BUG: f2py incorrectly translates dimension declarations. by ianthomas23 · Pull Request #17654 · numpy/numpy

ianthomas23 · 2020-10-27T15:05:06Z

In fortran functions passed to f2py, calculations that occur in the dimensions of array declarations are interpreted as floats. Valid
fortran requires integers. Problem only occurs with fortran functions not subroutines as only for the former does f2py generate fortran wrapper code that fails to compile.

Relevant code is in numpy.f2py.crackfortran.getlincoef(), which calculates and returns the coefficients to getarrlen() for writing to the fortran wrapper code.

To convert floats to ints I have opted for try..except blocks which are common in crackfortran.py rather than explicit is_int() checks and conversion using int().

Fixes gh-8062.

In fortran functions passed to f2py, calculations that occur in the dimensions of array declarations are interpreted as floats. Valid fortran requires integers. Problem only occurs with fortran functions not subroutines as only for the former does f2py generate fortran wrapper code that fails to compile. Relevant code is in numpy.f2py.crackfortran.getlincoef(), which calculates and returns the coefficients to getarrlen() for writing to the fortran wrapper code. Fixes numpygh-8062.

melissawm · 2020-10-28T00:21:06Z

Hello, @ianthomas23 ! Thanks for this. Note that this only happens if the expression inside the dimension declaration is actually simplified by f2py (which is not the case for a non-linear expression like a(0:n*n/2)). Fix seems OK to me, I don't know if there's a more elegant solution here than the try/except blocks :)

I think the Shippable failure is unrelated - don't understand what happened there.

mattip · 2020-10-28T10:55:54Z

s390x failure not connected to this PR

mattip · 2020-10-28T10:56:07Z

Thanks @ianthomas23

eric-wieser · 2020-10-28T11:29:45Z

numpy/f2py/crackfortran.py

+                    except:
+                        pass
+                    try:
+                        b = int(b)
+                    except:
+                        pass


Bare except is an antipattern, as this means that f2py now not only swallows Ctrl+C but give different compilation results depending on whether it was pressed. This should be except ValueError or at least except Exception?

Would you mind making a follow-up PR?

Sorry, that was my oversight. I can do a follow up if @ianthomas23 prefers.

For reasons I don't understand, LGTM caught this, but decided to give a CI tick anyway.

@jhelie, I think we might have discussed this before - have we forgotten a config option, or is this working as intended?

@eric-wieser, @melissawm: My apologies, I will submit a follow-up PR to change except to except ValueError. I'll wait until gh-17662 is accepted to avoid having to rebase.

You shouldn't need to rebase on that PR, since the lines aren't next to each other - github will sort out the merge automatically I'd expect

Say, aren't both try / except clauses redundant here?
The original values of a and b are created by myeval(), which raises an exception if the return type is not a float or int:

numpy/numpy/f2py/crackfortran.py

Lines 2105 to 2109 in ab22e00

def myeval(e, g=None, l=None):

r = eval(e, g, l)

if type(r) in [type(0), type(0.0)]:

return r

raise ValueError('r=%r' % (r))

eric-wieser · 2020-10-28T22:35:12Z

numpy/f2py/tests/test_crackfortran.py

+    code = """
+        function test(n, a)
+          integer, intent(in) :: n
+          real(8), intent(out) :: a(0:2*n/2)


What shape should a(0:n/2*2) give when n is odd?

Good question. Fortran uses integer maths is array declarations, so that shape here, in Python syntax, is (1+2*(n//2),)

I have submitted a new PR (gh-17687) to fix the original problem and it includes a test of this example.

What does -(-n)/2*2 give in fortran? (edited) The python integer division is not the same as the integer division used in most languages.

Corrections to PR numpygh-17654 following post-merge review. Fix now only casts floats to integers if they are definitely integers. Also, division that occurs in fortran array declarations is evaluated in an integer context, so f2py.crackfortran.myeval does the same, converting single to double slashes. Fixed numpygh-8062.

	def myeval(e, g=None, l=None):
	r = eval(e, g, l)
	if type(r) in [type(0), type(0.0)]:
	return r
	raise ValueError('r=%r' % (r))

Uh oh!

Conversation

ianthomas23 commented Oct 27, 2020

Uh oh!

melissawm commented Oct 28, 2020

Uh oh!

mattip commented Oct 28, 2020

Uh oh!

mattip commented Oct 28, 2020

Uh oh!

eric-wieser Oct 28, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

eric-wieser Nov 1, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

eric-wieser Oct 28, 2020 •

edited

Loading

eric-wieser Nov 1, 2020 •

edited

Loading