608130 - a few ecma3 tests SegFault with ppc64 jit when compiled with -AS3

Status: RESOLVED WONTFIX

(Tamarin Graveyard :: Baseline JIT (CodegenLIR), defect)

Description

[Chris Peyer]

The followin tests segfault on ppc64 non-interp:

ecma3/Boolean/e15_6_4_3_2.abc
ecma3/Boolean/e15_6_4_3_3.abc
ecma3/Expressions/e11_2_1_2.abc
ecma3/Expressions/e11_2_1_5.abc
ecma3/ObjectObjects/e15_2_2_1_rt.abc
ecma3/ObjectObjects/e15_2_1_1_rt.abc
ecma3/TypeConversion/e9_9_1_rt.abc

Result:
Segmentation Fault
exit code: -11

They must be compiled with -AS3.  To do this, move the ecma3/dir.asc_args file to somewhere else temporarily then run the tests using -f.

I can also provide the .abc files if needed.

Comment 1

[Brent Baker]

Attachment: gdb stacktrace

This is a stacktrace from running Boolean/e15_6_4_3_2.abc on mac PPC64.

This only happens when the jit is enabled.

Comment 2

[Brent Baker]

This is NOT an injection from the latest NJ merge:
http://hg.mozilla.org/tamarin-redux/rev/12e82003a10d

Comment 3

[Brent Baker]

(In reply to comment #2)
> This is NOT an injection from the latest NJ merge:
> http://hg.mozilla.org/tamarin-redux/rev/12e82003a10d

Same issue also happens in argo version of avmshell (ftp://10.60.48.47/milestones/tamarin-argo/4035-973d551be753/mac/)

Comment 4

[Brent Baker]

These tests are currently passing with the exact-gc patch queue on mac64-ppc. Lars suggested having Edwin take a look at this issue and I will be sending machine/source information to Edwin offlien.

Lars: "That suggests that the bug shows up in part because of heap dynamics, and since those are changing with exact tracing the bug might have hidden itself again.  Might want to ping Edwin about looking for the bug sooner rather than later."

Comment 5

[Brent Baker]

After looking at this some more, it is not true that all of the tests are now passing. They no longer crash the VM, but the vm is now not producing the correct results for all of the tests:

ecma3/Expressions/e11_2_1_2.abc : false.valueOf() =  FAILED! expected: false
  ecma3/Expressions/e11_2_1_2.abc : false.toString() = true FAILED! expected: false
  ecma3/Expressions/e11_2_1_5.abc : false.valueOf() =  FAILED! expected: false
  ecma3/Expressions/e11_2_1_5.abc : false.toString() = true FAILED! expected: false
  ecma3/ObjectObjects/e15_2_1_1_rt.abc : Object(false).valueOf() =  FAILED! expected: false

I have tried to isolate the failing code but it does not appear to fail in the same manner when the code is isolated.

Comment 6

[Brent Baker]

QRB: ping, this needs to be reviewed and targeted as Lars is about to land the exact-gc patch which will alter this issue again.

Comment 7

[Tamarin Bot]

changeset: 5666:0b62f92f7169
user:      Lars T Hansen <[e-mail]>
summary:   For 608130: skip some tests on non-interp PPC64 because they are simply not stable (r=lhansen)

http://hg.mozilla.org/tamarin-redux/rev/0b62f92f7169

Comment 8

[Edwin Smith]

I haven't had any time to look at this, and probably won't for a while.  How important is it?  (its targeted for Serrano, but has no priority set... uh oh).  The looming question is who will be maintaining PPC64.

Do we have good evidence this is PPC64 only and not PPC32?

Comment 9

[Andre Kruetzfeldt]

PPC is no longer supported.