Otherwise the VmData size will keep increasing in multi-threads mode.
Although the increasing VmData size doesn't mean there is enormous
memory leak, it does show that the thread stack is not handled properly
when threads exit.
Memory leak came from the master bitmap that is reused in each reflow. But the bitmap
width and height were not set to zero when bitmap was freed. When the master bitmap was
enlarged in some cases the size of the master bitmap would increase monotonically.
This patch fixes this issue by setting master bitmap width and height to zero at the
beginning of each reflow. And a memeory usage logger is added in koptreader.
Because the lua reader is single threaded on which both user inputloop
and background page rendering is processed. Although there is a pretty
good precache system to save user's time spending on waiting for the
rendering when going to the next page, user input is indeed blocked when
running the precache thing. The situation is even worse in koptreader as
reflowing on page would usually take several second, in this period
users cannot move to the next page view even it's already in the cache.
This patch will let precache run in the background in a seperate thread
so that the koptreader is still responsive when precaching the next
page. Now it only just works. Welcome to find out bugs in it.
Probably we should use fomit-frame-pointer explicitly even though
fomit-frame-pointer is included by default with -O2 and -O3 just like
the Linux kernel guys doing. I found append -O3 directly to `CFLAGES`
will do the trick. So I removed the KOPT_CFLAGS variable.
In some compilation platform if finite-math-only option is turned
on, math functions like exp and sqrt will be dynamically linked to
finite versions which cannot be located in Kindle's GLIBC. In my
toolchain the symbol __exp_finite cannot be found in GLIBC_2.4 so
gcc just use __exp_finite in GLIBC_2.15, which will cause a run time
error in Kindle saying "version GLIBC_2.15 not found"
Reflow can be configured to use bbox which could remove
side comments and defects in the margin and get much better reflowed
page in some documents. Currently the bbox is only set in PDFReader or
DJVUReader reader. But eventually there will be bbox setting routines in
KOPTReader.
Thirdparty should be cleaned and refetched and remake to get
this feature in operation.
Many users complained that the reflowing process is too slow. This is
rather a speed-quality dilemma. Trade-off can be made by decreasing the
dpi of source page. And by default quality is prefered.
By default the indentaion detection is enabled in K2pdfopt. This will
sometimes generate very poor reflowed page when there are sidenotes on
the page which would be treated as indentation. Disabling indentaion
detection will rescue the reflowed page from this situation.
