Semantics of char comparison operations is not clear

[osa1]

This is possibly me being silly.. I'm currently working on something where I have to convert chars to u32s for storage, and then convert back. My code is only correct if ∀ (c1 : char) (c2 : char) . c1.cmp(c2) == (c1 as u32).cmp(c2 as u32). In other words conversion from char to u32 must preserve the ordering relation of values. (there is probably a name given to this notion of relation-preserving mapping, but I don't know what it is.. anyone know?)

Looking at char documentation, it says char represents "Unicode scalar value"s, but other than that it doesn't say what does conversion to u32 do, or whether they're represented as u32 internally, and whether comparison is simple u32 comparison (i.e. comparison of unicode scalar values) or something more complicated. It's also not easy to learn about this by looking at the source, becuase PartialOrd implementation of char relies on >, < etc. implementations (1, 2), which are I think hard-coded somewhere in the compiler.

Simple u32 comparison is probably the only sensible implementation, but that's not entirely clear to me. When I click on the "unicode scalar value" link in the documentation it doesn't say how the values must be compared. Without knowing more about unicode or "unicode scalar value" I'm not sure if e.g. a character can have multiple unicode scalar values but those need to be equal, or something like that.

[oli-obk]

I was reasonably sure it is just a u32 comparison. But, let's check:

I was looking at the mir interpreter so see what it does, and it also just forwards to the host char comparison 🙃

So I looked at the codegen backend, and that codegens to icmp, irrespective of the type (unless it's float), so we can reasonably say that any builtin comparison that is not float comparison is an integral comparison of the bits of the value represented as its same-sized integer.

I guess we can simplify miri to just do that, too? cc @rust-lang/wg-const-eval

[RalfJung]

Fair question. :)
Yeah Miri could use u32 operations on the host (I also strongly assume that that's what we want), though I don't think that will simplify much: it should still go through to_char() to ensure we raise UB for values that are not valid chars.

[scottmcm]

@osa1 I think you're correct, and it's the same as a comparison on the u32.

Would you be interested in making a PR to updates the docs to that effect? Perhaps there could be a doc comment on impl Ord for char...