Draft RFC: variadic generics

[rust-highfive]

Issue by eddyb
Monday Oct 28, 2013 at 17:39 GMT

For earlier discussion, see rust-lang/rust#10124

This issue was labelled with: B-RFC in the Rust repository

---

The Problem

• fn types need a reform, and being able to define a trait with a variable number of type parameters would help
• working with functions which have a variable number of arguments is impossible right now (e.g. a generic bind method, f.bind(a, b)(c) == f(a, b, c)) and defining such functions may only be done (in a limited fashion) with macros
• tuples also have similar restrictions right now, there is no way to define a function which takes any tuple and returns the first element, for example

The Solution: Part One

C++11 sets a decent precedent, with its variadic templates, which can be used to define type-safe variadic functions, among other things.
I propose a similar syntax, a trailing ..T in generic formal type parameters:

type Tuple<..T> = (..T); // the parens here are the same as the ones around a tuple.
// use => expansion
Tuple<> => ()
Tuple<int> => (int,)
Tuple<A, B, C> => (A, B, C)

The simple example above only uses ..T, but not T itself.
The question which arises is this: what is T? C++11 has a special case for variadic parameter packs, but we can do better.
We have tuples. We can use them to store the actual variadic generic type parameters:

// Completely equivalent definition:
type Tuple<..T> = T;
// Detailed expansion of (..T) from above:
Tuple<> => (..()) => ()
Tuple<int> => (..(int,)) => (int,)
Tuple<A, B, C> => (..(A, B, C)) => (A, B, C)

The Solution: Part Two

Now that we know the answer is "tuples", everything else is about extending them.
The prefix .. operator would expand a tuple type:

// in another tuple type:
type AppendTuple<T, U> = (..T, U);
type PrependTuple<T, U> = (U, ..T);
AppendTuple<PrependTuple<Tuple<B>, A>, C> => (A, B, C)
// in a fn type's parameter list:
type VoidFn<..Args> = fn(..Args);
VoidFn<A, B, C> => fn(A, B, C);
// in a variadic generic parameter list:
Tuple<..(A, B, C), ..(D, E, F)> => (A, B, C, D, E, F)

Then we can do the same thing with values:

// in another tuple:
(..(true, false), ..(1, "foo"), "bar") => (true, false, 1, "foo", "bar")
// in a function call:
f(..(a, b), ..(c, d, e), f) => f(a, b, c, d, e, f)

There's only one piece missing: we're still not able to define a function which takes a variable number of arguments.
For this, I propose ..x: T (where T is a tuple type) in a pattern, which can be used to "capture" multiple arguments when used in fn formal arguments:

// Destructure a tuple.
let (head, ..tail) = foo;
// This function has the type fn(int, A, B, C), and can be called as such:
fn bar(_: int, ..x: (A, B, C)) -> R {...}

A type bound for ..T (i.e. impl<..T: Trait>) could mean that the tuple T has to satisfy the bound (or each type in T, but that's generally less useful).

Examples:

// The highly anticipated Fn trait.
trait Fn<Ret, ..Args> {
    fn call(self, .._: Args) -> Ret;
}
impl Fn<Ret, ..Args> for fn(..Args) -> Ret {
    fn call(self, ..args: Args) -> Ret {
        self(..args)
    }
}
impl Fn<Ret, ..Args> for |..Args| -> Ret {
    fn call(self, ..args: Args) -> Ret {
        self(..args)
    }
}

// Cloning tuples (all cons-list-like algorithms can be implemented this way).
impl<Head, ..Tail: Clone> Clone for (Head, ..Tail) {
    fn clone(&self @ &(ref head, ..ref tail)) -> (Head, ..Tail) {
        (head.clone(), ..tail.clone())
    }
}
impl Clone for () {
    fn clone(&self) -> () {
        ()
    }
}

// Restricting all types in a variadic tuple to one type.
trait ArrayLikeTuple<T> {
    fn len() -> uint;
}
impl<T> ArrayLikeTuple<T> for () {
    fn len() -> uint {0}
}
impl<T, ..Tail: ArrayLikeTuple<T>> ArrayLikeTuple<T> for (T, ..Tail) {
    fn len() -> uint {
        1 + Tail::len()
    }
}

// And using that trait to write variadic container constructors.
impl<T> Vec<T> {
    fn new<..Args: ArrayLikeTuple<T>>(..args: Args) -> Vec<T> {
        let v: Vec<T> = Vec::with_capacity(Args::len());
        // Use an unsafe move_iter-like pattern to move all the args
        // directly into the newly allocated vector. 
        // Alternatively, implement &move [T] and move_iter on that.
    }
}

// Zipping tuples, using a `Tuple` kind and associated items.
trait TupleZip<Other>: Tuple {
    type Result: Tuple;
    fn zip(self, other: Other) -> Result;
}
impl TupleZip<()> for () {
    type Result = ();
    fn zip(self, _: ()) -> () {}
}
impl<A, ATail: TupleZip<BTail>, B, BTail: Tuple> TupleZip<(B, ..BTail)> for (A, ..ATail) {
    type Result = ((A, B), ..ATail::Result);
    fn zip(self @ (a, ..a_tail), (b, ..b_tail): (B, ..BTail)) -> Result {
        ((a, b), ..a_tail.zip(b_tail))
    }
}

Todo

• better formal descriptions
• figure out the details of pattern matching
• more examples

[eddyb]

There is one issue with my initial approach: a reference to a "subtuple" of a larger tuple could have different padding than the tuple type made out of those fields.

One solution I may have is a &(ref head, ..ref tail) pattern turning &(A, B, C, D) into &A and (&B, &C, &D), i.e. a subtuple of references instead of a reference of a subtuple.
But I have no idea how inference or generics would work with it.

[alexchandel]

This is probably a prerequisite for taking the Cartesian product of iterators without a macro, like in rust-itertools/iproduct! and bluss/rust-itertools#10.

[Diggsey]

This should be expanded to also support using the ".." operator on fixed size arrays (behaves equivalent to a tuple where the elements are all the same type and the arity matches the length of the array).

In combination with #174 this would make it much easier to deal with variadic arguments of the same type. Something which even C++ is still missing.

[alexchandel]

This would need to be compatible with range syntax.

[gnzlbg]

How would one implement the cartesian product of two type tuples with this?

[eddyb]

@gnzlbg what exactly do you have in mind, Product<(A, B), (X, Y)> = ((A, X), (A, Y), (B, X), (B, Y))?

[gnzlbg]

@eddyb yes, but with variadics.

[goertzenator]

Having wrestled with c++11 templates and Rust macros, I really like this RFC. I hope it comes to fruition sooner rather than later.

[oblitum]

Me too, I really wished this to come before 1.0 and solved tuple impl etc.

[dgrunwald]

Unfortunately this syntax doesn't work, because the .. operator is already taken by range syntax. We could avoid this problem by using triple dots like C++. Which are also already taken (by inclusive range syntax), but only as infix operator and only in pattern syntax (though there were some requests to allow triple dots in expressions, too).

For the general idea: 👍, I was imagining variadic generics using pretty much the same approach. However, the devil is in the details that aren't yet worked out. C++ resolves expressions involving parameter pack expansion only after monomorphization; but I assume Rust would want to type-check the generic code? Wouldn't that require restrictions on where the unpacking operator is used? Also, consuming variadics in C++ usually involves a recursive function using template specialization. With this proposal, it seems like you'd need to create a trait for every variadic function, so that you can "specialize" using trait impls as in the Clone example?

Also, the ... operator in C++ allows "unpacking" complex expressions, not just directly unpacking the parameter pack. For example, in C++11 I can do:

template <typename... T> void consumer(T... t) { ... }

template <typename... T> void adapter(T... t)
{
   consumer(t.adapt()...);
   // expands to:
   // consumer(t1.adapt(), t2.adapt(), t3.adapt(), ...);
}

What's the easiest way to write the 3-line adapter function in Rust with this RFC? Do I have to write recursive Adapt implementations like the Clone impls in the example?

As for Vec::new: I think in cases where we don't actually need a type parameter pack, but just a variadic function with a single parameter type, it might be better to support something like this:
fn new(..elements: [T]) -> Vec<T> { ... }
This requires being able to pass unsized types as parameters - but that seems doable (and might already be planned?).

[alexchandel]

Why not the dereference operator?

type AppendTuple<T, U> = (*T, U);
f(*(a, b), *(c, d, e), f) => f(a, b, c, d, e, f);

[gnzlbg]

@dgrunwald :

Also, consuming variadics in C++ usually involves a recursive function using template specialization.

In C++1y there are fold-expressions that allow you to do a lot of things without recursion:

template<typename... Args>
bool all(Args... args) { return (... && args); }

bool b = all(true, true, true, false);

but I assume Rust would want to type-check the generic code? Wouldn't that require restrictions on where the unpacking operator is used?

You can easily type check that all parameters in the parameter pack implement a given trait. If one needs more complicated type-checking a variadic function is likely not the best solution.

[oblitum]

@gnzlbg that's C++1z actually.

[alexchandel]

The syntax between the value-level and type-level languages should be kept as close as possible. Already the + operator means the intersection of traits at the type level, even though * (or &) more accurately conveys a lattice meet.

@gnzlbg I don't think that form of duck-typing is compatible with Rust's trait-based generics.

Rather than having special fold syntax, it's cleaner to design some system compatible with HKTs for constraining the tuple's types, and then iterate over a tuple for fold any other operation:

fn fold_variadic<*T> (x: T) -> u32
    where *T: BitAnd<u32>
{
     x.iter().fold(!0u32, |b, a| b & a)
}

fold_variadic(0xFFF0, 0x0FFF) == 0x0FF0

In the above example, T is a tuple of variadic length, and * is a currying operator. This makes intuitive sense once you realize that type arguments are effectively pattern-matched in variadic generics. Consider the partial signature fold_variadic::<*T> and the invocation fold_variadic::<u32, u32, u32>, and notice how T is bound:

<*T> = <u32, u32, u32>
*T = u32, u32, u32

The curried/destructured tuple is matched to u32, u32, u32, so the tuple is (u32, u32, u32).

The interaction of a curried tuple with +, =, and :, as below, should also be specified.

where *T: BitAnd<u32>
// or
where *T = u32

A sensible semantic is distribution over the curried form. *T: BitAnd<u32> means every type in the tuple implements BitAnd<u32>, *T = u32 means every type in the (variable-length) tuple is u32.

[gnzlbg]

@oblitum you are correct.
@alexchandel my previous post was just to show that in future C++ you won't need template specialization to solve this problems in c++.

A recurring task in C++ meta-programming is zipping a tuple of values with a tuple of types (tags, or std::integral_constants). I wonder how this could be done in Rust.

[Vikaton]

Any update on this?

[semirix]

@Vikaton +1 I would be really interested to know as well. This is the only feature that I miss from C++.

[ticki]

How will this interact with type level numerals (in particular, polymorphism over arrays of size N)?

[canndrew]

There's been lots more discussion on this over on #1582 but I'll try to summarise here.

The main problem with all of these RFCs is the problem of memory layout for tuples. In general, in current Rust, a tuple of the form (A, B, C) does not contain a tuple of the form (B, C) anywhere in its representation. This means if you have an (A, B, C) you can't take a reference to the "sub-tuple" (B, C) because it doesn't exist. For example, if (A, B, C) is the tuple (u16, u16, u32) it may be represented as:

------------------------------------------------
| Byte | 0  | 1  | 2  | 3  | 4  | 5  | 6  | 7  |
|------|---------|---------|-------------------|
| Data | A (u16) | B (u16) |      C (u32)      |
------------------------------------------------

By comparison, a tuple (B, C) == (u16, u32) may be represented as.

------------------------------------------------
| Byte | 0  | 1  | 2  | 3  | 4  | 5  | 6  | 7  |
|------|---------|---------|-------------------|
| Data | B (u16) | padding |      C (u32)      |
------------------------------------------------

Note the different position of B.

There are two proposed solutions to this:

• Don't allow references to "sub-tuples" of tuples. Instead, when binding part of a tuple by reference, bind it as a tuple of references. This solves the problem but has the potential to be confusing. Under this scheme let (head, ...tail) = (0, 1, 2) will give you head == 0, tail == (1, 2) but let (ref head, ...ref tail) = (0, 1, 2) will give you head == &0, tail == (&1, &2) rather than tail == &(1, 2).
• Only allow expanding tuples into the tail (or alternatively the prefix) of a larger tuple. This means that (a, b, c, ...d) would be valid but using ... anywhere else would not. Eg. these would not be valid: (a, b, ...c, d), (...a, b, c, d), (a, b, ...c, ...d) etc. This still allows you to define anything you could without this restriction, it just becomes a lot more cumbersome for some usecases. Eg. instead of being able to write (...a, b) you'd have to write this. Additionally, this approach requires that one of two changes are made to tuple layouts. The first option is to pack tuples less space-efficiently so that (A, B, C) always contains a (B, C). For example (u16, u16, u32) would have to be packed as

--------------------------------------------------------------------
| Byte | 0  | 1  | 2  | 3  | 4  | 5  | 6  | 7  | 8  | 9  | 10 | 11 |
|------|---------|---------|---------|---------|--------------------
| Data | A (u16) | padding | B (u16) | padding |      C (u32)      |
--------------------------------------------------------------------

The second option is to accept RFC #1397 and reverse the layout order of tuples so that (u16, u16, u32) would look like

------------------------------------------------
| Byte | 0  | 1  | 2  | 3  | 4  | 5  | 6  | 7  |
|------|-------------------|---------|---------|
| Data |      C (u32)      | B (u16) | A (u16) |
------------------------------------------------

And its tail (u16, u32) would look like

------------------------------------------------
| Byte | 0  | 1  | 2  | 3  | 4  | 5  | 6  | 7  |
|------|-------------------|---------|---------|
| Data |      C (u32)      | B (u16) | padding |
------------------------------------------------

This comes at the cost of disallowing the compiler from generating code which overwrites padding at the end of tuples as the "padding" may actually contain the next element of a larger tuple.

[gnzlbg]

IIUC what you want is to be able to slice (A,B,C) into (B,C). I have some questions:

• Why/when isn't (&B, &C) enough? Is there a real use case for this? We cannot do this for structs, so why should we be able to do this for tuples?
• Why does the language have to define the layout of tuples? I would rather have it be "implementation defined" so that the implementation can reorder the elements of a tuple to minimize padding. Some C++ implementations of tuple actually do this [*]. We can, like for structs, use #[repr(C)] to get a specific layout.

I think it would be great to be able to slice (A,B,C) into (B,C) and still have (A,B,C) with a perfectly packed memory layout but I don't see how we could achieve both things.

[*]: https://rmf.io/cxx11/optimal-tuple-i/, https://rmf.io/cxx11/optimal-tuple-ii/, https://rmf.io/cxx11/optimal-tuple-iii/ , https://rmf.io/cxx11/optimal-tuple-iv/

[gnzlbg]

Meaning if the trade-off is between sugar for a particular use case (slicing (A,B,C) into (B,C) instead of just (&B,&C)) or a zero-cost abstraction (a perfectly packed tuple) I would chose the zero-cost abstraction any day.

[glaebhoerl]

FWIW the last option mentioned by @canndrew, involving #1397, seems like clearly the cleanest and nicest design to me. The only big question mark hanging over it in my mind is the backwards (in)compatibility impact of #1397 on already-existing unsafe code that uses size_of where only stride would now be correct.

[bluss]

I want to crosslink you to the fresh RFC clarification issue /pull/1592 because the commitment to having a specific position in the tuple possible to be unsized is relevant to the tuple representation discussion here.

[Diggsey]

@canndrew The name tuple refers to both those things: while the name is ambiguous I don't think anyone's been conflating the two: we already have (0u32, 0u32) : (u32, u32) and we don't need HKT to express that.

The expansion syntax would presumably work for both tuple types and tuple values, although the layout argument is only relevent for tuple values.

Also lifetimes are not an issue since lifetime arguments are always separated from type arguments, there's no need for a construct which supports both. It's possible to start with just our existing tuples, and then potentially later add a notion of "tuple lifetimes" or similar, which might look like ('a, 'b, 'c) and which would be unusable on their own (only as part of an expansion)

[eddyb]

@canndrew Actually, in my demo I've used a function which turns &(A, B) into (&A, &B), not unlike the C++ std::forward_as_tuple function @petrochenkov mentioned, and I agree that's better, if only because modelling the type is not too hard in a library, but strange in the compiler itself (the best thing it can do is just use a lang item which does the type transformation).

[canndrew]

@Diggsey

I don't think anyone's been conflating the two: we already have (0u32, 0u32) : (u32, u32) and we don't need HKT to express that.

No, but we'd need it to express something equivalent to (u32, u32) : (type, type) because at the moment we only have (u32, u32) : type. And @petrochenkov was conflating them here but then realised what he was doing in his next comment. It's an important distinction because (str, str) isn't a valid type, should never be read as a type and should never appear in a type position. And ('a, 'b) certainly isn't a valid type. But there might be things that are written like that when we have HKT and can talk about concepts like type -> type, (type, type) or (lifetime, lifetime).

@eddyb
When I mentioned a compiler intrinsic I meant that without syntax to split a tuple by reference it wouldn't be possible to implement the std::tuple::split function I mentioned above.

[eddyb]

@canndrew Ah, I agree then, we can use an intrinsic relying on a trait to produce the right return type in generic contexts, while the actual monomorphized implementation of the intrinsic would only care about concrete tuple types.

[gnzlbg]

Currently (u16, u32, u16) requires 12 bytes, while (u32, u16, u16) requires 8 bytes. I would expect them to be the same size since the compiler should be able to reorder the elements of the tuple to minimize padding automatically.

While I now see the advantage of &(B, C) vs (&B, &C) (one pointer vs two pointers) that prevents the optimization above.

Is there a way to get both optimizations?

[petrochenkov]

@canndrew

And @petrochenkov was conflating them here but then realised what he was doing in his next comment.

Well, the "conflation" was certainly intentional in my first comment, I do want to use tuple types for expressing type lists.
To conflate or not to conflate is a design choice, I don't think the alternative from my second comment is unambiguously better.

[vi]

Shouldn't ... be reserved for for i in 1...10 inclusive range syntax?
More dots, less problems: ....

[eddyb]

@vi start...end and ...end are taken but start... isn't used.

[ubsan]

This would also make our syntax line up with C++'s, which is... I don't know, a nice symmetry at least.

[canndrew]

start...end and ...end are taken but start... isn't used.

Having a... mean something completely different to a.., a...b, a..b, ...b and ..b seems potentially very confusing. I don't think the (a, b, c; d) syntax looks bad. I read the semicolon as a hard comma that says "everything after here is rolled into one thing". It's also similar to the [a; n] syntax.

[canndrew]

I decided to reopen #1582 largely because of this comment. I'll close it again if @eddyb or @nikomatsakis disagree with the decision.

[eddyb]

@canndrew It seems non-trivial to see at a glance what is being spread, and it's non-symmetric.
I agree the range syntax overlap is unfortunate, but (a, b...) and (a..., b) are obvious IMO.
The semicolon is also easier to miss and would look really strange for variadic functions:

fn foo<T: Tuple>(; args: Tuple) {...}

[leonardo-m]

Regarding this instead of looking at C++ design I suggest to look instead at the D language and its type lists (once named type tuples,
https://dlang.org/phobos/std_meta.html ), that support an array-like indexing syntax, a array-like slicing syntax, a array-like concatenation syntax, this is much better than what C++17 offers (one problem of the D is that they auto-flatten).

[eddyb]

What D does strongly requires either duck typing or value refinement / dependent types.
E.g. pair[0] having a different type than pair[1] - AFAIK you can't even do this in standard Haskell.

[leonardo-m]

Right, D type lists are dynamically typed at compile-time, just like D templates are half-dynamically typed... (half because in D you often use template constraints to perform half type checking) Unlike Rust. Thank you eddyb.

[gnzlbg]

@eddyb

What D does strongly requires either duck typing or value refinement / dependent types.
E.g. pair[0] having a different type than pair[1] - AFAIK you can't even do this in standard Haskell.

Type level integers would be enough. This is a minimal form of dependent types[*], but you could then use pair[Usize<0>{}] and pair[Usize<1>{}] to call two different trait methods, so that pair[] can return different types.

One could probably think of multiple combinations of more or less complex rules to make pair[0] work. Not that I like any, but one could be to make all integer literals dependently typed, of type std::Usize<VALUE>, which implements a (potentially magic) conversion to usize. Or also the opposite, give the std::Usize<VALUE> type a "magic" constructor from integer literals so that when calling pair[0] the VALUE parameter is deduced. And probably many more alternatives.

[*] EDIT: maybe we could already do something like this without dependent types at all by making all integer literals be of their corresponding type in the typenum crate, e.g. 3usize of type Usize3, and implement some magic on those.

EDIT 2: The main reason I don't like this "style" of dependent types is that it might generate an explosion in code size and compile times. Each literal integer value has a different type, each function taking one of those is a different function, etc. But I think something like this should be explored as an ergonomic improvement to type-level integers, and in a way that doesn't impact code-size or compile-times, after the minimal RFC goes through because I hate having to manually construct std::integral_constant<int, 3> in C++. Pinging @withoutboats since this is a bit offtopic here.

[eddyb]

What you are describing is what I call "value refinement" e.g. Index<0usize> - where that 0usize is a subtype of usize with only one value, 0. This is more powerful than just const generics and can generalize to runtime values just like dependent types, and also to ranges, variant types, etc.

[petrochenkov]

@leonardo-m

I think the current design of Rust tuple indexing with dot-number was a big design mistake

Nothing prevents adding a new syntax for this as well (especially together with type/value lists), e.g. x.[y], that can't be overloaded, requires y being a constant expression (index or maybe range), x being a variadic list, or tuple, or array, or struct (or even closure!) and returning the y-th element of x at compile time.

[eddyb]

s[Struct::field] is another related idea but that is more reflection oriented

[qraynaud]

I was planning to transcribe some C++ code I have in Rust. It is some playground project of mine that is somehow like yacc. It currently can generate correct parsing tables and parsers for any LALR(1) grammar for C++ and Javascript.

Somehow, I wanted to move the grammar analysis codebase to Rust and also the reference parser implementation. My concern is I'd like to be able to define my parsers like I did in C++ / JS: for each pattern in each rule in the grammar, you have to provide a method to construct the value it produces consuming the subrules / tokens values. I had an interface like this:

template </* … */>
class parser: base_parser</* … */> {
  void register_rules() {
    this->register_code("rule_name", 0, *this, /* pattern index in the rule */, &parser::rule_name_0);
  }

  float rule_name_0(int i, float f) {
    return i + f;
  }
}

The form with 4 args is providing a method, the 3rd being the object on which the method should be invoked on while with 3 args, you could pass on any function instead. What is important to me is that register_code is able to accept any function / method whatever the arguments if the arguments are all comprised in the types included in a variant type the parser is templated with and if they have less than 50 arguments (the limitation is caused by the fact that I used macros to implement this because when I wrote this code, c++11 was not really around). If you want to take a look at what I did, the code is here: https://bitbucket.org/qraynaud/jungle/src/default/src/libs/jungle/caller.hpp.

I have in the internal parser code a class that can invoke any of those methods using a provided array of values of the variant type I talked of before. Since everything relies on proper templating, the type checking is ensured at compilation for the most part. The variant still has to check it has the correct type value in it at execution time but it will fail properly if not.

I find this important because it makes the rules codebase easy to understand. You put an argument for every token / subrule your pattern contains (eg rule: expr "+" expr + "int" would result in a signature like (left: f64, plus: (), right: f64, i: u32)). I could give to the method the array of arguments directly (like an array of enums) but it feels wrong to me.

Since the basic parser code is not generated, I have the feeling that it is impossible to achieve anything even remotely like this in Rust as of today. It would be great if this spec could take into account those kind of usages to make them possible.

[tinyplasticgreyknight]

@qraynaud I'm not sure if I understand what you're looking for exactly, but would it be a useful workaround if your function always accepts only a single argument, which might be a tuple containing the "real" arguments?

Maybe I should read your original repository and get a feel for how this is all used :-)

[qraynaud]

@tinyplasticgreyknight : even so, every function would have a different tuple type and that would not work either for me right?

Each method for each parser rule's pattern has a different signature. I think that was not clear enough. What I implemented in C++ is similar to the apply method in JS (obj.method.apply(obj, [arg1, arg2, …])).

I don't really see how I can build the correct tuple, its signature being something like (..T) with each T being one type between a list of known types. With this idea I got my problem moved. Before it was calling a method using a vector of arguments, now it is constructing a tuple with a vector of arguments. Not knowing beforehand the types & the numbers of arguments…

[NateLing]

Will it implement in 2018?

[Centril]

@NateLing up for writing a formal RFC proposal about it? ;)