This adds more documentation about the viewport offset calculations
for refreshing rotated viewports.
The old behaviour also was (still) buggy, even after the latest round
of fixes. Now, the offsets when rotated are hopefully calculated correct.
This will convert any file name to lowercase before doing the comparison.
Note that this will only work for ASCII character range, a full Unicode
aware solution will be much more complicated. And in the end, file names
are byte arrays, not character strings ;-)
fixes#1183.
The strcoll() workaround we had in place for Kobo devices was (or has
become) ineffective. We had set self.strcoll to nil on Kobo devices -
but this was the instance variable. Setting it to nil effectively makes
the instance variable vanish, so when trying to access it later, it
was not there and got looked up via the metatable - which had the original
reference. Setting it to nil had no effect whatsoever.
We simplify that approach and set the replacement function where before we
had set this to nil.
This is a partial fix for issue #1183 (and explains a comment in issue #686
which says that the old fix did not work).
However, to really fix#1183 - if we want to do so - we would need a collate
function that normalizes uppercase/lowercase before compare.
The calculation did not correctly calculate the offset in rotated
modes when the rotation was 90 or 270 degrees. Under these circumstances,
getWidth() and getHeight() did return rotated values for the framebuffer,
while Device.viewport did return not rotated values.
Since commit 12a76fee33, we had a potential
bug on the event mechanism:
It introduced (besides the checkTasks method itself) a second run of the
checkTasks() method. In the second run, however, scheduled events were
not taken into consideration in how long to wait for input events
afterwards.
So when the after the first run of checkTasks() there were new scheduled
tasks added to the task queue, they were not properly scheduled and
and depended on an already existing scheduled event or an input event
to trigger.
This might have led to unexpected order of execution (though the order
is not guaranteed by the task scheduling anyway!) or to events triggering
not at all until the next input event.
And koreader release version is normalized to a 10 digits containing
a four digits year field followed by a two digits month field followed
by a four digits revision field like "2014041079".
Don't do regional footer refreshes when we're in flipping mode.
There's probably a more elegant/correct way to handle this, but this
does the job ;).
Fix#1140
the cache would behave badly when the same item was insert()ed twice:
it would add the size twice to memory consumption, but would never substract
it twice when purging the (actually single) object from cache. So the cache
would seem to fill up while in fact it wasn't.
