692215 - IM: Assertion failure: backing.firstCopy == stackPosition_, at ion/MIRGraph.cpp:362

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Christian Holler (:decoder)]

The following testcase asserts on ionmonkey revision acf3c1fb7c94 (run with --ion-eager), tested on 64 bit:


function test1() {
    var src = "switch(x) {\n";
    for (var i=-1; i<4; i++) {
        src += (i >= 0) ? src : "default:\n";
    }
}
test1();

Comment 1

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: fix

We were missing code to handle the case where we set a stack slot, but it's in the middle of a copy list. The copy list would become corrupted. This patch rewrites the nodes around the entry that is about to be overwritten.

Comment 2

[Chris Leary [:cdleary] (not checking bugmail)]

Comment on attachment 566384 [details] [diff] [review]
fix

Review of attachment 566384 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/ion/MIRGraph.cpp
@@ +226,5 @@
>          slots_[lowest].copyOf = NotACopy;
>          slots_[lowest].firstCopy = prev;
> +    } else if (var.isCopy()) {
> +        uint32 prev = var.copyOf;
> +        if (slots_[prev].firstCopy != slot) {

I think we would hit this case less if we iterated over the stack positions backwards in the addPredecessor. Probably doesn't matter though.

@@ +234,5 @@
> +                prev = slots_[prev].nextCopy;
> +                JS_ASSERT(prev != NotACopy);
> +            }
> +        }
> +        slots_[prev].firstCopy = var.nextCopy;

I kind of dislike this because we're abusing the firstCopy/nextCopy union, which is confusing to read. When you take the conditional you're actually clobbering a nextCopy field, and when you don't you're really setting the firstCopy field. Can we do a positive test? Like,

  uint32 backing = var.copyOf;
  if (slots_[backing].firstCopy == slot) {
    slots_[backing].firstCopy = var.nextCopy;
  } else {
    // ... iterative thing to set nextCopy on a list member to var.nextCopy...
  }

That way if they were made into non-union members it would still work alright.

Comment 3

[David Anderson [:dvander] - inactive, e-mail if emergency]

> That way if they were made into non-union members it would still work alright.

Thanks, I think this is clearer too.

http://hg.mozilla.org/projects/ionmonkey/rev/c7c0d34dc0d9

Comment 4

[Christian Holler (:decoder)]

A testcase for this bug was automatically identified at js/src/jit-test/tests/ion/bug692215.js.