Implementation-level interface for RNGs

This trait encapsulates the low-level functionality common to all generators, and is the "back end", to be implemented by generators. End users should normally use the rand::Rng trait which is automatically implemented for every type implementing RngCore.

Three different methods for generating random data are provided since the optimal implementation of each is dependent on the type of generator. There is no required relationship between the output of each; e.g. many implementations of fill_bytes consume a whole number of u32 or u64 values and drop any remaining unused bytes. The same can happen with the next_u32 and next_u64 methods, implementations may discard some random bits for efficiency.

Implementers should produce bits uniformly. Pathological RNGs (e.g. always returning the same value, or never setting certain bits) can break rejection sampling used by random distributions, and also break other RNGs when seeding them via SeedableRng::from_rng.

Algorithmic generators implementing SeedableRng should normally have portable, reproducible output, i.e. fix Endianness when converting values to avoid platform differences, and avoid making any changes which affect output (except by communicating that the release has breaking changes).

Typically an RNG will implement only one of the methods available in this trait directly, then use the helper functions from the impls module to implement the other methods.

Note that implementors of RngCore also automatically implement the TryRngCore trait with the Error associated type being equal to Infallible.

It is recommended that implementations also implement:

• Debug with a custom implementation which does not print any internal state (at least, CryptoRngs should not risk leaking state through Debug).
• Serialize and Deserialize (from Serde), preferably making Serde support optional at the crate level in PRNG libs.
• Clone, if possible.
• never implement Copy (accidental copies may cause repeated values).
• do not implement Default for pseudorandom generators, but instead implement SeedableRng, to guide users towards proper seeding. External / hardware RNGs can choose to implement Default.
• Eq and PartialEq could be implemented, but are probably not useful.

Example

A simple example, obviously not generating very random output:

#![allow(dead_code)]
use rand_core::{RngCore, impls};

struct CountingRng(u64);

impl RngCore for CountingRng {
    fn next_u32(&mut self) -> u32 {
        self.next_u64() as u32
    }

    fn next_u64(&mut self) -> u64 {
        self.0 += 1;
        self.0
    }

    fn fill_bytes(&mut self, dst: &mut [u8]) {
        impls::fill_bytes_via_next(self, dst)
    }
}