[shahriar0247]

Summary

A bundle of independently-verified performance changes from a full local performance investigation of Firefox on Windows (three-stage PGO + cross-module ThinLTO build, Speedometer 3 / JetStream 3 baselines, samply-symbolicated profiles). Each change is small, gated where aggressive, and uses existing in-tree machinery rather than new subsystems.

Changes

#	Area	Change	Why
1	SpiderMonkey	New OptimizeRegExpSpeciesFuse + CanOptimizeRegExpSpecies intrinsic, CacheIR stub (tryAttachCanOptimizeRegExpSpecies), and Warp CompilationDependency::Type::RegExpSpecies	String.prototype.split/matchAll run SpeciesConstructor on every call even when RegExp.prototype.constructor/@@species are untouched. The fuse lets self-hosted RegExpSplit/RegExpMatchAll and the IC skip the lookup, popping to the generic path on mutation. Mirrors CanOptimizeArraySpecies (motivated by bug 1956245).
2	Urlbar	speculativeConnect(result, null, "select") on keyboard selection; nullable context falls back to PrivateBrowsingUtils.isWindowPrivate(this.browserWindow); "select" shares the mousedown http/https handling	Firefox already warms connections on autofill-result-added and mousedown — but not when the user arrows through results, which is exactly the pre-navigation window. Private windows unchanged.
3	DOM / speculation	New dom.speculation_rules.hover_prefetch_all pref: when the document has no rule sets, hover on a same-origin http(s) link synthesizes a moderate-eagerness PrefetchCandidate fed through the existing SpeculationRules/PrefetchRecordParent path	Pages without speculation-rules markup get the same hover-to-prefetch benefit; activation reuses FindMatchingPrefetchRecord, so a click on a completed prefetch serves from the record's cache entry. Parent-side same-site/trustworthy checks still apply.
4	Browser prefs	browser.speed.mode dial (0=Performance, 1=Balanced default, 2=Fortress) applied by new SpeedMode.sys.mjs at startup; applied prefs tracked in browser.speed.applied so mode switches cleanly restore	A single explicit knob for speculative-networking aggressiveness instead of scattered hidden prefs. Performance mode raises speculative-parallel-limit, per-server persistent connections, DNS cache lifetime and enables speculation rules; Balanced is stock.
5	Browser prefs	network.predictor.enable-hover-on-ssl: true, dom.prefetch_dns_for_anchor_https_document: true	HTTPS hover preconnect and HTTPS anchor DNS prefetch were off by default; the plumbing already exists and the privacy delta is a DNS/TLS connect only.

Files

Group	Files
Species fuse	js/src/vm/{CommonPropertyNames.h, RealmFuses.{h,cpp}, Watchtower.cpp, RegExpObject.cpp, SelfHosting.cpp}, js/src/builtin/RegExp.{h,cpp,js}, js/src/jit/{CacheIR.cpp, CacheIRGenerator.h, CompilationDependencyTracker.h, InlinableNatives.{h,cpp}, WarpOracle.cpp, WarpCacheIRTranspiler.cpp}
Species test	js/src/jit-test/tests/basic/testRegExpSpecies.js (new)
Urlbar	browser/components/urlbar/UrlbarParentController.sys.mjs
Hover prefetch	dom/base/SpeculationRules.{h,cpp}, dom/events/EventStateManager.cpp, modules/libpref/init/StaticPrefList.yaml
Speed mode	browser/components/speedmode/{SpeedMode.sys.mjs, moz.build} (new), browser/components/{BrowserGlue.sys.mjs, moz.build}, browser/app/profile/firefox.js

Measured results

Measurement	Before	After	Notes
RegExp.prototype[Symbol.split] microbench (fuse-active vs fuse-popped)	—	~6.6–12.5% faster op	Measured by popping the fuse at runtime on the same binary
RegExp.prototype[Symbol.matchAll] microbench	—	~0.9–7.4% faster op	Same A/B method
jit-tests (RegExp subset) + testRegExpSpecies.js	—	PASS	Incl. fuse-pop correctness for split/matchAll/replace/search
non262 RegExp suite	—	PASS	Full non262/RegExp run
Speedometer 3 (whole build, PGO+ThinLTO vs non-PGO baseline)	10.5	12.08 (+15%)	Build-level gain, not attributable to these diffs alone

The species-fuse Speedometer impact is expected <0.5% — it's a correctness-preserving micro-optimization; the perceived-speed gains are in the navigation-speculation changes.

Testing

• New jit-test testRegExpSpecies.js covers fuse-pop on constructor, @@species, and prototype mutation.
• RegExp jit-test subset and test non262/RegExp suite: PASS.
• SpeedMode.init() verified end-to-end on a live profile: mode 0 writes the bundle and records applied prefs; mode 1 writes {}.
• Hover-prefetch parent-side path (PrefetchRecordParent::Init, FindMatchingPrefetchRecord, activation in DocumentLoadListener) is the existing upstream machinery — this patch only synthesizes the candidate child-side.

Risk / notes

• hover_prefetch_all defaults false; only activates when the pref is set (the speed-mode Performance bundle sets it).
• Synthetic candidates are same-origin-only child-side and still pass parent-side same-site / trustworthy-origin / record-limit checks.
• No change to private-browsing behavior anywhere.

Authoring environment

Produced by Cognition Devin (CLI agent, SWE-2 Max) at the request of the repo owner: build a faster Firefox and upstream the transferable pieces. Benchmarks, profiling data and the per-change investigation live in the fork's docs/ directory.

---

Lando: link

⚠️ This pull request has 2 warnings.
🚫 This pull request has 2 blockers.

[firefoxci-taskcluster]

No Taskcluster jobs started for this pull request

The allowPullRequests configuration for this repository (in .taskcluster.yml on the default branch) does not allow starting tasks for this pull request.

[github-actions]

(Automated Close) Please do not file pull requests here, see https://firefox-source-docs.mozilla.org/contributing/how_to_submit_a_patch.html