* 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
answering #1407.
Also : removed obsolete function ZSTD_setDStreamParameter()
which could only be used with one parameter (DStream_p_maxWindowSize).
Now replaced by ZSTD_DCtx_setWindowSize() (which exists since a few revisions)
isolate all logic associated with block decompression
into its own module.
zstd_decompress is still in charge
of context creation/destruction,
frames, headers, streaming, special blocks, etc.
Compressed blocks themselves are now handled within zstd_decompress_block .
We could undersize the literals buffer by up to 11 bytes,
due to a combination of 2 bugs:
* The literals buffer didn't have `WILDCOPY_OVERLENGTH` extra
space, like it is supposed to.
* We didn't check the literals buffer size in `ZSTD_sufficientBuff()`.
There were 2 competing set of debug functions
within zstd_internal.h and bitstream.h.
They were mostly duplicate, and required care to avoid messing with each other.
There is now a single implementation, shared by both.
Significant change :
The macro variable ZSTD_DEBUG does no longer exist,
it has been replaced by DEBUGLEVEL,
which required modifying several source files.
removed "cached" structure.
prices are now saved in the optimal table.
Primarily done for simplification.
Might improve speed by a little.
But actually, and surprisingly, also improves ratio in some circumstances.
* Expose the reference external sequences API for zstdmt.
Allows external sequences of any length, which get split when necessary.
* Reset the LDM window when the context is reset.
* Store the maximum number of LDM sequences.
* Sequence generation now returns the number of last literals.
* Fix sequence generation to not throw out the last literals when blocks of
more than 1 MB are encountered.
* `ZSTD_ldm_generateSequences()` generates the LDM sequences and
stores them in a table. It should work with any chunk size, but
is currently only called one block at a time.
* `ZSTD_ldm_blockCompress()` emits the pre-defined sequences, and
instead of encoding the literals directly, it passes them to a
secondary block compressor. The code to handle chunk sizes greater
than the block size is currently commented out, since it is unused.
The next PR will uncomment exercise this code.
* During optimal parsing, ensure LDM `minMatchLength` is at least
`targetLength`. Also don't emit repcode matches in the LDM block
compressor. Enabling the LDM with the optimal parser now actually improves
the compression ratio.
* The compression ratio is very similar to before. It is very slightly
different, because the repcode handling is slightly different. If I remove
immediate repcode checking in both branches the compressed size is exactly
the same.
* The speed looks to be the same or better than before.
Up Next (in a separate PR)
--------------------------
Allow sequence generation to happen prior to compression, and produce more
than a block worth of sequences. Expose some API for zstdmt to consume.
This will test out some currently untested code in
`ZSTD_ldm_blockCompress()`.
The deep fuzzer tests caught a subtle bug that was probably there for a long time.
The impact of the bug is not a crash, or any other clear error signal,
rather, it reduces performance, by cutting data into smaller blocks.
Eventually, the following test would fail because it produces too many 1-byte blocks,
requiring more space than buffer can provide :
`./zstreamtest_asan --mt -s3514 -t1678312 -i1678314`
The root scenario is as follows :
- Create context, initialize it using explicit parameters or a `cdict` to pin them down, set `pledgedSrcSize=1`
- The compression parameters will not be adapted, but `windowSize` and `blockSize` will be automatically set to `1`.
`windowSize` and `blockSize` are dynamic values, set within `ZSTD_resetCCtx_internal()`.
The automatic adaptation makes it possible to generate smaller contexts for smaller input sizes.
- Complete compression
- New compression with same context, using same parameters, but `pledgedSrcSize=ZSTD_CONTENTSIZE_UNKNOWN`
trigger "continue mode"
- Continue mode doesn't modify blockSize, because it used to depend on `windowLog` only,
but in fact, it also depends on `pledgedSrcSize`.
- The "old" blocksize (1) is still there,
next compression will use this value to cut input into blocks,
resulting in more blocks and worse performance than necessary performance.
Given the scenario, and its possible variants, I'm surprised it did not show up before.
But I suspect it did show up, it's just that it never triggered an error, because "worse performance" is not a trigger.
The above test is a special corner case, where performance is so impacted that it reaches an error case.
The fix works, but I'm not completely pleased.
I think the current code relies too much on implied relations between variables.
This will likely break again in the future when some related part of the code change.
Unfortunately, no time to make larger changes if we want to keep the release target for zstd v1.3.3.
So a longer term fix will have to be considered after the release.
To do : create a reliable test case which triggers this scenario for CI tests.
There was a flaw in the formula
which compared literal cost with match cost :
at a given position,
a non-null literal suite is going to be part of next sequence,
while if position ends a previous match, to immediately start another match,
next sequence will have a litlength of zero.
A litlength of zero has a non-null cost.
It follows that literals cost should be compared to match cost + litlength==0.
Not doing so gave a structural advantage to matches, which would be selected more often.
I believe that's what led to the creation of the strange heuristic which added a complex cost to matches.
The heuristic was actually compensating.
It was probably created through multiple trials, settling for best outcome on a given scenario (I suspect silesia.tar).
The problem with this heuristic is that it's hard to understand,
and unfortunately, any future change in the parser would impact the way it should be calculated and its effects.
The "proper" formula makes it possible to remove this heuristic.
Now, the problem is : in a head to head comparison, it's sometimes better, sometimes worse.
Note that all differences are small (< 0.01 ratio).
In general, the newer formula is better for smaller files (for example, calgary.tar and enwik7).
I suspect that's because starting statistics are pretty poor (another area of improvement).
However, for silesia.tar specifically, it's worse at level 22 (while being better at level 17, so even compression level has an impact ...).
It's a pity that zstd -22 gets worse on silesia.tar.
That being said, I like that the new code gets rid of strange variables,
which were introducing complexity for any future evolution (faster variants being in mind).
Therefore, in spite of this detrimental side effect, I tend to be in favor of it.
this version has same speed as branch `opt`
which is itself 5-10% slower than branch `dev`
(no identified reason)
It does not compress exactly the same as `opt` or `dev`,
maybe because it doesn't stop search after repcodes,
leading to sometimes better compression, sometimes worse
(by a small margin).
warning : _extDict path does not work for the time being
This means that benchmark module works,
but file module will fail with large files (and high compression level).
Objective is to fuse _extDict path into current one,
in order to have a single parser to maintain.
* Maximum window size in 32-bit mode is 1GB, since allocations for 2GB fail
on my Mac.
* Maximum window size in 64-bit mode is 2GB, since that is the largest
power of 2 that works with the overflow prevention.
* Allow `--long=windowLog` to set the window log, along with
`--zstd=wlog=#`. These options also set the window size during
decompression, but don't override `--memory=#` if it is set.
* Present a helpful error message when the window size is too large during
decompression.
* The long range matcher defaults to a hash log 7 less than the window log,
which keeps it at 20 for window log 27.
* Keep the default long range matcher window size and the default maximum
window size at 27 for the API and CLI.
* Add tests that use the maximum window size and hash size for compression
and decompression.
