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/Analysis/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 "llvm/Transforms/Utils/ValueMapper.h" 77 #include <algorithm> 78 #include <cassert> 79 #include <climits> 80 #include <cstdint> 81 #include <iterator> 82 #include <map> 83 #include <utility> 84 85 using namespace llvm; 86 using namespace llvm::PatternMatch; 87 88 #define DEBUG_TYPE "local" 89 90 STATISTIC(NumRemoved, "Number of unreachable basic blocks removed"); 91 92 //===----------------------------------------------------------------------===// 93 // Local constant propagation. 94 // 95 96 /// ConstantFoldTerminator - If a terminator instruction is predicated on a 97 /// constant value, convert it into an unconditional branch to the constant 98 /// destination. This is a nontrivial operation because the successors of this 99 /// basic block must have their PHI nodes updated. 100 /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch 101 /// conditions and indirectbr addresses this might make dead if 102 /// DeleteDeadConditions is true. 103 bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions, 104 const TargetLibraryInfo *TLI, 105 DeferredDominance *DDT) { 106 TerminatorInst *T = BB->getTerminator(); 107 IRBuilder<> Builder(T); 108 109 // Branch - See if we are conditional jumping on constant 110 if (auto *BI = dyn_cast<BranchInst>(T)) { 111 if (BI->isUnconditional()) return false; // Can't optimize uncond branch 112 BasicBlock *Dest1 = BI->getSuccessor(0); 113 BasicBlock *Dest2 = BI->getSuccessor(1); 114 115 if (auto *Cond = dyn_cast<ConstantInt>(BI->getCondition())) { 116 // Are we branching on constant? 117 // YES. Change to unconditional branch... 118 BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2; 119 BasicBlock *OldDest = Cond->getZExtValue() ? Dest2 : Dest1; 120 121 // Let the basic block know that we are letting go of it. Based on this, 122 // it will adjust it's PHI nodes. 123 OldDest->removePredecessor(BB); 124 125 // Replace the conditional branch with an unconditional one. 126 Builder.CreateBr(Destination); 127 BI->eraseFromParent(); 128 if (DDT) 129 DDT->deleteEdge(BB, OldDest); 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 (auto *SI = dyn_cast<SwitchInst>(T)) { 154 // If we are switching on a constant, we can convert the switch to an 155 // unconditional branch. 156 auto *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 BasicBlock *ParentBB = SI->getParent(); 201 DefaultDest->removePredecessor(ParentBB); 202 i = SI->removeCase(i); 203 e = SI->case_end(); 204 if (DDT) 205 DDT->deleteEdge(ParentBB, DefaultDest); 206 continue; 207 } 208 209 // Otherwise, check to see if the switch only branches to one destination. 210 // We do this by reseting "TheOnlyDest" to null when we find two non-equal 211 // destinations. 212 if (i->getCaseSuccessor() != TheOnlyDest) 213 TheOnlyDest = nullptr; 214 215 // Increment this iterator as we haven't removed the case. 216 ++i; 217 } 218 219 if (CI && !TheOnlyDest) { 220 // Branching on a constant, but not any of the cases, go to the default 221 // successor. 222 TheOnlyDest = SI->getDefaultDest(); 223 } 224 225 // If we found a single destination that we can fold the switch into, do so 226 // now. 227 if (TheOnlyDest) { 228 // Insert the new branch. 229 Builder.CreateBr(TheOnlyDest); 230 BasicBlock *BB = SI->getParent(); 231 std::vector <DominatorTree::UpdateType> Updates; 232 if (DDT) 233 Updates.reserve(SI->getNumSuccessors() - 1); 234 235 // Remove entries from PHI nodes which we no longer branch to... 236 for (BasicBlock *Succ : SI->successors()) { 237 // Found case matching a constant operand? 238 if (Succ == TheOnlyDest) { 239 TheOnlyDest = nullptr; // Don't modify the first branch to TheOnlyDest 240 } else { 241 Succ->removePredecessor(BB); 242 if (DDT) 243 Updates.push_back({DominatorTree::Delete, BB, Succ}); 244 } 245 } 246 247 // Delete the old switch. 248 Value *Cond = SI->getCondition(); 249 SI->eraseFromParent(); 250 if (DeleteDeadConditions) 251 RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI); 252 if (DDT) 253 DDT->applyUpdates(Updates); 254 return true; 255 } 256 257 if (SI->getNumCases() == 1) { 258 // Otherwise, we can fold this switch into a conditional branch 259 // instruction if it has only one non-default destination. 260 auto FirstCase = *SI->case_begin(); 261 Value *Cond = Builder.CreateICmpEQ(SI->getCondition(), 262 FirstCase.getCaseValue(), "cond"); 263 264 // Insert the new branch. 265 BranchInst *NewBr = Builder.CreateCondBr(Cond, 266 FirstCase.getCaseSuccessor(), 267 SI->getDefaultDest()); 268 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof); 269 if (MD && MD->getNumOperands() == 3) { 270 ConstantInt *SICase = 271 mdconst::dyn_extract<ConstantInt>(MD->getOperand(2)); 272 ConstantInt *SIDef = 273 mdconst::dyn_extract<ConstantInt>(MD->getOperand(1)); 274 assert(SICase && SIDef); 275 // The TrueWeight should be the weight for the single case of SI. 276 NewBr->setMetadata(LLVMContext::MD_prof, 277 MDBuilder(BB->getContext()). 278 createBranchWeights(SICase->getValue().getZExtValue(), 279 SIDef->getValue().getZExtValue())); 280 } 281 282 // Update make.implicit metadata to the newly-created conditional branch. 283 MDNode *MakeImplicitMD = SI->getMetadata(LLVMContext::MD_make_implicit); 284 if (MakeImplicitMD) 285 NewBr->setMetadata(LLVMContext::MD_make_implicit, MakeImplicitMD); 286 287 // Delete the old switch. 288 SI->eraseFromParent(); 289 return true; 290 } 291 return false; 292 } 293 294 if (auto *IBI = dyn_cast<IndirectBrInst>(T)) { 295 // indirectbr blockaddress(@F, @BB) -> br label @BB 296 if (auto *BA = 297 dyn_cast<BlockAddress>(IBI->getAddress()->stripPointerCasts())) { 298 BasicBlock *TheOnlyDest = BA->getBasicBlock(); 299 std::vector <DominatorTree::UpdateType> Updates; 300 if (DDT) 301 Updates.reserve(IBI->getNumDestinations() - 1); 302 303 // Insert the new branch. 304 Builder.CreateBr(TheOnlyDest); 305 306 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 307 if (IBI->getDestination(i) == TheOnlyDest) { 308 TheOnlyDest = nullptr; 309 } else { 310 BasicBlock *ParentBB = IBI->getParent(); 311 BasicBlock *DestBB = IBI->getDestination(i); 312 DestBB->removePredecessor(ParentBB); 313 if (DDT) 314 Updates.push_back({DominatorTree::Delete, ParentBB, DestBB}); 315 } 316 } 317 Value *Address = IBI->getAddress(); 318 IBI->eraseFromParent(); 319 if (DeleteDeadConditions) 320 RecursivelyDeleteTriviallyDeadInstructions(Address, TLI); 321 322 // If we didn't find our destination in the IBI successor list, then we 323 // have undefined behavior. Replace the unconditional branch with an 324 // 'unreachable' instruction. 325 if (TheOnlyDest) { 326 BB->getTerminator()->eraseFromParent(); 327 new UnreachableInst(BB->getContext(), BB); 328 } 329 330 if (DDT) 331 DDT->applyUpdates(Updates); 332 return true; 333 } 334 } 335 336 return false; 337 } 338 339 //===----------------------------------------------------------------------===// 340 // Local dead code elimination. 341 // 342 343 /// isInstructionTriviallyDead - Return true if the result produced by the 344 /// instruction is not used, and the instruction has no side effects. 345 /// 346 bool llvm::isInstructionTriviallyDead(Instruction *I, 347 const TargetLibraryInfo *TLI) { 348 if (!I->use_empty()) 349 return false; 350 return wouldInstructionBeTriviallyDead(I, TLI); 351 } 352 353 bool llvm::wouldInstructionBeTriviallyDead(Instruction *I, 354 const TargetLibraryInfo *TLI) { 355 if (isa<TerminatorInst>(I)) 356 return false; 357 358 // We don't want the landingpad-like instructions removed by anything this 359 // general. 360 if (I->isEHPad()) 361 return false; 362 363 // We don't want debug info removed by anything this general, unless 364 // debug info is empty. 365 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(I)) { 366 if (DDI->getAddress()) 367 return false; 368 return true; 369 } 370 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(I)) { 371 if (DVI->getValue()) 372 return false; 373 return true; 374 } 375 if (DbgLabelInst *DLI = dyn_cast<DbgLabelInst>(I)) { 376 if (DLI->getLabel()) 377 return false; 378 return true; 379 } 380 381 if (!I->mayHaveSideEffects()) 382 return true; 383 384 // Special case intrinsics that "may have side effects" but can be deleted 385 // when dead. 386 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 387 // Safe to delete llvm.stacksave if dead. 388 if (II->getIntrinsicID() == Intrinsic::stacksave) 389 return true; 390 391 // Lifetime intrinsics are dead when their right-hand is undef. 392 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 393 II->getIntrinsicID() == Intrinsic::lifetime_end) 394 return isa<UndefValue>(II->getArgOperand(1)); 395 396 // Assumptions are dead if their condition is trivially true. Guards on 397 // true are operationally no-ops. In the future we can consider more 398 // sophisticated tradeoffs for guards considering potential for check 399 // widening, but for now we keep things simple. 400 if (II->getIntrinsicID() == Intrinsic::assume || 401 II->getIntrinsicID() == Intrinsic::experimental_guard) { 402 if (ConstantInt *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0))) 403 return !Cond->isZero(); 404 405 return false; 406 } 407 } 408 409 if (isAllocLikeFn(I, TLI)) 410 return true; 411 412 if (CallInst *CI = isFreeCall(I, TLI)) 413 if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0))) 414 return C->isNullValue() || isa<UndefValue>(C); 415 416 if (CallSite CS = CallSite(I)) 417 if (isMathLibCallNoop(CS, TLI)) 418 return true; 419 420 return false; 421 } 422 423 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a 424 /// trivially dead instruction, delete it. If that makes any of its operands 425 /// trivially dead, delete them too, recursively. Return true if any 426 /// instructions were deleted. 427 bool 428 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V, 429 const TargetLibraryInfo *TLI) { 430 Instruction *I = dyn_cast<Instruction>(V); 431 if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI)) 432 return false; 433 434 SmallVector<Instruction*, 16> DeadInsts; 435 DeadInsts.push_back(I); 436 437 do { 438 I = DeadInsts.pop_back_val(); 439 salvageDebugInfo(*I); 440 441 // Null out all of the instruction's operands to see if any operand becomes 442 // dead as we go. 443 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 444 Value *OpV = I->getOperand(i); 445 I->setOperand(i, nullptr); 446 447 if (!OpV->use_empty()) continue; 448 449 // If the operand is an instruction that became dead as we nulled out the 450 // operand, and if it is 'trivially' dead, delete it in a future loop 451 // iteration. 452 if (Instruction *OpI = dyn_cast<Instruction>(OpV)) 453 if (isInstructionTriviallyDead(OpI, TLI)) 454 DeadInsts.push_back(OpI); 455 } 456 457 I->eraseFromParent(); 458 } while (!DeadInsts.empty()); 459 460 return true; 461 } 462 463 /// areAllUsesEqual - Check whether the uses of a value are all the same. 464 /// This is similar to Instruction::hasOneUse() except this will also return 465 /// true when there are no uses or multiple uses that all refer to the same 466 /// value. 467 static bool areAllUsesEqual(Instruction *I) { 468 Value::user_iterator UI = I->user_begin(); 469 Value::user_iterator UE = I->user_end(); 470 if (UI == UE) 471 return true; 472 473 User *TheUse = *UI; 474 for (++UI; UI != UE; ++UI) { 475 if (*UI != TheUse) 476 return false; 477 } 478 return true; 479 } 480 481 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively 482 /// dead PHI node, due to being a def-use chain of single-use nodes that 483 /// either forms a cycle or is terminated by a trivially dead instruction, 484 /// delete it. If that makes any of its operands trivially dead, delete them 485 /// too, recursively. Return true if a change was made. 486 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN, 487 const TargetLibraryInfo *TLI) { 488 SmallPtrSet<Instruction*, 4> Visited; 489 for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects(); 490 I = cast<Instruction>(*I->user_begin())) { 491 if (I->use_empty()) 492 return RecursivelyDeleteTriviallyDeadInstructions(I, TLI); 493 494 // If we find an instruction more than once, we're on a cycle that 495 // won't prove fruitful. 496 if (!Visited.insert(I).second) { 497 // Break the cycle and delete the instruction and its operands. 498 I->replaceAllUsesWith(UndefValue::get(I->getType())); 499 (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI); 500 return true; 501 } 502 } 503 return false; 504 } 505 506 static bool 507 simplifyAndDCEInstruction(Instruction *I, 508 SmallSetVector<Instruction *, 16> &WorkList, 509 const DataLayout &DL, 510 const TargetLibraryInfo *TLI) { 511 if (isInstructionTriviallyDead(I, TLI)) { 512 salvageDebugInfo(*I); 513 514 // Null out all of the instruction's operands to see if any operand becomes 515 // dead as we go. 516 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 517 Value *OpV = I->getOperand(i); 518 I->setOperand(i, nullptr); 519 520 if (!OpV->use_empty() || I == OpV) 521 continue; 522 523 // If the operand is an instruction that became dead as we nulled out the 524 // operand, and if it is 'trivially' dead, delete it in a future loop 525 // iteration. 526 if (Instruction *OpI = dyn_cast<Instruction>(OpV)) 527 if (isInstructionTriviallyDead(OpI, TLI)) 528 WorkList.insert(OpI); 529 } 530 531 I->eraseFromParent(); 532 533 return true; 534 } 535 536 if (Value *SimpleV = SimplifyInstruction(I, DL)) { 537 // Add the users to the worklist. CAREFUL: an instruction can use itself, 538 // in the case of a phi node. 539 for (User *U : I->users()) { 540 if (U != I) { 541 WorkList.insert(cast<Instruction>(U)); 542 } 543 } 544 545 // Replace the instruction with its simplified value. 546 bool Changed = false; 547 if (!I->use_empty()) { 548 I->replaceAllUsesWith(SimpleV); 549 Changed = true; 550 } 551 if (isInstructionTriviallyDead(I, TLI)) { 552 I->eraseFromParent(); 553 Changed = true; 554 } 555 return Changed; 556 } 557 return false; 558 } 559 560 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to 561 /// simplify any instructions in it and recursively delete dead instructions. 562 /// 563 /// This returns true if it changed the code, note that it can delete 564 /// instructions in other blocks as well in this block. 565 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB, 566 const TargetLibraryInfo *TLI) { 567 bool MadeChange = false; 568 const DataLayout &DL = BB->getModule()->getDataLayout(); 569 570 #ifndef NDEBUG 571 // In debug builds, ensure that the terminator of the block is never replaced 572 // or deleted by these simplifications. The idea of simplification is that it 573 // cannot introduce new instructions, and there is no way to replace the 574 // terminator of a block without introducing a new instruction. 575 AssertingVH<Instruction> TerminatorVH(&BB->back()); 576 #endif 577 578 SmallSetVector<Instruction *, 16> WorkList; 579 // Iterate over the original function, only adding insts to the worklist 580 // if they actually need to be revisited. This avoids having to pre-init 581 // the worklist with the entire function's worth of instructions. 582 for (BasicBlock::iterator BI = BB->begin(), E = std::prev(BB->end()); 583 BI != E;) { 584 assert(!BI->isTerminator()); 585 Instruction *I = &*BI; 586 ++BI; 587 588 // We're visiting this instruction now, so make sure it's not in the 589 // worklist from an earlier visit. 590 if (!WorkList.count(I)) 591 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI); 592 } 593 594 while (!WorkList.empty()) { 595 Instruction *I = WorkList.pop_back_val(); 596 MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI); 597 } 598 return MadeChange; 599 } 600 601 //===----------------------------------------------------------------------===// 602 // Control Flow Graph Restructuring. 603 // 604 605 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this 606 /// method is called when we're about to delete Pred as a predecessor of BB. If 607 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred. 608 /// 609 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI 610 /// nodes that collapse into identity values. For example, if we have: 611 /// x = phi(1, 0, 0, 0) 612 /// y = and x, z 613 /// 614 /// .. and delete the predecessor corresponding to the '1', this will attempt to 615 /// recursively fold the and to 0. 616 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred, 617 DeferredDominance *DDT) { 618 // This only adjusts blocks with PHI nodes. 619 if (!isa<PHINode>(BB->begin())) 620 return; 621 622 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify 623 // them down. This will leave us with single entry phi nodes and other phis 624 // that can be removed. 625 BB->removePredecessor(Pred, true); 626 627 WeakTrackingVH PhiIt = &BB->front(); 628 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) { 629 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt)); 630 Value *OldPhiIt = PhiIt; 631 632 if (!recursivelySimplifyInstruction(PN)) 633 continue; 634 635 // If recursive simplification ended up deleting the next PHI node we would 636 // iterate to, then our iterator is invalid, restart scanning from the top 637 // of the block. 638 if (PhiIt != OldPhiIt) PhiIt = &BB->front(); 639 } 640 if (DDT) 641 DDT->deleteEdge(Pred, BB); 642 } 643 644 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its 645 /// predecessor is known to have one successor (DestBB!). Eliminate the edge 646 /// between them, moving the instructions in the predecessor into DestBB and 647 /// deleting the predecessor block. 648 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, DominatorTree *DT, 649 DeferredDominance *DDT) { 650 assert(!(DT && DDT) && "Cannot call with both DT and DDT."); 651 652 // If BB has single-entry PHI nodes, fold them. 653 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 654 Value *NewVal = PN->getIncomingValue(0); 655 // Replace self referencing PHI with undef, it must be dead. 656 if (NewVal == PN) NewVal = UndefValue::get(PN->getType()); 657 PN->replaceAllUsesWith(NewVal); 658 PN->eraseFromParent(); 659 } 660 661 BasicBlock *PredBB = DestBB->getSinglePredecessor(); 662 assert(PredBB && "Block doesn't have a single predecessor!"); 663 664 bool ReplaceEntryBB = false; 665 if (PredBB == &DestBB->getParent()->getEntryBlock()) 666 ReplaceEntryBB = true; 667 668 // Deferred DT update: Collect all the edges that enter PredBB. These 669 // dominator edges will be redirected to DestBB. 670 std::vector <DominatorTree::UpdateType> Updates; 671 if (DDT && !ReplaceEntryBB) { 672 Updates.reserve(1 + (2 * pred_size(PredBB))); 673 Updates.push_back({DominatorTree::Delete, PredBB, DestBB}); 674 for (auto I = pred_begin(PredBB), E = pred_end(PredBB); I != E; ++I) { 675 Updates.push_back({DominatorTree::Delete, *I, PredBB}); 676 // This predecessor of PredBB may already have DestBB as a successor. 677 if (llvm::find(successors(*I), DestBB) == succ_end(*I)) 678 Updates.push_back({DominatorTree::Insert, *I, DestBB}); 679 } 680 } 681 682 // Zap anything that took the address of DestBB. Not doing this will give the 683 // address an invalid value. 684 if (DestBB->hasAddressTaken()) { 685 BlockAddress *BA = BlockAddress::get(DestBB); 686 Constant *Replacement = 687 ConstantInt::get(Type::getInt32Ty(BA->getContext()), 1); 688 BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement, 689 BA->getType())); 690 BA->destroyConstant(); 691 } 692 693 // Anything that branched to PredBB now branches to DestBB. 694 PredBB->replaceAllUsesWith(DestBB); 695 696 // Splice all the instructions from PredBB to DestBB. 697 PredBB->getTerminator()->eraseFromParent(); 698 DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList()); 699 700 // If the PredBB is the entry block of the function, move DestBB up to 701 // become the entry block after we erase PredBB. 702 if (ReplaceEntryBB) 703 DestBB->moveAfter(PredBB); 704 705 if (DT) { 706 // For some irreducible CFG we end up having forward-unreachable blocks 707 // so check if getNode returns a valid node before updating the domtree. 708 if (DomTreeNode *DTN = DT->getNode(PredBB)) { 709 BasicBlock *PredBBIDom = DTN->getIDom()->getBlock(); 710 DT->changeImmediateDominator(DestBB, PredBBIDom); 711 DT->eraseNode(PredBB); 712 } 713 } 714 715 if (DDT) { 716 DDT->deleteBB(PredBB); // Deferred deletion of BB. 717 if (ReplaceEntryBB) 718 // The entry block was removed and there is no external interface for the 719 // dominator tree to be notified of this change. In this corner-case we 720 // recalculate the entire tree. 721 DDT->recalculate(*(DestBB->getParent())); 722 else 723 DDT->applyUpdates(Updates); 724 } else { 725 PredBB->eraseFromParent(); // Nuke BB. 726 } 727 } 728 729 /// CanMergeValues - Return true if we can choose one of these values to use 730 /// in place of the other. Note that we will always choose the non-undef 731 /// value to keep. 732 static bool CanMergeValues(Value *First, Value *Second) { 733 return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second); 734 } 735 736 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an 737 /// almost-empty BB ending in an unconditional branch to Succ, into Succ. 738 /// 739 /// Assumption: Succ is the single successor for BB. 740 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { 741 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!"); 742 743 LLVM_DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into " 744 << Succ->getName() << "\n"); 745 // Shortcut, if there is only a single predecessor it must be BB and merging 746 // is always safe 747 if (Succ->getSinglePredecessor()) return true; 748 749 // Make a list of the predecessors of BB 750 SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB)); 751 752 // Look at all the phi nodes in Succ, to see if they present a conflict when 753 // merging these blocks 754 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 755 PHINode *PN = cast<PHINode>(I); 756 757 // If the incoming value from BB is again a PHINode in 758 // BB which has the same incoming value for *PI as PN does, we can 759 // merge the phi nodes and then the blocks can still be merged 760 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB)); 761 if (BBPN && BBPN->getParent() == BB) { 762 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) { 763 BasicBlock *IBB = PN->getIncomingBlock(PI); 764 if (BBPreds.count(IBB) && 765 !CanMergeValues(BBPN->getIncomingValueForBlock(IBB), 766 PN->getIncomingValue(PI))) { 767 LLVM_DEBUG(dbgs() 768 << "Can't fold, phi node " << PN->getName() << " in " 769 << Succ->getName() << " is conflicting with " 770 << BBPN->getName() << " with regard to common predecessor " 771 << IBB->getName() << "\n"); 772 return false; 773 } 774 } 775 } else { 776 Value* Val = PN->getIncomingValueForBlock(BB); 777 for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) { 778 // See if the incoming value for the common predecessor is equal to the 779 // one for BB, in which case this phi node will not prevent the merging 780 // of the block. 781 BasicBlock *IBB = PN->getIncomingBlock(PI); 782 if (BBPreds.count(IBB) && 783 !CanMergeValues(Val, PN->getIncomingValue(PI))) { 784 LLVM_DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() 785 << " in " << Succ->getName() 786 << " is conflicting with regard to common " 787 << "predecessor " << IBB->getName() << "\n"); 788 return false; 789 } 790 } 791 } 792 } 793 794 return true; 795 } 796 797 using PredBlockVector = SmallVector<BasicBlock *, 16>; 798 using IncomingValueMap = DenseMap<BasicBlock *, Value *>; 799 800 /// Determines the value to use as the phi node input for a block. 801 /// 802 /// Select between \p OldVal any value that we know flows from \p BB 803 /// to a particular phi on the basis of which one (if either) is not 804 /// undef. Update IncomingValues based on the selected value. 805 /// 806 /// \param OldVal The value we are considering selecting. 807 /// \param BB The block that the value flows in from. 808 /// \param IncomingValues A map from block-to-value for other phi inputs 809 /// that we have examined. 810 /// 811 /// \returns the selected value. 812 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB, 813 IncomingValueMap &IncomingValues) { 814 if (!isa<UndefValue>(OldVal)) { 815 assert((!IncomingValues.count(BB) || 816 IncomingValues.find(BB)->second == OldVal) && 817 "Expected OldVal to match incoming value from BB!"); 818 819 IncomingValues.insert(std::make_pair(BB, OldVal)); 820 return OldVal; 821 } 822 823 IncomingValueMap::const_iterator It = IncomingValues.find(BB); 824 if (It != IncomingValues.end()) return It->second; 825 826 return OldVal; 827 } 828 829 /// Create a map from block to value for the operands of a 830 /// given phi. 831 /// 832 /// Create a map from block to value for each non-undef value flowing 833 /// into \p PN. 834 /// 835 /// \param PN The phi we are collecting the map for. 836 /// \param IncomingValues [out] The map from block to value for this phi. 837 static void gatherIncomingValuesToPhi(PHINode *PN, 838 IncomingValueMap &IncomingValues) { 839 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 840 BasicBlock *BB = PN->getIncomingBlock(i); 841 Value *V = PN->getIncomingValue(i); 842 843 if (!isa<UndefValue>(V)) 844 IncomingValues.insert(std::make_pair(BB, V)); 845 } 846 } 847 848 /// Replace the incoming undef values to a phi with the values 849 /// from a block-to-value map. 850 /// 851 /// \param PN The phi we are replacing the undefs in. 852 /// \param IncomingValues A map from block to value. 853 static void replaceUndefValuesInPhi(PHINode *PN, 854 const IncomingValueMap &IncomingValues) { 855 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 856 Value *V = PN->getIncomingValue(i); 857 858 if (!isa<UndefValue>(V)) continue; 859 860 BasicBlock *BB = PN->getIncomingBlock(i); 861 IncomingValueMap::const_iterator It = IncomingValues.find(BB); 862 if (It == IncomingValues.end()) continue; 863 864 PN->setIncomingValue(i, It->second); 865 } 866 } 867 868 /// Replace a value flowing from a block to a phi with 869 /// potentially multiple instances of that value flowing from the 870 /// block's predecessors to the phi. 871 /// 872 /// \param BB The block with the value flowing into the phi. 873 /// \param BBPreds The predecessors of BB. 874 /// \param PN The phi that we are updating. 875 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB, 876 const PredBlockVector &BBPreds, 877 PHINode *PN) { 878 Value *OldVal = PN->removeIncomingValue(BB, false); 879 assert(OldVal && "No entry in PHI for Pred BB!"); 880 881 IncomingValueMap IncomingValues; 882 883 // We are merging two blocks - BB, and the block containing PN - and 884 // as a result we need to redirect edges from the predecessors of BB 885 // to go to the block containing PN, and update PN 886 // accordingly. Since we allow merging blocks in the case where the 887 // predecessor and successor blocks both share some predecessors, 888 // and where some of those common predecessors might have undef 889 // values flowing into PN, we want to rewrite those values to be 890 // consistent with the non-undef values. 891 892 gatherIncomingValuesToPhi(PN, IncomingValues); 893 894 // If this incoming value is one of the PHI nodes in BB, the new entries 895 // in the PHI node are the entries from the old PHI. 896 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) { 897 PHINode *OldValPN = cast<PHINode>(OldVal); 898 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) { 899 // Note that, since we are merging phi nodes and BB and Succ might 900 // have common predecessors, we could end up with a phi node with 901 // identical incoming branches. This will be cleaned up later (and 902 // will trigger asserts if we try to clean it up now, without also 903 // simplifying the corresponding conditional branch). 904 BasicBlock *PredBB = OldValPN->getIncomingBlock(i); 905 Value *PredVal = OldValPN->getIncomingValue(i); 906 Value *Selected = selectIncomingValueForBlock(PredVal, PredBB, 907 IncomingValues); 908 909 // And add a new incoming value for this predecessor for the 910 // newly retargeted branch. 911 PN->addIncoming(Selected, PredBB); 912 } 913 } else { 914 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) { 915 // Update existing incoming values in PN for this 916 // predecessor of BB. 917 BasicBlock *PredBB = BBPreds[i]; 918 Value *Selected = selectIncomingValueForBlock(OldVal, PredBB, 919 IncomingValues); 920 921 // And add a new incoming value for this predecessor for the 922 // newly retargeted branch. 923 PN->addIncoming(Selected, PredBB); 924 } 925 } 926 927 replaceUndefValuesInPhi(PN, IncomingValues); 928 } 929 930 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an 931 /// unconditional branch, and contains no instructions other than PHI nodes, 932 /// potential side-effect free intrinsics and the branch. If possible, 933 /// eliminate BB by rewriting all the predecessors to branch to the successor 934 /// block and return true. If we can't transform, return false. 935 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, 936 DeferredDominance *DDT) { 937 assert(BB != &BB->getParent()->getEntryBlock() && 938 "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!"); 939 940 // We can't eliminate infinite loops. 941 BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0); 942 if (BB == Succ) return false; 943 944 // Check to see if merging these blocks would cause conflicts for any of the 945 // phi nodes in BB or Succ. If not, we can safely merge. 946 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false; 947 948 // Check for cases where Succ has multiple predecessors and a PHI node in BB 949 // has uses which will not disappear when the PHI nodes are merged. It is 950 // possible to handle such cases, but difficult: it requires checking whether 951 // BB dominates Succ, which is non-trivial to calculate in the case where 952 // Succ has multiple predecessors. Also, it requires checking whether 953 // constructing the necessary self-referential PHI node doesn't introduce any 954 // conflicts; this isn't too difficult, but the previous code for doing this 955 // was incorrect. 956 // 957 // Note that if this check finds a live use, BB dominates Succ, so BB is 958 // something like a loop pre-header (or rarely, a part of an irreducible CFG); 959 // folding the branch isn't profitable in that case anyway. 960 if (!Succ->getSinglePredecessor()) { 961 BasicBlock::iterator BBI = BB->begin(); 962 while (isa<PHINode>(*BBI)) { 963 for (Use &U : BBI->uses()) { 964 if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) { 965 if (PN->getIncomingBlock(U) != BB) 966 return false; 967 } else { 968 return false; 969 } 970 } 971 ++BBI; 972 } 973 } 974 975 LLVM_DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB); 976 977 std::vector<DominatorTree::UpdateType> Updates; 978 if (DDT) { 979 Updates.reserve(1 + (2 * pred_size(BB))); 980 Updates.push_back({DominatorTree::Delete, BB, Succ}); 981 // All predecessors of BB will be moved to Succ. 982 for (auto I = pred_begin(BB), E = pred_end(BB); I != E; ++I) { 983 Updates.push_back({DominatorTree::Delete, *I, BB}); 984 // This predecessor of BB may already have Succ as a successor. 985 if (llvm::find(successors(*I), Succ) == succ_end(*I)) 986 Updates.push_back({DominatorTree::Insert, *I, Succ}); 987 } 988 } 989 990 if (isa<PHINode>(Succ->begin())) { 991 // If there is more than one pred of succ, and there are PHI nodes in 992 // the successor, then we need to add incoming edges for the PHI nodes 993 // 994 const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB)); 995 996 // Loop over all of the PHI nodes in the successor of BB. 997 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 998 PHINode *PN = cast<PHINode>(I); 999 1000 redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN); 1001 } 1002 } 1003 1004 if (Succ->getSinglePredecessor()) { 1005 // BB is the only predecessor of Succ, so Succ will end up with exactly 1006 // the same predecessors BB had. 1007 1008 // Copy over any phi, debug or lifetime instruction. 1009 BB->getTerminator()->eraseFromParent(); 1010 Succ->getInstList().splice(Succ->getFirstNonPHI()->getIterator(), 1011 BB->getInstList()); 1012 } else { 1013 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) { 1014 // We explicitly check for such uses in CanPropagatePredecessorsForPHIs. 1015 assert(PN->use_empty() && "There shouldn't be any uses here!"); 1016 PN->eraseFromParent(); 1017 } 1018 } 1019 1020 // If the unconditional branch we replaced contains llvm.loop metadata, we 1021 // add the metadata to the branch instructions in the predecessors. 1022 unsigned LoopMDKind = BB->getContext().getMDKindID("llvm.loop"); 1023 Instruction *TI = BB->getTerminator(); 1024 if (TI) 1025 if (MDNode *LoopMD = TI->getMetadata(LoopMDKind)) 1026 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1027 BasicBlock *Pred = *PI; 1028 Pred->getTerminator()->setMetadata(LoopMDKind, LoopMD); 1029 } 1030 1031 // Everything that jumped to BB now goes to Succ. 1032 BB->replaceAllUsesWith(Succ); 1033 if (!Succ->hasName()) Succ->takeName(BB); 1034 1035 if (DDT) { 1036 DDT->deleteBB(BB); // Deferred deletion of the old basic block. 1037 DDT->applyUpdates(Updates); 1038 } else { 1039 BB->eraseFromParent(); // Delete the old basic block. 1040 } 1041 return true; 1042 } 1043 1044 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI 1045 /// nodes in this block. This doesn't try to be clever about PHI nodes 1046 /// which differ only in the order of the incoming values, but instcombine 1047 /// orders them so it usually won't matter. 1048 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) { 1049 // This implementation doesn't currently consider undef operands 1050 // specially. Theoretically, two phis which are identical except for 1051 // one having an undef where the other doesn't could be collapsed. 1052 1053 struct PHIDenseMapInfo { 1054 static PHINode *getEmptyKey() { 1055 return DenseMapInfo<PHINode *>::getEmptyKey(); 1056 } 1057 1058 static PHINode *getTombstoneKey() { 1059 return DenseMapInfo<PHINode *>::getTombstoneKey(); 1060 } 1061 1062 static unsigned getHashValue(PHINode *PN) { 1063 // Compute a hash value on the operands. Instcombine will likely have 1064 // sorted them, which helps expose duplicates, but we have to check all 1065 // the operands to be safe in case instcombine hasn't run. 1066 return static_cast<unsigned>(hash_combine( 1067 hash_combine_range(PN->value_op_begin(), PN->value_op_end()), 1068 hash_combine_range(PN->block_begin(), PN->block_end()))); 1069 } 1070 1071 static bool isEqual(PHINode *LHS, PHINode *RHS) { 1072 if (LHS == getEmptyKey() || LHS == getTombstoneKey() || 1073 RHS == getEmptyKey() || RHS == getTombstoneKey()) 1074 return LHS == RHS; 1075 return LHS->isIdenticalTo(RHS); 1076 } 1077 }; 1078 1079 // Set of unique PHINodes. 1080 DenseSet<PHINode *, PHIDenseMapInfo> PHISet; 1081 1082 // Examine each PHI. 1083 bool Changed = false; 1084 for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I++);) { 1085 auto Inserted = PHISet.insert(PN); 1086 if (!Inserted.second) { 1087 // A duplicate. Replace this PHI with its duplicate. 1088 PN->replaceAllUsesWith(*Inserted.first); 1089 PN->eraseFromParent(); 1090 Changed = true; 1091 1092 // The RAUW can change PHIs that we already visited. Start over from the 1093 // beginning. 1094 PHISet.clear(); 1095 I = BB->begin(); 1096 } 1097 } 1098 1099 return Changed; 1100 } 1101 1102 /// enforceKnownAlignment - If the specified pointer points to an object that 1103 /// we control, modify the object's alignment to PrefAlign. This isn't 1104 /// often possible though. If alignment is important, a more reliable approach 1105 /// is to simply align all global variables and allocation instructions to 1106 /// their preferred alignment from the beginning. 1107 static unsigned enforceKnownAlignment(Value *V, unsigned Align, 1108 unsigned PrefAlign, 1109 const DataLayout &DL) { 1110 assert(PrefAlign > Align); 1111 1112 V = V->stripPointerCasts(); 1113 1114 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 1115 // TODO: ideally, computeKnownBits ought to have used 1116 // AllocaInst::getAlignment() in its computation already, making 1117 // the below max redundant. But, as it turns out, 1118 // stripPointerCasts recurses through infinite layers of bitcasts, 1119 // while computeKnownBits is not allowed to traverse more than 6 1120 // levels. 1121 Align = std::max(AI->getAlignment(), Align); 1122 if (PrefAlign <= Align) 1123 return Align; 1124 1125 // If the preferred alignment is greater than the natural stack alignment 1126 // then don't round up. This avoids dynamic stack realignment. 1127 if (DL.exceedsNaturalStackAlignment(PrefAlign)) 1128 return Align; 1129 AI->setAlignment(PrefAlign); 1130 return PrefAlign; 1131 } 1132 1133 if (auto *GO = dyn_cast<GlobalObject>(V)) { 1134 // TODO: as above, this shouldn't be necessary. 1135 Align = std::max(GO->getAlignment(), Align); 1136 if (PrefAlign <= Align) 1137 return Align; 1138 1139 // If there is a large requested alignment and we can, bump up the alignment 1140 // of the global. If the memory we set aside for the global may not be the 1141 // memory used by the final program then it is impossible for us to reliably 1142 // enforce the preferred alignment. 1143 if (!GO->canIncreaseAlignment()) 1144 return Align; 1145 1146 GO->setAlignment(PrefAlign); 1147 return PrefAlign; 1148 } 1149 1150 return Align; 1151 } 1152 1153 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign, 1154 const DataLayout &DL, 1155 const Instruction *CxtI, 1156 AssumptionCache *AC, 1157 const DominatorTree *DT) { 1158 assert(V->getType()->isPointerTy() && 1159 "getOrEnforceKnownAlignment expects a pointer!"); 1160 1161 KnownBits Known = computeKnownBits(V, DL, 0, AC, CxtI, DT); 1162 unsigned TrailZ = Known.countMinTrailingZeros(); 1163 1164 // Avoid trouble with ridiculously large TrailZ values, such as 1165 // those computed from a null pointer. 1166 TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1)); 1167 1168 unsigned Align = 1u << std::min(Known.getBitWidth() - 1, TrailZ); 1169 1170 // LLVM doesn't support alignments larger than this currently. 1171 Align = std::min(Align, +Value::MaximumAlignment); 1172 1173 if (PrefAlign > Align) 1174 Align = enforceKnownAlignment(V, Align, PrefAlign, DL); 1175 1176 // We don't need to make any adjustment. 1177 return Align; 1178 } 1179 1180 ///===---------------------------------------------------------------------===// 1181 /// Dbg Intrinsic utilities 1182 /// 1183 1184 /// See if there is a dbg.value intrinsic for DIVar before I. 1185 static bool LdStHasDebugValue(DILocalVariable *DIVar, DIExpression *DIExpr, 1186 Instruction *I) { 1187 // Since we can't guarantee that the original dbg.declare instrinsic 1188 // is removed by LowerDbgDeclare(), we need to make sure that we are 1189 // not inserting the same dbg.value intrinsic over and over. 1190 BasicBlock::InstListType::iterator PrevI(I); 1191 if (PrevI != I->getParent()->getInstList().begin()) { 1192 --PrevI; 1193 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI)) 1194 if (DVI->getValue() == I->getOperand(0) && 1195 DVI->getVariable() == DIVar && 1196 DVI->getExpression() == DIExpr) 1197 return true; 1198 } 1199 return false; 1200 } 1201 1202 /// See if there is a dbg.value intrinsic for DIVar for the PHI node. 1203 static bool PhiHasDebugValue(DILocalVariable *DIVar, 1204 DIExpression *DIExpr, 1205 PHINode *APN) { 1206 // Since we can't guarantee that the original dbg.declare instrinsic 1207 // is removed by LowerDbgDeclare(), we need to make sure that we are 1208 // not inserting the same dbg.value intrinsic over and over. 1209 SmallVector<DbgValueInst *, 1> DbgValues; 1210 findDbgValues(DbgValues, APN); 1211 for (auto *DVI : DbgValues) { 1212 assert(DVI->getValue() == APN); 1213 if ((DVI->getVariable() == DIVar) && (DVI->getExpression() == DIExpr)) 1214 return true; 1215 } 1216 return false; 1217 } 1218 1219 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value 1220 /// that has an associated llvm.dbg.declare or llvm.dbg.addr intrinsic. 1221 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII, 1222 StoreInst *SI, DIBuilder &Builder) { 1223 assert(DII->isAddressOfVariable()); 1224 auto *DIVar = DII->getVariable(); 1225 assert(DIVar && "Missing variable"); 1226 auto *DIExpr = DII->getExpression(); 1227 Value *DV = SI->getOperand(0); 1228 1229 // If an argument is zero extended then use argument directly. The ZExt 1230 // may be zapped by an optimization pass in future. 1231 Argument *ExtendedArg = nullptr; 1232 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 1233 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0)); 1234 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 1235 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0)); 1236 if (ExtendedArg) { 1237 // If this DII was already describing only a fragment of a variable, ensure 1238 // that fragment is appropriately narrowed here. 1239 // But if a fragment wasn't used, describe the value as the original 1240 // argument (rather than the zext or sext) so that it remains described even 1241 // if the sext/zext is optimized away. This widens the variable description, 1242 // leaving it up to the consumer to know how the smaller value may be 1243 // represented in a larger register. 1244 if (auto Fragment = DIExpr->getFragmentInfo()) { 1245 unsigned FragmentOffset = Fragment->OffsetInBits; 1246 SmallVector<uint64_t, 3> Ops(DIExpr->elements_begin(), 1247 DIExpr->elements_end() - 3); 1248 Ops.push_back(dwarf::DW_OP_LLVM_fragment); 1249 Ops.push_back(FragmentOffset); 1250 const DataLayout &DL = DII->getModule()->getDataLayout(); 1251 Ops.push_back(DL.getTypeSizeInBits(ExtendedArg->getType())); 1252 DIExpr = Builder.createExpression(Ops); 1253 } 1254 DV = ExtendedArg; 1255 } 1256 if (!LdStHasDebugValue(DIVar, DIExpr, SI)) 1257 Builder.insertDbgValueIntrinsic(DV, DIVar, DIExpr, DII->getDebugLoc(), 1258 SI); 1259 } 1260 1261 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value 1262 /// that has an associated llvm.dbg.declare or llvm.dbg.addr intrinsic. 1263 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII, 1264 LoadInst *LI, DIBuilder &Builder) { 1265 auto *DIVar = DII->getVariable(); 1266 auto *DIExpr = DII->getExpression(); 1267 assert(DIVar && "Missing variable"); 1268 1269 if (LdStHasDebugValue(DIVar, DIExpr, LI)) 1270 return; 1271 1272 // We are now tracking the loaded value instead of the address. In the 1273 // future if multi-location support is added to the IR, it might be 1274 // preferable to keep tracking both the loaded value and the original 1275 // address in case the alloca can not be elided. 1276 Instruction *DbgValue = Builder.insertDbgValueIntrinsic( 1277 LI, DIVar, DIExpr, DII->getDebugLoc(), (Instruction *)nullptr); 1278 DbgValue->insertAfter(LI); 1279 } 1280 1281 /// Inserts a llvm.dbg.value intrinsic after a phi that has an associated 1282 /// llvm.dbg.declare or llvm.dbg.addr intrinsic. 1283 void llvm::ConvertDebugDeclareToDebugValue(DbgInfoIntrinsic *DII, 1284 PHINode *APN, DIBuilder &Builder) { 1285 auto *DIVar = DII->getVariable(); 1286 auto *DIExpr = DII->getExpression(); 1287 assert(DIVar && "Missing variable"); 1288 1289 if (PhiHasDebugValue(DIVar, DIExpr, APN)) 1290 return; 1291 1292 BasicBlock *BB = APN->getParent(); 1293 auto InsertionPt = BB->getFirstInsertionPt(); 1294 1295 // The block may be a catchswitch block, which does not have a valid 1296 // insertion point. 1297 // FIXME: Insert dbg.value markers in the successors when appropriate. 1298 if (InsertionPt != BB->end()) 1299 Builder.insertDbgValueIntrinsic(APN, DIVar, DIExpr, DII->getDebugLoc(), 1300 &*InsertionPt); 1301 } 1302 1303 /// Determine whether this alloca is either a VLA or an array. 1304 static bool isArray(AllocaInst *AI) { 1305 return AI->isArrayAllocation() || 1306 AI->getType()->getElementType()->isArrayTy(); 1307 } 1308 1309 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set 1310 /// of llvm.dbg.value intrinsics. 1311 bool llvm::LowerDbgDeclare(Function &F) { 1312 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false); 1313 SmallVector<DbgDeclareInst *, 4> Dbgs; 1314 for (auto &FI : F) 1315 for (Instruction &BI : FI) 1316 if (auto DDI = dyn_cast<DbgDeclareInst>(&BI)) 1317 Dbgs.push_back(DDI); 1318 1319 if (Dbgs.empty()) 1320 return false; 1321 1322 for (auto &I : Dbgs) { 1323 DbgDeclareInst *DDI = I; 1324 AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress()); 1325 // If this is an alloca for a scalar variable, insert a dbg.value 1326 // at each load and store to the alloca and erase the dbg.declare. 1327 // The dbg.values allow tracking a variable even if it is not 1328 // stored on the stack, while the dbg.declare can only describe 1329 // the stack slot (and at a lexical-scope granularity). Later 1330 // passes will attempt to elide the stack slot. 1331 if (!AI || isArray(AI)) 1332 continue; 1333 1334 // A volatile load/store means that the alloca can't be elided anyway. 1335 if (llvm::any_of(AI->users(), [](User *U) -> bool { 1336 if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1337 return LI->isVolatile(); 1338 if (StoreInst *SI = dyn_cast<StoreInst>(U)) 1339 return SI->isVolatile(); 1340 return false; 1341 })) 1342 continue; 1343 1344 for (auto &AIUse : AI->uses()) { 1345 User *U = AIUse.getUser(); 1346 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1347 if (AIUse.getOperandNo() == 1) 1348 ConvertDebugDeclareToDebugValue(DDI, SI, DIB); 1349 } else if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1350 ConvertDebugDeclareToDebugValue(DDI, LI, DIB); 1351 } else if (CallInst *CI = dyn_cast<CallInst>(U)) { 1352 // This is a call by-value or some other instruction that 1353 // takes a pointer to the variable. Insert a *value* 1354 // intrinsic that describes the alloca. 1355 DIB.insertDbgValueIntrinsic(AI, DDI->getVariable(), 1356 DDI->getExpression(), DDI->getDebugLoc(), 1357 CI); 1358 } 1359 } 1360 DDI->eraseFromParent(); 1361 } 1362 return true; 1363 } 1364 1365 /// Propagate dbg.value intrinsics through the newly inserted PHIs. 1366 void llvm::insertDebugValuesForPHIs(BasicBlock *BB, 1367 SmallVectorImpl<PHINode *> &InsertedPHIs) { 1368 assert(BB && "No BasicBlock to clone dbg.value(s) from."); 1369 if (InsertedPHIs.size() == 0) 1370 return; 1371 1372 // Map existing PHI nodes to their dbg.values. 1373 ValueToValueMapTy DbgValueMap; 1374 for (auto &I : *BB) { 1375 if (auto DbgII = dyn_cast<DbgInfoIntrinsic>(&I)) { 1376 if (auto *Loc = dyn_cast_or_null<PHINode>(DbgII->getVariableLocation())) 1377 DbgValueMap.insert({Loc, DbgII}); 1378 } 1379 } 1380 if (DbgValueMap.size() == 0) 1381 return; 1382 1383 // Then iterate through the new PHIs and look to see if they use one of the 1384 // previously mapped PHIs. If so, insert a new dbg.value intrinsic that will 1385 // propagate the info through the new PHI. 1386 LLVMContext &C = BB->getContext(); 1387 for (auto PHI : InsertedPHIs) { 1388 BasicBlock *Parent = PHI->getParent(); 1389 // Avoid inserting an intrinsic into an EH block. 1390 if (Parent->getFirstNonPHI()->isEHPad()) 1391 continue; 1392 auto PhiMAV = MetadataAsValue::get(C, ValueAsMetadata::get(PHI)); 1393 for (auto VI : PHI->operand_values()) { 1394 auto V = DbgValueMap.find(VI); 1395 if (V != DbgValueMap.end()) { 1396 auto *DbgII = cast<DbgInfoIntrinsic>(V->second); 1397 Instruction *NewDbgII = DbgII->clone(); 1398 NewDbgII->setOperand(0, PhiMAV); 1399 auto InsertionPt = Parent->getFirstInsertionPt(); 1400 assert(InsertionPt != Parent->end() && "Ill-formed basic block"); 1401 NewDbgII->insertBefore(&*InsertionPt); 1402 } 1403 } 1404 } 1405 } 1406 1407 /// Finds all intrinsics declaring local variables as living in the memory that 1408 /// 'V' points to. This may include a mix of dbg.declare and 1409 /// dbg.addr intrinsics. 1410 TinyPtrVector<DbgInfoIntrinsic *> llvm::FindDbgAddrUses(Value *V) { 1411 auto *L = LocalAsMetadata::getIfExists(V); 1412 if (!L) 1413 return {}; 1414 auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L); 1415 if (!MDV) 1416 return {}; 1417 1418 TinyPtrVector<DbgInfoIntrinsic *> Declares; 1419 for (User *U : MDV->users()) { 1420 if (auto *DII = dyn_cast<DbgInfoIntrinsic>(U)) 1421 if (DII->isAddressOfVariable()) 1422 Declares.push_back(DII); 1423 } 1424 1425 return Declares; 1426 } 1427 1428 void llvm::findDbgValues(SmallVectorImpl<DbgValueInst *> &DbgValues, Value *V) { 1429 if (auto *L = LocalAsMetadata::getIfExists(V)) 1430 if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L)) 1431 for (User *U : MDV->users()) 1432 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U)) 1433 DbgValues.push_back(DVI); 1434 } 1435 1436 void llvm::findDbgUsers(SmallVectorImpl<DbgInfoIntrinsic *> &DbgUsers, 1437 Value *V) { 1438 if (auto *L = LocalAsMetadata::getIfExists(V)) 1439 if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L)) 1440 for (User *U : MDV->users()) 1441 if (DbgInfoIntrinsic *DII = dyn_cast<DbgInfoIntrinsic>(U)) 1442 DbgUsers.push_back(DII); 1443 } 1444 1445 bool llvm::replaceDbgDeclare(Value *Address, Value *NewAddress, 1446 Instruction *InsertBefore, DIBuilder &Builder, 1447 bool DerefBefore, int Offset, bool DerefAfter) { 1448 auto DbgAddrs = FindDbgAddrUses(Address); 1449 for (DbgInfoIntrinsic *DII : DbgAddrs) { 1450 DebugLoc Loc = DII->getDebugLoc(); 1451 auto *DIVar = DII->getVariable(); 1452 auto *DIExpr = DII->getExpression(); 1453 assert(DIVar && "Missing variable"); 1454 DIExpr = DIExpression::prepend(DIExpr, DerefBefore, Offset, DerefAfter); 1455 // Insert llvm.dbg.declare immediately after InsertBefore, and remove old 1456 // llvm.dbg.declare. 1457 Builder.insertDeclare(NewAddress, DIVar, DIExpr, Loc, InsertBefore); 1458 if (DII == InsertBefore) 1459 InsertBefore = &*std::next(InsertBefore->getIterator()); 1460 DII->eraseFromParent(); 1461 } 1462 return !DbgAddrs.empty(); 1463 } 1464 1465 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress, 1466 DIBuilder &Builder, bool DerefBefore, 1467 int Offset, bool DerefAfter) { 1468 return replaceDbgDeclare(AI, NewAllocaAddress, AI->getNextNode(), Builder, 1469 DerefBefore, Offset, DerefAfter); 1470 } 1471 1472 static void replaceOneDbgValueForAlloca(DbgValueInst *DVI, Value *NewAddress, 1473 DIBuilder &Builder, int Offset) { 1474 DebugLoc Loc = DVI->getDebugLoc(); 1475 auto *DIVar = DVI->getVariable(); 1476 auto *DIExpr = DVI->getExpression(); 1477 assert(DIVar && "Missing variable"); 1478 1479 // This is an alloca-based llvm.dbg.value. The first thing it should do with 1480 // the alloca pointer is dereference it. Otherwise we don't know how to handle 1481 // it and give up. 1482 if (!DIExpr || DIExpr->getNumElements() < 1 || 1483 DIExpr->getElement(0) != dwarf::DW_OP_deref) 1484 return; 1485 1486 // Insert the offset immediately after the first deref. 1487 // We could just change the offset argument of dbg.value, but it's unsigned... 1488 if (Offset) { 1489 SmallVector<uint64_t, 4> Ops; 1490 Ops.push_back(dwarf::DW_OP_deref); 1491 DIExpression::appendOffset(Ops, Offset); 1492 Ops.append(DIExpr->elements_begin() + 1, DIExpr->elements_end()); 1493 DIExpr = Builder.createExpression(Ops); 1494 } 1495 1496 Builder.insertDbgValueIntrinsic(NewAddress, DIVar, DIExpr, Loc, DVI); 1497 DVI->eraseFromParent(); 1498 } 1499 1500 void llvm::replaceDbgValueForAlloca(AllocaInst *AI, Value *NewAllocaAddress, 1501 DIBuilder &Builder, int Offset) { 1502 if (auto *L = LocalAsMetadata::getIfExists(AI)) 1503 if (auto *MDV = MetadataAsValue::getIfExists(AI->getContext(), L)) 1504 for (auto UI = MDV->use_begin(), UE = MDV->use_end(); UI != UE;) { 1505 Use &U = *UI++; 1506 if (auto *DVI = dyn_cast<DbgValueInst>(U.getUser())) 1507 replaceOneDbgValueForAlloca(DVI, NewAllocaAddress, Builder, Offset); 1508 } 1509 } 1510 1511 void llvm::salvageDebugInfo(Instruction &I) { 1512 // This function is hot. An early check to determine whether the instruction 1513 // has any metadata to save allows it to return earlier on average. 1514 if (!I.isUsedByMetadata()) 1515 return; 1516 1517 SmallVector<DbgInfoIntrinsic *, 1> DbgUsers; 1518 findDbgUsers(DbgUsers, &I); 1519 if (DbgUsers.empty()) 1520 return; 1521 1522 auto &M = *I.getModule(); 1523 auto &DL = M.getDataLayout(); 1524 1525 auto wrapMD = [&](Value *V) { 1526 return MetadataAsValue::get(I.getContext(), ValueAsMetadata::get(V)); 1527 }; 1528 1529 auto doSalvage = [&](DbgInfoIntrinsic *DII, SmallVectorImpl<uint64_t> &Ops) { 1530 auto *DIExpr = DII->getExpression(); 1531 DIExpr = 1532 DIExpression::prependOpcodes(DIExpr, Ops, DIExpression::WithStackValue); 1533 DII->setOperand(0, wrapMD(I.getOperand(0))); 1534 DII->setOperand(2, MetadataAsValue::get(I.getContext(), DIExpr)); 1535 LLVM_DEBUG(dbgs() << "SALVAGE: " << *DII << '\n'); 1536 }; 1537 1538 auto applyOffset = [&](DbgInfoIntrinsic *DII, uint64_t Offset) { 1539 SmallVector<uint64_t, 8> Ops; 1540 DIExpression::appendOffset(Ops, Offset); 1541 doSalvage(DII, Ops); 1542 }; 1543 1544 auto applyOps = [&](DbgInfoIntrinsic *DII, 1545 std::initializer_list<uint64_t> Opcodes) { 1546 SmallVector<uint64_t, 8> Ops(Opcodes); 1547 doSalvage(DII, Ops); 1548 }; 1549 1550 if (auto *CI = dyn_cast<CastInst>(&I)) { 1551 if (!CI->isNoopCast(DL)) 1552 return; 1553 1554 // No-op casts are irrelevant for debug info. 1555 MetadataAsValue *CastSrc = wrapMD(I.getOperand(0)); 1556 for (auto *DII : DbgUsers) { 1557 DII->setOperand(0, CastSrc); 1558 LLVM_DEBUG(dbgs() << "SALVAGE: " << *DII << '\n'); 1559 } 1560 } else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { 1561 unsigned BitWidth = 1562 M.getDataLayout().getIndexSizeInBits(GEP->getPointerAddressSpace()); 1563 // Rewrite a constant GEP into a DIExpression. Since we are performing 1564 // arithmetic to compute the variable's *value* in the DIExpression, we 1565 // need to mark the expression with a DW_OP_stack_value. 1566 APInt Offset(BitWidth, 0); 1567 if (GEP->accumulateConstantOffset(M.getDataLayout(), Offset)) 1568 for (auto *DII : DbgUsers) 1569 applyOffset(DII, Offset.getSExtValue()); 1570 } else if (auto *BI = dyn_cast<BinaryOperator>(&I)) { 1571 // Rewrite binary operations with constant integer operands. 1572 auto *ConstInt = dyn_cast<ConstantInt>(I.getOperand(1)); 1573 if (!ConstInt || ConstInt->getBitWidth() > 64) 1574 return; 1575 1576 uint64_t Val = ConstInt->getSExtValue(); 1577 for (auto *DII : DbgUsers) { 1578 switch (BI->getOpcode()) { 1579 case Instruction::Add: 1580 applyOffset(DII, Val); 1581 break; 1582 case Instruction::Sub: 1583 applyOffset(DII, -int64_t(Val)); 1584 break; 1585 case Instruction::Mul: 1586 applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_mul}); 1587 break; 1588 case Instruction::SDiv: 1589 applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_div}); 1590 break; 1591 case Instruction::SRem: 1592 applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_mod}); 1593 break; 1594 case Instruction::Or: 1595 applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_or}); 1596 break; 1597 case Instruction::And: 1598 applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_and}); 1599 break; 1600 case Instruction::Xor: 1601 applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_xor}); 1602 break; 1603 case Instruction::Shl: 1604 applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_shl}); 1605 break; 1606 case Instruction::LShr: 1607 applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_shr}); 1608 break; 1609 case Instruction::AShr: 1610 applyOps(DII, {dwarf::DW_OP_constu, Val, dwarf::DW_OP_shra}); 1611 break; 1612 default: 1613 // TODO: Salvage constants from each kind of binop we know about. 1614 continue; 1615 } 1616 } 1617 } else if (isa<LoadInst>(&I)) { 1618 MetadataAsValue *AddrMD = wrapMD(I.getOperand(0)); 1619 for (auto *DII : DbgUsers) { 1620 // Rewrite the load into DW_OP_deref. 1621 auto *DIExpr = DII->getExpression(); 1622 DIExpr = DIExpression::prepend(DIExpr, DIExpression::WithDeref); 1623 DII->setOperand(0, AddrMD); 1624 DII->setOperand(2, MetadataAsValue::get(I.getContext(), DIExpr)); 1625 LLVM_DEBUG(dbgs() << "SALVAGE: " << *DII << '\n'); 1626 } 1627 } 1628 } 1629 1630 unsigned llvm::removeAllNonTerminatorAndEHPadInstructions(BasicBlock *BB) { 1631 unsigned NumDeadInst = 0; 1632 // Delete the instructions backwards, as it has a reduced likelihood of 1633 // having to update as many def-use and use-def chains. 1634 Instruction *EndInst = BB->getTerminator(); // Last not to be deleted. 1635 while (EndInst != &BB->front()) { 1636 // Delete the next to last instruction. 1637 Instruction *Inst = &*--EndInst->getIterator(); 1638 if (!Inst->use_empty() && !Inst->getType()->isTokenTy()) 1639 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType())); 1640 if (Inst->isEHPad() || Inst->getType()->isTokenTy()) { 1641 EndInst = Inst; 1642 continue; 1643 } 1644 if (!isa<DbgInfoIntrinsic>(Inst)) 1645 ++NumDeadInst; 1646 Inst->eraseFromParent(); 1647 } 1648 return NumDeadInst; 1649 } 1650 1651 unsigned llvm::changeToUnreachable(Instruction *I, bool UseLLVMTrap, 1652 bool PreserveLCSSA, DeferredDominance *DDT) { 1653 BasicBlock *BB = I->getParent(); 1654 std::vector <DominatorTree::UpdateType> Updates; 1655 1656 // Loop over all of the successors, removing BB's entry from any PHI 1657 // nodes. 1658 if (DDT) 1659 Updates.reserve(BB->getTerminator()->getNumSuccessors()); 1660 for (BasicBlock *Successor : successors(BB)) { 1661 Successor->removePredecessor(BB, PreserveLCSSA); 1662 if (DDT) 1663 Updates.push_back({DominatorTree::Delete, BB, Successor}); 1664 } 1665 // Insert a call to llvm.trap right before this. This turns the undefined 1666 // behavior into a hard fail instead of falling through into random code. 1667 if (UseLLVMTrap) { 1668 Function *TrapFn = 1669 Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap); 1670 CallInst *CallTrap = CallInst::Create(TrapFn, "", I); 1671 CallTrap->setDebugLoc(I->getDebugLoc()); 1672 } 1673 new UnreachableInst(I->getContext(), I); 1674 1675 // All instructions after this are dead. 1676 unsigned NumInstrsRemoved = 0; 1677 BasicBlock::iterator BBI = I->getIterator(), BBE = BB->end(); 1678 while (BBI != BBE) { 1679 if (!BBI->use_empty()) 1680 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType())); 1681 BB->getInstList().erase(BBI++); 1682 ++NumInstrsRemoved; 1683 } 1684 if (DDT) 1685 DDT->applyUpdates(Updates); 1686 return NumInstrsRemoved; 1687 } 1688 1689 /// changeToCall - Convert the specified invoke into a normal call. 1690 static void changeToCall(InvokeInst *II, DeferredDominance *DDT = nullptr) { 1691 SmallVector<Value*, 8> Args(II->arg_begin(), II->arg_end()); 1692 SmallVector<OperandBundleDef, 1> OpBundles; 1693 II->getOperandBundlesAsDefs(OpBundles); 1694 CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, OpBundles, 1695 "", II); 1696 NewCall->takeName(II); 1697 NewCall->setCallingConv(II->getCallingConv()); 1698 NewCall->setAttributes(II->getAttributes()); 1699 NewCall->setDebugLoc(II->getDebugLoc()); 1700 II->replaceAllUsesWith(NewCall); 1701 1702 // Follow the call by a branch to the normal destination. 1703 BasicBlock *NormalDestBB = II->getNormalDest(); 1704 BranchInst::Create(NormalDestBB, II); 1705 1706 // Update PHI nodes in the unwind destination 1707 BasicBlock *BB = II->getParent(); 1708 BasicBlock *UnwindDestBB = II->getUnwindDest(); 1709 UnwindDestBB->removePredecessor(BB); 1710 II->eraseFromParent(); 1711 if (DDT) 1712 DDT->deleteEdge(BB, UnwindDestBB); 1713 } 1714 1715 BasicBlock *llvm::changeToInvokeAndSplitBasicBlock(CallInst *CI, 1716 BasicBlock *UnwindEdge) { 1717 BasicBlock *BB = CI->getParent(); 1718 1719 // Convert this function call into an invoke instruction. First, split the 1720 // basic block. 1721 BasicBlock *Split = 1722 BB->splitBasicBlock(CI->getIterator(), CI->getName() + ".noexc"); 1723 1724 // Delete the unconditional branch inserted by splitBasicBlock 1725 BB->getInstList().pop_back(); 1726 1727 // Create the new invoke instruction. 1728 SmallVector<Value *, 8> InvokeArgs(CI->arg_begin(), CI->arg_end()); 1729 SmallVector<OperandBundleDef, 1> OpBundles; 1730 1731 CI->getOperandBundlesAsDefs(OpBundles); 1732 1733 // Note: we're round tripping operand bundles through memory here, and that 1734 // can potentially be avoided with a cleverer API design that we do not have 1735 // as of this time. 1736 1737 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split, UnwindEdge, 1738 InvokeArgs, OpBundles, CI->getName(), BB); 1739 II->setDebugLoc(CI->getDebugLoc()); 1740 II->setCallingConv(CI->getCallingConv()); 1741 II->setAttributes(CI->getAttributes()); 1742 1743 // Make sure that anything using the call now uses the invoke! This also 1744 // updates the CallGraph if present, because it uses a WeakTrackingVH. 1745 CI->replaceAllUsesWith(II); 1746 1747 // Delete the original call 1748 Split->getInstList().pop_front(); 1749 return Split; 1750 } 1751 1752 static bool markAliveBlocks(Function &F, 1753 SmallPtrSetImpl<BasicBlock*> &Reachable, 1754 DeferredDominance *DDT = nullptr) { 1755 SmallVector<BasicBlock*, 128> Worklist; 1756 BasicBlock *BB = &F.front(); 1757 Worklist.push_back(BB); 1758 Reachable.insert(BB); 1759 bool Changed = false; 1760 do { 1761 BB = Worklist.pop_back_val(); 1762 1763 // Do a quick scan of the basic block, turning any obviously unreachable 1764 // instructions into LLVM unreachable insts. The instruction combining pass 1765 // canonicalizes unreachable insts into stores to null or undef. 1766 for (Instruction &I : *BB) { 1767 // Assumptions that are known to be false are equivalent to unreachable. 1768 // Also, if the condition is undefined, then we make the choice most 1769 // beneficial to the optimizer, and choose that to also be unreachable. 1770 if (auto *II = dyn_cast<IntrinsicInst>(&I)) { 1771 if (II->getIntrinsicID() == Intrinsic::assume) { 1772 if (match(II->getArgOperand(0), m_CombineOr(m_Zero(), m_Undef()))) { 1773 // Don't insert a call to llvm.trap right before the unreachable. 1774 changeToUnreachable(II, false, false, DDT); 1775 Changed = true; 1776 break; 1777 } 1778 } 1779 1780 if (II->getIntrinsicID() == Intrinsic::experimental_guard) { 1781 // A call to the guard intrinsic bails out of the current compilation 1782 // unit if the predicate passed to it is false. If the predicate is a 1783 // constant false, then we know the guard will bail out of the current 1784 // compile unconditionally, so all code following it is dead. 1785 // 1786 // Note: unlike in llvm.assume, it is not "obviously profitable" for 1787 // guards to treat `undef` as `false` since a guard on `undef` can 1788 // still be useful for widening. 1789 if (match(II->getArgOperand(0), m_Zero())) 1790 if (!isa<UnreachableInst>(II->getNextNode())) { 1791 changeToUnreachable(II->getNextNode(), /*UseLLVMTrap=*/false, 1792 false, DDT); 1793 Changed = true; 1794 break; 1795 } 1796 } 1797 } 1798 1799 if (auto *CI = dyn_cast<CallInst>(&I)) { 1800 Value *Callee = CI->getCalledValue(); 1801 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 1802 changeToUnreachable(CI, /*UseLLVMTrap=*/false, false, DDT); 1803 Changed = true; 1804 break; 1805 } 1806 if (CI->doesNotReturn()) { 1807 // If we found a call to a no-return function, insert an unreachable 1808 // instruction after it. Make sure there isn't *already* one there 1809 // though. 1810 if (!isa<UnreachableInst>(CI->getNextNode())) { 1811 // Don't insert a call to llvm.trap right before the unreachable. 1812 changeToUnreachable(CI->getNextNode(), false, false, DDT); 1813 Changed = true; 1814 } 1815 break; 1816 } 1817 } 1818 1819 // Store to undef and store to null are undefined and used to signal that 1820 // they should be changed to unreachable by passes that can't modify the 1821 // CFG. 1822 if (auto *SI = dyn_cast<StoreInst>(&I)) { 1823 // Don't touch volatile stores. 1824 if (SI->isVolatile()) continue; 1825 1826 Value *Ptr = SI->getOperand(1); 1827 1828 if (isa<UndefValue>(Ptr) || 1829 (isa<ConstantPointerNull>(Ptr) && 1830 SI->getPointerAddressSpace() == 0)) { 1831 changeToUnreachable(SI, true, false, DDT); 1832 Changed = true; 1833 break; 1834 } 1835 } 1836 } 1837 1838 TerminatorInst *Terminator = BB->getTerminator(); 1839 if (auto *II = dyn_cast<InvokeInst>(Terminator)) { 1840 // Turn invokes that call 'nounwind' functions into ordinary calls. 1841 Value *Callee = II->getCalledValue(); 1842 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 1843 changeToUnreachable(II, true, false, DDT); 1844 Changed = true; 1845 } else if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(&F)) { 1846 if (II->use_empty() && II->onlyReadsMemory()) { 1847 // jump to the normal destination branch. 1848 BasicBlock *NormalDestBB = II->getNormalDest(); 1849 BasicBlock *UnwindDestBB = II->getUnwindDest(); 1850 BranchInst::Create(NormalDestBB, II); 1851 UnwindDestBB->removePredecessor(II->getParent()); 1852 II->eraseFromParent(); 1853 if (DDT) 1854 DDT->deleteEdge(BB, UnwindDestBB); 1855 } else 1856 changeToCall(II, DDT); 1857 Changed = true; 1858 } 1859 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Terminator)) { 1860 // Remove catchpads which cannot be reached. 1861 struct CatchPadDenseMapInfo { 1862 static CatchPadInst *getEmptyKey() { 1863 return DenseMapInfo<CatchPadInst *>::getEmptyKey(); 1864 } 1865 1866 static CatchPadInst *getTombstoneKey() { 1867 return DenseMapInfo<CatchPadInst *>::getTombstoneKey(); 1868 } 1869 1870 static unsigned getHashValue(CatchPadInst *CatchPad) { 1871 return static_cast<unsigned>(hash_combine_range( 1872 CatchPad->value_op_begin(), CatchPad->value_op_end())); 1873 } 1874 1875 static bool isEqual(CatchPadInst *LHS, CatchPadInst *RHS) { 1876 if (LHS == getEmptyKey() || LHS == getTombstoneKey() || 1877 RHS == getEmptyKey() || RHS == getTombstoneKey()) 1878 return LHS == RHS; 1879 return LHS->isIdenticalTo(RHS); 1880 } 1881 }; 1882 1883 // Set of unique CatchPads. 1884 SmallDenseMap<CatchPadInst *, detail::DenseSetEmpty, 4, 1885 CatchPadDenseMapInfo, detail::DenseSetPair<CatchPadInst *>> 1886 HandlerSet; 1887 detail::DenseSetEmpty Empty; 1888 for (CatchSwitchInst::handler_iterator I = CatchSwitch->handler_begin(), 1889 E = CatchSwitch->handler_end(); 1890 I != E; ++I) { 1891 BasicBlock *HandlerBB = *I; 1892 auto *CatchPad = cast<CatchPadInst>(HandlerBB->getFirstNonPHI()); 1893 if (!HandlerSet.insert({CatchPad, Empty}).second) { 1894 CatchSwitch->removeHandler(I); 1895 --I; 1896 --E; 1897 Changed = true; 1898 } 1899 } 1900 } 1901 1902 Changed |= ConstantFoldTerminator(BB, true, nullptr, DDT); 1903 for (BasicBlock *Successor : successors(BB)) 1904 if (Reachable.insert(Successor).second) 1905 Worklist.push_back(Successor); 1906 } while (!Worklist.empty()); 1907 return Changed; 1908 } 1909 1910 void llvm::removeUnwindEdge(BasicBlock *BB, DeferredDominance *DDT) { 1911 TerminatorInst *TI = BB->getTerminator(); 1912 1913 if (auto *II = dyn_cast<InvokeInst>(TI)) { 1914 changeToCall(II, DDT); 1915 return; 1916 } 1917 1918 TerminatorInst *NewTI; 1919 BasicBlock *UnwindDest; 1920 1921 if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) { 1922 NewTI = CleanupReturnInst::Create(CRI->getCleanupPad(), nullptr, CRI); 1923 UnwindDest = CRI->getUnwindDest(); 1924 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(TI)) { 1925 auto *NewCatchSwitch = CatchSwitchInst::Create( 1926 CatchSwitch->getParentPad(), nullptr, CatchSwitch->getNumHandlers(), 1927 CatchSwitch->getName(), CatchSwitch); 1928 for (BasicBlock *PadBB : CatchSwitch->handlers()) 1929 NewCatchSwitch->addHandler(PadBB); 1930 1931 NewTI = NewCatchSwitch; 1932 UnwindDest = CatchSwitch->getUnwindDest(); 1933 } else { 1934 llvm_unreachable("Could not find unwind successor"); 1935 } 1936 1937 NewTI->takeName(TI); 1938 NewTI->setDebugLoc(TI->getDebugLoc()); 1939 UnwindDest->removePredecessor(BB); 1940 TI->replaceAllUsesWith(NewTI); 1941 TI->eraseFromParent(); 1942 if (DDT) 1943 DDT->deleteEdge(BB, UnwindDest); 1944 } 1945 1946 /// removeUnreachableBlocks - Remove blocks that are not reachable, even 1947 /// if they are in a dead cycle. Return true if a change was made, false 1948 /// otherwise. If `LVI` is passed, this function preserves LazyValueInfo 1949 /// after modifying the CFG. 1950 bool llvm::removeUnreachableBlocks(Function &F, LazyValueInfo *LVI, 1951 DeferredDominance *DDT) { 1952 SmallPtrSet<BasicBlock*, 16> Reachable; 1953 bool Changed = markAliveBlocks(F, Reachable, DDT); 1954 1955 // If there are unreachable blocks in the CFG... 1956 if (Reachable.size() == F.size()) 1957 return Changed; 1958 1959 assert(Reachable.size() < F.size()); 1960 NumRemoved += F.size()-Reachable.size(); 1961 1962 // Loop over all of the basic blocks that are not reachable, dropping all of 1963 // their internal references. Update DDT and LVI if available. 1964 std::vector <DominatorTree::UpdateType> Updates; 1965 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ++I) { 1966 auto *BB = &*I; 1967 if (Reachable.count(BB)) 1968 continue; 1969 for (BasicBlock *Successor : successors(BB)) { 1970 if (Reachable.count(Successor)) 1971 Successor->removePredecessor(BB); 1972 if (DDT) 1973 Updates.push_back({DominatorTree::Delete, BB, Successor}); 1974 } 1975 if (LVI) 1976 LVI->eraseBlock(BB); 1977 BB->dropAllReferences(); 1978 } 1979 1980 for (Function::iterator I = ++F.begin(); I != F.end();) { 1981 auto *BB = &*I; 1982 if (Reachable.count(BB)) { 1983 ++I; 1984 continue; 1985 } 1986 if (DDT) { 1987 DDT->deleteBB(BB); // deferred deletion of BB. 1988 ++I; 1989 } else { 1990 I = F.getBasicBlockList().erase(I); 1991 } 1992 } 1993 1994 if (DDT) 1995 DDT->applyUpdates(Updates); 1996 return true; 1997 } 1998 1999 void llvm::combineMetadata(Instruction *K, const Instruction *J, 2000 ArrayRef<unsigned> KnownIDs) { 2001 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata; 2002 K->dropUnknownNonDebugMetadata(KnownIDs); 2003 K->getAllMetadataOtherThanDebugLoc(Metadata); 2004 for (const auto &MD : Metadata) { 2005 unsigned Kind = MD.first; 2006 MDNode *JMD = J->getMetadata(Kind); 2007 MDNode *KMD = MD.second; 2008 2009 switch (Kind) { 2010 default: 2011 K->setMetadata(Kind, nullptr); // Remove unknown metadata 2012 break; 2013 case LLVMContext::MD_dbg: 2014 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg"); 2015 case LLVMContext::MD_tbaa: 2016 K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD)); 2017 break; 2018 case LLVMContext::MD_alias_scope: 2019 K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD)); 2020 break; 2021 case LLVMContext::MD_noalias: 2022 case LLVMContext::MD_mem_parallel_loop_access: 2023 K->setMetadata(Kind, MDNode::intersect(JMD, KMD)); 2024 break; 2025 case LLVMContext::MD_range: 2026 K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD)); 2027 break; 2028 case LLVMContext::MD_fpmath: 2029 K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD)); 2030 break; 2031 case LLVMContext::MD_invariant_load: 2032 // Only set the !invariant.load if it is present in both instructions. 2033 K->setMetadata(Kind, JMD); 2034 break; 2035 case LLVMContext::MD_nonnull: 2036 // Only set the !nonnull if it is present in both instructions. 2037 K->setMetadata(Kind, JMD); 2038 break; 2039 case LLVMContext::MD_invariant_group: 2040 // Preserve !invariant.group in K. 2041 break; 2042 case LLVMContext::MD_align: 2043 K->setMetadata(Kind, 2044 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD)); 2045 break; 2046 case LLVMContext::MD_dereferenceable: 2047 case LLVMContext::MD_dereferenceable_or_null: 2048 K->setMetadata(Kind, 2049 MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD)); 2050 break; 2051 } 2052 } 2053 // Set !invariant.group from J if J has it. If both instructions have it 2054 // then we will just pick it from J - even when they are different. 2055 // Also make sure that K is load or store - f.e. combining bitcast with load 2056 // could produce bitcast with invariant.group metadata, which is invalid. 2057 // FIXME: we should try to preserve both invariant.group md if they are 2058 // different, but right now instruction can only have one invariant.group. 2059 if (auto *JMD = J->getMetadata(LLVMContext::MD_invariant_group)) 2060 if (isa<LoadInst>(K) || isa<StoreInst>(K)) 2061 K->setMetadata(LLVMContext::MD_invariant_group, JMD); 2062 } 2063 2064 void llvm::combineMetadataForCSE(Instruction *K, const Instruction *J) { 2065 unsigned KnownIDs[] = { 2066 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope, 2067 LLVMContext::MD_noalias, LLVMContext::MD_range, 2068 LLVMContext::MD_invariant_load, LLVMContext::MD_nonnull, 2069 LLVMContext::MD_invariant_group, LLVMContext::MD_align, 2070 LLVMContext::MD_dereferenceable, 2071 LLVMContext::MD_dereferenceable_or_null}; 2072 combineMetadata(K, J, KnownIDs); 2073 } 2074 2075 template <typename RootType, typename DominatesFn> 2076 static unsigned replaceDominatedUsesWith(Value *From, Value *To, 2077 const RootType &Root, 2078 const DominatesFn &Dominates) { 2079 assert(From->getType() == To->getType()); 2080 2081 unsigned Count = 0; 2082 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 2083 UI != UE;) { 2084 Use &U = *UI++; 2085 if (!Dominates(Root, U)) 2086 continue; 2087 U.set(To); 2088 LLVM_DEBUG(dbgs() << "Replace dominated use of '" << From->getName() 2089 << "' as " << *To << " in " << *U << "\n"); 2090 ++Count; 2091 } 2092 return Count; 2093 } 2094 2095 unsigned llvm::replaceNonLocalUsesWith(Instruction *From, Value *To) { 2096 assert(From->getType() == To->getType()); 2097 auto *BB = From->getParent(); 2098 unsigned Count = 0; 2099 2100 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end(); 2101 UI != UE;) { 2102 Use &U = *UI++; 2103 auto *I = cast<Instruction>(U.getUser()); 2104 if (I->getParent() == BB) 2105 continue; 2106 U.set(To); 2107 ++Count; 2108 } 2109 return Count; 2110 } 2111 2112 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To, 2113 DominatorTree &DT, 2114 const BasicBlockEdge &Root) { 2115 auto Dominates = [&DT](const BasicBlockEdge &Root, const Use &U) { 2116 return DT.dominates(Root, U); 2117 }; 2118 return ::replaceDominatedUsesWith(From, To, Root, Dominates); 2119 } 2120 2121 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To, 2122 DominatorTree &DT, 2123 const BasicBlock *BB) { 2124 auto ProperlyDominates = [&DT](const BasicBlock *BB, const Use &U) { 2125 auto *I = cast<Instruction>(U.getUser())->getParent(); 2126 return DT.properlyDominates(BB, I); 2127 }; 2128 return ::replaceDominatedUsesWith(From, To, BB, ProperlyDominates); 2129 } 2130 2131 bool llvm::callsGCLeafFunction(ImmutableCallSite CS, 2132 const TargetLibraryInfo &TLI) { 2133 // Check if the function is specifically marked as a gc leaf function. 2134 if (CS.hasFnAttr("gc-leaf-function")) 2135 return true; 2136 if (const Function *F = CS.getCalledFunction()) { 2137 if (F->hasFnAttribute("gc-leaf-function")) 2138 return true; 2139 2140 if (auto IID = F->getIntrinsicID()) 2141 // Most LLVM intrinsics do not take safepoints. 2142 return IID != Intrinsic::experimental_gc_statepoint && 2143 IID != Intrinsic::experimental_deoptimize; 2144 } 2145 2146 // Lib calls can be materialized by some passes, and won't be 2147 // marked as 'gc-leaf-function.' All available Libcalls are 2148 // GC-leaf. 2149 LibFunc LF; 2150 if (TLI.getLibFunc(CS, LF)) { 2151 return TLI.has(LF); 2152 } 2153 2154 return false; 2155 } 2156 2157 void llvm::copyNonnullMetadata(const LoadInst &OldLI, MDNode *N, 2158 LoadInst &NewLI) { 2159 auto *NewTy = NewLI.getType(); 2160 2161 // This only directly applies if the new type is also a pointer. 2162 if (NewTy->isPointerTy()) { 2163 NewLI.setMetadata(LLVMContext::MD_nonnull, N); 2164 return; 2165 } 2166 2167 // The only other translation we can do is to integral loads with !range 2168 // metadata. 2169 if (!NewTy->isIntegerTy()) 2170 return; 2171 2172 MDBuilder MDB(NewLI.getContext()); 2173 const Value *Ptr = OldLI.getPointerOperand(); 2174 auto *ITy = cast<IntegerType>(NewTy); 2175 auto *NullInt = ConstantExpr::getPtrToInt( 2176 ConstantPointerNull::get(cast<PointerType>(Ptr->getType())), ITy); 2177 auto *NonNullInt = ConstantExpr::getAdd(NullInt, ConstantInt::get(ITy, 1)); 2178 NewLI.setMetadata(LLVMContext::MD_range, 2179 MDB.createRange(NonNullInt, NullInt)); 2180 } 2181 2182 void llvm::copyRangeMetadata(const DataLayout &DL, const LoadInst &OldLI, 2183 MDNode *N, LoadInst &NewLI) { 2184 auto *NewTy = NewLI.getType(); 2185 2186 // Give up unless it is converted to a pointer where there is a single very 2187 // valuable mapping we can do reliably. 2188 // FIXME: It would be nice to propagate this in more ways, but the type 2189 // conversions make it hard. 2190 if (!NewTy->isPointerTy()) 2191 return; 2192 2193 unsigned BitWidth = DL.getIndexTypeSizeInBits(NewTy); 2194 if (!getConstantRangeFromMetadata(*N).contains(APInt(BitWidth, 0))) { 2195 MDNode *NN = MDNode::get(OldLI.getContext(), None); 2196 NewLI.setMetadata(LLVMContext::MD_nonnull, NN); 2197 } 2198 } 2199 2200 namespace { 2201 2202 /// A potential constituent of a bitreverse or bswap expression. See 2203 /// collectBitParts for a fuller explanation. 2204 struct BitPart { 2205 BitPart(Value *P, unsigned BW) : Provider(P) { 2206 Provenance.resize(BW); 2207 } 2208 2209 /// The Value that this is a bitreverse/bswap of. 2210 Value *Provider; 2211 2212 /// The "provenance" of each bit. Provenance[A] = B means that bit A 2213 /// in Provider becomes bit B in the result of this expression. 2214 SmallVector<int8_t, 32> Provenance; // int8_t means max size is i128. 2215 2216 enum { Unset = -1 }; 2217 }; 2218 2219 } // end anonymous namespace 2220 2221 /// Analyze the specified subexpression and see if it is capable of providing 2222 /// pieces of a bswap or bitreverse. The subexpression provides a potential 2223 /// piece of a bswap or bitreverse if it can be proven that each non-zero bit in 2224 /// the output of the expression came from a corresponding bit in some other 2225 /// value. This function is recursive, and the end result is a mapping of 2226 /// bitnumber to bitnumber. It is the caller's responsibility to validate that 2227 /// the bitnumber to bitnumber mapping is correct for a bswap or bitreverse. 2228 /// 2229 /// For example, if the current subexpression if "(shl i32 %X, 24)" then we know 2230 /// that the expression deposits the low byte of %X into the high byte of the 2231 /// result and that all other bits are zero. This expression is accepted and a 2232 /// BitPart is returned with Provider set to %X and Provenance[24-31] set to 2233 /// [0-7]. 2234 /// 2235 /// To avoid revisiting values, the BitPart results are memoized into the 2236 /// provided map. To avoid unnecessary copying of BitParts, BitParts are 2237 /// constructed in-place in the \c BPS map. Because of this \c BPS needs to 2238 /// store BitParts objects, not pointers. As we need the concept of a nullptr 2239 /// BitParts (Value has been analyzed and the analysis failed), we an Optional 2240 /// type instead to provide the same functionality. 2241 /// 2242 /// Because we pass around references into \c BPS, we must use a container that 2243 /// does not invalidate internal references (std::map instead of DenseMap). 2244 static const Optional<BitPart> & 2245 collectBitParts(Value *V, bool MatchBSwaps, bool MatchBitReversals, 2246 std::map<Value *, Optional<BitPart>> &BPS) { 2247 auto I = BPS.find(V); 2248 if (I != BPS.end()) 2249 return I->second; 2250 2251 auto &Result = BPS[V] = None; 2252 auto BitWidth = cast<IntegerType>(V->getType())->getBitWidth(); 2253 2254 if (Instruction *I = dyn_cast<Instruction>(V)) { 2255 // If this is an or instruction, it may be an inner node of the bswap. 2256 if (I->getOpcode() == Instruction::Or) { 2257 auto &A = collectBitParts(I->getOperand(0), MatchBSwaps, 2258 MatchBitReversals, BPS); 2259 auto &B = collectBitParts(I->getOperand(1), MatchBSwaps, 2260 MatchBitReversals, BPS); 2261 if (!A || !B) 2262 return Result; 2263 2264 // Try and merge the two together. 2265 if (!A->Provider || A->Provider != B->Provider) 2266 return Result; 2267 2268 Result = BitPart(A->Provider, BitWidth); 2269 for (unsigned i = 0; i < A->Provenance.size(); ++i) { 2270 if (A->Provenance[i] != BitPart::Unset && 2271 B->Provenance[i] != BitPart::Unset && 2272 A->Provenance[i] != B->Provenance[i]) 2273 return Result = None; 2274 2275 if (A->Provenance[i] == BitPart::Unset) 2276 Result->Provenance[i] = B->Provenance[i]; 2277 else 2278 Result->Provenance[i] = A->Provenance[i]; 2279 } 2280 2281 return Result; 2282 } 2283 2284 // If this is a logical shift by a constant, recurse then shift the result. 2285 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) { 2286 unsigned BitShift = 2287 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U); 2288 // Ensure the shift amount is defined. 2289 if (BitShift > BitWidth) 2290 return Result; 2291 2292 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps, 2293 MatchBitReversals, BPS); 2294 if (!Res) 2295 return Result; 2296 Result = Res; 2297 2298 // Perform the "shift" on BitProvenance. 2299 auto &P = Result->Provenance; 2300 if (I->getOpcode() == Instruction::Shl) { 2301 P.erase(std::prev(P.end(), BitShift), P.end()); 2302 P.insert(P.begin(), BitShift, BitPart::Unset); 2303 } else { 2304 P.erase(P.begin(), std::next(P.begin(), BitShift)); 2305 P.insert(P.end(), BitShift, BitPart::Unset); 2306 } 2307 2308 return Result; 2309 } 2310 2311 // If this is a logical 'and' with a mask that clears bits, recurse then 2312 // unset the appropriate bits. 2313 if (I->getOpcode() == Instruction::And && 2314 isa<ConstantInt>(I->getOperand(1))) { 2315 APInt Bit(I->getType()->getPrimitiveSizeInBits(), 1); 2316 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue(); 2317 2318 // Check that the mask allows a multiple of 8 bits for a bswap, for an 2319 // early exit. 2320 unsigned NumMaskedBits = AndMask.countPopulation(); 2321 if (!MatchBitReversals && NumMaskedBits % 8 != 0) 2322 return Result; 2323 2324 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps, 2325 MatchBitReversals, BPS); 2326 if (!Res) 2327 return Result; 2328 Result = Res; 2329 2330 for (unsigned i = 0; i < BitWidth; ++i, Bit <<= 1) 2331 // If the AndMask is zero for this bit, clear the bit. 2332 if ((AndMask & Bit) == 0) 2333 Result->Provenance[i] = BitPart::Unset; 2334 return Result; 2335 } 2336 2337 // If this is a zext instruction zero extend the result. 2338 if (I->getOpcode() == Instruction::ZExt) { 2339 auto &Res = collectBitParts(I->getOperand(0), MatchBSwaps, 2340 MatchBitReversals, BPS); 2341 if (!Res) 2342 return Result; 2343 2344 Result = BitPart(Res->Provider, BitWidth); 2345 auto NarrowBitWidth = 2346 cast<IntegerType>(cast<ZExtInst>(I)->getSrcTy())->getBitWidth(); 2347 for (unsigned i = 0; i < NarrowBitWidth; ++i) 2348 Result->Provenance[i] = Res->Provenance[i]; 2349 for (unsigned i = NarrowBitWidth; i < BitWidth; ++i) 2350 Result->Provenance[i] = BitPart::Unset; 2351 return Result; 2352 } 2353 } 2354 2355 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be 2356 // the input value to the bswap/bitreverse. 2357 Result = BitPart(V, BitWidth); 2358 for (unsigned i = 0; i < BitWidth; ++i) 2359 Result->Provenance[i] = i; 2360 return Result; 2361 } 2362 2363 static bool bitTransformIsCorrectForBSwap(unsigned From, unsigned To, 2364 unsigned BitWidth) { 2365 if (From % 8 != To % 8) 2366 return false; 2367 // Convert from bit indices to byte indices and check for a byte reversal. 2368 From >>= 3; 2369 To >>= 3; 2370 BitWidth >>= 3; 2371 return From == BitWidth - To - 1; 2372 } 2373 2374 static bool bitTransformIsCorrectForBitReverse(unsigned From, unsigned To, 2375 unsigned BitWidth) { 2376 return From == BitWidth - To - 1; 2377 } 2378 2379 bool llvm::recognizeBSwapOrBitReverseIdiom( 2380 Instruction *I, bool MatchBSwaps, bool MatchBitReversals, 2381 SmallVectorImpl<Instruction *> &InsertedInsts) { 2382 if (Operator::getOpcode(I) != Instruction::Or) 2383 return false; 2384 if (!MatchBSwaps && !MatchBitReversals) 2385 return false; 2386 IntegerType *ITy = dyn_cast<IntegerType>(I->getType()); 2387 if (!ITy || ITy->getBitWidth() > 128) 2388 return false; // Can't do vectors or integers > 128 bits. 2389 unsigned BW = ITy->getBitWidth(); 2390 2391 unsigned DemandedBW = BW; 2392 IntegerType *DemandedTy = ITy; 2393 if (I->hasOneUse()) { 2394 if (TruncInst *Trunc = dyn_cast<TruncInst>(I->user_back())) { 2395 DemandedTy = cast<IntegerType>(Trunc->getType()); 2396 DemandedBW = DemandedTy->getBitWidth(); 2397 } 2398 } 2399 2400 // Try to find all the pieces corresponding to the bswap. 2401 std::map<Value *, Optional<BitPart>> BPS; 2402 auto Res = collectBitParts(I, MatchBSwaps, MatchBitReversals, BPS); 2403 if (!Res) 2404 return false; 2405 auto &BitProvenance = Res->Provenance; 2406 2407 // Now, is the bit permutation correct for a bswap or a bitreverse? We can 2408 // only byteswap values with an even number of bytes. 2409 bool OKForBSwap = DemandedBW % 16 == 0, OKForBitReverse = true; 2410 for (unsigned i = 0; i < DemandedBW; ++i) { 2411 OKForBSwap &= 2412 bitTransformIsCorrectForBSwap(BitProvenance[i], i, DemandedBW); 2413 OKForBitReverse &= 2414 bitTransformIsCorrectForBitReverse(BitProvenance[i], i, DemandedBW); 2415 } 2416 2417 Intrinsic::ID Intrin; 2418 if (OKForBSwap && MatchBSwaps) 2419 Intrin = Intrinsic::bswap; 2420 else if (OKForBitReverse && MatchBitReversals) 2421 Intrin = Intrinsic::bitreverse; 2422 else 2423 return false; 2424 2425 if (ITy != DemandedTy) { 2426 Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, DemandedTy); 2427 Value *Provider = Res->Provider; 2428 IntegerType *ProviderTy = cast<IntegerType>(Provider->getType()); 2429 // We may need to truncate the provider. 2430 if (DemandedTy != ProviderTy) { 2431 auto *Trunc = CastInst::Create(Instruction::Trunc, Provider, DemandedTy, 2432 "trunc", I); 2433 InsertedInsts.push_back(Trunc); 2434 Provider = Trunc; 2435 } 2436 auto *CI = CallInst::Create(F, Provider, "rev", I); 2437 InsertedInsts.push_back(CI); 2438 auto *ExtInst = CastInst::Create(Instruction::ZExt, CI, ITy, "zext", I); 2439 InsertedInsts.push_back(ExtInst); 2440 return true; 2441 } 2442 2443 Function *F = Intrinsic::getDeclaration(I->getModule(), Intrin, ITy); 2444 InsertedInsts.push_back(CallInst::Create(F, Res->Provider, "rev", I)); 2445 return true; 2446 } 2447 2448 // CodeGen has special handling for some string functions that may replace 2449 // them with target-specific intrinsics. Since that'd skip our interceptors 2450 // in ASan/MSan/TSan/DFSan, and thus make us miss some memory accesses, 2451 // we mark affected calls as NoBuiltin, which will disable optimization 2452 // in CodeGen. 2453 void llvm::maybeMarkSanitizerLibraryCallNoBuiltin( 2454 CallInst *CI, const TargetLibraryInfo *TLI) { 2455 Function *F = CI->getCalledFunction(); 2456 LibFunc Func; 2457 if (F && !F->hasLocalLinkage() && F->hasName() && 2458 TLI->getLibFunc(F->getName(), Func) && TLI->hasOptimizedCodeGen(Func) && 2459 !F->doesNotAccessMemory()) 2460 CI->addAttribute(AttributeList::FunctionIndex, Attribute::NoBuiltin); 2461 } 2462 2463 bool llvm::canReplaceOperandWithVariable(const Instruction *I, unsigned OpIdx) { 2464 // We can't have a PHI with a metadata type. 2465 if (I->getOperand(OpIdx)->getType()->isMetadataTy()) 2466 return false; 2467 2468 // Early exit. 2469 if (!isa<Constant>(I->getOperand(OpIdx))) 2470 return true; 2471 2472 switch (I->getOpcode()) { 2473 default: 2474 return true; 2475 case Instruction::Call: 2476 case Instruction::Invoke: 2477 // Can't handle inline asm. Skip it. 2478 if (isa<InlineAsm>(ImmutableCallSite(I).getCalledValue())) 2479 return false; 2480 // Many arithmetic intrinsics have no issue taking a 2481 // variable, however it's hard to distingish these from 2482 // specials such as @llvm.frameaddress that require a constant. 2483 if (isa<IntrinsicInst>(I)) 2484 return false; 2485 2486 // Constant bundle operands may need to retain their constant-ness for 2487 // correctness. 2488 if (ImmutableCallSite(I).isBundleOperand(OpIdx)) 2489 return false; 2490 return true; 2491 case Instruction::ShuffleVector: 2492 // Shufflevector masks are constant. 2493 return OpIdx != 2; 2494 case Instruction::Switch: 2495 case Instruction::ExtractValue: 2496 // All operands apart from the first are constant. 2497 return OpIdx == 0; 2498 case Instruction::InsertValue: 2499 // All operands apart from the first and the second are constant. 2500 return OpIdx < 2; 2501 case Instruction::Alloca: 2502 // Static allocas (constant size in the entry block) are handled by 2503 // prologue/epilogue insertion so they're free anyway. We definitely don't 2504 // want to make them non-constant. 2505 return !cast<AllocaInst>(I)->isStaticAlloca(); 2506 case Instruction::GetElementPtr: 2507 if (OpIdx == 0) 2508 return true; 2509 gep_type_iterator It = gep_type_begin(I); 2510 for (auto E = std::next(It, OpIdx); It != E; ++It) 2511 if (It.isStruct()) 2512 return false; 2513 return true; 2514 } 2515 } 2516