matklad's dream fastrand

[matklad]

I love fastrand, it's the closest to the ideal prng crate for me, but it is annoyingly not quite there. As I've been thinking about this today, I decided to write my thoughts down. This is very explicitly not a proposal for a change -- I think there's just a difference in values between what fastrand is, and what I want from my idiosyncratic crate. Still, might be interesting for someone. Or maybe one day I'll fork more_opinionated_fastrand or something :)

The overall theme is "remove magic, it's ok if the user needs to write some code themselves if that leads to greater clarity".

So, the list of not-suggested changes:

• remove interior mutability, the user can wrap in the Cell if they need two.
• bias creation API towards requiring a seed

#[derive(Debug, PartialEq, Eq)]
pub struct Rng { s: u64 }

impl Rng {
    pub fn new(seed: u64) -> Rng {
        Rng { s: seed }
    }
    #[cfg(feature = "std")]
    pub fn from_entropy() -> Rng {
        Rng::new(random_seed())
    }
    pub fn seed(&self) -> u64 {
        self.s
    }
}

/// get me a random without getrandom
#[cfg(feature = "std")]
fn random_seed() -> u64 {
    std::hash::Hasher::finish(&std::hash::BuildHasher::build_hasher(
        &std::collections::hash_map::RandomState::new(),
    ))
}

• remove magical Clone, it should either not exist, or create an identical Copy
• remove Default
• remove thread-local and global API, to bias towards using seed and passing rng explicitly. Again, the user can wrap the thing into a thread local if they need to.
• fn u32(impl RangeBounds) -> fn u32(); fn u32_range(impl RangeBounds). .. just looks weird at the call-site
• fn bool() for a unbiased coint flip, fn bool_ratio(num: u32, denom: u32) for num/denom biased flip