Fix segfault in ft2font

[cgohlke]

On Windows, Python 2.7, 32 and 64 bit, the sankey_demo_rankine.py example segfaults in the ft2font.pyd extension at https://github.com/matplotlib/matplotlib/blob/master/src/ft2font_wrapper.cpp#L667. Apparently on that platform std::vector of size 0 doesn't like to be indexed.
Please review: I did not check if not calling CALL_CPP("set_text", self->x->set_text(size, &codepoints[0], angle, flags, xys)) for size 0 has any side effects.

[tacaswell]

Another reason to release more often is we will get @cgohlke to look at our code more often....

[tacaswell]

and we really need to get appveyor setup.

[jkseppan]

Calling set_text with an empty text has the side effect of setting the bounding box to empty. One way font objects are used seems to be that you call font.set_text(...) and then font.get_width_height(), and if the objects get used many times, I guess the object could remember the bounding box of the previous call and give a nonempty size for an empty string.

[jkseppan]

So the problem is that codepoints[0] is illegal for an empty vector? If we can guarantee a C++11 compiler (probably not?) you could use codepoints.data(). Otherwise, I guess something like this would work:

uint32_t* codepoints_array = NULL;
if (size > 0) {
    codepoints_array = &codepoints[0];
}
CALL_CPP(... codepoints_array, ...));

[mdboom]

I don't know how much we can rely on C++11 -- we'd have to test it. I think gcc is fine, but clang and MSVC may be playing catch up still (not sure).

Anyway, your workaround seems like a good suggestion. I suspect what's happening here is that no buffer is allocated at all in the zero-length case, so &codepoints[0] is garbage.

[jkseppan]

Looks good to me, and passes tests. Did you check that this still fixes the segfault on Windows?

[cgohlke]

Yes, I tested this on Windows.