If you haven't read about the compile options for both ifort and CVF please do so.  With current 32-bit ifort, if you wish to evaluate simple expressions in double precision, as CVF did, /arch:ia32 should work.  It would be preferable to write what you mean in standard Fortran, e.g.
P=MIN(500.0d0,dble(P)+P0)-P0
(if you don't like Fortran 90), so that your result doesn't depend on the compile options.
You have used one bit of f77, so I won't try to show you an f66 version.
As mentioned further down, these compilers are evaluating your expressions at compile time, so run time settings will have no effect, and the current compiler is technically correct in doing what you asked rather than what you may have meant.