674391 - TI: Assertion failure: Unpossible rejoin!, at methodjit/InvokeHelpers.cpp:1328

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Christian Holler (:decoder)]

Attachment: Test case for shell (see README file inside).

The attached testcase asserts on TI revision bd40ad1be9d8 (run main.js with options -j -m -n -a). This seems very complex, cannot get it into a single file, but got it greatly reduced from its original form.

Comment 1

[Brian Hackett [Laid off!]]

WFM, I just get:

ReferenceError: expect is not defined
ReferenceError: startTest is not defined

Can you paste the stack traces from the last few calls to js_GC, and/or the output when setting the JMFLAGS environment variable to 'recompile'?

Comment 2

[Christian Holler (:decoder)]

Attachment: Stacks of last three calls to js_GC

Here are the stacks of the last three calls to js_GC. Will attach the other information afterwards. If that is not enough, I can also simply provide you a shell with gdb so you can debug it on the fuzzing box. It's an Ubuntu (64 bit).

Comment 3

[Christian Holler (:decoder)]

Attachment: Output of JMFLAGS="recompile"

Comment 4

[Brian Hackett [Laid off!]]

(In reply to comment #2)
> Created attachment 549326 [details]
> Stacks of last three calls to js_GC
> 
> Here are the stacks of the last three calls to js_GC. Will attach the other
> information afterwards. If that is not enough, I can also simply provide you
> a shell with gdb so you can debug it on the fuzzing box. It's an Ubuntu (64
> bit).

Ah, this does it, the problem is that js_IteratorNext can GC and throw us into the interpreter, but we don't rejoin from calls to it.  Thought that this function couldn't GC, maybe I should be eating my own dogfood though and using the reports for GC callers at sixgill.mozilla.org, since it's pretty clear what's going on there.

Comment 5

[Brian Hackett [Laid off!]]

http://hg.mozilla.org/projects/jaegermonkey/rev/e5b57c9ebbe9

Side note: IterNext used REJOIN_NONE instead of REJOIN_FALLTHROUGH because it used to be called in the middle of JSOP_FOR* opcodes.  Without the fix from bug 648175 that removed these opcodes and made ITERNEXT its own opcode, being able to rejoin from an IterNext would have been extremely complicated.

Comment 6

[Christian Holler (:decoder)]

A testcase for this bug was automatically identified at js/src/jit-test/tests/jaeger/recompile/bug674391.js.