950462 - Assertion failure: false (MOZ_ASSUME_UNREACHABLE(unexpected type)), at jit/Lowering.cpp:1696 or Crash [@ js::jit::LIRGenerator::visitToDouble]

Status: VERIFIED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Christian Holler (:decoder)]

The following testcase asserts on mozilla-central revision c049cb230d77 (run with --fuzzing-safe --ion-eager):


for (var i = 0; i < 3; i++)
  [,][i] = 3.14159;

Comment 1

[Christian Holler (:decoder)]

Attachment: [crash-signature] Machine-readable crash signature

Comment 2

[Christian Holler (:decoder)]

JSBugMon: Bisection requested, result:
autoBisect shows this is probably related to the following changeset:

The first bad revision is:
changeset:   http://hg.mozilla.org/mozilla-central/rev/1bd9d75fe43b
user:        Jan de Mooij
date:        Sat Dec 14 10:57:25 2013 +0100
summary:     Bug 949475 - Remove JOF_TYPESET from initializer ops. r=bhackett

This iteration took 341.696 seconds to run.

Comment 4

[Jan de Mooij [:jandem]]

Attachment: Patch

Yeah this is a regression from bug 949475 part 1, but the underlying bug is older. Before bug 949475, IonBuilder::jsop_newarray would do:

types::TemporaryTypeSet::DoubleConversion conversion =
    bytecodeTypes(pc)->convertDoubleElements(constraints());

Because we only monitored the result of JSOP_NEWARRAY if we had an array with singleton type, bytecodeTypes(pc) would be empty most of the time. NEWARRAY is no longer monitored and we no longer use bytecodeTypes(pc) here.

The problem is that the script is Ion-compiled multiple times. The first time we set the CONVERT_DOUBLE_ELEMENTS flag on the template object. The next time we don't set it but it's still set from the last compilation, and we end up trying to convert the magic hole value to double. This patch just clears the flag.

Comment 5

[Hannes Verschore [:h4writer]]

Comment on attachment 8356090 [details] [diff] [review]
Patch

Review of attachment 8356090 [details] [diff] [review]:
-----------------------------------------------------------------

r+ if you also do the same change to "jit/MCallOptimize.cpp" for inlineArray.

So if I correctly understand "templateObject->type()->convertDoubleElements() == AlwaysConvertToDoubles" should always agree with templateObject->getElementsHeader()::AlwaysConvertToDoubles (when we start using it). Since the second is derived from the first. Would it be too hard to adjust the header when the type changes? Because IIUC the type of the templateObject changes and as a result 'AlwaysConvertToDoubles' isn't correct anymore...

Comment 6

[Jan de Mooij [:jandem]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/b289ebad63f9

(In reply to Hannes Verschore [:h4writer], pto 13th - 17th of January from comment #5)
> So if I correctly understand
> "templateObject->type()->convertDoubleElements() == AlwaysConvertToDoubles"
> should always agree with
> templateObject->getElementsHeader()::AlwaysConvertToDoubles (when we start
> using it). Since the second is derived from the first. Would it be too hard
> to adjust the header when the type changes? Because IIUC the type of the
> templateObject changes and as a result 'AlwaysConvertToDoubles' isn't
> correct anymore...

Multiple objects can use the same TypeObject, so when a type is added, we don't know the template object with that type.

Comment 7

[Andrew McCreight [:mccr8]]

What security rating should this get, jandem?

Comment 8

[Jan de Mooij [:jandem]]

(In reply to Andrew McCreight [:mccr8] from comment #7)
> What security rating should this get, jandem?

We'll hit a MOZ_ASSUME_UNREACHABLE, not sure what compilers do there in opt builds. If we continue and return true somehow, other code may get confused but probably hard to exploit. Setting sec-moderate.

Comment 9

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/b289ebad63f9

Comment 10

[Christian Holler (:decoder)]

JSBugMon: This bug has been automatically verified fixed.