[ADT] Add zip_longest iterators.

[Meinersbur]

Summary

Like the already existing zip_shortest/zip_first iterators, zip_longest iterates over multiple iterators at once, but has as many iterations as the longest sequence.

This means some iterators may reach the end before others do. There are two variants: zip_longest_optional uses llvm::Optional's None value to mark a past-the-end value. zip_longest_default uses the type's default-initialized value for this purpose (which means one cannot distinguish the default value from being in the list and having reached the end). I included both variants in this patch, but we may use only one of the variants.

Neither variant is reverse-iteratable because the tuples iterated over would be different for different length sequences (IMHO for the same reason neither zip_shortest nor zip_first should be reverse-iteratable; one can still reverse the ranges individually if that's the expected behavior).

In contrast to zip_shortest/zip_first, zip_longest tuples contain rvalues instead of references. This is because llvm::Optional cannot contain reference types and the value-initialized default does not have a memory location a reference could point to.

The motivation for these iterators is to use C++ foreach to compare two lists of ordered attributes in D48100 (SemaOverload.cpp and ASTReaderDecl.cpp).

Idea by @hfinkel.

[dblaikie]

Idle thought: Could/should this be composed from other primitives that could be created. Such as an infinitely long (or maybe bounded, not sure) range built from another range and a functor of some kind? (or such a thing specifically tailored for Optional-or-default-T situations) & then zip two of those together & bound that resulting infinitely long range (or bound the input ranges, either way).

Maybe that's overgeneralization, but just an idea to see how others feel about it.

[Meinersbur]

Is the following what you are thinking about?

An zip_infinite begin-iterator, that never reaches the end (maybe with two flavors: zip_infinite_default and zip_infinite_optional) and return a None/value-initialized value for any iterator past the list. Then one can use that with an end iterator, such as end_any (-> zip_shortest) or end_all (-> zip_longest).

Unfortunately, in the zip_infinite_default-case, we cannot distinguish between the default value end the end of the list. We'd have to look into the iterator's list of iterators, which makes this implementation-specific again.

Alternatively, zip_infinite could return tuples of iterators, which end_any/end_all could compare to the end iterators. However, the current implementation of zip_shortest does not need to store the end-iterators for operator++ (because in contrast to the others, it does not need to check whether one of the sequences is past the the end), making the implementation less efficient than it needs to. (template specialization of shortest_iterator<zip_infinite<T>> maybe?)

include/llvm/ADT/STLExtras.h

860–888

The zip_*_range classes (including zippy) might no be needed: iterator_range could be used as well.

[dblaikie]

Nah - I was thinking of something where zip wouldn't need any changes,
potentially. One possibility:

• iterator adapter that takes a (potentially finite) range of T and

produces an infinite range of Optional<T> (with elements beyond the
original range being empty Optionals)

• same as the above, except potentially finite range of T to an

infinite range of T where the elements beyond the original range are
default constructed

• A range truncation helper that just does something like:

iterator_range trunc(R r, size_t t) { return {adl_begin(r),
std::next(adl_end(r), t)}; }

• then you do something like: trunc(zip(inf_opt(R1), inf_opt(R2)),

max(size(R1), size(R2)))

This allows the use of these infinite/padded ranges in other situations
apart from zipping.

But maybe that's niche enough/not worth generalizing over pre-emptively.
I'm not sure. Just a thought.

[Meinersbur]

Getting the size may require iterating the sequence in advance, which I tried to avoid in D48100. The current zip_shortest also does not require a value-initializable type and returns references to elements in the container, both of which would not be possible.

[dblaikie]

In D48348#1143128, @Meinersbur wrote:

Getting the size may require iterating the sequence in advance, which I tried to avoid in D48100.

Yeah, it's a tradeoff there, to be sure. Not sure if anyone else has thoughts on the design tradeoffs here - would hope other folks might chime in.

The current zip_shortest also does not require a value-initializable type and returns references to elements in the container, both of which would not be possible.

I believe returning references to elements is necessary to implement the iterator concept, unfortunately. This is usually dealt with in cases where there's no underlying sequence, by having the iterator contain a member of the desired type and returning references to that (with the restriction that references and pointers to that element are invalidated when the iterator is incremented - not sure if that requires downgrading the iterator category because of this, I don't think so though (I don't think the categories enshrine the invalidation semantics)).

I believe this requirement comes from the need for "i->x" to be valid when "(*i).x" is valid, and operator-> returns a pointer, so it has to have some storage to back it.

[Meinersbur]

ping. Any other opinions on this?

[hfinkel]

In D48348#1144008, @dblaikie wrote:

In D48348#1143128, @Meinersbur wrote:

Getting the size may require iterating the sequence in advance, which I tried to avoid in D48100.

Yeah, it's a tradeoff there, to be sure. Not sure if anyone else has thoughts on the design tradeoffs here - would hope other folks might chime in.

I'd prefer to not have to get the size in advance. The iterator as proposed seems like a cleaner abstraction. @chandlerc, any thoughts on this?

The current zip_shortest also does not require a value-initializable type and returns references to elements in the container, both of which would not be possible.

I believe returning references to elements is necessary to implement the iterator concept, unfortunately. This is usually dealt with in cases where there's no underlying sequence, by having the iterator contain a member of the desired type and returning references to that (with the restriction that references and pointers to that element are invalidated when the iterator is incremented - not sure if that requires downgrading the iterator category because of this, I don't think so though (I don't think the categories enshrine the invalidation semantics)).

I believe this requirement comes from the need for "i->x" to be valid when "(*i).x" is valid, and operator-> returns a pointer, so it has to have some storage to back it.

Hrmm. Can we have an Optional of a reference?

[Meinersbur]

In D48348#1144008, @dblaikie wrote:

In D48348#1143128, @Meinersbur wrote:

Getting the size may require iterating the sequence in advance, which I tried to avoid in D48100.

Yeah, it's a tradeoff there, to be sure. Not sure if anyone else has thoughts on the design tradeoffs here - would hope other folks might chime in.

IMHO this violates the don't-pay-for-what-you-don't-use principle.

In D48348#1185185, @hfinkel wrote:

Hrmm. Can we have an Optional of a reference?

Currently, trying to use Optional<int&> results in a compile error (two of Optional's constructors fall together). Note sure how the implementation of Optional can be (whether it makes sense to) changed.

[Meinersbur]

ping

[Meinersbur]

ping

[dblaikie]

I /think/ this still doesn't meet the iterator requirements (that if "(*i).j" is valid, then "i->j" must be valid too - [input.iterators]p2, Table 95, "a->m is equivalent to (*a).m if a is dereferencable"). I believe the way standard iterators (like istream_iterator) implement this is by having an instance of the value type inside the iterator that has the current value in it, so operator-> can return a pointer to that. Would that be possible/practical?

[Meinersbur]

In D48348#1310674, @dblaikie wrote:

I /think/ this still doesn't meet the iterator requirements (that if "(*i).j" is valid, then "i->j" must be valid too - [input.iterators]p2, Table 95, "a->m is equivalent to (*a).m if a is dereferencable"). I believe the way standard iterators (like istream_iterator) implement this is by having an instance of the value type inside the iterator that has the current value in it, so operator-> can return a pointer to that. Would that be possible/practical?

The type of *a is a std::tuple of the client-iterator's value types. Storing an internal scratch field inside the the zip-operator would allow operator-> to return a pointer. However, std::tuple has just two methods that could be called: swap and operator=. Both would be useless and operator-> should probably return a const std::tuple* to avoid messing with the internal field anyway. It still would not allow assigning to the tuple's elements.

Experimenting with llvm::zip and llvm::map_iterator shows that they don't meet the requirement either:

const SmallVector<unsigned, 6> pi{3, 1, 4, 1, 5, 9};
auto i = map_iterator(pi.begin(), [](unsigned v) -> std::pair<unsigned,char> {return (v == 3) ? std::make_pair(v,'a') : std::make_pair(v,'\0');} );
(*i).first;
i->first;

error: taking the address of a temporary object of type 'std::pair<unsigned int, char>' [-Waddress-of-temporary]
  PointerT operator->() { return &static_cast<DerivedT *>(this)->operator*(); }
                                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

      SmallVector<unsigned, 6> pi{3, 1, 4, 1, 5, 9};
      auto i = zip(pi,pi).begin();
      std::tuple<const unsigned&,const unsigned&> x{pi[0],pi[1]};
      (*i).operator=(x);
      i->operator=(x);

llvm/ADT/iterator.h:170:34: error: taking the address of a temporary object of type 'llvm::detail::zip_common<llvm::detail::zip_shortest<unsigned int *, unsigned int *>, unsigned int *, unsigned int *>::value_type' (aka 'std::tuple<unsigned int &, unsigned int &>') [-Waddress-of-temporary]
  PointerT operator->() { return &static_cast<DerivedT *>(this)->operator*(); }
                                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

To answer the question. Yes, it it possible to add a scratch std::tuple to zip_longest iterators, but there is no practical gain.

[dblaikie]

In D48348#1315100, @Meinersbur wrote:

In D48348#1310674, @dblaikie wrote:

I /think/ this still doesn't meet the iterator requirements (that if "(*i).j" is valid, then "i->j" must be valid too - [input.iterators]p2, Table 95, "a->m is equivalent to (*a).m if a is dereferencable"). I believe the way standard iterators (like istream_iterator) implement this is by having an instance of the value type inside the iterator that has the current value in it, so operator-> can return a pointer to that. Would that be possible/practical?

The type of *a is a std::tuple of the client-iterator's value types. Storing an internal scratch field inside the the zip-operator would allow operator-> to return a pointer. However, std::tuple has just two methods that could be called: swap and operator=. Both would be useless and operator-> should probably return a const std::tuple* to avoid messing with the internal field anyway. It still would not allow assigning to the tuple's elements.

Experimenting with llvm::zip and llvm::map_iterator shows that they don't meet the requirement either:

const SmallVector<unsigned, 6> pi{3, 1, 4, 1, 5, 9};
auto i = map_iterator(pi.begin(), [](unsigned v) -> std::pair<unsigned,char> {return (v == 3) ? std::make_pair(v,'a') : std::make_pair(v,'\0');} );
(*i).first;
i->first;

error: taking the address of a temporary object of type 'std::pair<unsigned int, char>' [-Waddress-of-temporary]
  PointerT operator->() { return &static_cast<DerivedT *>(this)->operator*(); }
                                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

      SmallVector<unsigned, 6> pi{3, 1, 4, 1, 5, 9};
      auto i = zip(pi,pi).begin();
      std::tuple<const unsigned&,const unsigned&> x{pi[0],pi[1]};
      (*i).operator=(x);
      i->operator=(x);

llvm/ADT/iterator.h:170:34: error: taking the address of a temporary object of type 'llvm::detail::zip_common<llvm::detail::zip_shortest<unsigned int *, unsigned int *>, unsigned int *, unsigned int *>::value_type' (aka 'std::tuple<unsigned int &, unsigned int &>') [-Waddress-of-temporary]
  PointerT operator->() { return &static_cast<DerivedT *>(this)->operator*(); }
                                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

To answer the question. Yes, it it possible to add a scratch std::tuple to zip_longest iterators, but there is no practical gain.

Fair enough - thanks for the explanation (it should be const and there are no const member functions to use, so there's no "i->u" or "(*i).u" that's valid to support).

[Meinersbur]

• Rebase to trunk
• Fix msvc build

[Meinersbur]

Thanks for accepting. I expected a discussion on whether zip_longest_default or zip_longest_optional is better (or make llvm::Optional support references and use zip_longest_optional with llvm::Optional<T&>). Should I commit both, as-is?

[dblaikie]

In D48348#1315209, @Meinersbur wrote:

Thanks for accepting. I expected a discussion on whether zip_longest_default or zip_longest_optional is better (or make llvm::Optional support references and use zip_longest_optional with llvm::Optional<T&>). Should I commit both, as-is?

Ah, sure - I'd probably prefer Optional over default (because the Optional makes it strictly more descriptive/non-lossy) & just call it zip_longest. We could have another iterator adapter that collapses optionals to default constructed instances in tuples, for instance... though I guess that'd be a bit single-use anyway.

unittests/ADT/IteratorTest.cpp

416–431

This testing is a bit convoluted - the bit testing, etc. Could it be done in a more direct manner, perhaps? (either stamping out/loop unrolling each value & testing the values directly against known constants (rather than against each other) or with a separate container containing the expected resulting pairs - and comparing those in a loop?)

(similarly below)

[Meinersbur]

• Rebase
• Remove zip_longest_default, rename zip_longest_optional to zip_longest
• Use adl_begin/adl_end instead of std::begin/std::end
• Deconvolute zipped_and_filtered test

[Meinersbur]

unittests/ADT/IteratorTest.cpp

416–431

This was the same patter as ZipIteratorTest above. Changed to use an explicit expected array.

[dblaikie]

unittests/ADT/IteratorTest.cpp

331

I'd probably do this test the way you wrote the other test - two vectors of different lengths of interesting things, then a vector of pairs with the hardcoded pairs of optionals in it - then loop and compare each one? Rather than the bit fiddling, counting, etc?

415

I probably wouldn't include this test unless there's some really specific relationship between zip and filter - it's not practical to test all the combinations of iterators with each other, etc - so only the specific interface of each one should generally be tested, I think?

[Meinersbur]

• Remove ZipLongestFilter tests
• Deconvolute ZipLongestBasic test

[dblaikie]

unittests/ADT/IteratorTest.cpp

333–334

Wonder if it's worth testing with two different (& incompatible) types - like const char* in the second vector, for instance?

Up to you.

[Meinersbur]

• Use StringRef for one side of zip_longest test case
• Use size_t to avoid signed/unsigned comparison compiler warning

[Meinersbur]

unittests/ADT/IteratorTest.cpp

333–334

Using StringRef which has an overloaded comparison operator. I get an ambiguity error if just using a literal string in the initializer_list of Optional<StringRef>, so there is an explicit conversion.

[dblaikie]

unittests/ADT/IteratorTest.cpp

333–334

Ah, cool - you might be able to get away with {...} instead of StringRef(...) if you like, but yeah, either's fine - thanks!

[Meinersbur]

• Use brace initialization for Optional<StringRef>

[Meinersbur]

unittests/ADT/IteratorTest.cpp

333–334

That worked. Thanks, looks nice like this.

[Meinersbur]

• Rename ZipLongestOptionalValueType to ZipLongestValueType
• Remove unused ZipLongestDefaultValueType

include/llvm/ADT/STLExtras.h

 iterator_range<early_inc_iterator_impl<detail::IterOfRange<RangeT>>>
 make_early_inc_range(RangeT &&Range) {
   using EarlyIncIteratorT =
       early_inc_iterator_impl<detail::IterOfRange<RangeT>>;
   return make_range(EarlyIncIteratorT(std::begin(std::forward<RangeT>(Range))),
                     EarlyIncIteratorT(std::end(std::forward<RangeT>(Range))));
 }
 
-// forward declarations required by zip_shortest/zip_first
+// forward declarations required by zip_shortest/zip_first/zip_longest
 template <typename R, typename UnaryPredicate>
 bool all_of(R &&range, UnaryPredicate P);
+template <typename R, typename UnaryPredicate>
+bool any_of(R &&range, UnaryPredicate P);
 
 template <size_t... I> struct index_sequence;
 
 template <class... Ts> struct index_sequence_for;
 
 namespace detail {
 
 using std::declval;

 /// be the shortest.
 template <typename T, typename U, typename... Args>
 detail::zippy<detail::zip_first, T, U, Args...> zip_first(T &&t, U &&u,
                                                           Args &&... args) {
   return detail::zippy<detail::zip_first, T, U, Args...>(
       std::forward<T>(t), std::forward<U>(u), std::forward<Args>(args)...);
 }
 
+namespace detail {
+template <typename Iter>
+static Iter next_or_end(const Iter &I, const Iter &End) {
+  if (I == End)
+    return End;
+  return std::next(I);
+}
+
+template <typename Iter>
+static auto deref_or_none(const Iter &I, const Iter &End)
+    -> llvm::Optional<typename std::remove_const<
+        typename std::remove_reference<decltype(*I)>::type>::type> {
+  if (I == End)
+    return None;
+  return *I;
+}
+
+template <typename... Iters> struct ZipLongestValueType {
+  using type = std::tuple<
+      llvm::Optional<typename std::remove_const<typename std::remove_reference<
+          decltype(*std::declval<Iters>())>::type>::type>...>;
+};
+
+template <typename... Iters>
+class zip_longest_iterator
+    : public iterator_facade_base<
+          zip_longest_iterator<Iters...>,
+          typename std::common_type<
+              std::forward_iterator_tag,
+              typename std::iterator_traits<Iters>::iterator_category...>::type,
+          typename ZipLongestValueType<Iters...>::type,
+          typename std::iterator_traits<typename std::tuple_element<
+              0, std::tuple<Iters...>>::type>::difference_type,
+          typename ZipLongestValueType<Iters...>::type *,
+          typename ZipLongestValueType<Iters...>::type> {
+public:
+  using value_type = typename ZipLongestValueType<Iters...>::type;
+
+private:
+  std::tuple<Iters...> iterators;
+  std::tuple<Iters...> end_iterators;
+
+  template <size_t... Ns>
+  bool test(const zip_longest_iterator<Iters...> &other,
+            index_sequence<Ns...>) const {
+    return llvm::any_of(
+        std::initializer_list<bool>{std::get<Ns>(this->iterators) !=
+                                    std::get<Ns>(other.iterators)...},
+        identity<bool>{});
+  }
+
+  template <size_t... Ns> value_type deref(index_sequence<Ns...>) const {
+    return value_type(
+        deref_or_none(std::get<Ns>(iterators), std::get<Ns>(end_iterators))...);
+  }
+
+  template <size_t... Ns>
+  decltype(iterators) tup_inc(index_sequence<Ns...>) const {
+    return std::tuple<Iters...>(
+        next_or_end(std::get<Ns>(iterators), std::get<Ns>(end_iterators))...);
+  }
+
+public:
+  zip_longest_iterator(std::pair<Iters &&, Iters &&>... ts)
+      : iterators(std::forward<Iters>(ts.first)...),
+        end_iterators(std::forward<Iters>(ts.second)...) {}
+
+  value_type operator*() { return deref(index_sequence_for<Iters...>{}); }
+
+  value_type operator*() const { return deref(index_sequence_for<Iters...>{}); }
+
+  zip_longest_iterator<Iters...> &operator++() {
+    iterators = tup_inc(index_sequence_for<Iters...>{});
+    return *this;
+  }
+
+  bool operator==(const zip_longest_iterator<Iters...> &other) const {
+    return !test(other, index_sequence_for<Iters...>{});
+  }
+};
+
+template <typename... Args> class zip_longest_range {
+public:
+  using iterator =
+      zip_longest_iterator<decltype(adl_begin(std::declval<Args>()))...>;
+  using iterator_category = typename iterator::iterator_category;
+  using value_type = typename iterator::value_type;
+  using difference_type = typename iterator::difference_type;
+  using pointer = typename iterator::pointer;
+  using reference = typename iterator::reference;
+
+private:
+  std::tuple<Args...> ts;
+
+  template <size_t... Ns> iterator begin_impl(index_sequence<Ns...>) const {
+    return iterator(std::make_pair(adl_begin(std::get<Ns>(ts)),
+                                   adl_end(std::get<Ns>(ts)))...);
+  }
+
+  template <size_t... Ns> iterator end_impl(index_sequence<Ns...>) const {
+    return iterator(std::make_pair(adl_end(std::get<Ns>(ts)),
+                                   adl_end(std::get<Ns>(ts)))...);
+  }
+
+public:
+  zip_longest_range(Args &&... ts_) : ts(std::forward<Args>(ts_)...) {}
+
+  iterator begin() const { return begin_impl(index_sequence_for<Args...>{}); }
+  iterator end() const { return end_impl(index_sequence_for<Args...>{}); }
+};
+} // namespace detail
+
+/// Iterate over two or more iterators at the same time. Iteration continues
+/// until all iterators reach the end. The llvm::Optional only contains a value
+/// if the iterator has not reached the end.
+template <typename T, typename U, typename... Args>
+detail::zip_longest_range<T, U, Args...> zip_longest(T &&t, U &&u,
+                                                     Args &&... args) {
+  return detail::zip_longest_range<T, U, Args...>(
+      std::forward<T>(t), std::forward<U>(u), std::forward<Args>(args)...);
+}
+
+namespace detail {
+template <typename Iter>
+static auto deref_or_default(const Iter &I, const Iter &End) ->
+    typename std::remove_const<
+        typename std::remove_reference<decltype(*I)>::type>::type {
+  if (I == End)
+    return {}; // value-initialize
+  return *I;
+}
+} // namespace detail
+
 /// Iterator wrapper that concatenates sequences together.
 ///
 /// This can concatenate different iterators, even with different types, into
 /// a single iterator provided the value types of all the concatenated
 /// iterators expose `reference` and `pointer` types that can be converted to
 /// `ValueT &` and `ValueT *` respectively. It doesn't support more
 /// interesting/customized pointer or reference types.
 ///
class concat_iterator
   }
 
   /// Returns null if the specified iterator is at the end. Otherwise,
   /// dereferences the iterator and returns the address of the resulting
   /// reference.
   template <size_t Index> ValueT *getHelper() const {
     auto &Begin = std::get<Index>(Begins);
     auto &End = std::get<Index>(Ends);
     if (Begin == End)
       return nullptr;
 
     return &*Begin;
   }
 
   /// Finds the first non-end iterator, dereferences, and returns the resulting
   /// reference.
   ///
   /// It is an error to call this with all iterators at the end.
   template <size_t... Ns> ValueT &get(index_sequence<Ns...>) const {
     // Build a sequence of functions to get from iterator if possible.
     ValueT *(concat_iterator::*GetHelperFns[])() const = {
         &concat_iterator::getHelper<Ns>...};
 
     // Loop over them, and return the first result we find.
     for (auto &GetHelperFn : GetHelperFns)
       if (ValueT *P = (this->*GetHelperFn)())
         return *P;
 
     llvm_unreachable("Attempted to get a pointer from an end concat iterator!");
   }
 
 public:
   /// Constructs an iterator from a squence of ranges.
   ///
   /// We need the full range to know how to switch between each of the
   /// iterators.
   template <typename... RangeTs>
   explicit concat_iterator(RangeTs &&... Ranges)
       : Begins(std::begin(Ranges)...), Ends(std::end(Ranges)...) {}
 
   using BaseT::operator++;
 
   concat_iterator &operator++() {
     increment(index_sequence_for<IterTs...>());
     return *this;

unittests/ADT/IteratorTest.cpp

TEST(ZipIteratorTest, ZipFirstBasic) {
   for (auto tup : zip_first(SmallVector<bool, 0>{1, 1, 0, 1}, pi)) {
     EXPECT_EQ(get<0>(tup), get<1>(tup) & 0x01);
     iters += 1;
   }
 
   EXPECT_EQ(iters, 4u);
 }
 
+TEST(ZipIteratorTest, ZipLongestBasic) {
+  using namespace std;
+  const vector<unsigned> pi{3, 1, 4, 1, 5, 9};
+  const vector<StringRef> e{"2", "7", "1", "8"};
+
+  {
+    // Check left range longer than right.
+    const vector<tuple<Optional<unsigned>, Optional<StringRef>>> expected{
+        {3, {"2"}}, {1, {"7"}}, {4, {"1"}}, {1, {"8"}}, {5, None}, {9, None}};
+    size_t iters = 0;
+    for (auto tup : zip_longest(pi, e)) {
+      EXPECT_EQ(tup, expected[iters]);
+      iters += 1;
+    }
+    EXPECT_EQ(iters, expected.size());
+  }
+
+  {
+    // Check right range longer than left.
+    const vector<tuple<Optional<StringRef>, Optional<unsigned>>> expected{
+        {{"2"}, 3}, {{"7"}, 1}, {{"1"}, 4}, {{"8"}, 1}, {None, 5}, {None, 9}};
+    size_t iters = 0;
+    for (auto tup : zip_longest(e, pi)) {
+      EXPECT_EQ(tup, expected[iters]);
+      iters += 1;
+    }
+    EXPECT_EQ(iters, expected.size());
+  }
+}
+
 TEST(ZipIteratorTest, Mutability) {
   using namespace std;
   const SmallVector<unsigned, 4> pi{3, 1, 4, 1, 5, 9};
   char message[] = "hello zip\0";
 
   for (auto tup : zip(pi, message, message)) {
     EXPECT_EQ(get<1>(tup), get<2>(tup));
     get<2>(tup) = get<0>(tup) & 0x01 ? 'y' : 'n';
TEST(ZipIteratorTest, Filter) {
 
   // Should have skipped pi[2].
   EXPECT_EQ(iters, 5u);
 
   // Ensure that in-place mutation works.
   EXPECT_TRUE(all_of(pi, [](unsigned n) { return (n & 0x01) == 0; }));
 }
 
 TEST(ZipIteratorTest, Reverse) {
   using namespace std;
   vector<unsigned> ascending{0, 1, 2, 3, 4, 5};
 
   auto zipped = zip_first(ascending, vector<bool>{0, 1, 0, 1, 0, 1});
   unsigned last = 6;
   for (auto tup : reverse(zipped)) {
     // Check that this is in reverse.
     EXPECT_LT(get<0>(tup), last);
     last = get<0>(tup);
     EXPECT_EQ(get<0>(tup) & 0x01, get<1>(tup));
   }
 
   auto odds = [](decltype(zipped)::value_type tup) { return get<1>(tup); };
   last = 6;
   for (auto tup : make_filter_range(reverse(zipped), odds)) {
     EXPECT_LT(get<0>(tup), last);
     last = get<0>(tup);
     EXPECT_TRUE(get<0>(tup) & 0x01);
     get<0>(tup) += 1;
   }
 
   // Ensure that in-place mutation works.
   EXPECT_TRUE(all_of(ascending, [](unsigned n) { return (n & 0x01) == 0; }));
 }