[robobun]

Problem

• bun run and bun build --no-bundle print a folded NaN (0 / 0, +"abc", a macro result) as the identifier NaN, so a local NaN captures it: for (let NaN = 0; NaN < 1; NaN++) console.log(0 / 0) prints 0, and export const NaN = 0 / 0 is a TDZ error (the NaN twin of ReferenceError: Cannot access uninitialized variable. #7263). --no-bundle also prints void 0 as undefined next to a local undefined.
• The cause is print_number / print_undefined (src/js_printer/lib.rs): they name the global with no scope check. renamer: reserve NaN, Infinity and undefined so inlined constants reach the globals #41909 reserved the names in the bundler renamer, but these paths do not rename.

Fix

• print_ast, the transpiler entry point, looks for a bound symbol named undefined, NaN or Infinity in the file and sets the printer option from that. The bundler still sets it from its renamer.
• With the option off, or inside a with body (esbuild's withNesting), the printer writes 0 / 0, 1 / 0 and void 0, in parentheses where precedence needs them.
• Other files print as before, except that non-minified output now writes an infinite literal as Infinity, like the bundler. The runtime transpiler cache version moves to 30.
• Verified: test/bundler/transpiler/transpiler.test.js (new block; 7 tests fail on 1.4.3), plus the suites listed in the notes.

Background

• The parser folds +"abc", void 0, 1e999, an unbound NaN identifier, and under bun run also 0 / 0, into constant nodes with no symbol. The printer must spell the value.
• bun run prints with minify_syntax, which already writes void 0 and 1 / 0; only NaN was exposed there.
• Every binding the transpiler prints has a Symbol in the file's table: declarations from the parser, or exports that data loaders (TOML keys) synthesize. Unresolved references are Kind::Unbound.

Notes

• Supersedes js_printer: guard undefined/NaN/Infinity identifiers against local shadows #36122 and js_printer: print 0 / 0 and void 0 where a local binding or with object shadows NaN / undefined #36124 (July 27). Both conflict with main since renamer: reserve NaN, Infinity and undefined so inlined constants reach the globals #41909 landed the renamer half of js_printer: guard undefined/NaN/Infinity identifiers against local shadows #36122. js_printer: print 0 / 0 and void 0 where a local binding or with object shadows NaN / undefined #36124 used the same symbol-table check; this version adds the with handling, the property-key case and tests for each door.
• Why a symbol scan in print_ast and not a parser flag: the parser is not the only producer of bindings on this path. Bun.Transpiler().transformSync(src, "toml") (and JSON, YAML, JSON5) turns top-level keys into var bindings in transpiler.rs without the parser, so a TOML file with keys NaN / Infinity and a nan / inf value printed var NaN = "n", x = NaN. A class expression's own name also bypasses declare_symbol. The scan covers every producer. Cost: about 4 ns per symbol (0.2 ms for 50k symbols), once per file.
• The ledger rows: { class NaN {} String(0 / 0) } gave the class source, switch (1) { case 1: let NaN = "local"; console.log(0 / 0) } gave local, and the macro face const g = (NaN = nan()) => NaN printed (NaN = NaN) => NaN (a TDZ error under bun run); it now prints (NaN = 0 / 0) => NaN. A function-level const NaN only looked handled because DCE dropped the unused binding; a kept let NaN captured there too.
• Parentheses: 0 / 0 and 1 / 0 are wrapped at Level::Multiply and above (x / (0 / 0), (0 / 0).toFixed(), 2 ** (0 / 0)), void 0 at Level::Prefix and above, as before.
• A NaN or Infinity property key that would print as 0 / 0 / 1 / 0 prints computed (number_property_key_must_be_computed).
• With minify_identifiers the transpiler path runs MinifyRenamer, which reserves these names since renamer: reserve NaN, Infinity and undefined so inlined constants reach the globals #41909, but a named export export let NaN is must_not_be_renamed, so the scan is used there too.
• Identifiers written inside a with body in the source stay identifiers (the parser does not fold them there), so with_nesting only affects synthesized constants.
• Not handled, same as esbuild: sloppy-mode direct eval("var NaN = 1") introducing the binding at runtime.
• NaN bits, measured on x86-64 with new Function (no Bun transpile) in both Bun and Node: f64[0] = 0 / 0 and -(0 / 0) store 7ff8000000000000, the same as the NaN global; only a runtime z / z gives the x86 default fff8…. On aarch64 the hardware default NaN is 7ff8… already. So Float64Array / Buffer.writeDoubleBE output does not change.
• transpiler.test.js "numeric property keys that overflow to Infinity": the whitespace-only minifier now prints {Infinity:1} (valid), so that block uses minify: { whitespace, syntax } to keep covering the computed [1/0] form, and gains a let Infinity; variant for the non-minified fallback.
• The new tests evaluate the printed code with new Function, which hands it to JavaScriptCore without a second pass through Bun's transpiler, so the --no-bundle door is checked on its own.
• Self-reviewed before opening: the review found the data-loader producer and asked for the module-level export const NaN case; both are addressed above.
• Suites run locally with the debug build: test/bundler/transpiler/transpiler.test.js, test/bundler/bundler_edgecase.test.ts, test/bundler/bundler_minify.test.ts, test/bundler/bundler_loader.test.ts, test/bundler/esbuild/{default,dce,lower}.test.ts, test/js/bun/transpiler/, test/js/bun/toml/, test/bundler/transpiler/{decorators,ts-use-define-for-class-fields}.test.ts.

[coderabbitai]

Warning

Review limit reached

• Run on-demand review

_{On-demand reviews are free for the next 12 days. After that, they cost $0.25 per reviewed file.}

Or wait 37 minutes for your next included review.

Check out review usage here.

View limit details

Limit details: You’ve used all 10 included reviews currently available.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Essentials

Run ID: aa975df6-d8ef-4384-98f9-b9441761a0a8

📥 Commits

Reviewing files that changed from the base of the PR and between d745f03 and a21943e.

📒 Files selected for processing (5)

• src/bundler/LinkerContext.rs
• src/bundler/linker_context/postProcessJSChunk.rs
• src/js_printer/lib.rs
• src/jsc/RuntimeTranspilerCache.rs
• test/bundler/transpiler/transpiler.test.js

---

_{Comment @coderabbitai help to get the list of available commands.}

[robobun]

^{Updated 11:22 AM PT - Sep 8th, 2026}

@robobun, your commit a21943e is building: #113184

[robobun]

Status: ready for review.

Reproduced on 1.4.3 (f42e98025) with the ledger lines: for (let NaN = 0; NaN < 1; NaN++) console.log(0 / 0, +"abc", Math.min(0 / 0, 5)) prints 0 0 0 under bun run, { class NaN {} console.log(String(0 / 0)) } prints the class source, and bun build --no-bundle prints console.log(undefined) inside { let undefined = "U"; ... }. The new block in test/bundler/transpiler/transpiler.test.js covers Bun.Transpiler (three configs), bun run, bun run with a module-level export const NaN, and bun build --no-bundle; 7 tests fail with USE_SYSTEM_BUN=1 and pass on this branch.

[robobun]

Closing as a duplicate of #36124. That PR was rebased onto main earlier today with the same symbol-table approach plus the with handling, and its CI is already green, so it is the one to review. The two extra test cases from here are noted on #36124.

[claude]

I reviewed this PR and didn't find any bugs. Because it changes the default printed form of infinite literals for the transpiler path and reworks the precedence/wrapping logic in print_number, a human look at the output-format shift would still be worthwhile.

What was reviewed:

• print_number NaN/Infinity branches — wrapping at Level::Multiply matches the property-access/** cases in the tests; -1 / 0 and -(0/0) still evaluate to the right values unwrapped.
• print_ast symbol scan — filters Kind::Unbound so a bare reference to global NaN doesn't trip it; covers non-parser producers (TOML loader, class-expression names) as claimed.
• with_nesting — incremented only around the SWith body, not the head expression; reset via the struct init in Printer::init.
• Cache version bump to 30 covers the changed output per REVIEW.md's cache-key rule.

Extended reasoning...

Overview

This PR fixes a correctness bug in the JS printer where folded constants (NaN, Infinity, undefined) were emitted as bare identifiers even when a local binding with that name existed, causing the local to capture the constant. The fix renames has_run_symbol_renamer → global_value_properties_unshadowed, adds a symbol-table scan in print_ast to set it for the transpiler path, adds a with_nesting counter so with bodies always fall back to arithmetic forms, and teaches print_number to emit 0 / 0 / 1 / 0 (and print_undefined to keep void 0) when the names are unsafe. The runtime transpiler cache version is bumped to 30. Five files changed: two bundler linker sites (mechanical rename), the printer, the cache version, and transpiler.test.js.

Security risks

None. This is a code-generation correctness fix with no auth, crypto, network, or filesystem surface. The only input is the AST/symbol table already produced by the parser; no new untrusted-input parsing is introduced.

Level of scrutiny

Medium-high. The JS printer is on the hot path for every file Bun runs or builds, and the change alters default output for the non-bundled transpiler: infinite literals that previously printed as 1 / 0 now print as Infinity when no shadowing binding exists (the existing "numeric property keys that overflow to Infinity" test block was updated to reflect this). That is a user-visible output-format change beyond the shadowed-name bugfix itself, and the precedence/wrapping rework in the Infinity branch (by_name boolean replacing the old has_run_symbol_renamer || minify_syntax condition) deserves a maintainer's eye even though it tests out correctly. The new per-file symbols_for_source.iter().flatten().any(...) scan is O(symbols) on every print_ast call; the PR description quantifies it (~4 ns/symbol) but a reviewer familiar with the printer's perf budget should confirm that's acceptable.

Other factors

The test coverage is unusually thorough: let/class/switch/label/param/class-expression-name/with binding forms, TOML data-loader synthesized bindings, the module-level export const NaN = 0 / 0 TDZ case, parenthesization at ** and property-access precedence, and end-to-end bun run / bun build --no-bundle spawns evaluated via new Function to bypass a second transpile. Tests follow harness conventions (bunExe/bunEnv/tempDir, await using, Promise.all on pipes, stdout/stderr asserted before exitCode, test.concurrent for the subprocess cases). The cache-version bump satisfies REVIEW.md's "any change to cached/serialized output bumps the format version constant" rule. The bug hunt exited on dry_streak with no findings and no ruled-out candidates. No CODEOWNERS cover the changed paths. I checked the print_space_before_identifier() call that now precedes 0 / 0 / (0 / 0) — it's conservative (may add an unneeded space) but never produces wrong adjacency, and the .x / ** follow-on cases are wrapped at Level::Multiply so no 0.x ambiguity arises.