startup: Cache packaged ESM ASAR resolutions

[Dmitriy Vasyura (dmitrivMS)]

Summary

• cache successful packaged ESM resolutions into node_modules.asar
• key entries by specifier, normalized parent path, conditions, and import attributes
• preserve closer-dependency precedence on the first resolution and avoid caching defaults or failures
• add a packaged ASAR integration fixture that verifies hits are reused but isolated by parent

The cache is process-local and lives for the process lifetime, so it also helps repeated dynamic imports after startup. CommonJS, built-ins, Electron modules, external extension paths, closer dependencies, and failed resolutions are unchanged.

Performance

The motivating main-process trace found 33 ASAR resolution attempts from out/main.js for seven unique specifier-parent pairs, including 15 resolutions of @vscode/spdlog and eight of @vscode/policy-watcher.

Packaged resolver benchmark

40 alternating baseline/cached Electron runs, excluding five warmups, shaped as 33 import sites over seven unique contexts:

Metric	Baseline	Cached	Change
Resolution workload median	44.184 ms	19.172 ms	-25.012 ms / -56.61%
Electron process median	135.250 ms	109.649 ms	-25.601 ms / -18.93%
Archive lookups	33	7	-78.79%

Main-process CPU profiles

Six packaged launches per variant, excluding one warmup:

Metric	Baseline	Cached	Change
Active main CPU median	378.879 ms	360.334 ms	-18.545 ms / -4.89%
Attributed JavaScript median	366.230 ms	347.248 ms	-18.982 ms / -5.18%
internalModuleStat median	12.605 ms	4.880 ms	-61.29%
Profiled elapsed median	1,926 ms	1,892 ms	-1.77%

The profiler adds roughly 33-35% overhead in this local Code OSS package, so the CPU deltas are more reliable than the profiled wall-time delta.

Full packaged startup

Ten unprofiled launches per variant, excluding one warmup:

• warm median: 1,422 ms -> 1,418 ms (-4 ms / -0.28%)
• warm mean: 1,425.67 ms -> 1,408.67 ms (-1.19%)
• run CV: 1.89% baseline, 1.55% cached

The median wall-time change is below run variance, so this PR does not claim a statistically significant end-to-end startup improvement. The demonstrated benefit is reduced resolver and main-process CPU, including for later repeated imports.

For the full-package A/B, the baseline forced the generated cache lookup to miss while retaining the small key/map overhead; every other application byte was identical. The standalone resolver benchmark compares the exact pre-change and changed bootstrap revisions.

Validation

• npm run typecheck-client
• targeted ESLint
• focused bootstrap ESM tests: 6 passing
• full vscode-win32-x64-min package build
• packaged Code OSS startup A/B
• git diff --check

[Copilot]

Copilot review overview

🟡 Changes recommended

The non-injective cache key can incorrectly reuse resolutions across distinct contexts.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review tier: Balanced
Findings: 1

New issues introduced by this change (1)

Severity	Finding
	src/​bootstrap-esm.ts — The flattened key is not injective: conditions [..., 'type', 'json'] with no attributes produce…

What changed in this PR

Adds process-local caching for packaged ESM ASAR resolutions.

Changes:

• Caches resolutions by specifier, parent, conditions, and attributes.
• Preserves closer-dependency precedence.
• Adds packaged-ASAR integration coverage.

A critical cache-key collision can return incorrect resolutions; collections need structural encoding and regression coverage. Direct fallback caching for packages without exports also lacks coverage.

File	Description
src/​vs/​code/​test/​node/​bootstrapESM.test.ts	Adds cache reuse and parent-isolation tests.
src/​bootstrap-esm.ts	Implements caching; requires collision-safe keys and fallback-path coverage.

Suppressed comments (1)

src/bootstrap-esm.ts:196

• The new test package declares exports, so it only exercises the self-reference cache write above; this direct fallback cache path for packages without exports remains untested. Add a second archived package without exports and assert that repeated imports produce one archive lookup followed by a cache hit.

					const result = { url, shortCircuit: true };
					resolutionCache.set(cacheKey, result);

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.