This primitive integer type is half the bit-width of a usize or isize. Its purpose is to have a narrower range of values, appropriate for defining multi-dimensional collections of values.
The classic example is a rectangular matrix (2D list), which can be implemented as a Vec<Vec<T>> or as a Vec<T> with "arithmetic addressing". The maximum width or height of a "2D Vec" must be usize::MAX.isqrt() which could be losslessly casted as usize_half
While it's not perfectly typesafe, you can get a 90% solution with #[cfg(target_ptr_width = "32")] type Idx = u16; #[cfg(target_ptr_width = "64")] type Idx = u32; and as such there's no need for std to provide what's ultimately a very niche usage datatype. std/core is for broadly used vocabulary types with roughly one obvious choice of implementation tradeoffs.
Only if the matrix is square. Otherwise the max for either width or height (but not both at the same time) is isize::MAX. (A "clever" solution would thus be packing both extents into a single integer, plus a few bits to store the width of one extent.)
What about using const generics (array) rather than conditional compilation?
pub struct usize_half([u8; size_of::<usize>() / 2]);
I expect this would be more cross-platform. The only problems I see are: