Switch to a macro `ZSTD_FALLTHROUGH;` instead of a comment. On supported
compilers this uses an attribute, otherwise it becomes a comment.
This is necessary to be compatible with clang's `-Wfall-through`, and
gcc's `-Wfall-through=2` which don't support comments. Without this the
linux build emits a bunch of warnings.
Also add a test to CI to ensure that we don't regress.
the new alignment setting is better for gcc-9 and gcc-10
by about ~+5%.
Unfortunately, it's worse for essentially all other compilers.
Make the new alignment setting conditional to gcc-9+.
changed strategy,
now unconditionally prefetch the first 2 cache lines,
instead of cache lines corresponding to the first and last bytes of the match.
This better corresponds to cpu expectation,
which should auto-prefetch following cachelines on detecting the sequential nature of the read.
This is globally positive, by +5%,
though exact gains depend on compiler (from -2% to +15%).
The only negative counter-example is gcc-9.
pipeline increased from 4 to 8 slots.
This change substantially improves decompression speed when there are long distance offsets.
example with enwik9 compressed at level 22 :
gcc-9 : 947 -> 1039 MB/s
clang-10: 884 -> 946 MB/s
I also checked the "cold dictionary" scenario,
and found a smaller benefit, around ~2%
(measurements are more noisy for this scenario).
* Switch to yearless copyright per FB policy
* Fix up SPDX-License-Identifier lines in `contrib/linux-kernel` sources
* Add zstd copyright/license header to the `contrib/linux-kernel` sources
* Update the `tests/test-license.py` to check for yearless copyright
* Improvements to `tests/test-license.py`
* Check `contrib/linux-kernel` in `tests/test-license.py`
Following #2545,
I noticed that one field in `seq_t` is optional,
and only used in combination with prefetching.
(This may have contributed to static analyzer failure to detect correct initialization).
I then wondered if it would be possible to rewrite the code
so that this optional part is handled directly by the prefetching code
rather than delegated as an option into `ZSTD_decodeSequence()`.
This resulted into this refactoring exercise
where the prefetching responsibility is better isolated into its own function
and `ZSTD_decodeSequence()` is streamlined to contain strictly Sequence decoding operations.
Incidently, due to better code locality,
it reduces the need to send information around,
leading to simplified interface, and smaller state structures.
When the output buffer is `NULL` with size 0, but the frame content size
is non-zero, we will write to the NULL pointer because our bounds check
underflowed.
This was exposed by a recent PR that allowed an empty frame into the
single-pass shortcut in streaming mode.
* Fix the bug.
* Fix another NULL dereference in zstd-v1.
* Overflow checks in 32-bit mode.
* Add a dedicated test.
* Expose the bug in the dedicated simple_decompress fuzzer.
* Switch all mallocs in fuzzers to return NULL for size=0.
* Fix a new timeout in a fuzzer.
Neither clang nor gcc show a decompression speed regression on x86-64.
On x86-32 clang is slightly positive and gcc loses 2.5% of speed.
Credit to OSS-Fuzz.
* All copyright lines now have -2020 instead of -present
* All copyright lines include "Facebook, Inc"
* All licenses are now standardized
The copyright in `threading.{h,c}` is not changed because it comes from
zstdmt.
The copyright and license of `divsufsort.{h,c}` is not changed.
The alignment is added before the loop, so this shouldn't hurt
performance in any case. The only way it hurts is if there is already
performance instability, and we force it to be stable but in the bad
case.
This consistently gets us into the good case with gcc-{7,8,9} on an
Intel i9-9900K and clang-9. gcc-5 is 5% worse than its best case but has
stable performance. We get consistently good behavior on my Macbook Pro
compiled with both clang and gcc-8. It ends up in the 50% from DSB and
50% from MITE case, but the performance is the same as the 85% DSB case,
so thats fine.
In the case that `op >= oend_w` it is possible that `diff < 8` because
the two buffers could be adjacent.
Credit to OSS-Fuzz, which found the bug. It isn't reproducible because
it depends on the memory layout.
* Bump `WILDCOPY_OVERLENGTH` to 16 to fix the wildcopy overread.
* Optimize `ZSTD_wildcopy()` by removing unnecessary branches and
unrolling the loop.
* Extract `ZSTD_overlapCopy8()` into its own function.
* Add `ZSTD_safecopy()` for `ZSTD_execSequenceEnd()`. It is
optimized for single long sequences, since that is the important
case that can end up in `ZSTD_execSequenceEnd()`. Without this
optimization, decompressing a block with 1 long match goes
from 5.7 GB/s to 800 MB/s.
* Refactor `ZSTD_execSequenceEnd()`.
* Increase the literal copy shortcut to 16.
* Add a shortcut for offset >= 16.
* Simplify `ZSTD_execSequence()` by pushing more cases into
`ZSTD_execSequenceEnd()`.
* Delete `ZSTD_execSequenceLong()` since it is exactly the
same as `ZSTD_execSequence()`.
clang-8 seeds +17.5% on silesia and +21.8% on enwik8.
gcc-9 sees +12% on silesia and +15.5% on enwik8.
TODO: More detailed measurements, and on more datasets.
Crdit to OSS-Fuzz for finding the wildcopy overread.
Summary: The idea behind wildcopy is that it can be cheaper to copy more bytes (say 8) than it is to copy less (say, 3). This change takes that further by exploiting some properties:
1. it's almost always OK to copy 16 bytes instead of 8, which means fewer copy instructions, and fewer branches
2. A 16 byte chunk size means that ~90% of wildcopy invocations will have a trip count of 1, so branch prediction will be improved.
Speedup on Xeon E5-2680v4 is in the range of 3-5%.
Measured wildcopy length distributions on silesia.tar:
level <=8 <=16 <=24 >24
1 78.05% 11.49% 3.52% 6.94%
3 82.14% 8.99% 2.44% 6.43%
6 85.81% 6.51% 2.92% 4.76%
8 83.02% 7.31% 3.64% 6.03%
10 84.13% 6.67% 3.29% 5.91%
15 77.58% 7.55% 5.21% 9.66%
16 80.07% 7.20% 3.98% 8.75%
Test Plan: benchmark silesia, make check
