This is a mega-commit - because most of it had to be done in one go
otherwise some commits would fail to compile - that attempts to fix a
few problems with Geany's includes as well as various other related
cleanups. After this change it's easier to use includes and there's
little worry about which order things are included in or who includes
what.
Overview of changes:
* Include config.h at the start of each source file if HAVE_CONFIG_H
is defined (and never in headers).
* Go through each source file and make the includes section generally
like this:
- Always config.h first as above
- Then if the file has a header with the same name, include that
- Then include in alphabetical order each other internal/geany header.
- Then include standard headers
- Then include non-standard system headers
- Then include GLib/GTK+ related stuff
* Doing as above makes it easier to find implicit header include
dependencies and it exposed quite a few weird problems with includes
or forward declarations, fix those.
* Make geany.h contain not much besides some defines.
- Add a little header file "app.h" for GeanyApp and move it there
- Move "app" global to new "app.h" file
- Move "ignore_callback" global to "callbacks.h"
- Move "geany_object" global to "geanyobject.h"
* Add an include in "geany.h" for "app.h" since GeanyApp used to be
defined there and some plugins included this header to access
GeanyApp.
* Include "gtkcompat.h" everywhere instead of gtk/gtk.h so that
everywhere sees the same definitions (not a problem in practice AFAIK
so this could be changed back if better that way.
* Remove forward declarations from previous commits as some people
apparently consider this bad style, despite that it reduces inter-
header dependencies.
TODO:
* As always, to test on win32
* As always, to test with not Autotools
* Test plugins better, both builtin and geany-plugins, likely API/ABI bump
* Test with various defines/flags that may change what is included
* win32.[ch] not really touched since I couldn't test
* Always define GEANY_FOO_H to 1 in the header guards
* Always put a G_BEGIN_DECLS/G_END_DECLS guard in every header for
consistency, even private ones where it doesn't matter.
* Always include either <glib.h>, <gtk/gtk.h> or some other header
that will provide G_BEGIN_DECLS before using it. In a lot of headers
that use glib.h and gtk/gtk.h stuff anyway, this resolves an implicit
dependency they had on them being included before that header.
* Always put a comment at the #endif part of the guard so it's
easier to see what it applies to.
* Always use an underscore between the header guard identifier's words
even though the filename doesn't have one.
Old implementation was not really fitting the updating needs and had a
bug making symbols disappear if they haven't changed but their parent
did (e.g. when a C++ constructor's signature changed).
New implementation does:
1) walk old tree, updating or removing rows;
2) add remaining tags.
It walks less than (new_tags + old_tags + new_tags) in the worst case,
thanks to some hash table-based caching; and also gets rid of the
"valid" column in the symbols tree, saving a few bytes in memory.
Finally, there is a ~7% performance gain (from 21 to 18ms) upon common
tree updates, sometimes more.
The main advantages of not clearing and rebuilding the whole list is
that it doesn't loose the folding and selection (as far as the selected
row(s) still exist after the update, of course), and it reduces
flickering upon update.
The current implementation works in 3-steps:
1) mark all rows as invalid;
2) insert/update the rows, according to the new tag list, marking them
as valid;
3) remove all rows that are still invalid.
This walks (rows) the first time, (tags*rows) the second and (rows) the
third. This also uses an extra column to store the row's validity.
A (probably) better implementation would be to:
1) walk the current rows, updating them if necessary, or removing them;
2) add the remaining tags that weren't there before.
This is probably faster in theory (and probably also in practice), but
it needs to refactor a lot the code to easily update *or* create a row,
what the current code does not provide.
Basically this is would be a two-pass update, walking (rows*tags) in
the first pass, and only the remaining non-added tags in the second.
git-svn-id: https://geany.svn.sourceforge.net/svnroot/geany/trunk@5562 ea778897-0a13-0410-b9d1-a72fbfd435f5
