Skip to content

perf(semantic): flatten hoisting_variables to avoid per-scope map allocation#23927

Merged
Dunqing merged 1 commit into
oxc-project:mainfrom
linyiru:perf/semantic-flatten-hoisting-variables
Jun 29, 2026
Merged

perf(semantic): flatten hoisting_variables to avoid per-scope map allocation#23927
Dunqing merged 1 commit into
oxc-project:mainfrom
linyiru:perf/semantic-flatten-hoisting-variables

Conversation

@linyiru

@linyiru linyiru commented Jun 29, 2026

Copy link
Copy Markdown
Contributor

Summary

SemanticBuilder::hoisting_variables is a nested map
FxHashMap<ScopeId, IdentHashMap<'a, SymbolId>>. Each entry(scope_id).or_default()
on a vacant scope heap-allocates a fresh inner IdentHashMap via the system
allocator, and subsequent inserts grow it. On antd.js this is the single largest
system-allocation site in semantic analysis (VariableDeclarator::bind /
Function::bind in binder.rs).

This flattens it to a single FxHashMap<(ScopeId, Ident<'a>), SymbolId>:

  • one insert / one point-probe per hoisted variable (was: outer probe → inner-map
    allocation → inner probe);
  • no per-scope inner map, so the per-scope system allocations disappear;
  • one hash instead of two.

Behaviour is identical: hoisting_variables is only ever inserted into (binder.rs)
and point-queried in check_redeclaration (builder.rs) — it is never iterated, so a
flat (scope, name) key is a drop-in.

Results — cargo allocs (system allocations)

fixture phase before after
antd.js semantic 1020 28 (−97%)
antd.js minifier 3041 1040 (−66%, rebuilds semantic internally)
pdf.mjs semantic 25 16
kitchen-sink.tsx semantic 41 36

allocs_semantic.snap / allocs_minifier.snap are updated accordingly. This is on the
hot path for tools that run SemanticBuilder repeatedly per module (e.g. Rolldown).

Relation to #19134

#19134 previously tried to move this temporary state (including hoisting_variables)
to a scratch arena; CodSpeed reported a ~3.6% regression and it was closed. This PR
takes a different, minimal approach: it does not introduce an arena, does not
change SemanticBuilder::new's signature, and touches no call sites. It removes a hash
and an allocation per hoisted variable (rather than adding arena threading), so it
should be CodSpeed-neutral-or-better — please let CI confirm.

Tests

  • cargo test -p oxc_semantic --lib --tests — 60 passed (incl. redeclaration / var-hoist tests)
  • cargo test -p oxc_minifier --lib --tests — 538 passed

AI disclosure

Prepared with AI assistance (Claude). I reviewed the change, located the hotspot with a
per-call-site allocation profile, verified the reduction with cargo allocs, and ran
the test suites above.

@codspeed-hq

codspeed-hq Bot commented Jun 29, 2026

Copy link
Copy Markdown

Merging this PR will not alter performance

✅ 62 untouched benchmarks
⏩ 9 skipped benchmarks1


Comparing linyiru:perf/semantic-flatten-hoisting-variables (6e9cd15) with main (dbf2719)2

Open in CodSpeed

Footnotes

  1. 9 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports.

  2. No successful run was found on main (eba0cf5) during the generation of this report, so dbf2719 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report.

@linyiru linyiru marked this pull request as ready for review June 29, 2026 06:49
…ocation

`hoisting_variables` was a nested `FxHashMap<ScopeId, IdentHashMap<'a, SymbolId>>`,
so `entry(scope_id).or_default()` heap-allocated a fresh inner `IdentHashMap` (system
allocator) for every scope that hoists a variable. On antd.js this was the single
largest system-allocation site in semantic analysis.

Flatten to `FxHashMap<(ScopeId, Ident<'a>), SymbolId>`: one insert / one probe per
hoisted var, no per-scope inner map, and one hash instead of two. Behaviour is
identical — the map is only ever inserted into and point-queried, never iterated.

`cargo allocs` (system allocations):
  semantic  antd.js  1020 -> 28    (-97%)
  minifier  antd.js  3041 -> 1040  (-66%; the minifier rebuilds semantic internally)
@linyiru linyiru force-pushed the perf/semantic-flatten-hoisting-variables branch from 3922016 to 6e9cd15 Compare June 29, 2026 06:49
@Dunqing Dunqing merged commit d5c916a into oxc-project:main Jun 29, 2026
38 checks passed
camc314 added a commit that referenced this pull request Jun 29, 2026
### 💥 BREAKING CHANGES

- 94fbacb ast: [**BREAKING**] Only export `AstBuilder` and `NONE` in
`builder` module (#23876) (overlookmotel)
- 8de5122 ecmascript: [**BREAKING**] Switch to new `AstBuilder` (#23834)
(overlookmotel)
- dc0ef38 transformer: [**BREAKING**] Switch to new `AstBuilder`
(#23831) (overlookmotel)
- 88f4455 str: [**BREAKING**] `Str` and `Ident` methods take
`&GetAllocator` (#23781) (overlookmotel)
- 36009dd allocator: [**BREAKING**] `GetAllocator::allocator` take
`&self` (#23676) (overlookmotel)
- bd74f9d allocator: [**BREAKING**] Rename `AllocatorAccessor` trait to
`GetAllocator` (#23675) (overlookmotel)

### 🚀 Features

- 326fe25 transformer_plugins: Support `typeof` `define` keys (#23605)
(Alexander Lichter)
- f2091b3 ast: Unify old and new `AstBuilder`s (#23875) (overlookmotel)
- cd1fd12 codegen: Expose `Codegen::print_string` API (#23785) (camc314)
- 785461b ast: Add custom builder methods to AST types (#23651)
(overlookmotel)
- 05d1357 ast: Add AST creation methods to AST types (#23650)
(overlookmotel)
- 2580eda str: Add `Str::from_str_in` and `Ident::from_str_in` methods
(#23767) (overlookmotel)
- 6883fcf minifier: Fold write-once falsy var to false in boolean
context (#23540) (Dunqing)
- fcbf993 allocator: Add `Vec::from_value_in` method (#23718)
(overlookmotel)
- 989ddb7 allocator: Add `Vec::from_box_in` method (#23717)
(overlookmotel)
- 9d1aa7f allocator: Improve `PartialEq` for `Vec` (#23716)
(overlookmotel)

### 🐛 Bug Fixes

- da0e5bf minifier: Don't reorder a closed-over TDZ read when inlining a
var (#23771) (Dunqing)
- 0b3021f allocator: Remove `Vec::from_box_in` (#23873) (overlookmotel)
- 0ab64ec ast: Silence deprecation warnings within files defining
deprecated `AstBuilder` methods (#23889) (overlookmotel)
- 8c07cad all: Enable `disable_old_builder` Cargo feature for `oxc_ast`
crate in tests (#23888) (overlookmotel)
- 3800f01 ast: Legacy `AstBuilder` methods take `self` not `&self`
(#23891) (overlookmotel)
- 869ac20 semantic/cfg: Connect for update exit to loop test (#23791)
(camc314)
- d3e92d5 semantic/cfg: Connect while branches from condition exit
(#23790) (camc314)
- 025045d ast: `ExportNamedDeclaration` plain builder methods return
boxed nodes (#23783) (overlookmotel)
- 7537c58 ast: Fix name of `AstBuilder` method for
`Expression::V8IntrinsicExpression` (#23766) (overlookmotel)
- 3f574f5 traverse: Fix unsoundness in `Traverse` walk functions
(#23745) (overlookmotel)
- 585760f parser: String in AST reference arena (#23721) (overlookmotel)
- 7231d55 allocator: Fix unsound lifetime extension in `Box::new_in`
(#23685) (overlookmotel)

### ⚡ Performance

- d5c916a semantic: Flatten hoisting_variables to avoid per-scope map
allocation (#23927) (Lawrence Lin)
- e71609d minifier: Bail member-expr folding before the side-effect walk
(#23924) (Lawrence Lin)
- e1f89ab minifier: Reduce string allocations folding addition (#23846)
(overlookmotel)
- 9f6ee3b isolated-declarations: Pool scope maps to avoid per-scope
alloc/rehash (#23761) (Boshen)
- 0b07c4c semantic: Avoid heap alloc for catch-clause binding ids
(#23911) (Lawrence Lin)
- c5eef8b regular_expression: Skip capturing-group pre-parse when
pattern has no `(` (#23908) (Lawrence Lin)
- b4f5b4b isolated_declarations: Remove redundant clone of formal
parameter pattern (#23912) (Lawrence Lin)
- 53d083f isolated_declarations: Use `TakeIn` not `CloneIn` (#23847)
(overlookmotel)
- 3ea9304 react_compiler: Use faster API to arena allocate strings
(#23849) (overlookmotel)
- a6d8e45 parser: Avoid span lookup for arrow expression body (#23788)
(camc314)
- e1886a0 transformer, minifier: Use `static_ident!` macro to create
static `Ident`s (#23727) (overlookmotel)
- 5527bef transformer/object-rest-spread: Reduce iteration (#23720)
(overlookmotel)
- 680ffbc transformer: Allocate AST nodes in arena directly (#23711)
(overlookmotel)
- 1c63c66 parser: Allocate AST nodes in arena directly (#23712)
(overlookmotel)
- 3855f0c minifier: Allocate AST nodes in arena directly (#23710)
(overlookmotel)
- d025887 isolated_declarations: Allocate AST nodes in arena directly
(#23709) (overlookmotel)
- 10b96c6 parser: Remove string search from parsing JSX element name
(#23713) (overlookmotel)

### 📚 Documentation

- 3d61dea all: Correct capitalization in comments (#23887)
(overlookmotel)
- aa1ad74 ast: Add `#[deprecated]` to legacy `AstBuilder` methods
(#23877) (overlookmotel)
- a4676db ast: Correct doc comment for `NONE` (#23765) (overlookmotel)
- 419ec80 syntax: Fix typo in doc comment (#23674) (overlookmotel)

### 🛡️ Security

- 3cdd18f deps: Update npm packages (#23690) (renovate[bot])

Co-authored-by: Boshen <1430279+Boshen@users.noreply.github.com>
Co-authored-by: Cameron <cameron.clark@hey.com>
camc314 pushed a commit that referenced this pull request Jul 3, 2026
…ocation (#23927)

## Summary

`SemanticBuilder::hoisting_variables` is a nested map
`FxHashMap<ScopeId, IdentHashMap<'a, SymbolId>>`. Each
`entry(scope_id).or_default()`
on a vacant scope heap-allocates a fresh inner `IdentHashMap` via the
**system**
allocator, and subsequent inserts grow it. On `antd.js` this is the
single largest
system-allocation site in semantic analysis (`VariableDeclarator::bind`
/
`Function::bind` in `binder.rs`).

This flattens it to a single `FxHashMap<(ScopeId, Ident<'a>),
SymbolId>`:

- one insert / one point-probe per hoisted variable (was: outer probe →
inner-map
  allocation → inner probe);
- no per-scope inner map, so the per-scope system allocations disappear;
- one hash instead of two.

Behaviour is identical: `hoisting_variables` is only ever inserted into
(`binder.rs`)
and point-queried in `check_redeclaration` (`builder.rs`) — it is never
iterated, so a
flat `(scope, name)` key is a drop-in.

## Results — `cargo allocs` (system allocations)

| fixture | phase | before | after |
| --- | --- | ---: | ---: |
| antd.js | semantic | 1020 | **28** (−97%) |
| antd.js | minifier | 3041 | **1040** (−66%, rebuilds semantic
internally) |
| pdf.mjs | semantic | 25 | 16 |
| kitchen-sink.tsx | semantic | 41 | 36 |

`allocs_semantic.snap` / `allocs_minifier.snap` are updated accordingly.
This is on the
hot path for tools that run `SemanticBuilder` repeatedly per module
(e.g. Rolldown).

## Relation to #19134

#19134 previously tried to move this temporary state (including
`hoisting_variables`)
to a scratch arena; CodSpeed reported a ~3.6% regression and it was
closed. This PR
takes a different, minimal approach: it does **not** introduce an arena,
does **not**
change `SemanticBuilder::new`'s signature, and touches no call sites. It
removes a hash
and an allocation per hoisted variable (rather than adding arena
threading), so it
should be CodSpeed-neutral-or-better — please let CI confirm.

## Tests

- `cargo test -p oxc_semantic --lib --tests` — 60 passed (incl.
redeclaration / var-hoist tests)
- `cargo test -p oxc_minifier --lib --tests` — 538 passed

## AI disclosure

Prepared with AI assistance (Claude). I reviewed the change, located the
hotspot with a
per-call-site allocation profile, verified the reduction with `cargo
allocs`, and ran
the test suites above.
camc314 added a commit that referenced this pull request Jul 3, 2026
### 💥 BREAKING CHANGES

- 94fbacb ast: [**BREAKING**] Only export `AstBuilder` and `NONE` in
`builder` module (#23876) (overlookmotel)
- 8de5122 ecmascript: [**BREAKING**] Switch to new `AstBuilder` (#23834)
(overlookmotel)
- dc0ef38 transformer: [**BREAKING**] Switch to new `AstBuilder`
(#23831) (overlookmotel)
- 88f4455 str: [**BREAKING**] `Str` and `Ident` methods take
`&GetAllocator` (#23781) (overlookmotel)
- 36009dd allocator: [**BREAKING**] `GetAllocator::allocator` take
`&self` (#23676) (overlookmotel)
- bd74f9d allocator: [**BREAKING**] Rename `AllocatorAccessor` trait to
`GetAllocator` (#23675) (overlookmotel)

### 🚀 Features

- 326fe25 transformer_plugins: Support `typeof` `define` keys (#23605)
(Alexander Lichter)
- f2091b3 ast: Unify old and new `AstBuilder`s (#23875) (overlookmotel)
- cd1fd12 codegen: Expose `Codegen::print_string` API (#23785) (camc314)
- 785461b ast: Add custom builder methods to AST types (#23651)
(overlookmotel)
- 05d1357 ast: Add AST creation methods to AST types (#23650)
(overlookmotel)
- 2580eda str: Add `Str::from_str_in` and `Ident::from_str_in` methods
(#23767) (overlookmotel)
- 6883fcf minifier: Fold write-once falsy var to false in boolean
context (#23540) (Dunqing)
- fcbf993 allocator: Add `Vec::from_value_in` method (#23718)
(overlookmotel)
- 989ddb7 allocator: Add `Vec::from_box_in` method (#23717)
(overlookmotel)
- 9d1aa7f allocator: Improve `PartialEq` for `Vec` (#23716)
(overlookmotel)

### 🐛 Bug Fixes

- da0e5bf minifier: Don't reorder a closed-over TDZ read when inlining a
var (#23771) (Dunqing)
- 0b3021f allocator: Remove `Vec::from_box_in` (#23873) (overlookmotel)
- 0ab64ec ast: Silence deprecation warnings within files defining
deprecated `AstBuilder` methods (#23889) (overlookmotel)
- 8c07cad all: Enable `disable_old_builder` Cargo feature for `oxc_ast`
crate in tests (#23888) (overlookmotel)
- 3800f01 ast: Legacy `AstBuilder` methods take `self` not `&self`
(#23891) (overlookmotel)
- 869ac20 semantic/cfg: Connect for update exit to loop test (#23791)
(camc314)
- d3e92d5 semantic/cfg: Connect while branches from condition exit
(#23790) (camc314)
- 025045d ast: `ExportNamedDeclaration` plain builder methods return
boxed nodes (#23783) (overlookmotel)
- 7537c58 ast: Fix name of `AstBuilder` method for
`Expression::V8IntrinsicExpression` (#23766) (overlookmotel)
- 3f574f5 traverse: Fix unsoundness in `Traverse` walk functions
(#23745) (overlookmotel)
- 585760f parser: String in AST reference arena (#23721) (overlookmotel)
- 7231d55 allocator: Fix unsound lifetime extension in `Box::new_in`
(#23685) (overlookmotel)

### ⚡ Performance

- d5c916a semantic: Flatten hoisting_variables to avoid per-scope map
allocation (#23927) (Lawrence Lin)
- e71609d minifier: Bail member-expr folding before the side-effect walk
(#23924) (Lawrence Lin)
- e1f89ab minifier: Reduce string allocations folding addition (#23846)
(overlookmotel)
- 9f6ee3b isolated-declarations: Pool scope maps to avoid per-scope
alloc/rehash (#23761) (Boshen)
- 0b07c4c semantic: Avoid heap alloc for catch-clause binding ids
(#23911) (Lawrence Lin)
- c5eef8b regular_expression: Skip capturing-group pre-parse when
pattern has no `(` (#23908) (Lawrence Lin)
- b4f5b4b isolated_declarations: Remove redundant clone of formal
parameter pattern (#23912) (Lawrence Lin)
- 53d083f isolated_declarations: Use `TakeIn` not `CloneIn` (#23847)
(overlookmotel)
- 3ea9304 react_compiler: Use faster API to arena allocate strings
(#23849) (overlookmotel)
- a6d8e45 parser: Avoid span lookup for arrow expression body (#23788)
(camc314)
- e1886a0 transformer, minifier: Use `static_ident!` macro to create
static `Ident`s (#23727) (overlookmotel)
- 5527bef transformer/object-rest-spread: Reduce iteration (#23720)
(overlookmotel)
- 680ffbc transformer: Allocate AST nodes in arena directly (#23711)
(overlookmotel)
- 1c63c66 parser: Allocate AST nodes in arena directly (#23712)
(overlookmotel)
- 3855f0c minifier: Allocate AST nodes in arena directly (#23710)
(overlookmotel)
- d025887 isolated_declarations: Allocate AST nodes in arena directly
(#23709) (overlookmotel)
- 10b96c6 parser: Remove string search from parsing JSX element name
(#23713) (overlookmotel)

### 📚 Documentation

- 3d61dea all: Correct capitalization in comments (#23887)
(overlookmotel)
- aa1ad74 ast: Add `#[deprecated]` to legacy `AstBuilder` methods
(#23877) (overlookmotel)
- a4676db ast: Correct doc comment for `NONE` (#23765) (overlookmotel)
- 419ec80 syntax: Fix typo in doc comment (#23674) (overlookmotel)

### 🛡️ Security

- 3cdd18f deps: Update npm packages (#23690) (renovate[bot])

Co-authored-by: Boshen <1430279+Boshen@users.noreply.github.com>
Co-authored-by: Cameron <cameron.clark@hey.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants