[NumPy-Tickets] [NumPy] #2047: f2py needs to be updated to use non-deprecated macros/functions

[NumPy Trac]

#2047: f2py needs to be updated to use non-deprecated macros/functions
------------------------+---------------------------------------------------
 Reporter:  charris     |       Owner:  pearu         
     Type:  task        |      Status:  needs_decision
 Priority:  normal      |   Milestone:  1.7.0         
Component:  numpy.f2py  |     Version:  devel         
 Keywords:              |  
------------------------+---------------------------------------------------
Changes (by pearu):

  * status:  new => needs_decision


Comment:

 There is one gotcha with fixing this issue.

 Consider the following example:
 {{{
   !file foo.f90
   subroutine foo (a)
     integer, dimension (2) :: a
     !f2py intent (inplace) a
     a(1) = 7
   end subroutine foo
 }}}
 and compile it with f2py: {{{f2py -c -m foo foo.f90}}}.
 The gotcha is illustrated with the following test:
 {{{
 $ python -c 'import foo, numpy; a=numpy.array([1,2], dtype=numpy.int16);
 print id(a),a,a.dtype;foo.foo(a); print id(a),a,a.dtype'
 45223856 [1 2] int16
 45223856 [7 2] int32
 }}}
 Notice that initially 16-bit numpy array is turned to 32-bit array *in
 place* while the id's of arrays are *not* changed. This non-pythonic
 "feature" is achieved by using swap_arrays function (in fortranobject.c).

 Regarding non-pythonic: while this is generally bad idea for Python
 programmers, Fortran programmers use this non-pythonic feature all the
 time, so swap_arrays is a compromise made many years ago when f2py was
 standalone.

 Since the aim of this issue is to avoid using swap_arrays, I am now stuck
 with what to do as numpy API does not provide equivalent method to the
 swap_arrays function.

 We have multiple choices:
   1) introduce swap_arrays function to numpy. Then this issue cannot be
 fixed for 1.7. And, perhaps more importantly, a non-pythonic function is
 introduced to numpy that probably needs a wider discussion.

   2) remove swap_arrays from numpy. Again, this is not a 1.7 issue as it
 will break numpy 1.6 based code that uses the swap_arrays feature.

   3) keep swap_arrays as it is and try to reimplement it using non-
 deprecated numpy macros/functions.

-- 
Ticket URL: <http://projects.scipy.org/numpy/ticket/2047#comment:7>
NumPy <http://projects.scipy.org/numpy>
My example project