It's used only for basic bittrees and fixed-size reverse bittree
because those showed a clear benefit on x86-64 with GCC and Clang.
The other methods were more mixed and thus are commented out but
they should be tested on other archs.
Now extra buffer space is reserved so that repeating bytes for
any single match will never need to copy from two places (both
the beginning and the end of the buffer). This simplifies
dict_repeat() and helps a little with speed.
This seems to reduce .lzma decompression time about 2 %, so
with .xz and CRC it could be slightly less. The small things
add up still.
It's not completely obvious if this is better in the decoder.
It should be good if compiler can avoid creating a branch
(like using CMOV on x86).
This also makes lzma_encoder.c use the new macros.
The new decoder resumes the first decoder loop in the Resumable mode.
Then, the code executes in Non-resumable mode until it detects that it
cannot guarantee to have enough input/output to decode another symbol.
The Resumable mode is how the decoder has always worked. Before decoding
every input bit, it checks if there is enough space and will save its
location to be resumed later. When the decoder has more input/output,
it jumps back to the correct sequence in the Resumable mode code.
When the input/output buffers are large, the Resumable mode is much
slower than the Non-resumable because it has more branches and is harder
for the compiler to optimize since it is in a large switch block.
Early benchmarking shows significant time improvement (8-10% on gcc and
clang x86) by using the Non-resumable code as much as possible.
The new "safe" range decoder mode is the same as old range decoder, but
now the default behavior of the range decoder will not check if there is
enough input or output to complete the operation. When the buffers are
close to fully consumed, the "safe" operations must be used instead. This
will improve speed because it will reduce the number of branches needed
for most of the range decoder operations.
Perhaps the generated files aren't even copyrightable but
using the same license for them as for the rest of the liblzma
keeps things more consistent for tools that look for license info.
The initial commit 5d018dc035
in 2007 had a comment in sha256.c that the code is based on
Crypto++ Library 5.5.1. In 2009 the Authors list in sha256.c
and the AUTHORS file was updated with information that the
code had come from Crypto++ but via 7-Zip. I know I had viewed
7-Zip's SHA-256 code but back then the C code has been identical
enough with Crypto++, so I don't why I thought the author info
would need that extra step via 7-Zip for this single file.
Another error is that I had mixed sha.* and shacal2.* files
when checking for author info in Crypto++. The shacal2.* files
aren't related to liblzma's sha256.c and thus Kevin Springle's
code in Crypto++ isn't either.
If liblzma is configured with --disable-clmul-crc
CFLAGS="-msse4.1 -mpclmul", then it will fail to compile because the
generic version must be used but the CRC tables were not included.
The code was using HAVE_FUNC_ATTRIBUTE_IFUNC instead of CRC_USE_IFUNC.
With ARM64, ifunc is incompatible because it requires non-inline
function calls for runtime detection.
Even though the proper name for the architecture is aarch64, this
project uses ARM64 throughout. So the rename is for consistency.
Additionally, crc32_arm64.h was slightly refactored for the following
changes:
* Added MSVC, FreeBSD, and macOS support in
is_arch_extension_supported().
* crc32_arch_optimized() now checks the size when aligning the
buffer.
* crc32_arch_optimized() loop conditions were slightly modified to
avoid both decrementing the size and incrementing the buffer
pointer.
* Use the intrinsic wrappers defined in <arm_acle.h> because GCC and
Clang name them differently.
* Minor spacing and comment changes.
The CRC_GENERIC is now split into CRC32_GENERIC and CRC64_GENERIC, since
the ARM64 optimizations will be different between CRC32 and CRC64.
For the same reason, CRC_ARCH_OPTIMIZED is split into
CRC32_ARCH_OPTIMIZED and CRC64_ARCH_OPTIMIZED.
ifunc will only be used with x86-64 CLMUL because the runtime detection
methods needed with ARM64 are not compatible with ifunc.
The CRC32 instructions in ARM64 can calculate the CRC32 result
for 8 bytes in a single operation, making the use of ARM64
instructions much faster compared to the general CRC32 algorithm.
Optimized CRC32 will be enabled if ARM64 has CRC extension
running on Linux.
Signed-off-by: Chenxi Mao <chenxi.mao2013@gmail.com>
Not all RISC-V processors support fast unaligned access so
it's better to read only one byte in the main loop. This can
be faster even on x86-64 when compared to reading 32 bits at
a time as half the time the address is only 16-bit aligned.
The downside is larger code size on archs that do support
fast unaligned access.
The new Filter ID is 0x0B.
Thanks to Chien Wong <m@xv97.com> for the initial version of the Filter,
the xz CLI updates, and the Autotools build system modifications.
Thanks to Igor Pavlov for his many contributions to the design of
the filter.
Now crc_simd_body() in crc_x86_clmul.h is only called once
in a translation unit, we no longer need to be so cautious
about ensuring the always-inline behavior.
CRC_CLMUL was split to CRC_ARCH_OPTIMIZED and CRC_X86_CLMUL.
CRC_ARCH_OPTIMIZED is defined when an arch-optimized version is used.
Currently the x86 CLMUL implementations are the only arch-optimized
versions, and these also use the CRC_x86_CLMUL macro to tell when
crc_x86_clmul.h needs to be included.
is_clmul_supported() was renamed to is_arch_extension_supported().
crc32_clmul() and crc64_clmul() were renamed to
crc32_arch_optimized() and crc64_arch_optimized().
This way the names make sense with arch-specific non-CLMUL
implementations as well.
A CLMUL-only build will have the crcxx_clmul() inlined into
lzma_crcxx(). Previously a jump to the extern lzma_crcxx_clmul()
was needed. Notes about shared liblzma on ELF platforms:
- On platforms that support ifunc and -fvisibility=hidden, this
was silly because CLMUL-only build would have that single extra
jump instruction of extra overhead.
- On platforms that support neither -fvisibility=hidden nor linker
version script (liblzma*.map), jumping to lzma_crcxx_clmul()
would go via PLT so a few more instructions of overhead (still
not a big issue but silly nevertheless).
There was a downside with static liblzma too: if an application only
needs lzma_crc64(), static linking would make the linker include the
CLMUL code for both CRC32 and CRC64 from crc_x86_clmul.o even though
the CRC32 code wouldn't be needed, thus increasing code size of the
executable (assuming that -ffunction-sections isn't used).
Also, now compilers are likely to inline crc_simd_body()
even if they don't support the always_inline attribute
(or MSVC's __forceinline). Quite possibly all compilers
that build the code do support such an attribute. But now
it likely isn't a problem even if the attribute wasn't supported.
Now all x86-specific stuff is in crc_x86_clmul.h. If other archs
The other archs can then have their own headers with their own
is_clmul_supported() and crcxx_clmul().
Another bonus is that the build system doesn't need to care if
crc_clmul.c is needed.
is_clmul_supported() stays as inline function as it's not needed
when doing a CLMUL-only build (avoids a warning about unused function).
It requires fast unaligned access to 64-bit integers
and a fast instruction to count leading zeros in
a 64-bit integer (__builtin_ctzll()). This perhaps
should be enabled on some other archs too.
Thanks to Chenxi Mao for the original patch:
https://github.com/tukaani-project/xz/pull/75 (the first commit)
According to the numbers there, this may improve encoding
speed by about 3-5 %.
This enables the 8-byte method on MSVC ARM64 too which
should work but wasn't tested.
Lasse Collin
Jia Tan
Chenxi Mao