* new MathRow class which contains the description of a MathData
object in terms of math class and spacing
+ macros and their arguments used in the MathData object are
linearized (replaced with their contents) so that all math insets
are typeset as a string together. To this end, we introduce a
method addToMathRow to InsetMath and MathData. This method allows
to linearize recursively a MathData object.
+ It is then necessary to set manually the dimension and position of
the macros and arguments.
+ the class class and spacing are computed using the MathClass helpers.
The MathRow data is cached in the MathData object in a bufferview-dependent
way (different dpi for different screens).
* delegate most of the work MathData::metrics/draw to MathRow metrics/draw.
The case of draw is trickier, since many draw() methods rely on their
metrics without any spacing added.
Math macros can be displayed on screen by providing a different
representation than the one used for latex output. This representation
is actually used by lyx even while it is being updated. This leads to
printing useless error messages on the terminal. For example, a macro
parameter has to be entered as \#1 and, if the macro is already used in
a math inset, lyx prints on terminal the error message "Math parse error:
missing token after \\" as soon as one hits the \ key, followed by
"MathMacroArgument::MathMacroArgument: wrong Argument id: -48" as soon as
one hits the # key. So, this is not a useful information and simply
clutters the terminal output. On the other hand, the input is sanitized
even if one stops input after hitting either \ or #, so that no further
messages are issued. Hence, those error messages are simply pointless.
Trying to spare a few cycles by avoiding computing metrics during
screen updates and export. See also 8f86ee74, 72cf7c8f, and e36a8903.
Guillaume will tell whether this also avoids crashing his documents ;)
Some macros defined in the lib/symbols file are classified are texmode.
But the MathMacro class was missing a currentMode method for returning
this information.
Revert to the strategy used at 8f86ee74 but not using mathedWordList
because it may be still uninitialized at load time. Instead, use the
globalMacros method for getting the same info.
There was a thinko at 8ec91e80, because globalMacros always returns
null for user defined macros.
It may happen that mathedWordList is not still updated at load time,
so we would still be using a bogus pointer. Better fetching the
necessary info from the global macro table.
The math macros system is quite complex. Macros are updated during
metrics calculation, so a missing update is very likely to cause a
crash. This commit tries to assure that they are updated at export
time, which also happens when the table of contents is updated.
Moreover, in order to circumvent a possible missing update, when
a math macro is detected we try to avoid using the sym_ member
of the MacroData class, as it may contain bogus values.
This requires to change many docstrings into std::strings. The logic behind that
is that they represent a fixed set of math fonts, and therefore “string” means
here “poor man's enum” rather than text (this is consistent with MetricsBase).
Profiling of scrolling inside a document over macro-instensive areas:
Before the patch:
44,1% BufferView::updateMetrics()
-> 34,8% InsetMathHull::metrics()
-> 9,8% FontSetChanger::FontSetChanger()
28,4% BufferView::draw()
After the patch:
35,3% BufferView::updateMetrics()
-> 27,2% InsetMathHull::metrics
-> 0,4% FontSetChanger::FontSetChanger()
47,5% BufferView::draw()
FontSetChanger::FontSetChanger() is made 41x less expensive (with reference
BV::draw()) just by removing this conversion. The remaining 0,4% could be
squished by replacing the strings with a proper enum, but this is premature. Of
course, this only treats the symptoms: there is no good reason that this
function is called 45500 times over the time of 40 repaints.
RefChanger temporarily assigns a value to a non-const reference of any
kind. RefChanger provides a flexible and uniform generalisation of the various
scope guards previously derived from the old Changer class in MetricsInfo.h.
As before, a temporary assignment lasts as long as the Changer object lives. But
the new Changer is movable. In particular, contorsions are no longer needed to
change a private field. Special code can be moved into the appropriate classes,
and it is no longer necessary to create a new class for each specific use.
Syntax change:
FontSetChanger dummy(mi.base, value);
-> Changer dummy = mi.base.changeFontSet(value);
New function for generating arbitrary Changers:
Changer dummy = make_change(ref, val, condition);
Bugfix:
* Fix the display of \displaystyle{\substack{\frac{xyz}{}}} (missing style
change).
It turns out that it is always better using the copy of the MacroData
for updating the macro_ pointer to avoid problems related to the cursor
position.
This can happen when a macro is copied and then the document where
it is defined is closed. In this case, the macro survives in the
cut stack but the the buffer pointer is dangling.
The MacroData pointer is updated by MathData::metrics() which is not
called when selecting a math inset with instant preview for math on.
Thus, we have to update it in the copy constructor otherwise a crash
is almost assured when hitting Ctrl+C.
In the test case the crash occured in mathml export of the temporary buffer,
because the macro was updated, and because one of the used other macros was
not copied, the macro argument was detached. However, the underlying problem
of the crash was a broken ArgumentProxy::mathMacro_ reference which became
invalid each time the ownng MathMacro was copied. In the bug test case the
copying happened due to resizing a std::vector, but any other copy would have
created the same problem. The crash did not always happen, because sometimes
the old freed memory was not immediately reused, so the invalid reference did
still point to usable data.
The fix is easy: Convert ArgumentProxy::mathMacro_ to a pointer and update it
always after creating a copy of the owner. The pimpl of MathMacro from the
previous commit helps here to distinguish between the data that can be
automatically copied (in MathMacro::Private) and the cleanup that needs to be
done manually (in MathMacro). This way, the manual copy constructor and
assigment operator of MathMacro does not need to be touched if a new member is
added.
The expanded cells of a mathmacro were previously stored in an InsetMathSqrt.
This was only used as a container for the MathData object in the first cell
of the sqrt inset, which contained the actual expanded arguments.
Funny enough, the only place were the inset property of expanded_ was really
used cannot be seen in the diff. It was MathMacro::kerning(), and this usage
was wrong, since InsetMathSqrt::kerning() always returns 0. Threfore, using
the correct type (MathData) for expanded_ does not only make the code more
readable, gets rid of an unneeded dependency, but also fixes a bug: Now the
correct kerning is returned for expanded cells. Also, expanded_ and
definition_ use the same type now, which looks nicely symmetric.
Previously, things like [ name ] where exported for computer algebra systems.
Now, the expanded macros are exported, which may still be wrong, but now the
CAS has at least a chance to understand what was meant.
Thanks to Scott for testing. If a macro is unknown (displayed in red), then
macro_ is 0. The LATTEST is now adjusted and works like in MathMacro::write()
where I stole it from.
Actually I wanted to do that in cc87f8100 but forgot to adjust the original
solution completely. Now we do not search for an arbitrary latexkeys instance
which just happens to have the same name as the macro, but we only use the
symbol that was explicitly set for global macros.
This is different from bug #8999, since in this case a new macro instance is
created. You still get a TeX capacity exceeded error if you try to typeset the
exported document, but this is the same as for bug #8999 and better than a
crash.
This was a regression of e86cdc40: A newly introduced member variable was
not initialized in the constructor, which made it quite random whether symbols
like \coloneqq where displayed correctly or as an empty edit box.
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?
Actually, the test case showed several problems:
- ERT insets did use layout "Standard", not "Plain Layout"
- The font scale was read correctly, but tex2lyx claimed that it did ignore
the option "scaled=0.95"
- If a third argument of the CJK environment was given, it caused the whole
environment to be put in ERT with a broken encoding. This is now fixed for
the bug test case by using the \font_cjk header variable, but the encoding
problem still exists for unsupported encodings. I'll file a separate bug
for that.
- The CJKutf8 package was not handled in the preamble parsing. Therefore the
chinese comment in the preamble was read with a wrong encoding, and guessing
the document language did not work.
The new file CJKutf8.tex was created by copying and modifying CJK.tex, but
unfortunately it is impossible to tell git to inherit the history of CJK.tex
for the new file (search the web for git svn copy if you want to know details).
as an image.
The math macro now exports as XHTML + MathML and can be viewed in
Firefox.
git-svn-id: svn://svn.lyx.org/lyx/lyx-devel/trunk@38190 a592a061-630c-0410-9148-cb99ea01b6c8
A dynamic_cast is necessary when:
- the object to be casted is from an external library because we can't add Qxxx::asXxxx() to Qt e.g.:
* QAbstractListModel to GuiIdListModel,
* QValidator to PathValidator,
* QWidget to TabWorkArea,
* QWidget to GuiWorkArea;
- the object is to be casted from an interface to the implementing class, because the Interface does not know by whom it is implemented:
* ProgressInterface to GuiProgress,
* Application to GuiApplication.
A dynamic_cast can be replaced by:
- already existing as***Inset() functions, e.g.:
* asHullInset(),
* asInsetMath()->asMacro(),
* asInsetText();
- a static_cast when we are sure this can't go wrong, e.g.:
* we are sure that CellData::inset->clone() is an InsetTableCell,
* in cases where we explicitly check it->lyxCode().
git-svn-id: svn://svn.lyx.org/lyx/lyx-devel/trunk@35855 a592a061-630c-0410-9148-cb99ea01b6c8
