LZ4_compress_HC_destSize() had a tendency
to discard its last match when this match overflowed specified dstBuffer limit.
The impact is generally moderate,
but occasionally huge,
typically when this last match is very large
(such as compressing a bunch of zeroes).
Issue #784 fixed for both Chain and Opt implementations.
Added a unit test suggested by @remittor checking this topic.
LZ4_decompress_safe_partial()
now also supports a scenario where
nb_bytes_to_generate is <= block_decompressed_size
And
nb_bytes_to_read is >= block_compressed_size.
Previously, the only supported scenario was
nb_bytes_to_read == block_compress_size.
Pay attention that,
if nb_bytes_to_read is > block_compressed_size,
then, necessarily, it requires that
nb_bytes_to_generate is <= block_decompress_size.
If both are larger, it will generate corrupted data.
which serves no more purpose.
The comment implies that the simple presence of this unused function was affecting performance,
and that's the reason why it was not removed earlier.
This is likely another side effect of instruction alignment.
It's obviously unreliable to rely on it in this way,
meaning that the impact will be different, positive of negative,
with any minor code change, and any minor compiler version change, even parameter change.
`LZ4_memcpy()` uses `__builtin_memcpy()` to ensure that clang/gcc
can inline the `memcpy()` calls in freestanding mode.
This is necessary for decompressing the Linux Kernel with LZ4.
Without an analogous patch decompression ran at 77 MB/s, and with
the patch it ran at 884 MB/s.
Similar work in the kernel:
https://patchwork.kernel.org/patch/11351499/
UBsan (+clang-10) complains about doing pointer arithmetic (adding 0)
to a nullpointer.
This patch is tested with clang-10+ubsan
Define 0-argument functions like foo(void) instead of foo(), in order
to avoid a warning with -Wold-style-definition. This makes it easier
to embed lz4.c in projects that compile with -Werror
-Wold-style-definition.
when src ptr is in very low memory area (< 64K),
the virtual reference to data in dictionary
might end up in a very high memory address.
Since it's not a "real" memory address,
just a virtual one, to calculate distance,
it doesn't matter : only distance matters.
The assert was to restrictive.
Fixed.
by enabling the fast decoder path.
Visual requires a different set of macro constants to detect x86 / x64.
On my laptop, decoding speed on x64 went up from 3.12 to 3.45 GB/s.
32-bit is less impressive, though still favorable,
with speed increasing from 2.55 to 2.60 GB/s.
So both cases are now enabled.
Suggested by Bartosz Taudul (@wolfpld).
Otherwise, the output from decoding LZ4-compressed input could be
platform dependent.
Also add a compile-time check to confirm the existing code's assumptions
that, if <stdint.h> isn't used, then sizeof(int) == 4.
Updates #792
PR#756 fixed the data corruption bug, but didn't clear `ip`. PR#760
fixed that off-by-one error, but missed the case where `ip == filledIp`,
which is harder for the fuzzers to find (it took 20 days not 1 day).
Verified this fixed the issue reported by OSS-Fuzz.
Credit to OSS-Fuzz.
When using an empty dictionary, we bail out of loading or attaching it in
ways that leave the working context in potentially slightly different states.
In particular, in some paths, we will cause the currentOffset to be non-zero,
while in others we would allow it to remain 0.
This difference in behavior is perfectly harmless, but in some situations, it
can produce slight differences in the compressed output. For sanity's sake,
we currently try to maintain a strict correspondence between the behavior of
the dict attachment and the dict loading paths. This patch restores them to
behaving identically.
This shouldn't have any negative side-effects, as far as I can tell. When
writing the dict attachment code, I tried to preserve zeroed currentOffsets
when possible, since they benchmarked as very slightly faster. However, the
case of attaching an empty dictionary is probably rare enought that it's
acceptable to minisculely degrade performance in that corner case.
It's now possible to select a custom LZ4_DISTANCE_MAX at compile time,
provided it's <= 65535.
However, in some cases (when compressing in byU16 mode),
the new distance wasn't respected,
as it used to implied that it was necessarily within range.
Added a distance check for this case.
Also : added a new TravisCI test which ensures that
custom LZ4_DISTANCE_MAX compiles correctly
and compresses correctly (relying on `assert()` to find outsized offsets).
When I did FAST_DEC_LOOP performance test, I found the
offset check is much more than v1.8.3
You will see the condition check difference via lzbench with dickens test case.
v1.8.3 34959
v.1.9.x 1055885
After investigate the code, we could see the difference.
v.1.8.3 SKIP the condition check if
if condition is true in:
https://github.com/lz4/lz4/blob/v1.8.3/lib/lz4.c#L1463
AND below condition is true
https://github.com/lz4/lz4/blob/v1.8.3/lib/lz4.c#L1478\
The offset check should be invoked.
v1.9.3
The offset check code will be invoked in every loop which lead to downgrade.
So the fix would be move this check to specific condition
to avoid useless condition check.
After this change, the call number is same as v1.8.3
and deprecate LZ4_decompress_fast(),
with deprecation warnings enabled by default.
Note that, as a consequence of the fix,
LZ4_decompress_fast is now __slower__ than LZ4_decompress_safe().
That's because, since it doesn't know the input buffer size,
it must progress more cautiously into the input buffer
to ensure to out-of-bound read.
between lz4.c and lz4hc.c .
was left in a strange state after the "amalgamation" patch.
Now only 3 directives remain,
same name across both implementations,
single definition place.
Might allow some light simplification due to reduced nb of states possible.
when one block was not compressible,
it would tag the context as `dirty`,
resulting in compression automatically bailing out of all future blocks,
making the rest of the frame uncompressible.
since Visual 2017,
worries about potential overflow, which are actually impossible.
Replaced (c * a) by (c ? a : 0).
Will likely replaced a * by a cmov.
Probably harmless for performance.
make it possible to generate LZ4-compressed block
with a controlled maximum offset (necessarily <= 65535).
This could be useful for compatibility with decoders
using a very limited memory budget (<64 KB).
Answer #154
it fails on x86 32-bit mode :
Visual reports an alignment of 8-bytes (even with alignof())
but actually only align LZ4_stream_t on 4 bytes.
The alignment check then fails, resulting in missed initialization.
- promoted LZ4_resetStream_fast() to stable
- moved LZ4_resetStream() into deprecate, but without triggering a compiler warning
- update all sources to no longer rely on LZ4_resetStream()
note : LZ4_initStream() proposal is slightly different :
it's able to initialize any buffer, provided that it's large enough.
To this end, it accepts a void*, and returns an LZ4_stream_t*.
For small offsets of size 1, 2, 4 and 8, we can set a single uint64_t,
and then use it to do a memset() variation. In particular, this makes
the somewhat-common RLE (offset 1) about 2-4x faster than the previous
implementation - we avoid not only the load blocked by store, but also
avoid the loads entirely.
Generally we want our wildcopy loops to look like the
memcpy loops from our libc, but without the final byte copy checks.
We can unroll a bit to make long copies even faster.
The only catch is that this affects the value of FASTLOOP_SAFE_DISTANCE.
We've already checked that we are more than FASTLOOP_SAFE_DISTANCE
away from the end, so this branch can never be true, we will have
already jumped to the second decode loop.
Use LZ4_wildCopy16 for variable-length literals. For literal counts that
fit in the flag byte, copy directly. We can also omit oend checks for
roughly the same reason as the previous shortcut: We check once that both
match length and literal length fit in FASTLOOP_SAFE_DISTANCE, including
wildcopy distance.
Add an LZ4_wildCopy16, that will wildcopy, potentially smashing up
to 16 bytes, and use it for match copy. On x64, this avoids many
blocked loads due to store forwarding, similar to issue #411.
Copy the main loop, and change checks such that op is always less
than oend-SAFE_DISTANCE. Currently these are added for the literal
copy length check, and for the match copy length check.
Otherwise the first loop is exactly the same as the second. Follow on
diffs will optimize the first copy loop based on this new requirement.
I also tried instead making a separate inlineable function for the copy
loop (similar to existing partialDecode flags, etc), but I think the
changes might be significant enough to warrent doubling the code, instead
pulling out common functionality to separate functions.
This is the basic transformation that will allow several following optimisations.
Dictionaries don't need to be > 4 bytes, they need to be >= 4 bytes. This test
was overly conservative.
Also removes the test in `LZ4_attach_dictionary()`.
Fixes a mismatch in behavior between loading into the context (via
`LZ4_loadDict()`) a very small (<= 4 bytes) non-contiguous dictionary, versus
attaching it with `LZ4_attach_dictionary()`.
Before this patch, this divergence could be reproduced by running
```
make -C tests fuzzer MOREFLAGS="-m32"
tests/fuzzer -v -s1239 -t3146
```
Making sure these two paths behave exactly identically is an easy way to test
the correctness of the attach path, so it's desirable that this remain an
unpolluted, high signal test.
so "funny" thing with cppcheck
is that no 2 versions give the same list of warnings.
On Mac, I'm using v1.81, which had all warnings fixed.
On Travis CI, it's v1.61, and it complains about a dozen more/different things.
On Linux, it's v1.72, and it finds a completely different list of a half dozen warnings.
Some of these seems to be bugs/limitations in cppcheck itself.
The TravisCI version v1.61 seems unable to understand %zu correctly, and seems to assume it means %u.
