zstd-rs/rust/README.md

# Rust rewrite

This directory contains the in-progress Rust replacement for the zstd library
and command-line program. During the migration, the crate is built as a static
library and linked into the original C test programs. Production C translation
units become declaration-only shims as their implementations move to Rust; the
original C tests remain unchanged and provide compatibility coverage.

## Component map

The crate is organized from low-level representation helpers toward the public
zstd ABI:

- Common primitives
  - `mem`, `bits`, `bitstream`, and `cpu` implement byte-order, bitstream, and
    target-feature operations used by the codecs.
  - `errors`, `debug`, `xxhash`, and `zstd_common` provide common exported ABI
    functions and state.
  - `common` contains shared frame constants and internal data types.
- Entropy coding
  - `entropy_common` reads FSE normalized counts and Huffman statistics.
  - `fse_decompress` builds FSE decoding tables and decodes FSE streams.
  - `fse_compress` normalizes counts, writes FSE headers, builds compression
    tables, and encodes FSE streams.
  - `huf_compress` builds Huffman compression tables, writes table headers,
    and encodes one- and four-stream Huffman payloads.
  - `huf_decompress` builds Huffman decoding tables and decodes X1 and X2
    Huffman streams.
- Compression primitives
  - `hist` counts byte frequencies for FSE and Huffman compression.
  - `zstd_presplit` chooses split points for full compression blocks.
  - `zstd_compress_literals` emits raw, RLE, and Huffman literal sections
    while preserving the compressor's Huffman-table repeat state.
  - `zstd_compress_stats` converts stored sequences into symbol codes,
    selects each block's symbol encoding types, compresses a seqStore's
    literals and sequences into a compressed-block body, builds the block
    entropy statistics shared with the superblock writer and the block
    splitter, and exports collected sequences in the public `ZSTD_Sequence`
    format.  Its C shims extract the sequence store, the entropy-table
    leaves, and the two `ZSTD_CCtx_params` scalars these paths read.
  - `zstd_compress_block_split` searches for profitable sequence-store
    partitions, while `zstd_compress` emits those partitions through the
    Rust single-block serializer. C retains split discovery's context setup
    and the outer block-dispatch decision. `zstd_compress` also owns ordinary
    sequence-block entropy emission, sequence collection, the legacy RLE
    compatibility gate, sequence-store construction, and its branch/fallback
    policy; C supplies only the private matchfinder, LDM
    block-preparation/consumption, external-sequence-producer, and
    state-preparation callbacks. Rust also owns the frame-chunk
    block loop, including block sizing, target/split/internal dispatch, output
    accounting, and frame-state
    updates; C supplies the private block-compression callbacks. Rust also
    owns the single-threaded buffered/stable stream state machine, including
    direct versus buffered output, pending-output draining, and frame reset
    policy. The external-sequence-and-literals block loop and public sequence
    conversion are Rust-owned as well; Rust also owns overflow-correction
    branch/order while C retains the private window, workspace, and index
    callbacks and only CCtx-facing adapters.
  - `zstd_compress_frame` serializes frame headers, skippable frames, and the
    last empty block; it takes scalar frame parameters so the C-owned
    `ZSTD_CCtx_params` layout never crosses the language boundary.
  - `zstd_compress` also owns the transparent single-threaded and multithreaded
    stream-initialization policy: dictionary selection, parameter resolution,
    initial buffer sizing, and the ordered setup decisions are projected into
    Rust while C retains the private contexts and mutation callbacks.
    The public sequence APIs likewise use Rust-owned validation, frame-header,
    checksum, and output-accounting orchestration around C-owned block state.
    The public end-of-frame path and the `ZSTD_compress2_c` fallback also use
    Rust-owned orchestration boundaries while C retains the context reset,
    stream adapter, checksum/epilogue, and trace callbacks.
  - `zstd_compress_params` owns the compression-level tables (formerly
    `clevels.h`), parameter bounds, clamping, validation, table selection,
    source/dictionary adjustment, and match-state/CDict size estimation.
    The C integration layer keeps the public `ZSTD_*` symbols and feeds the
    leaves configuration-owned scalars: the excluded-block-compressor
    strategy cascade, struct sizes, and sanitizer redzone policy.
  - `zstd_fast` and `zstd_double_fast` implement the single- and two-table
    fast block match finders, including attached and external dictionary paths.
  - `zstd_lazy` implements greedy, lazy, lazy2, and binary-tree matching,
    including row-based and dictionary search variants.
  - `zstd_opt_tree` maintains the binary-tree index used by optimal matching;
    `zstd_opt` owns the dynamic-programming price model, optimal parse, and
    sequence emission.
  - `zstd_ldm` implements long-distance-match parameter selection, table
    maintenance, sequence generation, sequence consumption, and sequence-store
    encoding.
- Dictionary building
  - `divsufsort` constructs the suffix array that drives the legacy `ZDICT`
    trainer (`ZDICT_trainFromBuffer_legacy`); `dict_builder_zdict`,
    `dict_builder_cover`, and `dict_builder_fastcover` own the sample analysis,
    training, and dictionary assembly.  The corresponding C translation units
    are declaration-only ABI shims.
- Runtime support
  - `threading` provides platform pthread wrappers required by zstd headers.
  - `pool` implements the bounded worker pool used by multithreaded compression.
  - The `zstdmt_compress` integration keeps job descriptors and synchronization
    private to C while Rust owns input-retention scans, reusable input-range
    overlap decisions, outer scheduling and end-directive adjustments,
    job-creation decisions, compression-job stage sequencing and error flow,
    pending-output decisions through scalar job projections, and frame
    progression's job-ring scan, error normalization, and active-worker
    accounting.  Rust also owns frame-block preparation ordering and the MT
    serial turn/skip policy, including LDM-before-checksum sequencing; C
    callbacks retain the private match-state window, workspace operations,
    synchronization, LDM state, and checksum state.
- Dictionary support
  - `zstd_ddict` owns, loads, copies, and references decode dictionaries.
- Legacy decoding
  - `legacy` hosts one frozen module per historical format; `legacy::zstd_v01`
    through `legacy::zstd_v07` port the seven self-contained historical
    decoders.  Their original C translation units remain declaration-only
    shims for the native build.
- Block decompression
  - `zstd_decompress_block` decodes literal and sequence sections, maintains
    FSE/Huffman repeat state, executes compressed-block sequences, and selects
    the short or long sequence decoder from projected configuration and history
    state. C retains the decoder-context layout and configuration projection;
    the public and fullbench block-decoder wrappers are Rust-owned.
  - `zstd_decompress` owns the public decompression context, one-shot,
    dictionary, parameter, and streaming state machines.  Its C shim retains
    the configuration-dependent context layout and platform details, plus
    legacy and trace leaves; Rust owns decoder storage allocation, custom
    memory dispatch, and the `ZSTD_copyDCtx` prefix copy up to the projected
    `inBuff` field.
- Command-line frontend
  - `zstd_cli` owns the Rust parser, safety policy, and dispatch.  It is built
    by the separate `cli/` static-library package only for program archives,
    so library builds do not acquire program-only dependencies.  The C
    `fileio` layer retains the format-specific codec callbacks, private
    asynchronous-pool adapters, metadata, zstd codec/error mapping, and
    adaptive-policy integration, while Rust owns the mixed-format probe/dispatch
    loop, zstd stream-compression I/O loop, scalar adaptive decisions,
    optional-format decompression loops, decompression result policy/final
    accounting, both shared- and separate-destination multi-file compression
    schedulers, and both shared- and separate-destination decompression
    schedulers. C retains destination-name construction and the private
    file/resource/format callbacks for these scheduler boundaries.
    Rust already owns the file preference policy, filename decisions,
    source/destination opening, dictionary buffers, asynchronous I/O pools,
    and pass-through copy leaf.
  - `timefn` provides the monotonic nanosecond clock behind `UTIL_time_t`,
    while `benchfn` owns the benchmark run/timing loop (`BMK_benchFunction`,
    `BMK_benchTimedFn`) and `benchzstd` owns benchmark orchestration and
    reporting. Both live in Rust; the C translation units are ABI shims.
    C test binaries (fullbench, fuzzer, zstreamtest, paramgrill, ...) link a
    helpers-only build of the archive, produced without the package's `cli`
    feature, because the parser layer requires the C `fileio` backend that
    tests do not compile.

Dictionary-ingestion dispatch, CCtx dictionary/prefix attachment dispatch, the
CLI zstd compression stream loop, scalar adaptive decisions, and the
shared- and separate-destination multi-file compression schedulers,
shared- and separate-destination decompression scheduling,
single-threaded stream initialization and the buffered/stable stream state
machine, MT stream initialization, MT outer scheduling and flush policy, MT
compression-job stage sequencing and error flow, MT frame-progression job
aggregation, frame-block preparation ordering, MT serial turn/skip and
LDM/checksum sequencing, overflow-correction policy/order, public sequence-API
orchestration, sequence-store and block policy, external-producer invocation
and success-path validation, external-sequence-store reset,
external-sequence/literals block loop, optional-format decompression loops,
block decoder wrappers, decompression result policy, and public end-of-frame
compression orchestration, public CCtx-copy pledge/frame-parameter policy,
public CCtx allocation validation and allocate-before-init ordering, public
static-CCtx workspace validation and initialization dispatch, and public CDict
constructor parameter selection, default-level normalization, and
workspace/object teardown ordering, public static-CDict workspace validation
and initialization dispatch, public advanced-CDict one-shot validation and
begin/end ordering, advanced-CDict parameter selection, dedicated-search
fallback, row-matchfinder resolution, and create/init/failure ordering,
public advanced-compression parameter validation/init policy, public
usingDict parameter selection, dictionary-presence handling, and default-level
normalization, and public usingCDict frame-policy construction and begin/end
sequencing, legacy public CDict-begin frame-policy construction and
unknown-source pledge, and the compressBegin_usingDict family’s unknown-source
parameter selection and default-level normalization, and public advanced-begin
parameter validation and init-then-begin ordering now run in Rust.
CDict advanced private workspace construction, private static-CCtx and
static-CDict workspace construction and dictionary-content allocation/loading,
and advanced-CDict dictionary-content loading remain in C. Rust now owns
advanced-CDict custom-memory validation, workspace-size query/allocation,
allocation/create/init cleanup ordering, the CCtx workspace-size formula, the
scalar CCtx-reset plan, post-match-state storage reservation order, and
match-state reset policy/order; C retains
private layout-size inputs, workspace resize/layout, private field publication,
and allocator callbacks. Private CCtx reset/matchfinder/workspace operations
and codec/adaptive-policy
callbacks remain in C. CDict initialization ordering and scalar publication,
shared compression-begin dictionary selection, CDict reset attach-versus-copy
selection, and CDict-begin parameter selection, initialization ordering, and
source-window policy now run in Rust behind private-state bridges; their C
callbacks and final begin/reset/attach/copy operations remain. The remaining
C paths must move before the rewrite is complete.
Keeping that boundary explicit prevents a passing hybrid build from being
mistaken for the final all-Rust result.

## Legacy decoding

Each `lib/legacy/zstd_v0N.c` file is a frozen snapshot of the entropy coders
and frame logic of one historical release. The Rust ports in `src/legacy/`
keep that property: every version owns its own frozen FSE/Huff0 and frame
logic, ported line by line, and must never reuse the modern entropy modules
or share code with other legacy versions. Outputs and error codes must be
byte-identical to the original C files. Their only shared dependency is the
`errors` module, matching the C files' `error_private.h` include.

Cargo features `legacy-v01` .. `legacy-v07` gate the per-version modules and
are never default features. All seven modules are now available; the build
systems derive the enabled feature list from the C configuration:

- `lib/Makefile` and `programs/Makefile` map `ZSTD_LEGACY_SUPPORT=N` to the
  features for versions >= N (0 disables legacy), matching the
  `ZSTD_LEGACY_FILES` selection in `lib/libzstd.mk`.
- `tests/Makefile` always enables all seven features because the test
  objects compile every `lib/legacy/*.c` file regardless of dispatch level.
- `build/meson` maps `legacy_level` like the makefiles; `build/cmake`
  enables all seven whenever `ZSTD_LEGACY_SUPPORT` is on because it always
  compiles all seven C files.

Every build system also encodes the legacy selection in the Rust target
directory name (for example `c1-d1-default-legacy5`), for the same reason the
HUF mode is encoded there: a cached archive built for one configuration must
never be linked into a build that expects another.

Each port adds `src/legacy/zstd_v0N.rs`, registers it in `src/legacy/mod.rs`
behind its feature, and reduces `lib/legacy/zstd_v0N.c` to a
declaration-only shim. For v0.1 the streaming `ZSTDv01_Dctx` state lives
entirely in Rust: C code only ever holds an opaque pointer, so the C-side
struct definition is gone.

## Compatibility boundary

The public ABI continues to come from the existing headers under `lib/`.
Exported Rust functions therefore use C layout and calling conventions. A C
source file whose implementation has moved to Rust remains in the original
makefile source list as a small shim so header configuration and platform
preprocessor behavior stay available during the transition.

The library, test, and program makefiles select an archive directory for the
active C configuration: enabled compression/decompression/dictionary-builder
modules, default or forced HUF X1/X2, and the matching Rust target for 32-bit
C binaries. The
native static archive flattens Rust object members rather than nesting a Rust
archive, while the native shared library retains all migrated Rust exports.
When the HUF mode changes, the test and program paths also rebuild cached C
outputs before linking. This prevents original C tests from using a stale or
configuration-incompatible implementation.

## Validation

Run focused Rust checks from this directory:

```sh
cargo fmt --check
cargo clippy --all-targets -- -D warnings
cargo test --all-targets
cargo build --release
```

The program-only Rust archive has its own feature matrix and should be checked
from `rust/cli` as well:

```sh
cargo clippy --all-targets -- -D warnings
cargo test --all-targets
cargo test --no-default-features --features cli,compression --all-targets
cargo test --no-default-features --features cli,decompression --all-targets
cargo test --no-default-features --all-targets
```

Then run original compatibility tests from the repository root, starting with
the narrow target for the component being migrated. For example:

```sh
make -C tests fuzzer
./tests/fuzzer -i1 --no-big-tests
make -C tests test-rust-lib-smoke
```

Broader `tests/Makefile` targets remain the authoritative integration gates as
more of the library and CLI are rewritten.