Under feature(capture_disjoint_fields), unsafe-checking has false positives under unsafe blocks

[pnkfelix]

Spawned off of issue #85435

I tried this code on nightly:

#![allow(incomplete_features)] // narrow diagnostic output
#![feature(capture_disjoint_fields)] // need this to witness the bug consistently in rustc's git history.

fn main() {
    let val: u8 = 5;
    let u8_ptr: *const u8 = &val;
    let _closure = || {
        unsafe {
            // Fails compilation with:                                                                                                                                                                                            
            // error[E0133]: dereference of raw pointer is unsafe and                                                                                                                                                             
            //               requires unsafe function or block                                                                                                                                                                    
            let tmp = *u8_ptr;
            tmp

            // Just dereferencing and returning directly compiles fine:                                                                                                                                                           
            // *u8_ptr
        }
    };
}

I expected to see this happen: Code accepted by compiler

Instead, this happened:

error[E0133]: dereference of raw pointer is unsafe and requires unsafe function or block
  --> src/main.rs:12:23
   |
12 |             let tmp = *u8_ptr;
   |                       ^^^^^^^ dereference of raw pointer
   |
   = note: raw pointers may be null, dangling or unaligned; they can violate aliasing rules and cause data races: all of these are undefined behavior

error: aborting due to previous error

For more information about this error, try `rustc --explain E0133`.

Note: This has same root cause as #85435. But, we're going to address #85435 with a targeted short-term fix (namely PR #85564), so that it won't hit stable. So I'm filing this bug to track fixing the root cause.

[pnkfelix]

searched nightlies: from nightly-2021-03-04 to nightly-2021-05-21
regressed nightly: nightly-2021-03-17
searched commits: from 107896c to f5d8117
regressed commit: f5d8117

bisected with cargo-bisect-rustc v0.6.0

Host triple: x86_64-unknown-linux-gnu
Reproduce with:

cargo bisect-rustc --access github --preserve

[pnkfelix]

cc @roxelo

[arora-aman]

So the solution to bug would could be to call the restrict capture precision functions in https://github.com/rust-lang/rust/blob/master/compiler/rustc_typeck/src/check/upvar.rs#L1589-L1593

• restrict_capture_precision
• restrict_repr_packed_field_ref_capture

I think this would not cause problems because:

• The reason we needed the fake reads was because we were trying to convert a less precise place into a closure capture.
  Eg: let (a, _) = x and we'd try convert x into a closure capture even though x.0 was being captured. So we added fake reads for the entirety of x.
• In the case of raw ptr dereference eg *(some_place), we are capturing some_place and matching on a more precise *(some_place) in the let. Restricting precision of the fake read shouldn't affect this scenario.

[pnkfelix]

Before we jump into the approach outlined by @arora-aman : we are expecting to move the unsafe-checking from MIR to THIR, according to rust-lang/compiler-team#402 . Won't that shift also resolve this issue?

[arora-aman]

I'm not aware of that, I'll read through that. But I don't think we should be adding fake reads of raw pointers dereference anyway.

As I understand it, the fake reads are less precise closure captures to facilitate pattern matching (in both let and match) and we shouldn't be capturing raw pointer dereference.

[roxelo]

It doesn't seem like rust-lang/compiler-team#402 is going to be done any time soon (and I am not entirely sure if that will fix the bug or not), wouldn't it be better for the time being to attempt @arora-aman's suggestion? It should be a quick fix

[LeSeulArtichaut]

We are expecting to move the unsafe-checking from MIR to THIR, according to rust-lang/compiler-team#402 . Won't that shift also resolve this issue?

Since yesterday (#85555), dereferences of raw pointers is implemented in the THIR unsafety checker (gated under -Z thir-unsafeck), and it correctly does not report an error here.

It doesn't seem like rust-lang/compiler-team#402 is going to be done any time soon

I'm not sure how much time it will take to make the THIR unsafety checker the default once it is fully implemented, but the implementation itself is going rather fast (see rust-lang/project-thir-unsafeck#7).

[nikomatsakis]

@arora-aman @roxelo can you spell out a bit more what is happening? What fake reads do we introduce?