Opt-in to `FulfillmentError` generation to avoid doing extra work in the new solver #125864

compiler-errors · 2024-06-01T19:13:50Z

In the new solver, we do additional trait solving in order to generate fulfillment errors, because all we have is the root obligation. This is problematic because there are many cases where we don't need the full error information, and instead are just calling ObligationCtxt::select_all_or_error to probe whether a predicate holds or not. This is also problematic because we use ObligationCtxts within the error reporting machinery itself, and so we can definitely cause stack overflows:

rust/compiler/rustc_trait_selection/src/solve/inspect/analyse.rs

Lines 75 to 84 in a94483a

    
           let ocx = ObligationCtxt::new(infcx); 
        
           ocx.register_obligations(obligations); 
        
           let errors = ocx.select_all_or_error(); 
        
           if errors.is_empty() { 
        
               Ok(Certainty::Yes) 
        
           } else if errors.iter().all(|e| !e.is_true_error()) { 
        
               Ok(Certainty::AMBIGUOUS) 
        
           } else { 
        
               Err(NoSolution) 
        
           }

So instead, make TraitEngine and ObligationCtxt generic over E: FulfillmentErrorLike<'tcx>, and introduce a new ScrubbedTraitError which only stores whether the failure was due to a "true error" or an ambiguity. Then, introduce ObligationCtxt::new_with_diagnostics for the callsites that actually inspect their FulfillmentErrors.

r? @lcnr

Number-wise, there are:

     39 ObligationCtxt::new
     32 ObligationCtxt::new_with_diagnostics
      1 ObligationCtxt::new_generic

calls to each ObligationCtxt constructor, which suggests that there are indeed a lot of callsites that don't care about diagnostics.

rustbot · 2024-06-01T19:13:59Z

Some changes occurred in engine.rs, potentially modifying the public API of ObligationCtxt.

cc @lcnr, @compiler-errors

Some changes occurred to the core trait solver

cc @rust-lang/initiative-trait-system-refactor

Some changes occurred in src/tools/clippy

cc @rust-lang/clippy

Some changes occurred to MIR optimizations

cc @rust-lang/wg-mir-opt

HIR ty lowering was modified

cc @fmease

compiler-errors · 2024-06-01T19:14:10Z

@bors try @rust-timer queue

compiler-errors · 2024-06-01T19:14:30Z

I don't expect this to be much of an optimization in the old solver, but worth checking I guess :^)

bors · 2024-06-01T19:15:20Z

⌛ Trying commit dee3d9a with merge c639bce...

…ng, r=<try> Opt-in to `FulfillmentError` generation to avoid doing extra work in the new solver In the new solver, we do additional trait solving in order to generate fulfillment errors, because all we have is the root obligation. This is problematic because there are many cases where we don't need the full error information, and instead are just calling `ObligationCtxt::select_all_or_error` to probe whether a predicate holds or not. This is also problematic because we use `ObligationCtxt`s within the error reporting machinery itself, and so we can definitely cause stack overflows: https://github.com/rust-lang/rust/blob/a94483a5f2bae907bc898fc7a8d9cc87db47b693/compiler/rustc_trait_selection/src/solve/inspect/analyse.rs#L75-L84 So instead, make `TraitEngine` and `ObligationCtxt` generic over `E: FulfillmentErrorLike<'tcx>`, and introduce a new `ScrubbedTraitError` which only stores whether the failure was due to a "true error" or an ambiguity. Then, introduce `ObligationCtxt::new_with_diagnostics` for the callsites that actually inspect their `FulfillmentError`s. r? `@lcnr`

fmease · 2024-06-02T20:39:27Z

Last try-build has never finished.

@bors try @rust-timer queue

compiler-errors · 2024-06-02T20:40:25Z

ty @fmease

bors · 2024-06-02T20:40:38Z

⌛ Trying commit bb84313 with merge 6c3f703...

…ng, r=<try> Opt-in to `FulfillmentError` generation to avoid doing extra work in the new solver In the new solver, we do additional trait solving in order to generate fulfillment errors, because all we have is the root obligation. This is problematic because there are many cases where we don't need the full error information, and instead are just calling `ObligationCtxt::select_all_or_error` to probe whether a predicate holds or not. This is also problematic because we use `ObligationCtxt`s within the error reporting machinery itself, and so we can definitely cause stack overflows: https://github.com/rust-lang/rust/blob/a94483a5f2bae907bc898fc7a8d9cc87db47b693/compiler/rustc_trait_selection/src/solve/inspect/analyse.rs#L75-L84 So instead, make `TraitEngine` and `ObligationCtxt` generic over `E: FulfillmentErrorLike<'tcx>`, and introduce a new `ScrubbedTraitError` which only stores whether the failure was due to a "true error" or an ambiguity. Then, introduce `ObligationCtxt::new_with_diagnostics` for the callsites that actually inspect their `FulfillmentError`s. r? `@lcnr`

bors · 2024-06-02T22:19:00Z

☀️ Try build successful - checks-actions
Build commit: 6c3f703 (6c3f703257411e9f17acbd032b1b15b1f8d4ad44)

rust-timer · 2024-06-03T00:22:56Z

Finished benchmarking commit (6c3f703): comparison URL.

Overall result: ❌ regressions - no 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.

@bors rollup=never
@rustbot label: -S-waiting-on-perf -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)	0.2%	[0.2%, 0.2%]	1
Regressions ❌ (secondary)	1.7%	[1.7%, 1.7%]	1
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	0.2%	[0.2%, 0.2%]	1

Max RSS (memory usage)

Results (secondary 2.1%)

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)	2.1%	[2.1%, 2.1%]	1
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-	-	0

Cycles

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

Binary size

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

Bootstrap: 669.121s -> 673.78s (0.70%)
Artifact size: 318.92 MiB -> 319.07 MiB (0.05%)

lcnr · 2024-06-03T09:35:44Z

compiler/rustc_borrowck/src/diagnostics/conflict_errors.rs

@@ -1347,7 +1347,7 @@ impl<'cx, 'tcx> MirBorrowckCtxt<'cx, 'tcx> {
            return;
        };
        // Try to find predicates on *generic params* that would allow copying `ty`
-        let ocx = ObligationCtxt::new(self.infcx);
+        let ocx = ObligationCtxt::new_with_diagnostics(self.infcx);


we use new_with_diagnostics because we care about leaves instead of the root goal?

Yes, I believe so.

compiler/rustc_borrowck/src/type_check/constraint_conversion.rs

compiler/rustc_hir_analysis/src/coherence/orphan.rs

lcnr · 2024-06-03T09:40:00Z

compiler/rustc_infer/src/infer/canonical/query_response.rs

        &self,
        inference_vars: CanonicalVarValues<'tcx>,
        answer: T,
-        fulfill_cx: &mut dyn TraitEngine<'tcx>,
+        fulfill_cx: &mut dyn TraitEngine<'tcx, E>,


shouldn't this always use scrubbed errors? we never emit errors inside of a canonical query after all 🤔

This was removed in a follow-up I believe

Wait no, it's because ScrubbedTraitError is defined in a downstream crate (rustc_trait_selection). And pulling it up to rustc_infer is a bit annoying, because then the impl FromFulfillmentError for OldSolverError is foreign because that type technically comes from rustc_data_structures, but I guess I could do it.

compiler/rustc_infer/src/infer/canonical/query_response.rs

lcnr · 2024-06-03T09:41:56Z

compiler/rustc_infer/src/traits/engine.rs

+pub trait FulfillmentErrorLike<'tcx>: Debug + 'tcx {
+    fn is_true_error(&self) -> bool;
+}


do we actually need this, i'd expect every user of is_true_error to use the concrete "non-diagnostics` errors

We use is_true_error for both scrubbed and real errors, since it's useful to avoid needing to call select_where_possible then select_all_or_error.

We could not do that, but it would mean there are far more places where we need to call new_with_diagnostics for little benefit.

compiler/rustc_trait_selection/src/solve/normalize.rs

compiler/rustc_trait_selection/src/traits/engine.rs

compiler/rustc_trait_selection/src/traits/normalize.rs

compiler/rustc_trait_selection/src/traits/query/type_op/implied_outlives_bounds.rs

lcnr

i like this change 😊 sorry for all the formatting nits xx

r=me after nits

…to new_with_diagnostics

compiler-errors · 2024-06-03T14:37:54Z

@bors r=lcnr

bors · 2024-06-03T14:37:57Z

📌 Commit a41c44f has been approved by lcnr

It is now in the queue for this repository.

bors · 2024-06-03T15:43:09Z

⌛ Testing commit a41c44f with merge eb5e244...

bors · 2024-06-03T18:09:26Z

☀️ Test successful - checks-actions
Approved by: lcnr
Pushing eb5e244 to master...

rust-timer · 2024-06-03T20:06:35Z

Finished benchmarking commit (eb5e244): comparison URL.

Overall result: no relevant changes - no action needed

@rustbot label: -perf-regression

Instruction count

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

Max RSS (memory usage)

Results (primary 3.7%, secondary 2.9%)

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)	3.7%	[2.8%, 4.7%]	2
Regressions ❌ (secondary)	2.9%	[2.9%, 2.9%]	1
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	3.7%	[2.8%, 4.7%]	2

Cycles

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

Binary size

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

Bootstrap: 667.912s -> 672.681s (0.71%)
Artifact size: 318.98 MiB -> 318.96 MiB (-0.01%)

rustbot assigned lcnr Jun 1, 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. WG-trait-system-refactor The Rustc Trait System Refactor Initiative (-Znext-solver) labels Jun 1, 2024

This comment has been minimized.

Sign in to view

rustbot added the S-waiting-on-perf Status: Waiting on a perf run to be completed. label Jun 1, 2024

compiler-errors force-pushed the opt-in-error-reporting branch from dee3d9a to bb84313 Compare June 1, 2024 19:18

This comment has been minimized.

Sign in to view

compiler-errors force-pushed the opt-in-error-reporting branch from e8c972e to e8fe56c Compare June 3, 2024 00:05

rustbot removed the S-waiting-on-perf Status: Waiting on a perf run to be completed. label Jun 3, 2024