968931 - Assertion failure: false (MOZ_ASSUME_UNREACHABLE(Modified registers between VM call and OsiPoint)), at jit/shared/CodeGenerator-shared.cpp:536

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Christian Holler (:decoder)]

The following testcase asserts on mozilla-central revision 1e9f169c9715 (run with --fuzzing-safe --ion-eager --ion-regalloc=backtracking):


enableOsiPointRegisterChecks();
eval('(function () {\
var s = "{}";\
for (var i = 0; i < 19; i++)\
    s += s;\
    eval(s);\
})();\
');

Comment 1

[Christian Holler (:decoder)]

This problem could be related to --ion-regalloc=backtracking only, then it wouldn't be a security problem. Locking s-s until this is confirmed.

Comment 2

[Jan de Mooij [:jandem]]

I'm unable to reproduce this with the revision + shell flags in comment 0, x86 debug build on OS X.

NI from decoder, this one shouldn't be too hard to figure out once I have access to a machine where this reproduces.

Comment 3

[Christian Holler (:decoder)]

This still reproduces easily for me. Jan can you try the following config flags on a 32 bit Linux?

--disable-threadsafe --enable-debug --enable-optimize --enable-valgrind (+ 32 bit stuff)

Comment 4

[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/f030f97fcf10
user:        Dan Gohman
date:        Thu Oct 24 20:34:54 2013 -0700
summary:     Bug 875656 - IonMonkey: Juggle registers around to reduce the number of temporaries needed by LConcat. r=bhackett

This iteration took 380.333 seconds to run.

Comment 5

[Christian Holler (:decoder)]

Needinfo from :sunfish based on comment 4 :)

Comment 6

[Dan Gohman [:sunfish]]

Looks like a backtracking allocator bug. I'll investigate.

Comment 7

[Dan Gohman [:sunfish]]

This also needs --ion-parallel-compile=off to reproduce reliably. It does appear to be backtracking only.

Here's a simplified testcase that also fails with --ion-eager --ion-regalloc=backtracking --ion-parallel-compile=off on x86 linux:

(function () {
var s = "{}";
for (var i = 0; i < 19; i++)
    s += s;
    eval(s);
})();

The OsiPoint that triggers this error is the one immediately after the Concat. The mismatch is with %edx, which is Concat's output register.

Comment 8

[Christian Holler (:decoder)]

JSBugMon: Cannot process bug: Unknown exception (check manually)

Comment 9

[Hannes Verschore [:h4writer]]

Attachment: bug968931-backtracking-assert

Comment 10

[Jan de Mooij [:jandem]]

Comment on attachment 8396528 [details] [diff] [review]
bug968931-backtracking-assert

Review of attachment 8396528 [details] [diff] [review]:
-----------------------------------------------------------------

Thanks for fixing this!

::: js/src/jit/LiveRangeAllocator.h
@@ +689,5 @@
> +        if (interval->index() == 0 && !reg->isTemp()) {
> +#ifdef CHECK_OSIPOINT_REGISTERS
> +            // We don't add the output register to the safepoint,
> +            // but it still might get added as one of the inputs.
> +            // So eagerly add this reg to the safepoint clobbered register.

Nit: s/clobbered register/clobbered registers

@@ +691,5 @@
> +            // We don't add the output register to the safepoint,
> +            // but it still might get added as one of the inputs.
> +            // So eagerly add this reg to the safepoint clobbered register.
> +            LSafepoint *safepoint = reg->ins()->safepoint();
> +            if (safepoint)

Nit: if (LSafepoint *safepoint = reg->ins()->safepoint())

::: js/src/jit/shared/CodeGenerator-shared.cpp
@@ +564,2 @@
>      // temps after calling into the VM. This is fine because no other
> +    // instructions (including this OsiPoint) will depend on them. The same

Nit: finish last sentence

Comment 11

[Hannes Verschore [:h4writer]]

Opening the bug, since this is a assertion is over-restrictive.

Comment 12

[Hannes Verschore [:h4writer]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/2dc1dab474ab

Comment 13

[Wes Kocher (:KWierso) (Not reading bugmail; email directly if needed)]

https://hg.mozilla.org/mozilla-central/rev/2dc1dab474ab