This PR contains:

• bugfix
• feature
• refactor
• documentation
• other

Are tests included?

• yes (bugfixes and features will not be merged without tests)
• no

Breaking Changes?

• yes (breaking changes will not be merged unless absolutely necessary)
• no

List any relevant issue numbers:

• resolves Regression: AMD output renames shadowed 'exports' to 'exports$1' in eval-containing scope #6357

Description

Note: I went through many iterations of fixes to this issue. Used cursor with opus 4.7 max thinking to help design and evaluate fixes.
This fix seemed contained and resonable to me. It goes back to the way things mostly were before #5947 for avoiding un-necessary rewrite of exports variables.
Additionaly it doesn't impact performance either

Since #5947 was merged for 4.53.0, Rollup renames every nested exports binding — even ones that cannot conflict with the output format's runtime exports identifier. The rename silently breaks code where an exports name appears inside a string, for example inside eval, because Rollup doesn't rewrite string contents.

Reproducer

var modules = {
  foo: (unused, exports) => {
    console.log(exports.bar);
    eval('exports.bar = 1');
  }
};
export default modules;

4.52.5 (expected):

var modules = {
  foo: (unused, exports) => {
    console.log(exports.bar);
    eval('exports.bar = 1');
  }
};
export { modules as default };

4.53.0 – 4.60.x (current, broken):

var modules = {
  foo: (unused, exports$1) => {
    console.log(exports$1.bar);
    eval('exports.bar = 1'); // still references outer `exports` at runtime!
  }
};
export { modules as default };

The parameter was renamed to exports$1, but the string argument to eval was not, so the two stop referring to the same binding. The reporter ran into this while running webpack-produced development bundles through Rollup, where webpack's runtime uses eval to execute module factories whose parameters are literally named exports.

As the reporter notes in #6357, the rename is conceptually wrong even without the eval: it's an over-deconfliction of a legal nested binding that never needed to change.

Root cause

#5947 did two things:

1. Added the output-format runtime names (module, exports, require, __filename, __dirname, interop helpers) to a new chunk-level RESERVED_USED_NAMES seed in Chunk.ts, so the top-level deconflict step sees them as already-used.
2. Added 'exports' to the universal RESERVED_NAMES set in src/utils/RESERVED_NAMES.ts.

The first change is fine — it's how the top-level const exports = 1 case has always been protected in CJS/AMD/UMD/IIFE/SYSTEM. The second change is the regression: RESERVED_NAMES is consulted by getSafeName for every scope. ChildScope.deconflict runs it on each nested scope's variables, so any nested binding or parameter named exports is forced into exports$1, regardless of whether the surrounding format even has a runtime exports (ES doesn't), or whether the scope references anything that actually needs it.

#5947 also dropped a pre-existing conditional in ChildScope.addUsedOutsideNames that added 'exports' to a scope's used names only when that scope accessed an outer binding that was being re-assigned through exports("name", value) in SystemJS output. With 'exports' now universally reserved, that conditional was no longer strictly needed — but removing it meant that the "I actually do need to avoid exports" decision stopped being driven by the scope's own behavior and started being driven by a global flag.

Fix

1. Remove 'exports' from src/utils/RESERVED_NAMES.ts. exports isn't a JavaScript reserved word; it's only a runtime identifier in certain output formats.
2. Restore the SystemJS-specific check in ChildScope.addUsedOutsideNames so that nested scopes in SystemJS output continue to reserve exports exactly when they reference an outer binding that will be assigned via exports("name", value) at render time.
3. Thread format and exportNamesByVariable through the one extra call site in deconflictChunk.

For CJS/AMD/UMD/IIFE the existing machinery still does the right thing without any extra work: re-assigned exports already carry renderBaseName === 'exports', and ChildScope.addUsedOutsideNames already pulls that name into the scope's usedNames via getBaseVariableName(). So a nested exports binding in a scope that references such an outer variable still gets renamed to exports$1 — and one that doesn't (the reporter's case) stays as exports.

The top-level protection in Chunk.ts (RESERVED_USED_NAMES) and the module.info.safeVariableNames cache added by #5947 are both untouched.

Tests

• New test/form/samples/no-rename-unrelated-exports-parameter/ covers the reporter's exact reproducer across all six formats (amd, cjs, es, iife, system, umd).
• Existing tests whose expected output was changed by refactor: store safe variable names in cache for subsequent usage #5947 to match the buggy behaviour are reverted to match the correct (pre-refactor: store safe variable names in cache for subsequent usage #5947) behaviour for nested scopes:
  • deconflict-format-specific-exports — this test is literally titled "only deconflict 'exports' for formats where it is necessary"; the expected outputs had drifted away from that invariant for nested scopes in every format. Restored.
  • deconflict-format-specific-globals — same fix for the nested-scope case.
  • supports-es5-shim, supports-es6-shim — the wrapper function of the bundled shim ((function (module, exports) { … }(…))) no longer has its exports parameter spuriously renamed.
  • chunking-form/chunk-assigment-in-dynamic — the IIFE (function (exports) { exports.preFaPrint = … }(c)) generated by @rollup/plugin-commonjs for the wrapped module no longer has its exports parameter renamed.

Top-level rename behaviour that was intentionally unified by #5947 (e.g. const exports = 1 at module top-level being renamed in ES output for caching consistency) is left unchanged.

Performance

Because #5947's explicit goal was performance, I validated that this change doesn't regress it. I built three versions of Rollup and ran 60 iterations (2 runs × 30, alternated to control for drift) of a 500-module synthetic fixture that exercises the deconflict hot path, generating all six output formats per iteration:

• pre-PR: commit immediately before refactor: store safe variable names in cache for subsequent usage #5947
• master: current master with refactor: store safe variable names in cache for subsequent usage #5947 applied (has the regression)
• fix: this PR

The benchmark is demonstrably sensitive enough to detect the original #5947 speedup (|t| ≈ 30–37 on generate-time metrics — well above the 95% confidence threshold of 2). Times in ms, mean ± standard error:

Per-format generate time:

Negative = faster. #5947's speedup (−20% generate, −24% to −27% per non-ES format) is fully preserved. This PR is additionally 3.9% faster on the generate phase (t=6.4), because (a) RESERVED_NAMES.has() has one fewer element to check on every getSafeName call, and (b) nested bindings that used to be spuriously renamed no longer go through getSafeName + MagicString rewrite.

On the ES format, #5947 actually made generate time slightly slower (+4.7%) by forcing ES builds to treat exports/module/require/__filename/__dirname as reserved at the top level even though ES has no runtime for them; this PR brings ES back to essentially pre-#5947 levels (40.7ms → 39.0ms) by lifting the unnecessary nested-scope reservation.