Rewrite handling of universe-leaking placeholder regions into outlives constraints #123720

amandasystems · 2024-04-10T12:50:43Z

This commit prepares for Polonius by moving handling of leak check/universe errors out of the inference step by rewriting any universe error into an outlives-static constraint.

This variant is a work in progress but seems to pass most tests.

Note that a few debug assertions no longer hold; a few extra eyes on those changes are appreciated!

rustbot · 2024-04-10T12:50:51Z

r? @davidtwco

rustbot has assigned @davidtwco.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

amandasystems · 2024-04-10T12:52:02Z

?r @nikomatsakis

(it's not 100% ready for review yet probably maybe, but I figured I might as well start the PR now)

compiler-errors · 2024-04-10T14:15:16Z

the syntax is r? nikomatsakis 😆

amandasystems · 2024-04-10T16:26:29Z

the syntax is r? nikomatsakis 😆

That’s funny because I tried that first, thought I made a mistake, and went back to “fix” it

amandasystems · 2024-04-12T09:19:38Z

Update: I had a brainwave during my investigations of the failing UI tests (there's 20-ish of them if you enable the new logic everywhere, depending on whether you count duplicates). I tried adding a new constraint category so that I could easily not blame it (ever, actually), and that reduces the errors down to one, tiny span diff:

---- [ui] tests/ui/higher-ranked/trait-bounds/due-to-where-clause.rs stdout ----
diff of stderr:

2         --> $DIR/due-to-where-clause.rs:2:5
3          |
4       LL |     test::<FooS>(&mut 42);
-          |     ^^^^^^^^^^^^ implementation of `Foo` is not general enough
+          |     ^^^^^^^^^^^^^^^^^^^^^ implementation of `Foo` is not general enough
6          |
7          = note: `FooS<'_>` must implement `Foo<'0>`, for any lifetime `'0`...
8          = note: ...but `FooS<'_>` actually implements `Foo<'1>`, for some specific lifetime `'1`

I'm pretty confident this works now. Removing the WIP label!

nikomatsakis

Efficiency nit

compiler/rustc_borrowck/src/constraints/mod.rs

amandasystems · 2024-04-12T15:01:52Z

I think it's ready for a perf run, @lqd!

lqd · 2024-04-12T16:04:24Z

@bors try @rust-timer queue

bors · 2024-04-12T16:05:35Z

⌛ Trying commit 5a1c148 with merge 1b9e1e6...

…=<try> Rewrite handling of universe-leaking placeholder regions into outlives constraints This commit prepares for Polonius by moving handling of leak check/universe errors out of the inference step by rewriting any universe error into an outlives-static constraint. This variant is a work in progress but seems to pass most tests. Note that a few debug assertions no longer hold; a few extra eyes on those changes are appreciated!

nikomatsakis

OK, so, I reviewed and left a (tiny) nit. I don't know that a perf run is really needed, this is pretty clearly grossly inefficient. That's precisely what I expected, of course.

r=me but I'd like to hand this off to somebody else to drive over the finish line

compiler/rustc_borrowck/src/constraints/mod.rs

bors · 2024-04-12T19:15:21Z

☀️ Try build successful - checks-actions
Build commit: 1b9e1e6 (1b9e1e65acb35d4d1d734c6461ace5c4e50a9fd0)

rust-timer · 2024-04-13T00:32:37Z

Finished benchmarking commit (1b9e1e6): comparison URL.

Overall result: ❌ regressions - ACTION NEEDED

Benchmarking this pull request likely means that it is perf-sensitive, so we're automatically marking it as not fit for rolling up. While you can manually mark this PR as fit for rollup, we strongly recommend not doing so since this PR may lead to changes in compiler perf.

Next Steps: If you can justify the regressions found in this try perf run, please indicate this with @rustbot label: +perf-regression-triaged along with sufficient written justification. If you cannot justify the regressions please fix the regressions and do another perf run. If the next run shows neutral or positive results, the label will be automatically removed.

@bors rollup=never
@rustbot label: -S-waiting-on-perf +perf-regression

Warning ⚠: The following benchmark(s) failed to build:

rustc
wg-grammar
webrender-2022

Instruction count

This is a highly reliable metric that was used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	113.9%	[0.3%, 3862.1%]	236
Regressions ❌ (secondary)	393.2%	[0.1%, 7855.4%]	184
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	113.9%	[0.3%, 3862.1%]	236

Max RSS (memory usage)

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-3.7%	[-10.5%, -0.7%]	57
Improvements ✅ (secondary)	-5.6%	[-7.4%, -3.5%]	21
All ❌✅ (primary)	-3.7%	[-10.5%, -0.7%]	57

Cycles

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	215.7%	[0.9%, 3636.4%]	98
Regressions ❌ (secondary)	1147.6%	[1.0%, 7887.5%]	61
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	215.7%	[0.9%, 3636.4%]	98

Binary size

This benchmark run did not return any relevant results for this metric.

Bootstrap: missing data
Artifact size: 315.97 MiB -> 305.14 MiB (-3.43%)

bors · 2024-04-16T16:16:26Z

☔ The latest upstream changes (presumably #124015) made this pull request unmergeable. Please resolve the merge conflicts.

amandasystems · 2024-06-18T13:47:17Z

This is my first crater run so I'm not sure how to read the logs, but most of those errors look spurious to me, if not all of them.

This version is a squash-rebased version of a series of exiermental commits, since large parts of them were broken out into PR rust-lang#125069. It explicitly handles universe violations in higher-kinded outlives constraints by adding extra outlives static constraints.

amandasystems · 2024-07-01T09:37:09Z

r=me with a nit

Nit addressed (and I agree, that's just the sort of bad naming I know has confused me before)

amandasystems · 2024-07-01T09:54:37Z

@nikomatsakis, @lqd: I have added a description for why this was done in commit 9be3a3d. Check if it looks ok to you!

nikomatsakis · 2024-07-02T16:19:41Z

@bors r+

bors · 2024-07-02T16:19:44Z

📌 Commit 9be3a3d has been approved by nikomatsakis

It is now in the queue for this repository.

bors · 2024-07-02T16:19:45Z

🌲 The tree is currently closed for pull requests below priority 100. This pull request will be tested once the tree is reopened.

bors · 2024-07-03T01:24:11Z

⌛ Testing commit 9be3a3d with merge 67f0d43...

bors · 2024-07-03T03:42:19Z

☀️ Test successful - checks-actions
Approved by: nikomatsakis
Pushing 67f0d43 to master...

rust-timer · 2024-07-03T06:07:27Z

Finished benchmarking commit (67f0d43): comparison URL.

Overall result: ❌ regressions - no action needed

@rustbot label: -perf-regression

Instruction count

This is a highly reliable metric that was used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	1.3%	[1.2%, 1.4%]	6
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-	-	0

Max RSS (memory usage)

This benchmark run did not return any relevant results for this metric.

Cycles

Results (primary 2.2%)

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	2.2%	[2.2%, 2.2%]	1
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	2.2%	[2.2%, 2.2%]	1

Binary size

This benchmark run did not return any relevant results for this metric.

Bootstrap: 696.719s -> 695.519s (-0.17%)
Artifact size: 327.57 MiB -> 327.72 MiB (0.05%)

lqd · 2024-07-08T08:37:45Z

This is my first crater run so I'm not sure how to read the logs, but most of those errors look spurious to me, if not all of them.

Agreed, most of the 12 "regressions" look like crater errors. Since these crates not tested in the run however, I've checked them all locally: they build with this PR without issues.

We'll keep an eye out for possible new issues, and beta crater run regressions, that could be bisected to this PR, just in case the "spurious regressed" contains real effects (these can also be crater errors, but also contains OOMs).

But things look currently fine in my opinion.

…mpletely, r=<try> [WIP] Remove placeholders completely This PR does shotgun surgery on borrowck to remove all special handling of placeholders, completely replacing them with a preprocessing step that rewrites placeholder leaks into constraints, removing constraint propagation of placeholders and all logic used to detect placeholder violations during error reporting. This finishes what rust-lang#123720 started. The new method works like this: 1. during SCC construction, some information about SCC membership and reachability is retained 2. just after SCC construction, a constraint `r - (from: to_invalid) - > 'static` is added when `r` is the representative of an SCC and 1. that SCC either has had its universe shrunk because it reaches a region with a smaller one (in which case `to_invalid` is the smallest-universed region it reaches), or if it reaches a region with a too large universe that isn't part of the SCC (in which case `to_invalid` is the region with a too large universe). In either case, `from` is also `r`. 2. some region `reaches` in `r`'s SCC reaches another placeholder, `reached`, in which case the added constraint is `r -> (reaches: reached) 'static`. Through clever choice of defaults (chosing minimum elements), `reached` will be `r` if at all possible. When tracing errors for diagnostics one of these special constraints along a path are treated much like a HTTP redirect: if we are explaining `from: to` and reach an edge with `reaches: invalid` we stop the search and start following `reaches: invalid` instead. When doing this the implicit edges `x: 'static` for every region `x` are ignored, since the search would otherwise be able to cheat by going through 'static and re-find the same edge again. A bunch of optimisations are possible: - Conservatively add constraints, e.g. one per SCC. May worsen error tracing! - as a final pass, allow fusing the annotations for the SCC after adding the extra constraints to remove unnecessary information and save memory. This could be done cheaply since we already iterate over the entire SCC. - currently, if constraints are added the entire set of SCCs are recomputed. This is of course rather wasteful, and we could maybe do better. There are a bunch of rather nice bonuses: - We now don't need to expose region indices in `MirTypeckRegionConstraints` to the entire crate. The only entry point is `placeholder_region()` so correctness of the indices is now guaranteed - A lot of things that were previously iterations over lists is now a single lookup - The constraint graph search functions are simple and at least one of them can now take a proper region as target rather than a predicate function. The only case that needs the predicate argument to `find_constraint_path_to()` is `find_sub_region_live_at()`, which may or may not be possible to work around. r? nikomatsakis

…mpletely, r=<try> [WIP] Remove placeholders completely This PR does shotgun surgery on borrowck to remove all special handling of placeholders, completely replacing them with a preprocessing step that rewrites placeholder leaks into constraints, removing constraint propagation of placeholders and all logic used to detect placeholder violations during error reporting. This finishes what rust-lang#123720 started. The new method works like this: 1. during SCC construction, some information about SCC membership and reachability is retained 2. just after SCC construction, a constraint `r - (from: to_invalid) - > 'static` is added when `r` is the representative of an SCC and 1. that SCC either has had its universe shrunk because it reaches a region with a smaller one (in which case `to_invalid` is the smallest-universed region it reaches), or if it reaches a region with a too large universe that isn't part of the SCC (in which case `to_invalid` is the region with a too large universe). In either case, `from` is also `r`. 2. some region `reaches` in `r`'s SCC reaches another placeholder, `reached`, in which case the added constraint is `r -> (reaches: reached) 'static`. Through clever choice of defaults (chosing minimum elements), `reached` will be `r` if at all possible. When tracing errors for diagnostics one of these special constraints along a path are treated much like a HTTP redirect: if we are explaining `from: to` and reach an edge with `reaches: invalid` we stop the search and start following `reaches: invalid` instead. When doing this the implicit edges `x: 'static` for every region `x` are ignored, since the search would otherwise be able to cheat by going through 'static and re-find the same edge again. A bunch of optimisations are possible: - ~~Conservatively add constraints, e.g. one per SCC. May worsen error tracing!~~ - as a final pass, allow fusing the annotations for the SCC after adding the extra constraints to remove unnecessary information and save memory. This could be done cheaply since we already iterate over the entire set of SCCs. - currently, if constraints are added the entire set of SCCs are recomputed. This is of course rather wasteful, and we could do better. Especially since SCCs are added in dependency order. This would require a fully separate SCC module since the dynamic SCC combo we'd need now shares almost no properties with regular SCC computation. Given that this is meant to be a temporary work-around, that seems like too much work. There are a bunch of rather nice bonuses: - We now don't need to expose region indices in `MirTypeckRegionConstraints` to the entire crate. The only entry point is `placeholder_region()` so correctness of the indices is now guaranteed - A lot of things that were previously iterations over lists is now a single lookup - The constraint graph search functions are simple and at least one of them can now take a proper region as target rather than a predicate function. The only case that needs the predicate argument to `find_constraint_path_to()` is `find_sub_region_live_at()`, which may or may not be possible to work around. r? nikomatsakis

rustbot assigned davidtwco Apr 10, 2024

rustbot added S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. T-compiler Relevant to the compiler team, which will review and decide on the PR/issue. labels Apr 10, 2024