LLVM API Documentation

CodeGen/Analysis.cpp
Go to the documentation of this file.
00001 //===-- Analysis.cpp - CodeGen LLVM IR Analysis Utilities -----------------===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This file defines several CodeGen-specific LLVM IR analysis utilties.
00011 //
00012 //===----------------------------------------------------------------------===//
00013 
00014 #include "llvm/CodeGen/Analysis.h"
00015 #include "llvm/Analysis/ValueTracking.h"
00016 #include "llvm/CodeGen/MachineFunction.h"
00017 #include "llvm/IR/DataLayout.h"
00018 #include "llvm/IR/DerivedTypes.h"
00019 #include "llvm/IR/Function.h"
00020 #include "llvm/IR/Instructions.h"
00021 #include "llvm/IR/IntrinsicInst.h"
00022 #include "llvm/IR/LLVMContext.h"
00023 #include "llvm/IR/Module.h"
00024 #include "llvm/Support/ErrorHandling.h"
00025 #include "llvm/Support/MathExtras.h"
00026 #include "llvm/Target/TargetLowering.h"
00027 using namespace llvm;
00028 
00029 /// ComputeLinearIndex - Given an LLVM IR aggregate type and a sequence
00030 /// of insertvalue or extractvalue indices that identify a member, return
00031 /// the linearized index of the start of the member.
00032 ///
00033 unsigned llvm::ComputeLinearIndex(Type *Ty,
00034                                   const unsigned *Indices,
00035                                   const unsigned *IndicesEnd,
00036                                   unsigned CurIndex) {
00037   // Base case: We're done.
00038   if (Indices && Indices == IndicesEnd)
00039     return CurIndex;
00040 
00041   // Given a struct type, recursively traverse the elements.
00042   if (StructType *STy = dyn_cast<StructType>(Ty)) {
00043     for (StructType::element_iterator EB = STy->element_begin(),
00044                                       EI = EB,
00045                                       EE = STy->element_end();
00046         EI != EE; ++EI) {
00047       if (Indices && *Indices == unsigned(EI - EB))
00048         return ComputeLinearIndex(*EI, Indices+1, IndicesEnd, CurIndex);
00049       CurIndex = ComputeLinearIndex(*EI, nullptr, nullptr, CurIndex);
00050     }
00051     return CurIndex;
00052   }
00053   // Given an array type, recursively traverse the elements.
00054   else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
00055     Type *EltTy = ATy->getElementType();
00056     for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) {
00057       if (Indices && *Indices == i)
00058         return ComputeLinearIndex(EltTy, Indices+1, IndicesEnd, CurIndex);
00059       CurIndex = ComputeLinearIndex(EltTy, nullptr, nullptr, CurIndex);
00060     }
00061     return CurIndex;
00062   }
00063   // We haven't found the type we're looking for, so keep searching.
00064   return CurIndex + 1;
00065 }
00066 
00067 /// ComputeValueVTs - Given an LLVM IR type, compute a sequence of
00068 /// EVTs that represent all the individual underlying
00069 /// non-aggregate types that comprise it.
00070 ///
00071 /// If Offsets is non-null, it points to a vector to be filled in
00072 /// with the in-memory offsets of each of the individual values.
00073 ///
00074 void llvm::ComputeValueVTs(const TargetLowering &TLI, Type *Ty,
00075                            SmallVectorImpl<EVT> &ValueVTs,
00076                            SmallVectorImpl<uint64_t> *Offsets,
00077                            uint64_t StartingOffset) {
00078   // Given a struct type, recursively traverse the elements.
00079   if (StructType *STy = dyn_cast<StructType>(Ty)) {
00080     const StructLayout *SL = TLI.getDataLayout()->getStructLayout(STy);
00081     for (StructType::element_iterator EB = STy->element_begin(),
00082                                       EI = EB,
00083                                       EE = STy->element_end();
00084          EI != EE; ++EI)
00085       ComputeValueVTs(TLI, *EI, ValueVTs, Offsets,
00086                       StartingOffset + SL->getElementOffset(EI - EB));
00087     return;
00088   }
00089   // Given an array type, recursively traverse the elements.
00090   if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
00091     Type *EltTy = ATy->getElementType();
00092     uint64_t EltSize = TLI.getDataLayout()->getTypeAllocSize(EltTy);
00093     for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
00094       ComputeValueVTs(TLI, EltTy, ValueVTs, Offsets,
00095                       StartingOffset + i * EltSize);
00096     return;
00097   }
00098   // Interpret void as zero return values.
00099   if (Ty->isVoidTy())
00100     return;
00101   // Base case: we can get an EVT for this LLVM IR type.
00102   ValueVTs.push_back(TLI.getValueType(Ty));
00103   if (Offsets)
00104     Offsets->push_back(StartingOffset);
00105 }
00106 
00107 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
00108 GlobalVariable *llvm::ExtractTypeInfo(Value *V) {
00109   V = V->stripPointerCasts();
00110   GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
00111 
00112   if (GV && GV->getName() == "llvm.eh.catch.all.value") {
00113     assert(GV->hasInitializer() &&
00114            "The EH catch-all value must have an initializer");
00115     Value *Init = GV->getInitializer();
00116     GV = dyn_cast<GlobalVariable>(Init);
00117     if (!GV) V = cast<ConstantPointerNull>(Init);
00118   }
00119 
00120   assert((GV || isa<ConstantPointerNull>(V)) &&
00121          "TypeInfo must be a global variable or NULL");
00122   return GV;
00123 }
00124 
00125 /// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
00126 /// processed uses a memory 'm' constraint.
00127 bool
00128 llvm::hasInlineAsmMemConstraint(InlineAsm::ConstraintInfoVector &CInfos,
00129                                 const TargetLowering &TLI) {
00130   for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
00131     InlineAsm::ConstraintInfo &CI = CInfos[i];
00132     for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
00133       TargetLowering::ConstraintType CType = TLI.getConstraintType(CI.Codes[j]);
00134       if (CType == TargetLowering::C_Memory)
00135         return true;
00136     }
00137 
00138     // Indirect operand accesses access memory.
00139     if (CI.isIndirect)
00140       return true;
00141   }
00142 
00143   return false;
00144 }
00145 
00146 /// getFCmpCondCode - Return the ISD condition code corresponding to
00147 /// the given LLVM IR floating-point condition code.  This includes
00148 /// consideration of global floating-point math flags.
00149 ///
00150 ISD::CondCode llvm::getFCmpCondCode(FCmpInst::Predicate Pred) {
00151   switch (Pred) {
00152   case FCmpInst::FCMP_FALSE: return ISD::SETFALSE;
00153   case FCmpInst::FCMP_OEQ:   return ISD::SETOEQ;
00154   case FCmpInst::FCMP_OGT:   return ISD::SETOGT;
00155   case FCmpInst::FCMP_OGE:   return ISD::SETOGE;
00156   case FCmpInst::FCMP_OLT:   return ISD::SETOLT;
00157   case FCmpInst::FCMP_OLE:   return ISD::SETOLE;
00158   case FCmpInst::FCMP_ONE:   return ISD::SETONE;
00159   case FCmpInst::FCMP_ORD:   return ISD::SETO;
00160   case FCmpInst::FCMP_UNO:   return ISD::SETUO;
00161   case FCmpInst::FCMP_UEQ:   return ISD::SETUEQ;
00162   case FCmpInst::FCMP_UGT:   return ISD::SETUGT;
00163   case FCmpInst::FCMP_UGE:   return ISD::SETUGE;
00164   case FCmpInst::FCMP_ULT:   return ISD::SETULT;
00165   case FCmpInst::FCMP_ULE:   return ISD::SETULE;
00166   case FCmpInst::FCMP_UNE:   return ISD::SETUNE;
00167   case FCmpInst::FCMP_TRUE:  return ISD::SETTRUE;
00168   default: llvm_unreachable("Invalid FCmp predicate opcode!");
00169   }
00170 }
00171 
00172 ISD::CondCode llvm::getFCmpCodeWithoutNaN(ISD::CondCode CC) {
00173   switch (CC) {
00174     case ISD::SETOEQ: case ISD::SETUEQ: return ISD::SETEQ;
00175     case ISD::SETONE: case ISD::SETUNE: return ISD::SETNE;
00176     case ISD::SETOLT: case ISD::SETULT: return ISD::SETLT;
00177     case ISD::SETOLE: case ISD::SETULE: return ISD::SETLE;
00178     case ISD::SETOGT: case ISD::SETUGT: return ISD::SETGT;
00179     case ISD::SETOGE: case ISD::SETUGE: return ISD::SETGE;
00180     default: return CC;
00181   }
00182 }
00183 
00184 /// getICmpCondCode - Return the ISD condition code corresponding to
00185 /// the given LLVM IR integer condition code.
00186 ///
00187 ISD::CondCode llvm::getICmpCondCode(ICmpInst::Predicate Pred) {
00188   switch (Pred) {
00189   case ICmpInst::ICMP_EQ:  return ISD::SETEQ;
00190   case ICmpInst::ICMP_NE:  return ISD::SETNE;
00191   case ICmpInst::ICMP_SLE: return ISD::SETLE;
00192   case ICmpInst::ICMP_ULE: return ISD::SETULE;
00193   case ICmpInst::ICMP_SGE: return ISD::SETGE;
00194   case ICmpInst::ICMP_UGE: return ISD::SETUGE;
00195   case ICmpInst::ICMP_SLT: return ISD::SETLT;
00196   case ICmpInst::ICMP_ULT: return ISD::SETULT;
00197   case ICmpInst::ICMP_SGT: return ISD::SETGT;
00198   case ICmpInst::ICMP_UGT: return ISD::SETUGT;
00199   default:
00200     llvm_unreachable("Invalid ICmp predicate opcode!");
00201   }
00202 }
00203 
00204 static bool isNoopBitcast(Type *T1, Type *T2,
00205                           const TargetLoweringBase& TLI) {
00206   return T1 == T2 || (T1->isPointerTy() && T2->isPointerTy()) ||
00207          (isa<VectorType>(T1) && isa<VectorType>(T2) &&
00208           TLI.isTypeLegal(EVT::getEVT(T1)) && TLI.isTypeLegal(EVT::getEVT(T2)));
00209 }
00210 
00211 /// Look through operations that will be free to find the earliest source of
00212 /// this value.
00213 ///
00214 /// @param ValLoc If V has aggegate type, we will be interested in a particular
00215 /// scalar component. This records its address; the reverse of this list gives a
00216 /// sequence of indices appropriate for an extractvalue to locate the important
00217 /// value. This value is updated during the function and on exit will indicate
00218 /// similar information for the Value returned.
00219 ///
00220 /// @param DataBits If this function looks through truncate instructions, this
00221 /// will record the smallest size attained.
00222 static const Value *getNoopInput(const Value *V,
00223                                  SmallVectorImpl<unsigned> &ValLoc,
00224                                  unsigned &DataBits,
00225                                  const TargetLoweringBase &TLI) {
00226   while (true) {
00227     // Try to look through V1; if V1 is not an instruction, it can't be looked
00228     // through.
00229     const Instruction *I = dyn_cast<Instruction>(V);
00230     if (!I || I->getNumOperands() == 0) return V;
00231     const Value *NoopInput = nullptr;
00232 
00233     Value *Op = I->getOperand(0);
00234     if (isa<BitCastInst>(I)) {
00235       // Look through truly no-op bitcasts.
00236       if (isNoopBitcast(Op->getType(), I->getType(), TLI))
00237         NoopInput = Op;
00238     } else if (isa<GetElementPtrInst>(I)) {
00239       // Look through getelementptr
00240       if (cast<GetElementPtrInst>(I)->hasAllZeroIndices())
00241         NoopInput = Op;
00242     } else if (isa<IntToPtrInst>(I)) {
00243       // Look through inttoptr.
00244       // Make sure this isn't a truncating or extending cast.  We could
00245       // support this eventually, but don't bother for now.
00246       if (!isa<VectorType>(I->getType()) &&
00247           TLI.getPointerTy().getSizeInBits() ==
00248           cast<IntegerType>(Op->getType())->getBitWidth())
00249         NoopInput = Op;
00250     } else if (isa<PtrToIntInst>(I)) {
00251       // Look through ptrtoint.
00252       // Make sure this isn't a truncating or extending cast.  We could
00253       // support this eventually, but don't bother for now.
00254       if (!isa<VectorType>(I->getType()) &&
00255           TLI.getPointerTy().getSizeInBits() ==
00256           cast<IntegerType>(I->getType())->getBitWidth())
00257         NoopInput = Op;
00258     } else if (isa<TruncInst>(I) &&
00259                TLI.allowTruncateForTailCall(Op->getType(), I->getType())) {
00260       DataBits = std::min(DataBits, I->getType()->getPrimitiveSizeInBits());
00261       NoopInput = Op;
00262     } else if (isa<CallInst>(I)) {
00263       // Look through call (skipping callee)
00264       for (User::const_op_iterator i = I->op_begin(), e = I->op_end() - 1;
00265            i != e; ++i) {
00266         unsigned attrInd = i - I->op_begin() + 1;
00267         if (cast<CallInst>(I)->paramHasAttr(attrInd, Attribute::Returned) &&
00268             isNoopBitcast((*i)->getType(), I->getType(), TLI)) {
00269           NoopInput = *i;
00270           break;
00271         }
00272       }
00273     } else if (isa<InvokeInst>(I)) {
00274       // Look through invoke (skipping BB, BB, Callee)
00275       for (User::const_op_iterator i = I->op_begin(), e = I->op_end() - 3;
00276            i != e; ++i) {
00277         unsigned attrInd = i - I->op_begin() + 1;
00278         if (cast<InvokeInst>(I)->paramHasAttr(attrInd, Attribute::Returned) &&
00279             isNoopBitcast((*i)->getType(), I->getType(), TLI)) {
00280           NoopInput = *i;
00281           break;
00282         }
00283       }
00284     } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(V)) {
00285       // Value may come from either the aggregate or the scalar
00286       ArrayRef<unsigned> InsertLoc = IVI->getIndices();
00287       if (std::equal(InsertLoc.rbegin(), InsertLoc.rend(),
00288                      ValLoc.rbegin())) {
00289         // The type being inserted is a nested sub-type of the aggregate; we
00290         // have to remove those initial indices to get the location we're
00291         // interested in for the operand.
00292         ValLoc.resize(ValLoc.size() - InsertLoc.size());
00293         NoopInput = IVI->getInsertedValueOperand();
00294       } else {
00295         // The struct we're inserting into has the value we're interested in, no
00296         // change of address.
00297         NoopInput = Op;
00298       }
00299     } else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) {
00300       // The part we're interested in will inevitably be some sub-section of the
00301       // previous aggregate. Combine the two paths to obtain the true address of
00302       // our element.
00303       ArrayRef<unsigned> ExtractLoc = EVI->getIndices();
00304       std::copy(ExtractLoc.rbegin(), ExtractLoc.rend(),
00305                 std::back_inserter(ValLoc));
00306       NoopInput = Op;
00307     }
00308     // Terminate if we couldn't find anything to look through.
00309     if (!NoopInput)
00310       return V;
00311 
00312     V = NoopInput;
00313   }
00314 }
00315 
00316 /// Return true if this scalar return value only has bits discarded on its path
00317 /// from the "tail call" to the "ret". This includes the obvious noop
00318 /// instructions handled by getNoopInput above as well as free truncations (or
00319 /// extensions prior to the call).
00320 static bool slotOnlyDiscardsData(const Value *RetVal, const Value *CallVal,
00321                                  SmallVectorImpl<unsigned> &RetIndices,
00322                                  SmallVectorImpl<unsigned> &CallIndices,
00323                                  bool AllowDifferingSizes,
00324                                  const TargetLoweringBase &TLI) {
00325 
00326   // Trace the sub-value needed by the return value as far back up the graph as
00327   // possible, in the hope that it will intersect with the value produced by the
00328   // call. In the simple case with no "returned" attribute, the hope is actually
00329   // that we end up back at the tail call instruction itself.
00330   unsigned BitsRequired = UINT_MAX;
00331   RetVal = getNoopInput(RetVal, RetIndices, BitsRequired, TLI);
00332 
00333   // If this slot in the value returned is undef, it doesn't matter what the
00334   // call puts there, it'll be fine.
00335   if (isa<UndefValue>(RetVal))
00336     return true;
00337 
00338   // Now do a similar search up through the graph to find where the value
00339   // actually returned by the "tail call" comes from. In the simple case without
00340   // a "returned" attribute, the search will be blocked immediately and the loop
00341   // a Noop.
00342   unsigned BitsProvided = UINT_MAX;
00343   CallVal = getNoopInput(CallVal, CallIndices, BitsProvided, TLI);
00344 
00345   // There's no hope if we can't actually trace them to (the same part of!) the
00346   // same value.
00347   if (CallVal != RetVal || CallIndices != RetIndices)
00348     return false;
00349 
00350   // However, intervening truncates may have made the call non-tail. Make sure
00351   // all the bits that are needed by the "ret" have been provided by the "tail
00352   // call". FIXME: with sufficiently cunning bit-tracking, we could look through
00353   // extensions too.
00354   if (BitsProvided < BitsRequired ||
00355       (!AllowDifferingSizes && BitsProvided != BitsRequired))
00356     return false;
00357 
00358   return true;
00359 }
00360 
00361 /// For an aggregate type, determine whether a given index is within bounds or
00362 /// not.
00363 static bool indexReallyValid(CompositeType *T, unsigned Idx) {
00364   if (ArrayType *AT = dyn_cast<ArrayType>(T))
00365     return Idx < AT->getNumElements();
00366 
00367   return Idx < cast<StructType>(T)->getNumElements();
00368 }
00369 
00370 /// Move the given iterators to the next leaf type in depth first traversal.
00371 ///
00372 /// Performs a depth-first traversal of the type as specified by its arguments,
00373 /// stopping at the next leaf node (which may be a legitimate scalar type or an
00374 /// empty struct or array).
00375 ///
00376 /// @param SubTypes List of the partial components making up the type from
00377 /// outermost to innermost non-empty aggregate. The element currently
00378 /// represented is SubTypes.back()->getTypeAtIndex(Path.back() - 1).
00379 ///
00380 /// @param Path Set of extractvalue indices leading from the outermost type
00381 /// (SubTypes[0]) to the leaf node currently represented.
00382 ///
00383 /// @returns true if a new type was found, false otherwise. Calling this
00384 /// function again on a finished iterator will repeatedly return
00385 /// false. SubTypes.back()->getTypeAtIndex(Path.back()) is either an empty
00386 /// aggregate or a non-aggregate
00387 static bool advanceToNextLeafType(SmallVectorImpl<CompositeType *> &SubTypes,
00388                                   SmallVectorImpl<unsigned> &Path) {
00389   // First march back up the tree until we can successfully increment one of the
00390   // coordinates in Path.
00391   while (!Path.empty() && !indexReallyValid(SubTypes.back(), Path.back() + 1)) {
00392     Path.pop_back();
00393     SubTypes.pop_back();
00394   }
00395 
00396   // If we reached the top, then the iterator is done.
00397   if (Path.empty())
00398     return false;
00399 
00400   // We know there's *some* valid leaf now, so march back down the tree picking
00401   // out the left-most element at each node.
00402   ++Path.back();
00403   Type *DeeperType = SubTypes.back()->getTypeAtIndex(Path.back());
00404   while (DeeperType->isAggregateType()) {
00405     CompositeType *CT = cast<CompositeType>(DeeperType);
00406     if (!indexReallyValid(CT, 0))
00407       return true;
00408 
00409     SubTypes.push_back(CT);
00410     Path.push_back(0);
00411 
00412     DeeperType = CT->getTypeAtIndex(0U);
00413   }
00414 
00415   return true;
00416 }
00417 
00418 /// Find the first non-empty, scalar-like type in Next and setup the iterator
00419 /// components.
00420 ///
00421 /// Assuming Next is an aggregate of some kind, this function will traverse the
00422 /// tree from left to right (i.e. depth-first) looking for the first
00423 /// non-aggregate type which will play a role in function return.
00424 ///
00425 /// For example, if Next was {[0 x i64], {{}, i32, {}}, i32} then we would setup
00426 /// Path as [1, 1] and SubTypes as [Next, {{}, i32, {}}] to represent the first
00427 /// i32 in that type.
00428 static bool firstRealType(Type *Next,
00429                           SmallVectorImpl<CompositeType *> &SubTypes,
00430                           SmallVectorImpl<unsigned> &Path) {
00431   // First initialise the iterator components to the first "leaf" node
00432   // (i.e. node with no valid sub-type at any index, so {} does count as a leaf
00433   // despite nominally being an aggregate).
00434   while (Next->isAggregateType() &&
00435          indexReallyValid(cast<CompositeType>(Next), 0)) {
00436     SubTypes.push_back(cast<CompositeType>(Next));
00437     Path.push_back(0);
00438     Next = cast<CompositeType>(Next)->getTypeAtIndex(0U);
00439   }
00440 
00441   // If there's no Path now, Next was originally scalar already (or empty
00442   // leaf). We're done.
00443   if (Path.empty())
00444     return true;
00445 
00446   // Otherwise, use normal iteration to keep looking through the tree until we
00447   // find a non-aggregate type.
00448   while (SubTypes.back()->getTypeAtIndex(Path.back())->isAggregateType()) {
00449     if (!advanceToNextLeafType(SubTypes, Path))
00450       return false;
00451   }
00452 
00453   return true;
00454 }
00455 
00456 /// Set the iterator data-structures to the next non-empty, non-aggregate
00457 /// subtype.
00458 static bool nextRealType(SmallVectorImpl<CompositeType *> &SubTypes,
00459                          SmallVectorImpl<unsigned> &Path) {
00460   do {
00461     if (!advanceToNextLeafType(SubTypes, Path))
00462       return false;
00463 
00464     assert(!Path.empty() && "found a leaf but didn't set the path?");
00465   } while (SubTypes.back()->getTypeAtIndex(Path.back())->isAggregateType());
00466 
00467   return true;
00468 }
00469 
00470 
00471 /// Test if the given instruction is in a position to be optimized
00472 /// with a tail-call. This roughly means that it's in a block with
00473 /// a return and there's nothing that needs to be scheduled
00474 /// between it and the return.
00475 ///
00476 /// This function only tests target-independent requirements.
00477 bool llvm::isInTailCallPosition(ImmutableCallSite CS,
00478                                 const TargetLowering &TLI) {
00479   const Instruction *I = CS.getInstruction();
00480   const BasicBlock *ExitBB = I->getParent();
00481   const TerminatorInst *Term = ExitBB->getTerminator();
00482   const ReturnInst *Ret = dyn_cast<ReturnInst>(Term);
00483 
00484   // The block must end in a return statement or unreachable.
00485   //
00486   // FIXME: Decline tailcall if it's not guaranteed and if the block ends in
00487   // an unreachable, for now. The way tailcall optimization is currently
00488   // implemented means it will add an epilogue followed by a jump. That is
00489   // not profitable. Also, if the callee is a special function (e.g.
00490   // longjmp on x86), it can end up causing miscompilation that has not
00491   // been fully understood.
00492   if (!Ret &&
00493       (!TLI.getTargetMachine().Options.GuaranteedTailCallOpt ||
00494        !isa<UnreachableInst>(Term)))
00495     return false;
00496 
00497   // If I will have a chain, make sure no other instruction that will have a
00498   // chain interposes between I and the return.
00499   if (I->mayHaveSideEffects() || I->mayReadFromMemory() ||
00500       !isSafeToSpeculativelyExecute(I))
00501     for (BasicBlock::const_iterator BBI = std::prev(ExitBB->end(), 2);; --BBI) {
00502       if (&*BBI == I)
00503         break;
00504       // Debug info intrinsics do not get in the way of tail call optimization.
00505       if (isa<DbgInfoIntrinsic>(BBI))
00506         continue;
00507       if (BBI->mayHaveSideEffects() || BBI->mayReadFromMemory() ||
00508           !isSafeToSpeculativelyExecute(BBI))
00509         return false;
00510     }
00511 
00512   return returnTypeIsEligibleForTailCall(ExitBB->getParent(), I, Ret, TLI);
00513 }
00514 
00515 bool llvm::returnTypeIsEligibleForTailCall(const Function *F,
00516                                            const Instruction *I,
00517                                            const ReturnInst *Ret,
00518                                            const TargetLoweringBase &TLI) {
00519   // If the block ends with a void return or unreachable, it doesn't matter
00520   // what the call's return type is.
00521   if (!Ret || Ret->getNumOperands() == 0) return true;
00522 
00523   // If the return value is undef, it doesn't matter what the call's
00524   // return type is.
00525   if (isa<UndefValue>(Ret->getOperand(0))) return true;
00526 
00527   // Make sure the attributes attached to each return are compatible.
00528   AttrBuilder CallerAttrs(F->getAttributes(),
00529                           AttributeSet::ReturnIndex);
00530   AttrBuilder CalleeAttrs(cast<CallInst>(I)->getAttributes(),
00531                           AttributeSet::ReturnIndex);
00532 
00533   // Noalias is completely benign as far as calling convention goes, it
00534   // shouldn't affect whether the call is a tail call.
00535   CallerAttrs = CallerAttrs.removeAttribute(Attribute::NoAlias);
00536   CalleeAttrs = CalleeAttrs.removeAttribute(Attribute::NoAlias);
00537 
00538   bool AllowDifferingSizes = true;
00539   if (CallerAttrs.contains(Attribute::ZExt)) {
00540     if (!CalleeAttrs.contains(Attribute::ZExt))
00541       return false;
00542 
00543     AllowDifferingSizes = false;
00544     CallerAttrs.removeAttribute(Attribute::ZExt);
00545     CalleeAttrs.removeAttribute(Attribute::ZExt);
00546   } else if (CallerAttrs.contains(Attribute::SExt)) {
00547     if (!CalleeAttrs.contains(Attribute::SExt))
00548       return false;
00549 
00550     AllowDifferingSizes = false;
00551     CallerAttrs.removeAttribute(Attribute::SExt);
00552     CalleeAttrs.removeAttribute(Attribute::SExt);
00553   }
00554 
00555   // If they're still different, there's some facet we don't understand
00556   // (currently only "inreg", but in future who knows). It may be OK but the
00557   // only safe option is to reject the tail call.
00558   if (CallerAttrs != CalleeAttrs)
00559     return false;
00560 
00561   const Value *RetVal = Ret->getOperand(0), *CallVal = I;
00562   SmallVector<unsigned, 4> RetPath, CallPath;
00563   SmallVector<CompositeType *, 4> RetSubTypes, CallSubTypes;
00564 
00565   bool RetEmpty = !firstRealType(RetVal->getType(), RetSubTypes, RetPath);
00566   bool CallEmpty = !firstRealType(CallVal->getType(), CallSubTypes, CallPath);
00567 
00568   // Nothing's actually returned, it doesn't matter what the callee put there
00569   // it's a valid tail call.
00570   if (RetEmpty)
00571     return true;
00572 
00573   // Iterate pairwise through each of the value types making up the tail call
00574   // and the corresponding return. For each one we want to know whether it's
00575   // essentially going directly from the tail call to the ret, via operations
00576   // that end up not generating any code.
00577   //
00578   // We allow a certain amount of covariance here. For example it's permitted
00579   // for the tail call to define more bits than the ret actually cares about
00580   // (e.g. via a truncate).
00581   do {
00582     if (CallEmpty) {
00583       // We've exhausted the values produced by the tail call instruction, the
00584       // rest are essentially undef. The type doesn't really matter, but we need
00585       // *something*.
00586       Type *SlotType = RetSubTypes.back()->getTypeAtIndex(RetPath.back());
00587       CallVal = UndefValue::get(SlotType);
00588     }
00589 
00590     // The manipulations performed when we're looking through an insertvalue or
00591     // an extractvalue would happen at the front of the RetPath list, so since
00592     // we have to copy it anyway it's more efficient to create a reversed copy.
00593     using std::copy;
00594     SmallVector<unsigned, 4> TmpRetPath, TmpCallPath;
00595     copy(RetPath.rbegin(), RetPath.rend(), std::back_inserter(TmpRetPath));
00596     copy(CallPath.rbegin(), CallPath.rend(), std::back_inserter(TmpCallPath));
00597 
00598     // Finally, we can check whether the value produced by the tail call at this
00599     // index is compatible with the value we return.
00600     if (!slotOnlyDiscardsData(RetVal, CallVal, TmpRetPath, TmpCallPath,
00601                               AllowDifferingSizes, TLI))
00602       return false;
00603 
00604     CallEmpty  = !nextRealType(CallSubTypes, CallPath);
00605   } while(nextRealType(RetSubTypes, RetPath));
00606 
00607   return true;
00608 }