problem with optimisation flags

[Olivier_M_]

Hi,
I have a code which behaves incorrectly if anything above -O0 is
specified.
More specifically, the piece of code that looks like being responsible is something like this:

do i=1,nx
do j=1,ny
flux(i,j) = "calculate x derivative of matrix field at (i,j)"
enddo
enddo

do i=1,nx
do j=1,ny
flux(i,j) = flux(i,j) +
"calculate y derivative of matrix field at (i,j)"
enddo
enddo

The problem occuring, after testing on suitable matrix fields, is that the two blocks of code are being swapped around at execution time, which should not happen, and does not yield the right answer. (By the way, I know calculating both derivatives in the same loop would avoid this problem, but I am not free to modify the code as I wish).
Is this expected behaviour on the compiler's behalf?
If so, is there a compilation flag which can allow me to apply optimisation without code "rearrangement"?

[TimP]

Intel compilers should be attempting loop interchange and fusion only at -O3. According to what you have presented, both would be needed, to make the code at all efficient.
I would not be surprised, in numerical differentiation, to find that the situation is more complicated than we can see from what you have presented. If you are using a current Intel compiler, could you file an example of the problem on your premier.intel.com account?