Monthly Archives: March 2018

Firefox Performance Update #4

Another week, another slew of Firefox performance updates for you fine folks!

Firefox contributors have been hacking away to make Firefox faster in a bunch of different ways. Have you seen a bug land recently that improved Firefox performance? Let me know so I can include it in an upcoming post and give these contributors the recognition they deserve!

So let’s get to it – here’s what’s been happening lately:

🌟 Trisha made it cheaper to open up the List All Tabs menu, especially when you have many, many tabs open!
Florian Quèze enabled the blank window perceived performance optimization on Nightly for Linux and Windows. This allows us to paint a blank window very quickly after Firefox starts, but before we’ve finished getting the UI figured out. We hypothesize that this will help with perceived start-time performance.
Mike Conley enabled tab warming by default on Nightly. This should hopefully help with perceived tab switch performance. Here’s a blog post that explains how it works. Initial data from Telemetry is quite compelling!
Emilio Cobos Álvarez made it so that we can calculate the dimensions of a new window before starting layout! This had a positive effect on our start-up, window opening and session restoration benchmark tests. While he was in the area, he also made the Tabs-in-Titlebar calculations that we do when opening new browser windows more efficient.
Gijs Kruitbosch got rid of a sync layout flush when using the Find Bar
Alex Gaynor set up the infrastructure that should allow us to asynchronously launch content processes! Stay tuned for the real impact when this bug is fixed.
Michal Novotny made it so that JAR channels stop doing main-thread IO! This is good because we use JAR files to package various parts of the browser.
Kartikaya Gupta made it so that Windows users no longer have to suffer main-thread jank when we’re processing user events! This was a frequently occurring problem because oftentimes the main thread in the parent process would be blocked on the compositor, which was busy compositing. This should help with general smoothness and responsiveness on Windows.
Kris Maglione got rid of some now-unneeded compartments code for unsupported add-on types. This resulted in some wins on our JavaScript benchmarks!
Johann Hofmann added a test that will ensure we limit the number of times we interact with the Preferences component during start-up. This will prevent us from accidentally adding more and more overhead in the future.

Firefox Performance Update #3

Hi! I’ve got another slew of Firefox performance work to report today.

Special thanks to the folks who submitted things through this form to let me know about performance work that’s taken place recently! If you’ve seen something fixed lately that’ll likely have a positive impact on Firefox performance, let me know about it!

So, without further ado, here are some of the folks who have made some nice improvements to Firefox performance lately. Thanks for making Firefox faster and better!

Andrew Osmond made it so that very long animated GIFs and APNGs no longer consume ridiculous amounts of memory! Instead of holding all of the decoded frames in memory, the animation is streamed like a video. This should make Whimsy users very happy indeed.
Matt Woodrow got rid of an unnecessary hash table that we were building when computing DisplayLists (which is part of painting). Earlier this month, Matt also made it so that we don’t unnecessarily build layers if nothing on screen actually changed during an invalidation. For both bugs, this means Firefox is doing less work – and less work means getting pixels to the screen faster!
Emilio Cobos Álvarez made it so that we can style large text files (for example, this log file) much, much faster than before.
Dão Gottwald made it so that we only have to load some front-end CSS once instead of multiple times in different scopes. He also got rid of the tabbrowser XBL binding, which resulted in some wins on our Talos performance testing benchmark for painting newly opened tabs.
Gijs Kruitbosch got rid of some old, dead code that we were running early during start-up for no good reason. Gijs also made it so that Telemetry doesn’t make us load the blocklist. Finally, he made it so that we’re more likely to load the blocklist asynchronously rather than synchronously. All of these should help reduce browser start time!
Paolo Amadini got rid of some synchronous layout flushes that could be experienced when opening submenus within our animated menu panels.
Ryan VanderMeulen bumped our default Windows compiler to VS2017 15.6 for Firefox 61. Just switching has resulted in some benchmark improvements for our Windows builds!
- Not bad for a compiler upgrade!
Marco Bonardo made it so that we lazily load a Places script, rather than loading it automatically for every single window that gets opened.

Firefox Performance Update #2

So I’ve had my eyes out, watching for bugfixes that are landing in the Firefox code base that will speed it up for our users.

But I can’t do it alone. So I have a request: Have you seen a bugfix land recently that’ll likely impact Firefox’s performance in a positive way for our users? If so, I want to hear about it, and you should fill in this form to let me know!

Here’s that link again, just in case you missed it.

Anyhow, without further ado – here’s a small selection of interesting bugs that have been fixed recently that should improve Firefox performance!

Rob Wood has made it so that Talos tests run on Nightly builds now produce Gecko Profiles for both Linux and MacOS! Windows is coming soon. This is great for engineers working on improving our performance on our internal benchmarks.
Xidorn Quan enabled Stylo for chrome-level CSS documents! This allows us to process style changes to the browser UI in parallel using the same techniques that we use for web content that shipped in Firefox Quantum.
Kris Maglione made it so that we don’t load the Onboarding script for every single tab.
Jan de Mooij landed a patch that made browser JavaScript run more quickly and consume less memory early on during main and content process start-up
Henri Sivonen took an old and slower chunk of C++ code that checks for RTL characters in a string, and replaced it with a newer, faster version written in Rust. This has dramatically reduced the processing time on text nodes for pages where we’re unsure if RTL characters exist. For example, for strings in Thai, we now spend 3% of the original time searching those strings for RTL characters! That’s a huge improvement!
- Substantial!

A Blog by Mike Conley

The personal blog of a Toronto based software mechanic, musician, sound designer, and theatre enthusiast.

Monthly Archives: March 2018

Firefox Performance Update #4

Firefox Performance Update #3

Firefox Performance Update #2