Bartosz Milewski, “Re-discovering Monads in C++”

Re-discovering
Monads in C++
Bartosz Milewski, Бартош Милевски
Twitter: @BartoszMilewski
Novosibirsk, November 2014

Data in Boxes
• Smart pointers: unique_ptr, shared_ptr
• Optional, expected
• Containers, ranges
• Streams, channels (telemetry, stock quotes)
• Futures
• Database queries

In and Out of Boxes
• Common patterns in data processing
• Extract data from box
• Process it
• Put it back in a box
for (int i = 0; i < len; ++i)
w.push_back(f(v[i]));

Separate Concerns
• Separate extraction/repacking from processing
• Get rid of unessential variables
• Iterators and algorithms: better but not much
transform(begin(v), end(v), back_inserter(w), f);

Declarative Approach
• Describe “what” rather than “how”
• Example: ranges
int total = accumulate(view::iota(1) |
view::transform([](int x){return x*x;}) |
view::take(10), 0);

Change of Perspective
• Take a function that acts on items
[](int x){return x*x;}
• Lift it using view::transform
view::transform([](int x){return x*x;})
• Lifted function acts on ranges of items {1, 2, 3, 4, …}
view::iota(1) |
view::transform([](int x){return x*x;})
• Returns a range of items {1, 4, 9, 16, …}

range
1, 2, 3, 4
range
1, 4, 9, 16
transform(square)
transform
x x*x
square

Laziness
• Infinite range {1, 2, 3, 4, …}
std::iota(1)
• Can’t process it eagerly!
• Transform it lazily, on demand
• Working with data that can’t fit in memory
• infinite ranges
• big data
• database queries: LINQ

Functor Pattern
• Function on types (type template)
• Take any type T and produce Functor<T>
• Function on functions
• Take any function f from t1 to t2
• Return a function from Functor<t1> to Functor<t2>
• Preserve identity
• Preserve composition

Range as Functor
• Function on types: T to range<T>
• Function on functions: view::transform
view::iota(1) |
view::transform([](int x){return x*x;}) |
view::take(10)

future as Functor
• A box that may eventually contain a value of type T
std::future<T> fut = std::async(…);
• Processing this value:
auto x = fut.get(); // may block
auto y = f(x);
• Lifting a function using next:
auto y = fut.next(f).get(); // proposed
• fut.next(f).next(g).get();

Lifting Values
• A pointed functor is a functor
• Function on types
• Function on functions (lifting functions)
• Lifting a single value
• A template function from T to Functor<T>

Examples
• Containers: Creating a singleton container
vector<double> {3.14}
• Lazy range: view::single
• Future: make_ready_future (proposed)
• Usage:
• return value when Functor expected
• terminate recursion

Lifting Multi-Arg Functions
• Applicative functor is a pointed functor:
• Function on types
• Lifting functions and values
• Lifting multi-argument functions to functions taking
multiple Functor arguments (of different types)

Lazy Applicative Range (1)
• Example: Apply 2-arg function plus to two ranges
zip_with(plus, r1, r2);
• Variadic template zip_with
• Takes a function f of n-arguments
• Takes n ranges of corresponding types
• Applies f to corresponding elements (until one of the
ranges runs out)
• Value lifting through repeat (infinite range)

Lazy Applicative Range (2)
• Apply n-argument function to all combinations of
values from n-ranges
• Could be an overload of view::transform ?
• Implemented as nested loop
• Value lifting through single

Applicative Law
• Lifting a 1-argument function
• As function: transform(f, r)
• As value (1): repeat(f)
• Applying it to range:
zip_with(apply, repeat(f), r)
• As value (2): single(f)
• Applying it to range:
transform(apply, single(f), r)

Applicative future
• Apply multi-argument function to multiple futures
• Not planned, instead when_all
• To apply the function, all arguments must be ready
• Problem: futures may return exceptions, so it’s a
combination of applicatives

Functor Factories
• Library writer provides functions that return
Functors
• User wants to define their own
• How do you chain Functor factories?
Functor<t2> makeT2(t1);
Functor<t3> makeT3(t2);

Future example
• Composing asynchronous calls
future<HANDLE> async_open(string &);
future<Buffer> async_read(HANDLE fh);
• Option 1: call get() twice
• Option 2: call next. Problem: Double future.
future<future<Buffer>> ffBuf =
async_open(“foo").next(&async_read);
• Solution: Collapse double future using unwrap()
async_open("foo").next(&async_read).unwrap()
.next(&async_process);
• Better solution (proposed): Overload next for functions returning futures

Range Example
• Functions returning ranges: Range factories
• Applying range factory to range factory using
transform results in a range of ranges
• Collapsing using flatten
• Combination of transform and flatten called
for_each

Pythagorean Triples
auto triples =
for_each(ints(1), [](int z) {
return for_each(ints(1, z), [=](int x) {
return for_each(ints(x, z), [=](int y) {
return yield_if(x*x + y*y == z*z,
std::make_tuple(x, y, z));
});
});
});
• yield is the same as single inside for_each
• yield_if is conditional yield (monad plus)

Monad
• Monad is an applicative functor
• function on types
• function on functions: function lifting
• value lifting
• multi-argument function lifting
• Flatten (reduce double Functor to single Functor)
• Or Bind (combination of lifting and flatten)
• And some monad laws

Current Problems
• Lack of an overall abstraction (a monad template)
• Random naming
• fmap, transform, next, then, Select (LINQ)
• pure, single, yield, await, make_ready_future
• bind, for_each, next, then, SelectMany
• Need syntactic sugar, like Haskell do notation
• Resumable functions (proposed)

Conclusion
• The same pattern fits many problems
• ranges, lazy ranges
• optional, expected
• futures
• LINQ (IEnumerable)
• state, continuation, input, output
• many more…

Bartosz Milewski, “Re-discovering Monads in C++”

More Related Content

What's hot

Similar to Bartosz Milewski, “Re-discovering Monads in C++”

More from Platonov Sergey

Recently uploaded

Bartosz Milewski, “Re-discovering Monads in C++”