Some results from the benchmark (compilation of en_US.UTF-8/Compose):
$ grep 'model name' /proc/cpuinfo
model name : Intel(R) Core(TM)2 Duo CPU E8400 @ 3.00GHz
model name : Intel(R) Core(TM)2 Duo CPU E8400 @ 3.00GHz
$ uname -a
Linux ran 3.16.1-1-ARCH #1 SMP PREEMPT Thu Aug 14 07:40:19 CEST 2014 x86_64 GNU/Linux
$ ./test/compose bench
compiled 1000 compose tables in 7.776488331s
So according to the above benchmark and valgrind --tool=massif, an
xkb_compose_table adds an overhead of about ~8ms time and ~130KB
resident memory.
For contrast, a plain US keymap adds an overhead of ~3ms time and 90KB
resident memory.
Signed-off-by: Ran Benita <ran234@gmail.com>
We added a -fvisibilty=default to the libtest.la CFLAGS. When automake
sees that the CFLAGS are different from those of libxkbcommon.la, it
figures it should recompile all of the $(libxkbcommon_la_SOURCES),
instead of reusing the already-compiled object files.
But actually, the convenience library is not a shared object, it's just
used as an archive, thus the -fvisibility is entirely pointless. So
avoid the recompilations by just removing it.
Signed-off-by: Ran Benita <ran234@gmail.com>
Remove the deprecated symbols that were used for ABI compatibility
during the transition period to the first stable version, 0.2.0.
The old *names* can still be used, programs which use the old names will
continue to work, as long as they were compiled against a stable
version (as they have been #defined to the new names from the start; see
xkbcommon/xkbcommon-compat.h). Namely, this will break binaries which:
1. Were compiled against a pre-stable version of libxkbcommon, and
2. Are linked against the next version of libxkbcommon, and
3. Expect to work.
This scenario is very unlikely, and will break in many other ways
anyway. Also, retaining support for these means including them in the
new symbol version file, which I would like to avoid.
Signed-off-by: Ran Benita <ran234@gmail.com>
Symbol versions provide a means by which ELF utilities can determine
whether a program is incompatible with a too-old library version so
that package management tools can autodetect version-based
dependencies and suggest upgrade paths.
[ran: swap xkbcommon.map and xkbcommon-x11.map]
Signed-off-by: Jan Engelhardt <jengelh@inai.de>
Signed-off-by: Ran Benita <ran234@gmail.com>
There is really no reason to deny these tests from different platforms
only for a few #defines.
The only linux-only test (or test program, it is not run by make check)
is interactive-evdev, which actually uses evdev.
Signed-off-by: Ran Benita <ran234@gmail.com>
It is like test/stringcomp, only instead of using
xkb_keymap_new_from_string(), it uses xkbcomp to upload the keymap to a
dummy Xvfb X server and then xkb_x11_keymap_new_from_device().
If any of these components are not present or fails, the test is shown
as skipped.
The test is messy, fragile, limited and depends on external tools, but I
will improve on that later -- it's better to have a test.
Signed-off-by: Ran Benita <ran234@gmail.com>
We need to validate some UTF-8, so this adds an is_valid_utf8()
function, which is probably pretty slow but should work correctly.
Signed-off-by: Ran Benita <ran234@gmail.com>
Add two tests:
./test/interactive-x11
which is like test/interactive-evdev, but should behave exactly like your
X keyboard and react to state and keymap changes - in other words, just
like typing in xterm. Press ESC to exit.
./test/x11
which currently should only print out the same keymap as
xkbcomp $DISPLAY out.xkb
(modulo some whitespace and some constructs we do not support.)
Signed-off-by: Ran Benita <ran234@gmail.com>
These are function to create an xkb_keymap directly from XKB requests
to the X server. This opens up the possibility for X clients to use
xcb + xcb-xkb + xkbcommon as a proper replacement for Xlib + xkbfile for
keyboard support.
The X11 support must be enabled with --enable-x11 for now.
The functions are in xkbcommon/xkbcommon-x11.h. It depends on a recent
libxcb with xkb enabled. The functions are in a new libxkbcommon-x11.so,
with a new pkg-config file, etc. so that the packages may be split, and
libxkbcommon.so itself remains dependency-free.
Why not just use the RMLVO that the server puts in the _XKB_RULES_NAMES
property? This does not account for custom keymaps, on-the-fly keymap
modifications, remote clients, etc., so is not a proper solution in
practice. Also, some servers don't even set it. Now, the client just
needs to recreate the keymap in response to a change in the server's
keymap (as Xlib clients do with XRefreshKeyboardMapping() and friends).
Signed-off-by: Ran Benita <ran234@gmail.com>
This helps ensure we don't ship a library with undefined symbols, which
can easily happen with multiple inter-dependent DSOs.
Signed-off-by: Ran Benita <ran234@gmail.com>
This makes it easier to share the private functions in other DSOs
without relying (too much) on dead code elimination, exported symbols,
etc.
Signed-off-by: Ran Benita <ran234@gmail.com>
Even though the %name-prefix is more sensible, byacc doesn't support it,
but both bison and byacc support the -p argument.
Signed-off-by: Ran Benita <ran234@gmail.com>
This was an oversight: even though we ship the pre-built files, it is
still good behavior to include the original generators / source files in
the tarball.
Signed-off-by: Ran Benita <ran234@gmail.com>
The scanner is very similar in structure to the one in xkbcomp/rules.c.
It avoids copying and has nicer error reporting.
It uses gperf to generate a hashtable for the keywords, which gives a
nice speed boost (compared to the naive strcasecmp method at least). But
since there's hardly a reason to regenerate it every time and require
people to install gperf, the output (keywords.c) is added here as well.
Here are some stats from test/rulescomp:
Before:
compiled 1000 keymaps in 4.052939625s
==22063== total heap usage: 101,101 allocs, 101,101 frees, 11,840,834 bytes allocated
After:
compiled 1000 keymaps in 3.519665434s
==26505== total heap usage: 99,945 allocs, 99,945 frees, 7,033,608 bytes allocated
Signed-off-by: Ran Benita <ran234@gmail.com>
This wraps the current mmap call and adds a fallback implementation for
systems which do not have mmap (e.g. mingw).
Signed-off-by: Ran Benita <ran234@gmail.com>
The current API doesn't allow the caller to create keymaps from mmap()'ed
files. The problem is, xkb_keymap_new_from_string() requires a terminating
0 byte. However, there is no way to guarantee that when using mmap() so a
user currently has to copy the whole file just to get the terminating zero
byte (assuming they cannot use xkb_keymap_new_from_file()).
This adds a new entry xkb_keymap_new_from_buffer() which takes a memory
location and the buffer size in bytes.
Internally, we depend on yy_scan_{string,byte}() helpers. According to
flex documentation these already copy the input string because they are
wrappers around yy_scan_buffer().
yy_scan_buffer() on the other hand has some insane requirements. The
buffer must be writeable and the last two bytes must be ASCII-NUL. But the
buffer may contain other 0 bytes just fine.
Because we don't want these constraints in our public API,
xkb_keymap_new_from_buffer() needs to create a copy of the input memory.
But it then calls yy_scan_buffer() directly. Hence, we have the same
number of buffer-copies as with *_from_string() but without the
terminating 0 requirement.
The explicit yy_scan_buffer() call is preferred over yy_scan_byte() so the
buffer-copy operation is not hidden somewhere in flex.
Maybe some day we no longer depend on flex and can have a zero-copy API. A
user could mmap() a file and it would get parsed right from this buffer.
But until then, we shouldn't expose this limitation in the API but instead
provide an API that some day can work with zero-copy.
Signed-off-by: David Herrmann <dh.herrmann@gmail.com>
[ran: rebased on top of my branch]
Conflicts:
Makefile.am
src/xkbcomp/xkbcomp.c
Some tests use linux/input.h (and epoll), but we're building on some
other kernels (e.g. debian freebsd). We could just copy the file but
it's GPL. We could also skip the tests (exit code 77) but it doesn't
really matter.
Signed-off-by: Ran Benita <ran234@gmail.com>
Make it a bit easier to experiment with other formats.
Add a struct xkb_keymap_format_operations, which currently contains the
keymap compilation and _get_as_string functions. Each format can
implement whatever it wants from these.
The current public entry points become wrappers which do some error
reporting, allocation etc., and calling to the specific format. The
wrappers are all moved to src/keymap.c, so there are no XKB_EXPORT's
under src/xkbcomp/ anymore.
The only format available now is normal text_v1.
This is all not very KISS, and adds some indirection, but it is helpful
and somewhat cleaner.
Signed-off-by: Ran Benita <ran234@gmail.com>
When running autoreconf, it's possible to give flags to the underlying
aclocal by declaring a ACLOCAL_AMFLAGS variable in the top level
Makefile.am.
Putting ${ACLOCAL_FLAGS} there allows the user to set an environment
variable up before running autogen.sh and pull in the right directories
to look for m4 macros, say an up-to-date version of the xorg-util macros.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Cc: Daniel Stone <daniel@fooishbar.org>
This runs a bunch of random keys against xkb_state_update_key() and
xkb_state_key_get_one_sym(), in a fairly unintelligent way.
It might be nice to check when modifying this code path, or changing it,
to see things haven't slowed down considerably. However, given the
numbers this benchmark gives, it is pretty clear that we are not going
to be the bottleneck for anything. So this can more-or-less be ignored.
Incidentally, this also turned out to be a poor man's fuzzer, because it
turned up the fix in the previous commit. Maybe we should consider
beefing it up with an actual 'break stuff' intention and running it as
part of 'make check'.
Signed-off-by: Ran Benita <ran234@gmail.com>
The ks_tables.h file is generated by makekeys.py from
xkbcommon-keysyms.h, which in turn is generated initially by 'make
update-keysyms'. The xkbcommon-keysyms.h file is commited to git and
distributed in the tarball. Since ks_tables.h should only ever change
when xkbcommon-keysyms.h changes, it is more sensible to update them
together and treat them the same, instead of generating ks_tables.h
every time for every builder with 'make', as we do now.
This means we don't need python as a build dependency (only the one
running update-keysyms, i.e. no one, needs this), and we can be
sure exactly the same file is used by everyone. We also don't need to
run makekeys.py on every build.
Signed-off-by: Ran Benita <ran234@gmail.com>
This removes the complicated and undocumented hash-table creation-helper
and replaces it with an autogenerated sorted array. The search uses simple
bsearch() now.
We also tried using gperf but it turned out to generate way to big
hashtables and when reducing the size it isn't really faster than
bsearch() anymore.
There are no users complaining about the speed of keysym lookups and we
have no benchmarks that tell that we are horribly slow. Hence, we can
safely use the simpler approach and drop all that old code.
Signed-off-by: Ran Benita <ran234@gmail.com>
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
Simple HTML docs generated from the doxygen comments.
After running 'make' or 'make doc', try firefox doc/html/index.html to
see it (if you have doxygen). It's also installed with 'make install'.
You can use --enable-docs or --disable-docs, or specifically
--with-doxygen or --without-doxygen (autodetected, default yes).
The docs are currently not distributed in the tarball, because I
couldn't make it work properly in all cases :/
Signed-off-by: Ran Benita <ran234@gmail.com>
Ran Benita
Jan Engelhardt
Jasper St. Pierre
David Herrmann
Daniel Stone
Damien Lespiau