Defined memory layout separate from a C ABI

Help
1

The following code doesn’t work

17:07:47|~/tmp
λ bat main.zig 
pub const S = extern struct {
    x: u256,
};

pub fn f(s: S) void {
    _ = s;
}

pub fn main() void {
    f(.{ .x = 92 });
}


17:08:02|~/tmp
λ zig build-exe main.zig
main.zig:2:8: error: extern structs cannot contain fields of type 'u256'
    x: u256,
       ^~~~
main.zig:2:8: note: only integers with 0, 8, 16, 32, 64 and 128 bits are extern compatible

This makes sense C doesn’t support u256 yet, so there isn’t an ABI which could be used to pass this struct between C and Zig. But what if I don’t care about C ABI per se, and only need a well-defined memory layout (eg, I am describing the data as it exists on the disk / on the network)? Is there a way for me to make a bit-castable Zig struct with an u256 field and a well-defined memory layout, without, at the same time, guaranteeing a particular ABI for cross-language interop?

2 Likes
2

You’re looking for packed.

3

I don’t think packed quite cuts it here: I don’t want my bools to turn into u1s, for example.

2 Likes
4

Why would your bool turn into u1?

const std = @import("std");

pub const S = packed struct {
    _: u256 = 0,
    a: bool,
};

pub fn f(s: S) void {
    _ = s;
}

pub fn main() void {
    const s: S = .{.a = true};
    f(s);
    std.debug.print("{any}\n", .{s});
}

output: example.S{ ._ = 0, .a = true }

The bool is still a bool. If you’re saying that you don’t want the bool to occupy a single bit, just add whatever padding you need.

pub const S = packed struct {
    _: u256 = 0,
    a: bool,
    padding: u7 = undefined,
};

Now for external code, it will look as if the bool is occupying 8 bits.

1 Like
5

But depending on your endianness it could be a 8-bit bool but with the wrong bit set.

6

Yes, when working at such low level, you do need to take into consideration endianness. But the idea was to work with, for instance, a network protocol. The protocol would specify which bit would be the one you’re looking for, so you just define your struct so that the bool matches whatever the protocol specifies.

7

Here comes the trouble. Speaking of network protocols, TCP/IP requires big-endian while most popular archs (x86, x86-64, aarch64) are little-endian. An explicit conversion is needed. Doing it for arbitrary packed struct is no fun.