perf: Cache already-checked types in the privacy visitor - #160317

xmakro · 2026-08-01T06:19:36Z

The privacy checker walks the full type of every expression and pattern in a module, re-walking the same type once per node it appears on. This caches the types that walked clean (no privacy error) and skips them next time. A walk's result depends only on the interned type and the fixed module being checked, so a type that walks clean once walks clean everywhere.

Only clean walks are cached, so nothing is lost: a type that errors is never cached and still fires at every span, and no dep-graph edges are dropped since the full walk already ran once in the same query.

petrochenkov · 2026-08-01T11:57:09Z

Something similar was tried in #147486, but the caching overhead was found to be too large in most cases.

petrochenkov · 2026-08-01T11:57:22Z

@bors try @rust-timer queue

petrochenkov · 2026-08-01T11:57:55Z

@bors try @rust-timer queue

Needs a rebase first.

petrochenkov · 2026-08-02T11:43:41Z

@bors try @rust-timer queue

…<try> perf: Cache already-checked types in the privacy visitor

rust-bors · 2026-08-02T13:48:32Z

☀️ Try build successful (CI)
Build commit: 278a694 (278a6943d9a56361f343eea016b348202494ff1a)
Base parent: 8c3a200 (8c3a200b704adb5f110249b68daa4ff386b36e29)

rust-timer · 2026-08-02T14:30:59Z

Finished benchmarking commit (278a694): comparison URL.

Overall result: ✅ improvements - no action needed

Benchmarking means the PR may be perf-sensitive. It's automatically marked not fit for rolling up. Overriding is possible but disadvised: it risks changing compiler perf.

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

Instruction count

Our most reliable metric. Used to determine the overall result above. However, even this metric can be noisy.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.5%	[-1.1%, -0.2%]	55
Improvements ✅ (secondary)	-0.5%	[-0.9%, -0.2%]	67
All ❌✅ (primary)	-0.5%	[-1.1%, -0.2%]	55

Max RSS (memory usage)

Results (primary 0.4%, secondary -0.8%)

A less reliable metric. May be of interest, but not used to determine the overall result above.

	mean	range	count
Regressions ❌ (primary)	2.7%	[2.7%, 2.7%]	1
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.7%	[-0.9%, -0.5%]	2
Improvements ✅ (secondary)	-0.8%	[-2.7%, -0.4%]	13
All ❌✅ (primary)	0.4%	[-0.9%, 2.7%]	3

Cycles

Results (primary 0.0%, secondary 0.7%)

A less reliable metric. May be of interest, but not used to determine the overall result above.

	mean	range	count
Regressions ❌ (primary)	0.8%	[0.4%, 1.7%]	5
Regressions ❌ (secondary)	1.2%	[0.5%, 2.7%]	9
Improvements ✅ (primary)	-0.7%	[-1.0%, -0.5%]	5
Improvements ✅ (secondary)	-0.5%	[-0.7%, -0.5%]	4
All ❌✅ (primary)	0.0%	[-1.0%, 1.7%]	10

Binary size

This perf run didn't have relevant results for this metric.

Bootstrap: 488.148s -> 489.283s (0.23%)
Artifact size: 390.38 MiB -> 390.33 MiB (-0.01%)

petrochenkov · 2026-08-03T13:33:12Z

I see how the type caching can be useful, but I'm not sure how much the whole GenericArgs caching contributes to the results.
Could we benchmark them separately, starting with the types?
@rustbot author

rustbot · 2026-08-03T13:33:16Z

Reminder, once the PR becomes ready for a review, use @rustbot ready.

xmakro · 2026-08-03T18:08:54Z

You are right, I measured a bit more and the majority of the win is type caching. I updated the PR to type cache only.

petrochenkov · 2026-08-03T18:36:44Z

@bors try @rust-timer queue

…<try> perf: Cache already-checked types in the privacy visitor

rust-bors · 2026-08-03T20:42:41Z

☀️ Try build successful (CI)
Build commit: 5b9f7d9 (5b9f7d9cc424d1f848e527a01fff9f8a090d64d0)
Base parent: 7c329d6 (7c329d6c76e11ca40c5673818ab0439c1be8962c)

rust-timer · 2026-08-03T21:28:28Z

Finished benchmarking commit (5b9f7d9): comparison URL.

Overall result: ✅ improvements - no action needed

Benchmarking means the PR may be perf-sensitive. It's automatically marked not fit for rolling up. Overriding is possible but disadvised: it risks changing compiler perf.

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

Instruction count

Our most reliable metric. Used to determine the overall result above. However, even this metric can be noisy.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.5%	[-1.1%, -0.2%]	58
Improvements ✅ (secondary)	-0.4%	[-0.9%, -0.2%]	71
All ❌✅ (primary)	-0.5%	[-1.1%, -0.2%]	58

Max RSS (memory usage)

Results (primary -0.9%, secondary -0.7%)

A less reliable metric. May be of interest, but not used to determine the overall result above.

	mean	range	count
Regressions ❌ (primary)	0.4%	[0.4%, 0.5%]	2
Regressions ❌ (secondary)	0.5%	[0.4%, 0.6%]	3
Improvements ✅ (primary)	-3.6%	[-3.6%, -3.6%]	1
Improvements ✅ (secondary)	-1.9%	[-4.7%, -0.5%]	3
All ❌✅ (primary)	-0.9%	[-3.6%, 0.5%]	3

Cycles

Results (primary -0.3%, secondary -0.2%)

A less reliable metric. May be of interest, but not used to determine the overall result above.

	mean	range	count
Regressions ❌ (primary)	0.8%	[0.5%, 1.0%]	4
Regressions ❌ (secondary)	1.2%	[0.7%, 1.6%]	4
Improvements ✅ (primary)	-1.4%	[-2.4%, -0.6%]	4
Improvements ✅ (secondary)	-1.2%	[-3.6%, -0.4%]	5
All ❌✅ (primary)	-0.3%	[-2.4%, 1.0%]	8

Binary size

This perf run didn't have relevant results for this metric.

Bootstrap: 491.018s -> 490.335s (-0.14%)
Artifact size: 390.29 MiB -> 390.25 MiB (-0.01%)

petrochenkov · 2026-08-04T10:22:51Z

Mostly the same as with GenericArgs.
@bors r+

Could you make a PR with GenericArgs caching once this one merges, we'll benchmark it separately and land it too if it's above the noise.

rust-bors · 2026-08-04T10:22:54Z

📌 Commit da4563a has been approved by petrochenkov

It is now in the queue for this repository.

🌲 The tree is currently closed for pull requests below priority 10. This pull request will be tested once the tree is reopened.

Reason for tree closure: manually handling queue due to backlog

JonathanBrouwer · 2026-08-04T12:40:05Z

@bors rollup=iffy
Including this in a special perf-sensitive rollup

…uwer Rollup of 10 perf-sensitive pull requests Successful merges: - #157281 (perf: skip irrelevant foreign impls when building the specialization graph) - #159403 (Next steps for FnDef binder changes (instantiate most FnDef binders)) - #159763 (Optimize crate resolution for large workspace) - #160033 (Speed up `EverInitializedPlaces`) - #160268 (perf: store the fulfillment engine inline in ObligationCtxt) - #160317 (perf: Cache already-checked types in the privacy visitor) - #160399 (interpret: skip deref-projection validity checks when they are not needed) - #160451 (Deduplicate target and host filesearch) - #160453 (Add fast path to `escape_string_symbol`) - #160454 (Add offload guard flags to typeck to prevent perf regressions)

Rollup merge of #160317 - xmakro:perf/privacy-accessible-type-cache, r=petrochenkov perf: Cache already-checked types in the privacy visitor The privacy checker walks the full type of every expression and pattern in a module, re-walking the same type once per node it appears on. This caches the types that walked clean (no privacy error) and skips them next time. A walk's result depends only on the interned type and the fixed module being checked, so a type that walks clean once walks clean everywhere. Only clean walks are cached, so nothing is lost: a type that errors is never cached and still fires at every span, and no dep-graph edges are dropped since the full walk already ran once in the same query.

JonathanBrouwer · 2026-08-05T06:40:47Z

Verifying that actual perf results after merge match expected results
@rust-timer build 76908bc

…uwer Rollup of 10 perf-sensitive pull requests Successful merges: - rust-lang/rust#157281 (perf: skip irrelevant foreign impls when building the specialization graph) - rust-lang/rust#159403 (Next steps for FnDef binder changes (instantiate most FnDef binders)) - rust-lang/rust#159763 (Optimize crate resolution for large workspace) - rust-lang/rust#160033 (Speed up `EverInitializedPlaces`) - rust-lang/rust#160268 (perf: store the fulfillment engine inline in ObligationCtxt) - rust-lang/rust#160317 (perf: Cache already-checked types in the privacy visitor) - rust-lang/rust#160399 (interpret: skip deref-projection validity checks when they are not needed) - rust-lang/rust#160451 (Deduplicate target and host filesearch) - rust-lang/rust#160453 (Add fast path to `escape_string_symbol`) - rust-lang/rust#160454 (Add offload guard flags to typeck to prevent perf regressions)

rust-timer · 2026-08-05T11:47:04Z

Finished benchmarking commit (76908bc): comparison URL.

Overall result: ✅ improvements - no action needed

Benchmarking means the PR may be perf-sensitive. It's automatically marked not fit for rolling up. Overriding is possible but disadvised: it risks changing compiler perf.

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

Instruction count

Our most reliable metric. Used to determine the overall result above. However, even this metric can be noisy.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.5%	[-1.1%, -0.2%]	57
Improvements ✅ (secondary)	-0.4%	[-0.9%, -0.1%]	72
All ❌✅ (primary)	-0.5%	[-1.1%, -0.2%]	57

Max RSS (memory usage)

Results (primary -0.7%, secondary -2.0%)

A less reliable metric. May be of interest, but not used to determine the overall result above.

	mean	range	count
Regressions ❌ (primary)	0.5%	[0.5%, 0.5%]	2
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-3.1%	[-3.1%, -3.1%]	1
Improvements ✅ (secondary)	-2.0%	[-3.9%, -0.6%]	5
All ❌✅ (primary)	-0.7%	[-3.1%, 0.5%]	3

Cycles

Results (primary 0.3%, secondary -0.1%)

A less reliable metric. May be of interest, but not used to determine the overall result above.

	mean	range	count
Regressions ❌ (primary)	0.9%	[0.4%, 1.6%]	6
Regressions ❌ (secondary)	0.7%	[0.4%, 1.0%]	6
Improvements ✅ (primary)	-0.7%	[-1.3%, -0.5%]	4
Improvements ✅ (secondary)	-0.9%	[-2.2%, -0.4%]	6
All ❌✅ (primary)	0.3%	[-1.3%, 1.6%]	10

Binary size

This perf run didn't have relevant results for this metric.

Bootstrap: 489.577s -> 496.509s (1.42%)
Artifact size: 390.24 MiB -> 390.30 MiB (0.02%)

xmakro · 2026-08-05T15:25:46Z

Could you make a PR with GenericArgs caching once this one merges, we'll benchmark it separately and land it too if it's above the noise.

I benchmarked the GenericArgs cache on top of the merged type cache with a local rustc-perf run. The result was neutral, nothing moved significantly. Most of what the cache hits is the interned empty GenericArgs list, whose walk is nearly free. So this is not worth a PR. Thanks for spotting this in the original PR.

rustbot added S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. T-compiler Relevant to the compiler team, which will review and decide on the PR/issue. labels Aug 1, 2026

petrochenkov self-assigned this Aug 1, 2026