C1x: Bit-fields (WG14 N1260)

Joseph Myers

Some of my comments here are based on those posted in reflector message 11100.

DR#315 discusses issues concerning the type of a bit-field. For C99 the conclusion for the DR was to leave most of these matters implementation-defined, with the possibility of revisiting the matter in C1x.

The fundamental question is: in the width of a bit-field part of its type, or is the type that with which it was declared (ignoring the width, and with the possibility that the absence of signed is interpreted as unsigned)?

For C90, this was considered in DRs #015, #120 and #122. The conclusion was that bit-fields have special types with the limited numbers of bits, that unsigned bit-fields promote accordingly (015, 122) and that when values are stored in bit-fields conversions are applied according to the special type (120).

In C90, only int, signed int and unsigned int bit-fields needed to be considered (and so bit-fields wider than int were irrelevant). C99 adds _Bool bit-fields, and the possibility of other implementation-defined types (of course, an implementation does not need to provide such types, and any implementation burden associated with bit-fields wider than int falls only on those implementations that choose to do so).

For C++, other integral and enumeration types are allowed for bit-fields; it is explicit that bit-fields do not have special types ([class.bit], "The bit-field attribute is not part of the type of the class member.").

The following are affected by the choice between having special types and no special types:

• Promotions of unsigned bit-fields narrower than int. It is however accepted for both C99 (in special wording suggested in DR#315, covering at least the case where the declared type is unsigned int) and C++ (special wording in [conv.prom]) that such bit-fields promote to int when it can represent all values of the bit-field.
• Promotions of signed bit-fields narrower than int but with a wider declared type such as long. In C++ these also promote to int.
• How integer values are stored in the bit-field if they are out of range for the declared type. If bit-fields have special types, the normal rules for conversions apply to a conversion to such a special type. If not, DR#120 notes that no semantics for such assignment and initialization are defined, and special wording would be needed to define it. (I do not know if C++ has such wording, but haven't found it if so.) Without special types, what says that storing 2 in an unsigned:1 bit-field (fully defined in the presence of special types) yields 0 (taking the least significant bit) rather than 1 (taking the next bit)?
• How floating-point values are stored in the bit-field if out of range for the bit-field (so converting to a special type yields undefined behavior, or an unspecified value if Annex F is implemented) but within range for the declared type (so in the absence of special types the conversion to the declared type is defined and then the previous issue of storing out-of-range integer values applies). Depending on how integer values are stored in the absence of special types, this might mean (supposing 16-bit unsigned short and 32-bit unsigned int) that storing a float in an unsigned int:16 bit-field must be implemented differently, and possibly less efficiently, than storing it in an unsigned short, because some cases would only be defined for the bit-field and not for the direct conversion to 16 bits.
• The presence of extended integer types. Apparently, if bit-fields have special types they must be extended integer types and so meet the implementation documentation requirements for such (even though bit-fields had special types in C90 and it didn't have the concept of extended integer types).
• The semantics of arithmetic on bit-fields wider than int (as raised in DR#315). This only affect implementations choosing to have such implementation-defined bit-field types.

The C99 standard text leaves the question of special types ambiguous.

C90 6.5.2.1 says:

The expression that specifies the width of a bit-field shall be an integral constant expression that has nonnegative value that shall not exceed the number of bits in an ordinary object of compatible type. If the value is zero, the declaration shall have no declarator.

C99 6.7.2.1, before TC2 (the changes in TC2 being irrelevant) says:

The expression that specifies the width of a bit-field shall be an integer constant expression that has nonnegative value that shall not exceed the number of bits in an object of the type that is specified if the colon and expression are omitted. If the value is zero, the declaration shall have no declarator.

"of compatible type" was changed to "of the type that is specified if the colon and expression are omitted", possibly to make the readings of the C90 DRs clearer from the text of the standard.

Both C90 and C99 also say that "A bit-field is interpreted as [an integral|a signed or unsigned integer] type consisting of the specified number of bits.", both suggesting separate types.

C90 says "A bit-field shall have a type that is a qualified or unqualified version of one of int, unsigned int, or signed int."; C99 says "A bit-field shall have a type that is a qualified or unqualified version of _Bool, signed int, unsigned int, or some other implementation-defined type.", both suggesting the bit-field has the underlying type instead of the new type.