RFR: 8019779 JDK 8 build failed due to hotspot crashed on Solaris 10u10 sparcv9/sparc with SS12u3 compiler

[Lois Foltan]

On 7/25/2013 2:46 PM, Vladimir Kozlov wrote:
> On 7/25/13 11:20 AM, Lois Foltan wrote:
>> Hi Vladimir,
>>
>> Thank you again for your review.  I do have an update.  The change is
>> currently on hold in order to further analyze the performance impact of
>> reducing the optimization level for these two files.
>
> Was you able to narrow down the code which is compiled incorrectly? 
> May be you can avoid the problem by changing that code slightly 
> without performance impact.
Yes, we are making progress on narrowing in on the inlining of certain 
functions that look suspect.  I agree, another potential option maybe to 
change the code.

Thank you,
Lois

> Thanks,
> Vladimir
>
>>
>> Thanks,
>> Lois
>>
>> On 7/24/2013 12:34 PM, Vladimir Kozlov wrote:
>>> Thank you, Lois, for clarification.
>>>
>>> Changes are good.
>>>
>>> Thanks,
>>> Vladimir
>>>
>>> On 7/24/13 9:20 AM, Lois Foltan wrote:
>>>>
>>>> On 7/24/2013 11:37 AM, Vladimir Kozlov wrote:
>>>>> Lois,
>>>>>
>>>>> Usually such problems happened only with fastdebug build. You lowed
>>>>> optimization level for product and optimized
>>>>> builds. Did it also fail with these versions? Sorry, I can't see bug
>>>>> report.
>>>>>
>>>> Hi Vladimir,  thank you for your review.  The bug was originally
>>>> reported against a fastdebug build, however, I was also
>>>> able to easily reproduce the failure with the product build as well.
>>>> Thus the necessity of making this change for
>>>> fastdebug, optimized and product.
>>>>
>>>>> Also you kept lower opt level for jvmtiClassFileReconstituter.o (\>=
>>>>> 510). Does it still have problem with SS12u3?
>>>>>
>>>> That is unknown, so I chose the safer route to continue to include
>>>> the optimization reduction for
>>>> jvmtiClassFileReconstituter.o for SS12u3. There also seemed to be
>>>> some historical precedence within the make files to
>>>> use (\>=<compiler version #>)  for past compiler version #'s when
>>>> bumping up to a new compiler version. Certainly, if
>>>> the C++ compiler optimization bug that plagued
>>>> jvmtiClassFileReconstituter.o is fixed, then it can be easily be 
>>>> changed
>>>> back to just reduced for 510.  However, I don't know if we can make
>>>> that assertion at this point.
>>>>
>>>>> Thanks,
>>>>> Vladimir
>>>>>
>>>>> On 7/24/13 7:49 AM, Lois Foltan wrote:
>>>>>> Please review the following fix:
>>>>>>
>>>>>>      open webrev at http://cr.openjdk.java.net/~hseigel/bug_8019779
>>>>>>
>>>>>> Bug: JDK 8 build failed due to hotspot crashed on Solaris 10u10
>>>>>> sparcv9/sparc with SS12u3 compiler
>>>>>>
>>>>>>      bug link at https://jbs.oracle.com/bugs/browse/JDK-8019779
>>>>>>
>>>>>> Summary of fix:
>>>>>>
>>>>>>      The JDK 8 build on Solaris using the new SS12u3 (CC V5.12)
>>>>>> compiler
>>>>>>      failed with a Hotspot crash at the point the build executes 
>>>>>> rmic.
>>>>>>      This crash was tracked down to a C++ compiler optimization 
>>>>>> issue
>>>>>>      when two specific files are compiled with -xO4.  As a work
>>>>>> around fix,
>>>>>>      knock down the optimization level of these two files
>>>>>> specifically for
>>>>>>      SS12u3. This bug will be reported/transferred to the C++
>>>>>> compiler in BugDB.
>>>>>>
>>>>>> Test Builds:
>>>>>>     Based on jdk8/build forests:
>>>>>>      JDK 8 full build with C++ SS12u1 with
>>>>>> --with-debug-level=[release and fastdebug] on Solaris sparc
>>>>>>      JDK 8 full build with C++ SS12u3 with
>>>>>> --with-debug-level=[release and fastdebug] on Solaris sparc
>>>>>>
>>>>>>     Based on hotspot-rt:
>>>>>>      Built Hotspot fastdebug, optimized, product with C++ SS12u1 on
>>>>>> Solaris sparcv9 and Solaris Intel
>>>>>>      Built Hotspot debug, fastdebug, optimized, product with C++
>>>>>> SS12u3 on Solaris sparcv9 and Solaris Intel
>>>>>>
>>>>>> Tests:
>>>>>>     JDK 8 full release built with C++ SS12u3 on Solaris sparc - ran
>>>>>> Hotspot's jtreg tests
>>>>>>     JDK 8 full fastdebug built with C++ SS12u3 on Solaris sparc -
>>>>>> ran JCK full test suite
>>>>>>     Hotspot fastdebug built with C++ SS12u3 on Solaris sparcv9 -
>>>>>> ran vm.quick.testlist
>>>>>>
>>>>>>
>>>>>> Thank you,
>>>>>> Lois
>>>>
>>