecessor it must be BB and merging 681 // is always safe 682 if (Succ->getSinglePredecessor()) return true; 683 684 // Make a list of the predecessors of BB 685 SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB)); 686 687 // Look at all the phi nodes in Succ, to see if they present a conflict when 688 // merging these blocks 689 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 690 PHINode *PN = cast<PHINode>(I); 691 692 // If the incoming value from BB is again a PHINode in 693 // BB which has the same incoming value for *PI as PN does, we can 694 // merge the phi nodes and then the blocks can still be merged 695 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB)); 696 if (BBPN && BBPN->getParent() == BB) { 697 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) { 698 BasicBlock *IBB = PN->getIncomingBlock(PI); 699 if (BBPreds.count(IBB) && 700 !CanMergeValues(BBPN->getIncomingValueForBlock(IBB), 701 PN->getIncomingValue(PI))) { 702 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in " 703 << Succ->getName() << " is conflicting with " 704 << BBPN->getName() << " with regard to common predecessor " 705 << IBB->getName() << "\n"); 706 return false; 707 } 708 } 709 } else { 710 Value* Val = PN->getIncomingValueForBlock(BB); 711 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) { 712 // See if the incoming value for the common predecessor is equal to the 713 // one for BB, in which case this phi node will not prevent the merging 714 // of the block. 715 BasicBlock *IBB = PN->getIncomingBlock(PI); 716 if (BBPreds.count(IBB) && 717 !CanMergeValues(Val, PN->getIncomingValue(PI))) { 718 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in " 719 << Succ->getName() << " is conflicting with regard to common " 720 << "predecessor " << IBB->getName() << "\n"); 721 return false; 722 } 723 } 724 } 725 } 726 727 return true; 728 } 729 730 using PredBlockVector = SmallVector<BasicBlock *, 16>; 731 using IncomingValueMap = DenseMap<BasicBlock *, Value *>; 732 733 /// \brief Determines the value to use as the phi node input for a block. 734 /// 735 /// Select between \p OldVal any value that we know flows from \p BB 736 /// to a particular phi on the basis of which one (if either) is not 737 /// undef. Update IncomingValues based on the selected value. 738 /// 739 /// \param OldVal The value we are considering selecting. 740 /// \param BB The block that the value flows in from. 741 /// \param IncomingValues A map from block-to-value for other phi inputs 742 /// that we have examined. 743 /// 744 /// \returns the selected value. 745 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB, 746 IncomingValueMap &IncomingValues) { 747 if (!isa<UndefValue>(OldVal)) { 748 assert((!IncomingValues.count(BB) || 749 IncomingValues.find(BB)->second == OldVal) && 750 "Expected OldVal to match incoming value from BB!"); 751 752 IncomingValues.insert(std::make_pair(BB, OldVal)); 753 return OldVal; 754 } 755 756 IncomingValueMap::const_iterator It = IncomingValues.find(BB); 757 if (It != IncomingValues.end()) return It->second; 758 759 return OldVal; 760 } 761 762 /// \brief Create a map from block to value for the operands of a 763 /// given phi. 764 /// 765 /// Create a map from block to value for each non-undef value flowing 766 /// into \p PN. 767 /// 768 /// \param PN The phi we are collecting the map for. 769 /// \param IncomingValues [out] The map from block to value for this phi. 770 static void gatherIncomingValuesToPhi(PHINode *PN, 771 IncomingValueMap &IncomingValues) { 772 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 773 BasicBlock *BB = PN->getIncomingBlock(i); 774 Value *V = PN->getIncomingValue(i); 775 776 if (!isa<UndefValue>(V)) 777 IncomingValues.insert(std::make_pair(BB, V)); 778 } 779 } 780 781 /// \brief Replace the incoming undef values to a phi with the values 782 /// from a block-to-value map. 783 /// 784 /// \param PN The phi we are replacing the undefs in. 785 /// \param IncomingValues A map from block to value. 786 static void replaceUndefValuesInPhi(PHINode *PN, 787 const IncomingValueMap &IncomingValues) { 788 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 789 Value *V = PN->getIncomingValue(i); 790 791 if (!isa<UndefValue>(V)) continue; 792 793 BasicBlock *BB = PN->getIncomingBlock(i); 794 IncomingValueMap::const_iterator It = IncomingValues.find(BB); 795 if (It == IncomingValues.end()) continue; 796 797 PN->setIncomingValue(i, It->second); 798 } 799 } 800 801 /// \brief Replace a value flowing from a block to a phi with 802 /// potentially multiple instances of that value flowing from the 803 /// block's predecessors to the phi. 804 /// 805 /// \param BB The block with the value flowing into the phi. 806 /// \param BBPreds The predecessors of BB. 807 /// \param PN The phi that we are updating. 808 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB, 809 const PredBlockVector &BBPreds, 810 PHINode *PN) { 811 Value *OldVal = PN->removeIncomingValue(BB, false); 812 assert(OldVal && "No entry in PHI for Pred BB!"); 813 814 IncomingValueMap IncomingValues; 815 816 // We are merging two blocks - BB, and the block containing PN - and 817 // as a result we need to redirect edges from the predecessors of BB 818 // to go to the block containing PN, and update PN 819 // accordingly. Since we allow merging blocks in the case where the 820 // predecessor and successor blocks both share some predecessors, 821 // and where some of those common predecessors might have undef 822 // values flowing into PN, we want to rewrite those values to be 823 // consistent with the non-undef values. 824 825 gatherIncomingValuesToPhi(PN, IncomingValues); 826 827 // If this incoming value is one of the PHI nodes in BB, the new entries 828 // in the PHI node are the entries from the old PHI. 829 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) { 830 PHINode *OldValPN = cast<PHINode>(OldVal); 831 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) { 832 // Note that, since we are merging phi nodes and BB and Succ might 833 // have common predecessors, we could end up with a phi node with 834 // identical incoming branches. This will be cleaned up later (and 835 // will trigger asserts if we try to clean it up now, without also 836 // simplifying the corresponding conditional branch). 837 BasicBlock *PredBB = OldValPN->getIncomingBlock(i); 838 Value *PredVal = OldValPN->getIncomingValue(i); 839 Value *Selected = selectIncomingValueForBlock(PredVal, PredBB, 840 IncomingValues); 841 842 // And add a new incoming value for this predecessor for the 843 // newly retargeted branch. 844 PN->addIncoming(Selected, PredBB); 845 } 846 } else { 847 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) { 848 // Update existing incoming values in PN for this 849 // predecessor of BB. 850 BasicBlock *PredBB = BBPreds[i]; 851 Value *Selected = selectIncomingValueForBlock(OldVal, PredBB, 852 IncomingValues); 853 854 // And add a new incoming value for this predecessor for the 855 // newly retargeted branch. 856 PN->addIncoming(Selected, PredBB); 857 } 858 } 859 860 replaceUndefValuesInPhi(PN, IncomingValues); 861 } 862 863 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an 864 /// unconditional branch, and contains no instructions other than PHI nodes, 865 /// potential side-effect free intrinsics and the branch. If possible, 866 /// eliminate BB by rewriting all the predecessors to branch to the successor 867 /// block and return true. If we can't transform, return false. 868 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB) { 869 assert(BB != &BB->getParent()->getEntryBlock() && 870 "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!"); 871 872 // We can't eliminate infinite loops. 873 BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0); 874 if (BB == Succ) return false; 875 876 // Check to see if merging these blocks would cause conflicts for any of the 877 // phi nodes in BB or Succ. If not, we can safely merge. 878 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false; 879 880 // Check for cases where Succ has multiple predecessors and a PHI node in BB 881 // has uses which will not disappear when the PHI nodes are merged. It is 882 // possible to handle such cases, but difficult: it requires checking whether 883 // BB dominates Succ, which is non-trivial to calculate in the case where 884 // Succ has multiple predecessors. Also, it requires checking whether 885 // constructing the necessary self-referential PHI node doesn't introduce any 886 // conflicts; this isn't too difficult, but the previous code for doing this 887 // was incorrect. 888 // 889 // Note that if this check finds a live use, BB dominates Succ, so BB is 890 // something like a loop pre-header (or rarely, a part of an irreducible CFG); 891 // folding the branch isn't profitable in that case anyway. 892 if (!Succ->getSinglePredecessor()) { 893 BasicBlock::iterator BBI = BB->begin(); 894 while (isa<PHINode>(*BBI)) { 895 for (Use &U : BBI->uses()) { 896 if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) { 897 if (PN->getIncomingBlock(U) != BB) 898 return false; 899 } else { 900 return false; 901 } 902 } 903 ++BBI; 904 } 905 } 906 907 DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB); 908 909 if (isa<PHINode>(Succ->begin())) { 910 // If there is more than one pred of succ, and there are PHI nodes in 911 // the successor, then we need to add incoming edges for the PHI nodes 912 // 913 const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB)); 914 915 // Loop over all of the PHI nodes in the successor of BB. 916 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 917 PHINode *PN = cast<PHINode>(I); 918 919 redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN); 920 } 921 } 922 923 if (Succ->getSinglePredecessor()) { 924 // BB is the only predecessor of Succ, so Succ will end up with exactly 925 // the same predecessors BB had. 926 927 // Copy over any phi, debug or lifetime instruction. 928 BB->getTerminator()->eraseFromParent(); 929 Succ->getInstList().splice(Succ->getFirstNonPHI()->getIterator(), 930 BB->getInstList()); 931 } else { 932 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) { 933 // We explicitly check for such uses in CanPropagatePredecessorsForPHIs. 934 assert(PN->use_empty() && "There shouldn't be any uses here!"); 935 PN->eraseFromParent(); 936 } 937 } 938 939 // If the unconditional branch we replaced contains llvm.loop metadata, we 940 // add the metadata to the branch instructions in the predecessors. 941 unsigned LoopMDKind = BB->getContext().getMDKindID("llvm.loop"); 942 Instruction *TI = BB->getTerminator(); 943 if (TI) 944 if (MDNode *LoopMD = TI->getMetadata(LoopMDKind)) 945 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 946 BasicBlock *Pred = *PI; 947 Pred->getTerminator()->setMetadata(LoopMDKind, LoopMD); 948 } 949 950 // Everything that jumped to BB now goes to Succ. 951 BB->replaceAllUsesWith(Succ); 952 if (!Succ->hasName()) Succ->takeName(BB); 953 BB->eraseFromParent(); // Delete the old basic block. 954 return true; 955 } 956 957 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI 958 /// nodes in this block. This doesn't try to be clever about PHI nodes 959 /// which differ only in the order of the incoming values, but instcombine 960 /// orders them so it usually won't matter. 961 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) { 962 // This implementation doesn't currently consider undef operands 963 // specially. Theoretically, two phis which are identical except for 964 // one having an undef where the other doesn't could be collapsed. 965 966 struct PHIDenseMapInfo { 967 static PHINode *getEmptyKey() { 968 return DenseMapInfo<PHINode *>::getEmptyKey(); 969 } 970 971 static PHINode *getTombstoneKey() { 972 return DenseMapInfo<PHINode *>::getTombstoneKey(); 973 } 974 975 static unsigned getHashValue(PHINode *PN) { 976 // Compute a hash value on the operands. Instcombine will likely have 977 // sorted them, which helps expose duplicates, but we have to check all 978 // the operands to be safe in case instcombine hasn't run. 979 return static_cast<unsigned>(hash_combine( 980 hash_combine_range(PN->value_op_begin(), PN->value_op_end()), 981 hash_combine_range(PN->block_begin(), PN->block_end()))); 982 } 983 984 static bool isEqual(PHINode *LHS, PHINode *RHS) { 985 if (LHS == getEmptyKey() || LHS == getTombstoneKey() || 986 RHS == getEmptyKey() || RHS == getTombstoneKey()) 987 return LHS == RHS; 988 return LHS->isIdenticalTo(RHS); 989 } 990 }; 991 992 // Set of unique PHINodes. 993 DenseSet<PHINode *, PHIDenseMapInfo> PHISet; 994 995 // Examine each PHI. 996 bool Changed = false; 997 for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I++);) { 998 auto Inserted = PHISet.insert(PN); 999 if (!Inserted.second) { 1000 // A duplicate. Replace this PHI with its duplicate. 1001 PN->replaceAllUsesWith(*Inserted.first); 1002 PN->eraseFromParent(); 1003 Changed = true; 1004 1005 // The RAUW can change PHIs that we already visited. Start over from the 1006 // beginning. 1007 PHISet.clear(); 1008 I = BB->begin(); 1009 } 1010 } 1011 1012 return Changed; 1013 } 1014 1015 /// enforceKnownAlignment - If the specified pointer points to an object that 1016 /// we control, modify the object's alignment to PrefAlign. This isn't 1017 /// often possible though. If alignment is important, a more reliable approach 1018 /// is to simply align all global variables and allocation instructions to 1019 /// their preferred alignment from the beginning. 1020 static unsigned enforceKnownAlignment(Value *V, unsigned Align, 1021 unsigned PrefAlign, 1022 const DataLayout &DL) { 1023 assert(PrefAlign > Align); 1024 1025 V = V->stripPointerCasts(); 1026 1027 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 1028 // TODO: ideally, computeKnownBits ought to have used 1029 // AllocaInst::getAlignment() in its computation already, making 1030 // the below max redundant. But, as it turns out, 1031 // stripPointerCasts recurses through infinite layers of bitcasts, 1032 // while computeKnownBits is not allowed to traverse more than 6 1033 // levels. 1034 Align = std::max(AI->getAlignment(), Align); 1035 if (PrefAlign <= Align) 1036 return Align; 1037 1038 // If the preferred alignment is greater than the natural stack alignment 1039 // then don't round up. This avoids dynamic stack realignment. 1040 if (DL.exceedsNaturalStackAlignment(PrefAlign)) 1041 return Align; 1042 AI->setAlignment(PrefAlign); 1043 return PrefAlign; 1044 } 1045 1046 if (auto *GO = dyn_cast<GlobalObject>(V)) { 1047 // TODO: as above, this shouldn't be necessary. 1048 Align = std::max(GO->getAlignment(), Align); 1049 if (PrefAlign <= Align) 1050 return Align; 1051 1052 // If there is a large requested alignment and we can, bump up the alignment 1053 // of the global. If the memory we set aside for the global may not be the 1054 // memory used by the final program then it is impossible for us to reliably 1055 // enforce the preferred alignment. 1056 if (!GO->canIncreaseAlignment()) 1057 return Align; 1058 1059 GO->setAlignment(PrefAlign); 1060 return PrefAlign; 1061 } 1062 1063 return Align; 1064 } 1065 1066 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign, 1067 const DataLayout &DL, 1068 const Instruction *CxtI, 1069 AssumptionCache *AC, 1070 const DominatorTree *DT) { 1071 assert(V->getType()->isPointerTy() && 1072 "getOrEnforceKnownAlignment expects a pointer!"); 1073 1074 KnownBits Known = computeKnownBits(V, DL, 0, AC, CxtI, DT); 1075 unsigned TrailZ = Known.countMinTrailingZeros(); 1076 1077 // Avoid trouble with ridiculously large TrailZ values, such as 1078 // those computed from a null pointer. 1079 TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1)); 1080 1081 unsigned Align = 1u << std::min(Known.getBitWidth() - 1, TrailZ); 1082 1083 // LLVM doesn't support alignments larger than this currently. 1084 Align = std::min(Align, +Value::MaximumAlignment); 1085 1086 if (PrefAlign > Align) 1087 Align = enforceKnownAlignment(V, Align, PrefAlign, DL); 1088 1089 // We don't need to make any adjustment. 1090 return Align; 1091 } 1092 1093 ///===---------------------------------------------------------------------===// 1094 /// Dbg Intrinsic utilities 1095 /// 1096 1097 /// See if there is a dbg.value intrinsic for DIVar before I. 1098 static bool LdStHasDebugValue(DILocalVariable *DIVar, DIExpression *DIExpr, 1099 Instruction *I) { 1100 // Since we can't guarantee that the original dbg.declare instrinsic 1101 // is removed by LowerDbgDeclare(), we need to make sure that we are 1102 // not inserting the same dbg.value intrinsic over and over. 1103 BasicBlock::InstListType::iterator PrevI(I); 1104 if (PrevI != I->getParent()->getInstList().begin()) { 1105 --PrevI; 1106 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI)) 1107 if (DVI->getValue() == I->getOperand(0) && 1108 DVI->getVariable() == DIVar && 1109 DVI->getExpression() == DIExpr) 1110 return true; 1111 } 1112 return false; 1113 } 1114 1115 /// See if there is a dbg.value intrinsic for DIVar for the PHI node. 1116 static bool PhiHasDebugValue(DILocalVariable *DIVar, 1117 DIExpression *DIExpr, 1118 PHINode *APN) { 1119 // Since we can't guarantee that the original dbg.declare instrinsic 1120 // is removed by LowerDbgDeclare(), we need to make sure that we are 1121 // not inserting the same dbg.value intrinsic over and over. 1122 SmallVector<DbgValueInst *, 1> DbgValues; 1123 findDbgValues(DbgValues, APN); 1124 for (auto *DVI : DbgValues) { 1125 assert(DVI->getValue() == APN); 1126 if ((DVI->getVariable() == DIVar) && (DVI->getExpression() == DIExpr)) 1127 return true; 1128 } 1129 return false; 1130 } 1131 1132 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value 1133 /// that has an associated llvm.dbg.declare or llvm.dbg.addr intrinsic. 1134 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII, 1135 StoreInst *SI, DIBuilder &Builder) { 1136 assert(DII->isAddressOfVariable()); 1137 auto *DIVar = DII->getVariable(); 1138 assert(DIVar && "Missing variable"); 1139 auto *DIExpr = DII->getExpression(); 1140 Value *DV = SI->getOperand(0); 1141 1142 // If an argument is zero extended then use argument directly. The ZExt 1143 // may be zapped by an optimization pass in future. 1144 Argument *ExtendedArg = nullptr; 1145 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 1146 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0)); 1147 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 1148 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0)); 1149 if (ExtendedArg) { 1150 // If this DII was already describing only a fragment of a variable, ensure 1151 // that fragment is appropriately narrowed here. 1152 // But if a fragment wasn't used, describe the value as the original 1153 // argument (rather than the zext or sext) so that it remains described even 1154 // if the sext/zext is optimized away. This widens the variable description, 1155 // leaving it up to the consumer to know how the smaller value may be 1156 // represented in a larger register. 1157 if (auto Fragment = DIExpr->getFragmentInfo()) { 1158 unsigned FragmentOffset = Fragment->OffsetInBits; 1159 SmallVector<uint64_t, 3> Ops(DIExpr->elements_begin(), 1160 DIExpr->elements_end() - 3); 1161 Ops.push_back(dwarf::DW_OP_LLVM_fragment); 1162 Ops.push_back(FragmentOffset); 1163 const DataLayout &DL = DII->getModule()->getDataLayout(); 1164 Ops.push_back(DL.getTypeSizeInBits(ExtendedArg->getType())); 1165 DIExpr = Builder.createExpression(Ops); 1166 } 1167 DV = ExtendedArg; 1168 } 1169 if (!LdStHasDebugValue(DIVar, DIExpr, SI)) 1170 Builder.insertDbgValueIntrinsic(DV, DIVar, DIExpr, DII->getDebugLoc(), 1171 SI); 1172 } 1173 1174 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value 1175 /// that has an associated llvm.dbg.declare or llvm.dbg.addr intrinsic. 1176 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII, 1177 LoadInst *LI, DIBuilder &Builder) { 1178 auto *DIVar = DII->getVariable(); 1179 auto *DIExpr = DII->getExpression(); 1180 assert(DIVar && "Missing variable"); 1181 1182 if (LdStHasDebugValue(DIVar, DIExpr, LI)) 1183 return; 1184 1185 // We are now tracking the loaded value instead of the address. In the 1186 // future if multi-location support is added to the IR, it might be 1187 // preferable to keep tracking both the loaded value and the original 1188 // address in case the alloca can not be elided. 1189 Instruction *DbgValue = Builder.insertDbgValueIntrinsic( 1190 LI, DIVar, DIExpr, DII->getDebugLoc(), (Instruction *)nullptr); 1191 DbgValue->insertAfter(LI); 1192 } 1193 1194 /// Inserts a llvm.dbg.value intrinsic after a phi that has an associated 1195 /// llvm.dbg.declare or llvm.dbg.addr intrinsic. 1196 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII, 1197 PHINode *APN, DIBuilder &Builder) { 1198 auto *DIVar = DII->getVariable(); 1199 auto *DIExpr = DII->getExpression(); 1200 assert(DIVar && "Missing variable"); 1201 1202 if (PhiHasDebugValue(DIVar, DIExpr, APN)) 1203 return; 1204 1205 BasicBlock *BB = APN->getParent(); 1206 auto InsertionPt = BB->getFirstInsertionPt(); 1207 1208 // The block may be a catchswitch block, which does not have a valid 1209 // insertion point. 1210 // FIXME: Insert dbg.value markers in the successors when appropriate. 1211 if (InsertionPt != BB->end()) 1212 Builder.insertDbgValueIntrinsic(APN, DIVar, DIExpr, DII->getDebugLoc(), 1213 &*InsertionPt); 1214 } 1215 1216 /// Determine whether this alloca is either a VLA or an array. 1217 static bool isArray(AllocaInst *AI) { 1218 return AI->isArrayAllocation() || 1219 AI->getType()->getElementType()->isArrayTy(); 1220 } 1221 1222 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set 1223 /// of llvm.dbg.value intrinsics. 1224 bool llvm::LowerDbgDeclare(Function &F) { 1225 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false); 1226 SmallVector<DbgDeclareInst *, 4> Dbgs; 1227 for (auto &FI : F) 1228 for (Instruction &BI : FI) 1229 if (auto DDI = dyn_cast<DbgDeclareInst>(&BI)) 1230 Dbgs.push_back(DDI); 1231 1232 if (Dbgs.empty()) 1233 return false; 1234 1235 for (auto &I : Dbgs) { 1236 DbgDeclareInst *DDI = I; 1237 AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress()); 1238 // If this is an alloca for a scalar variable, insert a dbg.value 1239 // at each load and store to the alloca and erase the dbg.declare. 1240 // The dbg.values allow tracking a variable even if it is not 1241 // stored on the stack, while the dbg.declare can only describe 1242 // the stack slot (and at a lexical-scope granularity). Later 1243 // passes will attempt to elide the stack slot. 1244 if (AI && !isArray(AI)) { 1245 for (auto &AIUse : AI->uses()) { 1246 User *U = AIUse.getUser(); 1247 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1248 if (AIUse.getOperandNo() == 1) 1249 ConvertDebugDeclareToDebugValue(DDI, SI, DIB); 1250 } else if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1251 ConvertDebugDeclareToDebugValue(DDI, LI, DIB); 1252 } else if (CallInst *CI = dyn_cast<CallInst>(U)) { 1253 // This is a call by-value or some other instruction that 1254 // takes a pointer to the variable. Insert a *value* 1255 // intrinsic that describes the alloca. 1256 DIB.insertDbgValueIntrinsic(AI, DDI->getVariable(), 1257 DDI->getExpression(), DDI->getDebugLoc(), 1258 CI); 1259 } 1260 } 1261 DDI->eraseFromParent(); 1262 } 1263 } 1264 return true; 1265 } 1266 1267 /// Finds all intrinsics declaring local variables as living in the memory that 1268 /// 'V' points to. This may include a mix of dbg.declare and 1269 /// dbg.addr intrinsics. 1270 TinyPtrVector<DbgInfoIntrinsic *> llvm::FindDbgAddrUses(Value *V) { 1271 auto *L = LocalAsMetadata::getIfExists(V); 1272 if (!L) 1273 return {}; 1274 auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L); 1275 if (!MDV) 1276 return {}; 1277 1278 TinyPtrVector<DbgInfoIntrinsic *> Declares; 1279 for (User *U : MDV->users()) { 1280 if (auto *DII = dyn_cast<DbgInfoIntrinsic>(U)) 1281 if (DII->isAddressOfVariable()) 1282 Declares.push_back(DII); 1283 } 1284 1285 return Declares; 1286 } 1287 1288 void llvm::findDbgValues(SmallVectorImpl<DbgValueInst *> &DbgValues, Value *V) { 1289 if (auto *L = LocalAsMetadata::getIfExists(V)) 1290 if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L)) 1291 for (User *U : MDV->users()) 1292 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U)) 1293 DbgValues.push_back(DVI); 1294 } 1295 1296 bool llvm::replaceDbgDeclare(Value *Address, Value *NewAddress, 1297 Instruction *InsertBefore, DIBuilder &Builder, 1298 bool Deref, int Offset) { 1299 auto DbgAddrs = FindDbgAddrUses(Address); 1300 for (DbgInfoIntrinsic *DII : DbgAddrs) { 1301 DebugLoc Loc = DII->getDebugLoc(); 1302 auto *DIVar = DII->getVariable(); 1303 auto *DIExpr = DII->getExpression(); 1304 assert(DIVar && "Missing variable"); 1305 DIExpr = DIExpression::prepend(DIExpr, Deref, Offset); 1306 // Insert llvm.dbg.declare immediately after InsertBefore, and remove old 1307 // llvm.dbg.declare. 1308 Builder.insertDeclare(NewAddress, DIVar, DIExpr, Loc, InsertBefore); 1309 if (DII == InsertBefore) 1310 InsertBefore = &*std::next(InsertBefore->getIterator()); 1311 DII->eraseFromParent(); 1312 } 1313 return !DbgAddrs.empty(); 1314 } 1315 1316 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress, 1317 DIBuilder &Builder, bool Deref, int Offset) { 1318 return replaceDbgDeclare(AI, NewAllocaAddress, AI->getNextNode(), Builder, 1319 Deref, Offset); 1320 } 1321 1322 static void replaceOneDbgValueForAlloca(DbgValueInst *DVI, Value *NewAddress, 1323 DIBuilder &Builder, int Offset) { 1324 DebugLoc Loc = DVI->getDebugLoc(); 1325 auto *DIVar = DVI->getVariable(); 1326 auto *DIExpr = DVI->getExpression(); 1327 assert(DIVar && "Missing variable"); 1328 1329 // This is an alloca-based llvm.dbg.value. The first thing it should do with 1330 // the alloca pointer is dereference it. Otherwise we don't know how to handle 1331 // it and give up. 1332 if (!DIExpr || DIExpr->getNumElements() < 1 || 1333 DIExpr->getElement(0) != dwarf::DW_OP_deref) 1334 return; 1335 1336 // Insert the offset immediately after the first deref. 1337 // We could just change the offset argument of dbg.value, but it's unsigned... 1338 if (Offset) { 1339 SmallVector<uint64_t, 4> Ops; 1340 Ops.push_back(dwarf::DW_OP_deref); 1341 DIExpression::appendOffset(Ops, Offset); 1342 Ops.append(DIExpr->elements_begin() + 1, DIExpr->elements_end()); 1343 DIExpr = Builder.createExpression(Ops); 1344 } 1345 1346 Builder.insertDbgValueIntrinsic(NewAddress, DIVar, DIExpr, Loc, DVI); 1347 DVI->eraseFromParent(); 1348 } 1349 1350 void llvm::replaceDbgValueForAlloca(AllocaInst *AI, Value *NewAllocaAddress, 1351 DIBuilder &Builder, int Offset) { 1352 if (auto *L = LocalAsMetadata::getIfExists(AI)) 1353 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L)) 1354 for (auto UI = MDV->use_begin(), UE = MDV->use_end(); UI != UE;) { 1355 Use &U = *UI++; 1356 if (auto *DVI = dyn_cast<DbgValueInst>(U.getUser())) 1357 replaceOneDbgValueForAlloca(DVI, NewAllocaAddress, Builder, Offset); 1358 } 1359 } 1360 1361 void llvm::salvageDebugInfo(Instruction &I) { 1362 SmallVector<DbgValueInst *, 1> DbgValues; 1363 auto &M = *I.getModule(); 1364 1365 auto MDWrap = [&](Value *V) { 1366 return MetadataAsValue::get(I.getContext(), ValueAsMetadata::get(V)); 1367 }; 1368 1369 if (isa<BitCastInst>(&I)) { 1370 findDbgValues(DbgValues, &I); 1371 for (auto *DVI : DbgValues) { 1372 // Bitcasts are entirely irrelevant for debug info. Rewrite the dbg.value 1373 // to use the cast's source. 1374 DVI->setOperand(0, MDWrap(I.getOperand(0))); 1375 DEBUG(dbgs() << "SALVAGE: " << *DVI << '\n'); 1376 } 1377 } else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { 1378 findDbgValues(DbgValues, &I); 1379 for (auto *DVI : DbgValues) { 1380 unsigned BitWidth = 1381 M.getDataLayout().getPointerSizeInBits(GEP->getPointerAddressSpace()); 1382 APInt Offset(BitWidth, 0); 1383 // Rewrite a constant GEP into a DIExpression. Since we are performing 1384 // arithmetic to compute the variable's *value* in the DIExpression, we 1385 // need to mark the expression with a DW_OP_stack_value. 1386 if (GEP->accumulateConstantOffset(M.getDataLayout(), Offset)) { 1387 auto *DIExpr = DVI->getExpression(); 1388 // GEP offsets are i32 and thus always fit into an int64_t. 1389 DIExpr = DIExpression::prepend(DIExpr, DIExpression::NoDeref, 1390 Offset.getSExtValue(), 1391 DIExpression::WithStackValue); 1392 DVI->setOperand(0, MDWrap(I.getOperand(0))); 1393 DVI->setOperand(2, MetadataAsValue::get(I.getContext(), DIExpr)); 1394 DEBUG(dbgs() << "SALVAGE: " << *DVI << '\n'); 1395 } 1396 } 1397 } else if (isa<LoadInst>(&I)) { 1398 findDbgValues(DbgValues, &I); 1399 for (auto *DVI : DbgValues) { 1400 // Rewrite the load into DW_OP_deref. 1401 auto *DIExpr = DVI->getExpression(); 1402 DIExpr = DIExpression::prepend(DIExpr, DIExpression::WithDeref); 1403 DVI->setOperand(0, MDWrap(I.getOperand(0))); 1404 DVI->setOperand(2, MetadataAsValue::get(I.getContext(), DIExpr)); 1405 DEBUG(dbgs() << "SALVAGE: " << *DVI << '\n'); 1406 } 1407 } 1408 } 1409 1410 unsigned llvm::removeAllNonTerminatorAndEHPadInstructions(BasicBlock *BB) { 1411 unsigned NumDeadInst = 0; 1412 // Delete the instructions backwards, as it has a reduced likelihood of 1413 // having to update as many def-use and use-def chains. 1414 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted. 1415 while (EndInst != &BB->front()) { 1416 // Delete the next to last instruction. 1417 Instruction *Inst = &*--EndInst->getIterator(); 1418 if (!Inst->use_empty() && !Inst->getType()->isTokenTy()) 1419 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType())); 1420 if (Inst->isEHPad() || Inst->getType()->isTokenTy()) { 1421 EndInst = Inst; 1422 continue; 1423 } 1424 if (!isa<DbgInfoIntrinsic>(Inst)) 1425 ++NumDeadInst; 1426 Inst->eraseFromParent(); 1427 } 1428 return NumDeadInst; 1429 } 1430 1431 unsigned llvm::changeToUnreachable(Instruction *I, bool UseLLVMTrap, 1432 bool PreserveLCSSA) { 1433 BasicBlock *BB = I->getParent(); 1434 // Loop over all of the successors, removing BB's entry from any PHI 1435 // nodes. 1436 for (BasicBlock *Successor : successors(BB)) 1437 Successor->removePredecessor(BB, PreserveLCSSA); 1438 1439 // Insert a call to llvm.trap right before this. This turns the undefined 1440 // behavior into a hard fail instead of falling through into random code. 1441 if (UseLLVMTrap) { 1442 Function *TrapFn = 1443 Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap); 1444 CallInst *CallTrap = CallInst::Create(TrapFn, "", I); 1445 CallTrap->setDebugLoc(I->getDebugLoc()); 1446 } 1447 new UnreachableInst(I->getContext(), I); 1448 1449 // All instructions after this are dead. 1450 unsigned NumInstrsRemoved = 0; 1451 BasicBlock::iterator BBI = I->getIterator(), BBE = BB->end(); 1452 while (BBI != BBE) { 1453 if (!BBI->use_empty()) 1454 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 1455 BB->getInstList().erase(BBI++); 1456 ++NumInstrsRemoved; 1457 } 1458 return NumInstrsRemoved; 1459 } 1460 1461 /// changeToCall - Convert the specified invoke into a normal call. 1462 static void changeToCall(InvokeInst *II) { 1463 SmallVector<Value*, 8> Args(II->arg_begin(), II->arg_end()); 1464 SmallVector<OperandBundleDef, 1> OpBundles; 1465 II->getOperandBundlesAsDefs(OpBundles); 1466 CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, OpBundles, 1467 "", II); 1468 NewCall->takeName(II); 1469 NewCall->setCallingConv(II->getCallingConv()); 1470 NewCall->setAttributes(II->getAttributes()); 1471 NewCall->setDebugLoc(II->getDebugLoc()); 1472 II->replaceAllUsesWith(NewCall); 1473 1474 // Follow the call by a branch to the normal destination. 1475 BranchInst::Create(II->getNormalDest(), II); 1476 1477 // Update PHI nodes in the unwind destination 1478 II->getUnwindDest()->removePredecessor(II->getParent()); 1479 II->eraseFromParent(); 1480 } 1481 1482 BasicBlock *llvm::changeToInvokeAndSplitBasicBlock(CallInst *CI, 1483 BasicBlock *UnwindEdge) { 1484 BasicBlock *BB = CI->getParent(); 1485 1486 // Convert this function call into an invoke instruction. First, split the 1487 // basic block. 1488 BasicBlock *Split = 1489 BB->splitBasicBlock(CI->getIterator(), CI->getName() + ".noexc"); 1490 1491 // Delete the unconditional branch inserted by splitBasicBlock 1492 BB->getInstList().pop_back(); 1493 1494 // Create the new invoke instruction. 1495 SmallVector<Value *, 8> InvokeArgs(CI->arg_begin(), CI->arg_end()); 1496 SmallVector<OperandBundleDef, 1> OpBundles; 1497 1498 CI->getOperandBundlesAsDefs(OpBundles); 1499 1500 // Note: we're round tripping operand bundles through memory here, and that 1501 // can potentially be avoided with a cleverer API design that we do not have 1502 // as of this time. 1503 1504 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split, UnwindEdge, 1505 InvokeArgs, OpBundles, CI->getName(), BB); 1506 II->setDebugLoc(CI->getDebugLoc()); 1507 II->setCallingConv(CI->getCallingConv()); 1508 II->setAttributes(CI->getAttributes()); 1509 1510 // Make sure that anything using the call now uses the invoke! This also 1511 // updates the CallGraph if present, because it uses a WeakTrackingVH. 1512 CI->replaceAllUsesWith(II); 1513 1514 // Delete the original call 1515 Split->getInstList().pop_front(); 1516 return Split; 1517 } 1518 1519 static bool markAliveBlocks(Function &F, 1520 SmallPtrSetImpl<BasicBlock*> &Reachable) { 1521 SmallVector<BasicBlock*, 128> Worklist; 1522 BasicBlock *BB = &F.front(); 1523 Worklist.push_back(BB); 1524 Reachable.insert(BB); 1525 bool Changed = false; 1526 do { 1527 BB = Worklist.pop_back_val(); 1528 1529 // Do a quick scan of the basic block, turning any obviously unreachable 1530 // instructions into LLVM unreachable insts. The instruction combining pass 1531 // canonicalizes unreachable insts into stores to null or undef. 1532 for (Instruction &I : *BB) { 1533 // Assumptions that are known to be false are equivalent to unreachable. 1534 // Also, if the condition is undefined, then we make the choice most 1535 // beneficial to the optimizer, and choose that to also be unreachable. 1536 if (auto *II = dyn_cast<IntrinsicInst>(&I)) { 1537 if (II->getIntrinsicID() == Intrinsic::assume) { 1538 if (match(II->getArgOperand(0), m_CombineOr(m_Zero(), m_Undef()))) { 1539 // Don't insert a call to llvm.trap right before the unreachable. 1540 changeToUnreachable(II, false); 1541 Changed = true; 1542 break; 1543 } 1544 } 1545 1546 if (II->getIntrinsicID() == Intrinsic::experimental_guard) { 1547 // A call to the guard intrinsic bails out of the current compilation 1548 // unit if the predicate passed to it is false. If the predicate is a 1549 // constant false, then we know the guard will bail out of the current 1550 // compile unconditionally, so all code following it is dead. 1551 // 1552 // Note: unlike in llvm.assume, it is not "obviously profitable" for 1553 // guards to treat `undef` as `false` since a guard on `undef` can 1554 // still be useful for widening. 1555 if (match(II->getArgOperand(0), m_Zero())) 1556 if (!isa<UnreachableInst>(II->getNextNode())) { 1557 changeToUnreachable(II->getNextNode(), /*UseLLVMTrap=*/ false); 1558 Changed = true; 1559 break; 1560 } 1561 } 1562 } 1563 1564 if (auto *CI = dyn_cast<CallInst>(&I)) { 1565 Value *Callee = CI->getCalledValue(); 1566 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 1567 changeToUnreachable(CI, /*UseLLVMTrap=*/false); 1568 Changed = true; 1569 break; 1570 } 1571 if (CI->doesNotReturn()) { 1572 // If we found a call to a no-return function, insert an unreachable 1573 // instruction after it. Make sure there isn't *already* one there 1574 // though. 1575 if (!isa<UnreachableInst>(CI->getNextNode())) { 1576 // Don't insert a call to llvm.trap right before the unreachable. 1577 changeToUnreachable(CI->getNextNode(), false); 1578 Changed = true; 1579 } 1580 break; 1581 } 1582 } 1583 1584 // Store to undef and store to null are undefined and used to signal that 1585 // they should be changed to unreachable by passes that can't modify the 1586 // CFG. 1587 if (auto *SI = dyn_cast<StoreInst>(&I)) { 1588 // Don't touch volatile stores. 1589 if (SI->isVolatile()) continue; 1590 1591 Value *Ptr = SI->getOperand(1); 1592 1593 if (isa<UndefValue>(Ptr) || 1594 (isa<ConstantPointerNull>(Ptr) && 1595 SI->getPointerAddressSpace() == 0)) { 1596 changeToUnreachable(SI, true); 1597 Changed = true; 1598 break; 1599 } 1600 } 1601 } 1602 1603 TerminatorInst *Terminator = BB->getTerminator(); 1604 if (auto *II = dyn_cast<InvokeInst>(Terminator)) { 1605 // Turn invokes that call 'nounwind' functions into ordinary calls. 1606 Value *Callee = II->getCalledValue(); 1607 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 1608 changeToUnreachable(II, true); 1609 Changed = true; 1610 } else if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(&F)) { 1611 if (II->use_empty() && II->onlyReadsMemory()) { 1612 // jump to the normal destination branch. 1613 BranchInst::Create(II->getNormalDest(), II); 1614 II->getUnwindDest()->removePredecessor(II->getParent()); 1615 II->eraseFromParent(); 1616 } else 1617 changeToCall(II); 1618 Changed = true; 1619 } 1620 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Terminator)) { 1621 // Remove catchpads which cannot be reached. 1622 struct CatchPadDenseMapInfo { 1623 static CatchPadInst *getEmptyKey() { 1624 return DenseMapInfo<CatchPadInst *>::getEmptyKey(); 1625 } 1626 1627 static CatchPadInst *getTombstoneKey() { 1628 return DenseMapInfo<CatchPadInst *>::getTombstoneKey(); 1629 } 1630 1631 static unsigned getHashValue(CatchPadInst *CatchPad) { 1632 return static_cast<unsigned>(hash_combine_range( 1633 CatchPad->value_op_begin(), CatchPad->value_op_end())); 1634 } 1635 1636 static bool isEqual(CatchPadInst *LHS, CatchPadInst *RHS) { 1637 if (LHS == getEmptyKey() || LHS == getTombstoneKey() || 1638 RHS == getEmptyKey() || RHS == getTombstoneKey()) 1639 return LHS == RHS; 1640 return LHS->isIdenticalTo(RHS); 1641 } 1642 }; 1643 1644 // Set of unique CatchPads. 1645 SmallDenseMap<CatchPadInst *, detail::DenseSetEmpty, 4, 1646 CatchPadDenseMapInfo, detail::DenseSetPair<CatchPadInst *>> 1647 HandlerSet; 1648 detail::DenseSetEmpty Empty; 1649 for (CatchSwitchInst::handler_iterator I = CatchSwitch->handler_begin(), 1650 E = CatchSwitch->handler_end(); 1651 I != E; ++I) { 1652 BasicBlock *HandlerBB = *I; 1653 auto *CatchPad = cast<CatchPadInst>(HandlerBB->getFirstNonPHI()); 1654 if (!HandlerSet.insert({CatchPad, Empty}).second) { 1655 CatchSwitch->removeHandler(I); 1656 --I; 1657 --E; 1658 Changed = true; 1659 } 1660 } 1661 } 1662 1663 Changed |= ConstantFoldTerminator(BB, true); 1664 for (BasicBlock *Successor : successors(BB)) 1665 if (Reachable.insert(Successor).second) 1666 Worklist.push_back(Successor); 1667 } while (!Worklist.empty()); 1668 return Changed; 1669 } 1670 1671 void llvm::removeUnwindEdge(BasicBlock *BB) { 1672 TerminatorInst *TI = BB->getTerminator(); 1673 1674 if (auto *II = dyn_cast<InvokeInst>(TI)) { 1675 changeToCall(II); 1676 return; 1677 } 1678 1679 TerminatorInst *NewTI; 1680 BasicBlock *UnwindDest; 1681 1682 if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) { 1683 NewTI = CleanupReturnInst::Create(CRI->getCleanupPad(), nullptr, CRI); 1684 UnwindDest = CRI->getUnwindDest(); 1685 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(TI)) { 1686 auto *NewCatchSwitch = CatchSwitchInst::Create( 1687 CatchSwitch->getParentPad(), nullptr, CatchSwitch->getNumHandlers(), 1688 CatchSwitch->getName(), CatchSwitch); 1689 for (BasicBlock *PadBB : CatchSwitch->handlers()) 1690 NewCatchSwitch->addHandler(PadBB); 1691 1692 NewTI = NewCatchSwitch; 1693 UnwindDest = CatchSwitch->getUnwindDest(); 1694 } else { 1695 llvm_unreachable("Could not find unwind successor"); 1696 } 1697 1698 NewTI->takeName(TI); 1699 NewTI->setDebugLoc(TI->getDebugLoc()); 1700 UnwindDest->removePredecessor(BB); 1701 TI->replaceAllUsesWith(NewTI); 1702 TI->eraseFromParent(); 1703 } 1704 1705 /// removeUnreachableBlocks - Remove blocks that are not reachable, even 1706 /// if they are in a dead cycle. Return true if a change was made, false 1707 /// otherwise. If `LVI` is passed, this function preserves LazyValueInfo 1708 /// after modifying the CFG. 1709 bool llvm::removeUnreachableBlocks(Function &F, LazyValueInfo *LVI) { 1710 SmallPtrSet<BasicBlock*, 16> Reachable; 1711 bool Changed = markAliveBlocks(F, Reachable); 1712 1713 // If there are unreachable blocks in the CFG... 1714 if (Reachable.size() == F.size()) 1715 return Changed; 1716 1717 assert(Reachable.size() < F.size()); 1718 NumRemoved += F.size()-Reachable.size(); 1719 1720 // Loop over all of the basic blocks that are not reachable, dropping all of 1721 // their internal references... 1722 for (Function::iterator BB = ++F.begin(), E = F.end(); BB != E; ++BB) { 1723 if (Reachable.count(&*BB)) 1724 continue; 1725 1726 for (BasicBlock *Successor : successors(&*BB)) 1727 if (Reachable.count(Successor)) 1728 Successor->removePredecessor(&*BB); 1729 if (LVI) 1730 LVI->eraseBlock(&*BB); 1731 BB->dropAllReferences(); 1732 } 1733 1734 for (Function::iterator I = ++F.begin(); I != F.end();) 1735 if (!Reachable.count(&*I)) 1736 I = F.getBasicBlockList().erase(I); 1737 else 1738 ++I; 1739 1740 return true; 1741 } 1742 1743 void llvm::combineMetadata(Instruction *K, const Instruction *J, 1744 ArrayRef<unsigned> KnownIDs) { 1745 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata; 1746 K->dropUnknownNonDebugMetadata(KnownIDs); 1747 K->getAllMetadataOtherThanDebugLoc(Metadata); 1748 for (const auto &MD : Metadata) { 1749 unsigned Kind = MD.first; 1750 MDNode *JMD = J->getMetadata(Kind); 1751 MDNode *KMD = MD.second; 1752 1753 switch (Kind) { 1754 default: 1755 K->setMetadata(Kind, nullptr); // Remove unknown metadata 1756 break; 1757 case LLVMContext::MD_dbg: 1758 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg"); 1759 case LLVMContext::MD_tbaa: 1760 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 1761 break; 1762 case LLVMContext::MD_alias_scope: 1763 K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD)); 1764 break; 1765 case LLVMContext::MD_noalias: 1766 case LLVMContext::MD_mem_parallel_loop_access: 1767 K->setMetadata(Kind, MDNode::intersect(JMD, KMD)); 1768 break; 1769 case LLVMContext::MD_range: 1770 K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD)); 1771 break; 1772 case LLVMContext::MD_fpmath: 1773 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 1774 break; 1775 case LLVMContext::MD_invariant_load: 1776 // Only set the !invariant.load if it is present in both instructions. 1777 K->setMetadata(Kind, JMD); 1778 break; 1779 case LLVMContext::MD_nonnull: 1780 // Only set the !nonnull if it is present in both instructions. 1781 K->setMetadata(Kind, JMD); 1782 break; 1783 case LLVMContext::MD_invariant_group: 1784 // Preserve !invariant.group in K. 1785 break; 1786 case LLVMContext::MD_align: 1787 K->setMetadata(Kind, 1788 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD)); 1789 break; 1790 case LLVMContext::MD_dereferenceable: 1791 case LLVMContext::MD_dereferenceable_or_null: 1792 K->setMetadata(Kind, 1793 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD)); 1794 break; 1795 } 1796 } 1797 // Set !invariant.group from J if J has it. If both instructions have it 1798 // then we will just pick it from J - even when they are different. 1799 // Also make sure that K is load or store - f.e. combining bitcast with load 1800 // could produce bitcast with invariant.group metadata, which is invalid. 1801 // FIXME: we should try to preserve both invariant.group md if they are 1802 // different, but right now instruction can only have one invariant.group. 1803 if (auto *JMD = J->getMetadata(LLVMContext::MD_invariant_group)) 1804 if (isa<LoadInst>(K) || isa<StoreInst>(K)) 1805 K->setMetadata(LLVMContext::MD_invariant_group, JMD); 1806 } 1807 1808 void llvm::combineMetadataForCSE(Instruction *K, const Instruction *J) { 1809 unsigned KnownIDs[] = { 1810 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope, 1811 LLVMContext::MD_noalias, LLVMContext::MD_range, 1812 LLVMContext::MD_invariant_load, LLVMContext::MD_nonnull, 1813 LLVMContext::MD_invariant_group, LLVMContext::MD_align, 1814 LLVMContext::MD_dereferenceable, 1815 LLVMContext::MD_dereferenceable_or_null}; 1816 combineMetadata(K, J, KnownIDs); 1817 } 1818 1819 template <typename RootType, typename DominatesFn> 1820 static unsigned replaceDominatedUsesWith(Value *From, Value *To, 1821 const RootType &Root, 1822 const DominatesFn &Dominates) { 1823 assert(From->getType() == To->getType()); 1824 1825 unsigned Count = 0; 1826 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 1827 UI != UE;) { 1828 Use &U = *UI++; 1829 if (!Dominates(Root, U)) 1830 continue; 1831 U.set(To); 1832 DEBUG(dbgs() << "Replace dominated use of '" << From->getName() << "' as " 1833 << *To << " in " << *U << "\n"); 1834 ++Count; 1835 } 1836 return Count; 1837 } 1838 1839 unsigned llvm::replaceNonLocalUsesWith(Instruction *From, Value *To) { 1840 assert(From->getType() == To->getType()); 1841 auto *BB = From->getParent(); 1842 unsigned Count = 0; 1843 1844 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 1845 UI != UE;) { 1846 Use &U = *UI++; 1847 auto *I = cast<Instruction>(U.getUser()); 1848 if (I->getParent() == BB) 1849 continue; 1850 U.set(To); 1851 ++Count; 1852 } 1853 return Count; 1854 } 1855 1856 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To, 1857 DominatorTree &DT, 1858 const BasicBlockEdge &Root) { 1859 auto Dominates = [&DT](const BasicBlockEdge &Root, const Use &U) { 1860 return DT.dominates(Root, U); 1861 }; 1862 return ::replaceDominatedUsesWith(From, To, Root, Dominates); 1863 } 1864 1865 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To, 1866 DominatorTree &DT, 1867 const BasicBlock *BB) { 1868 auto ProperlyDominates = [&DT](const BasicBlock *BB, const Use &U) { 1869 auto *I = cast<Instruction>(U.getUser())->getParent(); 1870 return DT.properlyDominates(BB, I); 1871 }; 1872 return ::replaceDominatedUsesWith(From, To, BB, ProperlyDominates); 1873 } 1874 1875 bool llvm::callsGCLeafFunction(ImmutableCallSite CS, 1876 const TargetLibraryInfo &TLI) { 1877 // Check if the function is specifically marked as a gc leaf function. 1878 if (CS.hasFnAttr("gc-leaf-function")) 1879 return true; 1880 if (const Function *F = CS.getCalledFunction()) { 1881 if (F->hasFnAttribute("gc-leaf-function")) 1882 return true; 1883 1884 if (auto IID = F->getIntrinsicID()) 1885 // Most LLVM intrinsics do not take safepoints. 1886 return IID != Intrinsic::experimental_gc_statepoint && 1887 IID != Intrinsic::experimental_deoptimize; 1888 } 1889 1890 // Lib calls can be materialized by some passes, and won't be 1891 // marked as 'gc-leaf-function.' All available Libcalls are 1892 // GC-leaf. 1893 LibFunc LF; 1894 if (TLI.getLibFunc(CS, LF)) { 1895 return TLI.has(LF); 1896 } 1897 1898 return false; 1899 } 1900 1901 void llvm::copyNonnullMetadata(const LoadInst &OldLI, MDNode *N, 1902 LoadInst &NewLI) { 1903 auto *NewTy = NewLI.getType(); 1904 1905 // This only directly applies if the new type is also a pointer. 1906 if (NewTy->isPointerTy()) { 1907 NewLI.setMetadata(LLVMContext::MD_nonnull, N); 1908 return; 1909 } 1910 1911 // The only other translation we can do is to integral loads with !range 1912 // metadata. 1913 if (!NewTy->isIntegerTy()) 1914 return; 1915 1916 MDBuilder MDB(NewLI.getContext()); 1917 const Value *Ptr = OldLI.getPointerOperand(); 1918 auto *ITy = cast<IntegerType>(NewTy); 1919 auto *NullInt = ConstantExpr::getPtrToInt( 1920 ConstantPointerNull::get(cast<PointerType>(Ptr->getType())), ITy); 1921 auto *NonNullInt = ConstantExpr::getAdd(NullInt, ConstantInt::get(ITy, 1)); 1922 NewLI.setMetadata(LLVMContext::MD_range, 1923 MDB.createRange(NonNullInt, NullInt)); 1924 } 1925 1926 void llvm::copyRangeMetadata(const DataLayout &DL, const LoadInst &OldLI, 1927 MDNode *N, LoadInst &NewLI) { 1928 auto *NewTy = NewLI.getType(); 1929 1930 // Give up unless it is converted to a pointer where there is a single very 1931 // valuable mapping we can do reliably. 1932 // FIXME: It would be nice to propagate this in more ways, but the type 1933 // conversions make it hard. 1934 if (!NewTy->isPointerTy()) 1935 return; 1936 1937 unsigned BitWidth = DL.getTypeSizeInBits(NewTy); 1938 if (!getConstantRangeFromMetadata(*N).contains(APInt(BitWidth, 0))) { 1939 MDNode *NN = MDNode::get(OldLI.getContext(), None); 1940 NewLI.setMetadata(LLVMContext::MD_nonnull, NN); 1941 } 1942 } 1943 1944 namespace { 1945 1946 /// A potential constituent of a bitreverse or bswap expression. See 1947 /// collectBitParts for a fuller explanation. 1948 struct BitPart { 1949 BitPart(Value *P, unsigned BW) : Provider(P) { 1950 Provenance.resize(BW); 1951 } 1952 1953 /// The Value that this is a bitreverse/bswap of. 1954 Value *Provider; 1955 1956 /// The "provenance" of each bit. Provenance[A] = B means that bit A 1957 /// in Provider becomes bit B in the result of this expression. 1958 SmallVector<int8_t, 32> Provenance; // int8_t means max size is i128. 1959 1960 enum { Unset = -1 }; 1961 }; 1962 1963 } // end anonymous namespace 1964 1965 /// Analyze the specified subexpression and see if it is capable of providing 1966 /// pieces of a bswap or bitreverse. The subexpression provides a potential 1967 /// piece of a bswap or bitreverse if it can be proven that each non-zero bit in 1968 /// the output of the expression came from a corresponding bit in some other 1969 /// value. This function is recursive, and the end result is a mapping of 1970 /// bitnumber to bitnumber. It is the caller's responsibility to validate that 1971 /// the bitnumber to bitnumber mapping is correct for a bswap or bitreverse. 1972 /// 1973 /// For example, if the current subexpression if "(shl i32 %X, 24)" then we know 1974 /// that the expression deposits the low byte of %X into the high byte of the 1975 /// result and that all other bits are zero. This expression is accepted and a 1976 /// BitPart is returned with Provider set to %X and Provenance[24-31] set to 1977 /// [0-7]. 1978 /// 1979 /// To avoid revisiting values, the BitPart results are memoized into the 1980 /// provided map. To avoid unnecessary copying of BitParts, BitParts are 1981 /// constructed in-place in the \c BPS map. Because of this \c BPS needs to 1982 /// store BitParts objects, not pointers. As we need the concept of a nullptr 1983 /// BitParts (Value has been analyzed and the analysis failed), we an Optional 1984 /// type instead to provide the same functionality. 1985 /// 1986 /// Because we pass around references into \c BPS, we must use a container that 1987 /// does not invalidate internal references (std::map instead of DenseMap). 1988 static const Optional<BitPart> & 1989 collectBitParts(Value *V, bool MatchBSwaps, bool MatchBitReversals, 1990 std::map<Value *, Optional<BitPart>> &BPS) { 1991 auto I = BPS.find(V); 1992 if (I != BPS.end()) 1993 return I->second; 1994 1995 auto &Result = BPS[V] = None; 1996 auto BitWidth = cast<IntegerType>(V->getType())->getBitWidth(); 1997 1998 if (Instruction *I = dyn_cast<Instruction>(V)) { 1999 // If this is an or instruction, it may be an inner node of the bswap. 2000 if (I->getOpcode() == Instruction::Or) { 2001 auto &A = collectBitParts(I->getOperand(0), MatchBSwaps, 2002 MatchBitReversals, BPS); 2003 auto &B = collectBitParts(I->getOperand(1), MatchBSwaps, 2004 MatchBitReversals, BPS); 2005 if (!A || !B) 2006 return Result; 2007 2008 // Try and merge the two together. 2009 if (!A->Provider || A->Provider != B->Provider) 2010 return Result; 2011 2012 Result = BitPart(A->Provider, BitWidth); 2013 for (unsigned i = 0; i < A->Provenance.size(); ++i) { 2014 if (A->Provenance[i] != BitPart::Unset && 2015 B->Provenance[i] != BitPart::Unset && 2016 A->Provenance[i] != B->Provenance[i]) 2017 return Result = None; 2018 2019 if (A->Provenance[i] == BitPart::Unset) 2020 Result->Provenance[i] = B->Provenance[i]; 2021 else 2022 Result->Provenance[i] = A->Provenance[i]; 2023 } 2024 2025 return Result; 2026 } 2027 2028 // If this is a logical shift by a constant, recurse then shift the result. 2029 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) { 2030 unsigned BitShift = 2031 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U); 2032 // Ensure the shift amount is defined. 2033 if (BitShift > BitWidth) 2034 return Result; 2035 2036 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps, 2037 MatchBitReversals, BPS); 2038 if (!Res) 2039 return Result; 2040 Result = Res; 2041 2042 // Perform the "shift" on BitProvenance. 2043 auto &P = Result->Provenance; 2044 if (I->getOpcode() == Instruction::Shl) { 2045 P.erase(std::prev(P.end(), BitShift), P.end()); 2046 P.insert(P.begin(), BitShift, BitPart::Unset); 2047 } else { 2048 P.erase(P.begin(), std::next(P.begin(), BitShift)); 2049 P.insert(P.end(), BitShift, BitPart::Unset); 2050 } 2051 2052 return Result; 2053 } 2054 2055 // If this is a logical 'and' with a mask that clears bits, recurse then 2056 // unset the appropriate bits. 2057 if (I->getOpcode() == Instruction::And && 2058 isa<ConstantInt>(I->getOperand(1))) { 2059 APInt Bit(I->getType()->getPrimitiveSizeInBits(), 1); 2060 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue(); 2061 2062 // Check that the mask allows a multiple of 8 bits for a bswap, for an 2063 // early exit. 2064 unsigned NumMaskedBits = AndMask.countPopulation(); 2065 if (!MatchBitReversals && NumMaskedBits % 8 != 0) 2066 return Result; 2067 2068 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps, 2069 MatchBitReversals, BPS); 2070 if (!Res) 2071 return Result; 2072 Result = Res; 2073 2074 for (unsigned i = 0; i < BitWidth; ++i, Bit <<= 1) 2075 // If the AndMask is zero for this bit, clear the bit. 2076 if ((AndMask & Bit) == 0) 2077 Result->Provenance[i] = BitPart::Unset; 2078 return Result; 2079 } 2080 2081 // If this is a zext instruction zero extend the result. 2082 if (I->getOpcode() == Instruction::ZExt) { 2083 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps, 2084 MatchBitReversals, BPS); 2085 if (!Res) 2086 return Result; 2087 2088 Result = BitPart(Res->Provider, BitWidth); 2089 auto NarrowBitWidth = 2090 cast<IntegerType>(cast<ZExtInst>(I)->getSrcTy())->getBitWidth(); 2091 for (unsigned i = 0; i < NarrowBitWidth; ++i) 2092 Result->Provenance[i] = Res->Provenance[i]; 2093 for (unsigned i = NarrowBitWidth; i < BitWidth; ++i) 2094 Result->Provenance[i] = BitPart::Unset; 2095 return Result; 2096 } 2097 } 2098 2099 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be 2100 // the input value to the bswap/bitreverse. 2101 Result = BitPart(V, BitWidth); 2102 for (unsigned i = 0; i < BitWidth; ++i) 2103 Result->Provenance[i] = i; 2104 return Result; 2105 } 2106 2107 static bool bitTransformIsCorrectForBSwap(unsigned From, unsigned To, 2108 unsigned BitWidth) { 2109 if (From % 8 != To % 8) 2110 return false; 2111 // Convert from bit indices to byte indices and check for a byte reversal. 2112 From >>= 3; 2113 To >>= 3; 2114 BitWidth >>= 3; 2115 return From == BitWidth - To - 1; 2116 } 2117 2118 static bool bitTransformIsCorrectForBitReverse(unsigned From, unsigned To, 2119 unsigned BitWidth) { 2120 return From == BitWidth - To - 1; 2121 } 2122 2123 /// Given an OR instruction, check to see if this is a bitreverse 2124 /// idiom. If so, insert the new intrinsic and return true. 2125 bool llvm::recognizeBSwapOrBitReverseIdiom( 2126 Instruction *I, bool MatchBSwaps, bool MatchBitReversals, 2127 SmallVectorImpl<Instruction *> &InsertedInsts) { 2128 if (Operator::getOpcode(I) != Instruction::Or) 2129 return false; 2130 if (!MatchBSwaps && !MatchBitReversals) 2131 return false; 2132 IntegerType *ITy = dyn_cast<IntegerType>(I->getType()); 2133 if (!ITy || ITy->getBitWidth() > 128) 2134 return false; // Can't do vectors or integers > 128 bits. 2135 unsigned BW = ITy->getBitWidth(); 2136 2137 unsigned DemandedBW = BW; 2138 IntegerType *DemandedTy = ITy; 2139 if (I->hasOneUse()) { 2140 if (TruncInst *Trunc = dyn_cast<TruncInst>(I->user_back())) { 2141 DemandedTy = cast<IntegerType>(Trunc->getType()); 2142 DemandedBW = DemandedTy->getBitWidth(); 2143 } 2144 } 2145 2146 // Try to find all the pieces corresponding to the bswap. 2147 std::map<Value *, Optional<BitPart>> BPS; 2148 auto Res = collectBitParts(I, MatchBSwaps, MatchBitReversals, BPS); 2149 if (!Res) 2150 return false; 2151 auto &BitProvenance = Res->Provenance; 2152 2153 // Now, is the bit permutation correct for a bswap or a bitreverse? We can 2154 // only byteswap values with an even number of bytes. 2155 bool OKForBSwap = DemandedBW % 16 == 0, OKForBitReverse = true; 2156 for (unsigned i = 0; i < DemandedBW; ++i) { 2157 OKForBSwap &= 2158 bitTransformIsCorrectForBSwap(BitProvenance[i], i, DemandedBW); 2159 OKForBitReverse &= 2160 bitTransformIsCorrectForBitReverse(BitProvenance[i], i, DemandedBW); 2161 } 2162 2163 Intrinsic::ID Intrin; 2164 if (OKForBSwap && MatchBSwaps) 2165 Intrin = Intrinsic::bswap; 2166 else if (OKForBitReverse && MatchBitReversals) 2167 Intrin = Intrinsic::bitreverse; 2168 else 2169 return false; 2170 2171 if (ITy != DemandedTy) { 2172 Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, DemandedTy); 2173 Value *Provider = Res->Provider; 2174 IntegerType *ProviderTy = cast<IntegerType>(Provider->getType()); 2175 // We may need to truncate the provider. 2176 if (DemandedTy != ProviderTy) { 2177 auto *Trunc = CastInst::Create(Instruction::Trunc, Provider, DemandedTy, 2178 "trunc", I); 2179 InsertedInsts.push_back(Trunc); 2180 Provider = Trunc; 2181 } 2182 auto *CI = CallInst::Create(F, Provider, "rev", I); 2183 InsertedInsts.push_back(CI); 2184 auto *ExtInst = CastInst::Create(Instruction::ZExt, CI, ITy, "zext", I); 2185 InsertedInsts.push_back(ExtInst); 2186 return true; 2187 } 2188 2189 Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, ITy); 2190 InsertedInsts.push_back(CallInst::Create(F, Res->Provider, "rev", I)); 2191 return true; 2192 } 2193 2194 // CodeGen has special handling for some string functions that may replace 2195 // them with target-specific intrinsics. Since that'd skip our interceptors 2196 // in ASan/MSan/TSan/DFSan, and thus make us miss some memory accesses, 2197 // we mark affected calls as NoBuiltin, which will disable optimization 2198 // in CodeGen. 2199 void llvm::maybeMarkSanitizerLibraryCallNoBuiltin( 2200 CallInst *CI, const TargetLibraryInfo *TLI) { 2201 Function *F = CI->getCalledFunction(); 2202 LibFunc Func; 2203 if (F && !F->hasLocalLinkage() && F->hasName() && 2204 TLI->getLibFunc(F->getName(), Func) && TLI->hasOptimizedCodeGen(Func) && 2205 !F->doesNotAccessMemory()) 2206 CI->addAttribute(AttributeList::FunctionIndex, Attribute::NoBuiltin); 2207 } 2208 2209 bool llvm::canReplaceOperandWithVariable(const Instruction *I, unsigned OpIdx) { 2210 // We can't have a PHI with a metadata type. 2211 if (I->getOperand(OpIdx)->getType()->isMetadataTy()) 2212 return false; 2213 2214 // Early exit. 2215 if (!isa<Constant>(I->getOperand(OpIdx))) 2216 return true; 2217 2218 switch (I->getOpcode()) { 2219 default: 2220 return true; 2221 case Instruction::Call: 2222 case Instruction::Invoke: 2223 // Can't handle inline asm. Skip it. 2224 if (isa<InlineAsm>(ImmutableCallSite(I).getCalledValue())) 2225 return false; 2226 // Many arithmetic intrinsics have no issue taking a 2227 // variable, however it's hard to distingish these from 2228 // specials such as @llvm.frameaddress that require a constant. 2229 if (isa<IntrinsicInst>(I)) 2230 return false; 2231 2232 // Constant bundle operands may need to retain their constant-ness for 2233 // correctness. 2234 if (ImmutableCallSite(I).isBundleOperand(OpIdx)) 2235 return false; 2236 return true; 2237 case Instruction::ShuffleVector: 2238 // Shufflevector masks are constant. 2239 return OpIdx != 2; 2240 case Instruction::Switch: 2241 case Instruction::ExtractValue: 2242 // All operands apart from the first are constant. 2243 return OpIdx == 0; 2244 case Instruction::InsertValue: 2245 // All operands apart from the first and the second are constant. 2246 return OpIdx < 2; 2247 case Instruction::Alloca: 2248 // Static allocas (constant size in the entry block) are handled by 2249 // prologue/epilogue insertion so they're free anyway. We definitely don't 2250 // want to make them non-constant. 2251 return !dyn_cast<AllocaInst>(I)->isStaticAlloca(); 2252 case Instruction::GetElementPtr: 2253 if (OpIdx == 0) 2254 return true; 2255 gep_type_iterator It = gep_type_begin(I); 2256 for (auto E = std::next(It, OpIdx); It != E; ++It) 2257 if (It.isStruct()) 2258 return false; 2259 return true; 2260 } 2261 } 2262