With statistic data of test data files of silesia
the chance of position beyond highThreshold is very
low (~1.3%@L8 in most cases, all <2.5%), and is in
"lowprob area". Add the branch hint so compiler can
get better pipiline codegen.
With this change it is observed ~1% of mozilla and
xml, and slight (0.3%~0.8%) but consistent uplift on
other files on Arm N1.
Signed-off-by: Jun He <jun.he@arm.com>
Change-Id: Id9ba1d5c767e975290b5c1bf0ecce906544f4ade
match is used for following sequence copy. It is
only updated when extDict is needed, which is a
low probability case. So it can be prefetched to
reduce cache miss.
The benchmarks on various Arm platforms showed
uplift from 1% ~ 14% with gcc-11/clang-14.
Signed-off-by: Jun He <jun.he@arm.com>
Change-Id: If201af4799d2455d74c79f8387404439d7f684ae
* Intial commit to address 3090. Added support to decompress empty block
* Update zstd_decompress_block.c
Addressed review comments for the case of 'set_basic'
* Update lib/decompress/zstd_decompress_block.c
Co-authored-by: Nick Terrell <nickrterrell@gmail.com>
* Update lib/decompress/zstd_decompress_block.c
Co-authored-by: Nick Terrell <nickrterrell@gmail.com>
Co-authored-by: Nick Terrell <nickrterrell@gmail.com>
ZSTD_seqSymbol is a structure with total of 64 bits
wide. So it can be loaded in one operation and
extract its fields by simply shifting or extracting
on aarch64.
GCC doesn't recognize this and generates more
unnecessary ldr/ldrb/ldrh operations that cause
performance drop.
With this change it is observed 2~4% uplift of
silesia and 2.5~6% of cantrbry @L8 on Arm N1.
Signed-off-by: Jun He <jun.he@arm.com>
Change-Id: I7748909204cf78a17eb9d4f2333692d53239daa8
This saves some 1.7Kb in rodata section (x86_64, zstd tool),
while assembler code stays the same except
the type of a few load/extend instructions.
Should not have negative performance implications.
* When dynamic dispatching to bmi2 add lzcnt and bmi to the
TARGET_ATTRIBUTE.
* Centralize the bmi2 TARGET_ATTRIBUTE definition to
BMI2_TARGET_ATTRIBUTE so we can change it in the future.
* Only enable bmi2 when both bmi1 & bmi2 are supported. There shouldn't
be any cases where bmi2 is supported but bmi1 isn't. But, since we are
using the instruction we should check bmi1 as well.
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.
