bitsizeof builtin operator

[chqrlie]

To help cram enumerations in bitfields, we could have a bitsizeof() builtin operator that computes the number of bits for an expression or type at compile time:

type TypeKind enum u8 {
    BuiltinType,
    PointerType,
    ArrayType,
    StructType,
    EnumType,
    FunctionType,
    AliasType,
    ModuleType,
}

type T struct { u32 b : 17 }

bitsizeof(TypeKind); // -> 3
bitsizeof(bool); // -> 1
bitsizeof(char); // -> 8
bitsizeof(i32); // -> 32
bitsizeof(T.b); // -> 17

Some corner cases are up for discussion:

bitsizeof(1); // -> 1 or 32 ?
bitsizeof(T); // -> 17 or 32 ?

[bvdberg]

It would avoid bugs where the enum grows, but the bitfield is not updated. I think we can add it as a keyword. It should only accept a type as input (or module.type). T.b is not needed for the bitfield case.

[chqrlie]

T.b is not an actual expression since T is not an instance but a type.

To ensure the consistency of derived flags structures, I guess bitsizeof(T) should evaluate to 17, the number of defined bits. This would make const u32 NumStmtBits = bitsizeof(StmtBits); much easier to maintain, and possibly not needed at all as we could write:

type ExprBits struct {
    u32 : bitsizeof(StmtBits);
    u32 kind : 5;
    u32 is_ctv : 1;
    u32 is_ctc : 1;
    u32 valtype : 2;
    u32 has_effect : 1;
}

The word bitsize was too simple to reserve as a keyword, what do you think of bitsizeof ? Does any other name come to mind?

GNU Pascal has BitSizeOf: https://www.gnu-pascal.de/gpc/BitSizeOf.html
FreePascal has BitSizeOf: https://wiki.freepascal.org/BitSizeOf - https://build.alb42.de/fpcbin/docu/rtl/system/bitsizeof.html
(Pascal identifiers are not case sensitive)

Zig has @bitSizeOf and @bitOffsetOf: https://ziglang.org/documentation/master/#bitSizeOf

[bvdberg]

A struct is always the full size, leftover bits from bitfields dont get merged to the next struct. So in case of T.b, it would be 32 bits.

numbits is indeed a name programmers might want to use. Other names i could think of:

• bitsof (like elemsof/sizeof), not very nice, but therefore unlikely to be used in code.

I also like bitsizeof, that might even be an accurate description of what it does ;)

[chqrlie]

A struct is always the full size, leftover bits from bitfields dont get merged to the next struct. So in case of T.b, it would be 32 bits.

You mean in the case of bitsizeof(T)? If we allow bitsizeof(T.b), it would unambiguously be 17. I see no advantage at making bitsizeof(T) 32 instead of counting the bits actually used in T and returning 17.

This is necessary for this idiom and makes it foolproof:

type ExprBits struct {
    u32 : bitsizeof(StmtBits);
    u32 kind : 5;
    u32 is_ctv : 1;
    u32 is_ctc : 1;
    u32 valtype : 2;
    u32 has_effect : 1;
}

numbits is indeed a name programmers might want to use. Other names i could think of:

• bitsof (like elemsof/sizeof), not very nice, but therefore unlikely to be used in code.

I also like bitsizeof, that might even be an accurate description of what it does ;)

bitsizeof seems consensual and consistent with other languages.

[bvdberg]

yes the bitfields in the AST were the use case i had in mind.

bitsizeof(T.b) is not something useful, so the argument I would like is only a Type or module.Type. Requesting bitsizeof() of (anonymous) Struct members has no real use-case.