Thursday, September 18, 2014

TenFourFox 33 sort of gets off the ground

This is being typed in TenFourFox 33, which bodes well, but this was a very hard port and I don't have full confidence in it. For one thing, it was a very large merge and overall I had to backout several changes that would break us. While irregexp, the new regular expression engine for JavaScript derived from Google V8, does appear to mostly work after some substantial time spent sniffing out all the test failures and endian problems, it has one big problem in that it can't grow its backtrack stack, used for complex match jobs, without crashing. I suspect an OS glitch; I can't find anything wrong with my generated code yet. Right now the system simply uses a large enough default stack to pass the tests and refuses to enlarge it further. This seems to work, but I can't guarantee it works with every possible addon (Ghostery in particular has had past problems with limitations on regular expressions) and it requires several megabytes of space by default even if it never actually uses it, potentially per compartment. If memory or compatibility pressure becomes too high we can run or at least force Firefox to run without native regular expression compilation, but although that works fine it is very slow by comparison.

In addition, WebRTC's new screen capture code doesn't compile at all on 10.4. I think I have it hacked so that it will work without it and we can still use WebRTC for webcams as we are doing now, but it's all very rickety. The rapid development of that code means further bustage in the near future.

Finally, the one thing I was really hoping for out of Fx 33+, asynchronous pan and zoom where you can move around and scroll the page at least while it is loading and JavaScript is running, won't compile on 10.4 either. Tiger completely lacks any method, even an undocumented one, for turning a CGEvent from an event tap into an NSEvent, and I am so far not able to find documentation on the opaque CGEvent object to be able to fake up enough of it to get scrolling to function. The old CGSEventRecord does apparently have undocumented support, but I can't find any documentation on that either, just some references in code to header files that do not exist in the SDKs that come with Xcode 2.5.

Overall, although it works, I'm not happy with what I had to do to make it work. I'm not sure I'm going to make a release of 33 (other than changesets) -- I might turn this around immediately into aurora 34 and try to get cracking on converting the MIPS IonMonkey JIT because if the stack limitations of irregexp become a big problem, I may need the rest of the Ion JIT to overcome the performance hit we will take by disabling native regular expressions. I think this should be the biggest priority for 38, the next ESR. There may not be much further we can take the browser after that if we even get that far.

No comments:

Post a Comment

Due to an increased frequency of spam, comments are now subject to moderation.