We rely on one variable _dmon.modify_watches shared between thread to
ensure that we don't lock with the dmon polling thread waiting indefinitely
and helding a lock.
To ensure that the polling thread sees modifications done to 'modify_watches'
we use an additional mutex that act as a memory barrier.
Should fix commit bb12f085f3.
When taking the critical section we should always send the
event to wakeup the events thread. In addition use
TryEnterCriticalSection to send the event only if needed
reducing the number of spurious events sent.
Avoid waiting with a finite timeout and wait indefinitely in
dmon thread. When we need to unwatch we send a signal to a special
event meant to wakeup the waiting thread.
Some asserts are placed in case that can effectively occur
so we remove the assertion and we return false. In turn we
adapt the logic accordingly so when false is returned to add
a watch we do not open that directory.
We no longer use in Lite XL recursive directory monitoring as
it was a source of problems.
To enforce this at compile time we use the preprocessor to
remove the implementation of recursive monitoring for the Linux
implementation only.
In order to get right the cursor position on text on mouse clicks it is
needed to take into account text's subpixel positioning.
This fix mostly corrects the problem but cursor positioning is still
somewhat inaccurate for long lines due to repeated commands to draw a text
along a line. Repeated draw text calls make the subpixel information
lost and small errors will add-up.
Seen with some fonts like FiraSans, github issue:
https://github.com/franko/lite-xl/issues/46
The fix works essentially by looking to the bounds of each glyph to
accurately ensure that there are no overlaps between the glyphs.
The construction of the font atlas was changed to make some related
improvements now that the bounds of each glyph are know. The main
changes are:
- no longer align glyph on the baseline but align them on their upper
bounds. We ensure this way that very tall fonts do not leak in the
upper part
- terminate the row based on x bounds of the glyph to be more accurate
- for each row keep trace of the y of the more larger along y of the
glyph. The value is used to start a new row to be sure that the new
now does not overlap with the previous one
- sort glyphs by y size before drawing them. In this way the space
utilization is better. The algorithm used is the very simple insert
sort. It behaves like O(n^2) with the number of characters but should
be ok since n is always small, typically below 128.
- compute the optimal image width and height for the given font's atlas
for optimal memory usage. As a bonus now the lite's code don't have to
try and repeat to get a good image size
Ensure we set the font's size with freetype to the intended size so that
hinting works as expected and perform subpixel scaling using affine
matrix coefficient.
Text looks more correct without the gamma correction. With gamma correction
text looks too bold in dark mode and too light in light mode. This
behavior is a sign that gamma correction is not the right thing to do.
The problem is seen with gamma at 1.5 used previously but is even worst
if a gamma value of 2.2, the theoretical value, is used.
Francesco Abbate
Guldoman
Jan200101
Adam Harrison