RFR: JDK-8167479: arm 32/64 slowdebug fails to build on unpack200

[David Holmes]

On 12/10/2016 9:57 PM, Magnus Ihse Bursie wrote:
> On 2016-10-12 13:07, David Holmes wrote:
>>
>>
>> On 12/10/2016 7:04 PM, Magnus Ihse Bursie wrote:
>>> On 2016-10-12 04:38, David Holmes wrote:
>>>> On 11/10/2016 7:41 PM, Erik Joelsson wrote:
>>>>> Hello,
>>>>>
>>>>> After the cleanup of C++ linking with the GCC toolchain. We missed
>>>>> updating one instance with the correct TOOLCHAIN configuration. This
>>>>> leads to build failures for arm ports with slowdebug. I have no
>>>>> idea why
>>>>> it only happens there, but correcting the TOOLCHAIN configuration (so
>>>>> that g++ is used to link instead of gcc) makes it build again.
>>>>
>>>> Unexplained errors raise alarm bells for me. This particular reference
>>>> relates to C++ exception handling which raises questions:
>>>>
>>>> - why only on this platform?
>>>> - why only in slowdebug?
>>>> - why only in this code? (which seems unrelated to exceptions)
>>>>
>>>> I agree the fix is to use g++ instead of gcc for linking, assuming the
>>>> exception related code should actually be there.
>>> The alarm bells ringed for me too. :-) I actually started debugging this
>>> one, but Erik beat me to it. :) I did however conclude that this error
>>> occured due to static linking (without it, everything went fine) and if
>>> I explicitely specified -lstdc++ as well, I got a more informative
>>> error, telling me that the static version of libgcc was not found. I
>>> didn't go further, since Erik found the (correct) solution, but I assume
>>> that the devkit we use to build for the arm platforms are missing some
>>> bits, which was only needed for slowdebug builds.
>>
>> Still too vague for me. What is special about this code?
> I have no idea. The question should probably be rephrased: "Why does
> this work on all the other platforms, when it's obviously incorrect?"
> And the answer to that is most likely "*shrug* ... pure luck?".

As I said there are three questions to answer:

 >>>> - why only on this platform?
 >>>> - why only in slowdebug?
 >>>> - why only in this code? (which seems unrelated to exceptions)

I just realized that "why this code" can be answered by this being the 
only bit of code not already being linked with g++. If we reverted the 
other g++ linker changes and ran this build I suspect we would see a lot 
more issues.

David

> /Magnus
>
>>
>> Thanks,
>> David
>>
>>> /Magnus
>>>
>>>>
>>>> Thanks,
>>>> David
>>>>
>>>>> The build error looks like this:
>>>>>
>>>>> /localhome/hg/jdk9-dev-CDR/build/linux-arm64-slowdebug/support/native/jdk.pack200/unpackexe/bands.o:(.data.DW.ref.__gxx_personality_v0[DW.ref.__gxx_personality_v0]+0x0):
>>>>>
>>>>>
>>>>> undefined reference to `__gxx_personality_v0'
>>>>> collect2: error: ld returned 1 exit status
>>>>>
>>>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8167479
>>>>>
>>>>> Webrev: http://cr.openjdk.java.net/~erikj/8167479/webrev.01/
>>>>>
>>>>> /Erik
>>>>>
>>>
>