[Webkit-unassigned] [Bug 103146] ARMv7 replaceWithJump ASSERT failure after r135330.

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=103146





--- Comment #15 from Filip Pizlo <fpizlo at apple.com>  2013-03-20 22:15:19 PST ---
(In reply to comment #14)
> > FWIW I was seeing this crash in QNX, not Linux, so I suspect the fix should be more generic.
> 
> Did you try the patch? Does it fix your platform? If so, we can make it more generic by adding your OS to the list.

I think that the reason why you guys are seeing this badness is that you're using the ExecutableAllocator and not ExecutableAllocatorFixedVMPool, or whatever it is called.

We used the FixedVMPool on both X86_64 and ARMv7.  We do it for two simple reasons:

- About ~16MB is more than enough, since JSC only JITs things when it absolutely needs to.

- Having all JIT memory within a confined pool means that we can consistently use compact jumps.

Maybe the right thing for y'all is to switch to using the fixed VM pool?

Maybe the right thing for the project is to switch *everything* over to the fixed pool, including x86-32, so we can simplify the code and all share the invariant that for a given JSGlobalData, any two slabs of JIT code will always be within a small enough distance from each other that a relatively compact jump can be emitted?  (Note that for example the choice of 16MB on ARMv7, and 1GB on X86_64, is a direct consequent of the largest jumpable distance using a single jump instruction, on those platforms.)

-- 
Configure bugmail: https://bugs.webkit.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.