codehaus


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Weird symbol resolution error with llvm JIT in gandiva jni



> On Dec 17, 2018, at 6:39 PM, Antoine Pitrou <antoine@xxxxxxxxxx> wrote:
> 
> 
> Is LoadLibraryPermanently(nullptr) called after the libgandiva_jni.so is
> loaded?  Perhaps you can try LoadLibraryPermanently("libgandiva_jni.so")
> explicitly?

We are already doing this inside gandiva.

https://github.com/apache/arrow/blob/master/cpp/src/gandiva/engine.cc#L61 <https://github.com/apache/arrow/blob/master/cpp/src/gandiva/engine.cc#L61>

> 
> 
> Le 17/12/2018 à 13:50, Ravindra Pindikura a écrit :
>> Antoine,
>> 
>> I set ARROW_GANDIVA_STATIC_LIBSTDCPP and modified the jni/CMakeLists.txt to export all symbols (i.e removed the version-script parameter in LINK_FLAGS).
>> 
>> $ nm ./debug/libgandiva_jni.so|grep  _ZNSt8ios_base4InitC1Ev
>> 0000000001df34f0 T _ZNSt8ios_base4InitC1Ev
>> 
>> $ nm --defined-only ./debug/libgandiva_jni.so|grep  _ZNSt8ios_base4InitC1Ev
>> 0000000001df34f0 T _ZNSt8ios_base4InitC1Ev
>> 
>> I still get the same “could not be resolved!” Error when I run the java test. It appears like when the library is loaded via jni, some symbols are getting hidden from the JIT.
>> 
>> Of the four sub-cases of the test :
>> 
>> 1. a cpp executable (shared/static), the decimal functions invoked from CPP fns
>> 2. a cpp executable (shared/static), the decimal functions invoked from JIT code
>> 3. a java executable, the decimal functions invoked from CPP Fns
>> 4. a java executable, the decimal functions invoked from JIT code
>> 
>> 1, 2 & 3 work good. Only 4 causes the symbol resolution issue.
>> 
>> Thanks & regards,
>> Ravindra.
>> 
>> 
>>> On Dec 16, 2018, at 6:34 PM, Antoine Pitrou <antoine@xxxxxxxxxx> wrote:
>>> 
>>> 
>>> In other words, see if:
>>> nm --defined-only libgandiva.a | grep _ZNSt8ios_base4InitC1Ev
>>> 
>>> outputs anything.
>>> 
>>> Regards
>>> 
>>> Antoine.
>>> 
>>> 
>>> Le 16/12/2018 à 14:03, Antoine Pitrou a écrit :
>>>> 
>>>> Does the static gandiva lib bundle libstdc++? Otherwise you'll probably
>>>> need to load it explicitly in the Java process.
>>>> 
>>>> Regards
>>>> 
>>>> Antoine.
>>>> 
>>>> 
>>>> Le 16/12/2018 à 04:53, Ravindra Pindikura a écrit :
>>>>> I think this happens if any of the cpp files that are used for IR generation (clang —emit-llvm) includes <iostream>. For decimal support, I’m generating IR code from arrow/util/decimal.cc <http://decimal.cc/> which includes <iostream> (many times in the header file chain).
>>>>> 
>>>>>> On Dec 16, 2018, at 9:14 AM, Ravindra Pindikura <ravindra@xxxxxxxxxx> wrote:
>>>>>> 
>>>>>> Hi,
>>>>>> 
>>>>>> I’ve been working on adding decimal support in gandiva - I’m hitting a symbol resolution error with std::ios_base::Init with the java tests on ubuntu. Any ideas on what I may be doing wrong ?
>>>>>> 
>>>>>> LLVM ERROR: Program used external function '_ZNSt8ios_base4InitC1Ev' which could not be resolved!
>>>>>> 
>>>>>> 
>>>>>> I’ve the same test (add two decimals) in three forms : 
>>>>>> 
>>>>>> 1. Cpp test executable with shared gandiva lib
>>>>>> 
>>>>>> Works on both ubuntu and Mac
>>>>>> 
>>>>>> 2. Cpp test executable with static gandiva lib
>>>>>> 
>>>>>> Works on both ubuntu and Mac
>>>>>> 
>>>>>> 3. Java test (internally uses jni over static gandiva lib)
>>>>>> 
>>>>>> Works on Mac, but fails on ubuntu !!
>>>>>> 
>>>>>> Ubuntu failure : https://travis-ci.org/apache/arrow/jobs/468562791
>>>>>> 
>>>>>> I tried switching to ORC JIT (from MC JIT) but that didn’t help. In the JIT, the symbol is not found in the process address space for ubuntu+java. For all the other cases, the symbol is found.
>>>>>> 
>>>>>> https://github.com/apache/arrow/pull/2942/files#diff-5ef29a5cabe384645dbf7a9c704aceeaR36
>>>>>> 
>>>>>> Thanks & regards,
>>>>>> Ravindra.
>>>>> 
>>>>> 
>>