Investigation of bug #9236 showed that crash to be due to a Paragraph's
holding a dangling pointer to an old and deleted Layout after the
DocumentClass was reset. Since the backtraces look almost identical, it
seems likely that we have the same problem here.
Since this crash seems almost always to involve tables, I looked at the
code in switchBetweenClasses() and found that the Paragraphs that belong
to "hidden" table cells are not seen by the initial recursion using a
ParIterator: It skips right over them. This was confirmed by test code
suggested by Enrico, with results reported in Trac.
The present patch attempts to deal with this problem in the second
recursion, over Insets. When we see an InsetTabular, we call a new
routine that recurses through the cells, looking for hidden ones. If it
finds a hidden one, it then resets the Layout for the cell's Paragraphs
(there should be only one, but we do not make any assumptions) to the
PlainLayout that belongs to the new DocumentClass. This is good enough,
since such cells never have content.
There is extensive discussion of the patch here:
https://www.mail-archive.com/lyx-devel@lists.lyx.org/msg185095.html
Additional testing by Enrico and me confirmed the existence of the
dangling pointer.
This branch implements string-wise metrics computation. The goal is to
have both good metrics computation (and font with proper kerning and
ligatures) and better performance than what we have with
force_paint_single_char. Moreover there has been some code
factorization in TextMetrics, where the same row-breaking algorithm
was basically implemented 3 times.
Globally, the new code is a bit shorter than the existing one, and it
is much cleaner. There is still a lot of potential for code removal,
especially in the RowPainter, which should be rewritten to use the new
Row information.
The bugs fixed and caused by this branch are tracked at ticket #9003:
http://www.lyx.org/trac/ticket/9003
What is done:
* Make TextMetrics methods operate on Row objects: breakRow and
setRowHeight instead of rowBreakPoint and rowHeight.
* Change breakRow operation to operate at strings level to compute
metrics The list of elements is stored in the row object in visual
ordering, not logical. This will eventually allow to get rid of the
Bidi class.
* rename getColumnNearX to getPosNearX (and change code accordingly).
It does not make sense to return a position relative to the start of
row, since nobody needs this.
* Re-implement cursorX and getPosNearX using row elements.
* Get rid of lyxrc.force_paint_single_char. This was a workaround that
is not necessary anymore.
* Implement proper string metrics computation (with cache). Remove
useless workarounds which disable kerning and ligatures.
* Draw also RtL text string-wise. This speeds-up drawing.
* Do not cut strings at selection boundary in RowPainter. This avoids
ligature/kerning breaking in latin text, and bad rendering problems
in Arabic.
* Remove homebrew Arabic and Hebrew support from Encoding.cpp. We now
rely on Qt to do handle complex scripts.
* Get rid of LyXRC::rtl_support, which does not have a real use case.
* Fix display of [] and {} delimiters in Arabic scripts.
Don't create an intermediate copy (found by Jean-Marc).
I doubt that this has anything to do with the mystery crash, but it works, and
following the standard patterns is better anyway.
FileName::tempName() created a new temp file name by using qt to create a
temporary file with a unique name, and then deleting that file and returning
the name. This was unsafe, since other processes or even other threads of the
running LyX could create files with the same name between deletion and then
using the temp name.
This is fixed by using the TempFile class instead. As a side effect, a few
cases where the temp files were not deleted after usage were fixed as well.
The only place that is still unsafe is createTmpDir().
The firs tinvolves a thinko in BibTeXInfo::expandFormat. We were previously
counting passes through this routine, which means: one for every character,
more or less. So long strings would hit the "recursion limit". But what
we are worried about is an infinite loop caused by misues of macros, so that
is what we need to count.
This prevents the error we were previously getting, but it reveals a huge
slowdown when one tries to open a citation inset with a large nubmer of keys.
So we also limit the number of keys we try to process, and the length of the
string we try to display, when we are generating citation information.
I'm convinced that there is a deeper problem in how citation information is
generated (see the bug tracker for more info), but that will require major
surgery and a file format change
This method is a proxy for LyXRC::preview that forces to forward
declare some wrapper around an enum...
Instead, two simple static methods previewMath() and previewText() are
introduced, that make the code much easier to follow.
If a separator inset is used after an environment with NextNoIndent
set to false, a blank line was already output. So, avoid outputting
another blank line or an odd looking line with only a '%' character
(as the previous blank line provides already a visual separation).
Multirow cells now have the same alignment rendering in LyX
as in the output. The alignment of a multirow can change as
long as the column is not of fixed width. If the column is
of fixed width, the multirow is left-aligned.
When setting a multirow, the alignment is copied from
the last cell in the selection.
This was not necessary when LyX was generously outputting newlines.
As it may happen that the output produced by writeExternal (the
result of an external inset) starts right at the beginning of a
line, if otexstream does not know that something is already on
the line, the iomanip-like variable 'breakln' fails to actually
break the line.
The algorithm used for breaking a paragraph in LaTeX export is changed
for avoiding spurious blank lines causing too much vertical space.
This change is tied to the introduction of a new inset (with two
different specializations) helping in either outputing LaTeX paragraph
breaks or separating environments in LyX. Both of the above goals were
previously achieved by the ---Separator--- layout and can now be
accomplished by the new inset in a more natural way. As an example,
after leaving an environment by hitting the Return key for two times,
a third return automatically inserts a parbreak inset, which is
equivalent to the old separator layout, i.e., it also introduces a
blank line in the output. If this blank line is not wanted, the
parbreak separator can be changed to a plain separator by a right
click of the mouse. Of course, an environment can still be separated
by the following one by using the Alt+P+Return shortcut (or the
corresponding menu key), but now the plain separator inset is used
instead of the old separator layout, such that no blank line occurs in
the LaTeX output.
Old documents are converted such that the LaTeX output remains unchanged.
As a result of this conversion, the old separator layout is replaced by
the new parbreak inset, which may also appear in places where the old
algorithm was introducing blank lines while the new one is not.
Note that not all blank lines were actually affecting the LaTeX output,
because a blank line is simply ignored by the TeX engine when it occurs
in the so called "vertical mode" (e.g., after an alignment environment).
The old ---Separator--- layout is now gone and old layout files using it
are also automatically converted.
Round trip conversions between old and new format should leave a document
unchanged. This means that the new behavior about paragraph breaking is
not "carried back" to the old format. Indeed, this would need introducing
special LaTeX commands in ERT that would accumulate in roundtrip
conversions, horribly cluttering the document. So, when converting a
modified document to old formats, the LaTeX output may slightly differ in
vertical spacing if the document is processed by an old version of LyX.
In other words, forward compatibility is guaranteed, but not backwards.
* remove unused class TexStream.
* remove unused virtual method Inset::cellXOffset
* remove second argument of FileDialog constructor, which was actually
not used
* remove some dead local code
* remove some unused private members of classes
* in InsetMathNest::updateBuffer, fix the logic of a test
for possible thread conflicts, of the sort Georg resolved at
6a30211f. I have made static variables const where possible,
and marked cases that looked potentially problematic with the
comment:
// FIXME THREAD
Many of these definitely are vulnerable to concurrent access, such
as the static variables declared at the start of output_latex.cpp.
Suppose, e.g., we were outputting latex and also displaying the
source of a different document.
I'd appreciate it if others could grep for "FIXME THREAD" and see
if some of these are harmless, or what.
This code was commented out at [ad94e7bd/lyxgit], since we thought it was not necessary anymore and then removed at [5aede959/lyxgit]. Bug #9042 is the evidence that we were wrong.
and Cursors. So just calling InsetText::addToToc for the cells causes
problems, because InsetText::addToToc then adds the cell inset itself
as part of the DocIterator. This then leads to assertions, such as bug
The solution is to refactor InsetText::addToToc so that we can call the
iterating part without adding the inset.
The problem was that collectBibKeys() was called before the recursive include
check was done. Now collectBibKeys() works even for recursive includes, and
your get a proper error message if you try to change the file name to the
parent file.
At efa0f198 I introduced a more flexible caption handling for long tables.
This was needed to fix tex2lyx bug #7412. Unfortunately this created bug #8933,
since I forgot to output captions to LaTeX which are not in any header or
footer. This was fixed at 2b941da7f, but this introduced bug #8992 (duplicate
caption output).
This commit is hopefully the final fix: Captions which are in header or footer
lines are output with the header or footer line, and captions in standard
lines are output standalone.
This makes the defaults of Inset::inheritFont() and Inset::resetFontEdit()
compatible. There is no user visible change except for the Chunk inset which
does not produce invalid LaTeX after editing operations anymore.
This is the safe version for 2.1.0, for later there are still open questions:
- All insets with ResetsFont true should be audited: Is this really needed,
or do they show similar editing problems as the Chunk inset?
- Does inheritFont() need to be customizable in the layout file as well?
- Is resetFontEdit() != !inheritFont() needed at all?
I did not use change tracking for the docs, since I updated all existing
translations.
This commit fixes a bug uncovered by the fix to #8082:
If you create a table, select all columns in a row,
and set as multicolumn, the right border used to be unset.
The new HTML clipboard export could cause error message boxes on copying
data to the clipboard (bug #8866). These are now suppressed, like all other
errors which might occur for preparing the clipboard data.
The main part of the fix (unicodesymbols) is from Jürgen. This commit fixes
tree problems:
- \; etc. were also used in text mode, but are math only
- all of those glyphs need to be forced with utf8
- actually, \; etc. are not the correct macros, since the encoded spaces are
breakable, but the math spaces are all protected. The sapce symbols are not
defined in the utf8 encodings.
- InsetBox.cpp:
a framebox without inner box and without a width is \fbox
a framebox with inner box is also \fbox
a framebox without inner box and with width is \framebox
- GuiBox.cpp: the width checkbox must be checked if there is width
* InsetBox and GuiBox: Use proper empty length instead of the broken -9.99col% trick
* some slight changes to the logic of GuiBox to make sure that values are set as needed.
* lengthToWidget(): handle properly the empty length case. All the other related Qt helpers did it already, it was probably an oversight. Also set the default_unit parameter as optional (not needed in this patch actually, but I got carried away :)
* allow generating LaTeX code for an empty length, since some broken code does that.
InsetInfos of the type ICON have a graphics Inset in their paragraph.
These insets were inserted without a font. This lead to an assertion in
Paragraph::fontSpan because the fontlist_ of the paragraph is empty.
This is mostly unused private class members.
There are also a few unused functions that got #if'ed out. I never know in this case whether the code should be nuked.
each failure.
There are several places I was not sure what to do. These are marked
by comments beginning "LASSERT:" so they can be found easily. At the
moment, they are at:
Author.cpp:105: // LASSERT: What should we do here?
Author.cpp:121: // LASSERT: What should we do here?
Buffer.cpp:4525: // LASSERT: Is it safe to continue here, or should we just return?
Cursor.cpp:345: // LASSERT: Is it safe to continue here, or should we return?
Cursor.cpp:403: // LASSERT: Is it safe to continue here, or should we return?
Cursor.cpp:1143: // LASSERT: There have been several bugs around this code, that seem
CursorSlice.cpp:83: // LASSERT: This should only ever be called from an InsetMath.
CursorSlice.cpp:92: // LASSERT: This should only ever be called from an InsetMath.
LayoutFile.cpp:303: // LASSERT: Why would this fail?
Text.cpp:995: // LASSERT: Is it safe to continue here?
As discussed on the list. No automatic contents detection is done, the user
needs to use the special paste menu instead. I used the new TempFile class
for safe temporary file handling.
The documentation would go into section 2.2 of UserGuide.lyx, but I am not
allowed to edit that document.
output, due to failure to clean the ids in the new citation stuff.
I've solved this by allowing the citation format information to contain
keys of the form "clean:key". This signals that we are to apply the
html::cleanAttr() function to the key before returning it. I.e., we
strip non-alphanumeric stuff, basically.
- fileformat change
- it was a pity that LyX did not yet support a simple rectangular frame without a defined width but LyX did this for e.g. oval frames
- \fbox and \mbox often occur in TeX files and can now be imported
Previously, the format used for included pdf files was the same as for
document export via ps2pdf. This caused unwanted conversion routes, e.g.
export via odt->pdf instead of dvi->ps->pdf.
I renamed the format for included graphics and not for exported documents,
since otherwise the command line syntax for export would change. This would
require more adaptions for the users, since with the chosen solution the
custom converters are almost always changed correctly in prefs2prefs(),
so that only custom external templates need manual adjustement.
work we do when calling plaintext() for the purpose of generating
material for the advanced search function.
Here again, not only were we parsing BibTeX files, since Julien's
(sensible) introduction of plaintext output for that inset, but we
were in fact writing (to disk) complete plaintext output for
included files every time we did such a search.
worth doing, as we were creating too much output for tooltips anyway.
But we need to ignore BibTeX insets altogether, as the collection of
the references, etc, is too slow.
so we can write a limited amount when using this for TOC and
tooltip output.
This should solve the problem with slowness that Kornel noticed,
which was caused by our trying to write an entire plaintext
bibliography every time we updated the TOC. We did that because
he had a bibliography inside a branch, and we use plaintext for
creating the tooltip that goes with the branch list.
Other related bugs were fixed along the way. E.g., it turns out
that, if someone had an InsetInclude inside a branch, then we would
have been writing a *plaintext file* for that inset every time we
updated the TOC. I wonder if some of the other reports of slowness
we have received might be due to this kind of issue?
This default argument is inserted iff no inset argument is present. This is useful particularly for mandatory arguments that need to have a sensible default value.
This only move the code, but does not change the displayed labels.
Thus for numerical citation, the label is set to the cite number;
for author-year citation, the abbreviated list of authors is used
i.e. "Smith et al. 2001".
Eventually, we might want to make the label customizable, or get
it from BibTeX.
I do not see where this is really useful (contrary to real plaintext). And it breaks spell checking with hunspell and 8bit disctionaries (part of bug #8526)
