* Flatten ZSTD_row_getMatchMask
* Remove the SIMD abstraction layer.
* Add big endian support.
* Align `hashTags` within `tagRow` to a 16-byte boundary.
* Switch SSE2 to use aligned reads.
* Optimize scalar path using SWAR.
* Optimize neon path for `n == 32`
* Work around minor clang issue for NEON (https://bugs.llvm.org/show_bug.cgi?id=49577)
* replace memcpy with MEM_readST
* silence alignment warnings
* fix neon casts
* Update zstd_lazy.c
* unify simd preprocessor detection (#3)
* remove duplicate asserts
* tweak rotates
* improve endian detection
* add cast
there is a fun little catch-22 with gcc: result from pmovmskb has to be cast to uint32_t to avoid a zero-extension
but must be uint16_t to get gcc to generate a rotate instruction..
* more casts
* fix casts
better work-around for the (bogus) warning: unary minus on unsigned
Call `ZSTD_enforceMaxDist()` before each block with the beginning of the
block. This ensures that `lowLimit` is updated to `dictLimit` whenever
the ext-dict is out of range, so we can use prefix mode for speed.
This can cause non-determinism because prefix mode and ext-dict mode
match finders can return different results. It can also hurt speed
because ext-dict match finders are slower.
The scenario is:
1. Compress large data with a dictionary.
2. The dictionary goes out of bounds, so we invalidate it.
3. However, we still have `lowLimit < dictLimit`, since it is
never updated.
4. We will call the ext-dict match finder instead of the prefix one.
The repcode checks disallowed repcodes that are equal to `windowLow`.
This is slightly inefficient, but isn't a problem on its own. Together
with the next commit, it cause non-determinism.
`ZSTD_insertBt1()` has a speed optimization that skips the prefix of
very long matches.
https://github.com/facebook/zstd/blob/40def70387f99b239f3f566ba399275a8fd44cde/lib/compress/zstd_opt.c#L476
This optimization is based off the length longest match found. However,
when indices are reset, we only ensure that we can reference the whole
window starting from `ip`. If the previous block ended with a long match
then `nextToUpdate` could be much less than `ip`. It might be far enough
back that `nextToUpdate < maxDist`, so it doesn't have a full window of
data to reference. This can cause non-determinism bugs, because we may
find a match that is beyond `ip - maxDist`, and may sometimes be
un-referencable, and that match triggers the speed optimization.
The fix is to base the `windowLow` off of the `target` of
`ZSTD_updateTree_internal()`, because anything below that value will be
obsolete by the time `ZSTD_updateTree_internal()` completes.
and restored limit to 256 when in 64-bit mode
(it was reduced to 200 to give more room for 32-bit).
This should fix test instability issues
using lot of threads in 32-bit environments.
When running armv6 userspace on armv8 hardware with a 64 bit Linux kernel,
the mode 2 caused SIGBUS (unaligned memory access).
Running all our arm builds in the build farm
only on armv8 simplifies administration a lot.
Depending on compiler and environment, this change might slow down
memory accesses (did not benchmark it). The original analysis is 6 years old.
Fixes#2632
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.
Linearly back off the frequency of overflow correction based on the
number of times the `ZSTD_window_t` has been overflow corrected. This
will still allow the fuzzer to quickly find overflow correction bugs,
while also keeping good speed for larger inputs.
Additionally, the `nbOverflowCorrections` variable can be useful for
debugging coredumps, since we can inspect the `ZSTD_CCtx` to see if
overflow correction has happened yet.
I've verified this fixes the timeouts in OSS-Fuzz (176 seconds -> 6
seconds). I've also verified that fuzzers and `fuzzer` and `zstreamtest`
still catch the row-hash overflow correction bug.
The FAQ covers the questions asked in Issue #2566. It first covers why
you would want to use a dictionary, then what a dictionary is, and
finally it tells you how to train a dictionary, and clarifies some of
the parameters.
There is definitely more that could be said about some of the advanced
trainers, but this should be a good start.
This flag forces zstd to always load the prefix in ext-dict mode, even
if it happens to be contiguous, to force determinism. It also applies to
dictionaries that are re-processed.
A determinism test case is also added, which fails without
`ZSTD_c_deterministicRefPrefix` and passes with it set.
Question: Should this be the default behavior? It isn't in this PR.
