Date: 2022-06-17

Thomas Köppe <tkoeppe@google.com>

ISO/IEC JTC1 SC22 WG14 N2994

To: WG14, WG21 liaison (SG22)

Comma omission and comma deletion

Contents

  1. Revision history
  2. Summary
  3. The problem
  4. Defining the goal
  5. Alternative designs
  6. Discussion and proposal
  7. Impact
  8. Implementation experience
  9. Proposed wording
  10. Compatibility with C++
  11. Acknowledgements

Revision history

Note: The essential content of this paper has already been applied to C++ (as part of C++20). After an initial presentation of this paper to WG14, the earlier version N2034 has been added to SD3 as a potential future extension for C. The wording of this paper is taken from the final wording in C++ (comprising P0306R4 and P1042R1), adapted for the C standard.

Summary

This is a proposal to make variadic macros easier to use with no arguments by adding a new special functional macro __VA_OPT__.

The problem

Function-style macros that can have variable arguments suffer from a number of ill-specified corner cases. Consider the following macro definitions:

#define F(X, ...) f(10, X, __VA_ARGS__) #define G(X)      f(10, X) #define H(...)    f(10, __VA_ARGS__)

Invocations of these macros are surprising:

InvocationEffectNotes
F(a, b, c)f(10, a, b, c)variable arguments are “b, c
F(a, )f(10, a, )variable arguments contain zero tokens
syntax error
F(a)ill-formedviolates 16.3p12 (no variable arguments)
G(a)f(10, a)
H(a, b, c)f(10, a, b, c)variable arguments are “a, b, c
H(a)f(10, a)variable arguments are “a
H()f(10, )variable arguments are “”
syntax error

There are two problems:

  1. When the macro definition ends in ...), the invocation must contain at least as many commas as the macro has mandatory parameters. This makes the invocation F(a) invalid.
  2. In the case of zero mandatory parameters, it is not possible to distinguish between zero arguments in the invocation and one (variable) argument that is empty. The root of this problem is the unfortunate human convention by which a list of n elements is presented with n − 1 infix separators. This convention does not degenerate to the case n = 0, and empty lists are always a special case.

However, it is quite natural for a macro invocation with variable arguments to degenerate to the case where there are no arguments. In the example, we would like F(a) to be replaced with f(10, a). A more realistic example is a custom diagnostic facility such as the following:

#define ERROR(msg, ...) std::printf("[" __FILE__ ":%d] " msg, __LINE__, __VA_ARGS__) ERROR("%d errors.\n", 0);  // OK, std::printf("[" "file.cpp" ":%d] " "%d errors.\n", 7, 0); ERROR("No errors.\n");     // Error

The complication arises when we consider H(). We may perhaps wish it to be replaced with f(10). However, we may also wish to have a macro such as

#define ADD_COMMA(...) , __VA_ARGS__

which always produces a comma, even when invoked with no arguments. The difference is that we consider H() to have zero arguments, whereas we consider ADD_COMMA() to have one, empty argument.

Defining the goals

We would like to make the preprocessor more expressive to allow users to write macros for all of the situations described above. This requires two distinct changes, one simple and the other complex.

Goal 1. Allow the omission of the comma before the variable arguments in the invocation (i.e. allow F(a) rather than requiring F(a, )).

Goal 2. Provide a mechanism to express a replacement text that contains the variable arguments but which contains a separating comma only if the variable arguments are non-empty (i.e. allow both f(10, a) and f(10, a, b) as possible replacements of F). At the same time, continue to provide a mechanism that unconditionally contains comma before the (possibly empty) variable arguments, like ADD_COMMA above.

This behaviour of Goal 1 is already supported by many popular compilers as a non-conforming extension. It is a non-breaking change, since the current syntax F(a) is ill-formed. Goal 2 is much harder to solve, since there is no single simple enhancement of the existing semantics that satisfies all possible use cases.

We will step through a series of possible solutions (inspired by existing vendor extensions) and analyse their shortcomings, before presenting the proposed solution.

Alternative designs

1. Delete any commas if there are no variable arguments

This approach does not add any new syntax. It merely solves Goal 1 above by allowing a variadic macro invocation to not contain any variable arguments. However, under this approach, the absence of variable arguments is taken as a request to delete an existing comma immediately preceding the __VA_ARGS__ token:

F(a, )   =>   f(10, a, )   // has variable argument, replaced as-is F(a)     =>   f(10, a)     // no variable arguments, final comma deleted H()      =>   f(10, )      // has (empty) variable argument, replaced as-is

This is a minimal, unsurprising extension. However, it suffers from the major draw-back that it offers no mechanism to delete a trailing comma from a variadic macro with zero mandatory parameters.

A variant of this extension is currently provided by MSVC++ and Embarcadero compilers, which always delete the comma, even in the case of zero mandatory arguments. Another possible extension is to provide those semantics under a new name (e.g. __VA_ARGS_FOO__).

2. Hijack existing syntax to opt in to comma deletion

This approach also allows the omission of the variable arguments, and in addition it reuses the concatenation operator ## to control comma deletion explicitly:

#define F1(X, ...) f(10, X, __VA_ARGS__) #define F2(X, ...) f(10, X, ## __VA_ARGS__) #define H1(...) f(10, __VA_ARGS__) #define H2(...) f(10, ## __VA_ARGS__) F1(a, b)   =>   f(10, a, b)   // standard F1(a, )    =>   f(10, a, )    // standard (empty variable arguments) F1(a)      =>   f(10, a, )    // extension (no variable arguments), final comma unaffected F2(a, b)   =>   f(10, a, b)   // variable arguments not empty F2(a, )    =>   f(10, a, )    // variable arguments present (though empty), final comma unaffected F2(a)      =>   f(10, a)      // no variable arguments, final comma deleted H1(a)      =>   f(10, a)      // standard H1()       =>   f(10, )       // standard (empty variable arguments) H2(a)      =>   f(10, a)      // variable arguments not empty H2()       =>   f(10)         // variable arguments present (though empty), final comma deleted

This extension is somewhat difficult to explain, but it generally Does What You Want. The complete omission of variable arguments is required for comma deletion (compare F2(a, ) and F2(a)), though omission of the variable arguments alone is not enough to delete the comma (compare F1(a, ) and F1(a)), but the case of zero mandatory parameters is special, and in that case it is mere absence of tokens from the variable arguments that enables the comma deletion when the ## operator is used.

The downside of this extension is three-fold: 1) Parsing this syntax requires look-ahead, adding cost to the translation. 2) The extension reuses an unrelated piece of syntax, muddling the language. 3) The extension hides its dependency on the presence or absence of the variable arguments and whether the variable arguments contain tokens in subtle and non-explicit ways.

3. “Named pack” style

A rather more different approach abandons the use of C99’s __VA_ARGS__ token in favour of something like #define F(X, Args...) or #define F(X, ...Args). GCC has long provided the former (where the replacement text would use Args for the variable arguments, and , ##Args (with mandatory whitespace after the comma!) requests comma deletion). The template-pack-like syntax ...Args does not appear to be used by any preprocessor and may provide a less obstructed extension route (e.g. one could say that x, y, ...Args always has comma deletion semantics).

However, all these approaches seem undesirable. First off, they are a departure, and perhaps even a regression, from the direction taken by C99 and its __VA_ARGS__ token. Second, this design would only satisfy those needs that require comma deletion, leaving use cases like the above ADD_COMMA to use the existing syntax. Thus there would be two parallel but dissimilar constructions living side by side, which seems inelegant and wasteful.

4. Omit comma from the macro definition

Note: This idea came up during the discussion of N2034 in WG14.

We could use syntax like #define F(X ...) f(10, __VA_ARGS__) to request comma deletion (note the absence of a comma before the ellipsis in the definition). This approach does not degenerate to the case of macros with no named parameters, though. Moreover, WG14 felt that this was too clever and too subtle, whereas the proposed solution below is highly visible and explicit. Also, an unrelated difference between C++ and C is that C++ allows omitting the final comma from the parameter list of a variable function declaration. While this has nothing to do with the preprocessor, the semantics of the optional comma of that feature are the opposite of this present consideration, which is unnecessarily confusing.

5. A new token

All of the considered extensions so far have in common that they end up creating a parallel set of constructions which are identical to the existing macro facilities except when the macro is invoked with no variable arguments, and they all provide some automatic mechanism to determine when to delete a comma. However, none of them are quite explicit about what they are doing.

For the next idea, we consider adding a new token. Let us call it __VA_ARGS_OPT__, with the semantics that wherever it appears in the replacement text, it is replaced with the variable arguments (just like __VA_ARGS__), but additionally, whenever the variable arguments do contain tokens, a comma is prepended:

#define F(X, ...) f(10, X __VA_ARGS_OPT__) F(a, b)  =>   f(10, a, b)  // __VA_ARGS_OPT__ => ', b' F(a, )   =>   f(10, a)     // empty variable arguments, __VA_ARGS_OPT__ => '' F(a)     =>   f(10, a)     // empty variable arguments, __VA_ARGS_OPT__ => ''

In this approach, we have separated Goals 1 and 2 entirely; whether a leading (!) comma is inserted now only depends on whether the variable arguments contain tokens, not on whether they are present at all.

Discussion and proposal

We already said that solutions 1, 2 and 3 are ultimately inelegant, since they create a redundant structure that replicates existing facilities and only differs in subtle details. Solution 4 (a new token) feels cleaner and more orthogonal. In the words of Richard Smith:

“I remain unconvinced that implicitly adding or removing a comma is a good idea. We need the user to tell us which behavior they want.”

We can do a little better than solution 4. Our proposal is to add a new, special kind of functional macro __VA_OPT__. This macro may only be used in the replacement text of a variadic macro:

#define F(X, Y, ...) a b c __VA_OPT__(content)

The semantics are as follows: If the variable arguments contain no tokens, then __VA_OPT__(content) is replaced by no tokens (more precisely, by a placemarker). Otherwise, it is replaced by content, which can contain any admissible replacement text, including __VA_ARGS__.

The canonical use case of __VA_OPT__ is for an optional separator:

#define LOG(msg, ...) printf(msg __VA_OPT__(,) __VA_ARGS__) LOG("hello world")   // => printf("hello world") LOG("hello world", ) // => printf("hello world") LOG("hello %d", n)   // => printf("hello %d", n)

However, this mechanism allows other constructions, too:

#define SDEF(sname, S, ...) S sname __VA_OPT__(= { __VA_ARGS__ }) SDEF(foo);           // => S foo; SDEF(bar, 1, 2, 3);  // => S foo = { 1, 2, 3 };

#define LOG(...) \ printf("at line=%d" __VA_OPT__(": "), __LINE__); \ __VA_OPT__(printf(__VA_ARGS__);) \ printf("\n") LOG(); // => printf("at line=%d", 123); printf("\n"); LOG("All well in zone %n", n); // => printf("at line=%d: ", 123); printf("All well in zone %n", n); printf("\n");

Impact

The proposal is a pure extension of the preprocessor. Syntax that was previously not allowed becomes admissible under the proposed changes.

Implementation experience

The proposed extension to allow omission of the variable arguments has been implemented by many compilers. Faisal Vali implemented a version of the proposed __VA_OPT__ extension in Clang.

Proposed wording

Change paragraph 6.10.3p4 as follows.

If the identifier-list in the macro definition does not end with an ellipsis, the number of arguments (including those arguments consisting of no preprocessing tokens) in an invocation of a function-like macro shall equal the number of parameters in the macro definition. Otherwise, there shall be moreat least as many arguments in the invocation thanas there are parameters in the macro definition (excluding the ...). There shall exist a ) preprocessing token that terminates the invocation.

Change paragraph 6.10.3p5 as follows.

The identifieridentifiers __VA_ARGS__ and __VA_OPT__ shall occur only in the replacement-list of a function-like macro that uses the ellipsis notation in the parameters.

Change paragraph 6.10.3p12 as follows.

If there is a ... in the identifier list in the macro definition, then the trailing arguments (if any), including any separating comma preprocessing tokens, are merged to form a single item: the variable arguments. The number of arguments so combined is such that, following merger, the number of arguments is one more than the number of parameters in the macro definition (excluding the ...), except that if there are as many arguments as named parameters, the macro invocation behaves as if a comma token had been appended to the argument list such that variable arguments are formed that contain no pp-tokens.

Insert a new piece of syntax at the beginning of 6.10.3.1.

Syntax

va-opt-replacement:

          __VA_OPT__ ( pp-tokensopt )

Subsection 6.10.3.1 currently lacks “Constraints” and “Semantics” partitions and is therefore somewhat unclear about the distinction. We propose to add this structure in passing, as well as a description. First, insert a new “Description” part, and a new paragraph, immediately after the above new syntax.

Description

Argument substitution is a process during macro expansion in which identifiers corresponding to the parameters of the macro definition and the special constructs __VA_ARGS__ and __VA_OPT__ are replaced with token sequences from the arguments of the macro invocation and possibly of the argument of the feature __VA_OPT__. The latter process allows to control a substitute token sequence that is only expanded if the argument list that corresponds to a trailing ... of the parameter list is present and has a non-empty substitution.

Next, insert a new “Constraints” part, and two new paragraphs. Please see below for an optional edit suggestion.

Constraints

The identifier __VA_OPT__ shall always occur as part of the preprocessing token sequence va-opt-replacement; its closing ) is determined by skipping intervening pairs of matching left and right parentheses in its pp-tokens. The pp-tokens of a va-opt-replacement shall not contain __VA_OPT__. The pp-tokens shall form a valid replacement list for the current function-like macro.

A va-opt-replacement is treated as if it were a parameter.

Next, add a new part “Semantics” above the existing wording, and modify as follows. Change paragraph 6.10.3.1p1 (now p4) as follows.

Semantics

After the arguments for the invocation of a function-like macro have been identified, argument substitution takes place. AFor each parameter in the replacement list, unless that is neither preceded by a # or ## preprocessing token ornor followed by a ## preprocessing token (see below), is replaced by the preprocessing tokens naming the parameter are replaced by a token sequence determined as follows:

Drafting note: The above clarifies that __VA_OPT__ expansions are not rescanned prior to rescanning of the replacement list containing the instance of __VA_OPT__.

Append an example after 6.10.3.1p1 (now p4).

Example:

#define LPAREN() ( #define G(Q) 42 #define F(R, X, ...) __VA_OPT__(G R X) ) int x = F(LPAREN(), 0, <:-);     // replaced by int x = 42;

Change paragraph 6.10.3.1p2 (now p5) as follows.

An identifier __VA_ARGS__ that occurs in the replacement list shall beis treated as if it were a parameter, and the variable arguments shall form the preprocessing tokens used to replace it.

Optional suggestion: Now that we have distinct sections for syntax and semantics, we could split this paragraph, and turn the first part into a new paragraph at the end of the “Constraints” section, “An identifier __VA_ARGS__ that occurs in the replacement list is treated as if it were a parameter.”, and change the remaining p4 into ‘The identifier __VA_ARGS__ is replaced by the tokens of the variable arguments.

Then append new paragraphs to subsection 6.10.3.1 as follows.

The preprocessing token sequence for the corresponding argument of a va-opt-replacement is defined as follows. If a (hypothetical) substitution of __VA_ARGS__ as neither an operand of # nor ## consists of no preprocessing tokens, the argument consists of a single placemarker preprocessing token (6.10.3.3, 6.10.3.4). Otherwise, the argument consists of the results of the expansion of the contained pp-tokens as the replacement list of the current function-like macro before removal of placemarker tokens, rescanning, and further replacement.

Note: The placemarker tokens are removed before stringization (6.10.3.2), and can be removed by rescanning and further replacement (6.10.3.4).

Example:

#define F(...) f(0 __VA_OPT__(,) __VA_ARGS__) #define G(X, ...) f(0, X __VA_OPT__(,) __VA_ARGS__) #define SDEF(sname, ...) S sname __VA_OPT__(= { __VA_ARGS__ }) #define EMP F(a, b, c) // replaced by f(0, a, b, c) F() // replaced by f(0) F(EMP) // replaced by f(0) G(a, b, c) // replaced by f(0, a, b, c) G(a, ) // replaced by f(0, a) G(a) // replaced by f(0, a) SDEF(foo); // replaced by S foo; SDEF(bar, 1, 2); // replaced by S bar = { 1, 2 }; #define H1(X, ...) X __VA_OPT__(##) __VA_ARGS__ // error: ## may not appear at the beginning of a replacement list (6.10.3.3) #define H2(X, Y, ...) __VA_OPT__(X ## Y,) __VA_ARGS__ H2(a, b, c, d) // replaced by ab, c, d #define H3(X, ...) #__VA_OPT__(X##X X##X) H3(, 0) // replaced by "" #define H4(X, ...) __VA_OPT__(a X ## X) ## b H4(, 1) // replaced by a b #define H5A(...) __VA_OPT__()/**/__VA_OPT__() #define H5B(X) a ## X ## b #define H5C(X) H5B(X) H5C(H5A()) // replaced by ab

Change paragraph 6.10.3.2p2 as follows.

If, in the replacement list, a parameter is immediately preceded by a # preprocessing token, both are replaced by a single character string literal preprocessing token that contains the spelling of the preprocessing token sequence for the corresponding argument (excluding placemarker tokens). Let the stringizing argument be the preprocessing token sequence for the corresponding argument with placemarker tokens removed. Each occurrence of white space between the stringizing argument’s preprocessing tokens becomes a single space character in the character string literal. White space before the first preprocessing token and after the last preprocessing token comprising the stringizing argument is deleted. Otherwise, the original spelling of each preprocessing token in the stringizing argument is retained in the character string literal, except for special handling for producing the spelling of string literals and character literals: a \ character is inserted before each " and \ character of a character literal or string literal (including the delimiting " characters), except that it is implementation-defined whether a \ character is inserted before the \ character beginning a universal character name. If the replacement that results is not a valid character string literal, the behavior is undefined. The character string literal corresponding to an empty stringizing argument is "". […]

Compatibility with C++

The entire proposal (rationale, implementation experience and wording) applies almost verbatim to the C++ language as well. Indeed, this paper originated from a C++ proposal, which was eventually adopted (as WG21 P0306R4 and amended by WG21 P1042R1) (following feedback from WG14) and subsequent editorial improvements.

An earlier version of this proposal (N2034) was presented to WG14 at the 2016 London meeting and received favourably, resulting in an entry in SD3 to solve the same problem in a future revision of the C language. We would like to ask the WG14 liaison to present this updated revision to WG14 for inclusion in the next revision of C.

Acknowledgements

Many thanks to Dawn Perchik, David Krauss, Hubert S. Tong and Richard Smith for valuable discussion, guidance, suggestions, examples and review, to Aaron Ballman and Jens Gustedt for significant help in adapting the wording for WG14, and to Faisal Vali for implementing the feature and clarifying several important details! Thanks also go to the members of WG14 for their hospitality and a very productive discussion.