Reduce memory usage for Path values with a single Ident, or where we can reconstruct the Span by joshtriplett · Pull Request #163744 · rust-lang/rust

joshtriplett · 2026-10-04T09:18:16Z

Make Path an enum, with a variant for just a single Ident (the common case).

The majority of Path values in the compiler just need a single Ident, with no arguments, and no separate Span differing from the one in the Ident. For instance, local variables, item names, argument names, and so on.

However, Path stored them as a span and a pointer to a vector (including len/capacity) of PathSegment structs, each containing an Ident, a NodeId, and an Option<Box<GenericArgs>>. And, to make it even worse, these typically have allocated capacity for four PathSegment structs despite only having one. So, in total, a one-Ident Path took up:

16 bytes for a Span and ThinVec pointer
16 bytes for the ThinVec length and capacity
4 PathSegments, each 24 bytes,
for a total of 128 bytes, not counting any allocator overhead.

On large crates like aws-sdk-ec2, this can be a substantial fraction of the memory usage of the AST. (And the HIR, but this commit doesn't try to deal with that yet.)

Turn Path into an enum, with one variant Path::Ident for the single-Ident no-args case, and the other variant Path::General for any case with multiple segments, zero segments, any generic arguments, or a Span that doesn't match the Ident.

For aws-sdk-ec2 (release-2026-10-02), 55% of all Path values (806434/1460321) can use Path::Ident.

In order to keep Path the same size (16 bytes) and not grow all the structures containing it, also avoid storing a Span for Path values where we can reconstruct it from the segments.

If the span we would have stored matches the span from the first segment to the last, use a Path::NoSpan variant that just stores the segments.

This works for the vast majority of Path values that otherwise used Path::General. We can then box the ones that remain Path::General, in order to keep Path 16 bytes and avoid growing it or the structures that contain it.

For aws-sdk-ec2 (release-2026-10-02), after the 55% of Path values that can use Path::Ident, another 42% (615812/1460321) can use Path::NoSpan, leaving less than 3% (38075/1460321) that still need Path::General.

Add input-stats measurement of Path distributions, to collect this data:

ast-stats - Path Ident: 806434  Other segs 0: 0  1: 41816  2: 47050  3: 130254  4+: 434767  NoSpan: 615812

100% human-written code.

…case) The majority of Path values in the compiler just need a single Ident, with no arguments, and no separate Span differing from the one in the Ident. For instance, local variables, item names, argument names, and so on. However, Path stored them as a span and a pointer to a vector (including len/capacity) of PathSegment structs, each containing an Ident, a NodeId, and an Option<Box<GenericArgs>>. And, to make it even worse, these typically have allocated capacity for *four* PathSegment structs despite only having *one*. So, in total, a one-Ident Path took up: - 16 bytes for a Span and ThinVec pointer - 16 bytes for the ThinVec length and capacity - 4 PathSegments, each 24 bytes, for a total of 128 bytes, not counting any allocator overhead. On large crates like aws-sdk-ec2, this can be a substantial fraction of the memory usage of the AST. (And the HIR, but this commit doesn't try to deal with that yet.) Turn Path into an enum, with one variant `Path::Ident` for the single-Ident no-args case, and the other variant `Path::General` for any case with multiple segments, zero segments, any generic arguments, or a span that doesn't match the ident. For aws-sdk-ec2 (release-2026-10-02), 55% of all Path values (806434/1460321) can use Path::Ident. This commit *temporarily* increases the size of Path to 24 bytes; a subsequent commit will re-shrink it to 16 bytes.

… the segments If the span we would have stored matches the span from the first segment to the last, use a `Path::NoSpan` variant that just stores the segments. This works for the vast majority of Path values that otherwise used `Path::General`. We can then box the ones that remain `Path::General`, in order to keep `Path` 16 bytes and avoid growing it or the structures that contain it.

joshtriplett · 2026-10-04T09:18:45Z

@bors try @rust-timer queue

Reduce memory usage for Path values with a single Ident, or where we can reconstruct the Span

joshtriplett · 2026-10-04T09:51:28Z

(I can fix up clippy and other things after this gets a perf report.)

rust-bors · 2026-10-04T10:22:36Z

☀️ Try build successful (CI)
Build commit: 9fc4eb4 (9fc4eb4b9dc2f54609bbd11ba261542f1142a8a8)
Base parent: 4ddbc06 (4ddbc06ea09abcda34f80baf31cc9bc2686b0ae6)

rust-timer · 2026-10-04T11:06:00Z

Finished benchmarking commit (9fc4eb4): comparison URL.

Overall result: ❌✅ regressions and improvements - please read:

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.

Next, please: If you can, justify the regressions found in this try perf run in writing along with @rustbot label: +perf-regression-triaged. If not, fix the regressions and do another perf run. Neutral or positive results will clear the label automatically.

@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)	0.6%	[0.1%, 2.0%]	194
Regressions ❌ (secondary)	0.6%	[0.1%, 2.0%]	110
Improvements ✅ (primary)	-0.3%	[-0.4%, -0.2%]	4
Improvements ✅ (secondary)	-0.4%	[-1.0%, -0.1%]	16
All ❌✅ (primary)	0.5%	[-0.4%, 2.0%]	198

Max RSS (memory usage)

Results (primary -1.3%, secondary -3.9%)

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

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-1.3%	[-2.0%, -1.0%]	5
Improvements ✅ (secondary)	-3.9%	[-5.3%, -2.1%]	13
All ❌✅ (primary)	-1.3%	[-2.0%, -1.0%]	5

Cycles

Results (primary -0.1%, secondary -1.1%)

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

	mean	range	count
Regressions ❌ (primary)	3.1%	[2.8%, 3.4%]	2
Regressions ❌ (secondary)	2.3%	[2.2%, 2.6%]	3
Improvements ✅ (primary)	-2.3%	[-2.9%, -1.5%]	3
Improvements ✅ (secondary)	-2.9%	[-4.5%, -2.2%]	6
All ❌✅ (primary)	-0.1%	[-2.9%, 3.4%]	5

Binary size

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

Bootstrap: 491.424s -> 491.272s (-0.03%)
Artifact size: 408.69 MiB -> 408.85 MiB (0.04%)

joshtriplett · 2026-10-04T17:36:09Z

Well, great improvement on Max RSS. Had hoped it would be neutral or better on performance, but it's definitely a tiny hit. Will investigate and see if I can improve it.

…ctually erroring lower_import_res wanted the Span but only used it to report an error; pass in the &Path instead and defer calling `.span()`.

joshtriplett · 2026-10-04T19:16:28Z

I've already found a number of wins to avoid the perf hits; working on those now.

We don't need to call `self.hi_span()` in this case because we know it'll be `self.prefix.span()`.

… actually erroring

Avoid doing `path.span().with_hi(path.span().hi())` along some paths.

This provides a non-trivial performance improvement.

joshtriplett · 2026-10-05T12:48:38Z

I've eliminated most of the regressions. rayon-rs/either#146 and rayon-rs/either#147 will help once available, as will mozilla/thin-vec#99 .

joshtriplett · 2026-10-06T00:43:26Z

Several additional potential wins worth exploring after this PR, which I'd like to avoid stacking on top because they'd make it larger to review:

Add a fast path to parsing that avoids allocating a ThinVec of segments in the common case of a single identifier. This would require unrolling one iteration of parse_path_segments into parse_path_inner, creating a Path::Ident in that case, or continuing on to accumulate the remaining segments if there are more. The primary complication here, and the reason I didn't go on to do it directly in this PR, is that this also needs to handle the recovery cases (or at least detect and punt them to parse_path_segments).
Make Path generic, so it also supports a QPath that stores a QSelf directly in Path::General. This would shrink all the structures that keep a QSelf side-by-side with a Path. Any Path associated with a QSelf will use Path::General.
Make UsePathList store &Path rather than &[PathSegment], so that creating a UsePathList doesn't require forcing a less ideal Path representation. AFAICT, UsePathList uses that for diagnostics, but currently always has to obtain it in order to reference it, even in builds with no diagnostics.
Optimizing HIR paths. These don't have excess capacity (unlike AST paths), because they use an arena slice rather than a ThinVec, but they could still store the simple ident case much more efficiently.

…_ident` No expected performance impact (just printing code), just a cleanup.

rustbot · 2026-10-06T01:09:32Z

Changes to the size of AST and/or HIR nodes.

cc @nnethercote

The parser was modified, potentially altering the grammar of (stable) Rust
which would be a breaking change.

cc @fmease

rustfmt is developed in its own repository. If possible, consider making this change to rust-lang/rustfmt instead.

cc @rust-lang/rustfmt

Some changes occurred to diagnostic attributes.

cc @mejrs

Some changes occurred in compiler/rustc_builtin_macros/src/autodiff.rs

cc @ZuseZ4

These commits modify the Cargo.lock file. Unintentional changes to Cargo.lock can be introduced when switching branches and rebasing PRs.

If this was unintentional then you should revert the changes before this PR is merged.
Otherwise, you can ignore this comment.

Some changes occurred in compiler/rustc_attr_ir

cc @jdonszelmann, @JonathanBrouwer

clippy is developed in its own repository. If possible, consider making this change to rust-lang/rust-clippy instead.

cc @rust-lang/clippy

Some changes occurred in compiler/rustc_attr_parsing

cc @jdonszelmann, @JonathanBrouwer

rustbot · 2026-10-06T01:09:35Z

r? @camelid

rustbot has assigned @camelid.
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

Why was this reviewer chosen?

The reviewer was selected based on:

Owners of files modified in this PR: compiler
compiler expanded to 77 candidates
Random selection from 18 candidates

joshtriplett · 2026-10-06T01:09:41Z

r? @nnethercote

joshtriplett · 2026-10-06T01:15:00Z

I'd recommend reviewing commit by commit. The first three commits are the bulk of the semantic change, which converts callers very mechanically to make sure the conversion is always obvious. The rest is optimization to avoid performance regression, and the optimizations tend to be pretty self-contained and obvious.

nnethercote

The code changes seem fine; very tedious but quite mechanical. A couple of nits below.

My main concern is the perf results don't yet seem to justify the effort. Instruction counts get a bit worse, memory usage improves a little but not that much. If any of the pending future improvements are likely to make a big difference, it would be good to see them here as well.

View changes since this review

nnethercote · 2026-10-06T05:18:15Z

-            segment.ident.stable_hash(hcx, hasher);
-        }
+        self.num_segments().stable_hash(hcx, hasher);
+        self.iter_idents().for_each(|ident| ident.stable_hash(hcx, hasher));


Pre-existing, but it seems weird/wrong that args isn't hashed!

@nnethercote From what I can tell, this is because hashing for AST paths was introduced for paths used in attributes, which can never have generic args. It's a hazard that would be an issue if those hashes were used anywhere else, though. I'll send a separate PR fixing it.

nnethercote · 2026-10-06T05:30:23Z

 ast-stats ----------------------------------------------------------------
 ast-stats Total                  6_624                   109
+ast-stats ----------------------------------------------------------------
+ast-stats - Path Ident: 17  General segs 0: 0  1: 1  2: 2  3: 3  4+: 1  Span reconstructible: 7


This line doesn't really fit. Should it remain temporary, local-only code?

I can easily drop it if undesired. I think there's value in having these statistics readily available as part of the AST size stats. I could probably manage to fit it into another part of the input-stats output and/or in another form, if that would help.

As the author of the AST stats code... I suspect the value of these statistics will recede rapidly once you finish working on this stuff :)

Currently we have the count/byte measurements in a particular table form, with labelled columns, and a total line. And then there's this single line of different non-byte measurements tacked on the end, in a form that sort of mirrors the table but is also different. It looks like there's a bug in the table-writing code. If it was in a more clearly separated and labelled "miscelleneous stats" table that would be much better.

@nnethercote That's completely fair. I'll turn it into a proper separate table, and make it a little less pithily abbreviated.

… segments

petrochenkov · 2026-10-06T09:20:26Z

I'd really prefer not doing this unless there's a very clear perf benefit.

joshtriplett · 2026-10-06T15:08:37Z

I'll work on some additional things atop this to show a more definitive win.

rust-bors · 2026-10-07T20:47:45Z

☔ The latest upstream changes (presumably #163945) made this pull request unmergeable. Please resolve the merge conflicts by rebasing.

joshtriplett added 3 commits October 4, 2026 02:04

Track path sizes in input-stats

205f029

rustbot added A-attributes Area: Attributes (`#[…]`, `#![…]`) 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 Oct 4, 2026

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 Oct 4, 2026

This comment has been minimized.

Sign in to view

rust-bors Bot pushed a commit that referenced this pull request Oct 4, 2026

Auto merge of #163744 - joshtriplett:path-unsegment, r=<try>

9fc4eb4

Reduce memory usage for Path values with a single Ident, or where we can reconstruct the Span

This comment has been minimized.

Sign in to view

rustbot added perf-regression Performance regression. and removed S-waiting-on-perf Status: Waiting on a perf run to be completed. labels Oct 4, 2026

joshtriplett added 4 commits October 4, 2026 12:08

rustc_ast_lowering: Cache computed path spans

cd21804

Inline segments_span

cb76f45

rustc_ast_lowering: Cache spans in restriction lowering

c0b3c9e

rustc_ast_lowering: Defer calling .span() for an error path until a…

ab3bb3f

…ctually erroring lower_import_res wanted the Span but only used it to report an error; pass in the &Path instead and defer calling `.span()`.

joshtriplett added 6 commits October 4, 2026 12:21

rustc_resolve: Don't recompute use_tree.span() in a loop

230042e

UseTree::span(): Add a fast path for UseTreeKind::Simple(None)

b3e1209

We don't need to call `self.hi_span()` in this case because we know it'll be `self.prefix.span()`.

report_path_args: Defer error span computation until actually erroring

30a9c82

check_stability_and_deprecation: Defer error span computation until…

f16bcbc

… actually erroring

MetaItem::from_tokens: Avoid redundant span calculations

bec0098

Avoid doing `path.span().with_hi(path.span().hi())` along some paths.

Inline Segment::from

ec9c4a6

This provides a non-trivial performance improvement.

rustbot added the T-rustfmt Relevant to the rustfmt team, which will review and decide on the PR/issue. label Oct 5, 2026

This comment has been minimized.

Sign in to view

Fix rustfmt for changes to Path

93bc654

joshtriplett force-pushed the path-unsegment branch from 8f87557 to 93bc654 Compare October 5, 2026 20:28

joshtriplett mentioned this pull request Oct 5, 2026

Add #[inline] to Either::fold rayon-rs/either#147

Merged

This comment has been minimized.

Sign in to view

Clean up rustc_hir_pretty::print_attr_item using `PathSegment::from…

b12956f

…_ident` No expected performance impact (just printing code), just a cleanup.

joshtriplett force-pushed the path-unsegment branch from a99858e to b12956f Compare October 6, 2026 01:03

joshtriplett marked this pull request as ready for review October 6, 2026 01:09

rustbot added S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. and removed S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. labels Oct 6, 2026

rustbot assigned camelid Oct 6, 2026

rustbot assigned nnethercote and unassigned camelid Oct 6, 2026

nnethercote reviewed Oct 6, 2026

View reviewed changes

joshtriplett added 3 commits October 6, 2026 00:40

rustc_ast: Path::is_empty: Add explanations for NoSpan and Ident

e0c81d4

rustc_ast: Update Path documentation, discuss non-canonical forms

3ef66bd

rustc_ast: Add an explanation for why Path's span may differ from its…

410f4f5

… segments

nnethercote added S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Oct 8, 2026

petrochenkov self-assigned this Oct 8, 2026

Conversation

joshtriplett commented Oct 4, 2026 • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

joshtriplett commented Oct 4, 2026

Uh oh!

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

joshtriplett commented Oct 4, 2026

Uh oh!

rust-bors Bot commented Oct 4, 2026

Uh oh!

This comment has been minimized.

rust-timer commented Oct 4, 2026

Overall result: ❌✅ regressions and improvements - please read:

Uh oh!

joshtriplett commented Oct 4, 2026

Uh oh!

joshtriplett commented Oct 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

joshtriplett commented Oct 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

This comment has been minimized.

joshtriplett commented Oct 6, 2026

Uh oh!

This comment has been minimized.

rustbot commented Oct 6, 2026

Uh oh!

rustbot commented Oct 6, 2026

Uh oh!

joshtriplett commented Oct 6, 2026

Uh oh!

joshtriplett commented Oct 6, 2026

Uh oh!

nnethercote left a comment • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

petrochenkov commented Oct 6, 2026

Uh oh!

joshtriplett commented Oct 6, 2026

Uh oh!

rust-bors Bot commented Oct 7, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

joshtriplett commented Oct 4, 2026 •

edited by rustbot

Loading

joshtriplett commented Oct 4, 2026 •

edited

Loading

joshtriplett commented Oct 5, 2026 •

edited

Loading

nnethercote left a comment •

edited by rustbot

Loading