lyx_mirror/development/PAINTING_ANALYSIS

# -*- org -*-
Understanding the painting process

This file tries to describe the state of the metrics/painting
mechanism, and identify the improvements that could be made. The first
section can be read alone, although the context for them is really
given in the following ones.

Please keep this file up to date as the code evolves!!!

Abbreviations:
bv: BufferView
pm: ParagraphMetrics
tm::TextMetrics

* Questions / Ideas

These questions are consequences of the description made in the
following section. Some actions are proposed.

** SinglePar update

The flag Update::SinglePar is set in many places but never acted on.
Instead, metrics update is skipped whenever the recorded height of
current paragraph did not change and Update::Force was not specified.
Is it better to keep that (which incurs extra work) or to condition it
on Update::SinglePar? If the first solution is kept, the flag
SingleParUpdate shall be removed.

Moreover, I fail to see (yet) where the 'single' part of the program
is acted on.

** Buffer::change issues

When calling Buffer::changed outside of bv::processUpdateFlags,
how do we know that the update strategy is set correctly? It is
possible to reset the strategy at the end of bv::draw. What would be
a good value? NoScreenUpdate?

On a related note, what is the semantics of a call to
Buffer::changed(false)? What does the caller mean?

** How to avoid redraw with FitCursor when the cursor is already OK?

In this case, we invoke Buffer::change(false) with drawing disabled
and NoScreenUpdate strategy.

In the draw phase, bv::checkCursorScrollOffset (the horizontal
scrolling machinery) will change the strategy to FullScreenUpdate if
the current row needs further scrolling.

When the update strategy it kept to NoScreenUpdate, there is currently
a no-draw full repaint, which should not be necessary. It would be
possible to avoid that if the call to checkCursorScrollOffset was done
in bv::processUpdateFlags instead of bv::draw.

The global idea would be to extend FitCursor to cover also horizontal
cursor.


* Proposals


* Clean-up of drawing code

The goal is to make painting with drawing disable fast enough that it
can be used after every metrics computation. Then we can separate real
drawing from metrics.

** DONE RowPainter

Inset position is set in paintInset, paintOnlyInsets, and paintText.
This should be done only once in paintInset

** DONE TextMetrics::drawParagraph

We can really simplify the code when drawing is disabled only
paintInset needs to be called.
 + do right at the start when drawing is already disabled
 + do it in the loop for rows that are not visible on screen.

The only thing we want to do here is to set inset positions (for
text). The other insets still use the painter with drawing disabled.

** DONE Painter::text

We cannot remove (or make private) the version that uses a
FontInfo because it is used by PainterInfo::draw. Document this and
remove unused arguments rtl and double spacing. This would become a specialized helper.
Proposed solution: keep the existing function, but private and without
optional arguments.

Avoid to return (and thus compute) the width of strings?
 + used by InsetSpecialChar (fixable)
 + used by textDecoration() in text(): more difficult to fix

Idea: add a version of text where wordspacing and textwidth (giving
the width of strings) are required parameters and remove optional
version.

==> more versions, no optional parameters.

** DONE When a document ends with a newline, add the bottom margin anyway

The code that tests for a newline was added at 6bb98d07 in 2007.

** When a paragraph ends with a newline, compute correctly the height of the extra row.
** Rewrite TextMetrics::completionPosAndDim using row information

Currently it uses setRowHeight in a very weird way. In particular the
topBottomSpace parameter should be removed after that.

** Rewrite TextMetrics::editXY, checkInsetHit using row information (getPosNearX)?

   The helper version should return a Row::Element instead of an InsetTable.

** TODO make Inset::display() more useful

[This has been started in the features/betterbreak branch. Time will
tell whether it really helps. The question in particular is the
handling of separator insets]

Extending the DisplayType enum would allow to remove special cases
from the code.

The enumeration could be like
: Inline = 0
: BreakBefore = 1 // break row before this inset
: BreakAfter = 2 // break row after this inset
: CanBreakAfter = 4 // optionally break row after this inset
: AlignLeft = 8
: AlignRight = 16
: NoBoundary = 32 // do not allow cursor to go at the end of the row
:                 //before display inset
: Display = BreakBefore|BreakAfter

A display equation would be Display, other could be Display|AlignLeft
BreakAfter can be used by Newline or separator insets
CanBreakAfter can be used by the optional hyphen InsetSpecialChar.

** Set inset position during metrics phase

In order to do that, a no-paint drawing will be initiated after every
redoParagraph. This code path will need to be made as fast as possible.

Effect: avoid depending on actual drawing having taken place. In turn,
it will allow to do drawing on paint events, like any reasonable
application would do.

** Cleanup after complete metrics
   Then the following can be done:
   + remove hack in InsetMathNest::drawSelection
   + remove painting when not inside in drawParagraph
   + remove Cursor::inCoordCache?

** Use Row for MathData

It may not be so difficult. Implement x2pos and pos2x from
the TM:cursorX and TM::getPosNearX, and use them for both text and
math.

Will the strings display OK if drawing string-wise?

Then it would be possible to streamline drawing with disabled painter.

** Paint directly to screen

Instead of using an intermediary pixmap. I have no idea of how
difficult it will prove.
One benefit will be that subpixel aliasing will work again (#9972)

** Merging bv::updateMetrics and tm::metrics

While the full metrics computation tries hard to limit the number of
paragraphs that are rebroken, the version that is used for inner inset
does not try any such optimization. This can be very bad for very tall
insets. We should re-use the bv::updateMetrics logic:
 + transfer all the logic of bv::updateMetrics to tm.
 + Main InsetText should not be special.

The difficuly for a tall table cell for example, is that it may be
necessary to break the whole contents to know the width of the cell.


* Description of current drawing mechanism

** Two stage drawing

There are two parts to drawing the work area:

 + the metrics phase computes the size of insets and breaks the
   paragraphs into rows. It stores the dimension of insets (both
   normal and math) in bv::coordCache.

 + the drawing phase draws the contents and caches the inset
   positions. Since the caching of positions is useful in itself,
   there is a provision for drawing "without" drawing when the only
   thing we want is to cache inset positions
   (Painter::setDrawingEnabled).

The machinery is controlled via bv::processUpdateFlags. This method is
called at the end of bv::mouseEventDispatch and in
GuiApplication::dispatch, via the updateCurrentView method. There are
also several calls in code related to dialogs. We should evaluate
whether this is correct.

Depending on the Update::flags passed to the method, it sets an update
strategy in (NoScreenUpdate, SingleParUpdate, FullScreenUpdate,
DecorationUpdate). It triggers a recomputation of the metrics when either:

 + Update::Force has been specified
 + Update::FitCursor has been specified and there is a need to scroll
   the display.
 + the current paragraph, after rebreak, does not have the same height as in
   existing metrics. Note that the Update::SinglePar flag is *never*
   taken into account.

The screen is drawn (with appropriate update strategy), except when
update flag is Update::None.


** Metrics computation

This is triggered by bv::updateMetrics, which calls tm::redoParagraph for
  + all visible paragraphs
  + paragraph above the screen (up to one page)
  + paragraphs below the screen (up to one page again)

The paragraphs outside of the screen are required to make PageUp/Down
work.

tm::redoParagraph will call Inset::metrics for each inset. In the case
of text insets, this will invoke recursively tm::metrics, which redoes
all the paragraphs of the inset.


** Drawing the work area.

This is done in bv::draw. This method is triggered mainly by
Buffer::changed, which draws all the work areas that show the given buffer.

Note that, When Buffer::changed is called outside of
bv::processUpdateFlags, it is not clear whether the update strategy
has been set to a reasonable value beforehand.

The action depends on the update strategy:

 + NoScreenUpdate: repaint the whole screen with drawing disabled.
   This is only needed to update the inset positions. I am not sure
   when this is necessary, actually. This is triggered when:
   - Update::FitCursor is set but the cursor is already visible. It is
     not clear why something needs to be done in this case, since this
     should be equivalent to Update::None.
   - this is also set when update flag is Update::None, but this value
     is AFAICS not acted on in the code (meaning that nothing happens
     at all in this case).

 + FullScreenUpdate: repaint the whole screen. This is set when:
   - updateMetrics has been invoked.
   - scroll value of current row has changed (although this should not
     be necessary).

 + DecorationUpdate: actually like FullScreenUpdate, although it is
   intended to only update inset decorations. This triggers when:
   - Update::Decoration is set (without Update::Force) as flag.
   - An hovered inset is detected.

 + SingleParUpdate: only tries to repaint current paragraph in a way
   that is not yet very clear to me.