We found that movemask is not used properly or consumes too much CPU.
This effort helps to optimize the movemask emulation on ARM.
For level 8-9 we saw 3-5% improvements. For level 10 we say 1.5%
improvement.
The key idea is not to use pure movemasks but to have groups of bits.
For rowEntries == 16, 32 we are going to have groups of size 4 and 2
respectively. It means that each bit will be duplicated within the group
Then we do AND to have only one bit set in the group so that iteration
with lowering bit `a &= (a - 1)` works as well.
Also, aarch64 does not have rotate instructions for 16 bit, only for 32
and 64, that's why we see more improvements for level 8-9.
vshrn_n_u16 instruction is used to achieve that: vshrn_n_u16 shifts by
4 every u16 and narrows to 8 lower bits. See the picture below. It's
also used in
[Folly](c570259008/folly/container/detail/F14Table.h (L446)).
It also uses 2 cycles according to Neoverse-N{1,2} guidelines.
64 bit movemask is already well optimized. We have ongoing experiments
but were not able to validate other implementations work reliably faster.
This commit avoids checking whether a hashtable write is safe in two of the
three match-found paths in `ZSTD_compressBlock_fast_noDict_generic`. This pro-
duces a ~0.5% speed-up in compression.
A comment in the code describes why we can skip this check in the other two
paths (the repcode check and the first match check in the unrolled loop).
A downside is that in the new position where we make this check, we have not
yet computed `mLength`. We therefore have to avoid writing *possibly* dangerous
positions, rather than the old check which only avoids writing *actually*
dangerous positions. This leads to a miniscule loss in ratio (remember that
this scenario can only been triggered in very negative levels or under incomp-
ressibility acceleration).
credit to oss-fuzz
This issue could happen when using the new Sequence Compression API in Explicit Delimiter Mode
with a too small dstCapacity.
In which case, there was one place where the buffer size wasn't checked.
discovered by oss-fuzz
It's a bug in the test itself :
ZSTD_compressBound() as an upper bound of the compress size
only works for data compressed "normally".
But in situations where many flushes are forcefully introduced,
this creates many more blocks,
each of which has a potential to increase the size by 3 bytes.
In extreme cases (lots of small incompressible blocks), the expansion can go beyond ZSTD_compressBound().
This situation is similar when using the CompressSequences() API
with Explicit Block Delimiters.
In which case, each explicit block acts like a deliberate flush.
When employed by a fuzzer, it's possible to generate scenarios like the one described above,
with tons of incompressible blocks of small sizes,
thus going beyond ZSTD_compressBound().
fix : when using Explicit Block Delimiters, use a larger bound, to account for this scenario.
credit to oss-fuzz
In rare circumstances, the block-splitter might cut a block at the exact beginning of a repcode.
In which case, since litlength=0, if the repcode expected 1+ literals in front, its signification changes.
This scenario is controlled in ZSTD_seqStore_resolveOffCodes(),
and the repcode is transformed into a raw offset when its new meaning is incorrect.
In more complex scenarios, the previous block might be emitted as uncompressed after all,
thus modifying the expected repcode history.
In the case discovered by oss-fuzz, the first block is emitted as uncompressed,
so the repcode history remains at default values: 1,4,8.
But since the starting repcode is repcode3, and the literal length is == 0,
its meaning is : = repcode1 - 1.
Since repcode1==1, it results in an offset value of 0, which is invalid.
So that's what the `assert()` was verifying : the result of the repcode translation should be a valid offset.
But actually, it doesn't matter, because this result will then be compared to reality,
and since it's an invalid offset, it will necessarily be discarded if incorrect,
then the repcode will be replaced by a raw offset.
So the `assert()` is not useful.
Furthermore, it's incorrect, because it assumes this situation cannot happen, but it does, as described in above scenario.
which properly represents the maximum bit size of compressed literals (11) as defined in the specification.
To be preferred from HUF_TABLELOG_DEFAULT which represents the same value but by accident.
Name selected to keep the same convention as existing width definitions,
MLFSELog, LLFSELog and OffFSELog.
In rare cases, the default huffman depth selector is a bit too harsh,
requiring brutal adaptations to the tree,
resulting is some loss of compression ratio.
This new heuristic avoids the worse cases, favoring compression ratio.
As an example, compression of a specific distribution of 771 literals
is now improved to 441 bytes, from 601 bytes before.
