generic float inference type is unified with f64 prematurely

[pnkfelix]

I don't exactly know what terminology to use to describe this bug; feel free to update the title to reflect reality...

This was something I encountered while using bjz's cgmath library ; this test is derived from that library (but independent of it).

pub struct Matrix4<S>;
pub trait POrd<S> {}

pub fn translate<S: POrd<S>>(_: S) -> Matrix4<S> { unimplemented!() }

impl POrd<f32> for f32 {}
impl POrd<f64> for f64 {}

fn main() {
    // This is a generic float, not constrained to f64.
    let x = 1.0;
    // (And it might occur arbitrarily far from its use below.)

    // error: mismatched types: expected `Matrix4<f32>` but found `Matrix4<f64>` (expected f32 but found f64)
    let m : Matrix4<f32> = translate(x);
    //                     ^~~~~~~~~~~~
    println!("m: {:?}", m);
}

Transcript of compilation attempt:

% rustc -O -o bug bug.rs
bug.rs:15:28: 15:40 error: mismatched types: expected `Matrix4<f32>` but found `Matrix4<f64>` (expected f32 but found f64)
bug.rs:15     let m : Matrix4<f32> = translate(x);
                                     ^~~~~~~~~~~~
error: aborting due to previous error

1. Ideally we would actually manage to properly assign f32 to the type of x, since I do not see any reason for the generic float type to be constrained to f64. (Note that if you add a : f32 type annotation to the let x then compilation terminates successfully.)
2. But even if we cannot accomplish that goal for some hypothetical reason, we still need to a better job with the error message here.

Note also that the impl POrd<f64> for f64 {} is necessary to reproduce the poor-quality error message given here. If you remove it, you get this instead: error: failed to find an implementation of trait POrd<<generic float #0>> for <generic float #0> (which also is not great, but at least then I have a hint that my problem originates with some use of a generic float literal somewhere in my code.

[steveklabnik]

Inference is totally different now, does this issue still persist, @pnkfelix ?

[apasel422]

The following compiles successfully now:

#[derive(Debug)]
pub struct Matrix4<S>(S);
pub trait POrd<S> {}

pub fn translate<S: POrd<S>>(_: S) -> Matrix4<S> { unimplemented!() }

impl POrd<f32> for f32 {}
impl POrd<f64> for f64 {}

fn main() {
    let x = 1.0;
    let m : Matrix4<f32> = translate(x);
    println!("m: {:?}", m);
}