Compilation failure for no_std & alloc feature

[michael-p]

Dear all,

It looks like the rand crate does not compile for no_std targets when the alloc feature is enabled:

$ cargo build --no-default-features --features "alloc"
   Compiling rand_core v0.6.1 (/Users/michaelp/Desktop/rand/rand_core)
   Compiling rand v0.8.2 (/Users/michaelp/Desktop/rand)
error[E0599]: no method named `powf` found for type `f64` in the current scope
   --> src/seq/index.rs:393:40
    |
393 |             let key = rng.gen::<f64>().powf(1.0 / weight);
    |                                        ^^^^ method not found in `f64`

The reason is that f64::powf() and friends are not available in a no_std context.
The two options I see are to either make the functions sample_weighted and sample_efraimidis_spirakis available only when std feature is enabled, or alternatively use the pow function from the libm crate (this is what most other no_std crates like nalgebra use for their math needs).

Thank you!

[dhardy]

You're right. That should have been covered by the tests.

I'll open a PR to fix the tests. As to fixing the code, probably best to disable these features? rand_distr depends on num-traits (and thus libm), but rand always tries to minimise dependencies.

[jiayihu]

I'm having the same issue but with rand_distr, but my understanding is that since it depends on num-traits it's intended to forcely require these f64 methods to work.

Example of error:

error[E0599]: no method named `powi` found for type `f64` in the current scope
   --> /Users/jiayihu/.cargo/registry/src/github.com-1ecc6299db9ec823/rand_distr-0.4.0/src/binomial.rs:115:27
    |
115 |             let mut r = q.powi(self.n as i32);
    |                           ^^^^ method not found in `f64`
    |
    = help: items from traits can only be used if the trait is in scope
    = note: the following traits are implemented but not in scope; perhaps add a `use` for one of them:
            candidate #1: `use num_traits::float::FloatCore;`
            candidate #2: `use num_traits::Float;`
            candidate #3: `use num_traits::real::Real;`

[vks]

@jiayihu Indeed, rand_distr only works on no_std with num_traits and libm. (I just realized that the readme still claims that no_std is not supported for rand_distr, we should fix that.)

The reason for this is that no_std only supports elementary floating point operations (addition, multiplication etc.), but no exponentials or powers and similar, which require a C standard library in rustc's implementation.

[vks]

I just realized that the readme still claims that no_std is not supported for rand_distr, we should fix that.

I opened #1100 to address this.

[jiayihu]

but no exponentials or powers and similar, which require a C standard library in rustc's implementation

Yes that happens often with embedded in my experience. If I'm correct, you're supposed to provide these implementations. Just as an example, I'm currently having these two functions implemented myself so that the linker can find them when needed:

#[no_mangle]
pub extern "C" fn fminf(a: f32, b: f32) -> f32 {
    if a < b {
        a
    } else {
        b
    }
}

#[no_mangle]
pub extern "C" fn fmaxf(a: f32, b: f32) -> f32 {
    if a > b {
        a
    } else {
        b
    }
}

So in the case of rand_distr, you are supposed to provide the impl of the f64 yourself. Ideally these are provided by libm, but it's still quite incomplete. Also I believe these are called Rust core intrinsics, correct me otherwise, in case anyone needs good keywords while searching Google :)