Re: Seg-fault in gsl-0.5/integration/qpsrt.c

[Brian Gough <bjg at network-theory dot co dot uk>]

Thanks -- those expressions do look like a problem (it was all
translated from the fortran which accounts for the strange logic).
I'll fix that and extend the tests to see if these sorts of things can
be caught with Checkergcc.

To turn off inline functions when building the library comment out the
#define HAVE_INLINE that appears in config.h after running ./configure.

Brett Viren writes:
 > Hello.
 > 
 > I guess this is for Brian Gough.
 > 
 > I think I spotted a problem in gsl-0.5/integration/qpsrt.c.
 > 
 > It turned up when using gsl_integration_qag to integrate a 5
 > dimensional integral containing various gaussian and poisson functions
 > along most dimensions.  
 > 
 > I had a bit of trouble getting gdb to halt just before failure because
 > the segfault comes after several thousand calls to
 > gsl_integration_qag.  Also, I had some problems understanding what was
 > going on due to inlined functions and optimization (can these be
 > turned off?).  But anyways, letting the segfault occur when running
 > under gdb pointed out the err as beeing at qpsrt.c:79 which is
 > right...
 > 
 >   while (errmin >= elist[order[k]] && k > i - 2)
 >     { 
 >       order[k+1] = order[k] ; 
 >       k-- ; 
 >     }			/* <---- ...here */
 > 
 > I don't understand why gdb sees the segfault at the end of the while
 > loop and I am also a bit confused at the logic in this file, but I
 > tried reversing the two tests, namely:
 > 
 >   while (k > i - 2 && errmin >= elist[order[k]]) 
 > 
 > and the segfault no longer occured.  To me, this seems the proper
 > order regardless, (keeping in mind I don't really understand the logic
 > of qpsrt.c <grin>).  There could also be the same problem at the while
 > loop just above on line 61:
 > 
 >   while (errmax < elist[order[i]] && i < top) 
 > 
 > But it was not being tickled by my code.
 > 
 > I tried to better isolate the problem to a test code, but was unable
 > to duplicate the condition.  If anyone wants my whole code to test it,
 > I'll supply it.
 > 
 > 
 > Thanks,
 > -Brett.
 > 
 > PS: I am not on this list, so I'll check for any responce in the
 > archive, or else just CC me.
 >