So the tests use the libc function libc::CPU_COUNT, which counts how many bits are flipped to 1 in the cpuset. The source code that calls the syscall uses the struct libc::cpuset_t, and the size of this set struct is used to write the specified bytes to the mask pointer.

The source of this struct is this:

pub struct cpuset_t {
    #[cfg(all(any(freebsd15, freebsd14), target_pointer_width = "64"))]
    __bits: [c_long; 16],
    #[cfg(all(any(freebsd15, freebsd14), target_pointer_width = "32"))]
    __bits: [c_long; 32],
    #[cfg(all(not(any(freebsd15, freebsd14)), target_pointer_width = "64"))]
    __bits: [c_long; 4],
    #[cfg(all(not(any(freebsd15, freebsd14)), target_pointer_width = "32"))]
    __bits: [c_long; 8],
}

So on newer FreeBSD version, this struct is 128 bytes large, otherwise 32 bytes. During testing I have found out that the current compilation results in the 32 bytes variant. This means we can only "use" 256 CPUs since the size of the mask is: 32 * 8 = 256. Miri's maximum number is 1024 and the test also sets and expects this value.

I have no idea how to alter compilation to use the 128 size variant, which would result in 1024 bits being used. But the test passes when using -Zmiri-num-cpus=256.

@rustbot ready