Use Highway SIMD by Jarred-Sumner · Pull Request #19134 · oven-sh/bun

Jarred-Sumner · 2025-04-19T23:49:55Z

This PR introduces the Google Highway SIMD library to shrink the performance gap between the baseline build and the non-baseline build of Bun.

Migrate several string-related functions from Zig's @Vector to Highway.
- Implements several core string operations in C++ using Highway primitives (src/bun.js/bindings/highway_strings.cpp):
  - indexOfChar (replaces memchr, strchr)
  - indexOfAnyChar (replaces manual Zig @Vector loops)
  - indexOfSubstring (replaces memmem)
  - indexOfInterestingCharacterInStringLiteral
  - indexOfNeedsEscapeForJavaScriptString (handles quotes, backslash, control chars, and $ for template literals)
  - scanCharFrequency (for identifier character frequency analysis, not used yet.)
  - copyU16ToU8 (was hoping this would have some impact, but did not)
  - WebSocket masking (fillWithSkipMask)
  - Specialized search functions for lexer/parser optimization (e.g., indexOfNewlineOrNonASCII, indexOfNewlineOrNonASCIIOrHashOrAt, indexOfSpaceOrNewlineOrNonASCII, containsNewlineOrNonASCIIOrQuote).
- Provides Zig FFI wrappers (src/highway.zig) for these C++ functions.
Integration and Replacement:
- Replaces numerous existing Zig implementations and C library calls in src/string_immutable.zig with the faster Highway-backed functions.
- Replaces manual @Vector usage in src/js_lexer.zig, src/http/websocket_http_client.zig, src/glob/GlobWalker.zig, and src/shell/braces.zig with Highway or standard library functions.
- Significantly refactors the JS lexer's comment and pragma scanning logic (// #__PURE__, // @jsx, // # sourceMappingURL=) to leverage the new specialized Highway search functions, resulting in a ~40% performance improvement for parsing JS files with inline sourcemaps.

Implementation Details:

Uses Highway's dynamic dispatch mechanism (HWY_DYNAMIC_DISPATCH) to select the best SIMD implementation at runtime.
Overrides libc memmem on Linux and macOS

How did you verify your code works?

Adds new tests specifically for comment parsing (test/bundler/bundler_comments.test.ts).

robobun · 2025-04-19T23:50:05Z

^{Updated 11:25 PM PT - Apr 21st, 2025}

❌ @Jarred-Sumner, your commit 5824427 has 1 failures in Build #15307:

test/js/bun/spawn/spawn.test.ts - timeout on 🐧 3.21 aarch64

🧪 try this PR locally:

bunx bun-pr 19134

@zackradisic

cc @zackradisic

dylan-conway · 2025-04-21T23:03:16Z

+    // FUTURE TODO: normalize tabs
+    // Some tests involving githubactions depend on tabs not being normalized right now.


what does this mean? why does checking for newlines or latin1 need special handling for tabs?

dylan-conway · 2025-04-21T23:05:55Z

+    // Scalar check for the remainder
+    for (; i < search_len; ++i) {
+        const uint8_t char_ = start_ptr[i];
+        if (char_ > 127 || char_ < 0x20) {


Suggested change

if (char_ > 127 || char_ < 0x20) {

if (char_ > 127 || char_ == '\n' || char_ == '\r') {

originally this check was if (char_ > 127 || char_ < 0x20 || char_ == '\n' || char_ == '\r') {. is it correct to accept anything less than a space?

< 0x20 covers "\n" and "\r"

dylan-conway · 2025-04-21T23:08:05Z

+        return null;
+    }
+
+    return highway_index_of_space_or_newline_or_non_ascii(


this needs to be updated to return null when it isn't found

dylan-conway · 2025-04-21T23:08:39Z

+
+    for (; i < search_len; ++i) {
+        const uint8_t char_ = start_ptr[i];
+        if (char_ <= ' ' || char_ > 127) {


Suggested change

if (char_ <= ' ' || char_ > 127) {

if (char_ == ' ' || char_ == '\n' || char == '\r' || char_ > 127) {

another case where we accept anything less than a space (or equal). not sure if correct

equivalent to status quo

dylan-conway · 2025-04-21T23:29:11Z

+    return result;
+}
+
+pub fn indexOfNewlineOrNonASCII(haystack: string) ?usize {


Most of these functions need their names updated. They will return non-null for more than they say

i'll rename it to be indexOfInterestingCharacter

dylan-conway · 2025-04-22T00:15:55Z

+                if (PragmaArg.scan(.skip_space_first, lexer.start + offset_for_errors, "jsxFrag", chunk, allow_newline)) |span| {
+                    lexer.jsx_pragma._jsxFrag = span;
+                    return "jsxFrag".len +
+                        if (span.range.len > 0) @as(usize, @intCast(span.range.len)) else 0;


would need a lot of review, but in the future it looks possible to change Range.len to be unsigned

Co-authored-by: Dylan Conway <35280289+dylan-conway@users.noreply.github.com>

…nt (#41284) ### Problem - A `//` comment that mentions `/*#__PURE__*/` as text marks the next call as pure. Bun removes an unused pure call at runtime and in `bun build`, without minification. So `` // `/*#__PURE__*/` `` followed by `console.log(...)` deletes the `console.log`. In `@angular/build` this deleted `source.appendRight(...)` and corrupted the Angular build output (angular/angular-cli#33973). - The cause is `scan_single_line_comment` in `src/js_parser/lexer.rs:2087`. On any `#` or `@` in a `//` comment it calls `scan_pragma`, which sets `has_pure_comment_before` when `__PURE__` follows. The position in the comment does not matter. Fixes #41275. Fixes #19355. ### Fix - In a `//` comment, `@__PURE__` or `#__PURE__` is an annotation only when it is the first word: only whitespace may sit between `//` and the marker. Text after the marker is still allowed. Block comments are unchanged. - This deviates from esbuild, terser and rollup, which match the marker anywhere in a comment. Those tools only act on it when minifying or tree shaking. Bun acts on it always, so a false match deletes user code. Line directives such as `// @ts-ignore`, `//# sourceMappingURL=` and `// eslint-disable-next-line` use the same first-word rule. Block comments keep the esbuild rule because a `/* */` comment cannot quote a `/*#__PURE__*/`, and `/** @__PURE__ */` must keep working. - Four existing tests in `test/bundler/bundler_comments.test.ts` put text before the marker and asserted removal. They came in with the SIMD lexer change (#19134) to exercise the scanner, and now assert that the call is kept. `test/bundler/transpiler/transpiler.test.js:1589` was a victim of this bug: the `@__PURE__` text in the commented-out lines above it deleted the first `err(...)` call, which asserted a TSX-only error for the `ts` loader. It now asserts the `ts` behavior, as in esbuild's `ts_parser_test.go`. - Verified: `test/bundler/bundler_comments.test.ts` (4 new cases) and `test/cli/install/bun-run.test.ts` (1 new runtime case). Both fail on the release binary. Also ran `test/bundler/esbuild/dce.test.ts`, `default.test.ts`, `tsconfig.test.ts`, `bundler_minify.test.ts`, `bundler_jsx.test.ts`, `transpiler/transpiler.test.js`, `transpiler/react-compiler.test.ts`. ### Background - A pure annotation (`/* @__PURE__ */ f()`) tells a bundler that the call has no side effects. If the result is unused, the call can be removed. - Bun's `dead_code_elimination` parser feature is on by default, also for `bun run`. The SExpr visitor (`src/js_parser/visit/visit_stmt.rs:1430`) calls `SideEffects::simplify_unused_expr` with no minify gate, and `scan_side_effects.rs` drops a call marked `IfUnused` whose arguments have no side effects. The docs describe pure annotation support as "Always active". - `has_pure_comment_before` is reset at the start of `Lexer::next()` and read in `parse_expr_common` (`src/js_parser/parse/mod.rs:75`), which marks the next `ECall` or `ENew` as `CanBeUnwrapped::IfUnused`. <details><summary>Notes</summary> Repro under bun 1.4.1: ``` $ bun -e 'function repro() { // `/*#__PURE__*/` console.log("This should not be deleted"); } console.log(repro.toString());' function repro() {} ``` esbuild 0.21.5 attaches the annotation to the same input but keeps the call unless `--minify-syntax` is passed. Alternative considered: gate the removal of a pure-marked expression statement on `minify_syntax`, as esbuild does. That changes documented runtime behavior and unminified `bun build` output for every correctly annotated call, so this PR narrows what counts as an annotation instead. A `/* see @__PURE__ for details */` block comment before a call still marks it pure. That is the esbuild rule and is not changed here. A quoted `/*#__PURE__*/` cannot occur inside a block comment, so the real-world reports all come from `//` comments. Self-reviewed: the review found the dead `transpiler.test.js` assertion and asked for the esbuild deviation to be stated. Both are addressed above. It also suggested gating `bun test` on `ignore_dce_annotations`. That is a separate change and is not in this PR. Unrelated failures seen locally with the debug build: two `bun-build-api.test.ts` bytecode tests that time out at 5s under the debug build. </details>  --- **no test proof** · iteration 5 · platform-specific test(s) that do not run on this machine, deferring to CI, which covers all platforms: test/cli/install/bun-run.test.ts

Add highway simd

f9fdbf6

Jarred-Sumner force-pushed the jarred/highwaysimd branch from 968b6e8 to f9fdbf6 Compare April 20, 2025 02:29

Jarred-Sumner added 19 commits April 19, 2025 20:04

more

c841fe1

Update highway_strings.cpp

56c4894

a

1358ad5

Update highway_strings.cpp

8e7ca7d

Update highway_strings.cpp

e4560ff

Clean this up

3f5fe4f

Migrate copyU16IntoU8

796da49

Update comment

34ac127

Fix

400bf03

Fix test failures, add custom memmem

db5b74d

Remove unnecessary usage of @vector

b0b1a97

cc @zackradisic

Update shell.zig

81bdcc1

Update ban-words.test.ts

c3b274f

Migrate WebSocket client fill

2b8e4a5

Update highway_strings.cpp

e8877ce

Update highway_strings.cpp

bb574fe

skip

a7eec19

Update js_lexer.zig

beb14f7

40% faster at parsing JS that that has inline sourcemaps

8909435

Jarred-Sumner force-pushed the jarred/highwaysimd branch from a2f70b1 to 8909435 Compare April 21, 2025 10:19

Fix

11c6b49

Jarred-Sumner marked this pull request as ready for review April 21, 2025 11:18

Jarred-Sumner changed the title ~~use highway simd~~ Use Highway SIMD Apr 21, 2025

dylan-conway reviewed Apr 21, 2025

View reviewed changes

dylan-conway added 2 commits April 21, 2025 16:13

Update highway.zig

e6756fc

Update bundler_comments.test.ts

21afd8e

dylan-conway reviewed Apr 21, 2025

View reviewed changes

Comment thread src/highway.zig Outdated

dylan-conway reviewed Apr 21, 2025

View reviewed changes

dylan-conway reviewed Apr 22, 2025

View reviewed changes

Jarred-Sumner and others added 2 commits April 21, 2025 22:47

Update src/highway.zig

55e70a9

Co-authored-by: Dylan Conway <35280289+dylan-conway@users.noreply.github.com>

bun run zig-format

5824427

Jarred-Sumner merged commit 0471254 into main Apr 22, 2025

Jarred-Sumner deleted the jarred/highwaysimd branch April 22, 2025 06:28

robobun mentioned this pull request Sep 3, 2026

lexer: only treat __PURE__ as an annotation when it starts a // comment #41284

Merged

robobun mentioned this pull request Oct 13, 2026

lexer: end a JSX pragma value at a line break in block comments too #45140

Open

Navigation Menu

Repository navigation

Use Highway SIMD - #19134

Use Highway SIMD#19134
Jarred-Sumner merged 25 commits into
mainfrom
jarred/highwaysimd

Jarred-Sumner commented Apr 19, 2025 •

edited

Loading

Uh oh!

robobun commented Apr 19, 2025 •

edited

Loading

Uh oh!

dylan-conway Apr 21, 2025

Uh oh!

dylan-conway Apr 21, 2025

Uh oh!

Jarred-Sumner Apr 22, 2025

Uh oh!

dylan-conway Apr 21, 2025

Uh oh!

dylan-conway Apr 21, 2025

Uh oh!

Jarred-Sumner Apr 22, 2025

Uh oh!

Uh oh!

dylan-conway Apr 21, 2025

Uh oh!

Jarred-Sumner Apr 22, 2025

Uh oh!

dylan-conway Apr 22, 2025

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		// FUTURE TODO: normalize tabs
		// Some tests involving githubactions depend on tabs not being normalized right now.

	if (char_ > 127 \|\| char_ < 0x20) {
	if (char_ > 127 \|\| char_ == '\n' \|\| char_ == '\r') {

	if (char_ <= ' ' \|\| char_ > 127) {
	if (char_ == ' ' \|\| char_ == '\n' \|\| char == '\r' \|\| char_ > 127) {

Conversation

Jarred-Sumner commented Apr 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

How did you verify your code works?

Uh oh!

robobun commented Apr 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

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!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Jarred-Sumner commented Apr 19, 2025 •

edited

Loading

robobun commented Apr 19, 2025 •

edited

Loading