There are various changes involved into making this work:
1. Zone size (zsz) is no longer clamped to the timeline viewport area.
This clamping has to be removed to prevent otherwise uncollapsed zones
from apparently becoming small near the viewport borders. Such a small
zone would then be collapsed, resulting in unwanted popping.
Interesingly, only the CPU zones were clamped before. GPU zones were
not.
2. Iteration over visible zones has to start before the visible timeline
viewport area. Without this some zones that would be otherwise
included in the collapsed area (started by a previous zone) may be
fully visible. This causes child zones to be drawn and produces
unwanted popping. (At this point threshold for continuing collapsed
area is greater than threshold for starting it.)
3. Since the iteration now starts before timeline visible area, it may so
happen that everything found will be in a small slice of timeline that
is outside the screen. To fix this, the end time of last found item is
checked against the viewport start time.
It is always valid to access *(zitend-1), as it is in each case done
after null set check (it == zitend).
Similar but simpler fix was also applied to per-thread call stack samples.
The reasoning is that you want to use the color to see where a zone of
a particular type is placed. When collapsed zones go back to displaying
thread color, you may mistake such region of collapsed zones for something
they aren't.
Previously on demand mode was determined by frame offset parameter being
greater than zero. However, if the application is not pumping frames with
FrameMark macro, the frame index will never increase and the frame offset
parameter stay at zero. It is not possible to distinguish on demand traces
from normal ones in this scenario.
Fix by explicitly saving the on demand flag in trace file and employ the
previous logic to set the flag when importing older traces.
