-
Notifications
You must be signed in to change notification settings - Fork 12.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Refine scopes around temporaries generated in local accesses #92508
Refine scopes around temporaries generated in local accesses #92508
Conversation
(rust-highfive has picked a reviewer for you, use r? to override) |
r? @pnkfelix |
Author of #91032 here; I saw this show up as a reference when I was looking over comments in my PR. I'll add a couple review comments, but I wanted to leave a few comments here too. It doesn't look like this PR fixes all of #69663. For example, I don't think this handles the case where you explicitly This PR seems pretty orthogonal to mine (#91032), so I think it makes sense to merge both of them. My PR mostly addresses the drop case, while yours seems to handle temporaries in indexing or Anyway, this is awesome to see this work! |
@@ -358,6 +367,34 @@ impl ScopeTree { | |||
self.rvalue_scopes.insert(var, lifetime); | |||
} | |||
|
|||
pub fn record_local_access_scope(&mut self, var: hir::ItemLocalId, proposed_lifetime: Scope) { | |||
debug!("record_local_access_scope(sub={:?}, sup={:?})", var, proposed_lifetime); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Any reason not to have this as var={:?}, proposed_lifetime={:?}
instead of sub={:?}, sup={:?}
? I tend to find it tricky to remember what sub and sup are when I just see those two terms.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Right, updated with your suggestion.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Awesome, thanks!
compiler/rustc_passes/src/region.rs
Outdated
} | ||
let mut ref_level = SmallVec::<[_; 4]>::new(); | ||
let mut ops_iter = ops.into_iter().rev(); | ||
ops_iter.next().unwrap(); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It might be worth a comment here saying that ops_iter
always has at least one element because all of the arms from the previous match
either push an item or return.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I simplified the tracking a little bit, so that it does not need to track the local variable itself.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ah, even better. Thanks!
async move { | ||
match client.status() { | ||
async { | ||
match Client(Box::new(true)).status() { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is this just cleaning up the test, or were changes needed so that it still tested the same behavior as before? In other words, did the error from the previous version go away due to your changes?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The error goes away with this change. However, I checked the issue #64130 and it seems that this test is intended for prompting hints for remedies when values living through yield points make generators non-Send
or non-Sync
. Therefore, I figured that it is good to keep this test by reproducing the previous errors.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sounds good!
@@ -0,0 +1,51 @@ | |||
// edition:2018 | |||
// run-pass |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
build-pass
is probably enough here, or probably even check-pass
(which I think is the default if you don't do anything).
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think build-pass
is better. It seems that there might be changes to MIR outputs, at least to intermediate MIR outputs.
@@ -0,0 +1,32 @@ | |||
// edition:2018 | |||
// run-pass |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Here too, it looks like build-pass
or check-pass
should be enough.
For #91032, I tried to use check-pass
when I could, and then build-pass
when I had a case where the new algorithm might remove a type that would still be in the MIR, since build-pass
runs the check that the types MIR found were a subset of those found in typeck.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It is updated to perform build-pass
test.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks!
Thank you @eholk !
That is true. I think I incorrectly stated that this PR will fix #69663 entirely. This is mistakenly included as there are quite a few issues that this PR may affect. I have labelled it to be relevant instead. Hopefully our works could together improve the usability of generators. 😄 |
9d8c451
to
2b56bcc
Compare
2b56bcc
to
071b196
Compare
Looking at this now; thanks for your patience, @dingxiangfei2009 ! |
/// all temporaries are dropped while it may be the case that | ||
/// only a subset of temporaries is requested to be dropped. | ||
/// For instance, `x[y.z()]` indexing operation can discard the temporaries | ||
/// generated in evaluating `y.z()`. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
For instance,
x[y.z()]
indexing operation can discard the temporaries generated in evaluatingy.z()
.
I'll admit, I still need to reload my mental cache here, but: Is this actually true? Aren't there cases where the value produced by y.z()
may hold references to those temporaries, so you'd need to keep them alive across the evaluation of the indexing operation?
I'll do some investigation to double-check my understanding here, and see if I can convince myself that you are correct.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
(While talking to niko about this, he notes the example of mutex.lock().something()
as a potentially problematic example...)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think this was initially bending my mind as well. Let me try to elaborate. So the indexing operation, either through Index
or IndexMut
, produces borrows with lifetimes that must outlives self
.
Let us take Index
as an example. In this example, we attempt to allow indexing into A
with a temporary reference &()
, producing a value of type &B<'_>
. The result of this indexing will need the temporary index &()
to be kept alive
use std::ops::Index;
struct A;
struct B<'a>(&'a ());
impl<'a> B<'a> {
fn foo(&self) {}
}
impl<'a> Index<&'a ()> for A {
type Output = B<'a>;
fn index<'s>(&'s self, idx: &'a ()) -> &'s Self::Output {
unimplemented!() // (1)
}
}
fn main() {
let a = A;
a[&()].foo();
}
At (1)
, we are expected to construct a borrow with a lifetime 's
. However, the Index
trait definition does allow us to impose Self::Output: 's
or any outliving relation between 'a
and 's
in this scope. Overall, this means that if the index, which is '_ ()
in this example, has a lifetime 'a
, 's
must outlives 'a
.
A side track
I think there should be an implicit bound ofSelf::Output: 's
to be satisfied, but the current compiler does not complain as long as the return value is !
.
Regarding the example from @nikomatsakis, let us example the following elaborated code.
use std::ops::{Index, IndexMut, DerefMut};
use std::cell::Cell;
use std::sync::Mutex;
struct A<'a>(Cell<&'a mut ()>);
impl<'a> Index<&'a mut ()> for A<'a> {
type Output = ();
fn index(&self, idx: &'a mut ()) -> &Self::Output {
&*self.0.replace(idx)
}
}
impl<'a> IndexMut<&'a mut ()> for A<'a> {
fn index_mut(&mut self, idx: &'a mut ()) -> &mut Self::Output {
self.0.replace(idx)
}
}
struct B(());
impl B {
fn new() -> B {
B(())
}
fn foo(&mut self) -> &mut () {
&mut self.0
}
}
fn main() {
let t = &mut ();
let mut a = A(Cell::new(t));
let i = Mutex::new(());
a[i.lock().unwrap().deref_mut()] = (); // <- temporary value is freed at the end of this statement
a[B::new().foo()] = (); // <- same reason
a[&mut ()] = ();
}
The expression a[i.lock().unwrap().deref_mut()] = ()
is currently rejected because the temporary mutex guard mutex.lock()
is preserved only until the end of this assignment.
In any case, I am also experimenting ways to produce examples that requiring living temporaries after index to be extra sure.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Okay.
It looks like you are making an argument based on the lifetimes embedded in the Idx
type of trait Index<Idx>
.
I think I can see the reasoning there; its something like "the definition of the Index
trait: pub trait Index<Idx: ?Sized> { type Output: ?Sized; fn index(&self, index: Idx) -> &Self::Output; }
, will force any free lifetimes associated with the Idx
to be bound at the scope of the impl<'a, 'b, ...> Index<...> for Concrete<...>
for any concrete type Concrete<...>
."
My follow-up question then is: Is there a case where the Idx
type doesn't have any lifetimes in it, but the scope of temporaries of the expression computing Idx
are nonetheless significant (and where the change suggested here could break code)?
My current suspicion is that the most natural cases where that could arise is that it could only happen in unsafe code that is incorrect (i.e., unsafe code that should be using lifetimes to express the constraints between the temporaries and the dynamic extent of the produced value, but is sidestepping doing so, and in the process opens up potential issues where user code can violate those now implicit constraints).
- I think we have precedent for making changes like that, but I also think we need to be careful whenever we do so.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
(In case its not clear, I'm going to poke around a bit more and see if I can make my concerns here concrete. My current thinking is that you may indeed be safe in the assertions you are making here, but I want to make sure I understand the edge cases really well.)
☔ The latest upstream changes (presumably #92007) made this pull request unmergeable. Please resolve the merge conflicts. |
071b196
to
f06d576
Compare
☔ The latest upstream changes (presumably #93893) made this pull request unmergeable. Please resolve the merge conflicts. |
Co-authored-by: Felix S <[email protected]>
f06d576
to
348584d
Compare
While playing around with the example provoided by @dingxiangfei2009 over here, I happened upon this small variant of what they posted which changes behavior under this PR: use std::ops::{Index, IndexMut};
use std::cell::Cell;
struct A<'a>(Cell<&'a mut ()>);
impl<'a> Index<&'a mut ()> for A<'a> {
type Output = ();
fn index(&self, idx: &'a mut ()) -> &Self::Output {
&*self.0.replace(idx)
}
}
impl<'a> IndexMut<&'a mut ()> for A<'a> {
fn index_mut(&mut self, idx: &'a mut ()) -> &mut Self::Output {
self.0.replace(idx)
}
}
fn main() {
let t = &mut ();
let mut a = A(Cell::new(t));
a[&mut ()] = ();
} Namely, the nightly (and stable) compiler accept the above code, but the compiler patched with your PR rejects it:
That implies to me that this PR cannot land as is. I talked to @nikomatsakis a little while about this, and they were thinking that a better way to go here would be a more holistic approach and attacking RFC 66 "Better temporary lifetimes", (potentially on an edition boundary, if need be), rather than trying to make small adustments with as-yet unknown impact. |
@pnkfelix Thank you for providing this example. I did not realize that some temporaries like I can restructure the PR to discard the scope rule concerning indexing and keep the scope rule for borrows of local variable. From here, I can continue to look at RFC 66. This would at least admit the program in the original issue so that in the future while working on RFC 66 the allowed programs here can be used as reference. By the way, I am interested in attacking RFC 66 because it has the potential to replace the work here. It does seem like a great effort, so I will greatly appreciate any mentorship. 😄 In fact, if there is any zulip thread about this RFC, I can join the discussion there. |
@pnkfelix Actually, I would like to revisit your example again. I extend your example by duplicating the last expression. use std::ops::{Index, IndexMut};
use std::cell::Cell;
struct A<'a>(Cell<&'a mut ()>);
impl<'a> Index<&'a mut ()> for A<'a> {
type Output = ();
fn index(&self, idx: &'a mut ()) -> &Self::Output {
&*self.0.replace(idx)
}
}
impl<'a> IndexMut<&'a mut ()> for A<'a> {
fn index_mut(&mut self, idx: &'a mut ()) -> &mut Self::Output {
self.0.replace(idx)
}
}
fn main() {
let t = &mut ();
let mut a = A(Cell::new(t));
a[&mut ()] = (); // <- Error
a[&mut ()] = ();
} This, on stable, gives this error.
Of cause, any access to use std::ops::{Index, IndexMut};
use std::cell::Cell;
struct A<'a>(Cell<&'a mut ()>);
const UNIT: () = ();
impl<'a> Index<usize> for A<'a> {
type Output = ();
fn index(&self, _: usize) -> &Self::Output {
&UNIT
}
}
impl<'a> Index<&'a mut ()> for A<'a> {
type Output = ();
fn index(&self, idx: &'a mut ()) -> &Self::Output {
&*self.0.replace(idx)
}
}
impl<'a> IndexMut<&'a mut ()> for A<'a> {
fn index_mut(&mut self, idx: &'a mut ()) -> &mut Self::Output {
self.0.replace(idx)
}
}
fn main() {
let t = &mut ();
let mut a = A(Cell::new(t));
a[&mut ()] = ();
a[0]; // <- same error here
}
// Similarly, even without the `a[0];` line, implementing Drop also leads to the same compilation error.
impl<'a> Drop for A<'a> {
fn drop(&mut self) {
println!("bye")
}
} However, I still intend to remove the scope rules for indexing. I just wanted to share that this example is very thought-provoking. |
Okay, that might be good to look at. I might recommend closing this PR and opening a fresh one if you go down that route, just because it might be confusing to have metadata linking from #72956 claiming that this PR (PR #92508) fixes that issue (issue #72956) when in fact the scope of PR #92508 has been reduced to not cover that issue. What do you think about that idea? |
@dingxiangfei2009 regarding RFC 66, I'd be open to working with you on that. |
☔ The latest upstream changes (presumably #94427) made this pull request unmergeable. Please resolve the merge conflicts. |
by reading the latest comments, I think I'll flip the review status to waiting on author to reflect that there seems to be some design work (in case, feel free to adjust, thanks!) @rustbot author |
@dingxiangfei2009 - can you please address the design concerns FYI: when a PR is ready for review, send a message containing apiraino commented on Mar 24 |
@dingxiangfei2009 @rustbot label: +S-inactive |
Fix #57017
Fix #72956
Related to #69663
Sorry for the long sabbatical. This PR takes inspiration from the comments by @pnkfelix and further extends the idea to at least allow refined scopes for temporaries generated while evaluating expressions like
x.status()
. In details, expressions likex.status()
generates temporaries for place or place references to a local variablex
, but they could be discarded at an earlier opportunity. What is left to improve the region analysis in a generator body is to identify regions where we can restrict those temporaries. They are tentatively set toif
andmatch
sub-expressions.I have come to a rather late realization that #91032 could be solving the related problems in a more comprehensive setting. However, I would like to gather some feedback and join the relevant discussion with this opportunity.