[SciPy-dev] scipy_core Solaris Build: Good and Bad News

[Travis Oliphant]

Christopher Hanley wrote:

>>Can you run the code under gdb (or another debugger) to determine 
>>exactly where the bus error is occurring?
>>
>>    
>>
>
>Well, GDB core dumps with the Python we have installed on Solaris. 
>Using DBX this is the most useful error message I have been able to get 
>out so far:
>
> >>> scipy.test()
>   Found 2 tests for scipy.base.umath
>   Found 23 tests for scipy.base.function_base
>   Found 3 tests for scipy.base.getlimits
>   Found 9 tests for scipy.base.twodim_base
>   Found 3 tests for scipy.base.matrix
>   Found 44 tests for scipy.base.shape_base
>   Found 42 tests for scipy.base.type_check
>   Found 3 tests for scipy.basic.helper
>   Found 4 tests for scipy.base.index_tricks
>   Found 0 tests for __main__
>.......................signal BUS (invalid address alignment) in 
>ufunc_frompyfunc at 0xfeb56674
>0xfeb56674: ufunc_frompyfunc+0x01bc:    st      %l6, [%i5 + 0x4]
>
>I realize that this probably isn't particularly helpful.  Todd has 
>suggested I build my own private copy of Python with  --with-pydebug so 
>that I can better step through the code.
>  
>

Actually, this confirms my suspicion about the problem.  In this one 
function I'm playing
games to use only one malloc call as explained before.    This is now 
(hopefully) fixed in SVN.

I don't do the same thing anywhere else (well except for allocating 
buffers, but these are always increments of the data type size and so 
should be fine).

All the pointers should now be on aligned addresses.

-Travis