In some cases, it might happen that this method is called in cases
where no metrics is know for the current paragraph or where its
position is not set.
Take care of these cases to avoid assertions.
Remove setting of inset positions in the method, but make sure that
updateMetrics(false) is always called to get everything right.
In the new code, updateMetrics(bool) os the method that sets
everything right with minimal effort.
Using SinglePar does not majke sense here since the paragraph is not
modified and it might even not have a position yet.
This fixes a crash in BufferView::singleParUpdate, which is not
prepared (yet) to such situations.
Sometimes quick selection-scrolling could cause a crash because the
position of some paragraphs is not computed. To fix that, in
BufferView::showCursor, make sure that the metrics are always kept
clean using updateMetrics(false), which is lighweight.
As a consequence, the 'update' parameter of showCursor and
scrollDocView is not needed anymore. Its removal is mechanical and
accounts for most of this commit.
The only other significant change is that, when creating synthetic
mouse events and relying on scroll() for small moves, the full metrics
recomputation is replaced by the lighter version.
More work is still to come on this code, but this should be going in
the right direction.
This is one of the places where a dialog is shown (indicating that
some bibtems have been renamed) while the metrics are not up to date.
Then a draw operation can be triggered too early and a crash can ensue.
Use ad-hoc solution for this case and use Buffer::setBusy(). The
Alert::warning helper cannot really do that since it doe snot know the
current buffer.
tabBarClicked() should really only reset focus if a click in a different
TabBarGroup happened. And no need to issue workAreaChanged() again.
This has lead to extra and erroneous focus changes (see #12954)
\cprotect'ed insets do not need forced local font switch; in fact files
with this break in compilation. E.g. a caption with an URL containing
the char '#' and caption text set to a secondary language.
Add a Cursor-less version of Undo::recordUndoBufferParams (we don't
use an optional arument to avoir #includ'ing Cursor.h).
The version with CursorData parameter has been kept for the case of
local dictionary lfuns that have some kind of locality.
Remove Cursor argument to Undo::recordUndoFullBuffer().
The only difference from regular letter is the alignment of
"Send To Address" layout, but it no longer reflect the output
corrctly, probably because LyX revert the alignment of layouts
in RTL context now.
This inset can change dimensions when edited (like macros can).
Therefore Update::SinglePar should be set with editXY() like it is
already set for edit().
With this patch, 3 draw strategies (set in prefs with
\draw_strategy partial|backingstore|full) are available:
- "partial": only draw the parts of text that have changed since last
paint event (default for X11 and windows)
- "backingstore": the same, but drawing happens on an offspring
pixmap; this breaks subpixel rendering (default for Wayland and
macOS)
- "full": the screen is fully redrawn at each paint event and should
therefore always be correct; this is presumably slower but
introducing it will allow to test it.
This would deserve a proper UI eventually.
The code was expecting that all properties were specified
in the argument and not only some. Simply do not process
those properties not specified in the argument.
Reorder the tests so that, when the flag Update::SinglePar has been
specified, it does not take precedence over Update::Force flag.
This fixes the crash with inset-split.
This bug is pretty old, but it was not visible because the grey area
below the document would overwrite it.
It would be better to close the frame at the end of the document, but
this is for later.
The LFUN buffer-reset-export states to start afresh the export for the
next run, but in fact, the boolean was never reset, so that this was
done for all subsequent runs unless the buffer had been closed and
reopened.
Now the lfun does what it states and only acts on the next pass.
The IDs are used to link to specific elements (for instance, the User's guide nomenclature).
The beginning of the list (<ul> or <ol>) has the same magic ID as the first item of the list.
The effect is closer to what LyX offers. Previously, the percentage was of the HTML container (often, the whole page), meaning that images were oversized. Now, the scale is respected in the same way as LyX, but its bounding box is off. Overall, the result is better, but not what users expect.
The code is not ready for situations where some paragraphs that are
not visible have metrics available. This has been made visible by
a4d9315b, which removes a couple of full metrics updates: the screen
was not correctly scrolled after a scrollToCursor() call.
In PararagraphMetrics, some methods are added to be able to handle the
fact that paragraphs have or do not have a position.
In TextMetrics, a new method returns the first visible paragraph.
Finally, in BufferView::updateMetrics, the paragraphs' positions are
reset (in the case where everything is not cleared) and some care is
taken to skip the ones that are not relevant.
The assumption outside of this method is that all the paragraphs that
are in the TextMetrics are visible (we are talking about top-level
TextMetrics here). This could be changed (in order to avoid
recomputing paragraph metrics), but the cost is high in terms of
complexity and it is not clear that the gain in terms of performance
would be important.
When using a back store, it is obviously important to resize it
whenever the buffer view is resized. The new code path added in
527984ed skipped this part, so that the window was wrong when creating
a new window.
Commit 5577e877 introduces forceUpdateBuffer() to delay actual
updatBuffer() calls to a central place.
In Cursor::upDownInText, the updateNeeded flag (which triggers
Update::Force update flag) is set to true when Cursor::checkDepm
returns true. This is fine, except that checkDepm also leads to a
forceBufferUpdate() call.
The logic in the method was such that, when updateNeeded is already
true when entering, forceBufferUpdate() will be called even though
checkDepm did not request it.
Fixes a slight delay with selecting in MergedManual (the manual of
manuals).
The handling of Update::FitCursor traditionnally recomputes all
metrics once or twice. Now that updateMetrics(false) ensures that all
paragraphs that require it have a metrics, we might as well use that.
Take this occasion to move handling of Update::SinglePar earlier,
before the check for Update::ForceDraw.
The missing metrics computation meant that, when loading a file, the
workarea would not be scrolled correctly to show the cursor.
The resizeBufferView does not create an additional metrics
computation, it just triggers it earlier.
This is triggered by selecting part of an equation under Xcb platform
(x11). Indeed, the current selection has to be sent to the X11 server
and BufferView::requestSelection() is called. Eventually
cap::grabSelection() is the function that creates this partial
MathData object.
Fixes ticket #12947.
The code that handles this special macro insertion (in
Cursor::mathMacroClose())) is very weird: it inserts the contents of
the macro _after_ the macro and later, since the macro is 'greedy', it
will eat this contents and put it in its first macro cell.
Does it make sense to you? It does not ake sense to me either. Anyway,
this is the point where one should make sure that the buffer is set.
Fixes ticket #12682.
Make sure that math insets have a proper buffer. To this end, make the
Buffer* parameter of InsetMath mandatory and fix the compilation
errors that ensue.
In order to ensure that MathData objects have a valid buffer, the
default MathData() constructor is deleted. This means that a buffer
shall be specified for each MathData object created.
This is fairly mechanical, actually. In particular, in most
InsetMathXxx cases, in MathData and in MathParser, the available
buffer_ member is used.
More specific cases:
- lyxfind.cpp takes the buffer from the Cursor
- calls to vector<MathData>::resize take an additional
MathData(buffer_) parameter. There are cases where resize actually
remove cells, and in this case clear() or even erase() have been
used.
- in InsetMathMacroTemplate, the optional parameters of the
constructors cannot be allowed anymore (a default value cannot
depend on another parameter). Therefore there a now two constructors
instead.
- in MathAutoCorrect.cpp, the MathData objects are not bound to a
buffer, so that std::nullptr is used instead.
- in MathExtern, use a buffer when one is specified, std::nulptr
instead.
Set the buffer of contents that is added to a MathData object through
MathData::insert() (both versions)
MathData::push_back()
asArray()
Also in math macros, initialize look_ with the relevant buffer.
This reduces the number of insets hat do not have a proper buffer.
See #13050 for discussion of this issue.
Now that SingleParUpdate does not always lead to a full screen update
when the height of the paragraph changes (see new behavior of
updateMatrics(bool)), it is necessary to make sure that the grey area
below the main page is always repainted.
When a buffer is reloaded, its content may remain the same, but the
memory allocation is new, so that the inset pointers in cursors are
now wrong. This requires to sanitize the cursors held by the buffer
views.
Before the biginset branch, some full metrics computation call that is
now removed probably did that as a side effect. Now we have to be more
precise.
To this effect, introduce WorkAreaManager::sanitizeCursors() and use
it in Buffer::reload().
When the width of the window did not change, computing full metrics is
not necessary in BufferView::resize(), but it is better to redraw the
screen (especially with Wayland).
Typical use cases are when using M-x to open the minibuffer or going
in and out of an equation (which shows/hides the math toolbars).
See description here:
https://marc.info/?l=lyx-devel&m=171243435229412&w=2
cutSelectionHelper did not request a metrics update when the selection
was inner to a paragraph. The new code is better, but it was not
necessary before the biginset branch because of a full metrics
computation that hid this missing case.
Annotate function LFUN_FINISHED_xxx to indicate that they do not
require a full metrics computation.
Remove an "optimization" that meant that when the cursor changed
inset, a full metrics computation was requested.
Part of bug #12297
When entering/leaving a math hull inset, a Update::Force flag was set,
in case the metrics of the inset would change because of a switch
between normal and preview representation.
When entering the inset, this code is now used only when the inset was
in preview mode.
In both cases, Update::Force is replaced with Update::SinglePar.
This requites in Text::dispatch to honor Update::SinglePar when it has
been set by some lfun, even when singleparupdate is false.
Part of bug #12297.
Entering a math inset triggers a work area reize because the math
toobars appear automatically. However, by default these toolbars are
at the bottom of the screen and their presence does not change the
typesetting of paragraphs. Therefore it is useful to avoid a call to
updateMetrics() in the case where the width of the work area did not
change.
Part of bug #12297.
The number of metrics updates when loading file and showing it in a
new work area is unreasonable.
The first avoided call to updateMetrics() was an explicit resize in
BufferView::init(). Instead, an assertion is suppressed by exiting
early BufferView::processUpdateFlags() when BufferView::ready()
returns false. This is a new method introduced to factor in some
existing tests.
Two other metrics computations are avoided by setting the enclosing
View object busy() while creating the new tab. To make this work
properly, GuiWorkArea::scheduleRedraw has to return early in this
case.
When saving an unnamed document or invoking "Save as...", call
setBusy(false) earlier so that repainting occurs correctly.
Fixes bug #12976.
With the branch-test.lyx file from #12297, loading takes forever when
previews are activated. This is because each preview element causes a
full screen metrics recomputation.
This commit just skips these calls and only does one when all previews
have been obtained. As a result, computing the previews takes 1 second
instead of 25 seconds on branch-test.lyx.
Part of bug #12297.
The most visible part of this commit is the move of part of
BufferView::updateMetrics to a new TextMetrics::updateMetrics. This
new method makes sure that metrics are known for all visible paragraphs
(starting from anchor), and that the positions of the paragraphs have
been recorded.
This method is called up to 3 times in BufferView::updateMetrics:
* unconditionally, to update all visible metrics,
* then, if the bottom of the document is visible and too high, after
updating the anchor ypos,
* and similarly if the top of the document is visible and too low.
This fixes for example the case where one jumps to Section 5.3 at the
end of Tutorial and 'scroll_below_document' is false.
Some now redundant code is removed from BufferView::scrollToCursor.
The anchor-setting code in BufferView::draw is not clearly useful, but
left here just in case. It generates a debug warning, though.
Part of bug #12297.
Replace flag parameter for updateMetrics() by a `force' boolean. When
it is false, the method keeps the metrics of paragraphs that are still
visible in WorkArea instead of computing everything afresh. All it has
to do is update their positions.
Add code to updateMetrics() to update the value of the anchor pit/ypos
(similar to the one in draw()).
Update processUpdateFlags() to use this when update flag is ForceDraw.
Modify scrollDocView() to just change the anchor paragraph position
when the scrolling operation would re-use some of the existing
paragraphs.
The time needed to update the metrics when scrolling with mouse in the
branch-test.lyx document is now divided by 20!
Part of bug #12297.
This can make a big difference for a very large branch that contains
lots of equations.
This is complementary to the previous patch, since instead of reducing
the number of calls to updatePosCache, we make it faster.
In the same test of scrolling with mouse wheel through the
branch-test.lyx document, one finds a 23% improvement for
BufferView::updateMetrics().
Part of bug #12297.
The setting of insets positions was done twice in updateMetrics.
When one of the paragraph is a huge branch, this can be very expensive.
This leads to a 17% improvement on updateMetrics time on a scrolling test.
Part of bug #12297
The idea of single par update is to try to re-break only the paragraph
containing the cursor (if this paragraph contains insets etc.,
re-breaking will recursively descend).
The existing single paragraph update mechanism was tailored to work
only at top level. Indeed changing a paragraph nested into an inset may
lead to larger changes.
This commit tries a rather naive approach that seems to work well: we
need a full redraw if either
1/ the height has changed
or
2/ the width has changed and it was equal to the text metrics width;
the goal is to catch the case of a one-row inset that grows with
its contents, but optimize the case of typing in a short paragraph
part of a larger inset.
NOTE: if only the height has changed, then it should be
possible to update all metrics at minimal cost. However,
since this is risky, we do not try that right now.
Part of bug #12297.
if the undo element we want to add only changes stuff that was already
modified by the previous one on undo stack (in the same group), then
skip it. There is nothing to gain in adding it to the stack.
The typical use case is when doing a search and replace in a large
document that does many replacements in each paragraph. In this case,
the same paragraph would be stored repeatedly.
Fixes bug #12564.
