add docs

src/conv.rs

@@ -74,10 +74,12 @@ impl F8 {
 		Self::merge(m as u8 & M_STORAGE_MAX, e as u8)
 	}
 
+	/// Reinterpret `bits` as an [`F8`]. Equivalent to [`std::mem::transmute`] but safe.
 	pub const fn from_bits(bits: u8) -> Self {
 		Self(bits)
 	}
 
+	/// Reinterpret `self` as an [`u8`]. Equivalent to [`std::mem::transmute`] but safe.
 	pub const fn to_bits(self) -> u8 {
 		self.0
 	}
@@ -89,6 +91,12 @@ impl F8 {
 
 	/// Split `self` into `(base, scale)` such that `self = base * 2.pow(scale)`.
 	///
+	/// Guarantees:
+	/// * `ldexp(val.frexp()) == val`.
+	/// * `base` will have at most [M_BITS]+1 low-order bits set.
+	/// * `scale.abs()` will have at most [E_BITS]+1 low-order bits set.
+	///
+	/// # Example
 	/// ```
 	/// # use f8::F8;
 	/// # let val = F8::from_bits(42);

src/lib.rs

@@ -1,14 +1,24 @@
+//! Software 8-bit floating-point math library. Not for production use.
+
 mod conv;
 mod fmt;
 mod ops;
 
+/// Mantissa width of [`F8`].
 pub const M_BITS: u8 = 5;
+
+/// Exponent width of [`F8`].
 pub const E_BITS: u8 = 3;
+
+/// The value such that `2.pow(E_CAP)` is just over the [`F8`] limit.
 const E_CAP: u8 = 4;
+
+/// Largest exponent value of [`F8`].
 pub const E_MAX: u8 = E_CAP - 1;
 
 static_assertions::const_assert_eq!(M_BITS + E_BITS, 8);
 
+/// The largest integer up to and including which all integers are representable exactly.
 pub const EXACT_INT_MAX: u8 = if E_MAX > M_BITS {
 	2 << M_BITS
 } else {
@@ -21,6 +31,7 @@ const E_BIAS: u8 = E_STORAGE_MAX - E_MAX;
 const M_MASK: u8 = M_STORAGE_MAX;
 const E_MASK: u8 = E_STORAGE_MAX << M_BITS;
 
+/// 8-bit unsigned binary floating-point type, with [`M_BITS`] mantissa bits and [`E_BITS`] exponent bits.
 #[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
 #[repr(transparent)]
 pub struct F8(u8);