Crash in `- [SWWindowContentsService _windowUpdateNotification:]`

[numist]

Reported by @bavarious, who adds that DictationIM crashes shortly after Switch. Both are Accessibility applications, but it may be a red herring since Switch is crashing in code unrelated to the Accessibility APIs. I suspect that Switch crashing may have left AX in a bad enough mood to take it out on others.

From the report:

I believe the problem happened after I typed ↑ whilst the context switcher was visible.

Unable to reproduce locally; maybe something else on the system was bound to ⌥↑?

From the crash group:

Thread 9 Crashed:
0   libobjc.A.dylib                      0x00007fff8375d7a2 objc_retain + 18
1   Switch                               0x00000001032d74c6 __53-[SWWindowContentsService _windowUpdateNotification:]_block_invoke (SWWindowContentsService.m:169)
2   libdispatch.dylib                    0x00007fff823a21bb _dispatch_call_block_and_release + 12
3   libdispatch.dylib                    0x00007fff8239f28d _dispatch_client_callout + 8
4   libdispatch.dylib                    0x00007fff823a1673 _dispatch_queue_drain + 451
5   libdispatch.dylib                    0x00007fff823a29c1 _dispatch_queue_invoke + 110
6   libdispatch.dylib                    0x00007fff823a0f87 _dispatch_root_queue_drain + 75
7   libdispatch.dylib                    0x00007fff823a2177 _dispatch_worker_thread2 + 40
8   libsystem_pthread.dylib              0x00007fff833d2ef8 _pthread_wqthread + 314
9   libsystem_pthread.dylib              0x00007fff833d5fb9 start_wqthread + 13

-Os seems to have collapsed some of the stack frames, but that code is dispatching windowContentService:updatedContent:forWindow: to all of the window content service's subscribers. Not sure which pointer was bad when objc_retain got to it, but it looks like a dead object problem.

This crash happened at least twice in less than an hour for one person, so it appears to be reproducible when conditions are right.

[numist]

Sounds like this isn't as reproducible as I'd hoped. Multi-dispatch hasn't been without its problems (#65 and numist/NNKit#9 come immediately to mind) so it's not unreasonable to expect that there's still some way for an object to slip through the cracks.

[numist]

But if there is, I can't find it yet.

Here's the line at 169:

[(id<SWWindowContentsSubscriber>)self.subscriberDispatcher windowContentService:self updatedContent:contentContainer.content forWindow:contentContainer.window];

The crash report adds:

Selector name found in current argument registers: subscriberDispatcher

SWWindowContentsService.m assembly from v0.0.8 when that selector is in the register bank:

mov        rsi, qword [ds:0x100031488]                           ; @selector(subscriberDispatcher)
call       r14                                                   ; imp___got__objc_msgSend (cached)
mov        rdi, rax                                              ; move result to arg1 of next call
call       imp___stubs__objc_retainAutoreleasedReturnValue       ; crash, presumably.
mov        qword [ss:rbp-0x50+var_16], rax
mov        rax, qword [ds:r13+0x28]
mov        qword [ss:rbp-0x50+var_24], rax
mov        rsi, qword [ds:0x1000315a8]                           ; subscriberDispatcher no longer in registers

subscriberDispatcher itself looks like:

push       rbp
mov        rbp, rsp
mov        rax, qword [ds:_OBJC_IVAR_$_NNService._subscriberDispatcher]
mov        rax, qword [ds:rdi+rax]
pop        rbp
ret

There's not a whole lot to go wrong here; the property is readonly and set in NNService's initializer.

[numist]

SWAppDelegate should implement [BITCrashManagerDelegate applicationLogForCrashManager:] to provide more context for this (and future) crashes.

[numist]

Obsoleted by #139, which removes the NNKit dependency.