Keep weak-symbol discovery and private dctx/ddict extraction in the C adapter,
but project the exact trace callback types into Rust and invoke them from the
Rust trace lifecycle. Preserve the nullable-hook behavior, trace context
width, dictionary-cold timing, record field layout, size conversions, and the
no-trace build path. Add focused Rust policy tests for callback, record, and
null-projection behavior.
Test Plan:
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo check --manifest-path rust/Cargo.toml --tests
- ulimit -v 41943040; make -j1
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo clippy --manifest-path rust/cli/Cargo.toml --all-targets -- -D warnings
- ulimit -v 41943040; make -j1 -C tests test
Remove the C-only DDict offset globals and route production
ZSTD_copyDDictParameters() through the existing C-produced DCtx view. This
keeps the private decoder layout and optional fuzzing fields owned by the
canonical projection while preserving the DDict entropy, prefix, window, and
checksum state updates. Keep the synthetic offset writer test-only so the
Rust unit tests remain independently linkable.
Test Plan:
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo check --manifest-path rust/cli/Cargo.toml --all-targets
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo test --manifest-path rust/cli/Cargo.toml --all-targets (207 passed)
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo clippy --manifest-path rust/cli/Cargo.toml --all-targets -- -D warnings
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 make -j1
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 make -j1 -C tests test
- git diff --check
Add the existing heap-mode bridge declaration beside the other decompression shim prototypes so the new Rust policy seam remains warning-clean under the native C build.
Test Plan:
- git diff --cached --check
- capped full test suite passed before this warning-only fix; final capped native rebuild follows
Route the configured default maximum window size through the Rust policy layer while keeping the build-time C configuration value and decoder layout in C. The explicit projection preserves overridden builds exactly and gives the scalar policy a focused null-input contract and tests.
Test Plan:
- git diff --check -- lib/decompress/zstd_decompress.c rust/src/zstd_decompress.rs
- capped serial cargo check, clippy, nightly fmt, C syntax checks, and focused default-window tests (passed in the worker)
`ZSTD_HEAPMODE` was still returned directly from C, and the stack entry
projection also copied the raw macro. That left the one-shot stack-versus-heap
classification in the C configuration shim even though Rust owns the public
decompression branch and stack action.
Keep the build-time override in C, but pass it through a one-field repr(C)
projection to Rust. Rust normalizes values below one to the stack mode and
values at or above one to the heap mode, preserving the existing behavior
while making the branch policy explicit. The C stack object, private DCtx
layout, platform initialization, and trace boundary remain C-owned; moving
those would either duplicate build configuration or cross the requested ABI
boundary. The stack projection now uses the same normalized result as the
one-shot branch.
ABI layout assertions cover the scalar projection, and focused Rust tests cover
negative, zero, positive, maximum, and missing-configuration inputs.
Test Plan:
- `git diff --cached --check` -- passed
- `rustfmt --edition 2021 --check rust/src/zstd_decompress.rs` -- passed
- Cargo, Make, native tests, and heavy commands intentionally not run per the
task restriction; integration verification remains pending.
ZSTD_NO_FORWARD_PROGRESS_MAX is a compile-time override that was returned
directly by the C adapter. Keep the active build value in C, where
configuration belongs, but pass it through an ABI-checked scalar projection to
Rust. Rust now owns the policy boundary and preserves the exact configured
threshold, while the existing stalled-stream comparison, error mapping, and
private DCtx layout remain unchanged.
Test Plan:
- `git diff --cached --check` -- passed
- Cargo, make, builds, native tests, and heavy verification were not run per
request.
Keep the C build-time mutual-exclusion check and private decoder-context assembly in the C shim, but pass the selected force-short/force-long policy through Rust. Rust now owns normalization to the runtime, short, or long sequence decoder mode, with ABI assertions and focused null/build-policy tests.
Test Plan: Not run in this atomic commit; the capped serial Rust, native, smoke, and original test-suite verification follows.
Keep the target-specific ZSTD_LEGACY_SUPPORT macro in the C translation unit, but pass its scalar value through a Rust policy bridge before legacy dispatch. Rust now accepts only the historical supported range 1 through 7 and normalizes all other values, including a missing projection, to disabled. ABI assertions and focused tests preserve the compile-time configuration contract.
Test Plan: Not run in this atomic commit; the capped serial Rust, native, smoke, and original test-suite verification follows.
The stack-backed one-shot decompression entry point used to make its own
heap-mode rejection, static DCtx initialization, allocation-error mapping,
and dispatch ordering in C. That kept scalar control flow beside the private
DCtx layout even though Rust already owns the decompression policy and context
leaf.
Rust now consumes an ABI-checked scalar projection and callback set. It makes
the branch and ordering decisions, maps a failed static initializer to the
same memory-allocation error, and invokes the existing C-owned context leaf.
The C shim retains the local stack object, private DCtx layout, and
`ZSTD_initStaticDCtx` implementation. Recording-callback tests cover heap
rejection, initializer failure, and successful initialization-to-dispatch
ordering without fabricating the private context layout.
Test Plan:
- `git diff --cached --check` -- passed
- `rustfmt --edition 2021 --check rust/src/zstd_decompress.rs` -- passed
- Full capped Rust/native verification remains pending until the parallel
compression seam is integrated.
The decoder context remains allocated and laid out by C because optional members
change with the active build configuration. Previously, the C adapter also
owned the platform leaf: its DYNAMIC_BMI2 branch queried CPU features and
assigned the private bmi2 member directly. That kept the policy in C and made
the Rust boundary depend on the context layout.
Move the leaf to Rust and give it an opaque scalar address plus the active
configuration value. The C adapter now only extracts the optional address when
that member exists and forwards it, while DYNAMIC_BMI2=0 passes NULL. Rust
performs the dynamic decision, queries CPU support, and writes the scalar. The
existing 68-word DCtx view is unchanged. Focused unit coverage verifies that
disabled dynamic dispatch leaves the scalar untouched and enabled dispatch
publishes the CPU feature result.
Test Plan:
- `cc -fsyntax-only -Werror -DDYNAMIC_BMI2=0 -Ilib -Ilib/common -Ilib/compress -Ilib/decompress -Ilib/dict -Ilib/legacy lib/decompress/zstd_decompress.c` -- passed
- `cc -fsyntax-only -Werror -DDYNAMIC_BMI2=1 -Ilib -Ilib/common -Ilib/compress -Ilib/decompress -Ilib/dict -Ilib/legacy lib/decompress/zstd_decompress.c` -- passed
- `rustfmt --edition 2021 --check rust/src/zstd_decompress.rs` -- passed
- `git diff --check -- lib/decompress/zstd_decompress.c rust/src/zstd_decompress.rs` -- passed
Call the authoritative DDict constructor directly from the Rust decompression
policy after converting the shared repr(C) allocator projection. Remove the
C forwarding declaration and wrapper so dictionary creation no longer crosses
an unnecessary Rust-to-C-to-C boundary, while compile-time size and alignment
checks keep the allocator conversion tied to the ABI contract.
Test Plan:
- `ulimit -v 41943040; cargo +nightly fmt --manifest-path rust/Cargo.toml --all -- --check`
- `ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo clippy --manifest-path rust/Cargo.toml --all-targets -- -D warnings`
- `ulimit -v 41943040; make -j1`
- `ulimit -v 41943040; make -j1 -C tests test`
- `ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo clippy --manifest-path rust/cli/Cargo.toml --all-targets -- -D warnings`
- `ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo test --manifest-path rust/cli/Cargo.toml --all-targets`
The integrated checks ran at the combined working-tree tip, including the
parallel MT checksum seam that remains as the next commit. Standalone root
Rust unit linking remains unavailable because the crate imports C-owned bridge
symbols without a Cargo build/link setup.
Keep the heapmode guard, stack-owned ZSTD_DCtx, static initialization, and
private context layout in C. Once the context is initialized, Rust now owns the
one-shot stack-context policy: it clears the projected static-size marker and
dispatches the existing ZSTD_decompressDCtx implementation through the opaque
context pointer.
Add matching C/Rust view-layout assertions and focused callback tests covering
static-size clearing, argument forwarding, and null-context rejection.
Test Plan:
- `ulimit -v 41943040; cargo +nightly fmt --manifest-path rust/Cargo.toml --all -- --check`
- `ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo clippy --manifest-path rust/Cargo.toml --all-targets -- -D warnings`
- `ulimit -v 41943040; make -j1`
- `ulimit -v 41943040; make -j1 -C tests test` (all upstream tests completed successfully)
- Standalone Rust unit linking remains unavailable without the C-owned bridge symbols; the integrated build and upstream C/Rust oracle supplied the runtime/link coverage.
Define the decompression trace projection with the exact unsigned-long-long
representation used by the public trace API. The generic U64 typedef can be
a different C type with the same width, which made the Rust trace view fail to
compile when assigned the DCtx trace field.
Test Plan:
- ulimit -v 41943040; make -j1
Keep weak-symbol lookup and the conditional traceCtx/DCtx layout in the C
adapter, but project only the trace slot, callback pointers, context pointer,
and dictionary metadata into Rust. Rust now decides begin/end eligibility and
builds the ABI-compatible trace record, preserving null callbacks, streaming
and size conversion semantics. Existing call sites remain in their
success-only branches, so checksum/error paths still emit no trace end.
Test Plan:
- rustfmt --edition 2021 --check rust/src/zstd_decompress.rs
- git diff --check and git diff --cached --check
- Static rg/diff inspection of weak callbacks, traceCtx lifecycle, dictionary
metadata, and success-only trace-end call sites
- No cargo, make, native, fuzz, or heavy tests run by request
Move the hidden fullbench block-decoder entrypoints and the public
ZSTD_decompressBlock compatibility wrapper into Rust. The C translation unit
now only projects the configuration-dependent private DCtx fields into the
Rust block context, preserving the existing ABI and decoder feature modes.
Test Plan:
- cargo test --manifest-path rust/Cargo.toml --all-targets -- --test-threads=1
- cargo clippy --manifest-path rust/Cargo.toml --tests -- -D warnings
- make -B -C lib -j2 lib
- make -B -C tests -j2 test-rust-lib-smoke
- make -B -C tests -j2 test-cli-tests
Legacy detection, frame-size queries, one-shot decode, and buffered streaming
were still routed through the C header dispatcher even though every historical
v0.1 through v0.7 decoder is already ported under Rust feature gates. The C
implementation also duplicated the support-level boundary independently from
the Rust archive selection.
Move the dispatcher and stream lifecycle to Rust using the existing DCtx field
projection. Rust now selects the version, preserves the null-buffer and
position semantics, initializes and frees buffered legacy contexts, and calls
the frozen per-version modules. C retains only the configuration leaf that
reports ZSTD_LEGACY_SUPPORT plus the private DCtx layout and trace helpers, so
library, program, and test builds can still choose different support levels.
Test Plan:
- `cargo test --manifest-path rust/Cargo.toml --no-default-features --features compression,decompression,dict-builder,legacy-v01,legacy-v02,legacy-v03,legacy-v04,legacy-v05,legacy-v06,legacy-v07 --all-targets -- --test-threads=1` -- 528 passed.
- Native `test-legacy`, `test-invalidDictionaries`, `test-decodecorpus`, and `test-zstd` -- passed.
- Strict C syntax checks with legacy support 0, 4, 5, and 7 -- passed.
- `cargo clippy` default, benches, and tests for library and CLI -- passed.
Port sequence-decoder selection and offset-history policy into Rust. The
Rust block decoder now handles short/long selection, offset-table analysis,
history thresholds, and prefetch sequencing; C projects only the decoder mode
and private context fields required by the ABI.
Test Plan:
- cargo test --manifest-path rust/Cargo.toml --all-targets -- --test-threads=1
- cargo clippy --manifest-path rust/Cargo.toml --all-targets -- -D warnings
- make -B -C lib -j2 lib ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT=1 ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG=0
- make -B -C lib -j2 lib ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT=0 ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG=1
- make -B -C tests -j2 test-zstd
Move the legacy buffered-stream dispatcher out of the C decompression adapter
while keeping ZSTD_DCtx_s private and C-owned. Rust now consumes the existing
ZSTD_rustDctxView projection, detects the legacy version at input.pos, preserves
static-context rejection and stream-stage reset behavior, and handles legacy
context replacement through the version-gated v0.4-v0.7 Rust stream APIs.
The Rust boundary retains the C helper's NULL normalization, dictionary
lifetime sentinel, previous-version cleanup semantics, input/output position
updates, and version-unsupported and no-legacy-support fallbacks. The C legacy
free helper remains unchanged; only the dispatcher declaration and body were
removed from the adapter. Focused tests cover unsupported versions, stage
reset, context reuse/switching, and partial v0.4 input/output progress.
Test Plan:
- `cargo clippy --manifest-path rust/Cargo.toml --no-default-features --features compression` -- passed
- Same clippy command with `--benches` and `--tests` -- passed
- `cargo +nightly fmt --manifest-path rust/Cargo.toml`, followed by all three clippy commands -- passed
- Decompression/legacy clippy with `-D warnings` -- passed
- All 107 Rust tests with v0.4-v0.7 features -- passed
- Focused legacy dispatcher tests: 4 passed
- `make -C lib -j2 lib-mt` and `make -C lib -j2 lib-nomt` -- passed
- `make -C tests -j2 test-legacy` -- built, then failed in the existing simple one-shot path with one v0.8 output byte mismatch (`e0` vs expected `e2`) before the streaming check; the new dispatcher symbol was not called on that failing path
Move long-distance matching and high-level decompression from C shims into
Rust. The decoder now owns context, dictionary, parameter, one-shot, and
buffered streaming state while C retains allocation/configuration, legacy,
and trace leaves.
Move CLI parsing, safety policy, and dispatch into a separate Rust static
archive. Keeping it separate prevents library builds from retaining FIO
symbols, while C continues to own file opening, replacement, and I/O.
Program targets now select matching compression/decompression archives.
The remaining C boundary is intentional: high-level compression, optimal
parsing, dictionary building, legacy callbacks, and CLI file I/O still need
migration.
Test Plan:
- cargo test --all-targets (native and i686)
- cargo test --all-targets in rust/cli (native and i686)
- CLI crate compression-only and decompression-only feature tests
- native and i686 fuzzer/zstreamtest runs, plus legacy and dictionary tests
- ZSTD_C_PREDICT and ZSTD_HEAPMODE=0 fuzzer coverage
- library, dynamic-link, and program-target build/round-trip matrix
Refs: rust/README.md
Move literal decoding, sequence-table construction, FSE sequence decoding, and
sequence execution into Rust. The C shim retains ownership of the configured
decoder context and passes only the leaf state needed by the block codec.
Correct the two high offset-code bases while porting the tables. Their prior
values made valid large-window streams decode as corrupted data; the new unit
test fixes the exact values and a native zstream regression exercises them.
High-level frame and streaming context control remains C for now.
Test Plan:
- cargo test --all-targets
- cargo test --target i686-unknown-linux-gnu --all-targets
- cargo clippy && cargo clippy --benches && cargo clippy --tests
- cargo +nightly fmt
- make -B -C tests -j2 fuzzer zstreamtest invalidDictionaries poolTests
- ./tests/fuzzer -s5346 -i1 --no-big-tests
- ./tests/zstreamtest -i3000 -s334462
- ./tests/invalidDictionaries
- timeout 20s stdbuf -oL ./tests/poolTests
Refs: rust/README.md
Move ZSTD_DDict allocation, ownership, construction, and public ABI exports
into Rust. The C decoder keeps ZSTD_loadDEntropy until its table loader moves.
The C shim now exposes C-computed DCtx field offsets. This keeps dictionary
state correct across conditional C layouts, including fuzz fields and 32-bit
static-BMI2 builds, without duplicating decoder-context layout in Rust.
Test Plan:
- cargo fmt, cargo clippy, cargo clippy --benches, cargo clippy --tests
- cargo test --all-targets and cargo build --release
- cargo test/build --target i686-unknown-linux-gnu
- C fuzzer, zstreamtest, invalidDictionaries, and fuzzer32 smoke runs
- fuzz-enabled and i686 -mbmi2 C/Rust dictionary roundtrip harnesses
Refs: rust/README.md
Move Huffman decoding-table construction and X1/X2 single- and four-stream
decoding into Rust. The original C translation unit is now a declaration shim,
so unchanged C callers and tests resolve the public decoder ABI from the Rust
archive.
The portable implementation preserves the C workspace and error contracts,
retains the existing assembly-only internal fast loops, and mirrors the C
32-bit decode cadence so X2 writes remain within its required output window.
Forced X1/X2 modes follow the matching archive configuration from the build
integration.
Test Plan:
- cargo clippy; cargo clippy --benches; cargo clippy --tests
- cargo +nightly fmt, then repeat the Clippy checks
- cargo test --all-targets for default, forced X1, and forced X2 modes
- cargo build and cargo test for i686-unknown-linux-gnu in all three modes
- Strict normal, X1, and X2 C shim compilation
- Original fuzzer and fullbench compatibility targets
Refs: rust/README.md
Depends-on: d89ebb31
Mark that `huf_decompress_amd64.S` supports BTI and PAC, which it trivially does because it is empty for aarch64.
The issue only requested BTI markings, but it also makes sense to mark PAC, which is the only other feature.
Also run add a test for this mode to the ARM64 QEMU test. Before this PR it warns on `huf_decompress_amd64.S`, after it doesn't.
Fixes Issue #3841.
in this new method, when an `offset==0` is detected,
it's converted into (size_t)(-1), instead of 1.
The logic is that (size_t)(-1) is effectively an extremely large positive number,
which will not pass the offset distance test at next stage (`execSequence()`).
Checked the source code, and offset is always checked (as it should),
using a formula which is not vulnerable to arithmetic overflow:
```
RETURN_ERROR_IF(sequence.offset > (size_t)(oLitEnd - virtualStart),
```
The benefit is that such a case (offset==0) is always detected as corrupted data
as opposed to relying on the checksum to detect the error.
The patch adds a validation to ensure that the last field, which is
reserved, must be all-zeroes in ZSTD_decodeSeqHeaders. This prevents
potential corruption from going undetected.
Fixes an issue where corrupted input could lead to undefined behavior
due to improper validation of reserved bits.
Signed-off-by: aimuz <mr.imuz@gmail.com>
This PR introduces no functional changes. It attempts to change all
macros currently using `{ }` or some variant of that to to
`do { } while (0)`, and introduces trailing `;` where necessary.
There were no bugs found during this migration.
The bug in Visual Studios warning on this has been fixed since VS2015.
Additionally, we have several instances of `do { } while (0)` which have
been present for several releases, so we don't have to worry about
breaking peoples builds.
Fixes Issue #3830.
`HUF_DecompressFastArgs_init()` was adding 0 to NULL. Fix it by exiting
early for empty outputs. This is no change in behavior, because the
function was already exiting 0 in this case, just slightly later.
* Rename `ilimit` to `ilowest` and set it equal to `src` instead of
`src + 6 + 8`. This is safe because the fast decoding loops guarantee
to never read below `ilowest` already. This allows the fast decoder to
run for at least two more iterations, because it consumes at most 7
bytes per iteration.
* Continue the fast loop all the way until the number of safe iterations
is 0. Initially, I thought that when it got towards the end, the
computation of how many iterations of safe might become expensive. But
it ends up being slower to have to decode each of the 4 streams
individually, which makes sense.
This drastically speeds up the Huffman decoder on the `github` dataset
for the issue raised in #3762, measured with `zstd -b1e1r github/`.
| Decoder | Speed before | Speed after |
|----------|--------------|-------------|
| Fallback | 477 MB/s | 477 MB/s |
| Fast C | 384 MB/s | 492 MB/s |
| Assembly | 385 MB/s | 501 MB/s |
We can also look at the speed delta for different block sizes of silesia
using `zstd -b1e1r silesia.tar -B#`.
| Decoder | -B1K ∆ | -B2K ∆ | -B4K ∆ | -B8K ∆ | -B16K ∆ | -B32K ∆ | -B64K ∆ | -B128K ∆ |
|----------|--------|--------|--------|--------|---------|---------|---------|----------|
| Fast C | +11.2% | +8.2% | +6.1% | +4.4% | +2.7% | +1.5% | +0.6% | +0.2% |
| Assembly | +12.5% | +9.0% | +6.2% | +3.6% | +1.5% | +0.7% | +0.2% | +0.03% |
gcc in the linux kernel was not unrolling the inner loops of the Huffman
decoder, which was destroying decoding performance. The compiler was
generating crazy code with all sorts of branches. I suspect because of
Spectre mitigations, but I'm not certain. Once the loops were manually
unrolled, performance was restored.
Additionally, when gcc couldn't prove that the variable left shift in
the 4X2 decode loop wasn't greater than 63, it inserted checks to verify
it. To fix this, mask `entry.nbBits & 0x3F`, which allows gcc to eliete
this check. This is a no op, because `entry.nbBits` is guaranteed to be
less than 64.
Lastly, introduce the `HUF_DISABLE_FAST_DECODE` macro to disable the
fast C loops for Issue #3762. So if even after this change, there is a
performance regression, users can opt-out at compile time.
ddidderr
Pavel P