Pattern guard can consume value that is being matched

[ghost]

Following compiles, and crashes when trying to print moved from value:

fn main() {
    let a = Some("...".to_owned());
    let b = match a {
        Some(_) if { drop(a); false } => None,
        x                             => x,
    };
    println!("{:?}", b);
}

[apasel422]

CC rust-lang/rfcs#107 (comment), #27282.

[alexcrichton]

triage: T-lang I-nominated

I thought we fixed this :(

[Aatch]

@alexcrichton unfortunately no. I fixed a very similar issue involving multiple guards, but not this case.

I did try to fix it, but it essentially requires SEME to work. This is because preventing this case introduces edges between successive patterns in the CFG and consequentially causes code like this:

match foo {
  (ref a, 0) => (),
  (ref a, 1) => ()
  _ => ()
}

to be rejected because the borrow in the first pattern overlaps the borrow in the second pattern. The reason why this has to be this way is because both move checking and borrow checking are done using the CFG, right now the CFG for a match "visits" all the arms in parallel, with pattern guards chaining into each other where appropriate. Improving it further requires chaining from the guard to the next pattern instead, but also requires indicating that the borrow in the pattern ends at both the "else" branch of the guard and the end of the block for the arm. In other words, we need regions with multiple exits aka SEME.

[arielb1]

Can't we just borrow the discriminant in all guards separately (cc @pnkfelix)? Actually, that might cause conflicts with ref mut borrows, but that conflict can be ignored (actually, we may want to mark patterns as aliasable in guards).

This issue won't be really fixed with MIR if we are talking about bypassing-exhaustiveness rather than

[Aatch]

@arielb1 did you accidentally comment on the wrong issue? I'm not sure how discriminants and exhaustiveness come in here.

[nikomatsakis]

triage: P-high

[nikomatsakis]

I think this is one of those "fixed by moving region/borrowck to MIR"

[brson]

@pnkfelix Is this still on your radar? It's been assigned a while.

[brson]

@rust-lang/lang nominated for retriage. Still looks bad. Can we make progress now?

[alexcrichton]

As an update, still segfaults both on normal and MIR trans (example above)

[aturon]

Moving to compiler team; we believe this is blocked on MIR borrowck.

[pnkfelix]

triage: P-medium

[nikomatsakis]

After some discussion in @rust-lang/compiler, the consensus was that we should increase the priority of this issue. The preferred fix is still MIR borrowck. Assigning to @pnkfelix as he has been doing work in that direction.

triage: P-high

[matthewjasper]

Discussed in the weekly NLL triage.
All of the examples here fail to compile with NLL. This also has a test added from #47549. So we think this can be closed.

[landaire]

A friend encountered this case which is not considered a hard error on the latest stable or nightly, but is a clear UAF (check the Miri interpreter):

fn main() {
    let x = String::from("sadsadsadsadsa");

    let x = match 0 {
        0 if { y(x) } => unreachable!(),
        _ => x,
    };

    println!("Hello, world!");
}

fn y(p: String) -> bool {
    false
}

Playground link

Relevant warning:

warning[E0382]: use of moved value: `x`
 --> src/main.rs:6:14
  |
2 |     let x = String::from("sadsadsadsadsa");
  |         - move occurs because `x` has type `std::string::String`, which does not implement the `Copy` trait
...
5 |         0 if { y(x) } => unreachable!(),
  |                  - value moved here
6 |         _ => x,
  |              ^ value used here after move
  |
  = warning: this error has been downgraded to a warning for backwards compatibility with previous releases
  = warning: this represents potential undefined behavior in your code and this warning will become a hard error in the future

This only becomes a hard error caught by NLL if you explicitly add #![feature(nll)]: https://play.rust-lang.org/?version=nightly&mode=debug&edition=2018&gist=b87d208ad05b96a6f3f2552c2b948ea0

I'm not sure if this was intended to be considered a hard error now with NLL, but if not, I think it's worth revisiting.

[matthewjasper]

I think this was closed before we decided to not close fixed by NLL issues until they're hard errors.

Reopening.

[eddyb]

I'm a bit surprised about #![feature(nll)] having any effect on Rust 2018, is that intentional?
If so, where can I read more about that decision?

[landaire]

I think the downgrade is performed here:

rust/src/librustc_mir/borrow_check/mod.rs

Lines 394 to 411 in 2f1bc91

	if tcx.migrate_borrowck() {
	// When borrowck=migrate, check if AST-borrowck would
	// error on the given code.
	// rust-lang/rust#55492, rust-lang/rust#58776 check the base def id
	// for errors. AST borrowck is responsible for aggregating
	// `signalled_any_error` from all of the nested closures here.
	let base_def_id = tcx.closure_base_def_id(def_id);
	match tcx.borrowck(base_def_id).signalled_any_error {
	SignalledError::NoErrorsSeen => {
	// if AST-borrowck signalled no errors, then
	// downgrade all the buffered MIR-borrowck errors
	// to warnings.
	for err in mbcx.errors_buffer.iter_mut() {
	downgrade_if_error(err);
	}

And references these issues related to the downgrade #55492 #58776

For the explicit #![feature(nll)], it looks like this has to do with if tcx.migrate_borrowck(), which going down the call graph leads us to:

rust/src/librustc/ty/context.rs

Lines 1555 to 1583 in fe5f42c

	/// What mode(s) of borrowck should we run? AST? MIR? both?
	/// (Also considers the `#![feature(nll)]` setting.)
	pub fn borrowck_mode(&self) -> BorrowckMode {
	// Here are the main constraints we need to deal with:
	//
	// 1. An opts.borrowck_mode of `BorrowckMode::Migrate` is
	// synonymous with no `-Z borrowck=...` flag at all.
	//
	// 2. We want to allow developers on the Nightly channel
	// to opt back into the "hard error" mode for NLL,
	// (which they can do via specifying `#![feature(nll)]`
	// explicitly in their crate).
	//
	// So, this precedence list is how pnkfelix chose to work with
	// the above constraints:
	//
	// * `#![feature(nll)]` *always* means use NLL with hard
	// errors. (To simplify the code here, it now even overrides
	// a user's attempt to specify `-Z borrowck=compare`, which
	// we arguably do not need anymore and should remove.)
	//
	// * Otherwise, if no `-Z borrowck=...` then use migrate mode
	//
	// * Otherwise, use the behavior requested via `-Z borrowck=...`
	if self.features().nll { return BorrowckMode::Mir; }
	self.sess.opts.borrowck_mode
	}

Can't seem to find the relevant discussion for the decision here though.

[eddyb]

@landaire Yeah but shouldn't Rust 2018 turn on the NLL borrowck in its full mode?
Unless BorrowckMode::Mir was never on stable and the Compare mode was enabled for 2015 recently?
(after having been the mode on the 2018 edition all this time)

[matthewjasper]

Unless BorrowckMode::Mir was never on stable

it wasn't.

[pnkfelix]

@eddyb wrote:

I'm a bit surprised about #![feature(nll)] having any effect on Rust 2018, is that intentional?
If so, where can I read more about that decision?

and also wrote:

@landaire Yeah but shouldn't Rust 2018 turn on the NLL borrowck in its full mode?
Unless BorrowckMode::Mir was never on stable and the Compare mode was enabled for 2015 recently?
(after having been the mode on the 2018 edition all this time)

See also #58781.

In particular is mention there of making full NLL mode the default for the 2018 edition on its own (vs doing that for all editions en masse).

p.s. the default is BorrowckMode::Migrate, not Compare. We should kill off Compare mode in the short term, as its not serving a real purpose anymore.