Further improve build_decode_table() performance by splitting the "fill
direct entries" and "fill subtable pointers and subtables" steps into
separate loops and making some other optimizations.
Improve libdeflate's worst-case performance decompressing malicious
DEFLATE streams by about 14x, bringing it within a factor of about 2x of
zlib, by skipping rebuilding the decode tables for the static Huffman
codes when they're already loaded into the decompressor.
This improves performance decompressing a stream of all empty static
Huffman blocks from about 0.36 MB/s to 175 MB/s, or the original
reproducer given on the Github issue from about 3.3 MB/s to 219 MB/s.
A regression test is added for these cases as well as the empty dynamic
Huffman blocks case to verify worst-case performance comparable to zlib.
Resolves https://github.com/ebiggers/libdeflate/issues/33
NEON intrinsics cannot be used when compiling for an ARM CPU without
hardware floating point support, e.g. the Debian armel port. In this
case arm_neon.h cannot even be included as it causes an #error.
[Based on a patch by Adrian Bunk <bunk@debian.org>, but changed to check
for __ARM_FP instead of !__SOFTFP__ to be consistent with arm_neon.h,
and added a comment.]
To match common shared library packaging conventions: name the shared
library libdeflate.so.0, with matching soname, and make libdeflate.so
a symlink that points to it.
All 64bit PowerPC CPUs handle unaligned accesses reasonably fast, so
set UNALIGNED_ACCESS_IS_FAST.
Decompression of the snappy html test case is almost 50% faster on
POWER9 with this patch applied.
Make it compatible with the new code organization, make it run the
test_checksums program for each implementation, and run each
implementation in both 64-bit and 32-bit modes.
Now that we detect CPU features on 32-bit x86, allow the SSE2
implementation of Adler-32 to be selected at runtime based on the
presence of the SSE2 feature.
Use 'volatile' for the CPU feature masks and dispatched function
pointers. We don't need memory barriers for them, so 'volatile' is good
enough to stop the compiler from inserting bogus reads/writes.
Move the x86 and ARM-specific code into their own directories to prevent
it from cluttering up the main library. This will make it a bit easier
to add new architecture-specific code.
But to avoid complicating things too much for people who aren't using
the provided Makefile, we still just compile all .c files for all
architectures (irrelevant ones end up #ifdef'ed out), and the headers
are included explicitly for each architecture so that an
architecture-specific include path isn't needed. So, now people just
need to compile both lib/*.c and lib/*/*.c instead of only lib/*.c.
Replace COMPILER_SUPPORTS_TARGET_INTRINSICS with macros for the
individual features, since COMPILER_SUPPORTS_TARGET_INTRINSICS was
x86-specific and would cause confusion when we try to use intrinsics in
'target' functions for other architectures.
