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