Re: [Swig-devel] generated code for Java breaks strict aliasing rules

[Bob M. <rma...@ee...>]

William S Fulton wrote:
> Bob Marinier wrote:
>   
>> I've attached a simple example demonstrating my problem.  When I build
>> it with SWIG 1.3.29, g++ 4.1.1 and the -O2 and -Wall flags, I get the
>> following warnings:
>>
>> $ swig -c++ -java example.i
>> $ g++ -c -fpic example_wrap.cxx -I/opt/sun-jdk-1.4.2.10/include
>> -I/opt/sun-jdk-1.4.2.10/include/linux -Wall -O2
>> example_wrap.cxx: In function 'jlong Java_exampleJNI_new_1Shape(JNIEnv*,
>> _jclass*)':
>> example_wrap.cxx:206: warning: dereferencing type-punned pointer will
>> break strict-aliasing rules
>> example_wrap.cxx: In function 'void
>> Java_exampleJNI_delete_1Shape(JNIEnv*, _jclass*, jlong)':
>> example_wrap.cxx:216: warning: dereferencing type-punned pointer will
>> break strict-aliasing rules
>>
>> If I finish building and test it (as shown below), everything works fine
>> in this simple case:
>>
>> $ g++ -shared example_wrap.o -o libexample.so
>> $ javac *.java
>> $ java main
>> Creating some objects:
>>    Created shape Shape@f5da06
>>
>> Guess I'll clean up now
>> Goodbye
>>
>> However, I have some real code that is a lot more complicated, and it
>> crashes if I compile with -O2.  Adding the -fno-strict-aliasing flag
>> "fixes" the problem, but this is undesirable because it disables some
>> optimizations.
>>
>> The offending lines in the example look look like this:
>>
>> *(Shape **)&jresult = result;
>>
>> To fix this, the line should be changed to:
>>
>> *(Shape **)(void *)&jresult = result;
>>
>> That is, add an intermediate cast to void*.  This eliminates the warning
>> (and presumably the crash, although I don't want to manually change
>> hundreds of lines of code to confirm that in my complex code, and I
>> haven't devised a simple test case that crashes.  However, we had some
>> similar code that we wrote ourselves with the same problem, and adding
>> the intermediate cast did fix it).
>>
>>     
>
> Adding the cast does not fix the problem. All it does is mask a useful
> compiler warning telling you that the code will crash. Look for the
> discussion on aliasing in Java.html for more info. If you see some slow
> down in the wrapper code due to using -fno-strict-aliasing, I'd like to
> see this. You only have to use -fno-strict-aliasing on the SWIG
> generated code. There is discussion of aliasing on the swig-devel
> mailing list earlier this year which you might find interesting.
>
> William
>   

Thanks for the info.  It turns out our old code isn't fixed after all :)

Bob