1 //===-- Analysis.cpp - CodeGen LLVM IR Analysis Utilities -----------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file defines several CodeGen-specific LLVM IR analysis utilities. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/CodeGen/Analysis.h" 14 #include "llvm/Analysis/ValueTracking.h" 15 #include "llvm/CodeGen/MachineFunction.h" 16 #include "llvm/CodeGen/TargetInstrInfo.h" 17 #include "llvm/CodeGen/TargetLowering.h" 18 #include "llvm/CodeGen/TargetSubtargetInfo.h" 19 #include "llvm/IR/DataLayout.h" 20 #include "llvm/IR/DerivedTypes.h" 21 #include "llvm/IR/Function.h" 22 #include "llvm/IR/Instructions.h" 23 #include "llvm/IR/IntrinsicInst.h" 24 #include "llvm/IR/LLVMContext.h" 25 #include "llvm/IR/Module.h" 26 #include "llvm/Support/ErrorHandling.h" 27 #include "llvm/Support/MathExtras.h" 28 #include "llvm/Target/TargetMachine.h" 29 #include "llvm/Transforms/Utils/GlobalStatus.h" 30 31 using namespace llvm; 32 33 /// Compute the linearized index of a member in a nested aggregate/struct/array 34 /// by recursing and accumulating CurIndex as long as there are indices in the 35 /// index list. 36 unsigned llvm::ComputeLinearIndex(Type *Ty, 37 const unsigned *Indices, 38 const unsigned *IndicesEnd, 39 unsigned CurIndex) { 40 // Base case: We're done. 41 if (Indices && Indices == IndicesEnd) 42 return CurIndex; 43 44 // Given a struct type, recursively traverse the elements. 45 if (StructType *STy = dyn_cast<StructType>(Ty)) { 46 for (auto I : llvm::enumerate(STy->elements())) { 47 Type *ET = I.value(); 48 if (Indices && *Indices == I.index()) 49 return ComputeLinearIndex(ET, Indices + 1, IndicesEnd, CurIndex); 50 CurIndex = ComputeLinearIndex(ET, nullptr, nullptr, CurIndex); 51 } 52 assert(!Indices && "Unexpected out of bound"); 53 return CurIndex; 54 } 55 // Given an array type, recursively traverse the elements. 56 else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 57 Type *EltTy = ATy->getElementType(); 58 unsigned NumElts = ATy->getNumElements(); 59 // Compute the Linear offset when jumping one element of the array 60 unsigned EltLinearOffset = ComputeLinearIndex(EltTy, nullptr, nullptr, 0); 61 if (Indices) { 62 assert(*Indices < NumElts && "Unexpected out of bound"); 63 // If the indice is inside the array, compute the index to the requested 64 // elt and recurse inside the element with the end of the indices list 65 CurIndex += EltLinearOffset* *Indices; 66 return ComputeLinearIndex(EltTy, Indices+1, IndicesEnd, CurIndex); 67 } 68 CurIndex += EltLinearOffset*NumElts; 69 return CurIndex; 70 } 71 // We haven't found the type we're looking for, so keep searching. 72 return CurIndex + 1; 73 } 74 75 /// ComputeValueVTs - Given an LLVM IR type, compute a sequence of 76 /// EVTs that represent all the individual underlying 77 /// non-aggregate types that comprise it. 78 /// 79 /// If Offsets is non-null, it points to a vector to be filled in 80 /// with the in-memory offsets of each of the individual values. 81 /// 82 void llvm::ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL, 83 Type *Ty, SmallVectorImpl<EVT> &ValueVTs, 84 SmallVectorImpl<EVT> *MemVTs, 85 SmallVectorImpl<uint64_t> *Offsets, 86 uint64_t StartingOffset) { 87 // Given a struct type, recursively traverse the elements. 88 if (StructType *STy = dyn_cast<StructType>(Ty)) { 89 // If the Offsets aren't needed, don't query the struct layout. This allows 90 // us to support structs with scalable vectors for operations that don't 91 // need offsets. 92 const StructLayout *SL = Offsets ? DL.getStructLayout(STy) : nullptr; 93 for (StructType::element_iterator EB = STy->element_begin(), 94 EI = EB, 95 EE = STy->element_end(); 96 EI != EE; ++EI) { 97 // Don't compute the element offset if we didn't get a StructLayout above. 98 uint64_t EltOffset = SL ? SL->getElementOffset(EI - EB) : 0; 99 ComputeValueVTs(TLI, DL, *EI, ValueVTs, MemVTs, Offsets, 100 StartingOffset + EltOffset); 101 } 102 return; 103 } 104 // Given an array type, recursively traverse the elements. 105 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 106 Type *EltTy = ATy->getElementType(); 107 uint64_t EltSize = DL.getTypeAllocSize(EltTy).getFixedValue(); 108 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) 109 ComputeValueVTs(TLI, DL, EltTy, ValueVTs, MemVTs, Offsets, 110 StartingOffset + i * EltSize); 111 return; 112 } 113 // Interpret void as zero return values. 114 if (Ty->isVoidTy()) 115 return; 116 // Base case: we can get an EVT for this LLVM IR type. 117 ValueVTs.push_back(TLI.getValueType(DL, Ty)); 118 if (MemVTs) 119 MemVTs->push_back(TLI.getMemValueType(DL, Ty)); 120 if (Offsets) 121 Offsets->push_back(StartingOffset); 122 } 123 124 void llvm::ComputeValueVTs(const TargetLowering &TLI, const DataLayout &DL, 125 Type *Ty, SmallVectorImpl<EVT> &ValueVTs, 126 SmallVectorImpl<uint64_t> *Offsets, 127 uint64_t StartingOffset) { 128 return ComputeValueVTs(TLI, DL, Ty, ValueVTs, /*MemVTs=*/nullptr, Offsets, 129 StartingOffset); 130 } 131 132 void llvm::computeValueLLTs(const DataLayout &DL, Type &Ty, 133 SmallVectorImpl<LLT> &ValueTys, 134 SmallVectorImpl<uint64_t> *Offsets, 135 uint64_t StartingOffset) { 136 // Given a struct type, recursively traverse the elements. 137 if (StructType *STy = dyn_cast<StructType>(&Ty)) { 138 // If the Offsets aren't needed, don't query the struct layout. This allows 139 // us to support structs with scalable vectors for operations that don't 140 // need offsets. 141 const StructLayout *SL = Offsets ? DL.getStructLayout(STy) : nullptr; 142 for (unsigned I = 0, E = STy->getNumElements(); I != E; ++I) { 143 uint64_t EltOffset = SL ? SL->getElementOffset(I) : 0; 144 computeValueLLTs(DL, *STy->getElementType(I), ValueTys, Offsets, 145 StartingOffset + EltOffset); 146 } 147 return; 148 } 149 // Given an array type, recursively traverse the elements. 150 if (ArrayType *ATy = dyn_cast<ArrayType>(&Ty)) { 151 Type *EltTy = ATy->getElementType(); 152 uint64_t EltSize = DL.getTypeAllocSize(EltTy).getFixedValue(); 153 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) 154 computeValueLLTs(DL, *EltTy, ValueTys, Offsets, 155 StartingOffset + i * EltSize); 156 return; 157 } 158 // Interpret void as zero return values. 159 if (Ty.isVoidTy()) 160 return; 161 // Base case: we can get an LLT for this LLVM IR type. 162 ValueTys.push_back(getLLTForType(Ty, DL)); 163 if (Offsets != nullptr) 164 Offsets->push_back(StartingOffset * 8); 165 } 166 167 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V. 168 GlobalValue *llvm::ExtractTypeInfo(Value *V) { 169 V = V->stripPointerCasts(); 170 GlobalValue *GV = dyn_cast<GlobalValue>(V); 171 GlobalVariable *Var = dyn_cast<GlobalVariable>(V); 172 173 if (Var && Var->getName() == "llvm.eh.catch.all.value") { 174 assert(Var->hasInitializer() && 175 "The EH catch-all value must have an initializer"); 176 Value *Init = Var->getInitializer(); 177 GV = dyn_cast<GlobalValue>(Init); 178 if (!GV) V = cast<ConstantPointerNull>(Init); 179 } 180 181 assert((GV || isa<ConstantPointerNull>(V)) && 182 "TypeInfo must be a global variable or NULL"); 183 return GV; 184 } 185 186 /// getFCmpCondCode - Return the ISD condition code corresponding to 187 /// the given LLVM IR floating-point condition code. This includes 188 /// consideration of global floating-point math flags. 189 /// 190 ISD::CondCode llvm::getFCmpCondCode(FCmpInst::Predicate Pred) { 191 switch (Pred) { 192 case FCmpInst::FCMP_FALSE: return ISD::SETFALSE; 193 case FCmpInst::FCMP_OEQ: return ISD::SETOEQ; 194 case FCmpInst::FCMP_OGT: return ISD::SETOGT; 195 case FCmpInst::FCMP_OGE: return ISD::SETOGE; 196 case FCmpInst::FCMP_OLT: return ISD::SETOLT; 197 case FCmpInst::FCMP_OLE: return ISD::SETOLE; 198 case FCmpInst::FCMP_ONE: return ISD::SETONE; 199 case FCmpInst::FCMP_ORD: return ISD::SETO; 200 case FCmpInst::FCMP_UNO: return ISD::SETUO; 201 case FCmpInst::FCMP_UEQ: return ISD::SETUEQ; 202 case FCmpInst::FCMP_UGT: return ISD::SETUGT; 203 case FCmpInst::FCMP_UGE: return ISD::SETUGE; 204 case FCmpInst::FCMP_ULT: return ISD::SETULT; 205 case FCmpInst::FCMP_ULE: return ISD::SETULE; 206 case FCmpInst::FCMP_UNE: return ISD::SETUNE; 207 case FCmpInst::FCMP_TRUE: return ISD::SETTRUE; 208 default: llvm_unreachable("Invalid FCmp predicate opcode!"); 209 } 210 } 211 212 ISD::CondCode llvm::getFCmpCodeWithoutNaN(ISD::CondCode CC) { 213 switch (CC) { 214 case ISD::SETOEQ: case ISD::SETUEQ: return ISD::SETEQ; 215 case ISD::SETONE: case ISD::SETUNE: return ISD::SETNE; 216 case ISD::SETOLT: case ISD::SETULT: return ISD::SETLT; 217 case ISD::SETOLE: case ISD::SETULE: return ISD::SETLE; 218 case ISD::SETOGT: case ISD::SETUGT: return ISD::SETGT; 219 case ISD::SETOGE: case ISD::SETUGE: return ISD::SETGE; 220 default: return CC; 221 } 222 } 223 224 ISD::CondCode llvm::getICmpCondCode(ICmpInst::Predicate Pred) { 225 switch (Pred) { 226 case ICmpInst::ICMP_EQ: return ISD::SETEQ; 227 case ICmpInst::ICMP_NE: return ISD::SETNE; 228 case ICmpInst::ICMP_SLE: return ISD::SETLE; 229 case ICmpInst::ICMP_ULE: return ISD::SETULE; 230 case ICmpInst::ICMP_SGE: return ISD::SETGE; 231 case ICmpInst::ICMP_UGE: return ISD::SETUGE; 232 case ICmpInst::ICMP_SLT: return ISD::SETLT; 233 case ICmpInst::ICMP_ULT: return ISD::SETULT; 234 case ICmpInst::ICMP_SGT: return ISD::SETGT; 235 case ICmpInst::ICMP_UGT: return ISD::SETUGT; 236 default: 237 llvm_unreachable("Invalid ICmp predicate opcode!"); 238 } 239 } 240 241 ICmpInst::Predicate llvm::getICmpCondCode(ISD::CondCode Pred) { 242 switch (Pred) { 243 case ISD::SETEQ: 244 return ICmpInst::ICMP_EQ; 245 case ISD::SETNE: 246 return ICmpInst::ICMP_NE; 247 case ISD::SETLE: 248 return ICmpInst::ICMP_SLE; 249 case ISD::SETULE: 250 return ICmpInst::ICMP_ULE; 251 case ISD::SETGE: 252 return ICmpInst::ICMP_SGE; 253 case ISD::SETUGE: 254 return ICmpInst::ICMP_UGE; 255 case ISD::SETLT: 256 return ICmpInst::ICMP_SLT; 257 case ISD::SETULT: 258 return ICmpInst::ICMP_ULT; 259 case ISD::SETGT: 260 return ICmpInst::ICMP_SGT; 261 case ISD::SETUGT: 262 return ICmpInst::ICMP_UGT; 263 default: 264 llvm_unreachable("Invalid ISD integer condition code!"); 265 } 266 } 267 268 static bool isNoopBitcast(Type *T1, Type *T2, 269 const TargetLoweringBase& TLI) { 270 return T1 == T2 || (T1->isPointerTy() && T2->isPointerTy()) || 271 (isa<VectorType>(T1) && isa<VectorType>(T2) && 272 TLI.isTypeLegal(EVT::getEVT(T1)) && TLI.isTypeLegal(EVT::getEVT(T2))); 273 } 274 275 /// Look through operations that will be free to find the earliest source of 276 /// this value. 277 /// 278 /// @param ValLoc If V has aggregate type, we will be interested in a particular 279 /// scalar component. This records its address; the reverse of this list gives a 280 /// sequence of indices appropriate for an extractvalue to locate the important 281 /// value. This value is updated during the function and on exit will indicate 282 /// similar information for the Value returned. 283 /// 284 /// @param DataBits If this function looks through truncate instructions, this 285 /// will record the smallest size attained. 286 static const Value *getNoopInput(const Value *V, 287 SmallVectorImpl<unsigned> &ValLoc, 288 unsigned &DataBits, 289 const TargetLoweringBase &TLI, 290 const DataLayout &DL) { 291 while (true) { 292 // Try to look through V1; if V1 is not an instruction, it can't be looked 293 // through. 294 const Instruction *I = dyn_cast<Instruction>(V); 295 if (!I || I->getNumOperands() == 0) return V; 296 const Value *NoopInput = nullptr; 297 298 Value *Op = I->getOperand(0); 299 if (isa<BitCastInst>(I)) { 300 // Look through truly no-op bitcasts. 301 if (isNoopBitcast(Op->getType(), I->getType(), TLI)) 302 NoopInput = Op; 303 } else if (isa<GetElementPtrInst>(I)) { 304 // Look through getelementptr 305 if (cast<GetElementPtrInst>(I)->hasAllZeroIndices()) 306 NoopInput = Op; 307 } else if (isa<IntToPtrInst>(I)) { 308 // Look through inttoptr. 309 // Make sure this isn't a truncating or extending cast. We could 310 // support this eventually, but don't bother for now. 311 if (!isa<VectorType>(I->getType()) && 312 DL.getPointerSizeInBits() == 313 cast<IntegerType>(Op->getType())->getBitWidth()) 314 NoopInput = Op; 315 } else if (isa<PtrToIntInst>(I)) { 316 // Look through ptrtoint. 317 // Make sure this isn't a truncating or extending cast. We could 318 // support this eventually, but don't bother for now. 319 if (!isa<VectorType>(I->getType()) && 320 DL.getPointerSizeInBits() == 321 cast<IntegerType>(I->getType())->getBitWidth()) 322 NoopInput = Op; 323 } else if (isa<TruncInst>(I) && 324 TLI.allowTruncateForTailCall(Op->getType(), I->getType())) { 325 DataBits = std::min((uint64_t)DataBits, 326 I->getType()->getPrimitiveSizeInBits().getFixedSize()); 327 NoopInput = Op; 328 } else if (auto *CB = dyn_cast<CallBase>(I)) { 329 const Value *ReturnedOp = CB->getReturnedArgOperand(); 330 if (ReturnedOp && isNoopBitcast(ReturnedOp->getType(), I->getType(), TLI)) 331 NoopInput = ReturnedOp; 332 } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(V)) { 333 // Value may come from either the aggregate or the scalar 334 ArrayRef<unsigned> InsertLoc = IVI->getIndices(); 335 if (ValLoc.size() >= InsertLoc.size() && 336 std::equal(InsertLoc.begin(), InsertLoc.end(), ValLoc.rbegin())) { 337 // The type being inserted is a nested sub-type of the aggregate; we 338 // have to remove those initial indices to get the location we're 339 // interested in for the operand. 340 ValLoc.resize(ValLoc.size() - InsertLoc.size()); 341 NoopInput = IVI->getInsertedValueOperand(); 342 } else { 343 // The struct we're inserting into has the value we're interested in, no 344 // change of address. 345 NoopInput = Op; 346 } 347 } else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) { 348 // The part we're interested in will inevitably be some sub-section of the 349 // previous aggregate. Combine the two paths to obtain the true address of 350 // our element. 351 ArrayRef<unsigned> ExtractLoc = EVI->getIndices(); 352 ValLoc.append(ExtractLoc.rbegin(), ExtractLoc.rend()); 353 NoopInput = Op; 354 } 355 // Terminate if we couldn't find anything to look through. 356 if (!NoopInput) 357 return V; 358 359 V = NoopInput; 360 } 361 } 362 363 /// Return true if this scalar return value only has bits discarded on its path 364 /// from the "tail call" to the "ret". This includes the obvious noop 365 /// instructions handled by getNoopInput above as well as free truncations (or 366 /// extensions prior to the call). 367 static bool slotOnlyDiscardsData(const Value *RetVal, const Value *CallVal, 368 SmallVectorImpl<unsigned> &RetIndices, 369 SmallVectorImpl<unsigned> &CallIndices, 370 bool AllowDifferingSizes, 371 const TargetLoweringBase &TLI, 372 const DataLayout &DL) { 373 374 // Trace the sub-value needed by the return value as far back up the graph as 375 // possible, in the hope that it will intersect with the value produced by the 376 // call. In the simple case with no "returned" attribute, the hope is actually 377 // that we end up back at the tail call instruction itself. 378 unsigned BitsRequired = UINT_MAX; 379 RetVal = getNoopInput(RetVal, RetIndices, BitsRequired, TLI, DL); 380 381 // If this slot in the value returned is undef, it doesn't matter what the 382 // call puts there, it'll be fine. 383 if (isa<UndefValue>(RetVal)) 384 return true; 385 386 // Now do a similar search up through the graph to find where the value 387 // actually returned by the "tail call" comes from. In the simple case without 388 // a "returned" attribute, the search will be blocked immediately and the loop 389 // a Noop. 390 unsigned BitsProvided = UINT_MAX; 391 CallVal = getNoopInput(CallVal, CallIndices, BitsProvided, TLI, DL); 392 393 // There's no hope if we can't actually trace them to (the same part of!) the 394 // same value. 395 if (CallVal != RetVal || CallIndices != RetIndices) 396 return false; 397 398 // However, intervening truncates may have made the call non-tail. Make sure 399 // all the bits that are needed by the "ret" have been provided by the "tail 400 // call". FIXME: with sufficiently cunning bit-tracking, we could look through 401 // extensions too. 402 if (BitsProvided < BitsRequired || 403 (!AllowDifferingSizes && BitsProvided != BitsRequired)) 404 return false; 405 406 return true; 407 } 408 409 /// For an aggregate type, determine whether a given index is within bounds or 410 /// not. 411 static bool indexReallyValid(Type *T, unsigned Idx) { 412 if (ArrayType *AT = dyn_cast<ArrayType>(T)) 413 return Idx < AT->getNumElements(); 414 415 return Idx < cast<StructType>(T)->getNumElements(); 416 } 417 418 /// Move the given iterators to the next leaf type in depth first traversal. 419 /// 420 /// Performs a depth-first traversal of the type as specified by its arguments, 421 /// stopping at the next leaf node (which may be a legitimate scalar type or an 422 /// empty struct or array). 423 /// 424 /// @param SubTypes List of the partial components making up the type from 425 /// outermost to innermost non-empty aggregate. The element currently 426 /// represented is SubTypes.back()->getTypeAtIndex(Path.back() - 1). 427 /// 428 /// @param Path Set of extractvalue indices leading from the outermost type 429 /// (SubTypes[0]) to the leaf node currently represented. 430 /// 431 /// @returns true if a new type was found, false otherwise. Calling this 432 /// function again on a finished iterator will repeatedly return 433 /// false. SubTypes.back()->getTypeAtIndex(Path.back()) is either an empty 434 /// aggregate or a non-aggregate 435 static bool advanceToNextLeafType(SmallVectorImpl<Type *> &SubTypes, 436 SmallVectorImpl<unsigned> &Path) { 437 // First march back up the tree until we can successfully increment one of the 438 // coordinates in Path. 439 while (!Path.empty() && !indexReallyValid(SubTypes.back(), Path.back() + 1)) { 440 Path.pop_back(); 441 SubTypes.pop_back(); 442 } 443 444 // If we reached the top, then the iterator is done. 445 if (Path.empty()) 446 return false; 447 448 // We know there's *some* valid leaf now, so march back down the tree picking 449 // out the left-most element at each node. 450 ++Path.back(); 451 Type *DeeperType = 452 ExtractValueInst::getIndexedType(SubTypes.back(), Path.back()); 453 while (DeeperType->isAggregateType()) { 454 if (!indexReallyValid(DeeperType, 0)) 455 return true; 456 457 SubTypes.push_back(DeeperType); 458 Path.push_back(0); 459 460 DeeperType = ExtractValueInst::getIndexedType(DeeperType, 0); 461 } 462 463 return true; 464 } 465 466 /// Find the first non-empty, scalar-like type in Next and setup the iterator 467 /// components. 468 /// 469 /// Assuming Next is an aggregate of some kind, this function will traverse the 470 /// tree from left to right (i.e. depth-first) looking for the first 471 /// non-aggregate type which will play a role in function return. 472 /// 473 /// For example, if Next was {[0 x i64], {{}, i32, {}}, i32} then we would setup 474 /// Path as [1, 1] and SubTypes as [Next, {{}, i32, {}}] to represent the first 475 /// i32 in that type. 476 static bool firstRealType(Type *Next, SmallVectorImpl<Type *> &SubTypes, 477 SmallVectorImpl<unsigned> &Path) { 478 // First initialise the iterator components to the first "leaf" node 479 // (i.e. node with no valid sub-type at any index, so {} does count as a leaf 480 // despite nominally being an aggregate). 481 while (Type *FirstInner = ExtractValueInst::getIndexedType(Next, 0)) { 482 SubTypes.push_back(Next); 483 Path.push_back(0); 484 Next = FirstInner; 485 } 486 487 // If there's no Path now, Next was originally scalar already (or empty 488 // leaf). We're done. 489 if (Path.empty()) 490 return true; 491 492 // Otherwise, use normal iteration to keep looking through the tree until we 493 // find a non-aggregate type. 494 while (ExtractValueInst::getIndexedType(SubTypes.back(), Path.back()) 495 ->isAggregateType()) { 496 if (!advanceToNextLeafType(SubTypes, Path)) 497 return false; 498 } 499 500 return true; 501 } 502 503 /// Set the iterator data-structures to the next non-empty, non-aggregate 504 /// subtype. 505 static bool nextRealType(SmallVectorImpl<Type *> &SubTypes, 506 SmallVectorImpl<unsigned> &Path) { 507 do { 508 if (!advanceToNextLeafType(SubTypes, Path)) 509 return false; 510 511 assert(!Path.empty() && "found a leaf but didn't set the path?"); 512 } while (ExtractValueInst::getIndexedType(SubTypes.back(), Path.back()) 513 ->isAggregateType()); 514 515 return true; 516 } 517 518 519 /// Test if the given instruction is in a position to be optimized 520 /// with a tail-call. This roughly means that it's in a block with 521 /// a return and there's nothing that needs to be scheduled 522 /// between it and the return. 523 /// 524 /// This function only tests target-independent requirements. 525 bool llvm::isInTailCallPosition(const CallBase &Call, const TargetMachine &TM) { 526 const BasicBlock *ExitBB = Call.getParent(); 527 const Instruction *Term = ExitBB->getTerminator(); 528 const ReturnInst *Ret = dyn_cast<ReturnInst>(Term); 529 530 // The block must end in a return statement or unreachable. 531 // 532 // FIXME: Decline tailcall if it's not guaranteed and if the block ends in 533 // an unreachable, for now. The way tailcall optimization is currently 534 // implemented means it will add an epilogue followed by a jump. That is 535 // not profitable. Also, if the callee is a special function (e.g. 536 // longjmp on x86), it can end up causing miscompilation that has not 537 // been fully understood. 538 if (!Ret && ((!TM.Options.GuaranteedTailCallOpt && 539 Call.getCallingConv() != CallingConv::Tail && 540 Call.getCallingConv() != CallingConv::SwiftTail) || 541 !isa<UnreachableInst>(Term))) 542 return false; 543 544 // If I will have a chain, make sure no other instruction that will have a 545 // chain interposes between I and the return. 546 // Check for all calls including speculatable functions. 547 for (BasicBlock::const_iterator BBI = std::prev(ExitBB->end(), 2);; --BBI) { 548 if (&*BBI == &Call) 549 break; 550 // Debug info intrinsics do not get in the way of tail call optimization. 551 // Pseudo probe intrinsics do not block tail call optimization either. 552 if (BBI->isDebugOrPseudoInst()) 553 continue; 554 // A lifetime end, assume or noalias.decl intrinsic should not stop tail 555 // call optimization. 556 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI)) 557 if (II->getIntrinsicID() == Intrinsic::lifetime_end || 558 II->getIntrinsicID() == Intrinsic::assume || 559 II->getIntrinsicID() == Intrinsic::experimental_noalias_scope_decl) 560 continue; 561 if (BBI->mayHaveSideEffects() || BBI->mayReadFromMemory() || 562 !isSafeToSpeculativelyExecute(&*BBI)) 563 return false; 564 } 565 566 const Function *F = ExitBB->getParent(); 567 return returnTypeIsEligibleForTailCall( 568 F, &Call, Ret, *TM.getSubtargetImpl(*F)->getTargetLowering()); 569 } 570 571 bool llvm::attributesPermitTailCall(const Function *F, const Instruction *I, 572 const ReturnInst *Ret, 573 const TargetLoweringBase &TLI, 574 bool *AllowDifferingSizes) { 575 // ADS may be null, so don't write to it directly. 576 bool DummyADS; 577 bool &ADS = AllowDifferingSizes ? *AllowDifferingSizes : DummyADS; 578 ADS = true; 579 580 AttrBuilder CallerAttrs(F->getContext(), F->getAttributes().getRetAttrs()); 581 AttrBuilder CalleeAttrs(F->getContext(), 582 cast<CallInst>(I)->getAttributes().getRetAttrs()); 583 584 // Following attributes are completely benign as far as calling convention 585 // goes, they shouldn't affect whether the call is a tail call. 586 for (const auto &Attr : {Attribute::Alignment, Attribute::Dereferenceable, 587 Attribute::DereferenceableOrNull, Attribute::NoAlias, 588 Attribute::NonNull, Attribute::NoUndef}) { 589 CallerAttrs.removeAttribute(Attr); 590 CalleeAttrs.removeAttribute(Attr); 591 } 592 593 if (CallerAttrs.contains(Attribute::ZExt)) { 594 if (!CalleeAttrs.contains(Attribute::ZExt)) 595 return false; 596 597 ADS = false; 598 CallerAttrs.removeAttribute(Attribute::ZExt); 599 CalleeAttrs.removeAttribute(Attribute::ZExt); 600 } else if (CallerAttrs.contains(Attribute::SExt)) { 601 if (!CalleeAttrs.contains(Attribute::SExt)) 602 return false; 603 604 ADS = false; 605 CallerAttrs.removeAttribute(Attribute::SExt); 606 CalleeAttrs.removeAttribute(Attribute::SExt); 607 } 608 609 // Drop sext and zext return attributes if the result is not used. 610 // This enables tail calls for code like: 611 // 612 // define void @caller() { 613 // entry: 614 // %unused_result = tail call zeroext i1 @callee() 615 // br label %retlabel 616 // retlabel: 617 // ret void 618 // } 619 if (I->use_empty()) { 620 CalleeAttrs.removeAttribute(Attribute::SExt); 621 CalleeAttrs.removeAttribute(Attribute::ZExt); 622 } 623 624 // If they're still different, there's some facet we don't understand 625 // (currently only "inreg", but in future who knows). It may be OK but the 626 // only safe option is to reject the tail call. 627 return CallerAttrs == CalleeAttrs; 628 } 629 630 /// Check whether B is a bitcast of a pointer type to another pointer type, 631 /// which is equal to A. 632 static bool isPointerBitcastEqualTo(const Value *A, const Value *B) { 633 assert(A && B && "Expected non-null inputs!"); 634 635 auto *BitCastIn = dyn_cast<BitCastInst>(B); 636 637 if (!BitCastIn) 638 return false; 639 640 if (!A->getType()->isPointerTy() || !B->getType()->isPointerTy()) 641 return false; 642 643 return A == BitCastIn->getOperand(0); 644 } 645 646 bool llvm::returnTypeIsEligibleForTailCall(const Function *F, 647 const Instruction *I, 648 const ReturnInst *Ret, 649 const TargetLoweringBase &TLI) { 650 // If the block ends with a void return or unreachable, it doesn't matter 651 // what the call's return type is. 652 if (!Ret || Ret->getNumOperands() == 0) return true; 653 654 // If the return value is undef, it doesn't matter what the call's 655 // return type is. 656 if (isa<UndefValue>(Ret->getOperand(0))) return true; 657 658 // Make sure the attributes attached to each return are compatible. 659 bool AllowDifferingSizes; 660 if (!attributesPermitTailCall(F, I, Ret, TLI, &AllowDifferingSizes)) 661 return false; 662 663 const Value *RetVal = Ret->getOperand(0), *CallVal = I; 664 // Intrinsic like llvm.memcpy has no return value, but the expanded 665 // libcall may or may not have return value. On most platforms, it 666 // will be expanded as memcpy in libc, which returns the first 667 // argument. On other platforms like arm-none-eabi, memcpy may be 668 // expanded as library call without return value, like __aeabi_memcpy. 669 const CallInst *Call = cast<CallInst>(I); 670 if (Function *F = Call->getCalledFunction()) { 671 Intrinsic::ID IID = F->getIntrinsicID(); 672 if (((IID == Intrinsic::memcpy && 673 TLI.getLibcallName(RTLIB::MEMCPY) == StringRef("memcpy")) || 674 (IID == Intrinsic::memmove && 675 TLI.getLibcallName(RTLIB::MEMMOVE) == StringRef("memmove")) || 676 (IID == Intrinsic::memset && 677 TLI.getLibcallName(RTLIB::MEMSET) == StringRef("memset"))) && 678 (RetVal == Call->getArgOperand(0) || 679 isPointerBitcastEqualTo(RetVal, Call->getArgOperand(0)))) 680 return true; 681 } 682 683 SmallVector<unsigned, 4> RetPath, CallPath; 684 SmallVector<Type *, 4> RetSubTypes, CallSubTypes; 685 686 bool RetEmpty = !firstRealType(RetVal->getType(), RetSubTypes, RetPath); 687 bool CallEmpty = !firstRealType(CallVal->getType(), CallSubTypes, CallPath); 688 689 // Nothing's actually returned, it doesn't matter what the callee put there 690 // it's a valid tail call. 691 if (RetEmpty) 692 return true; 693 694 // Iterate pairwise through each of the value types making up the tail call 695 // and the corresponding return. For each one we want to know whether it's 696 // essentially going directly from the tail call to the ret, via operations 697 // that end up not generating any code. 698 // 699 // We allow a certain amount of covariance here. For example it's permitted 700 // for the tail call to define more bits than the ret actually cares about 701 // (e.g. via a truncate). 702 do { 703 if (CallEmpty) { 704 // We've exhausted the values produced by the tail call instruction, the 705 // rest are essentially undef. The type doesn't really matter, but we need 706 // *something*. 707 Type *SlotType = 708 ExtractValueInst::getIndexedType(RetSubTypes.back(), RetPath.back()); 709 CallVal = UndefValue::get(SlotType); 710 } 711 712 // The manipulations performed when we're looking through an insertvalue or 713 // an extractvalue would happen at the front of the RetPath list, so since 714 // we have to copy it anyway it's more efficient to create a reversed copy. 715 SmallVector<unsigned, 4> TmpRetPath(llvm::reverse(RetPath)); 716 SmallVector<unsigned, 4> TmpCallPath(llvm::reverse(CallPath)); 717 718 // Finally, we can check whether the value produced by the tail call at this 719 // index is compatible with the value we return. 720 if (!slotOnlyDiscardsData(RetVal, CallVal, TmpRetPath, TmpCallPath, 721 AllowDifferingSizes, TLI, 722 F->getParent()->getDataLayout())) 723 return false; 724 725 CallEmpty = !nextRealType(CallSubTypes, CallPath); 726 } while(nextRealType(RetSubTypes, RetPath)); 727 728 return true; 729 } 730 731 static void collectEHScopeMembers( 732 DenseMap<const MachineBasicBlock *, int> &EHScopeMembership, int EHScope, 733 const MachineBasicBlock *MBB) { 734 SmallVector<const MachineBasicBlock *, 16> Worklist = {MBB}; 735 while (!Worklist.empty()) { 736 const MachineBasicBlock *Visiting = Worklist.pop_back_val(); 737 // Don't follow blocks which start new scopes. 738 if (Visiting->isEHPad() && Visiting != MBB) 739 continue; 740 741 // Add this MBB to our scope. 742 auto P = EHScopeMembership.insert(std::make_pair(Visiting, EHScope)); 743 744 // Don't revisit blocks. 745 if (!P.second) { 746 assert(P.first->second == EHScope && "MBB is part of two scopes!"); 747 continue; 748 } 749 750 // Returns are boundaries where scope transfer can occur, don't follow 751 // successors. 752 if (Visiting->isEHScopeReturnBlock()) 753 continue; 754 755 append_range(Worklist, Visiting->successors()); 756 } 757 } 758 759 DenseMap<const MachineBasicBlock *, int> 760 llvm::getEHScopeMembership(const MachineFunction &MF) { 761 DenseMap<const MachineBasicBlock *, int> EHScopeMembership; 762 763 // We don't have anything to do if there aren't any EH pads. 764 if (!MF.hasEHScopes()) 765 return EHScopeMembership; 766 767 int EntryBBNumber = MF.front().getNumber(); 768 bool IsSEH = isAsynchronousEHPersonality( 769 classifyEHPersonality(MF.getFunction().getPersonalityFn())); 770 771 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); 772 SmallVector<const MachineBasicBlock *, 16> EHScopeBlocks; 773 SmallVector<const MachineBasicBlock *, 16> UnreachableBlocks; 774 SmallVector<const MachineBasicBlock *, 16> SEHCatchPads; 775 SmallVector<std::pair<const MachineBasicBlock *, int>, 16> CatchRetSuccessors; 776 for (const MachineBasicBlock &MBB : MF) { 777 if (MBB.isEHScopeEntry()) { 778 EHScopeBlocks.push_back(&MBB); 779 } else if (IsSEH && MBB.isEHPad()) { 780 SEHCatchPads.push_back(&MBB); 781 } else if (MBB.pred_empty()) { 782 UnreachableBlocks.push_back(&MBB); 783 } 784 785 MachineBasicBlock::const_iterator MBBI = MBB.getFirstTerminator(); 786 787 // CatchPads are not scopes for SEH so do not consider CatchRet to 788 // transfer control to another scope. 789 if (MBBI == MBB.end() || MBBI->getOpcode() != TII->getCatchReturnOpcode()) 790 continue; 791 792 // FIXME: SEH CatchPads are not necessarily in the parent function: 793 // they could be inside a finally block. 794 const MachineBasicBlock *Successor = MBBI->getOperand(0).getMBB(); 795 const MachineBasicBlock *SuccessorColor = MBBI->getOperand(1).getMBB(); 796 CatchRetSuccessors.push_back( 797 {Successor, IsSEH ? EntryBBNumber : SuccessorColor->getNumber()}); 798 } 799 800 // We don't have anything to do if there aren't any EH pads. 801 if (EHScopeBlocks.empty()) 802 return EHScopeMembership; 803 804 // Identify all the basic blocks reachable from the function entry. 805 collectEHScopeMembers(EHScopeMembership, EntryBBNumber, &MF.front()); 806 // All blocks not part of a scope are in the parent function. 807 for (const MachineBasicBlock *MBB : UnreachableBlocks) 808 collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB); 809 // Next, identify all the blocks inside the scopes. 810 for (const MachineBasicBlock *MBB : EHScopeBlocks) 811 collectEHScopeMembers(EHScopeMembership, MBB->getNumber(), MBB); 812 // SEH CatchPads aren't really scopes, handle them separately. 813 for (const MachineBasicBlock *MBB : SEHCatchPads) 814 collectEHScopeMembers(EHScopeMembership, EntryBBNumber, MBB); 815 // Finally, identify all the targets of a catchret. 816 for (std::pair<const MachineBasicBlock *, int> CatchRetPair : 817 CatchRetSuccessors) 818 collectEHScopeMembers(EHScopeMembership, CatchRetPair.second, 819 CatchRetPair.first); 820 return EHScopeMembership; 821 } 822