Fix #602

[mscroggs]

After this PR:

• If real or conj is called on a real number, nothing is done
• If imag is called on a real number, 0 is returned

This is a slightly hacky fix for #602.

Happy for this to be closed if a better fix exists

edit: this feels even hackier now that it's been edited to make mypy pass

Close #602. Close https://github.com/FEniCS/dolfinx/issues/2757

[garth-wells]

What is real is the generated code supposed to do? Is it an artefact, or should it be creal?

[chrisrichardson]

It should not be creal as the input is double not _Complex double. It should not be anything. But ufl is asking for the real part of a number that is already real.

[mscroggs]

What is real is the generated code supposed to do? Is it an artefact, or should it be creal?

real is replaced with creal for double complex and crealf for float complex. It's currently not replaced for real types. This PR makes a replacement happen for these types

[chrisrichardson]

Maybe use "" instead of None, then it will just use parentheses, which have no effect. Also a bit hacky.

[mscroggs]

Maybe use "" instead of None, then it will just use parentheses, which have no effect. Also a bit hacky.

mypy also didn't like None. I went with "SPECIAL_CASE ()" and "SPECIAL_CASE 0 (as the latter will also allow dealing with imag(real type) = 0. But open to better suggestions

[chrisrichardson]

How about picking it up a bit sooner (before consulting the math table):

if c.args[0].dtype == L.DataType.REAL:
    if c.function in ("real", "conj"):
        return "(" + self.c_format(c.args[0]) + ")"
    elif c.function == "imag":
        return "0"

[chrisrichardson]

Or maybe we should handle it in lnodes.py before it gets here...

def _math_function(op, *args):
    name = op._ufl_handler_name_
    dtype = args[0].dtype
    if name in ("conj", "real") and dtype == DataType.REAL:
        assert len(args) == 1
        return args[0]
    if name == "imag" and dtype == DataType.REAL:
        assert len(args) == 1
        return LiteralFloat(0.0)
    return MathFunction(op._ufl_handler_name_, args)

[chrisrichardson]

I think this is the least "hacky", let's hope it passes all the tests...