In this version, the idea is to record undo at the place where the document is modified, which is definitely cleaner.
1/ in Buffer::updateBuffer, add a recordUndo, with the caveat that a
const_cast has to be used
2/ in GuiApplication::dispatch, add an extra undo group that
encompasses the updateBuffer call. Some other undo groups may be
redundant now, but it is not a problem since they do not cost
anything.
http://marc.info/?l=lyx-devel&m=133876924408431&w=2
The problem here is that the copy_params() routine in FindAndReplace.cpp
created a new DocumentClass, but it never updated its Buffer to reflect
that new DocumentClass. So its Paragraphs still contained points to the
Layouts in the old DocumentClass which, since ead697d4b6, gets garbage
collected once it is no longer needed. So the Layout doesn't exist, and
we crash.
For some unknown reason, the paste code likes to call Cursor::setSelection(), which is not appropriate here since we are not trying to set a selection at this point.
objects. The problem that led to the leak is that these objects can be held in
memory long after the Buffer that created them is gone, mostly due to their
use in the CutStack. So they were previously held in a storage facility, the
DocumentClassBundle. Unfortunately, they were now being created too often,
especially by cloning. It's not really a leak, because they're accessible, but
we weren't ever destroying them.
This new approach uses a shared_ptr instead.
Thanks to Vincent for pointing out const_pointer_cast.
While cppcheck did not turn out any suspicious error messages, using
the "performance" flag highlighted several nitpicks in three categories
* do not use it++ for iterators, ++it is better
* do not use size() to test for emptyness, empty() is here
* do not use "const T" as a function parameter, "const & T" is better
I doubt that any of these is a real performance problem, but the code is cleaner anyway.
When there is a single document open, the tabbar gets hidden. The only and selected tab is still visible though as a break in the line delining the tabwidget. To fix this we paint the tabwidget ourselves and remove the presence of the hidden tab.
When the socket does not exist anymore, we should not try to access it. In
trying to do so, std::map will create a new shared_ptr but this pointer
doesn't point at anything. To prevent this, we explicitly check whether
the socket is available.
This reverts the previous fix in [0a33374c/lyxgit] and fixes it
differently.
We always have to call 'notifyCursorLeaves', but we only have to make sure
that we call the 'fixIfBroken()' functions first.
We rely on the 'or' operator to prevent us from calling
'notifyCursorLeaves' if one of the two cursors is broken. This doesn't
work when using the '|' operator. The compiler 'optimizes' the code in
such a way that we always call notifyCursorLeaves anyway. Using the '||'
operator fixes this.
The undo range is extended to encompass all the paragraph after the end that
are a non-zero depth. The reason is that these paragraphs may see their depth
reduced if the last paragraph sees its depth reduced.
Note that there is a memory cost, since we store paragraphs than may not be
modified in practice. This may matter for some pathological files (people
putting all their document at depth 1 for a weird reason?)
I tried to be careful, but this can introduce new bugs. Please test.
Only user of this function is code that has been commented out, comment
out this function as well.
Signed-off-by: Lars Gullik Bjønnes <larsbj@gullik.org>
The check whether fname != oldname does not work for unnamed buffers.
Therefore, we check explicitly whether the Buffer that already exists with
the same name isn't really the buffer we are trying to save.
Text::outerFont looks recursively for paragraphs before the current one
which has a lower depth. If such a paragraph cannot be found, depthHook
and outerHook return the current paragraph. As such, we end up in an
infinite loop. So, if we find a par_depth that was the same as the
previous one, we apparently can't find a suitable paragraph and we should
quit the loop.
In this case, the working dir has already a slash at the end.
Appending another one would result in a path ending with a double slash,
which has a special meaning when used in TEXINPUTS (all subdirs would
be recursively scanned). So, avoid doing that.
We want the key as id, not the label (which is optional).
We also need a kind of namespace for the citation ids.
We should also clean the id tag before using it.
