Improved handling of dependence analysis for arrays in OpenMP

[nmnobre]

I realise this might be hard to solve, but in cases such as:

do ji = 1, npti, 1
  do jk = 1, nlay_i, 1
    ztmelts = -rTmlt * sz_i_1d(ji,jk)
    ztmp(jk) = ztmelts / MIN(ztmelts, t_i_1d(ji,jk) - rt0)
  enddo
  zperm = MAX(0._wp, 3.e-08_wp * MINVAL(ztmp) ** 3)
end do

ztmp prevents parallelisation of the outer loop via a WaW dependence.

However, we can see that ztmp is overwritten exactly over the same range in each iteration of the outer loop and that the only use is after the write in the program's text. So, at most, only the last iteration's writes might matter outside the loop. Furthermore, if I force parallelisation by ignoring the dependence analysis results, ztmp isn't marked private and remains shared. I know I'm asking for trouble here, but do we think this might be supported one day?

[hiker]

I can have a look, but I am rather busy, will be a week or two probably before I can get to that. I also haven't touch the dependency handling code for a few months, so it will take me some time to get back into it.

[sergisiso]

This pattern is found in several more places in NEMO like zdfsh2.f90 and dynvor.f90::vor_een.

[sergisiso]

It needs:

• [ ] The transformation validate to let it continue
• [ ] The new infer_sharing_clauses to mark the arrays as private (currently all arrays are shared)