This is to make adding support for new formats easier.
See the file pic_jpeg.h for the description of the interface that each
image format must support.
On some distributions (e.g. Ubuntu) the shell used for scripts is dash,
not bash and we explicitly rely on bash's features in the Makefile, so
we must guarantee that we are running with bash (not dash or anything
else).
think about setting the alignment trap to a mode that includes warn,
that simply kills the performance because of the storm of faults the
syslog gets...
MG 2012.03 pulling GLIBC_2.15 symbols from libm (because -ffast-math asks
for it).
I didn't see this on my own TC because it's using a much older glibc version (2.9).
instead of HOST (and allow it to be set from the env), use $(MAKE)
instead of make to allow using the jobserver properly, and remove the
dash from commands where we do care about the return code (or inhibit
errors the usual way: rm -f instead of -rm)
Enaling link time optimizations for kpdfview binary and crengine library
causes another slight performance increase --- almost negligible (1-2ms
per page) but still noticeable by precise measurements.
Thanks to NiLuJe who pointed out that our generic -march=armv6 can be
replaced (for K2/K3/DX/DXG) with a more specific optimization:
-march=armv6j -mtune=arm1136jf-s -mfpu=vfp. This I have now done and
also passed ARM_CFLAGS value to CXXFLAGS which is then passed to
crengine build. Tested, works fine. The performance improvement is
negligible (a few ms per page, but seems to be consistently better, i.e.
not just plus/minus fluctuations).
Thanks to NiLuJe who pointed out that we are building mupdf in debug
mode. Switching to "release" build reduced the size of the kpdfview
binary and did not cause any performance degradation (but no noticeable
improvement either --- the page handling times seem to be exactly the
same, i.e. fluctuating a couple of ms in both directions).
1. Remove unused "-lsdtc++" from the compilation stage as no linking is
done then.
2. Add our standard CFLAGS to the compilation of cre.cpp which brings in
-O3 -march=armv6 which enables optimizations.
It seemed strange that we compile a cpp file with gcc (as opposed to
g++), but I left it as is for now.
Instead of calling lfs.mkdir() to create "./history" and "./screenshots"
at runtime it is easier to create them at package build time.
I hesitated whether to add "./clipboard" to this list but decided
against it as we can perhaps change current directory and then
all the code manipulating clipboard would break, so I left it as is.
1. Don't build xmltools in the emulator
2. Remove duplicate --disable-desktopfiles
3. Don't build LFS support. The largest DjVu file I have ever published
was a highres facsimile edition of the London Walton Polyglot (1657)
which was a "mere" 1GB in size and I don't think anyone produced
anything bigger. Besides, storing DjVu files >2GB in size (even if
they existed, which I doubt) on a Kindle with only 3GB total storage
space is _exceedingly_ unlikely.
