Add hint to value()

[jordens]

The compiler is typically not smart enough to optimize away bounds checks if the arbitrary-int new()/value() API is not optimized away.

E.g. take this example:

#![no_std]

pub struct U<const N: usize>(u32);

impl<const N: usize> U<N> {
    const MAX: u32 = u32::MAX >> (32 - N);
    
    #[inline]
    pub const fn new(value: u32) -> Self {
        assert!(value <= Self::MAX);
        Self(value)
    }

    #[inline]
    pub const fn value(&self) -> u32 {
        // unsafe { core::hint::assert_unchecked(self.0 < 1 << N); }
        self.0
    }
}

#[inline(never)]
fn f(x: U<2>) -> i32 {
    const A: [i32; 4] = [1,8,9,0];
    A[x.value() as usize]
}

pub fn g(x: u32) -> i32 {
    f(U::new(x))
}

https://rust.godbolt.org/z/cYx6j666W

f() does a full bounds check even though U<2> is known to be an in-bounds index for a len-4 array.

Adding the soundness promise that the return of value() does not exceed MAX (matching the assert in new()) enables the compiler to optimize away the check regardless of the value of the argument.

This is currently behind a feature gate (hint) but IMO it could just be blanket-added since the point of the crate is precisely to make that promise.

[danlehmann]

Thanks a lot! I didn't know about this assert, but this is really great to know.

I'm testing the impact on binary size on a fairly large project with a lot of arbitrary-int usage. I'm sure there are some nice gains to be had with this.

One issue I see though: The most common place to access value() is I think through the const fn value() is line 133, which isn't covered by your patch. I'm trying to add that, but that does require a tiny breach of API contract (it's in a part where no bounds on T exist, so we can't compare). This might require a bump to a 2.0, eventhough I can't imagine anyone would want to use UInt with T not being a primitive int

[danlehmann]

I think we can use the hint cfg check as a sneaky way to get a small api change in. If it is present, const fn value can move to be a peer of const fn try_new(). I can upload that patch on top of yours

[danlehmann]

See #49 for the follow-up patch

[jordens]

Oh yes. I completely missed that. Thanks!