You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@harmony.apache.org by Salikh Zakirov <Sa...@Intel.com> on 2006/11/22 18:00:51 UTC

[drlvm][x86_64] DRLVM broken after HARMONY-2223

Hi,

DRLVM is broken on Linux/x86_64 (SUSE9) after following commit

 [r478172] Applied HARMONY-2223 [drlvm][jit][performance] Inlined int64
arithmetic
    routines - speedup of 64 arithmetics. Tested on SUSE9, observed ~5%
    gain on the microbenchmark.

with the following diagnostics:

/files/sszakiro/harmony/drlvm/trunk/vm/port/src/encoder/ia32_em64t/enc_tabl.cpp:1540:
static void EncoderBase::buildMnemonicDesc(const MnemonicInfo*):
Assertion `opcodesHashMap[minfo->mn][hash] == NOHASH' failed.

This assertion failure is observed on C Unit tests and when trying to
run java with any arguments.


Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Alex Astapchuk <al...@gmail.com>.
Tim Ellison wrote:
> Alex Astapchuk wrote:
>> the fix attached to the H-2223.
>> Sorry again for not tested on intel64.
>>
>> p.s. gosh!..
>> is there any way to make the classlib build in separate dirs like drlvm
>> does??
>>
>> it really wastes a huuuuge time to rebuild all the classlib thingies -
>> when I
>> want to test on separate platforms - <sigh> or switching configs
>> back and forth and back again.
> 
> Are you working on a shared mount? Trying to figure out how you have
> multiple architectures on a single workspace.

Yes, I normally keep workspace on NFS mount and have samba-mapped it on 
Win machines.



-- 
Thanks,
   Alex


Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by "Geir Magnusson Jr." <ge...@pobox.com>.
w00t!

Tim Ellison wrote:
> Gregory Shimansky wrote:
>> Tim Ellison wrote:
>>> Alex Astapchuk wrote:
>>>> the fix attached to the H-2223.
>>>> Sorry again for not tested on intel64.
>>>>
>>>> p.s. gosh!..
>>>> is there any way to make the classlib build in separate dirs like drlvm
>>>> does??
>>>>
>>>> it really wastes a huuuuge time to rebuild all the classlib thingies -
>>>> when I
>>>> want to test on separate platforms - <sigh> or switching configs
>>>> back and forth and back again.
>>> Are you working on a shared mount? Trying to figure out how you have
>>> multiple architectures on a single workspace.
>> It is quite common to have the same home directory to be mounted on NFS
>> among several different machines.
> 
> Understood -- just checking we are talking about the same problem.  I
> have no objection to putting the build artifacts into arch-specific subdirs.
> 
> Regards,
> Tim
> 

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Tim Ellison <t....@gmail.com>.
Gregory Shimansky wrote:
> Tim Ellison wrote:
>> Alex Astapchuk wrote:
>>> the fix attached to the H-2223.
>>> Sorry again for not tested on intel64.
>>>
>>> p.s. gosh!..
>>> is there any way to make the classlib build in separate dirs like drlvm
>>> does??
>>>
>>> it really wastes a huuuuge time to rebuild all the classlib thingies -
>>> when I
>>> want to test on separate platforms - <sigh> or switching configs
>>> back and forth and back again.
>>
>> Are you working on a shared mount? Trying to figure out how you have
>> multiple architectures on a single workspace.
> 
> It is quite common to have the same home directory to be mounted on NFS
> among several different machines.

Understood -- just checking we are talking about the same problem.  I
have no objection to putting the build artifacts into arch-specific subdirs.

Regards,
Tim

-- 

Tim Ellison (t.p.ellison@gmail.com)
IBM Java technology centre, UK.

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Gregory Shimansky <gs...@gmail.com>.
Tim Ellison wrote:
> Alex Astapchuk wrote:
>> the fix attached to the H-2223.
>> Sorry again for not tested on intel64.
>>
>> p.s. gosh!..
>> is there any way to make the classlib build in separate dirs like drlvm
>> does??
>>
>> it really wastes a huuuuge time to rebuild all the classlib thingies -
>> when I
>> want to test on separate platforms - <sigh> or switching configs
>> back and forth and back again.
> 
> Are you working on a shared mount? Trying to figure out how you have
> multiple architectures on a single workspace.

It is quite common to have the same home directory to be mounted on NFS 
among several different machines.

-- 
Gregory


Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Tim Ellison <t....@gmail.com>.
Alex Astapchuk wrote:
> the fix attached to the H-2223.
> Sorry again for not tested on intel64.
> 
> p.s. gosh!..
> is there any way to make the classlib build in separate dirs like drlvm
> does??
> 
> it really wastes a huuuuge time to rebuild all the classlib thingies -
> when I
> want to test on separate platforms - <sigh> or switching configs
> back and forth and back again.

Are you working on a shared mount? Trying to figure out how you have
multiple architectures on a single workspace.

Regards,
Tim

-- 

Tim Ellison (t.p.ellison@gmail.com)
IBM Java technology centre, UK.

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Alex Astapchuk <al...@gmail.com>.
the fix attached to the H-2223.
Sorry again for not tested on intel64.

p.s. gosh!..
is there any way to make the classlib build in separate dirs like drlvm 
does??

it really wastes a huuuuge time to rebuild all the classlib thingies - 
when I
want to test on separate platforms - <sigh> or switching configs
back and forth and back again.

Any chance?

-- 
Thanks,
   Alex





Salikh Zakirov wrote:
> Hi,
> 
> DRLVM is broken on Linux/x86_64 (SUSE9) after following commit
> 
>  [r478172] Applied HARMONY-2223 [drlvm][jit][performance] Inlined int64
> arithmetic
>     routines - speedup of 64 arithmetics. Tested on SUSE9, observed ~5%
>     gain on the microbenchmark.
> 
> with the following diagnostics:
> 
> /files/sszakiro/harmony/drlvm/trunk/vm/port/src/encoder/ia32_em64t/enc_tabl.cpp:1540:
> static void EncoderBase::buildMnemonicDesc(const MnemonicInfo*):
> Assertion `opcodesHashMap[minfo->mn][hash] == NOHASH' failed.
> 
> This assertion failure is observed on C Unit tests and when trying to
> run java with any arguments.
> 
> 




Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Alex Astapchuk <al...@gmail.com>.
Geir Magnusson Jr. wrote:
> 
> 
> Alex Astapchuk wrote:
>> Geir Magnusson Jr. wrote:
>>> I have a small problem with this - seems like we're rushing patches 
>>> in without much review or discussion.   For a while there we were 
>>> able to test on a few platforms first, lazily, because there was more 
>>> than 1 hour between patch submission and commit.
>>
>> Actually, the patch was there for a week - since Nov.17.
> 
> Really? When I took a look at the comment and patch log, things were 
> going fast and furious on 23 Nov, including a patch at 5:35am.

It's still all the same patch since the Nov. 17. It was onyl attached
several times to the same JIRA.
In the final version I only added few comment words.


>>
>> Thought, it's still my fault that I did not check the em64t build :-(
>> Sorry about that, I'm in working on the fix right now.
> 
> It's ok - things will break.  I was just want to make sure activities 
> like this remain visible on the dev list.
> 
> geir
> 
>>
> 


-- 
Thanks,
   Alex


Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by "Geir Magnusson Jr." <ge...@pobox.com>.

Alex Astapchuk wrote:
> Geir Magnusson Jr. wrote:
>> I have a small problem with this - seems like we're rushing patches in 
>> without much review or discussion.   For a while there we were able to 
>> test on a few platforms first, lazily, because there was more than 1 
>> hour between patch submission and commit.
> 
> Actually, the patch was there for a week - since Nov.17.

Really? When I took a look at the comment and patch log, things were 
going fast and furious on 23 Nov, including a patch at 5:35am.

> 
> Thought, it's still my fault that I did not check the em64t build :-(
> Sorry about that, I'm in working on the fix right now.

It's ok - things will break.  I was just want to make sure activities 
like this remain visible on the dev list.

geir

> 

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Alex Astapchuk <al...@gmail.com>.
Geir Magnusson Jr. wrote:
> I have a small problem with this - seems like we're rushing patches in 
> without much review or discussion.   For a while there we were able to 
> test on a few platforms first, lazily, because there was more than 1 
> hour between patch submission and commit.

Actually, the patch was there for a week - since Nov.17.

Thought, it's still my fault that I did not check the em64t build :-(
Sorry about that, I'm in working on the fix right now.

-- 
Thanks,
   Alex

> 
> Was there any on-list discussion about this?  did I simply miss it?
> 
> geir
> 
> 
> Salikh Zakirov wrote:
>> Hi,
>>
>> DRLVM is broken on Linux/x86_64 (SUSE9) after following commit
>>
>>  [r478172] Applied HARMONY-2223 [drlvm][jit][performance] Inlined int64
>> arithmetic
>>     routines - speedup of 64 arithmetics. Tested on SUSE9, observed ~5%
>>     gain on the microbenchmark.
>>
>> with the following diagnostics:
>>
>> /files/sszakiro/harmony/drlvm/trunk/vm/port/src/encoder/ia32_em64t/enc_tabl.cpp:1540: 
>>
>> static void EncoderBase::buildMnemonicDesc(const MnemonicInfo*):
>> Assertion `opcodesHashMap[minfo->mn][hash] == NOHASH' failed.
>>
>> This assertion failure is observed on C Unit tests and when trying to
>> run java with any arguments.
>>
> 




Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Alexey Varlamov <al...@gmail.com>.
Let ne answer in one message:
Salikh,
indeed I did not test on real x86_64 OS, only tried on 32-bit OSes on
x86 and x86_64 hardware. Will try to resolve this ASAP, sorry for
inconvenience!

Geir,
I would not call it rush - the JIRA was there for a few days actually,
with all the patches. Maybe I'm a bit too enthusiastic to see our perf
scores improved, but I really don't think it went headlong.

Alexei,

I generally agree that increased caution should be exercised, but
don't think we should say "full stop" now.

--
Regards,
Alexey

2006/11/23, Alexei Fedotov <al...@gmail.com>:
> Since many guys reported different regressions, I believe it is good
> time to delay a bit DRLVM features and optimizations and accept only
> regression bug fixes. What do you think?
>
> The first thing to do is to understand what's broken. To my
> understanding the most obvious thing which is broken is Vladimir's
> CruiseControl which runs tests with DRLVM. I do not see automatic
> notifications.
>
> I believe there are more regressions. For example, in my yesterday's
> test run 400 class library tests failed on Linux. The guy who
> descriptively named himself "test" reported the same statistics at
> http://harmonytest.org. I need some time to stop and mediate about it
> (and file JIRAs of course).
>
> --
> Thank you,
> Alexei
>
> On 11/22/06, Geir Magnusson Jr. <ge...@pobox.com> wrote:
> > I have a small problem with this - seems like we're rushing patches in
> > without much review or discussion.   For a while there we were able to
> > test on a few platforms first, lazily, because there was more than 1
> > hour between patch submission and commit.
> >
> > Was there any on-list discussion about this?  did I simply miss it?
> >
> > geir
> >
> >
> > Salikh Zakirov wrote:
> > > Hi,
> > >
> > > DRLVM is broken on Linux/x86_64 (SUSE9) after following commit
> > >
> > >  [r478172] Applied HARMONY-2223 [drlvm][jit][performance] Inlined int64
> > > arithmetic
> > >     routines - speedup of 64 arithmetics. Tested on SUSE9, observed ~5%
> > >     gain on the microbenchmark.
> > >
> > > with the following diagnostics:
> > >
> > > /files/sszakiro/harmony/drlvm/trunk/vm/port/src/encoder/ia32_em64t/enc_tabl.cpp:1540:
> > > static void EncoderBase::buildMnemonicDesc(const MnemonicInfo*):
> > > Assertion `opcodesHashMap[minfo->mn][hash] == NOHASH' failed.
> > >
> > > This assertion failure is observed on C Unit tests and when trying to
> > > run java with any arguments.
> > >
> >
>

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Alexei Fedotov <al...@gmail.com>.
Alexey,

Sorry I've missed your question.

I got 400 different test failures in several subsequent test runs.
Many tests failed one or two times. The difference from Tatiana's
report could be explained in a following way: I ran tess in one JVM.

Today I've got better results, though a test run isn't complete yet
SuSE 9
22     [junit] TEST
org.apache.harmony.tests.internal.net.www.protocol.http.HttpURLConnectionTest
FAILED
 2     [junit] TEST javax.swing.JCheckBoxMenuItemTest FAILED

Win2003 server
     23     [junit] TEST
org.apache.harmony.tests.internal.net.www.protocol.http.HttpURLConnectionTe
st FAILED
     23     [junit] TEST
org.apache.harmony.prefs.tests.java.util.prefs.PreferencesTest FAILED
     23     [junit] TEST
org.apache.harmony.nio.tests.java.nio.channels.SocketChannelTest
FAILED
      2     [junit] TEST javax.swing.Timer_MultithreadedTest FAILED

-- 
Thank you,
Alexei

On 11/22/06, Alexey Varlamov <al...@gmail.com> wrote:
> 23 Nov 2006 12:20:05 +0600, Egor Pasko <eg...@gmail.com>:
> > On the 0x22A day of Apache Harmony Alexei Fedotov wrote:
> > > Since many guys reported different regressions, I believe it is good
> > > time to delay a bit DRLVM features and optimizations and accept only
> > > regression bug fixes. What do you think?
> >
> > I suggest to revert HARMONY-2223 until Alex has a better fix. Alex
> > is working on it now.
> >
> > > The first thing to do is to understand what's broken. To my
> > > understanding the most obvious thing which is broken is Vladimir's
> > > CruiseControl which runs tests with DRLVM. I do not see automatic
> > > notifications.
> >
> > when is it going to be repaired? Can we afford to stop implementing
> > features until then?
> >
> > > I believe there are more regressions. For example, in my yesterday's
> > > test run 400 class library tests failed on Linux. The guy who
> > > descriptively named himself "test" reported the same statistics at
> > > http://harmonytest.org. I need some time to stop and mediate about it
> > > (and file JIRAs of course).
> >
> > that's weird, I ran Classlib tests on Linux yesterday, and they passed
> > in default mode with a known exception: HttpURLConnectionTest. Maybe,
> > the reason of your observation is some recent commit (maybe, even
> > HARMONY-2223)?
>
> I've just rechecked - got 18669/1/2 cases, with all failures known.
> Alexei, could you please confirm your result?
>
> >
> > > --
> > > Thank you,
> > > Alexei
> > >
> > > On 11/22/06, Geir Magnusson Jr. <ge...@pobox.com> wrote:
> > > > I have a small problem with this - seems like we're rushing patches in
> > > > without much review or discussion.   For a while there we were able to
> > > > test on a few platforms first, lazily, because there was more than 1
> > > > hour between patch submission and commit.
> > > >
> > > > Was there any on-list discussion about this?  did I simply miss it?
> > > >
> > > > geir
> > > >
> > > >
> > > > Salikh Zakirov wrote:
> > > > > Hi,
> > > > >
> > > > > DRLVM is broken on Linux/x86_64 (SUSE9) after following commit
> > > > >
> > > > >  [r478172] Applied HARMONY-2223 [drlvm][jit][performance] Inlined int64
> > > > > arithmetic
> > > > >     routines - speedup of 64 arithmetics. Tested on SUSE9, observed ~5%
> > > > >     gain on the microbenchmark.
> > > > >
> > > > > with the following diagnostics:
> > > > >
> > > > > /files/sszakiro/harmony/drlvm/trunk/vm/port/src/encoder/ia32_em64t/enc_tabl.cpp:1540:
> > > > > static void EncoderBase::buildMnemonicDesc(const MnemonicInfo*):
> > > > > Assertion `opcodesHashMap[minfo->mn][hash] == NOHASH' failed.
> > > > >
> > > > > This assertion failure is observed on C Unit tests and when trying to
> > > > > run java with any arguments.
> > > > >
> > > >
> > >
> >
> > --
> > Egor Pasko
> >
> >
>

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Alexey Varlamov <al...@gmail.com>.
23 Nov 2006 12:20:05 +0600, Egor Pasko <eg...@gmail.com>:
> On the 0x22A day of Apache Harmony Alexei Fedotov wrote:
> > Since many guys reported different regressions, I believe it is good
> > time to delay a bit DRLVM features and optimizations and accept only
> > regression bug fixes. What do you think?
>
> I suggest to revert HARMONY-2223 until Alex has a better fix. Alex
> is working on it now.
>
> > The first thing to do is to understand what's broken. To my
> > understanding the most obvious thing which is broken is Vladimir's
> > CruiseControl which runs tests with DRLVM. I do not see automatic
> > notifications.
>
> when is it going to be repaired? Can we afford to stop implementing
> features until then?
>
> > I believe there are more regressions. For example, in my yesterday's
> > test run 400 class library tests failed on Linux. The guy who
> > descriptively named himself "test" reported the same statistics at
> > http://harmonytest.org. I need some time to stop and mediate about it
> > (and file JIRAs of course).
>
> that's weird, I ran Classlib tests on Linux yesterday, and they passed
> in default mode with a known exception: HttpURLConnectionTest. Maybe,
> the reason of your observation is some recent commit (maybe, even
> HARMONY-2223)?

I've just rechecked - got 18669/1/2 cases, with all failures known.
Alexei, could you please confirm your result?

>
> > --
> > Thank you,
> > Alexei
> >
> > On 11/22/06, Geir Magnusson Jr. <ge...@pobox.com> wrote:
> > > I have a small problem with this - seems like we're rushing patches in
> > > without much review or discussion.   For a while there we were able to
> > > test on a few platforms first, lazily, because there was more than 1
> > > hour between patch submission and commit.
> > >
> > > Was there any on-list discussion about this?  did I simply miss it?
> > >
> > > geir
> > >
> > >
> > > Salikh Zakirov wrote:
> > > > Hi,
> > > >
> > > > DRLVM is broken on Linux/x86_64 (SUSE9) after following commit
> > > >
> > > >  [r478172] Applied HARMONY-2223 [drlvm][jit][performance] Inlined int64
> > > > arithmetic
> > > >     routines - speedup of 64 arithmetics. Tested on SUSE9, observed ~5%
> > > >     gain on the microbenchmark.
> > > >
> > > > with the following diagnostics:
> > > >
> > > > /files/sszakiro/harmony/drlvm/trunk/vm/port/src/encoder/ia32_em64t/enc_tabl.cpp:1540:
> > > > static void EncoderBase::buildMnemonicDesc(const MnemonicInfo*):
> > > > Assertion `opcodesHashMap[minfo->mn][hash] == NOHASH' failed.
> > > >
> > > > This assertion failure is observed on C Unit tests and when trying to
> > > > run java with any arguments.
> > > >
> > >
> >
>
> --
> Egor Pasko
>
>

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Egor Pasko <eg...@gmail.com>.
On the 0x22A day of Apache Harmony Alexei Fedotov wrote:
> Since many guys reported different regressions, I believe it is good
> time to delay a bit DRLVM features and optimizations and accept only
> regression bug fixes. What do you think?

I suggest to revert HARMONY-2223 until Alex has a better fix. Alex
is working on it now.

> The first thing to do is to understand what's broken. To my
> understanding the most obvious thing which is broken is Vladimir's
> CruiseControl which runs tests with DRLVM. I do not see automatic
> notifications.

when is it going to be repaired? Can we afford to stop implementing
features until then?

> I believe there are more regressions. For example, in my yesterday's
> test run 400 class library tests failed on Linux. The guy who
> descriptively named himself "test" reported the same statistics at
> http://harmonytest.org. I need some time to stop and mediate about it
> (and file JIRAs of course).

that's weird, I ran Classlib tests on Linux yesterday, and they passed
in default mode with a known exception: HttpURLConnectionTest. Maybe,
the reason of your observation is some recent commit (maybe, even
HARMONY-2223)?

> -- 
> Thank you,
> Alexei
> 
> On 11/22/06, Geir Magnusson Jr. <ge...@pobox.com> wrote:
> > I have a small problem with this - seems like we're rushing patches in
> > without much review or discussion.   For a while there we were able to
> > test on a few platforms first, lazily, because there was more than 1
> > hour between patch submission and commit.
> >
> > Was there any on-list discussion about this?  did I simply miss it?
> >
> > geir
> >
> >
> > Salikh Zakirov wrote:
> > > Hi,
> > >
> > > DRLVM is broken on Linux/x86_64 (SUSE9) after following commit
> > >
> > >  [r478172] Applied HARMONY-2223 [drlvm][jit][performance] Inlined int64
> > > arithmetic
> > >     routines - speedup of 64 arithmetics. Tested on SUSE9, observed ~5%
> > >     gain on the microbenchmark.
> > >
> > > with the following diagnostics:
> > >
> > > /files/sszakiro/harmony/drlvm/trunk/vm/port/src/encoder/ia32_em64t/enc_tabl.cpp:1540:
> > > static void EncoderBase::buildMnemonicDesc(const MnemonicInfo*):
> > > Assertion `opcodesHashMap[minfo->mn][hash] == NOHASH' failed.
> > >
> > > This assertion failure is observed on C Unit tests and when trying to
> > > run java with any arguments.
> > >
> >
> 

-- 
Egor Pasko


Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Vladimir Ivanov <iv...@gmail.com>.
On 11/24/06, Anton Luht <an...@gmail.com> wrote:
>
> Hello,
>
> > Probably we should introduce tags to differentiate test sets. What do
> > you think? (This would require from you to set a proper tag from
> > CruiseControl when uploading results).
>
> I've added 'hut' (Harmony unit tests) and 'eut' (Eclipse unit tests),
> tagged existing runs and put tags to the first page (list of runs).
> Hope that Vladimir will add tags  to Cruise Control.


not to CC - just to the upload script :)


--
> Regards,
> Anton Luht,
> Intel Java & XML Engineering
>

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Alexei Fedotov <al...@gmail.com>.
Thanks! I like it!
-- 
Thank you,
Alexei

On 11/24/06, Anton Luht <an...@gmail.com> wrote:
> Hello,
>
> > Probably we should introduce tags to differentiate test sets. What do
> > you think? (This would require from you to set a proper tag from
> > CruiseControl when uploading results).
>
> I've added 'hut' (Harmony unit tests) and 'eut' (Eclipse unit tests),
> tagged existing runs and put tags to the first page (list of runs).
> Hope that Vladimir will add tags  to Cruise Control.
>
>
> --
> Regards,
> Anton Luht,
> Intel Java & XML Engineering
>

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Anton Luht <an...@gmail.com>.
Hello,

> Probably we should introduce tags to differentiate test sets. What do
> you think? (This would require from you to set a proper tag from
> CruiseControl when uploading results).

I've added 'hut' (Harmony unit tests) and 'eut' (Eclipse unit tests),
tagged existing runs and put tags to the first page (list of runs).
Hope that Vladimir will add tags  to Cruise Control.


-- 
Regards,
Anton Luht,
Intel Java & XML Engineering

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Vladimir Ivanov <iv...@gmail.com>.
On 11/24/06, Alexei Fedotov <al...@gmail.com> wrote:
>
> Anton, thanks!
>
> Vladimir,
> Probably we should introduce tags to differentiate test sets. What do
> you think? (This would require from you to set a proper tag from
> CruiseControl when uploading results).


I told Anton some time ago that we need to differ test suites. If it will be
by tag - OK.
It will cost almost nothing for me.


On 11/23/06, Anton Luht <an...@gmail.com> wrote:
> > Hello, Alexei,
> >
> > > The guy who
> > > descriptively named himself "test" reported the same statistics at
> > > http://harmonytest.org. I need some time to stop and mediate about it
> > > (and file JIRAs of course).
> >
> > At harmonytest.org , those runs that have significant failure rate
> > (>10) are not HUT but EUT (Eclipse tests) results.
> >
> > 'Test' is me - I've tested how EUT results are imported :)
> >
> > --
> > Regards,
> > Anton Luht,
> > Intel Java & XML Engineering
> >
>
>
> --
> Thank you,
> Alexei
>

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Alexei Fedotov <al...@gmail.com>.
Anton, thanks!

Vladimir,
Probably we should introduce tags to differentiate test sets. What do
you think? (This would require from you to set a proper tag from
CruiseControl when uploading results).


On 11/23/06, Anton Luht <an...@gmail.com> wrote:
> Hello, Alexei,
>
> > The guy who
> > descriptively named himself "test" reported the same statistics at
> > http://harmonytest.org. I need some time to stop and mediate about it
> > (and file JIRAs of course).
>
> At harmonytest.org , those runs that have significant failure rate
> (>10) are not HUT but EUT (Eclipse tests) results.
>
> 'Test' is me - I've tested how EUT results are imported :)
>
> --
> Regards,
> Anton Luht,
> Intel Java & XML Engineering
>


-- 
Thank you,
Alexei

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Anton Luht <an...@gmail.com>.
Hello, Alexei,

> The guy who
> descriptively named himself "test" reported the same statistics at
> http://harmonytest.org. I need some time to stop and mediate about it
> (and file JIRAs of course).

At harmonytest.org , those runs that have significant failure rate
(>10) are not HUT but EUT (Eclipse tests) results.

'Test' is me - I've tested how EUT results are imported :)

-- 
Regards,
Anton Luht,
Intel Java & XML Engineering

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by Alexei Fedotov <al...@gmail.com>.
Since many guys reported different regressions, I believe it is good
time to delay a bit DRLVM features and optimizations and accept only
regression bug fixes. What do you think?

The first thing to do is to understand what's broken. To my
understanding the most obvious thing which is broken is Vladimir's
CruiseControl which runs tests with DRLVM. I do not see automatic
notifications.

I believe there are more regressions. For example, in my yesterday's
test run 400 class library tests failed on Linux. The guy who
descriptively named himself "test" reported the same statistics at
http://harmonytest.org. I need some time to stop and mediate about it
(and file JIRAs of course).

-- 
Thank you,
Alexei

On 11/22/06, Geir Magnusson Jr. <ge...@pobox.com> wrote:
> I have a small problem with this - seems like we're rushing patches in
> without much review or discussion.   For a while there we were able to
> test on a few platforms first, lazily, because there was more than 1
> hour between patch submission and commit.
>
> Was there any on-list discussion about this?  did I simply miss it?
>
> geir
>
>
> Salikh Zakirov wrote:
> > Hi,
> >
> > DRLVM is broken on Linux/x86_64 (SUSE9) after following commit
> >
> >  [r478172] Applied HARMONY-2223 [drlvm][jit][performance] Inlined int64
> > arithmetic
> >     routines - speedup of 64 arithmetics. Tested on SUSE9, observed ~5%
> >     gain on the microbenchmark.
> >
> > with the following diagnostics:
> >
> > /files/sszakiro/harmony/drlvm/trunk/vm/port/src/encoder/ia32_em64t/enc_tabl.cpp:1540:
> > static void EncoderBase::buildMnemonicDesc(const MnemonicInfo*):
> > Assertion `opcodesHashMap[minfo->mn][hash] == NOHASH' failed.
> >
> > This assertion failure is observed on C Unit tests and when trying to
> > run java with any arguments.
> >
>

Re: [drlvm][x86_64] DRLVM broken after HARMONY-2223

Posted by "Geir Magnusson Jr." <ge...@pobox.com>.
I have a small problem with this - seems like we're rushing patches in 
without much review or discussion.   For a while there we were able to 
test on a few platforms first, lazily, because there was more than 1 
hour between patch submission and commit.

Was there any on-list discussion about this?  did I simply miss it?

geir


Salikh Zakirov wrote:
> Hi,
> 
> DRLVM is broken on Linux/x86_64 (SUSE9) after following commit
> 
>  [r478172] Applied HARMONY-2223 [drlvm][jit][performance] Inlined int64
> arithmetic
>     routines - speedup of 64 arithmetics. Tested on SUSE9, observed ~5%
>     gain on the microbenchmark.
> 
> with the following diagnostics:
> 
> /files/sszakiro/harmony/drlvm/trunk/vm/port/src/encoder/ia32_em64t/enc_tabl.cpp:1540:
> static void EncoderBase::buildMnemonicDesc(const MnemonicInfo*):
> Assertion `opcodesHashMap[minfo->mn][hash] == NOHASH' failed.
> 
> This assertion failure is observed on C Unit tests and when trying to
> run java with any arguments.
>