Only force gnureadline in setup.py with bdist_wheel

[xolox]

The asymmetry (inconsistency) of unconditionally adding gnureadline as a dependency of binary (but not source) distributions in setup.py can cause issues like reported at paylogic/pip-accel/issues/34. Would you be willing to accept this pull request to remove the asymmetry? For reference, it seems that this was introduced in ba0cd52 with the following comment:

adjust dependency check for gnureadline on OS X

• always depend when making a bdist (for consistency)
• otherwise only depend if missing (e.g. pip install -e)

I guess it's kind of ironic that the original reasoning was consistency and now I'm calling this inconsistent :-). I'm interested to hear whether you agree about the asymmetry though. If you need more context about why I want to fix the asymmetry please refer to paylogic/pip-accel/issues/34, I'm posting a detailed explanation there.

[minrk]

The reason for bdist is that making a wheel on a machine that has readline should not exclude gnureadline from the dependency list, because it's still a dependency on the machines on which the wheel (or egg, etc.) will be installed.

[minrk]

This PR makes the behavior of creating a bdist less consistent because the resulting bdist is different depending on the computer building the bdist. In master, bdists always have the same dependencies, regardless of the compiling host. There is a fundamental difference between installing from sdists and bdists because when installing from sdist, the environment can be checked at the destination. With bdists, dependencies are computed without sufficient knowledge of the destination, so the least common denominator choice is taken.

Since IPython actually only cares about wheel bdists, and pip-accel uses bdist_dumb for some reason, if you change the check from startswith('bdist') to == 'bdist_wheel', it should serve both our purposes.

[xolox]

@minrk: Thanks for your initial reply and follow up! Last weekend I came to the same conclusion: The goal of ba0cd52 seemed to be to make sure that binary wheels would be generated including the proper dependencies (which makes sense of course). I'll change the pull request.

[minrk]

At the time, it was true of eggs as well, and it remains the right behavior for all bdists intended for distribution. But the only official bdists of IPython from now on are wheels, so the reality is that most bdists other than our own wheels are meant for private re-use (e.g. pip-accel), where the least-common-denominator behavior is not helpful.

[xolox]

@minrk: Thanks for the feedback and for the quick merge!