Unsafe checking skips pointer dereferences in unused places

[scottmcm]

Seen in #79735

I tried this code:

fn foo(ptr: *const bool) {
    let _ = *ptr;
}

https://play.rust-lang.org/?version=nightly&mode=debug&edition=2018&gist=44e42a069bc7601365f5f21a102178d9

Which, oddly, happily compiled despite the syntactic pointer dereference.

Back in 1.21, however, that gave the expected error:

error[E0133]: dereference of raw pointer requires unsafe function or block
 --> <source>:2:13
  |
2 |     let _ = *ptr;
  |             ^^^^ dereference of raw pointer

error: aborting due to previous error

https://rust.godbolt.org/z/rP93nf

The conversation in the 2020-12-15 lang team meeting implied that this was an unexpected regression, so opening and nomination for further discussion.

[scottmcm]

Probably P-low as it's sound as-is, and has been this way for over 3 years.

[LeSeulArtichaut]

Why is this sound? Because the dereference gets optimized away?

[camelid]

Why is this sound? Because the dereference gets optimized away?

I think so. If you look at the MIR (in debug mode, so this should only have 'guaranteed' optimizations):

// WARNING: This output format is intended for human consumers only
// and is subject to change without notice. Knock yourself out.
fn foo(_1: *const bool) -> () {
    debug ptr => _1;                     // in scope 0 at src/lib.rs:1:8: 1:11
    let mut _0: ();                      // return place in scope 0 at src/lib.rs:1:26: 1:26
    scope 1 {
    }

    bb0: {
        _0 = const ();                   // scope 0 at src/lib.rs:1:26: 3:2
        return;                          // scope 0 at src/lib.rs:3:2: 3:2
    }
}

You'll see that the dereference never happens. I think this is because the wildcard pattern (_) has special semantics.

[scottmcm]

Because the dereference gets optimized away?

It's because _ is magic. If you instead try this:

fn foo(ptr: *const bool) {
    *ptr;
}

the error is still there even though LLVM also optimizes away that dereference.

[LeSeulArtichaut]

Assigning P-medium as discussed as part of the Prioritization WG procedure.

[spastorino]

This was discussed during today's compiler meeting