seg fault in parallel loop using a common block variable in call - depends on ifort version

hallevison · ‎12-11-2012

Hi:

I am getting a sef fault that I do not understand. Perhaps the best thing is to show you the structure of the code.

[fortran]

integer i

     real*8 rfblow,rfbhi,drfb,rsmall(NTPMAX),vrelv,mass(NTPMAX)
     real*16 rsmall16(NTPMAX),ratq,rsmr16

      common/c_rsm/rfblow,rfbhi,drfb,rsmall,rsmall16
[/fortran]

later

[fortran]

!$OMP PARALLEL DO SCHEDULE(DYNAMIC) DEFAULT(NONE)

!$OMP& PRIVATE(i,rsmr16,vrelv,ratq)

!$OMP& SHARED(rsmall16,mass)

do i=2,nbod

[/fortran]

later in the same loop

[fortran]
                     write(*,*) 'Here #C0.1',i
                     write(*,*) 'Here #C0.3',rsmr16
                     write(*,*) 'Here #C0.2',rsmall16(i)
                     write(*,*) 'Here #C0.4',vrelv
                     write(*,*) 'Here #C0.5',mass(i)
                     write(*,*) 'Here #C0.6',ratq

                     call size_qstarrat(i,rsmall16(i),rsmr16,vrelv,
     &                    mass(i),ratq)
[/fortran]

Then I have the subroutine

[fortran]      subroutine size_qstarrat(id,rtarg,rimp,vrel2,mtr,ratq)

      integer id
      real*8 vrel2,mtr,grnd_MT,tglob

      real*16 rtarg,rimp,dcomet16,ratq
      real*16 qcap,qstar,qstarcgs,rtargcm
      real*16 ratm,pdis,ratmlr,ratmf
      real*16 mtarg,mimp,mnew,rnew

c-----
c... Executable code

      write(*,*) 'Here #Q0.1',id
      write(*,*) 'Here #Q0.2',rimp
      write(*,*) 'Here #Q0.3',rtarg
      write(*,*) 'Here #Q0.4',vrel2
      write(*,*) 'Here #Q0.5',mtr
      write(*,*) 'Here #Q0.6',ratq
[/fortran]

The output when I run it looks like:

Here #C0.1           6
Here #C0.3 1.337014580143983178473131948054959E-0012
Here #C0.2 7.797292446069711940208707975960478E-0007
Here #C0.4 1.039694216870462E-002
Here #C0.5 1.185758329154166E-006
Here #C0.6 0.000000000000000000000000000000000E+0000
Here #Q0.1           6
Here #Q0.2 1.337014580143983178473131948054959E-0012
forrtl: severe (174): SIGSEGV, segmentation fault occurred
Image              PC                Routine            Line        Source
stryAJtmp          00000000004215A9 size_qstarrat_           4438 stryAJtmp.f
stryAJtmp          000000000041C432 col_disk_                1987 stryAJtmp.f
libiomp5.so        00002AC0E7BE4073 Unknown               Unknown Unknown

In other words, I get a seg fault when I try to access rtarg in the subroutine. This is rsmall16(i) in the main routine which prints out fine before the call (it is #c0.2). So, for some reason the subroutine is not getting the address correct address for rsmall16(i). Note that rsmall16(i) is part of a common block, but the subroutine does not know about that. I need to keep it that way because I use the subroutine other places when the input is not part of the common block.

I should note that I have several machines with different versions of ifort. The machines with version 11.1 20100414
DO NOT HAVE THIS PROBLEM, but my machines with version 12.1.3 20120212 and 13.0.0 20120731 do.

I am compiling it with:

ifort -mcmodel=medium -i-dynamic -heap-arrays -openmp -fpe0 -traceback -w -recursive -o stryAJtmp stryAJtmp.f -L/usr/local/lib/ -L/home/hal/lib/ -I/home/hal/SWIFT_P/ -L/home/hal/SWIFT_P/ -lswift -lfxdr -lMT_rand_OMP

and my libraires are also compiled with these flags. Also, I have KMP_STACKSIZE=64m and OMP_NUM_THREADS 1

I tried to reproduce the problem with a simple code:

[fortran]

      real*16 x(100)
      common/c_test/x

      do i=1,100
         x(i) = i
      enddo

!$OMP    PARALLEL DO SCHEDULE(DYNAMIC) DEFAULT(NONE)
!$OMP&      PRIVATE(i)
!$OMP&      SHARED(x)
      do i=1,100
         call sub(x(i))
      enddo
!$OMP END PARALLEL DO

      stop
      end

c---

      subroutine sub(x)
      real*16 x

      write(*,*) x

      return
      end
[/fortran]

but, unfortunately, this worked fine.

Thanks for any insight that you can give me.

hallevison · ‎12-12-2012

Hi: I have been able to make this particular instance of the problem go away if I replace the subroutine call with [fortran] rifort16 = rsmall16(i) ! needed to get around ifort problem call size_qstarrat(i,rifort16,rsmr16,vrelv, & mass(i),ratq) [/fortran] But that just moves the issue to the next place where I use rsmall16 as a input parameter for a subroutine. Interestingly, only rsmall16 seems to be a problem despite the fact that I use other arrays from common blocks in calls.