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