Saturday, July 25, 2015

Updating you on 38 just-in-time

Did you see what I did there? For the past two weeks my free time apart from work and the Master's degree has been sitting in a debugger trying to fix JavaScript, which is just murder on my dating life. Here is the current showstopper bug-roll for 38.1.1b1:

  • The Faceblech bug with the new IonPower JavaScript JIT compiler is squashed, I think, after repairing some conformance test failures which in turn appear to have repaired Forceblah. In my defence, the two bugs in question were incredibly weird edge cases and these tests are not part of the usual JIT test suite, so I guess we'll have to run them as well in future. This also repairs an issue with Instagrump which is probably the same underlying issue since Faceboink owns them also.

    The silver lining after all that was that I was considering disabling inlining in the JIT prior to release, which worked around the "badness," but also cut the engine speed in about half. (Still faster than JaegerMonkey!) To make this a bit less of a hit, I tuned the thresholds for starting the twin JITs and got about 10% improvement without inlining. With inlining back on, it's still faster by about 4% and change -- the G5 now achieves a score of nearly 5800 on V8, up from 5560. I also tweaked our foreground finalization patch for generational GC so that we should be able to get the best of both worlds. Overall you should see even better performance out of this next beta.

  • I have a presumptive fix for the webfont "ATSUI puke" on the New York Times, but it's not implemented or well-tested yet. This is a crash on 10.5, so I consider it a showstopper and it will be fixed before the next beta. (It affects 31.8 also but I will not be making another 31 release unless there is a Mozilla ESR chemspill.)

  • The modified strip7 tool required for building 38.x has a serious bug in it that causes it to crash trying to strip certain symbols. I have fixed this bug and builders will need to install this new version (remember: do not replace your normal strip with this one; it is intentionally loose with the Mach-O specification). I will be uploading it sometime this week along with an updated gdb7 that has better debugger performance and repairs a bug with too eagerly disabling register display while single-stepping Ion code.

These bugs are not considered showstoppers, but I do acknowledge them and I plan to fix them either for the final release or the next version of 38:

  • I can confirm saved passwords do not appear in the preferences panel. They do work, though, and can be saved, so this is more of an issue with managing them; while it's possible to do so manually it requires some inconvenient screwing around with your profile, so I consider this the highest priority of the non-showstopper bugs.

  • Checkboxes on the dropdown menus from the Console tabs do not appear. This specific manifestation is purely cosmetic because they work normally otherwise, but this may be an indication there is a similar issue with dropdowns and context menus elsewhere, so I do want to fix this as well.

Other miscellaneous changes include some adjustments to HTML5 media streaming and I have decided to reduce the default window and tab undos back to 31's level (6 and 2 respectively) so that the browser still gives up tenured memory a bit more easily. Unfortunately, there is not enough time to get MP3 support fully functional for final release. I plan to get this completed in a future version of 38.x, but it will not be officially supported until then (you can still toggle tenfourfox.mp3.enabled to use the minimp3 driver for those sites it does work with as long as you remember that seeking within a track doesn't work yet).

The localizer elves have French, German, Spanish, Italian, Russian and Finnish installers available. Our Japanese localization appears to have dropped off the web, so if you can help us, o-negai shimasu! Swedish just needs a couple of strings to be finished. We do not yet have Polish or Asturian, which we used to, so if you can help on any of these languages, please visit issue 42 where Chris is coordinating these efforts. A big thank you to all of our localizers!

Once the localizations are all in, the Google Code project will be frozen to prepare for the wiki and issue tracker moving to Github ahead of Google Code going read-only on 24 August. Downloads will remain on SourceForge, but everything else will go to Github, including the source tree when we eventually drop source parity. I was hoping to have an Elcapitanspoof up in time for 38's final release, but we'll see if I have time to do the graphics.

Watch for the next beta to come out by next weekend with any luck, which gives us enough time if there needs to be a third emergency release prior to the final (weekend prior to 11 August).

Finally, I am pleased to note we are now no longer the only PowerPC JavaScript JIT out there, though we are the only one I know of for Mozilla SpiderMonkey. IBM has been working on a port of Google V8 to PowerPC for some time, both AIX and Linux, which recently became an official part of the Google V8 repository (i.e., the PPC port is now officially supported). If you've been looking at nabbing a POWER8 with that money burning a hole in your pocket, it even works with the new Power ISA little endian mode, of which we dare not speak. Since uppsala, Floodgap's main server, is a POWER6 running AIX and should be able to run this, I might give it a spin sometime when I have a few spare cycles. However, before some of the freaks amongst you get excited and think this means Google Chrome on OS X/ppc is just around the corner, there's still an awful lot more work required to get it operational than just the JavaScript engine, and it won't be me that works on it. It does mean, however, that things like node.js will now work on a Power-based server with substantially less fiddling around, and that might be very helpful for those of you who run Power boxes like me.

Monday, July 6, 2015

Beta 1 aftermath

So let's sum up the first 38 beta.

Confirmed bugs: Facesuck seems to be totally whacked in Ion mode (it works fine in Baseline only mode). IonPower passes all the JIT tests, though, so this must be something that the test suite does not cover. I'm investigating some other conformance test suites and corrected a couple other variances so far between Baseline and Ion but none of them appear to be what's ailing Faceblech yet.

Also, we have another web font that makes ATSUI puke, except it has an inconveniently null PostScript name so we can't filter it with the existing method. Fortunately Tobias had come up with an alternative font filter system some time ago that should work with 10.4.

Not confirmed (yet?): a few people have reported that memory usage skyrockets upon quit and the browser crashes (inevitably after exceeding its addressing space), on a variety of systems, both 10.4 and 10.5. I can't reproduce this on any of the test machines.

I need to do more looking into the stored passwords question.

Since we're out of runway, i.e., ESR31, and we need one more beta before release, I'm going to keep working on the Facebork problem (or at least try to fix it by fixing something else) until July 24. If we can't do it by then, I guess we launch without IonPower, which is unfortunate and will regress JavaScript performance, but we will still at least have Baseline. Faceburp is just too big to debug in place, so I need you folks to find another site that has similar problems. I haven't been able to yet myself.