| Project: | ISO JTC1/SC22/WG21: Programming Language C++ |
|---|---|
| Number: | P1818R1 |
| Date: | 2020-01-12 |
| Audience: | EWGI |
| Revises: | P1818R0 |
| Author: | Lawrence Crowl |
| Contact | Lawrence@Crowl.org |
Generalizing narrowing conversions can admit new primitive numeric types without further updates to the standard. Giving priority to select widening conversions in overload resolution will ease the construction and use of overloaded functions.
Introduction
Problem
Workarounds for Library Users
Explicit Argument Casting
Static Function Declaration
Local Extern Declaration
Workarounds for Library Authors
Add Additional Overloads
Make Lossy Conversions Explicit
Common Type Template
Solution
Proposal
Consequences
Wording
5.11 Keywords [lex.key]
7.3 Standard conversions [conv]
7.3.6 Integral promotions [conv]
7.4 Usual arithmetic conversions [expr.arith.conv]
9.3.4 List-initialization [dcl.init.list]
10.3.4 Constructors [class.ctor]
11.3.3 Best viable function [over.match.best]
11.3.3.1 Implicit conversion sequences [over.best.ics]
15.11 Predefined macro names [cpp.predefined]
Revision History
References
Overload resolution relies on finding the best implicit conversion sequence from each argument type to its corresponding parameter type in all candidate functions. When no one function has all the best sequences, the function call is ambiguous and therefore ill-formed.
While many standard conversions may appear in a conversion sequence, only one user-defined conversion may appear.
The standard distinguishes between information-preserving (widening) conversions and information-destroying (narrowing) conversions in two ways. First, the standard promotions are widening conversions. Second, initializer-list initialization defines some conversions as narrowing conversions and permits them only when the source is constexpr and the value is within the value set of the destination type.
Consider the following declarations.
float atan2( float, float );
double atan2( double, double );
long double atan2( long double, long double );
Of the nine possible call argument-type combinations (32), six are ambiguous. However, all combinations have a best overload, which preserves all the information in the arguments and operates at the least cost.
int main() {
float f; double d; long double ld;
atan2( f, f ); // matches float
atan2( f, d ); // ambiguous, want double
atan2( f, ld ); // ambiguous, want long double
atan2( d, f ); // ambiguous, want double
atan2( d, d ); // matches double
atan2( d, ld ); // ambiguous, want long double
atan2( ld, f ); // ambiguous, want long double
atan2( ld, d ); // ambiguous, want long double
atan2( ld, ld ); // matches long double
}
The problem extends to user-defined types and functions as well.
class cardinal {
unsigned int c;
public:
cardinal();
};
class integral {
int c;
public:
integral();
integral( cardinal );
operator cardinal();
};
class rational {
integral n, d;
public:
rational();
rational( cardinal );
operator cardinal();
rational( integral );
operator integral();
};
cardinal operator+( cardinal, cardinal );
integral operator+( integral, integral );
rational operator+( rational, rational );
int main() {
cardinal c; integral i; rational r;
c + c; // matches cardinal
c + i; // ambiguous, want integral
c + r; // ambiguous, want rational
i + c; // ambiguous, want integral
i + i; // matches integral
i + r; // ambiguous, want rational
r + c; // ambiguous, want rational
r + i; // ambiguous, want rational
r + r; // matches rational
}
A related problem is that adding a new overload into a header may introduce an ambiguity in client code. Such problems may not be found until well after products have shipped.
C++ programs will be easier to write and more robust to environmental changes if the language provides a better mechanism for finding the 'best' overload.
The user of library must necessarily be able to work around any ambiguity.
The typical workaround to ambiguity for library users is to add explicit casts to the call sites.
int main() {
atan2( f, f );
atan2( static_cast<double>(f), d );
atan2( static_cast<long double>(f), ld );
atan2( d, static_cast<double>(f) );
atan2( d, d );
atan2( static_cast<long double>(d), ld );
atan2( ld, static_cast<long double>(f) );
atan2( ld, static_cast<long double>(d) );
atan2( ld, ld );
}
Unfortunately, this approach now binds one particular overload to the call. A more appropriate overload added later will not be found.
One workaround to this problem is to define a local static function with exactly the needed arguments..
static long double atan2( float f, double d )
return atan2( static_cast<double>(f), d );
}
int main() {
float f; double d;
atan2( f, d );
}
Unfortunately, as above, this approach now binds one particular overload to the call. A more appropriate overload added later will not be found.
One workaround to this problem is Daveed Vandervorde's technique that adds a local extern function declaration to force a particular overload.
int main() {
float f; double d;
extern long double atan2( double, double );
return atan2( f, d );
}
This technique is effective, but not well known. It lacks generality in that it does not apply to member functions. More importantly, as more overloads are used within the function, the ambiguity problem resurfaces. Finally, as above, this approach now binds one particular overload to the call. A more appropriate overload added later will not be found.
Library authors can anticipate some problems.
The primary workaround is to add more overloaded functions.
double atan2( float f, double d )
{ return atan2( static_cast<double>(c), i ); }
long double atan2( float f, long double ld )
{ return atan2( static_cast<long double>(c), r ); }
double atan2( double d, float f )
{ return atan2( i, static_cast<double>(c) ); }
long double atan2( double d, long double ld )
{ return atan2( static_cast<long double>(i), r ); }
long double atan2( long double ld, float f )
{ return atan2( r, static_cast<long double>(c) ); }
long double atan2( long double ld, double d )
{ return atan2( r, static_cast<long double>(i) ); }
Unfortunately, number of additional overloads needed grows dramatically with increasing number of types and parameters. This growth places a specification burden on the library author. It also places a burden on the library user, because the number of overloads that must be excluded by a call also grows.
More problematically, independently developed types with the same functions will not have the additional overloads because the authors could not have anticipated the need.
Problematic conversion can be excluded from overloading by making them explicit.
explicit integral( rational );
Unfortunately, this approach requires users to cast arguments even when there would otherwise be no ambiguity. Furthermore, with this approach, the argument is bound to one type. A better function overload introduced later will not be selected because the type of the argument has been frozen into the cast.
A computed explicit, i.e. explicit(expr), does not work
because the function definition does not know the context of the call.
That is, the function definition cannot anticipate competing widening overloads.
Another approach to solving the problem is to write templates that convert arguments to a common type (e.g. [P0880R2]).
Unfortunately, this approach has difficulty with argument-dependent lookup and namespaces.
In the examples above, there is always one least-common information-preserving overload. There are two mechanisms that make identifying this overload possible. First, we prefer widening conversions over narrowing conversions. Second, we prefer the widening conversion that covers the least distance.
In [N3387], Jens Maurer applied these principles to the built-in integer types by adjusting the rules for integer conversion rank and promotion. The paper was not persued. We intend to generalize the approach to user-defined types.
We propose to introduce a distinction between narrowing and widening conversions in C++ programs. We then propose to alter overload resolution rules to prefer widening conversion over narrowing conversions.
For built-in types:
For user-defined types:
widening,
to be used where explicit may appear,
which declares a user-defined conversion to be a widening conversion.widening compile-time computable,
i.e. widening(boolean_expression).widening trait.
Notionally, this would be something like
std::is_widening_convertable< T, R >
The proposal only removes ambiguity, it does not introduce it. So, all existing correct code is unchanged.
The two's complement representation for built-in integer types implies signed→unsigned conversions are not widening. Likewise, unsigned→signed conversions are not widening.
When programmers change existing user-defined conversions to widening conversions, option 1 will not introduce user-defined conversions where none existed before. It is unclear whether option 2 may do so.
The parameter types have widening conversions, adding new overloads reduces the chance that existing calls will become ambiguous.
In the case where an existing call would be ambiguous, overload resolution would become more expensive. This expense must be evaluated with respect to the workarounds.
All wording edits are relative to N4800 Working Draft, Standard for Programming Language C++.
THE WORDING IS NOT YET COMPLETE.
Add widening to Table 5 — Keywords.
Paragraph 2 defines promotion of unsigned short to int, which loses the invariant that the value is not negative.
Conversion from a signed integer type to an unsigned integer type may change the numeric value. Better would be to convert to the smallest signed type that contains all the values of both of the arguments.
Bullet (1.5.3) says:
Otherwise, if the operand that has unsigned integer type has rank greater than or equal to the rank of the type of the other operand, the operand with signed integer type shall be converted to the type of the operand with unsigned integer type.
Bullet (1.5.4) says:
Otherwise, if the type of the operand with signed integer type can represent all of the values of the type of the operand with unsigned integer type, the operand with unsigned integer type shall be converted to the type of the operand with signed integer type.
Bullet (1.5.5) says:
Otherwise, both operands shall be converted to the unsigned integer type corresponding to the type of the operand with signed integer type.
Paragraph 7 defines narrowing conversion.
Add predefined feature test macros to table 17.
This paper revises P1818R0.
bool parameter to widening.