Localizing bugs in DIRAC run using debugger

The DIRAC program system is continuously evolving. From time to time, however, it gets - unintentionally - spoiled with various programming bugs. Here we would like to share knowledge on how to localize crashes/bugs in the code.

We believe that this short tutorial is useful not only for DIRAC developers, but also for ordinary users, who likewise could send us precise localization and specification of the bug.

Introduction

Let us assume that we have serially compiled DIRAC code. Make sure that DIRAC is compiled with “-g” flag for all Fortran, C and C++ compilers. Anyway, this flag is the the default part in all combination of compiler flags.

If you intend to run many debugging runs, please choose –type=Debug in the setup flags. This means omitting optimization flags by the CMake buildup system and speed up of (repeating) DIRAC compilation. Note that some bugs (crashes) might “disappear” when the optimization is completely turned off, because not all compilers have bug-free optimization procedure.

Ensure that proper debugger, installed on your machine, is called with the “pam” script. For example, for the GNU compilers set this is the gdb-debugger. Also commercial compilers are providing their own debuggers: Intel suite brings the idbg, the PGI suite has the pgdbg debugger etc.

Debugging with the “pam” script

Launching DIRAC with the “pam –debug ...” is suitable for one-shoot debugging. Pam script calls the debugger, which afterwards shows up its command line. Aftewards the user can insert breakpoints, stops into the code and repeatedly run ‘dirac.x’ inside the debugger.

Stack printing

There is the possibility to print out the stack content.

Any ‘QUIT’ crash of the code causes a SegFault termination you can investigate with the debugger - you can ‘visit’ individual procedures within the stack sequence and examine their variables.

For example, beeing the “gdb” debugger session:

> run
(Dirac runs and crashes somewhere)
> backtrace
(show the stack)
> up {n}
(go to the n-the previous function in the call stack; see help up)
> print <VARIABLE>
(whatever you want to inspect).

Oposite way can be achiewed using ‘down’ command.

Debugging in the plain mode with the GNU/gdb debugger

We give a simple demonstrative session of an thorough and effective debugging. Most recent version of the widely used the GNU/gdb debugger is recommended.

Make a working directory with enough disk space. Provide a link of the $DIRAC/dirac.x executable file there. This is because you will repeat modifications of DIRAC source files and subsequent compilation, so it is easier to have linked executable rather than copying it.

Make sure that the ‘DIRAC.INP’, ‘MOLECULE.INP’, gdb_demo and other necessary files (for example, the ‘DFCOEF’ if you are restarting SCF iterations) are in the working directory as well.

Demonstration of GNU/gdb debugging commands

This gdb_demo file (create&copy from this web-page) contains some commands (and macros) of the gdb debugger:

###  load the executable file with all symbols
file dirac.x

### This settings is necessary for complex breakpoint conditions in fortran...
#set language fortran
set language c
set language auto

### Write where I am...
echo In directory:
shell echo pwd
echo \n Files in dir:
shell ls -a

##################################################################
### Delete  DIRAC files in the scratch directory
### - needed for restarting and for further catching of bugs..
##################################################################
shell /bin/rm -r DF*  AOPROPER BSSMAT
echo \n After deletion remaining files in dir:
shell ls -a

### clean the desk: delete all previous breakpoints
d b

###################################
# Example of macros
###################################
define my_macro
printf "in GMOTRA begin NZ=%d N2BBASXQ=%d SPINFR=%d\n",nz,n2bbasxq,spinfr
end

define my_macro1
printf "in GMOTRA begin SPINFR2=%d\n",spinfr2
end

### demo of a conditional breakpoint
b gmotra_ if (nz==1 | (n2bbasxq==0 && spinfr==0))
command
 my_macro
 my_macro1
 echo type continue or sipmly c...\n
# c
end

### example of the unconditional breakpoint
b dirone.F:1399
command
 if ( (nfsym.eq.1 && n2bbasxq.gt.10) | nz.eq.1 | spinfr.eq.0)
  printf "NFSYM=%d\n", NFSYM
  if (nfsym.eq.2)
    printf "NTMO(1)=%d NTMO(2)=%d\n",NTMO(1),NTMO(2)
  else
    printf "One symmetry...NTMO(1)=%d \n",NTMO(1)
  end
 else
  printf "... else branch...."
 end

 ##### while cycle demo #####
 printf" Few values of the WORK(KTMAT) array, KTMAT=%d... \n",KTMAT
 set $indx=0
 while ($indx.lt.5)
  printf "WORK(%d)=%lf\n",ktmat+$indx,(*WORK)(ktmat+$indx)
  #  p ktmat+$indx
  set $indx=$indx+1
 end
end

In the working space type gdb dirac.x to run the dirac.x executable in the debugging mode. When the debugger’s command line appears, type source gdb_demo to load the debugger source file.

Each time when you modify and recompile DIRAC please retype again file dirac.x in the gdb mode to reload the fresh executable.

If you modify the gdb_demo debugger source file apply again the gdb command source gdb_demo to reload it.

More GNU/gdb know-how is given at this web-page.

Other debuggers

To debug the dirac.x program compiled with the Intel suite we recommended to employ own Intel debugger, idb, which is works with the same set of commands as the GNU/gdb counterpart. The same holds for Portland compilers. Combining different compilers with different debuggers does not work well.

For parallel debugging we recommend special (commercial) software, which is able to follow individual threads, like the totalview debugger. Otherwise in a multithreaded dirac.x run each thread can be caught and debugged with any serial debugger.