-
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
Clean up logic around live locals in generator analysis #71956
Clean up logic around live locals in generator analysis #71956
Conversation
debug!("always_live = {:?}", always_live_locals); | ||
|
||
// Locals that are always live or ones that need to be stored across | ||
// suspension points are not eligible for overlap. |
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.
Note: this is an incorrect comment added by me during a refactor. I didn't understand the purpose of this code. Because we compress the matrix to only save GeneratorSavedLocal
s below, I believe the intersect
is unnecessary?
One follow-up question is whether it's sound/desirable to consider liveness when computing storage conflicts. I believe |
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.
Great!
This is not exactly equivalent to the old logic; some generators are now smaller than before. I believe this was because the old logic was too conservative, but I'm not as familiar with the constraints as the original implementers were, so I could be wrong. For example, I don't see a reason the size of the mixed_sizes future couldn't be 5K. It went from 7K to 6K in this PR.
It would be nice to have deeper introspection mechanisms for generators so we can debug this sort of stuff more easily. It would also help with implementing further layout optimizations.
// Ignore unreachable blocks. | ||
if data.terminator().kind == TerminatorKind::Unreachable { | ||
continue; | ||
} |
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 correct in the presence of inline assembly (which is a statement)? It might access arbitrary locals and then diverge, leading to an Unreachable
terminator.
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.
(and yeah, inline assembly should really be a terminator)
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 you're correct, although this is what the current implementation does. We should probably just remove this?
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.
Probably, yeah. I'll try to exploit this for a bit, it's easy to create MIR like this in practice (luckily almost no-one uses inline asm in async
constructs).
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.
Good point.. I think we needed this for some optimizations, but we can amend this to "unreachable and has no statements"
// If the local is moved out of, or if it gets marked `StorageDead`, consider it no | ||
// longer initialized. | ||
PlaceContext::NonUse(NonUseContext::StorageDead) | ||
| PlaceContext::NonMutatingUse(NonMutatingUseContext::Move) => self.trans.kill(local), |
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.
So, this is now ignoring borrowed variables here, where the old analysis was not. With the equations you wrote down in the generator transform, I wouldn't think that this causes the size difference – in fact, I'd expect this to only affect movable generators, but it seems that that's incorrect?
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.
To clarify, you're saying that the version on stable is marking locals as needing to be saved across yield points if they are borrowed even in movable generators? I thought this was an oversight based on the comment next to if !movable {
in the old code. Do we always need to worry about borrows across yield points?
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's definitely fine to ignore them in movable generators as that will cause rather acute UB when the generator is in fact moved around in memory.
I'm just wondering why the previous analysis cared about borrows if they don't impact immovable generators anyways, which is what we wanted to hold off on. And I also was trying to understand why the resulting generators are smaller.
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 previous analysis cared about them because it's possible for something to require storage even if it's been de-initialized. The address could have been captured from a borrow and unsafe code could use that to re-initialize the value.
By the definition of this MaybeInitializedLocals analysis, we shouldn't care about it, but we should still assume in the generator transform that anything borrowed and MaybeStorageLive might still require storage, even if it's not initialized.
Lemme know if you have something particular in mind. |
let loc = Location { block, statement_index }; | ||
trace!("record conflicts at {:?}", loc); | ||
init.seek_before_primary_effect(loc); | ||
borrowed.seek_before_primary_effect(loc); |
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.
So, previously, the StorageConflictVisitor
would only look at the "before-effects", and these seek_before_primary_effect
calls seems to match that, but the reason that was correct was that MaybeRequiresStorage
was marking all locals that will be used by a statement/terminator as live in the before-effect, but now it looks like that is no longer the case since neither MaybeInitializedLocals
nor MaybeBorrowedLocals
use before-effects. Is this still sound, or should this check after the primary effect was applied as well? I'm not really 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.
There is indeed a bug here. I'm missing a seek_before_primary_effect(terminator_loc)
.
The comment below addresses your broader concern. If there's N locations in a block, we also have N dataflow effects in the block (as long as we don't use "before" effects). That means N+1 possibly unique dataflow states, since you need to consider boundary conditions (before the first effect is applied, after the last effect is applied). If we seek before all possible locations, then seek after the final location (via seek_to_block_end
), we will have observed all of them.
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, that makes a lot of sense. Thanks for the explanation!
// Locals that are always live conflict with all other locals. | ||
// | ||
// FIXME: Why do we need to handle locals without `Storage{Live,Dead}` specially here? | ||
// Shouldn't it be enough to know whether they are initialized? |
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.
If they don't have Storage{Live,Dead} then they are considered always StorageLive. If they are also borrowed then they could be reinitialized indirectly.
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, my point is that all conflicts, even between locals that are always StorageLive
, will be detected in the loop below. If such a local is never live and initialized or borrowed at the same time as another local, doesn't it suggest that we were too conservative with our StorageLive
annotations? Once again, I'm worried about broken MIR here, so I didn't try to change this now, but I think it's worth exploring.
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.
Ohh sorry, yeah, we should get rid of the special casing here. I think this was a silly micro-optimization I did early on when the analysis was simply "look for locals that are StorageLive
at the same time".
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'm gonna change this in a follow-up to try and minimize the scope of this PR in case we need to bisect something.
// | ||
// requires_storage := init | borrowed | ||
// | ||
// FIXME: This function is called in a loop, so it might be better to pass in a temporary |
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.
or maybe you could call union_row_with
twice?
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.
Whoops. Yeah, that's clearly better.
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.
This actually doesn't work. We only union into rows for locals that also require storage at this point
// FIXME: This function is called in a loop, so it might be better to pass in a temporary | ||
// bitset rather than cloning here. | ||
let mut requires_storage = init.get().clone(); | ||
requires_storage.union(borrowed.get()); |
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.
actually, I think you can do better than this by intersecting with storage_live ⋃ always_live_locals
, that's what the old code did.
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.
That's true. We could probably be using the actual liveness
here, but I was worried about ill-formed MIR. I can try that in another PR, but for now storage_live | always_live
is a fine approximation.
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.
Actually, once again borrowed | init
is strictly more precise. init
marks locals as uninitialized when we see StorageDead
.
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, and (crucially) so does borrowed
, so you're right.
// Ignore unreachable blocks. | ||
if data.terminator().kind == TerminatorKind::Unreachable { | ||
continue; | ||
} |
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.
Good point.. I think we needed this for some optimizations, but we can amend this to "unreachable and has no statements"
|
||
// Visit every reachable statement and terminator. The exact order does not matter. When two | ||
// locals are live at the same point in time, add an entry in the conflict matrix. | ||
for (block, data) in traversal::preorder(body) { |
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, this API is so much nicer! 💯
4123f6a
to
6e1416f
Compare
Rebased on top of master. I addressed some comments in the last two commits and responded to some others. I don't wanna change how we handle |
6e1416f
to
987647b
Compare
☔ The latest upstream changes (presumably #69171) made this pull request unmergeable. Please resolve the merge conflicts. |
Sorry for the slowness on my end, r=me after rebase |
Instead of using a bespoke dataflow analysis, `MaybeRequiresStorage`, for computing locals that need to be stored across yield points and that have conflicting storage, use a combination of simple, generally applicable dataflow analyses. In this case, the formula for locals that are live at a yield point is: live_across_yield := (live & init) | (!movable & borrowed) and the formula for locals that require storage (and thus may conflict with others) at a given point is: requires_storage := init | borrowed `init` is `MaybeInitializedLocals`, a direct equivalent of `MaybeInitializedPlaces` that works only on whole `Local`s. `borrowed` and `live` are the pre-existing `MaybeBorrowedLocals` and `MaybeLiveLocals` analyses respectively.
The generator transform needs to inspect all possible dataflow states. This can be done with half the number of bitset union operations if we can assume that the relevant analyses do not use "before" effects.
...when determining what locals are live. A local cannot be borrowed before it is `storage_live` and `MaybeBorrowedLocals` already invalidates borrows on `StorageDead`. Likewise, a local cannot be initialized before it is marked StorageLive and is marked as uninitialized after `StorageDead`.
987647b
to
3ff9317
Compare
@bors r=tmandry rollup=never In case we need to bisect. |
📌 Commit 3ff9317 has been approved by |
⌛ Testing commit 3ff9317 with merge 803443d39d917bcb03be70779f4a398b3fa7eafe... |
💔 Test failed - checks-azure |
Another msys2 failure @bors retry |
☀️ Test successful - checks-azure |
Revert rust-lang#71956 ...since it caused unsoundness in rust-lang#73137. Also adds a reduced version of rust-lang#73137 to the test suite. The addition of the `MaybeInitializedLocals` dataflow analysis has not been reverted, but it is no longer used. Presumably there is a more targeted fix, but I'm worried that other bugs may be lurking. I'm not yet sure what the root cause of rust-lang#73137 is. This will need to get backported to beta. r? @tmandry
…-morse,Mark-Simulacrum [beta] backport This is a beta backport rollup of the following: * [beta] Revert heterogeneous SocketAddr PartialEq impls rust-lang#73318 * Fix emcc failure for wasm32. rust-lang#73213 * Revert rust-lang#71956 rust-lang#73153 * [beta] Update cargo rust-lang#73141 * Minor: off-by-one error in RELEASES.md rust-lang#72914 * normalize adt fields during structural match checking rust-lang#72897 * Revert pr 71840 rust-lang#72989 * rust-lang/cargo#8361 * e658200 from rust-lang#72901 r? @ghost
…ir, r=tmandry Check for assignments between non-conflicting generator saved locals This is to prevent future changes to the generator transform from reintroducing the problem that caused rust-lang#73137. Namely, a store between two generator saved locals whose storage does not conflict. My ultimate goal is to introduce a modified version of rust-lang#71956 that handles this case properly. r? @tmandry
…ir, r=tmandry Check for assignments between non-conflicting generator saved locals This is to prevent future changes to the generator transform from reintroducing the problem that caused rust-lang#73137. Namely, a store between two generator saved locals whose storage does not conflict. My ultimate goal is to introduce a modified version of rust-lang#71956 that handles this case properly. r? @tmandry
…ir, r=tmandry Check for assignments between non-conflicting generator saved locals This is to prevent future changes to the generator transform from reintroducing the problem that caused rust-lang#73137. Namely, a store between two generator saved locals whose storage does not conflict. My ultimate goal is to introduce a modified version of rust-lang#71956 that handles this case properly. r? @tmandry
Resolves #69902. Requires #71893.
I've found it difficult to make changes in the logic around live locals in
generator/transform.rs
. It uses a custom dataflow analysis,MaybeRequiresStorage
, that AFAICT computes whether a local is either initialized or borrowed. That analysis is usingbefore
effects, which we should try to avoid if possible because they are harder to reason about than ones only using the unprefixed effects. @pnkfelix has suggested removing "before" effects entirely to simplify the dataflow framework, which I might pursue someday.This PR replaces
MaybeRequiresStorage
with a combination of the existingMaybeBorrowedLocals
and a newMaybeInitializedLocals
.MaybeInitializedLocals
is justMaybeInitializedPlaces
with a coarser resolution: it works on whole locals instead of move paths. As a result, I was able to simplify the logic incompute_storage_conflicts
andlocals_live_across_suspend_points
.This is not exactly equivalent to the old logic; some generators are now smaller than before. I believe this was because the old logic was too conservative, but I'm not as familiar with the constraints as the original implementers were, so I could be wrong. For example, I don't see a reason the size of the
mixed_sizes
future couldn't be 5K. It went from 7K to 6K in this PR.r? @jonas-schievink @tmandry