551173 - SPARC failures in acceptance suite

Status: VERIFIED FIXED

(Tamarin Graveyard :: Baseline JIT (CodegenLIR), defect, P1)

Description

[Brent Baker]

After spot merging the following nanojit change into tamarin (https://bugzilla.mozilla.org/attachment.cgi?id=428637), running acceptance has a couple of new failures.

The following acceptance tests crash:
ecma3/Statements/etry_007.abc
e4x/XML/e13_4_2.abc
e4x/XML/e13_4_4_18.abc
e4x/XML/e13_4_4_19.abc
e4x/XML/e13_4_4_23.abc
e4x/XML/e13_4_4_3.abc

dbx avmshell_d
(dbx) run ecma3/Statements/etry_007.abc
t@1 (l@1) signal SEGV (no mapping at the fault address) in
avmplus::AvmCore::atomWriteBarrier at line 4377 in file "AvmCore.cpp"
 4377          decr_atom(*address);
 
All of the XML tests crash with this error, must have the JIT forced:
(dbx) run -Ojit acceptance/e4x/XML/e13_4_2.abc
t@1 (l@1) signal SEGV (no mapping at the fault address) in
avmplus::ScopeChain::getSize at line 96 in file "ScopeChain-inlines.h"
  96       return _scopeTraits->size;
 
 
In /DEBUGGER/ builds the following fails:
ecma3/Statements/block.abc
 
(dbx) run ecma3/Statements/block.abc
t@1 (l@1) signal SEGV (no mapping at the fault address) in
avmplus::ScopeTypeChain::getScopeIsWithAt at line 67 in file
"ScopeChain-inlines.h"
  67       return (_scopes[i] & ISWITH) != 0;

Comment 1

[Leon Sha]

These failures are caused by a compiler bug. The patch will come out in several weeks. I have an ugly workaround for this bug. I suggest we wait the compiler patch.

Comment 2

[Brent Baker]

Attachment: test/acceptance/testconfig.txt

mark the failures on solaris SPARC as known until the compiler bug is fixed

Comment 3

[Dan Smith]

Targeting for 10.1.1.

Comment 4

[Brent Baker]

Need to rebase the known failures now that tr and tr-argo have anew nanojit. In the recent merge I know that at least one additional testcase is causing a core dump: 

ecma3/Date/e15_9_5_4.as

Comment 5

[Brent Baker]

Submitted a new SPARC bug, #552650, it does not appear to be the same issue to me but cross commenting anyways.

Comment 7

[Brent Baker]

from bug 552650 
(In reply to comment #3)
> I just confirmed, it is the same bug as bug 551173. In mozilla tree we use -xO2
> to compile Assembler.cpp instead of -xO5 as a workaround.

I modified the configure script in tamarin to use -x02 on all files and was not able to workaround the problem.

Comment 8

[Brent Baker]

Leon has there been any update on when a SunStudio compiler patch will be released that addresses this issue? SPARC compiling and testing has been offline for a long time and I am a little concerned.

Comment 9

[Leon Sha]

Attachment: Use -xO2 instead of -xO5 to workaround sunstudio 12u1 bugs.

The latest patch for sunstudio 12u1 didn't include the fix for this bug. Sorry for the wrong information. I post this patch to use -xO2 instead of -xO5 as a workaround. I guess last -x02 not working because "-xO2" should be used not "-x02". It is the difference between "0" and "O". Brent, can you try this patch and have a review? Thanks a lot.

Comment 10

[Brent Baker]

Comment on attachment 438652 [details] [diff] [review]
Use -xO2 instead of -xO5 to workaround sunstudio 12u1 bugs.

Looks good. Patch works in tamarin-redux. I have run most of the acceptance passes at this stage cleanly, will report back once they are all done. 

Hand applying the patch to tamarin-redux-argo does not fix the issue, I would guess that there are some necessary changes in tr that are not in tr-argo in order to get SPARC working. 

QRB: what is the solaris SPARC schedule for fp 10.1? I am guessing that there needs to be some back porting of nanojit changes into tamarin-redux-argo so that we can get SPARC working in that branch properly.

Comment 11

[Leon Sha]

http://hg.mozilla.org/tamarin-redux/rev/461cb64b9048

Comment 12

[Leon Sha]

(In reply to comment #10)

> Hand applying the patch to tamarin-redux-argo does not fix the issue, I would
> guess that there are some necessary changes in tr that are not in tr-argo in
> order to get SPARC working. 
I guess tamarin-redux-argo using different build system other that tamarin-redux.

Comment 13

[Brent Baker]

(In reply to comment #12)
> I guess tamarin-redux-argo using different build system other that
> tamarin-redux.

tr-argo and tr use the same build system, however tamarin-redux-argo version of nanojit is currently at the following revision:
http://hg.mozilla.org/projects/nanojit-central/rev/3fa60904a28e

So I am not sure if the issue in tr-argo is because SPARC compilation requires additional changes passed this nanojit rev.

Will ping QRB directly about this issue.

Comment 14

[Brent Baker]

Hi Brent,

I have just try this tree. I applied the following three changesets.
http://hg.mozilla.org/tamarin-redux/rev/461cb64b9048
http://hg.mozilla.org/tamarin-redux/rev/203096443c6f
http://hg.mozilla.org/tamarin-redux/rev/8fa983fa3812
The acceptance test goes smoothly.

Regards,
Leon Sha

Brent:
I have confirmed that applying the above patches to tamarin-argo rev3974 will fix the SPARC shell and all acceptance runs pass (hybrid, interp, jit, debug, debugger, release etc). I am marking this issue as "REOPENED" until there is a decision from QRB about back porting this into argo

QRB: I recommend that these changes be back ported from redux into the argo branch.

Comment 15

[Trevor Baker]

recommend taking these changes to facilitate testing bug 558428

Comment 16

[Dan Smith]

Targeting for 10.1, requesting additional review by Edwin Smith.

Comment 17

[Brent Baker]

Patch pushed into tr-argo: 3983:13caa279c63a

Comment 18

[Brent Baker]

Found 2 issues in the testsuite:

debugger builds running the following testcases are EXTREMELY slow. I ended up killing the run after 5 minutes
- as3/sampling/Callback.abc
- as3/sampling/FunctionEntry.abc

Callback does slowly spit out callback information, but after 5 minutes I stopped it, should run much faster 
-bash-3.00$ $AVM as3/sampling/Callback.abc
Sampling api: callback
callback1 called at 360
callback1 called at 394
callback1 called at 394
callback1 called at 394
callback1 called at 394
callback1 called at 394
callback1 called at 394
callback1 called at 394

Comment 19

[Brent Baker]

There are also 3 failures in test/acceptance/spidermonkey/js1_5/Regress/regress-58116

Comment 20

[Brent Baker]

Marked testcases as skip (sampling) and as expectedfail (spidermonkey)

tr/tr-argo: 826390e43bfb
tr only: 093e73b8f3be

Comment 21

[Brent Baker]

(In reply to comment #19)
> There are also 3 failures in
> test/acceptance/spidermonkey/js1_5/Regress/regress-58116

These failures are tracked as bug #551227

Comment 22

[Trevor Baker]

Brent, please verify and resolve.  the slow debugger issue wa logged as bug 560536

Comment 23

[Brent Baker]

Solaris buildslave is back online compiling and running acceptance suite. Marking as Verified/fixed

Slow running tests are being skipped (tracked via 560536)