llvm dev - Nov 2010 - [LLVMdev] Bug in DragonEgg or LLVM

The following code using OpenMP pragmas , when compiled with gcc 4.5 + LLVM 2.8
+ DragonEgg 2.8 and ran, produces segmentation fault.

//-----------------------------------------------------------
#define LOOPCOUNT 10000

int main()
{
  int bit_and = 1;
  int logics[LOOPCOUNT];
  int i;

  for (i = 0; i < LOOPCOUNT; ++i)
  {
      logics[i] = 1;
  }

#pragma omp parallel for schedule(dynamic,1) reduction(&:bit_and)
  for (i = 0; i < LOOPCOUNT; ++i)
  {
      bit_and = (bit_and & logics[i]);
  }
}
//-----------------------------------------------------------

Examination with valgrind produces these results:

==5064== Thread 2:
==5064== Use of uninitialised value of size 8
==5064==    at 0x40074E: main.omp_fn.0 (in test113)
==5064==    by 0x4C44FE1: gomp_thread_start (team.c:115)
==5064==    by 0x3A1EA07760: start_thread (in /lib64/libpthread-2.12.1.so)
==5064==    by 0x3A1E2E14FC: clone (in /lib64/libc-2.12.1.so)
==5064== 
==5064== Invalid read of size 4
==5064==    at 0x40074E: main.omp_fn.0 (in test113)
==5064==    by 0x4C44FE1: gomp_thread_start (team.c:115)
==5064==    by 0x3A1EA07760: start_thread (in /lib64/libpthread-2.12.1.so)
==5064==    by 0x3A1E2E14FC: clone (in /lib64/libc-2.12.1.so)
==5064==  Address 0x0 is not stack'd, malloc'd or (recently) free'd
==5064== 
==5064== 
==5064== Process terminating with default action of signal 11 (SIGSEGV)
==5064==  Access not within mapped region at address 0x0
==5064==    at 0x40074E: main.omp_fn.0 (in test113)
==5064==    by 0x4C44FE1: gomp_thread_start (team.c:115)
==5064==    by 0x3A1EA07760: start_thread (in /lib64/libpthread-2.12.1.so)
==5064==    by 0x3A1E2E14FC: clone (in /lib64/libc-2.12.1.so)


As I understand the segmentation fault arises because in the file 
"team.c" in function "gomp_thread_start" the program tries
to read memory through an uninitialized pointer.

Has any body encountered this problem before or has any ideas how to fix it or
knows which particular part of DragonEgg or LLVM could be causing this issue?
I would be very grateful for any help provided.

Hi,
> The following code using OpenMP pragmas , when compiled with gcc 4.5 + LLVM
2.8 + DragonEgg 2.8 and ran, produces segmentation fault.
I just tried it on x86-64 linux with latest dragonegg from svn, and it executes
fine, with no valgrind complaints.  After compiling at -O0 valgrind does say

==7251== Syscall param exit_group(status) contains uninitialised byte(s)
==7251==    at 0x4ED8598: _Exit (_exit.c:33)
==7251==    by 0x4E674A4: exit (exit.c:93)
==7251==    by 0x4E4CD94: (below main) (libc-start.c:258)

but that's because you forgot to return a value from main, so rubbish is
returned as the exit code.

How did you build the program and on what platform?

Ciao,

Duncan.

Hi,

Are there some practical techniques or tools for how to debug this 
issue? Perhaps to visualise the graph, or report on how the cycle arose?

Cheers,
George