Deref coercions and &-to-* coercions don't compose

[carlpaten]

Minimal repro:

use std::ops::Deref;

struct A (u8);

impl Deref for A {
    type Target = u8;
    fn deref(&self) -> &Self::Target {
        &self.0
    }
}

fn main() {
    let a = A(0);

    //compiles
    let b: &u8 = &a;
    let c: *const u8 = b;

    //error
    let d: *const u8 = &a;
}

Compile error:

coerce.rs:15:24: 15:26 error: mismatched types:
 expected `*const u8`,
    found `&A`
(expected u8,
    found struct `A`) [E0308]
coerce.rs:15     let c: *const u8 = &a;
                                    ^~
coerce.rs:15:24: 15:26 help: run `rustc --explain E0308` to see a detailed explanation
error: aborting due to previous error

[carlpaten]

@bluss thanks for walking me through this.

[bluss]

RFC 401 says that coercions are transitive, but I also realize that it does not consider deref coercions at all.

[oli-obk]

While I'm not sure if this should be fixed, we can at least fix the diagnostic to tell to user to use &*a instead of &a.

cc @rust-lang/wg-diagnostics

[estebank]

This should probably be added to

rust/src/librustc_typeck/check/demand.rs

Lines 581 to 587 in 8ad83af

	pub fn check_for_cast(
	&self,
	err: &mut DiagnosticBuilder<'_>,
	expr: &hir::Expr<'_>,
	checked_ty: Ty<'tcx>,
	expected_ty: Ty<'tcx>,
	) -> bool {

[eddyb]

cc @nikomatsakis Should we close this issue based on only a diagnostics improvement?

[nikomatsakis]

A good question. I do sort of think that coercions "should" compose, but I also suspect there are complications, and am somewhat tempted to say it's "working well enough" and leave it closed.