Document #: | xxx |
Date: | 2021-19-06 |
Project: | Programming Language C++ SG17 |
Reply-to: |
Mihail Naydenov <mihailnajdenov@gmail.com> |
This paper suggests alternative way of declaring parameter lists, one that let us omit parameter types.
Lambda expressions are often used in a specific context, which can be used to infer some of their required elements. Because of this, there is a strong desire to have the shortest possible form of such expressions. This resulted in few proposals, like for example Abbreviated Lambdas (P0573R2)1. P0573R2 did not pass for few reasons,2 here is a summary:
It is not hard to notice, two of the issue are related to the semantics of the Abbreviated Lambdas body, one is related to the parameters.
This paper also makes the observation, often it is the params that are contributing to the verbosity of a lambda the most, if we focus on the day-to-day uses and not the “perfect forwarding” scenario:
Parameter types dominate the lambda expression in common cases.
An argument can be made, we could use
auto&&
and get rid ofconst
, saving significant number of characters. This is not entirely true, because,auto&&
will be aconst
reference only if the original object isconst
. Most of the times this is not true - the object is notconst
, yet we want to immutably operate on it. By usingauto&&
we would change the meaning of our lambda, no matter how we look at it.
Interestingly, not only we will gain the most if we get rid of the types, sans “forwarding”, but it seems, the problems we had in attempt to do so are mostly technical - no easy way to differentiate b/w normal parameter list and one with no types. The body issues on the other hand are significantly more involved, while at the same time giving smaller verbosity reduction in day-to-day code.
This paper suggest splitting the issues in “body issue” and “params issue” and deals only with the latter.
Introduce a slightly different syntax to denote Abbreviated Parameters:
Double parenthesis indicated Abbreviated Parameters.
If an parameters list starts with double parentheses ((
, then all single identifier are not types, but variables instead:
auto something(string, vector); //< declaration of function with two parameters of type `string` and `vector`.
// param names are omitted
auto something((string, vector)); //< declaration of function with two parameters of deduced type.
// param names are `string` and `vector` (this proposal)
Using multiple identifiers results in exactly the same declarations no matter if single or double parentheses are used:
auto something(vector v); //< same declarations
auto something((vector v)); //<
auto something(const vector& v); //< same declarations
auto something((const vector& v)); //<
Mixing single and multiple identifiers is possible:
In other words, types are optional. Parameters of deduced type are templated parameters. Functions and lambdas, containing such parameters are templates. The above declaration is equivalent to:
Function template. <qualifiers>
are discussed in Part 2.
Mixing with regular template parameters is also possible:
template<class T> void func((T a, b));
[]<class T>((T a, b)) {};
// equivalent to
template<class T, class U> void func(T a, U <qualifiers> b);
[]<class T, class U>(T a, U <qualifiers> b) {};
As well as omitting identifiers, partially or completely:
void func((,));
[]((a,,c,)) {};
// equivalent to
template<class T, class U> void func(T <qualifiers>, U <qualifiers>);
[](auto <qualifiers> a, auto <qualifiers>, auto <qualifiers> c, auto <qualifiers>) {};
If there are no identifiers and no commas, the list is empty:
If the parameters list does not contain deduced types, or extra commas, the declaration is not a template:
Default arguments can be supplied as usual:
void func((a = 12, b = "hello"s));
// equivalent to
template<class T=int, class U=std::string> void func(T <qualifiers> = 12, U <qualifiers> b = "hello"s);
The syntax can not be used to declare function pointers or signatures:
// using FuncPtr = void(*)((string)); // error
// using FuncPtr = void(*)((std::string)); // error
// using FuncPtr = void(*)((const std::string&)); // error
The main reason to disallow this is to be consistent with current auto
rules:
Although, it could be argued auto
in using
declaration can be allowed to mean a template param for the using
declaration itself:
As this is completely out of scope for the current proposal, double parentheses are allowed only where auto
placeholder is allowed.
Similarly, function declarations inside a block are not allowed :
This should come to no surprise, because function templates in block scope are not possible today. However, even if we allow these, double parenths declarations will still be disallowed, because it will be a breaking change: auto something((string));
is already a valid expression (a constructor call to something
object, taking string
variable as an argument).
This covers Part 1 of the proposal - allow a 3th way of declaring function templates, one that will let us to omit the type (or its placeholder in the form of auto
) completely. Part 2 will discuss the options we have with regard to the deduced type and its decorations.
As stated in the beginning, getting rid of the parameters type is one of the best “bang for the buck” in the attempt to lower verbosity:
// given
struct fruit {...};
std::vector<std::string> fruits{"apples", "oranges","cherries"};
std::vector<std::vector<fruit>> baskets{...};
std::transform(fruits.begin(), fruits.end(), baskets.begin(), fruits.begin(),
// from
[](const std::string& fruit, const std::vector<fruit>& basket) { return fruit + ": " + std::to_string(basket.size());
// or
[](const auto& fruit, const auto& basket) { return fruit + ": " + std::to_string(basket.size());
// to (this proposal)
[]((fruit, basket)) { return fruit + ": " + std::to_string(basket.size());
});
// `auto&&` is discussed in Part 2
If we take a theoretical “abbreviated body”, the gains are much less in terms of reducing verbosity, at least in this example, though I would argue, the example is fairly representative in general:
std::transform(fruits.begin(), fruits.end(), baskets.begin(), fruits.begin(),
[](const auto& fruit, const auto& basket) { return fruit + ": " + std::to_string(basket.size();
});
std::transform(fruits.begin(), fruits.end(), baskets.begin(), fruits.begin(),
[](const auto& fruit, const auto& basket) => fruit + ": " + std::to_string(basket.size()
);
As you can see, if we ignore the generic, forwarding case, the gains are almost non-existent - a constant 7 characters total, 6 of them from return
, 1 for the semicolon.
This is not to say the generic case is to be ignored - quite the contrary. Most of the benefits in an “abbreviated lambdas” will come from a new, smarter “abbreviated body”. An “abbreviated body” will solve more problems then simply limiting verbosity as it will lower the complexity of advanced topics that only look simple on a first glance. This paper is not against any “abbreviated body” solutions and it does not block or contradict ony of them.
In any case however, “abbreviated parameters” have enough practical benefit for “abbreviating” the lambda expression as a whole to be worth pursuing on their own.
Interestingly, if the current proposal is applied to range-for
, it solves the main issue, which led to the rejection of the for(elem : range)
syntax.
The problem was that declarations like this can be misleading5:
int i;
for(i=0; i<size; i++) //< these are different - reuse-vs-introduce
for(int i=0; i<size; i++) //<
something i;
for(i : container) //< are these different as well?
for(int i : container) //<
Having a syntax that already means variable introduction (of deduced type) solves this issue completely:
Double parentheses always mean variable introduction.
This part deals with what the actual deduced type should be.
The often regarded “right solution” for that is to use auto&&
- it is what p0573 suggested and was previously proposed for the for(elem : list)
idea as well.
This proposal argues against it, though not strongly. It is a valid option and the current proposal, at its core (Part 1), will be equally well served by any type deduction. With that out of the way, let’s see pros and cons of other options.
auto&&
The single biggest downside of auto&&
is its mutable reference semantics as default. These are not desirable defaults.
Lets return to our fruits example:
std::transform(fruits.begin(), fruits.end(), baskets.begin(), fruits.begin(),
[](auto&& fruit, auto&& basket) { return fruit += ": " + std::to_string(basket.size());
});
If the user incidentally assigns to fruit
, it will break std::transform
preconditions.
It could be argued, the user can use fruits.cbegin()
instead of fruit.begin()
, but this requires the same amount of attention as writing const
for the lambda params. If we talk about safe default, auto&&
is not one.
Besides in some cases cbegin()
is not even an option:
std::sort(fruits.begin(), fruits.end(),
[](auto&& lhs, auto&& rhs) { return lhs.price() < rhs.price(); }
);
What if price()
is not the correct method (or overload!) as it, for example, fetches price from an external source, does some other processing, and is undesirable in this context for performance or other reasons? We created a problem where there was none before. If this is example looks contrived, then why do we have begin()
and cbegin()
in the first place?
It gets even more interesting, if we take into account how predominant const
lambdas are in practice. For example, from all standard algorithms, only one is usable with mutating lambda - for_each
! Even if the algorithm itself is mutating, the function object itself is not, instead the return value is used to write to desired destination (either to the same container or not, resulting in either mutation or copy).
An argument can also be made about forwarding scenarios, where non-const ref-ref is the desired type. However, forwarding specifically is excursively a library/wrapper feature. These scenarios are by definition less common then “the general usage” - a “library” is meant for code-reuse, and code-reuse is “write once, reuse many times”. This is, the day-to-day usage of a language is always more common then writing a library for it.
const
backIf we consider immutability is always a good thing, how do we get it back?
We either need full verbosity:
Or, well, partial verbosity, if we extend the possible definitions to allow “naked const
”
Considering, for many types, we can just use non-reference auto
today, we did not achieve much.
For an Alternative solution see Appendix.
This paper suggests, a const reference and/or const copy as the best options for the deduced type.
More specifically, const reference to be the default, but, if deemed feasible, allow optimizers to use copy instead for improved performance.
std::string s;
std::vector<std::string> v;
std::for_each(s.begin(), s.end(), []((c)){ ... }); //< decltype(c) is `const char`
std::for_each(v.begin(), v.end(), []((s)){ ... }); //< decltype(s) is `const std::sting&`
Because in both cases we are dealing with a const
arg. the effect of the transformation to copy will be close to invisible to the user. The only observable side-effect will be if the user, or a subroutine, takes and stores the address of the argument, with the assumption, its lifetime is greater then the lambda itself. In that case the address can become invalid or not, depending on the argument type.
How much this is a problem in practice is hard to jude, but this paper argues, it is extremely unlikely, a user will intentionally take the address of a param (or call a function which does so) AND use a signature, which does not have reference/pointer symbol on the param:
It is unlikely, the user expects, arg
storage to outlive the call to func
.
This goes back to the fact - there are no hidden references in C++. Sure, arg could be a reference, but no reason for the user to expect that. There is no previous experience to create such expectations!
Also, the implementation can be such that, the user can query and/or assert if the conversion to reference will take place. Something along the lines of static_assert(sizeof(arg) < TO_CREF_THRESHOLD)
. This will allow 100% safe code, albeit with increased complexity.
Having said that, I do see this feature as potentially controversial, that is why, it is suggested only if feasible - const reference alone is good enough default. Besides, if the user wants to have a copy instead, a simple auto
is still an option.
While on the subject of copying, why aren’t the arguments passed by copy, the way they are passed in functions and lambdas already, if there are no cvref decorations?
There are multiple arguments against that.
The pro-arguments are much weaker:
The first argument is weak, because we have a precedent with lambda captures. Captures can be though as “arguments, passed on lambda definition”. Not only that, but the immutability of lambdas has overwillingly positive reception in the community. In a way, by making the parameters of a lambda immutable, we increase the overall consistency of the lambda expression as a whole - now all arguments, passed to a lambda, either on definition or call time, are immutable.
The second argument is also weak, because it does not affect how we write code - the reference is there purely for performance reasons, exactly how it is used today.
The issue about param address is already discussed.
A note about coroutines. This paper does not explore this space, but suggests, it is good idea to have different rules for deduction for coroutines. This is because references and coroutines do not match well. It looks like capture by copy to be much better option, however this is a separate discussion and for now coroutines should not support double parenths syntax. We simply don’t have enough experience with coroutines to be sure of the right defaults. Coroutines are significantly enough different, then either lambda or functions, and require a separate evaluation. Extending the current proposal can be done on later date.
The above scenario handles “observation” functions only. If the user wants to modify the original argument passed (or want to ensure it has its address passed to the function), then using both &
or &&
are proposed to serve these purposes:
auto& incr((&src, delta)) { return src += delta; } //< `src` is of type auto&, `delta` is of type `const auto(&)`
[]((&&a)){ return some_func(std::forward<decltype(a)>(a)); } //< `a` is of type auto&&
By covering the 3 scenarios above (“observation”, “mutation”, “forwarding”) we solve all the issues listed above.
*The immutable reference is an implementation detail, an optimization, and is not semantically relevant.
In should be noted, so far auto&&
as a default was ever only suggested in a very limited context - in a for-loop and lambdas (or lambda-like, single expression functions). Current proposal is not limited to those type of uses alone and anticipates typeless params potentially in normal functions as well. This changes the requirements a bit, as being able to express all common parameter usages (“observation”, “reference”, “forwarding”) is of much greater value. This is, while the niche use case of for-loop or lambda can be handled by one “good enough” type, a more universal support of deduced params require greater flexibility and expressiveness.
For example, consider helper functions, local to a cpp file, or a private function in class. In these scenarios it is quite possible, the user might not use concepts to define such functions and opt for the “quick-and-dirty” approach as the functions are not really part of an API and the extra effort to “do it right” might not be worth it. Or may be this is a changing code, in a prototyping stage - the user is unsure what the requirements are. In these scenarios it will be highly unlikely, the “forwarding reference” to be the right type (even less so for all arguments!), and is quite likely, the user will want some control over the parameters, at least in terms of mutability:
namespace {
void assignPedalCrankFromRotorSpeed((&pedal, rotor)) { pedal = rotor / CHAIN_FACTOR; }
...
}
A local helper function
auto&&
As said in the beginning, auto&&
is reasonable solution as well. Its main strength is its simplicity both in terms of definition and implementation, as well as the fact “it just works” in code. There is a charm to that.
The “just works” part is considered especially relevant in cases where a mutating param must bind to a proxy object:
std::vector<int> vi(10);
std::vector<bool> vb(10);
std::for_each(vi.begin(), vi.end(), [](auto& v) { v = 1; }); //< works
std::for_each(vb.begin(), vb.end(), [](auto& v) { v = true; }); //< fails: MUST be auto&&
There is no denying, “it will be nice” to be able to write a minimal expression, which will work in both cases, even more so, considering the cases where proxy objects are used are expected to increase somewhat with ranges
.
However, this is only true from modifiable references, normal observation works with no added verbosity:
std::vector<int> vi(10);
std::vector<bool> vb(10);
std::for_each(vi.begin(), vi.end(), []((v)) { if(v == 1) ... ; }); //< works with current proposal as well
std::for_each(vb.begin(), vb.end(), []((v)) { if(v == true) ...; }); //<
This leaves us back to the fundamental question: Is it worth trading const? Consider:
std::vector<int> vi(10);
std::vector<bool> vb(10);
std::for_each(vi.begin(), vi.end(), []((v)) { if(v = 1) ... ; }); //< "just works"?
std::for_each(vb.begin(), vb.end(), []((v)) { if(v = true) ...; }); //<
Do we want to drop safer code for more “easy” one? Do we want to not guard against the above code, so that we can have “nice, universal syntax” in the limited context of mutation alone? It is an honest question, and although this paper argues for const, the opposite still has merit.
Related to that also is the question, do we want less clear code if favor of marginally less verbose.
Is this “better”:
then
&&
?&&
“tip the scale” and make the declaration “verbose”?This paper expresses doubt. Where auto
gives no additional information and can be cut out, &&
does convey information!
Of course “people will learn”, but it is not that simple. Where now (with &&
) one can instantly see what kind of function he/she is dealing with (is it forwarding or not), when there is no such indication, all bets are off until the entire body is read:
auto function((a, b, c)) {
...
// 10 lines later
something(a);
...
// another 20 lines
other_thing(std::forward<B>(b));
...
// yet another 30 lines
return do_it(std::forward<C>(c));
}
What are the chances, later a colleague comes in and:
...
// another 20 lines
other_thing(std::forward<B>(b));
...
// yet another 30 lines
c.omn(b); //< ok?
...
return do_it(std::forward<C>(c));
}
What if b
is moved by forward
? Did the colleague noticed the forward
at all, somewhere in the middle of the function?
Mistakes like this are best handled by static analysis for sure, but having more expressive parameters does not hurt either:
The &&
is a clear hint, there might be forward at some point!
Less is not always more.
default to auto&&
+ Simple implementation
+ Universal use in all cases
+ Minimal verbosity
- Less safe (or more verbose)
- Less clear
- Learning curve for confident use
default to const auto(&)
, allow &
and &&
+ More safe
+ More clear
+ More control
+ Minimal-to-No learning curve, consistent with current practices overall
+ (Better performance if conversion to const auto
is feasible)
- More involved implementation
- Marginally more verbose in some cases
- “Pedantic”?
- (Even more involved with const auto
conversion)
The above summary shows, no solution is really “better” - there are trade-offs in both cases.
This paper only hopes, we make the most informed decision when choosing one trade-off over another.
The auto&&
vs const auto&
debade can change considerably if the core of the proposal (Part 1) is extended to allow specifying mutability for all arguments:
This shifts the optics about using auto&&
as it becomes easier to gain immutability (as well as potential “maybe-reference” implementation).
Having such option does not contradict the const reference default as proposed, but makes many aspects of it redundant, like for example allowing &
and &&
syntax:
Some of the arguments pro explicit references still stand, but are greatly dampened, that is way this is more of an “Alternative solution”, not an “extension” - const ref default + parenth-const is not propsed.
Parenth-const creates some interesting possibilities, unrelated to abbreviated parameters, like allowing us to have all-const parameters in normal functions:
void function((std::string& s, std::vector<int>>& v, int i) const ); //< all arguments are immutable
//And/Or may be const in front
void function(const(std::string& s, std::vector<int>>& v, int i)); //< all arguments are immutable
As you can see, no abbreviated parameters in sight - we only take advantage of having to type just one const
for all params. Of course, the value of this feature greatly depends on the number of parameters (and if we have a mix of const and non-const params), but is interesting nevertheless.
Arguably, the most interesting results are when this is applied to the theoretical for-range
extension:
// `auto&&` is the default + parenth-const + range-for extension
for((e : list)const) //< some verbosity for observation, but clear and safer
// equivalent to
for(const auto&& e : std::as_const(list)) //< and immutable container!
for((e : list)) //< minimal-verbosity mutation and universal
// equivalent to
for(auto&& e : list)
As shown, this could solve the age-old problem where the list itself is not immutable, leading to undesired side-effects in some situations like when used with ref-counted containers.
Should be noted, when e
is by default deduced to const auto&
we can simply re-define the expression to have immutable container, if there is no reference on the element variable. In other words, we can get this feature (a const container) for free, because of the immutable default. The downside is that we lose minimal-verbosity mutation and we have to use &&
to be universal.
// `const auto&` is default _alone_ + range-for extension
for((e : list)) //< minimal-verbosity observation
// equivalent to
for(const auto& e : std::as_const(list)) //< and immutable container!
for((&&e : list)) //< to be universal, we need &&
// equivalent to
for(auto&& e : list)
auto&&
is not the primary proposal?- More verbose for the common, observation case.
- Hidden mutable reference is still the default!
- Introduces more new rules - not exactly a “natural extension” of preceding art.
~ Interesting, but questionable value outside abbreviated parameters. My be with the exception of range-for
?
+ Allows all mutating cases to be at minimal-verbosity possible.
+ Simpler, no need to introduce both &
and &&
.
Unsurprisingly the value of the alternative solution relies once again on how we much we value immutability and what price we want to pay for it.
Do we want default const and can live with “references noise” when modifying or want clear code when modifying and have const as an explicit opt-in, with the usual verbosity that comes with it.
Abbreviated Lambdas: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0573r2.html↩︎
Barry Revzin blog: https://brevzin.github.io/c++/2020/01/15/abbrev-lambdas/↩︎
Abbreviated Lambdas: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0573r2.html↩︎
Change scope of lambda trailing-return-type: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2021/p2036r1.html↩︎
range-for discussion: https://groups.google.com/a/isocpp.org/g/std-proposals/c/nKRCQVvCxD8↩︎