OPENMP : OpenMP threads don't respect synchronization

Question

I parallelize a Fortran code with OpenMP. I'm still debugging (there may be some remaining bugs but sequential version is OK) and the behaviour of the OpenMP threads is strange, there are threads that do not wait the others at OMP BARRIER or OMP END DO statements.

For example, in one (simplified) subroutine one has :

SUBROUTINE calcT (T)

USE mod_param_simul
USE mod_transport

REAL(8), INTENT(INOUT), DIMENSION(0:nr+1,0:nth+1) :: T

INTEGER :: i, j
REAL(8), DIMENSION(:,:), SAVE, ALLOCATABLE :: Work

!$OMP SINGLE
ALLOCATE (Work(nr,nth) )
!$OMP END SINGLE

!$OMP DO COLLAPSE(2)
DO j = 1, nth
   DO i = 1, nr
      Work(i,j) = aTdiffW(i,j)*(T(i-1,j  )-T(i,j) ) + aTdiffE(i,j)*(T(i+1,j  )-T(i,j) ) &
                + aTdiffN(i,j)*(T(i  ,j+1)-T(i,j) ) + aTdiffS(i,j)*(T(i  ,j-1)-T(i,j) ) &
                + bsrc_Tdiff(i,j)
   END DO
END DO
!$OMP END DO

!$OMP DO COLLAPSE(2)
DO j = 1, nth
   DO i = 1, nr
      T(i,j) = T(i,j) + Work(i,j)
   END DO
END DO
!$OMP END DO

!$OMP SINGLE
DEALLOCATE (Work)
!$OMP END SINGLE

RETURN
END SUBROUTINE calcT

I already have checked a lot of things :

• OpenMP visibility of all variables in the application
• SAVE attribute for variables, especially arrays, defined in Fortran modules
• bounds of arrays, of loop indexes
• dynamic allocation of arrays
• use valgrind -- memcheck to monitor the code ; nothing relevant found
• use recent releases of several Fortran compilers (Intel ifort, GNU gfortran, PGI pgf90) with full debugging options. I get crash of the code in the second loop nest (write/flush statements inserted before/after), updating array T with error messages like

Program received signal SIGSEGV: Segmentation fault - invalid memory reference.

or

forrtl: severe (408): fort: (8): Attempt to fetch from allocatable variable WORK when it is not allocated

Edit : I add more informations

Intel ifort  2021.3.0 20210609 ; options : -fpp  -stand f08 -e08 -free -g -O0 -init=huge -init=snan -init=arrays -ftrapuv -debug all -check all -warn general -implicitnone -no-ftz -fpe0 -fp-model precise     -traceback -fno-alias -fno-fnalias -O0 -qopenmp

Runtime error :

forrtl: severe (408): fort: (8): Attempt to fetch from allocatable variable WORK when it is not allocated

Image              PC                Routine            Line        Source             
a.out       000000000074806F  Unknown               Unknown  Unknown
a.out       000000000043EC20  mod_energy_mp_calct         529  energy.f08

and line 529 is just before line 531 which is T(i,j) = T(i,j) + Work(i,j)

GNU Fortran (GCC) 10.3.0 options : -x f95-cpp-input -std=f2008 -pedantic -ffree-form -g -fbacktrace -ffree-line-length-none -fcheck=all -finit-real=snan -ffpe-trap=invalid,zero,overflow -fimplicit    -none -Wimplicit-interface -Wconversion -Wunderflow -Wall -Wextra -fbounds-check -O0 -fopenmp

Runtime error :

Program received signal SIGSEGV: Segmentation fault - invalid memory reference.

Backtrace for this error:
#0  0x7f0741fb53ff in ???
#1  0x40e667 in __mod_energy_MOD_calct
    at /home/........../energy.f08:531
#2  0x46dae7 in MAIN__._omp_fn.0
    at /home/........../main.f08:290
#3  0x7f07429da091 in GOMP_parallel
    at ../../../gcc-10.3.0/libgomp/parallel.c:171
#4  0x4177e6 in MAIN__
    at /home/........../main.f08:81
#5  0x46d653 in main
    at /home/........../main.f08:16

Moreover, Nr and Nth are declared as INTEGER with SAVE attribute in a module that is used in this routine. The dimensions of T, (0:Nr+1,0:Nth+1), are for the inner domain and the boundary cells. T is updated in its inner domain thanks to array Work that is computed with all values of T, boundary conditions included.

The result is always the same. It can occur in this routine or in others with very similar constructions. I do not see what is wrong with respect to the OpenMP norm.

So, I'd be grateful if someone can give me some advices because I ran out of idea.

Regards,