source: doc/proposals/flags.md @ 18f7858

## Flag Enums ##

A common programming problem is to represent a value from a set of boolean flags, each of which can be either on or off. C already has enums and bitfields, which can be naturally used to represent the individual flags, but are un-ergonomic to combine together. This proposal introduces "flag enums", a variant of the usual enums specialized to represent flags in a more ergonomic way.

As an example, a flag enum for the TCP control bits could be defined as follows:

        ```
        enum TCP_Flags {
                FIN,
                SYN,
                RST,
                PSH,
                ACK,
                URG,
                ECE,
                CWR,
                NS
        } __attribute__((flag));
        ```

The `__attribute__` syntax is ugly, but represents the smallest backwards compatibility break; a new SUE for enum flags (e.g. `flag enum TCP_Flags { ... };` or even `flag TCP_Flags { ... };`) might also be reasonable.

A flag enum would be different than a normal enum in two ways: it would auto-generate discriminant values differently, and it would have a number of bitwise operators defined on it by default.

Normal enums generate their discriminant values sequentially starting at zero (`0, 1, 2, 3, ...`), while a flag enum would generate its discriminant values as successive powers of two starting at `1`. E.g. the `TCP_Flags` declaration above would codegen to an enum like below:

        ```
        enum TCP_Flags {
                FIN = 0x1,
                SYN = 0x2,
                RST = 0x4,
                PSH = 0x8,
                ACK = 0x10,
                URG = 0x20,
                ECE = 0x40,
                CWR = 0x80,
                NS = 0x100
        };
        ```

The precise rule used would be that if no enum discriminant is given, the discriminant is the smallest power of two larger than the previous discriminant (`1` if there is no previous discriminant). This would allow some flexibility for cases like these:

        ```
        enum FunFlags {
                NONE = 0,           // Named empty value
                FOO,  // == 0x1
                BAZ = 0x6,          // Multi-bit flag: 0x4 | 0x2
                BAR,  // == 0x8
                FOOBAR = FOO | BAR  // Named combination flag
        } __attribute__((flag));
        ```

Secondly, we would auto-generate a number of useful operators for any flag enum, as follows:
* The default constructor for any flag enum would be defined, and would produce a flag with an underlying value of 0.
* Assignment from and equality/inequality to `zero_t` should also be defined based on the underlying enum value.
* The bitwise operators `?&?, ?|?, ?^?, ~?` and their assignment variants `?&=?, ?|=?, ?^=?` shall be defined with the semantics of the underlying enum value; `?-?` and `?-=?` should also be defined such that `a - b == a & ~b` (a set difference operation).

With these operations defined, flag enums would support a full set of useful flag operations, using existing, known syntax, as follows:

        ```
        FunFlags f = some_val();
        if ( f ) { sout | "f has some flag(s) set"; }
        if ( f & FOO ) { sout | "f has FOO set"; }
        f |= FOO; // set FOO
        f -= FOO; // unset FOO
        f ^= FOO; // toggle FOO
        ```

In each of the cases above, `FOO` could be replaced by `(BAR | BAZ)` to do the same operation or test on multiple flags at once.

### Alternative/Additional Features ###

#### User-defined enum discriminant iterator ####
It may be useful to provide a more general method for changing the enum discriminant assignment function, e.g. the flag enum discriminants could be defined by something like the following:

        ```
        enum(@ << 1) TCP_Flags {  // each discriminant is left-shifted by 1 from the previous
                FIN = 0x1,  // first flag is 1
                SYN,
                ACK,
                ...
        }
        ```

#### Member expression for enums ####
As a more ergonomic way to set and unset enum flags, we could define a member expression for flags enums. Since only unions and structs can have member expressions now, this change would be backwards compatible. Basically, given a `FunFlags f`, `f.FOO` would return a proxy object which could be implicitly converted to `bool` (with semantics `f & FOO`, i.e. "check if `FOO` is set on `f`"), as well as having `bool` assigned to it (with semantics `f |= FOO` on true or `f -= FOO` on false, i.e. "set or unset `FOO` on `f` as appropriate"). With this member function, the operations above can be expressed as follows (possibly more ergonomically):

        ```
        FunFlags f = some_val();
        if ( f.FOO ) { sout | "f has FOO set"; }
        f.FOO = true;    // set FOO
        f.FOO = false;   // unset FOO
        f.FOO = ! f.FOO; // toggle FOO
        ```

### Related Work ###
C# has the [`[Flags]`][1] enum attribute, but their proposal does not go as far; specifically, the flag discriminants must be manually specified, and they do not automatically implement the bitwise operators on the flags. 

Java has [`EnumSet`][2] which represents the set of flags for a given enum (C++ [`bitset`][3] can be used similarly). The main disadvantage of applying this approach to Cforall is that C enum types already implicitly convert to int, and the bitwise operators already have interpretations on enums with `int` results based on this conversion. As such, all flags need to be wrapped in a set to be used type-safely with the bitwise operators.

[1]: https://msdn.microsoft.com/en-us/library/system.enum.hasflag(v=vs.110).aspx
[2]: http://docs.oracle.com/javase/7/docs/api/java/util/EnumSet.html
[3]: http://en.cppreference.com/w/cpp/utility/bitset