Le 30/05/2016 14:33, Jean-Marc Lasgouttes a écrit :
Le 27/05/2016 23:08, Guillaume Munch a écrit :
The protocol is to change the width of the window (all other things
equal).
Qt4: 75% of the time is spent (in number of instructions) inside
QTextEngine::shapeText(). Of these 75%, 2/3 is requested by metrics
calculation (QTextLine::setLineWidth() from GuiFontMetrics::breakAt())
and 1/3 by painting (GuiPainter::text()). Ultimately, 74% of the time
is spent in in FT_Load_Glyph and FT_Render_Glyph from
libfreetype.so.6.12.1 (thus, including the metrics phase).
At some time, I had the idea of caching the QTextLine object in the row
element. I suspect this would be a good thing to do.
I do not see the point. You already keep all the useful information.
Keeping the QtextLine object is not going to make setLineWidth run faster.
Qt5: QTextEngine::shapeText() only accounts for 13.22% of the cycles,
and costs 28 times less per call (69729 instructions vs 1932642).
QTextLine::setLineWidth() only costs 4% total. GuiPainter::text() costs
21%, with time spent in a certain QRasterPaintEngine::drawCachedGlyphs().
This difference is very surprising, there has to be something that is
different.
Conclusion: using QTextLine::setLineWidth() breaking rows is very costly
with qt4.8.7, and much less with qt5. I agree with you that I do not
remember seeing this slowness before, and I do not know what exactly
could trigger this behaviour.
I have to double-check, but this seems indeed slower now than it used to
be. What is your OS?
What tool did you use to assess performance?
Ubuntu 16.04, callgrind.
Now if you tried qt5 you could see what fast means :)
