P3176R0
The Oxford variadic comma

Note: In this proposal, the term ellipsis parameter refers to non-template variadic parameters, i.e. "C-style varargs", i.e. as in (/* */, ...).

1. Introduction

[P1219R2] "Homogeneous variadic function parameters" proposed a new feature, where the declarator (int...) would be interpreted as a homogeneous function template parameter pack. As part of this, it made it mandatory to separate an ellipsis parameter from preceding parameters using a comma.

The proposal did not pass, but the change to commas achieved strong consensus by EWGI in Kona 2019.

EWGI Poll: Make vararg declaration comma mandatory.

SF	F	N	A	SA
10	6	1	0	0

Going back to 2016, [P0281R0] "Remove comma elision in variadic function declarations" also proposed to make the variadic comma mandatory. Due to its impact on existing code, the proposal did not pass.

My proposal continues where [P1219R2] and [P0281R0] have left off. However, I call for deprecation, not for removal, resulting in the following behavior:

// OK, function template parameter packtemplate<class... Ts> void a(Ts...);// OK, abbreviated function template parameter packvoid b(auto...);// OK, ellipsis parameter, compatible with Cvoid c(int, ...); void d(...); // Deprecated, ellipsis parameter, ill-formed in Cvoid e(int...);void f(int x...);// Ill-formed, but unambiguousvoid g(int... args);

[P1161R3] "Deprecate uses of the comma operator in subscripting expressions" has followed the same strategy. It deprecated array[0, 1] in C++20 which freed up the syntax for [P2128R6] "Multidimensional subscript operator" in C++23.

1.1. History and C compatibility

The active version of the C standard ISO/IEC9899:2018 permits functions which accept a variable number of arguments. Such functions must have a parameter-type-list containing at least one parameter, and the ellipsis parameter must be comma-separated:

parameter-type-list:
        parameter-list
        parameter-list , ...

This syntax is unchanged since C89. C has never permitted the comma to be omitted.

Such ellipsis parameters were originally introduced in C++ along with function prototypes, but without a separating comma. Only int printf(char*...) would have been well-formed in pre-standard C++, unlike int printf(char*, ...).

For the purpose of C compatibility, C++ later allowed the separating comma, resulting in the syntax (unchanged since C++98):

parameter-declaration-clause:
        parameter-declaration-list_opt ..._opt
        parameter-declaration-list , ...

With the introduction of function template parameter packs in C++11, there arose a syntactical ambiguity for a declarator (T...) between:

• a parameter pack of type T... and

• a single T followed by an ellipsis parameter.

Currently, this ambiguity is resolved in favor of function template parameter packs, if that is well-formed.

2. Motivation

The declarator syntax (int...) interferes with future standardization. This has already impacted [P1219R2] and will impact any future proposal which attempts to utilize this syntax.

Furthermore, variadic function templates are a much common and flexible way of writing variadic functions in C++. Many users associate (int...) with a pack, not with an ellipsis parameter. Not parsing this syntax as a parameter pack is confusing and has potential for mistakes:

template<class Ts>
void f(Ts...); // well-formed, but equivalent to (Ts, ...)

Users who are not yet familiar with variadic templates can easily make the mistake of omitting ... after class. Another possible mistake can occur when writing generic lambdas or abbreviated function templates:

// abbreviated variadic function template
void g(auto ...args);
// abbreviated non-variadic function template with ellipsis parameter
void g(auto args...);

These two forms look awfully similar and are both well-formed with different meaning. The latter should be deprecated.

Lastly, as explained above, only the declarator (int, ...) is compatible with C, not (int...). The latter syntax is arguably pointless nowadays and only exists for the sake of compatibility with pre-standard C++.

2.1. T......

C++ also allows (auto......) or (T......), which is a declarator containing a function template parameter pack, followed by an ellipsis parameter. Such parameter-declaration-clauses are confusing because conceptually, they consist of two separate constructs, but the syntax strongly suggests that all . apply to auto or T.

Note: In other words, the declarator (auto......) is equivalent to (auto..., ...).

Note: Some users believe that this construct is entirely useless because one cannot provide any arguments to the ellipsis parameter. However, arguments can be provided if T... is template type pack which belongs to the surrounding class template.

3. Impact on the standard

This proposal is a pure deprecation.

No new features are proposed or removed, no semantics are altered, and no existing code is impacted.

4. Impact on existing code

No code becomes ill-formed. Any deprecated uses of T... can be converted to T, .... This transformation is simple enough to be performed by automatic tooling.

I conjecture that a non-trivial amount of code would be deprecated. It is hard to find the exact amount because (T...) could be using an ellipsis parameter or a parameter pack; one doesn’t know without compiling. This prevents simple syntax-based approaches to searching for occurrences. However, I was able to find a few dozen occurrences of T...... using a GitHub code search for /\(.+\.\.\.\.\.\.\).+\{/ language:c++ and other patterns.

5. Design considerations

5.1. Why deprecate, not remove?

I do not propose to remove support for ellipsis parameters without a preceding comma because there is no feature proposed which requires such removal. A deprecation is sufficient at this time.

Sudden removal has already been attempted by [P0281R0] and did not pass. It is unreasonable to suddenly break a large amount of C++ code with little motivation.

The approach of [P1161R3] has more grace.

5.2. What about (...)?

The declarator (...) should remain valid. It is compatible with C and unambiguous.

5.3. parameter-declaration-clause

In the discussion of this proposal, some users have expressed discontent with the current production for parameter-declaration-clause. It would be possible to isolate the deprecated case with the following editorial change:

parameter-declaration-clause:
        parameter-declaration-list_opt
        parameter-declaration-list ...
        parameter-declaration-list , ...
        ...

Of these productions, the second (highlighted) is deprecated. If the variadic comma was made mandatory, this rule would simply be removed and the rest left intact.

I believe that this editorial change is beneficial to the proposal and improves readability of the parameter-declaration-clause production in general.

6. Proposed wording

The proposed wording is relative to [N4950].

6.1. Editorial changes

In subclause 9.3.4.6 [dcl.fct], modify paragraph 3 as follows:

parameter-declaration-clause:
        parameter-declaration-list_opt ..._opt
        parameter-declaration-list_opt
        parameter-declaration-list ...
        parameter-declaration-list , ...
        ...

Update Annex A, subclause 7 [gram.dcl] accordingly.

6.2. Normative changes

In subclause 9.3.4.6 [dcl.fct], add a paragraph:

A parameter-declaration-clause of the form parameter-declaration-list ... is deprecated.

In Annex D [depr], add a subclause:

D.X Non-comma-separated ellipsis parameters [depr.ellipsis.comma]

A parameter-declaration-clause of the form parameter-declaration-list ... is deprecated.
[Example:

void a(...);                // OK
void b(auto...);            // OK
void c(auto, ...);          // OK

void d_depr(auto......);    // deprecated
void d_okay(auto..., ...);  // OK

void e_depr(auto x...);     // deprecated
void d_okay(auto x, ...);   // OK

void f_depr(auto...x...);   // deprecated
void f_okay(auto...x, ...); // OK

void g_depr(int...);        // deprecated
void g_okay(int, ...);      // OK

void h_depr(int x...);      // deprecated
void h_okay(int x, ...);    // OK

— end example]

7. Acknowledgements

I thank Christof Meerwald for contributing proposed wording.

I thank Arthur O’Dwyer for providing extensive editorial feedback, including the suggestion to update parameter-declaration-clause.