Luis de Bethencourt [Wed, 17 Feb 2016 15:05:49 +0000 (15:05 +0000)]
Revert "orcprogram-c: avoid running out of bounds of varnames"
This reverts commit
951091788b8496868bf86b2f853f5214659e782a.
This isn't a proper fix and the risk of bringing in a regression is too
high for the small and incomplete improvement.
Sebastian Dröge [Wed, 30 Dec 2015 18:52:58 +0000 (20:52 +0200)]
configure: Only check for Android's liblog on Android targets
https://bugzilla.gnome.org/show_bug.cgi?id=760008
Koop Mast [Sun, 20 Dec 2015 23:43:49 +0000 (00:43 +0100)]
configure: Make -Bsymbolic check work with clang.
Update the -Bsymbolic check with the version glib has. This version
works with clang.
https://bugzilla.gnome.org/show_bug.cgi?id=759713
Luis de Bethencourt [Mon, 14 Dec 2015 13:46:52 +0000 (13:46 +0000)]
orcprogram-c: avoid running out of bounds of varnames
These two for loops use varnames[i], since varname has 48 items avoid
running out of bounds by running the for loop up to that limit.
CID 1147004
Tim-Philipp Müller [Tue, 20 Oct 2015 10:28:52 +0000 (11:28 +0100)]
doc: update from entropywave locations to gstreamer
The entropywave website isn't even up any longer.
Docs have been uploaded to gstreamer website too now.
Tim-Philipp Müller [Tue, 20 Oct 2015 09:58:23 +0000 (10:58 +0100)]
doc: update generated opcode tables
Tim-Philipp Müller [Thu, 15 Oct 2015 23:05:03 +0000 (00:05 +0100)]
tools: remove unused test.orc
Tim-Philipp Müller [Thu, 15 Oct 2015 10:15:01 +0000 (11:15 +0100)]
build: fix detection of AMD Kn host cpu
Not that anyone noticed in all these years, so
the check is probably bogus.
Wim Taymans [Mon, 7 Sep 2015 13:30:36 +0000 (15:30 +0200)]
executor: fix load of parameters smaller than 64 bits
Parameters less than 64 bits are passed in just one param register.
Fixes crashes on ldresnearb and friends in emulated code.
Wim Taymans [Mon, 7 Sep 2015 11:43:55 +0000 (13:43 +0200)]
compiler: also prefer the backup function when no target
If we don't have a compiler target (like for ppc64), prefer to use the
backup function instead of doing full emulation.
Wim Taymans [Wed, 12 Aug 2015 11:16:28 +0000 (13:16 +0200)]
test-limits: improve test without target
When there is no target available, the unit test would fail with a
non-fatal error (because emulation would be done). The non-fatal error
would however not be overwritten by the real error when we exceed the
limits.
First relax the first compilation test to check for FATAL errors. Then
reset the program (and clear any non-fatal errors) before triggering
the next error.
Sebastian Dröge [Wed, 24 Jun 2015 20:26:31 +0000 (22:26 +0200)]
Back to development
Sebastian Dröge [Wed, 24 Jun 2015 20:19:53 +0000 (22:19 +0200)]
Release 0.4.24
Wim Taymans [Mon, 22 Jun 2015 14:42:58 +0000 (16:42 +0200)]
parse: increase array size
Increase the size of the array to hold our variable name and use
snprintf to avoid overflows.
Wim Taymans [Mon, 22 Jun 2015 11:13:36 +0000 (13:13 +0200)]
program: only reuse constants of the same size and value
When reusing an existing constant, make sure not only the value but also
the size is the same. We then also need to make a name based on the size
and value so that we can store the same value for different sizes.
Fixes problem with the bayer functions that used 255 as a 16 and 8 bit
constant. This was not detected in the unit test because both the backup
and asm functions do the same error, but it could be seen when a new
backup function was checked against a older ORC.
Tim-Philipp Müller [Fri, 22 May 2015 13:11:32 +0000 (14:11 +0100)]
testsuite: fix reading .orc files with win32 line endings on windows
Just read them in binary form with win32 line endings and
parse them to the orc parser as-is now that the parser
can handle win32 line endings.
Tim-Philipp Müller [Fri, 22 May 2015 12:55:46 +0000 (13:55 +0100)]
testsuite: add parsing test case for win32 line endings
Should handle \r\n as well as \n instead of relying on
the file reader to translate it.
Tim-Philipp Müller [Fri, 22 May 2015 13:07:09 +0000 (14:07 +0100)]
orcc: fix reading .orc files with win32 line endings on windows
We used to open the files in text/translation mode, which
makes fread() convert all windows-style line endings (\r\n)
to unix style line endings (\n) under the hood. However, the
file size we get from the seek is apparently still the original
binary size, and then we just get less data because the \r have
been removed. This means we put our zero terminator in the wrong
place, and there might be some junk between the end of the
text data and the zero terminator.
Fix this by reading the file in binary mode (who knows what
else the text mode might mess up or trip over, e.g. unicode
chars in comments or such) and pass it to the orc parser
as-is now that the parser can handle win32 line endings.
Spotted by Ray Molenkamp.
Tim-Philipp Müller [Fri, 22 May 2015 12:57:30 +0000 (13:57 +0100)]
orcparse: handle win32 \r\n line endings
Tim-Philipp Müller [Fri, 10 Apr 2015 11:41:10 +0000 (12:41 +0100)]
tests: use AM_TESTS_ENVIRONMENT
So test environment gets used also with the
automake test runner. The code has fallbacks
so this doesn't actually fix anything in
particular.
Luis de Bethencourt [Thu, 12 Mar 2015 13:50:33 +0000 (13:50 +0000)]
test: protect against out-of-bounds
Same as commit
8bc018cbbe25521f91ecba0fc98185d2eaac650d
CID #1146994
CID #1146995
Luis de Bethencourt [Thu, 12 Mar 2015 13:43:43 +0000 (13:43 +0000)]
test: protect against going out of bounds
Conditions should never happen but protect against going out of bounds in the
args array.
CID #1146993
Edward Hervey [Sun, 8 Mar 2015 06:51:34 +0000 (07:51 +0100)]
orcc: Remove duplicate codepath
The code is identical regardless of the variable value
CID #1226451
https://bugzilla.gnome.org/show_bug.cgi?id=745830
Edward Hervey [Fri, 27 Feb 2015 08:02:12 +0000 (09:02 +0100)]
testsuite: Put a limit to the memcpy test
Some systems have massive L3 cache that can exceed 64MB, which would
result in attempting to read/write about our allocated memory.
If we can deal with 64MB... there's a good chance we can handle above.
Vincent Penquerc'h [Fri, 30 Jan 2015 10:30:07 +0000 (10:30 +0000)]
orccodemem: fix mmap leak on error path
Not that these seem to be freed at all on non error
Sebastian Dröge [Wed, 17 Dec 2014 09:14:11 +0000 (10:14 +0100)]
Back to development
Sebastian Dröge [Wed, 17 Dec 2014 09:13:17 +0000 (10:13 +0100)]
Release 0.4.23
Wim Taymans [Wed, 10 Dec 2014 12:33:37 +0000 (13:33 +0100)]
neon: add support for all constants
Use mov and orr to construct all constants.
Wim Taymans [Tue, 9 Dec 2014 15:58:56 +0000 (16:58 +0100)]
neon: add support for 64 bit parameters
Wim Taymans [Tue, 9 Dec 2014 15:54:19 +0000 (16:54 +0100)]
neon: vtrn and vzip are undefined when src == dest
Make sure we use different src and dest registers for vtrn and vzip
because otherwise they are undefined.
Wim Taymans [Fri, 5 Dec 2014 11:01:21 +0000 (12:01 +0100)]
orcc: allow setting custom backup function
Add a new .backup keyword that instructs the orc compiler to call our
custom backup function instead of generating one. This is interesting if
the generated backup function is slower than a plain C implementation.
Jerome Laheurte [Wed, 3 Dec 2014 10:03:52 +0000 (11:03 +0100)]
Fix high-order bit of first register in VPUSH/VPOP generation
CID 1256553
CID 1256554
https://bugzilla.gnome.org/show_bug.cgi?id=727464
Jerome Laheurte [Wed, 26 Nov 2014 12:29:08 +0000 (13:29 +0100)]
Preserve NEON/VFP registers across subroutines according to ARM PCS (5.1.2.1)
https://bugzilla.gnome.org/show_bug.cgi?id=727464
Tim-Philipp Müller [Sat, 8 Nov 2014 13:23:12 +0000 (13:23 +0000)]
orcc: program-c: fix 64-bit parameter loading (loadpq) on big-endian systems
When passing 64-bit parameters through OrcExecutor, we
have to split them up into two 32-bit parameters for
backwards compatibility reasons. When generating C code,
make sure that we split up 64-bit parameters in the same
way as loadpq will read them back later. The lower 32 bits
should end up in params[ORC_VAR_D1+i] and the higher bits
should end up in params[ORC_VAR_T1+i]. The way it was done
so far, the higher/lower bits ended up swapped on big endian
systems, and then got deserialised in swapped order by loadpq.
This resulted in bogus parameters being used.
In particular, this broke the gstreamer volume element and
its unit tests on big endian systems when handling samples
in F64 format (i.e. doubles).
https://bugzilla.gnome.org/show_bug.cgi?id=739354
Wim Taymans [Fri, 26 Sep 2014 16:16:18 +0000 (18:16 +0200)]
orc: include orcprogram only once
Wim Taymans [Fri, 26 Sep 2014 14:32:25 +0000 (16:32 +0200)]
opcode: mark copy opcodes and don't emit mov for them
There is no need to emit a mov operation when the source and dest
registers of a copy operation are different, that's kindof the point of
the copy operation. Saves some useless movs whenever copyX is used.
Wim Taymans [Fri, 26 Sep 2014 13:02:21 +0000 (15:02 +0200)]
compiler: use src register as dest2
Also try to use the src register as the second destination register for
opcodes with 2 results. Saves one unnecessary mov for splitX opcodes.
Wim Taymans [Fri, 26 Sep 2014 10:11:14 +0000 (12:11 +0200)]
compiler: update max_used_temp_reg before use
It should not be smaller than min_temp_reg or we will fail to search for
a free register.
Wim Taymans [Fri, 26 Sep 2014 08:09:17 +0000 (10:09 +0200)]
sse: improve splitXX on sse3
We can implement splitXX with byte shuffles on sse3
Wim Taymans [Mon, 22 Sep 2014 08:14:04 +0000 (10:14 +0200)]
parse: cleanup
Be more elegant in parsing and passing the parameters to
orc_program_append_str_2().
Wim Taymans [Mon, 22 Sep 2014 08:11:35 +0000 (10:11 +0200)]
test: add more tests
Wim Taymans [Fri, 19 Sep 2014 16:07:10 +0000 (18:07 +0200)]
compiler: keep track of temp variables that contain param/const
Keep track of what const/param is loaded into a temp variable. We can
then see if the const is already loaded in a variable and simply reuse
it instead of using a new temp. We must also make sure we load constants
of different size in different temp variables.
Wim Taymans [Fri, 19 Sep 2014 20:57:03 +0000 (22:57 +0200)]
Revert reuse constants
This reverts commit
2ece711c615d327be1e74bec4844186ced58b0ef.
This reverts commit
2ae41b98a35405dbfd35209e7d68428be77a2e62.
This reverts commit
106dacc2620126dedb99d2b981ba2e5a7d4591f9.
This reverts commit
f6697e30681f46f6f3f632ddbbcab98107eec7e0.
These commits are more complicated than needed. A more simple solution
is possible by tracking the registers that hold a certain constant and
simply reusing that register when we can. We can then load the same
constant in different registers of different sizes if needed.
Wim Taymans [Fri, 19 Sep 2014 13:39:22 +0000 (15:39 +0200)]
parse: make contants per size
When we automatically create constants from the opcode arguments, make
different constants per size and rewrite the program to use the newly
created constant.
Wim Taymans [Fri, 19 Sep 2014 13:38:21 +0000 (15:38 +0200)]
parse: fix opcode arg size
Wim Taymans [Fri, 19 Sep 2014 10:23:44 +0000 (12:23 +0200)]
executor: use the size of the opcode arg to get multiplier
When loading a constant, select the multiplier based on the size of the
opcode argument. This is safer because we have existing code that declares
the constants of the wrong size.
Wim Taymans [Thu, 18 Sep 2014 17:32:18 +0000 (19:32 +0200)]
executor: fully expand const vars into the temp register
When loading a const in a register, always make it cover the full
register so that it containts enough values to be used in x2 and
x4 instructions.
Wim Taymans [Fri, 19 Sep 2014 08:55:52 +0000 (10:55 +0200)]
parse: register constants with the right size
When we detect a constant while parsing a line, register it with the
same size as the argument of the opcode.
Wim Taymans [Thu, 18 Sep 2014 12:19:06 +0000 (14:19 +0200)]
program-c: handle X2 and X4 with size < 2 variables
when the variable has a size < 2 it will have been declared as a simple
int8 or int16, when we use this in a X2 or X4 instruction, use the
variable directly instead of trying to use the x fields.
Wim Taymans [Thu, 18 Sep 2014 10:27:01 +0000 (12:27 +0200)]
compiler: generate only 1 temp for constants
Generate only 1 temp variable for constants. We do this by keeping track
of what variable we used to replace the constant.
Previously, orc would allocate a new temp variable every time a constant
was used, this wastes registers and is not needed.
Wim Taymans [Thu, 18 Sep 2014 08:21:48 +0000 (10:21 +0200)]
compiler: reuse _try_get_constant in _get_constant
Wim Taymans [Wed, 17 Sep 2014 14:06:35 +0000 (16:06 +0200)]
opcodes: avoid undefined operations on signed ints
Cast to unsigned int when doing addition and subtraction that can
overflow.
Cast to unsigned int when doing the lower part of a multiplication.
Cast to unsigned int when doing left shifts into the sign bit.
Regenerate the emulation code. This should avoid multiple warnings
reported with clang and -fsanitize=undefined.
See https://bugzilla.gnome.org/show_bug.cgi?id=728738
Wim Taymans [Wed, 17 Sep 2014 13:36:47 +0000 (15:36 +0200)]
compiler: do << and | on unsigned ints only
Wim Taymans [Wed, 17 Sep 2014 13:28:57 +0000 (15:28 +0200)]
compiler: avoid nagative array index
For instructions that don't have a second variable, don't try to index
the variable array with a -1 index.
Wim Taymans [Wed, 17 Sep 2014 13:19:46 +0000 (15:19 +0200)]
utils: convert to uin32 before doing |
In ORC_READ_UINT32_LE, convert each byte to uint32 before or-ing them
together. This avoids | with signed ints.
Wim Taymans [Wed, 17 Sep 2014 08:32:27 +0000 (10:32 +0200)]
docs: remove old env variable
ORC_CPU_FLAGS has been removed quite some time ago
Fabian Deutsch [Tue, 2 Sep 2014 08:38:17 +0000 (10:38 +0200)]
TMPDIR becomes the last option
Previously $TMPDIR and /tmp were the first options to try to use for
intermediate files. To prevent name collisions, now user specific dirs ($HOME
and $XDG_RUNTIME_DIR) are prefered over the user unspecific dirs ($TMP and
/tmp).
https://bugzilla.gnome.org/show_bug.cgi?id=735871
Sebastian Dröge [Wed, 27 Aug 2014 09:24:28 +0000 (12:24 +0300)]
Back to development
Sebastian Dröge [Wed, 27 Aug 2014 09:24:03 +0000 (12:24 +0300)]
Release 0.4.22
Jan Schmidt [Tue, 26 Aug 2014 05:29:50 +0000 (15:29 +1000)]
Add fallback CPU feature detection for Android
On Android, /proc/self/auxv might not be readable (except
when debuggable=true in the build, annoyingly), so do what
the android cpufeatures detection code does and fall back to
/proc/cpuinfo string matching.
Without this, release builds run really slowly, due to ORC always
doing emulation.
Jan Schmidt [Tue, 26 Aug 2014 05:29:28 +0000 (15:29 +1000)]
Remove declarations for deleted functions from orcutils.h
Jan Schmidt [Tue, 26 Aug 2014 03:10:04 +0000 (13:10 +1000)]
Detect android liblog and use android's log functions for debug output
Thiago Santos [Fri, 27 Jun 2014 16:30:03 +0000 (13:30 -0300)]
orccompiler: free error message
Edward Hervey [Fri, 27 Jun 2014 09:33:16 +0000 (11:33 +0200)]
orcarray: Use size_t for pointer arithmetics
Fixes these warnings on windows x86_64:
orcarray.c:80:33: error: cast from pointer to integer of different size [-Werror=pointer-to-int-cast]
orcarray.c:80:22: error: cast to pointer from integer of different size [-Werror=int-to-pointer-cast]
Sebastian Dröge [Mon, 23 Jun 2014 17:14:26 +0000 (19:14 +0200)]
orccompiler: If compiling a program twice, first free the results of the previous compilation
Sebastian Dröge [Mon, 23 Jun 2014 17:14:18 +0000 (19:14 +0200)]
orcc: Fix some memory leaks in the compiler
Colin Walters [Mon, 23 Jun 2014 10:59:35 +0000 (06:59 -0400)]
autogen: Honor NOCONFIGURE=1
See https://github.com/cgwalters/build-api
Sebastian Dröge [Mon, 23 Jun 2014 08:14:47 +0000 (10:14 +0200)]
Back to development
Sebastian Dröge [Mon, 23 Jun 2014 08:06:29 +0000 (10:06 +0200)]
Release 0.4.21
Sebastian Dröge [Mon, 23 Jun 2014 07:59:43 +0000 (09:59 +0200)]
configure: Pass -version-info to libtool again
Sebastian Dröge [Mon, 23 Jun 2014 07:39:26 +0000 (09:39 +0200)]
Back to development
Sebastian Dröge [Mon, 23 Jun 2014 07:38:51 +0000 (09:38 +0200)]
Release 0.4.20
Sebastian Dröge [Mon, 23 Jun 2014 07:20:03 +0000 (09:20 +0200)]
build: Modernize autotools build system
Vincent Penquerc'h [Thu, 1 May 2014 14:47:41 +0000 (15:47 +0100)]
Add some checks on the number of variables per type
We want to ensure no more than, say, 8 constants are added
to a program. Adding more will violate pervasive assumptions
in the code, and may lead to various buffer overflows. By
trapping these add creation time, we prevent these issues
without cluttering the code with range checks. The user is
assumed non malicious here.
Add a test to check we can add up to and including the limit
for a type, but no more.
Tim-Philipp Müller [Thu, 5 Jun 2014 14:45:57 +0000 (15:45 +0100)]
Update .gitignore
Tim-Philipp Müller [Wed, 4 Jun 2014 22:25:52 +0000 (23:25 +0100)]
orccodemem: fix list corruption when splitting memory chunks in two
When splitting orc memory chunks in two, the 'prev' pointer of the
new memory chunk to be inserted was wrong, causing invalid memory
access and double frees later down the line.
https://bugzilla.gnome.org/show_bug.cgi?id=731227
Sebastian Dröge [Tue, 20 May 2014 08:07:57 +0000 (10:07 +0200)]
orc: Close file after usage in the copied around file reading function
CID 1147011.
Sebastian Dröge [Mon, 19 May 2014 09:09:55 +0000 (11:09 +0200)]
orcarray: Use fabs() instead of abs(), the latter is for integers
orcarray.c:254:15: error: using integer absolute value function 'abs' when
argument is of floating point type [-Werror,-Wabsolute-value]
if (abs(a[i] - b[i]) < MIN_NONDENORMAL_D) continue;
Sebastian Dröge [Fri, 25 Apr 2014 18:42:02 +0000 (20:42 +0200)]
orcprogram-c: Convert shifts in swap opcodes to operations on unsigned integers
Shifting into the sign bit is undefined, so let's stop doing that.
Sebastian Dröge [Tue, 22 Apr 2014 16:05:17 +0000 (18:05 +0200)]
memcpy_speed: Use unsigned integers, overflows are undefined for signed integers
Sebastian Dröge [Fri, 18 Apr 2014 08:49:59 +0000 (10:49 +0200)]
Back to development
Sebastian Dröge [Fri, 18 Apr 2014 08:32:23 +0000 (10:32 +0200)]
Release 0.4.19
Vincent Penquerc'h [Thu, 17 Apr 2014 08:16:46 +0000 (09:16 +0100)]
orccodemem: mask group/other bits for temporary file
Ensures another user can't inject code into your process when
winning a race with the ORC temp file code.
Coverity 1147013
Vincent Penquerc'h [Wed, 16 Apr 2014 11:07:40 +0000 (12:07 +0100)]
orcprogam-{mips,mmx,sse}: bail out if get_align_var fails
Coverity 1146979, 1146980, 1146977, 1146978, 1146973
Vincent Penquerc'h [Wed, 16 Apr 2014 11:02:13 +0000 (12:02 +0100)]
orcprogram: bail out on error
Coverity 1146972
Sebastian Dröge [Tue, 15 Apr 2014 08:24:56 +0000 (10:24 +0200)]
orcbytecode: Don't replace program name without freeing the old name if any
Sebastian Dröge [Tue, 15 Apr 2014 08:17:04 +0000 (10:17 +0200)]
orcc: Fix indention of generated code
Sebastian Dröge [Mon, 14 Apr 2014 22:20:49 +0000 (00:20 +0200)]
orc: Update autogenerated files
Sebastian Dröge [Mon, 14 Apr 2014 21:12:57 +0000 (23:12 +0200)]
orcprogram-c: Fix swaplq C implementation
Sebastian Dröge [Mon, 14 Apr 2014 20:39:15 +0000 (22:39 +0200)]
orcbytecodes: Fix filename of generated file
We don't want to override orcbytecode.h but orcbytecodes.h
Sebastian Dröge [Mon, 14 Apr 2014 19:53:40 +0000 (21:53 +0200)]
orcarray: Fix compiler warning with gcc 4.9
In file included from ../orc/orcprogram.h:5:0,
from ../orc/orc.h:5,
from ../orc-test/orctest.h:5,
from orcarray.c:6:
../orc/orcutils.h:124:37: error: 'data' may be used uninitialized in this function [-Werror=maybe-uninitialized]
#define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset)))
^
orcarray.c:42:9: note: 'data' was declared here
void *data;
Sebastian Dröge [Mon, 14 Apr 2014 16:54:57 +0000 (18:54 +0200)]
memcpy_speed: Properly define ALIGN macro and fix memory leaks
With the previous definition we would use memory *before* our
allocated memory area.
Sebastian Dröge [Sun, 13 Apr 2014 20:01:38 +0000 (22:01 +0200)]
bench10: Fix syntax errors
Sebastian Dröge [Sun, 13 Apr 2014 19:58:16 +0000 (21:58 +0200)]
test.orc: Fix syntax errors
Sebastian Dröge [Sun, 13 Apr 2014 19:57:59 +0000 (21:57 +0200)]
benchmorc: Fix memory leaks in test
Sebastian Dröge [Sun, 13 Apr 2014 19:56:01 +0000 (21:56 +0200)]
orcparse: Guard against unknown tokens
Sebastian Dröge [Sun, 13 Apr 2014 19:55:52 +0000 (21:55 +0200)]
orcprogram: Free init function name
Sebastian Dröge [Sun, 13 Apr 2014 19:38:48 +0000 (21:38 +0200)]
tests: Fix various memory leaks in tests
Sebastian Dröge [Sun, 13 Apr 2014 19:36:43 +0000 (21:36 +0200)]
orcprogram: Free OrcCode together with the program
Sebastian Dröge [Sun, 13 Apr 2014 19:22:37 +0000 (21:22 +0200)]
orcprogram: Free type names of variables when freeing a program