There were several problems since the conversion to gtk3
* widget wasn't displayed at all - apparently this was because I had overriden the draw function of the gtklayout (sheet)
-> fixed by adding a draw callback instead
* several gtk warnings about allocating sizes to widgets without using *get_preferred_width/height
-> fixed by adding the itemedit dropdowns to the layout only when they are requested
* the arrow button didn't want to resize small enough
-> fixed by adding a custom style to drops the button's internal padding
* there were several sizing issues which ironically got fixed dropping much of the custom sizing code
gtk_statusbar_set_has_resize_grip
gdk_drawable_get_screen
gdk_device_get_core_pointer
There is no replacement for it in gtk3 and it's probably
not needed any more either.
The old code directly accessed a member variable of GtkEntry. This
direct access is no longer possible, but there is no public function
to get/set this variable. There is a private one used by
GtkComboBox internally) but it's not exported so it can't be used either.
As an interim solution I peeked at how hesiod handled it in his earlier
(unmerged) gtk3 branch. He replaced this private access with setting
gtk_widget_set_visible, something I wouldn't have considered in this
situation. The issue is in the register2 code, so for now it's not
too important. We can investigate this more when the register rewrite
is picked up again.
Only gnc-dense-cal.[ch] is left untouched because it's use
of color is tightly coupled to the use of GtkStyle. This
latter needs conversion as well, but would be too much to
add in one commit.
Additionally the register code uses GtkStyle to extract
colors. This has been partially converted to using GdkRGBA
with the inconvenient side effect that depending on how
a color is retrieved it has to be freed with gdk_rgba_free
or not. This may be cleaned up later.
The biggest problem was that a pointer to an out of scope struct tm was
passed to the wrapped function. With opt level 2, clang doesn't bother
setting the contents of the struct since it goes out of scope without being
used. This caused the transaction report to never report anything since
the start and end times it got were ridiculous.
Also most functions that take a struct tm pointer can change the contents
of the struct (if only to normalize it) so pass the new values back to Scheme.
Finally all calls to gnc_localtime and gnc_gmtime from Scheme leaked a struct tm
and calling gnc_tm_free from Scheme is a really bad idea so don't wrap it.
Mike Alexander brought this up with a test case that failed to round down
sufficiently; he found that reducing the rounding denominator by 2 sufficed
to make his test case pass.
In fact the sizing of the replacement denominator by shifting the larger of
the numerator or denominator by an arbitrary 62 bits was not correct most
of the time, so instead we begin with a shift of the full lower leg worth,
try to do the conversion, and if the conversion is still “big” shift the
larger value one more and try the operation again, repeating until the
result will fit in a GncNumeric.
I noticed improper screen redraws when switching cells while the
window was scrolled down. The drawing logic is pretty conflated
so instead of hunting this down, I have made it so that the full
visible part of the windows gets redrawn each time the user
switches to a different cell.
This is slightly more expensive, but should still be ok.
