ENH: make subplot reuse gridspecs if they fit

[jklymak]

PR Summary

Right now, if we call

import matplotlib.pyplot as plt
fig = plt.figure()

ax = fig.add_subplot(3, 4, 1)
ax2 = fig.add_subplot(3, 2, (3, 5))
print(ax.get_subplotspec().get_gridspec())
print(ax2.get_subplotspec().get_gridspec())

plt.show()

The gridspecs will be different. Thats unfortunate, because there is no reason for these two subplots to not share the same gridspec, and if they do, then constrained_layout will work with them. Note this change will mean that many calls to pyplot.subplot and pyplot.subplot2grid will now work with constrained_layout=True as well.

This PR changes that so they two subplots use the same gridspec.

Note The more fine-grained subplot call must come first, or the grid specs will not "fit", so this is pretty simple minded. However, the PR doesn't make things worse, it just makes two gridspecs like it did before. i.e. if we switch the order of the calls we get two gridspecs:

import matplotlib.pyplot as plt
fig = plt.figure()

ax2 = fig.add_subplot(3, 2, (3, 5))
ax = fig.add_subplot(3, 4, 1)
print(ax.get_subplotspec().get_gridspec())
print(ax2.get_subplotspec().get_gridspec())

plt.show()

PR Checklist

• Has Pytest style unit tests
• Code is PEP 8 compliant
• New features are documented, with examples if plot related
• Documentation is sphinx and numpydoc compliant
• Added an entry to doc/users/next_whats_new/ if major new feature (follow instructions in README.rst there)
• Documented in doc/api/api_changes.rst if API changed in a backward-incompatible way

[jklymak]

I suppose this is a breaking change if someone was checking the results of ax.get_geometry() in their code and expected the geometry to stay the same:

ax1 = fig.add_subplot(1, 2, 1)
ax2 = fig.add_subplot(1, 1, 1)
print(ax2.get_geometry())

used to return (1, 1, 1), but now returns (1, 2, 1). Note thats get_geometry is pretty crappy function anyways, as it does not return the full geometry specification for an axes; i.e. the actual geometry of ax2 is (1, 2, (1, 2))

[anntzer]

Perhaps a better idea would be to just always create a new gridspec with shape (lcm(gs1.nrows, gs2.nrows), lcm(gs1.ncols, gs2.ncols)) and reparent all axes to this gridspec? This would avoid the order dependency.

[jklymak]

Yes, that could work. Wasn't 100% sure this whole PR was a good idea, rather than just telling people to not use add_subplot, but happy to come back to it...

[anntzer]

I'm not really sure it's worth it either (well I don't think the implementation would be that complex), just pointing out the possibility.

[jklymak]

Taking this up again in a simpler form in #17347