Lifetimes bug on beta 1.31: the associated type may not live long enough

A reduced test case (play):

#![crate_type="lib"]

pub trait Database<'a> {
    type Guard: 'a;
}

pub struct Stateful<'a, D: 'a>(&'a D);

impl<'b, D: for <'a> Database<'a>> Stateful<'b, D> {
    pub fn callee<'a>(&'a self) -> <D as Database<'a>>::Guard {
        unimplemented!()
    }
    pub fn caller<'a>(&'a self) -> <D as Database<'a>>::Guard {
        let state = self.callee();
        unimplemented!()
    }
}

Error diagnostic output in details block below

Details

error[E0311]: the associated type `<D as Database<'_>>::Guard` may not live long enough
  --> src/lib.rs:14:13
   |
14 |         let state = self.callee();
   |             ^^^^^
   |
   = help: consider adding an explicit lifetime bound for `<D as Database<'_>>::Guard`
note: the associated type `<D as Database<'_>>::Guard` must be valid for the block suffix following statement 0 at 14:9...
  --> src/lib.rs:14:9
   |
14 | /         let state = self.callee();
15 | |         unimplemented!()
16 | |     }
   | |_____^
note: ...so that variable is valid at time of its declaration
  --> src/lib.rs:14:13
   |
14 |         let state = self.callee();
   |             ^^^^^

error[E0311]: the associated type `<D as Database<'_>>::Guard` may not live long enough
  --> src/lib.rs:14:13
   |
14 |         let state = self.callee();
   |             ^^^^^
   |
   = help: consider adding an explicit lifetime bound for `<D as Database<'_>>::Guard`
note: the associated type `<D as Database<'_>>::Guard` must be valid for the block at 13:63...
  --> src/lib.rs:13:63
   |
13 |       pub fn caller<'a>(&'a self) -> <D as Database<'a>>::Guard {
   |  _______________________________________________________________^
14 | |         let state = self.callee();
15 | |         unimplemented!()
16 | |     }
   | |_____^
note: ...so that references are valid when the destructor runs
  --> src/lib.rs:14:13
   |
14 |         let state = self.callee();
   |             ^^^^^

error[E0311]: the associated type `<D as Database<'_>>::Guard` may not live long enough
  --> src/lib.rs:14:21
   |
14 |         let state = self.callee();
   |                     ^^^^^^^^^^^^^
   |
   = help: consider adding an explicit lifetime bound for `<D as Database<'_>>::Guard`
note: the associated type `<D as Database<'_>>::Guard` must be valid for the method call at 14:21...
  --> src/lib.rs:14:21
   |
14 |         let state = self.callee();
   |                     ^^^^^^^^^^^^^
note: ...so type `<D as Database<'_>>::Guard` of expression is valid during the expression
  --> src/lib.rs:14:21
   |
14 |         let state = self.callee();
   |                     ^^^^^^^^^^^^^

error[E0311]: the associated type `<D as Database<'_>>::Guard` may not live long enough
  --> src/lib.rs:14:21
   |
14 |         let state = self.callee();
   |                     ^^^^^^^^^^^^^
   |
   = help: consider adding an explicit lifetime bound for `<D as Database<'_>>::Guard`
note: the associated type `<D as Database<'_>>::Guard` must be valid for the destruction scope surrounding statement at 14:9...
  --> src/lib.rs:14:9
   |
14 |         let state = self.callee();
   |         ^^^^^^^^^^^^^^^^^^^^^^^^^^
note: ...so that references are valid when the destructor runs
  --> src/lib.rs:14:21
   |
14 |         let state = self.callee();
   |                     ^^^^^^^^^^^^^

error: aborting due to 4 previous errors

Original bug report

Versions

rustc 1.31.0-beta.4 (04da282 2018-11-01)
cargo 1.31.0-beta (efb7972a0 2018-10-24)

Expected behavior

Everything compiling correctly like it does on stable 1.30

Actual behavior

A lot of lifetime-related errors:

error[E0311]: the associated type `<D as trie::Database<'_>>::Guard` may not live long enough                                                                                                                                    
   --> /home/mike/dev/etcdev/sputnikvm/stateful/src/lib.rs:536:23                                                                                                                                                                
    |                                                                                                                                                                                                                            
536 |         let account = state.get(&address);                                                                                                                                                                                 
    |                       ^^^^^^^^^^^^^^^^^^^                                                                                                                                                                                  
    |                                                                                                                                                                                                                            
    = help: consider adding an explicit lifetime bound for `<D as trie::Database<'_>>::Guard`                                                                                                                                    
note: the associated type `<D as trie::Database<'_>>::Guard` must be valid for the method call at 536:23...                                                                                                                      
   --> /home/mike/dev/etcdev/sputnikvm/stateful/src/lib.rs:536:23                                                                                                                                                                
    |                                                                                                                                                                                                                            
536 |         let account = state.get(&address);                                                                                                                                                                                 
    |                       ^^^^^^^^^^^^^^^^^^^                                                                                                                                                                                  
note: ...so that a type/lifetime parameter is in scope here                                                                                                                                                                      
   --> /home/mike/dev/etcdev/sputnikvm/stateful/src/lib.rs:536:23                                                                                                                                                                
    |                                                                                                                                                                                                                            
536 |         let account = state.get(&address);                                                                                                                                                                                 
    |                       ^^^^^^^^^^^^^^^^^^^

It has been an issue in nightly 1.31 for a long time now, but I haven't reported the issue while it was just nightly. Now the regression is in beta channel and may land in stable, breaking existing code.

My own findings on the subject: regression is related to an old lifetime system, it's not reproducible with nll feature on, so on nightly the code can be compiled with nll, but not without it. Same holds for beta now except nll isn't quite there yet.

The subjected codebase: https://github.com/ETCDEVTeam/sputnikvm/tree/master/stateful

visiting for T-compiler triage. P-high. Assigning to self for initial investigation.

@sorpaas

Based on crater run, looks like this also affects:

etclient, cc @sorpaas
evm-stateful, cc @sorpaas
sputnikvm-stateful, cc @splix
ETCDEVTeam/sputnikvm-dev (cc @splix ?)
TheWaWaR/cita-create-genesis cc @TheWaWaR

Bisected against when https://source.that.world/source/evm-rs.git stopped building.

Successfully built: rustc 1.30.0-nightly (4141a40 2018-09-25)
Error during compilation: rustc 1.30.0-nightly (ae7fe84 2018-09-26)

Heres the bors merge record from that time frame:

% git log 4141a40..ae7fe84 --format=oneline --author=bors
ae7fe84 Auto merge of #54561 - matthiaskrgr:clippy_up, r=kennytm
c7df1f5 Auto merge of #54453 - nikomatsakis:nll-issue-53121-shred-outlives, r=pnkfelix
6846f22 Auto merge of #51946 - japaric:emit-stack-sizes, r=nikomatsakis
e783d2b Auto merge of #54199 - nikomatsakis:predicate_may_hold-failure, r=eddyb
c3a1a0d Auto merge of #53824 - ljedrz:begone_onevector, r=michaelwoerister
a2b27c1 Auto merge of #54497 - ralexstokes:stabilize_pattern_parentheses, r=nikomatsakis
beff387 Auto merge of #54548 - nrc:update, r=kennytm
0b2eae7 Auto merge of #54164 - mikhail-m1:54131, r=pnkfelix
92aff72 Auto merge of #54575 - pietroalbini:rollup, r=pietroalbini
%

The most obvious culprit to my eye is PR #54453. That PR might have been intended to be restricted to NLL, I'm not sure (it touches code outside of rustc_mir, certainly). It does seem like the crate in question is not opting into NLL nor the 2018 editon.

Update: of course they aren't opting into NLL. This issue is tagged NLL-fixed-by-NLL

Here is a reduced test case (play):

#![crate_type="lib"]

pub trait Database<'a> {
    type Guard: 'a;
}

pub struct Stateful<'a, D: 'a>(&'a D);

impl<'b, D: for <'a> Database<'a>> Stateful<'b, D> {
    pub fn callee<'a>(&'a self) -> <D as Database<'a>>::Guard {
        unimplemented!()
    }
    pub fn caller<'a>(&'a self) -> <D as Database<'a>>::Guard {
        let state = self.callee();
        unimplemented!()
    }
}

I am not sure if I'm going to have time to look into this further before I go on leave (for ~2 weeks), and yet it seems like a potentially important thing to address. Unassigning self.

@nikomatsakis

I have now confirmed that PR #54453 is what injected this regression.

cc @nikomatsakis

Also, it's interesting that NLL makes this test case pass. Off hand, I'm not sure entirely sure why that is.

OK, so I investigated NLL a bit. The reason is actually somewhat interesting.

NLL introduces an additional liveness check -- it says:

For any variable X live at the position P, all regions that appear in X's type must include the point P.

This rule is much "cruder" than the outlives check here. In particular, if you have a variable x: <D as Database<'a>>::Guard, then for it to pass this liveness check, 'a must include all points where x occurs.

We could adopt a similar distinction in lexical region checking with relative ease -- basically a "live at" relation that is more conservative than outlives. I'm not sure how well justified it is though.

(This is a bit weird because of eager normalization: if we were able to normalize <D as Database<'a>>::Guard to another type (e.g., u32), then we would not impose this requirement on 'a, because we would be representing its type as merely u32. This will change with lazy normalization. It is conceivable that this could cause programs to stop type-checking. I guess we'll worry about it then.)

@nikomatsakis

@nikomatsakis says they know what the problem is and hope to have a fix soon (today).

Hmm, so, I have a fix. Though looking more closely at the test, I think I see a nicer fix.

The problem is that we now observe two potentially applicable rules when trying to find bounds for <T as Database<'0>>::Guard:

<T as Database<'a>>::Guard: 'a // from the where clauses
for<'b> { <T as Database<'b>>::Guard: 'b } // from the trait definition

Because of this, we decide to take no action to influence inference, which means that '0 winds up unconstrained, leading to the ultimate error.

My current fix basically overrides this with a coarser rule (one which works fine, however, in this example).

A smarter fix would be to observe that the first rule (the one talking about 'a) is in fact "specialization" of the latter, in which case we could conclude that we should relate '0 to the bound we need.

I am trying to decide which approach is lowest risk -- the current fix, as I wrote, actually makes us compatible with how NLL is handling things. I am planning to overhaul this whole setup in the not-that-distinct future, since the current solver is essentially "unnecessarily stupid".

Some options:

Land my current fix, which is relatively simple and (I think) low risk.
- Perhaps follow-up with an attempt to create a test case and a FIXME?
Try to do the better fix

Given the deadline, I'm leaning towards the first path.

Rollup merge of rust-lang#56043 - nikomatsakis:issue-55756-via-outliv…

Testing the more conservative fix from #56043:

@mersinvald

@mersinvald

It has been an issue in nightly 1.31 for a long time now, but I haven't reported the issue while it was just nightly. Now the regression is in beta channel and may land in stable, breaking existing code.

btw, in the future, we would definitely appreciate bug reports against nightly! That is indeed part of the point of nightly, to help us find regressions before they make it into beta =)

@Mark-Simulacrum

I don't know how to test the final two cases that @Mark-Simulacrum cited -- I guess I need to checkout some github repositories? -- but based on the logs they seem like the same root problem and hence I expect it is fixed.

Yes, you'd need to build https://github.com/ETCDEVTeam/sputnikvm-dev and https://github.com/TheWaWaR/cita-create-genesis to check them locally.

Rollup merge of rust-lang#56043 - nikomatsakis:issue-55756-via-outliv…

@Mark-Simulacrum

@Mark-Simulacrum
https://github.com/TheWaWaR/cita-create-genesis
Confirm success build both on:

rustc 1.30.1 (1433507eb 2018-11-07)
rustc 1.32.0-nightly (5aff30734 2018-11-19)

@mersinvald

@mersinvald

It has been an issue in nightly 1.31 for a long time now, but I haven't reported the issue while it was just nightly. Now the regression is in beta channel and may land in stable, breaking existing code.

btw, in the future, we would definitely appreciate bug reports against nightly! That is indeed part of the point of nightly, to help us find regressions before they make it into beta =)

Sure, I understand I should've report it once encountered. Would be less stressful for both parties. Thanks for fixing the issue!

Navigation Menu

Repository navigation

Lifetimes bug on beta 1.31: the associated type may not live long enough #55756

Original bug report

Versions

Expected behavior

Actual behavior

1 remaining item

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Lifetimes bug on beta 1.31: the associated type may not live long enough #55756

Description

Original bug report

Versions

Expected behavior

Actual behavior

Activity

pnkfelix commented on Nov 8, 2018

Mark-Simulacrum commented on Nov 9, 2018

pnkfelix commented on Nov 12, 2018

pnkfelix commented on Nov 12, 2018

pnkfelix commented on Nov 12, 2018

pnkfelix commented on Nov 12, 2018

pnkfelix commented on Nov 12, 2018

pnkfelix commented on Nov 12, 2018

1 remaining item

nikomatsakis commented on Nov 13, 2018

nikomatsakis commented on Nov 13, 2018

nagisa commented on Nov 15, 2018

nikomatsakis commented on Nov 15, 2018

nikomatsakis commented on Nov 19, 2018

nikomatsakis commented on Nov 19, 2018

nikomatsakis commented on Nov 19, 2018

Mark-Simulacrum commented on Nov 19, 2018

TheWaWaR commented on Nov 20, 2018

mersinvald commented on Nov 20, 2018

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions