+ NoopInput = Op;
+ } else if (isa<PtrToIntInst>(I)) {
+ // Look through ptrtoint.
+ // Make sure this isn't a truncating or extending cast. We could
+ // support this eventually, but don't bother for now.
+ if (!isa<VectorType>(I->getType()) &&
+ TLI.getPointerTy().getSizeInBits() ==
+ cast<IntegerType>(I->getType())->getBitWidth())
+ NoopInput = Op;
+ } else if (isa<TruncInst>(I) &&
+ TLI.allowTruncateForTailCall(Op->getType(), I->getType())) {
+ DataBits = std::min(DataBits, I->getType()->getPrimitiveSizeInBits());
+ NoopInput = Op;
+ } else if (isa<CallInst>(I)) {
+ // Look through call (skipping callee)
+ for (User::const_op_iterator i = I->op_begin(), e = I->op_end() - 1;
+ i != e; ++i) {
+ unsigned attrInd = i - I->op_begin() + 1;
+ if (cast<CallInst>(I)->paramHasAttr(attrInd, Attribute::Returned) &&
+ isNoopBitcast((*i)->getType(), I->getType(), TLI)) {
+ NoopInput = *i;
+ break;
+ }
+ }
+ } else if (isa<InvokeInst>(I)) {
+ // Look through invoke (skipping BB, BB, Callee)
+ for (User::const_op_iterator i = I->op_begin(), e = I->op_end() - 3;
+ i != e; ++i) {
+ unsigned attrInd = i - I->op_begin() + 1;
+ if (cast<InvokeInst>(I)->paramHasAttr(attrInd, Attribute::Returned) &&
+ isNoopBitcast((*i)->getType(), I->getType(), TLI)) {
+ NoopInput = *i;
+ break;
+ }
+ }
+ } else if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(V)) {
+ // Value may come from either the aggregate or the scalar
+ ArrayRef<unsigned> InsertLoc = IVI->getIndices();
+ if (ValLoc.size() >= InsertLoc.size() &&
+ std::equal(InsertLoc.begin(), InsertLoc.end(), ValLoc.rbegin())) {
+ // The type being inserted is a nested sub-type of the aggregate; we
+ // have to remove those initial indices to get the location we're
+ // interested in for the operand.
+ ValLoc.resize(ValLoc.size() - InsertLoc.size());
+ NoopInput = IVI->getInsertedValueOperand();
+ } else {
+ // The struct we're inserting into has the value we're interested in, no
+ // change of address.
+ NoopInput = Op;
+ }
+ } else if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) {
+ // The part we're interested in will inevitably be some sub-section of the
+ // previous aggregate. Combine the two paths to obtain the true address of
+ // our element.
+ ArrayRef<unsigned> ExtractLoc = EVI->getIndices();
+ ValLoc.append(ExtractLoc.rbegin(), ExtractLoc.rend());
+ NoopInput = Op;
+ }
+ // Terminate if we couldn't find anything to look through.
+ if (!NoopInput)
+ return V;
+
+ V = NoopInput;
+ }
+}
+
+/// Return true if this scalar return value only has bits discarded on its path
+/// from the "tail call" to the "ret". This includes the obvious noop
+/// instructions handled by getNoopInput above as well as free truncations (or
+/// extensions prior to the call).
+static bool slotOnlyDiscardsData(const Value *RetVal, const Value *CallVal,
+ SmallVectorImpl<unsigned> &RetIndices,
+ SmallVectorImpl<unsigned> &CallIndices,
+ bool AllowDifferingSizes,
+ const TargetLoweringBase &TLI) {
+
+ // Trace the sub-value needed by the return value as far back up the graph as
+ // possible, in the hope that it will intersect with the value produced by the
+ // call. In the simple case with no "returned" attribute, the hope is actually
+ // that we end up back at the tail call instruction itself.
+ unsigned BitsRequired = UINT_MAX;
+ RetVal = getNoopInput(RetVal, RetIndices, BitsRequired, TLI);
+
+ // If this slot in the value returned is undef, it doesn't matter what the
+ // call puts there, it'll be fine.
+ if (isa<UndefValue>(RetVal))
+ return true;
+
+ // Now do a similar search up through the graph to find where the value
+ // actually returned by the "tail call" comes from. In the simple case without
+ // a "returned" attribute, the search will be blocked immediately and the loop
+ // a Noop.
+ unsigned BitsProvided = UINT_MAX;
+ CallVal = getNoopInput(CallVal, CallIndices, BitsProvided, TLI);
+
+ // There's no hope if we can't actually trace them to (the same part of!) the
+ // same value.
+ if (CallVal != RetVal || CallIndices != RetIndices)
+ return false;
+
+ // However, intervening truncates may have made the call non-tail. Make sure
+ // all the bits that are needed by the "ret" have been provided by the "tail
+ // call". FIXME: with sufficiently cunning bit-tracking, we could look through
+ // extensions too.
+ if (BitsProvided < BitsRequired ||
+ (!AllowDifferingSizes && BitsProvided != BitsRequired))
+ return false;
+
+ return true;
+}
+
+/// For an aggregate type, determine whether a given index is within bounds or
+/// not.
+static bool indexReallyValid(CompositeType *T, unsigned Idx) {
+ if (ArrayType *AT = dyn_cast<ArrayType>(T))
+ return Idx < AT->getNumElements();
+
+ return Idx < cast<StructType>(T)->getNumElements();
+}
+
+/// Move the given iterators to the next leaf type in depth first traversal.
+///
+/// Performs a depth-first traversal of the type as specified by its arguments,
+/// stopping at the next leaf node (which may be a legitimate scalar type or an
+/// empty struct or array).
+///
+/// @param SubTypes List of the partial components making up the type from
+/// outermost to innermost non-empty aggregate. The element currently
+/// represented is SubTypes.back()->getTypeAtIndex(Path.back() - 1).
+///
+/// @param Path Set of extractvalue indices leading from the outermost type
+/// (SubTypes[0]) to the leaf node currently represented.
+///
+/// @returns true if a new type was found, false otherwise. Calling this
+/// function again on a finished iterator will repeatedly return
+/// false. SubTypes.back()->getTypeAtIndex(Path.back()) is either an empty
+/// aggregate or a non-aggregate
+static bool advanceToNextLeafType(SmallVectorImpl<CompositeType *> &SubTypes,
+ SmallVectorImpl<unsigned> &Path) {
+ // First march back up the tree until we can successfully increment one of the
+ // coordinates in Path.
+ while (!Path.empty() && !indexReallyValid(SubTypes.back(), Path.back() + 1)) {
+ Path.pop_back();
+ SubTypes.pop_back();
+ }
+
+ // If we reached the top, then the iterator is done.
+ if (Path.empty())
+ return false;
+
+ // We know there's *some* valid leaf now, so march back down the tree picking
+ // out the left-most element at each node.
+ ++Path.back();
+ Type *DeeperType = SubTypes.back()->getTypeAtIndex(Path.back());
+ while (DeeperType->isAggregateType()) {
+ CompositeType *CT = cast<CompositeType>(DeeperType);
+ if (!indexReallyValid(CT, 0))
+ return true;
+
+ SubTypes.push_back(CT);
+ Path.push_back(0);
+
+ DeeperType = CT->getTypeAtIndex(0U);
+ }
+
+ return true;
+}
+
+/// Find the first non-empty, scalar-like type in Next and setup the iterator
+/// components.
+///
+/// Assuming Next is an aggregate of some kind, this function will traverse the
+/// tree from left to right (i.e. depth-first) looking for the first
+/// non-aggregate type which will play a role in function return.
+///
+/// For example, if Next was {[0 x i64], {{}, i32, {}}, i32} then we would setup
+/// Path as [1, 1] and SubTypes as [Next, {{}, i32, {}}] to represent the first
+/// i32 in that type.
+static bool firstRealType(Type *Next,
+ SmallVectorImpl<CompositeType *> &SubTypes,
+ SmallVectorImpl<unsigned> &Path) {
+ // First initialise the iterator components to the first "leaf" node
+ // (i.e. node with no valid sub-type at any index, so {} does count as a leaf
+ // despite nominally being an aggregate).
+ while (Next->isAggregateType() &&
+ indexReallyValid(cast<CompositeType>(Next), 0)) {
+ SubTypes.push_back(cast<CompositeType>(Next));
+ Path.push_back(0);
+ Next = cast<CompositeType>(Next)->getTypeAtIndex(0U);