This causes some kind of mainloop breakage, if you call plt.subplot_tool() as the first thing before plt.show(). Only one "Figure 1" appears, not the configuration tool, and closing it deadlocks the prompt.

Should be fixed now -- basically, because the toolfig is not pyplot-managed, I needed to call show() on it explicitly.

I guess it's still a bit weird, because it starts the mainloop now, which shows the figure immediately instead of on plt.show() (if you haven't called plt.ion()).

Actually, it doesn't quite act like plt.ion(), as regular figures do not open automatically after, but also plt.show() blocks until the subplot tool is closed, even though it opens without it.

General comment/opinion:

I'm not fond of trying to fake dialogs using figures. The main motivation is that we do not need to maintain dialogs for each backend.

But:

• Figures are not dialogs, which causes awkward intercations (see comment above)
• IMHO the look&feel of the matplotlib widgets is poor.

I'd prefer native dialogs and for that:

• Define a narrowly limited scope of what we want to make configurable.
• Implement this with native widgets for each backend.

This is code that will rarely change and once we have done the one-time effort of implementing it, it's not significantly more maintainance burden than the mpl widgets approach.

@QuLogic Can you clarify what is the failing case? TBH I always have a bit of trouble with following ion()/ioff()...

@timhoffm Just to be clear, I have a strong preference for native widgets too (after all I wrote #8683 :-)). However right now the matplotlib figure "is what we have", and this PR doesn't make things worse in that respect.

@anntzer Correct, that's why I wrote "General comment".

Concerning this PR, I cannot easily review the change because I don't follow the details of figure creation. Would need to dig into it a bit further.

Well, it's not really ion mode, just that it sort of half-emulates it.

Previously, if you ran plt.subplot_tool(); plt.show(), then the tool would implicitly create a figure to work on, and both would show up when plt.show() was called. plt.show() would not exit until both were closed.

With this PR, it's 'half-interactive', i.e., the tool shows itself immediately, but the figure only shows on plt.show(). Then inconsistently, the plt.show() does not exit until both are closed.

Granted, it is a bit of an edge case to call subplot_tool first with nothing open, so maybe this is a tradeoff we'd be willing to make.

I'm fairly tempted to claim that indeed opening a subplot_tool without having a figure first is simply not something we really need to support. (I don't think there's a simple way to register somewhere that a non-pyplot window needs to be show()ed when pyplot.show() is called.)

I have a slight preference for using the zero time timeout to defer calling show until after the event loop gets a chance to spin. In addition to just being a funny edge case, it makes it more likely that someone could do

def my_function():
    plt.configure_subplots()
    time.sleep(32) # or actual work

and be left with a "frozen" window.

However, if there is consensus that this a weird edge case I'm not going to block merging over this.

Pushing to 3.4 as I don't think this is urgent.

This may be nice to get in at least before the big MEP22 work.