Add lzma_outq_clear_cache2() which may leave one buffer allocated
in the cache.
Add lzma_outq_outbuf_memusage() to get the memory needed for
a single lzma_outbuf. This is now used internally in outqueue.c too.
Track both the total amount of memory allocated and the amount of
memory that is in active use (not in cache).
In lzma_outbuf, allow storing the current input position that
matches the current output position. This way the main thread
can notice when no more output is possible without first providing
more input.
Allow specifying return code for lzma_outq_read() in a finished
lzma_outbuf.
If lzma_index_append() failed (most likely memory allocation failure)
it could have gone unnoticed and the resulting .xz file would have
an incorrect Index. Decompressing such a file would produce the
correct uncompressed data but then an error would occur when
verifying the Index field.
Now it limits the input and output buffer sizes that are
passed to a raw decoder. This way there's no need to check
if the sizes can grow too big or overflow when updating
Compressed Size and Uncompressed Size counts. This also means
that a corrupt file cannot cause the raw decoder to process
useless extra input or output that would exceed the size info
in Block Header (and thus cause LZMA_DATA_ERROR anyway).
More importantly, now the size information is verified more
carefully in case raw decoder returns LZMA_OK. This doesn't
really matter with the current single-threaded .xz decoder
as the errors would be detected slightly later anyway. But
this helps avoiding corner cases in the upcoming threaded
decompressor, and it might help other Block decoder uses
outside liblzma too.
The test files bad-1-lzma2-{9,10,11}.xz test these conditions.
With the single-threaded .xz decoder the only difference is
that LZMA_DATA_ERROR is detected in a difference place now.
Previously lzma_lzma_props_encode() and lzma_lzma2_props_encode()
assumed that the options pointers must be non-NULL because the
with these filters the API says it must never be NULL. It is
good to do these checks anyway.
This broke 32-bit builds due to a pointer type mismatch.
This bug was introduced with the output-size-limited encoding
in 625f4c7c99.
Thanks to huangqinjin for the bug report.
OpenBSD does not allow to change the group of a file if the user
does not belong to this group. In contrast to Linux, OpenBSD also
fails if the new group is the same as the old one. Do not call
fchown(2) in this case, it would change nothing anyway.
This fixes an issue with Perl Alien::Build module.
https://github.com/PerlAlien/Alien-Build/issues/62
Sometimes the version number from "less -V" contains a dot,
sometimes not. xzless failed detect the version number when
it does contain a dot. This fixes it.
Thanks to nick87720z for reporting this. Apparently it had been
reported here <https://bugs.gentoo.org/489362> in 2013.
Due to architectural limitations, address space available to a single
userspace process on MIPS32 is limited to 2 GiB, not 4, even on systems
that have more physical RAM -- e.g. 64-bit systems with 32-bit
userspace, or systems that use XPA (an extension similar to x86's PAE).
So, for MIPS32, we have to impose stronger memory limits. I've chosen
2000MiB to give the process some headroom.
When the uncompressed size is known to be exact, after decompressing
the stream exactly comp_size bytes of input must have been consumed.
This is a minor improvement to error detection.
The caller must still not specify an uncompressed size bigger
than the actual uncompressed size.
As a downside, this now needs the exact compressed size.
Right now this is just a planned extra-compact format for use
in the EROFS file system in Linux. At this point it's possible
that the format will either change or be abandoned and removed
completely.
The special thing about the encoder is that it uses the
output-size-limited encoding added in the previous commit.
EROFS uses fixed-sized blocks (e.g. 4 KiB) to hold compressed
data so the compressors must be able to create valid streams
that fill the given block size.
With this it is possible to encode LZMA1 data without EOPM so that
the encoder will encode as much input as it can without exceeding
the specified output size limit. The resulting LZMA1 stream will
be a normal LZMA1 stream without EOPM. The actual uncompressed size
will be available to the caller via the uncomp_size pointer.
One missing thing is that the LZMA layer doesn't inform the LZ layer
when the encoding is finished and thus the LZ may read more input
when it won't be used. However, this doesn't matter if encoding is
done with a single call (which is the planned use case for now).
For proper multi-call encoding this should be improved.
This commit only adds the functionality for internal use.
Nothing uses it yet.
Previously this required using --force but that has other
effects too which might be undesirable. Changing the behavior
of --keep has a small risk of breaking existing scripts but
since this is a fairly special corner case I expect the
likehood of breakage to be low enough.
I think the new behavior is more logical. The only reason for
the old behavior was to be consistent with gzip and bzip2.
Thanks to Vincent Lefevre and Sebastian Andrzej Siewior.
Omit the -q option from xz, gzip, and bzip2. With xz this shouldn't
matter. With gzip it's important because -q makes gzip replace SIGPIPE
with exit status 2. With bzip2 it's important because with -q bzip2
is completely silent if input is corrupt while other decompressors
still give an error message.
Avoiding exit status 2 from gzip is important because bzip2 uses
exit status 2 to indicate corrupt input. Before this commit xzgrep
didn't recognize corrupt .bz2 files because xzgrep was treating
exit status 2 as SIGPIPE for gzip compatibility.
zstd still needs -q because otherwise it is noisy in normal
operation.
The code to detect real SIGPIPE didn't check if the exit status
was due to a signal (>= 128) and so could ignore some other exit
status too.
This is a minor fix since this affects only the situation when
the files differ and the exit status is something else than 0.
In such case there could be SIGPIPE from a decompression tool
and that would result in exit status of 2 from xzdiff/xzcmp
while the correct behavior would be to return 1 or whatever
else diff or cmp may have returned.
This commit omits the -q option from xz/gzip/bzip2/lzop arguments.
I'm not sure why the -q was used in the first place, perhaps it
hides warnings in some situation that I cannot see at the moment.
Hopefully the removal won't introduce a new bug.
With gzip the -q option was harmful because it made gzip return 2
instead of >= 128 with SIGPIPE. Ignoring exit status 2 (warning
from gzip) isn't practical because bzip2 uses exit status 2 to
indicate corrupt input file. It's better if SIGPIPE results in
exit status >= 128.
With bzip2 the removal of -q seems to be good because with -q
it prints nothing if input is corrupt. The other tools aren't
silent in this situation even with -q. On the other hand, if
zstd support is added, it will need -q since otherwise it's
noisy in normal situations.
Thanks to Étienne Mollier and Sebastian Andrzej Siewior.
Before this commit all output queue buffers were allocated as
a single big allocation. Now each buffer is allocated separately
when needed. Used buffers are cached to avoid reallocation
overhead but the cache will keep only one buffer size at a time.
This should make things work OK in the decompression where most
of the time the buffer sizes will be the same but with some less
common files the buffer sizes may vary.
While this should work fine, it's still a bit preliminary
and may even get reverted if it turns out to be useless for
decompression.
When Intel CET is enabled, we need to include <cet.h> in assembly codes
to mark Intel CET support and add _CET_ENDBR to indirect jump targets.
Tested on Intel Tiger Lake under CET enabled Linux.
I don't want to use \c in macro arguments but groff_man(7)
suggests that \f has better portability. \f would be needed
for the .TP strings for portability reasons anyway.
Thanks to Bjarni Ingi Gislason.
A few are simply omitted, most are converted to "for example"
and surrounded with commas. Sounds like that this is better
style, for example, man-pages(7) recommends avoiding such
abbreviations except in parenthesis.
Thanks to Bjarni Ingi Gislason.
Docs of ancient troff/nroff mention \(em (em-dash) but not \(en
and \- was used for both minus and en-dash. I don't know how
portable \(en is nowadays but it can be changed back if someone
complains. At least GNU groff and OpenBSD's mandoc support it.
Thanks to Bjarni Ingi Gislason for the patch.
Output is from: test-groff -b -e -mandoc -T utf8 -rF0 -t -w w -z
[ "test-groff" is a developmental version of "groff" ]
Input file is ./src/scripts/xzgrep.1
<src/scripts/xzgrep.1>:20 (macro RB): only 1 argument, but more are expected
<src/scripts/xzgrep.1>:23 (macro RB): only 1 argument, but more are expected
<src/scripts/xzgrep.1>:26 (macro RB): only 1 argument, but more are expected
<src/scripts/xzgrep.1>:29 (macro RB): only 1 argument, but more are expected
<src/scripts/xzgrep.1>:32 (macro RB): only 1 argument, but more are expected
"abc..." does not mean the same as "abc ...".
The output from nroff and troff is unchanged except for the space
between "file" and "...".
Signed-off-by: Bjarni Ingi Gislason <bjarniig@rhi.hi.is>
Summary:
mandoc -T lint xzgrep.1 :
mandoc: xzgrep.1:79:2: WARNING: skipping paragraph macro: PP empty
There is no change in the output of "nroff" and "troff".
Signed-off-by: Bjarni Ingi Gislason <bjarniig@rhi.hi.is>
Output is from: test-groff -b -e -mandoc -T utf8 -rF0 -t -w w -z
[ "test-groff" is a developmental version of "groff" ]
Input file is ./src/xz/xz.1
<src/xz/xz.1>:408 (macro BR): only 1 argument, but more are expected
<src/xz/xz.1>:1009 (macro BR): only 1 argument, but more are expected
<src/xz/xz.1>:1743 (macro BR): only 1 argument, but more are expected
<src/xz/xz.1>:1920 (macro BR): only 1 argument, but more are expected
<src/xz/xz.1>:2213 (macro BR): only 1 argument, but more are expected
Output from nroff and troff is unchanged, except for a font change of a
full stop (.).
Signed-off-by: Bjarni Ingi Gislason <bjarniig@rhi.hi.is>
DJGPP 2.05 added support for thousands separators but it's
broken at least under WinXP with Finnish locale that uses
a non-breaking space as the thousands separator. Workaround
by disabling thousands separators for DJGPP builds.
The comment didn't match the value of RC_SYMBOLS_MAX and the value
itself was slightly larger than actually needed. The only harm
about this was that memory usage was a few bytes larger.
strerror() needs <string.h> which happened to be included via
tuklib_common.h -> tuklib_config.h -> sysdefs.h if HAVE_CONFIG_H
was defined. This wasn't tested without config.h before so it
had worked fine.
The previous commit broke crc32_tablegen.c.
If the whole package is built without config.h (with defines
set on the compiler command line) this should still work fine
as long as these headers conform to C99 well enough.
string.h is used unconditionally elsewhere in the project and
configure has always stopped if limits.h is missing, so these
headers must have been always available even on the weirdest
systems.
The dependency on po4a is optional. It's never required to install
the translated man pages when xz is built from a release tarball.
If po4a is missing when building from xz.git, the translated man
pages won't be generated but otherwise the build will work normally.
The translations are only updated automatically by autogen.sh and
by "make mydist". This makes it easy to keep po4a as an optional
dependency and ensures that I won't forget to put updated
translations to a release tarball.
The translated man pages aren't installed if --disable-nls is used.
The installation of translated man pages abuses Automake internals
by calling "install-man" with redefined dist_man_MANS and man_MANS.
This makes the hairy script code slightly less hairy. If it breaks
some day, this code needs to be fixed; don't blame Automake developers.
Also, this adds more quotes to the existing shell script code in
the Makefile.am "-hook"s.
Perhaps it's too drastic but on the other hand it will let me
learn about possible problems if people report the errors.
This won't be backported to the v5.2 branch.
See the code comment for reasoning. It's far from perfect but
hopefully good enough for certain cases while hopefully doing
nothing bad in other situations.
At presets -5 ... -9, 4020 MiB vs. 4096 MiB makes no difference
on how xz scales down the number of threads.
The limit has to be a few MiB below 4096 MiB because otherwise
things like "xz --lzma2=dict=500MiB" won't scale down the dict
size enough and xz cannot allocate enough memory. With
"ulimit -v $((4096 * 1024))" on x86-64, the limit in xz had
to be no more than 4085 MiB. Some safety margin is good though.
This is hack but it should be useful when running 32-bit xz on
a 64-bit kernel that gives full 4 GiB address space to xz.
Hopefully this is enough to solve this:
https://bugzilla.redhat.com/show_bug.cgi?id=1196786
FreeBSD has a patch that limits the result in tuklib_physmem()
to SIZE_MAX on 32-bit systems. While I think it's not the way
to do it, the results on --memlimit-compress have been good. This
commit should achieve practically identical results for compression
while leaving decompression and tuklib_physmem() and thus
lzma_physmem() unaffected.
xz --flush-timeout=2000, old version:
1. xz is started. The next flush will happen after two seconds.
2. No input for one second.
3. A burst of a few kilobytes of input.
4. No input for one second.
5. Two seconds have passed and flushing starts.
The first second counted towards the flush-timeout even though
there was no pending data. This can cause flushing to occur more
often than needed.
xz --flush-timeout=2000, after this commit:
1. xz is started.
2. No input for one second.
3. A burst of a few kilobytes of input. The next flush will
happen after two seconds counted from the time when the
first bytes of the burst were read.
4. No input for one second.
5. No input for another second.
6. Two seconds have passed and flushing starts.
The same code sequence repeats so it's nicer as a separate function.
Note that in one case there was no test for opt_mode != MODE_TEST,
but that was only because that condition would always be true, so
this commit doesn't change the behavior there.
When input blocked, xz --flush-timeout=1 would wake up every
millisecond and initiate flushing which would have nothing to
flush and thus would just waste CPU time. The fix disables the
timeout when no input has been seen since the previous flush.
Using the aligned methods requires more care to ensure that
the address really is aligned, so it's nicer if the aligned
methods are prefixed. The next commit will remove the unaligned_
prefix from the unaligned methods which in liblzma are used in
more places than the aligned ones.
Add a configure option --enable-unsafe-type-punning to get the
old non-conforming memory access methods. It can be useful with
old compilers or in some other less typical situations but
shouldn't normally be used.
Omit the packed struct trick for unaligned access. While it's
best in some cases, this is simpler. If the memcpy trick doesn't
work, one can request unsafe type punning from configure.
Because CRC32/CRC64 code needs fast aligned reads, if no very
safe way to do it is found, type punning is used as a fallback.
This sucks but since it currently works in practice, it seems to
be the least bad option. It's never needed with GCC >= 4.7 or
Clang >= 3.6 since these support __builtin_assume_aligned and
thus fast aligned access can be done with the memcpy trick.
Other things:
- Support GCC/Clang __builtin_bswapXX
- Cleaner bswap fallback macros
- Minor cleanups
This adds a configure option --enable-path-for-scripts=PREFIX
which defaults to empty except on Solaris it is /usr/xpg4/bin
to make POSIX grep and others available. The Solaris case had
been documented in INSTALL with a manual fix but it's better
to do this automatically since it is needed on most Solaris
systems anyway.
Thanks to Daniel Richard G.
LZMA_TIMED_OUT is *internally* used as a value for lzma_ret
enumeration. Previously it was #defined to 32 and cast to lzma_ret.
That way it wasn't visible in the public API, but this was hackish.
Now the public API has eight LZMA_RET_INTERNALx members and
LZMA_TIMED_OUT is #defined to LZMA_RET_INTERNAL1. This way
the code is cleaner overall although the public API has a few
extra mysterious enum members.
Or any off_t which isn't very big (like signed 64 bit integer
that most system have). A small off_t could overflow if the
file being decompressed had long enough run of zero bytes,
which would result in corrupt output.
Now memcpy() or GNU C packed structs for unaligned access instead
of type punning. See the comment in this commit for details.
Avoiding type punning with unaligned access is needed to
silence gcc -fsanitize=undefined.
New functions: unaliged_readXXne and unaligned_writeXXne where
XX is 16, 32, or 64.
I should have always known this but I didn't. Here is an example
as a reminder to myself:
int mycopy(void *dest, void *src, size_t n)
{
memcpy(dest, src, n);
return dest == NULL;
}
In the example, a compiler may assume that dest != NULL because
passing NULL to memcpy() would be undefined behavior. Testing
with GCC 8.2.1, mycopy(NULL, NULL, 0) returns 1 with -O0 and -O1.
With -O2 the return value is 0 because the compiler infers that
dest cannot be NULL because it was already used with memcpy()
and thus the test for NULL gets optimized out.
In liblzma, if a null-pointer was passed to memcpy(), there were
no checks for NULL *after* the memcpy() call, so I cautiously
suspect that it shouldn't have caused bad behavior in practice,
but it's hard to be sure, and the problematic cases had to be
fixed anyway.
Thanks to Jeffrey Walton.
Now the widths of the check names is used to adjust the width
of the Check column. This way there no longer is a need to restrict
the widths of the check names to be at most ten terminal-columns.
"xz -dcfv not_an_xz_file" crashed (all four options are
required to trigger it). It caused xz to call
lzma_get_progress(&strm, ...) when no coder was initialized
in strm. In this situation strm.internal is NULL which leads
to a crash in lzma_get_progress().
The bug was introduced when xz started using lzma_get_progress()
to get progress info for multi-threaded compression, so the
bug is present in versions 5.1.3alpha and higher.
Thanks to Filip Palian <Filip.Palian@pjwstk.edu.pl> for
the bug report.
FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION is #defined when liblzma
is being built for fuzz testing.
Most fuzzed inputs would normally get rejected because of incorrect
CRC32 and the actual header decoding code wouldn't get fuzzed.
Disabling CRC32 checks avoids this problem. The fuzzer program
must still use LZMA_IGNORE_CHECK flag to disable verification of
integrity checks of uncompressed data.
It ended up printing an uninitialized char-array when trying to
print the check names (column 7) on the "totals" line.
This also changes the column 12 (minimum xz version) to
50000002 (xz 5.0.0) instead of 0 when there are no valid
input files.
Thanks to kidmin for the bug report.
The 0 got treated specially in a buggy way and as a result
the function did nothing. The API doc said that 0 was supposed
to return LZMA_PROG_ERROR but it didn't.
Now 0 is treated as if 1 had been specified. This is done because
0 is already used to indicate an error from lzma_memlimit_get()
and lzma_memusage().
In addition, lzma_memlimit_set() no longer checks that the new
limit is at least LZMA_MEMUSAGE_BASE. It's counter-productive
for the Index decoder and was actually needed only by the
auto decoder. Auto decoder has now been modified to check for
LZMA_MEMUSAGE_BASE.
It returned LZMA_PROG_ERROR, which was done to avoid zero as
the limit (because it's a special value elsewhere), but using
LZMA_PROG_ERROR is simply inconvenient and can cause bugs.
The fix/workaround is to treat 0 as if it were 1 byte. It's
effectively the same thing. The only weird consequence is
that then lzma_memlimit_get() will return 1 even when 0 was
specified as the limit.
This fixes a very rare corner case in xz --list where a specific
memory usage limit and a multi-stream file could print the
error message "Internal error (bug)" instead of saying that
the memory usage limit is too low.
Only one definition was visible in a translation unit.
It avoided a few casts and temp variables but seems that
this hack doesn't work with link-time optimizations in compilers
as it's not C99/C11 compliant.
Fixes:
http://www.mail-archive.com/xz-devel@tukaani.org/msg00279.html
It's available in glibc (GNU/Linux, GNU/kFreeBSD). It's better
than sysconf(_SC_NPROCESSORS_ONLN) because sched_getaffinity()
gives the number of cores available to the process instead of
the total number of cores online.
As a side effect, this commit fixes a bug on GNU/kFreeBSD where
configure would detect the FreeBSD-specific cpuset_getaffinity()
but it wouldn't actually work because on GNU/kFreeBSD it requires
using -lfreebsd-glue when linking. Now the glibc-specific function
will be used instead.
Thanks to Sebastian Andrzej Siewior for the original patch
and testing.
This is the sane thing to do. The conflict with OpenSSL
on some OSes and especially that the OS-provided versions
can be significantly slower makes it clear that it was
a mistake to have the external SHA-256 support enabled by
default.
Those who want it can now pass --enable-external-sha256 to
configure. INSTALL was updated with notes about OSes where
this can be a bad idea.
The SHA-256 detection code in configure.ac had some bugs that
could lead to a build failure in some situations. These were
fixed, although it doesn't matter that much now that the
external SHA-256 is disabled by default.
MINIX >= 3.2.0 uses NetBSD's libc and thus has SHA256_Init
in libc instead of libutil. Support for the libutil version
was removed.
When optimizing, GCC can reorder code so that an uninitialized
value gets used in a comparison, which makes Valgrind unhappy.
It doesn't happen when compiled with -O0, which I tend to use
when running Valgrind.
Thanks to Rich Prohaska. I remember this being mentioned long
ago by someone else but nothing was done back then.
The patch is quite long but it's mostly about adding new #ifdefs
to omit code when encoders or decoders have been disabled.
This adds two new #defines to config.h: HAVE_ENCODERS and
HAVE_DECODERS.
People shouldn't rely on the presets when decoding raw streams,
but xz uses the presets as the starting point for raw decoder
options anyway.
lzma_encocder_presets.c was renamed to lzma_presets.c to
make it clear it's not used solely by the encoder code.
lzma_index_dup() calls index_dup_stream() which, in case of
an error, calls index_stream_end() to free memory allocated
by index_stream_init(). However, it illogically didn't
actually free the memory. To make it logical, the tree
handling code was modified a bit in addition to changing
index_stream_end().
Thanks to Evan Nemerson for the bug report.
This reverts commit 7a11c4a8e5.
It is a problem when libc has pipe2() but the kernel is too
old to have pipe2() and thus pipe2() fails. In xz it's pointless
to have a fallback for non-functioning pipe2(); it's better to
avoid pipe2() completely.
Thanks to Michael Fox for the bug report.
The sandboxing is used conditionally as described in main.c.
This isn't optimal but it was much easier to implement than
a full sandboxing solution and it still covers the most common
use cases where xz is writing to standard output. This should
have practically no effect on performance even with small files
as fork() isn't needed.
C and locale libraries can open files as needed. This has been
fine in the past, but it's a problem with things like Capsicum.
io_sandbox_enter() tries to ensure that various locale-related
files have been loaded before cap_enter() is called, but it's
possible that there are other similar problems which haven't
been seen yet.
Currently Capsicum is available on FreeBSD 10 and later
and there is a port to Linux too.
Thanks to Loganaden Velvindron for help.
The earlier version compiled but didn't actually work
since sysconf(_SC_PHYS_PAGES) always fails (or so I was told).
Thanks to Ole André Vadla Ravnås for the patch and testing.
In FreeBSD, cpuset_getaffinity() is the preferred way to get
the number of available cores.
Thanks to Rui Paulo for the patch. I edited it slightly, but
hopefully I didn't break anything.
It's a problem at least on OpenBSD which doesn't support
O_NONBLOCK on e.g. /dev/null. I'm not surprised if it's
a problem on other OSes too since this behavior is allowed
in POSIX-1.2008.
The code relying on this behavior was committed in June 2013
and included in 5.1.3alpha released on 2013-10-26. Clearly
the development releases only get limited testing.
I know that soname != app version, but I skip AGE=1
in -version-info to make the soname match the liblzma
version anyway. It doesn't hurt anything as long as
it doesn't conflict with library versioning rules.
This way an invalid filter chain is detected at the Stream
encoder initialization instead of delaying it to the first
call to lzma_code() which triggers the initialization of
the actual filter encoder(s).
This avoids the possibility of "File name too long" when
creating a temp file when the input file name is very long.
This also means that other users on the system can no longer
see the input file names in /tmp (or whatever $TMPDIR is)
since the temporary directory will have a generic name. This
usually doesn't matter since on many systems one can see
the arguments given to all processes anyway.
The number X chars to mktemp where increased from 6 to 10.
Note that with some shells temp files or dirs won't be used at all.
The behavior of grep -ql varies:
- GNU grep behaves like grep -q.
- OpenBSD grep behaves like grep -l.
POSIX doesn't make it 100 % clear what behavior is expected.
Anyway, using both -q and -l at the same time makes no sense
so both options simply should never be used at the same time.
Thanks to Christian Weisgerber.
POSIX supports $< only in inference rules (suffix rules).
Using it elsewhere is a GNU make extension and doesn't
work e.g. with OpenBSD make.
Thanks to Christian Weisgerber for the patch.
Note that this slightly changes how lzma_block_header_decode()
has been documented. Earlier it said that the .version is set
to the lowest required value, but now it says that the .version
field is kept unchanged if possible. In practice this doesn't
affect any old code, because before this commit the only
possible .version was 0.
The Maj macro is used where multiple things are added
together, so making Maj a sum of two expressions allows
some extra freedom for the compiler to schedule the
instructions.
I learned this trick from
<http://www.hackersdelight.org/corres.txt>.
This looks weird because the rotations become sequential,
but it helps quite a bit on both 32-bit and 64-bit x86:
- It requires fewer instructions on two-operand
instruction sets like x86.
- It requires one register less which matters especially
on 32-bit x86.
I hope this doesn't hurt other archs.
I didn't invent this idea myself, but I don't remember where
I saw it first.
This way a branch isn't needed for each operation
to choose between blk0 and blk2, and still the code
doesn't grow as much as it would with full unrolling.
Lasse Collin
Ed Maste
jiat75
Ville Skyttä
Alexander Bluhm
Ivan A. Melnikov
H.J. Lu
Adam Borowski
Bjarni Ingi Gislason
Antoine Cœur
Pavel Raiskup
Ben Boeckel