ddidderr/zstd-rs
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
Files
25f2aa9502cd49c7b9a60d3e67693a99bbeca606
zstd-rs/rust/src/huf_decompress.rs
T
ddidderr 27932113fd feat(rust): port Huffman decompression 
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
2026-07-10 20:48:51 +02:00

1398 lines
43 KiB
Rust
Raw Blame History

#![allow(non_snake_case)]
//! Huffman decoding tables and the single/four-stream HUF decoders.
//!
//! This is the Rust implementation of `lib/decompress/huf_decompress.c`.
//! The C version has a second, architecture-specific fast path. The table
//! format and the portable decoder below are shared by both implementations;
//! retaining the portable path keeps the C ABI and bitstream behaviour while
//! avoiding unchecked SIMD/assembly reads in the migration layer.
use crate::bitstream::{
BIT_DStream_status, BIT_DStream_t, BIT_endOfDStream, BIT_initDStream, BIT_lookBitsFast,
BIT_reloadDStream, BIT_skipBits,
};
use crate::entropy_common::{HUF_readStats_wksp, HUF_TABLELOG_MAX};
use crate::errors::{ERR_isError, ZstdErrorCode, ERROR};
use crate::mem::{MEM_32bits, MEM_64bits, MEM_readLE16, MEM_write16};
use std::os::raw::{c_int, c_uint, c_void};
#[cfg(all(
feature = "huf-force-decompress-x1",
feature = "huf-force-decompress-x2"
))]
compile_error!("Huffman X1 and X2 decoder force features are mutually exclusive");
const HUF_SYMBOLVALUE_MAX: usize = 255;
const HUF_DECODER_FAST_TABLELOG: u32 = 11;
const HUF_READ_STATS_WORKSPACE_SIZE_U32: usize = 219;
/* This is the C macro sequence used by both portable fast decode loops:
*
* X2, X1, X2, X0
*
* where X2 only runs on 64-bit targets, X1 runs on 64-bit targets (or with
* HUF_TABLELOG_MAX <= 12), and X0 always runs. Keeping this as data makes
* the 32-bit behaviour explicit: two decodes, not four. */
#[inline]
fn huf_fast_sequence_slots(is_64_bits: bool) -> [bool; 4] {
[
is_64_bits,
is_64_bits || HUF_TABLELOG_MAX <= 12,
is_64_bits,
true,
]
}
#[derive(Copy, Clone, Debug, PartialEq, Eq)]
#[allow(dead_code)] // Each variant is constructed only in its feature-selected build.
enum HufForcedDecoder {
X1,
X2,
}
/* `lib/libzstd.mk` maps HUF_FORCE_DECOMPRESS_X{1,2} to these mutually
* exclusive Cargo features. The C and Rust halves must select the same table
* layout, especially in minified builds where X1 is forced by default. */
#[cfg(all(
feature = "huf-force-decompress-x1",
not(feature = "huf-force-decompress-x2")
))]
const fn huf_forced_decoder() -> Option<HufForcedDecoder> {
Some(HufForcedDecoder::X1)
}
#[cfg(all(
feature = "huf-force-decompress-x2",
not(feature = "huf-force-decompress-x1")
))]
const fn huf_forced_decoder() -> Option<HufForcedDecoder> {
Some(HufForcedDecoder::X2)
}
#[cfg(not(any(
feature = "huf-force-decompress-x1",
feature = "huf-force-decompress-x2"
)))]
const fn huf_forced_decoder() -> Option<HufForcedDecoder> {
None
}
/* These are the C `sizeof()` values on every supported ABI: all fields are
* byte or four-byte aligned. The caller-owned workspace is part of the C
* contract, so preserve the exact rejection thresholds even though the Rust
* implementation keeps its scratch state on the stack. */
const HUF_READ_DTABLE_X1_WORKSPACE_SIZE: usize = (HUF_TABLELOG_MAX as usize + 1) * 4
+ (HUF_TABLELOG_MAX as usize + 1) * 4
+ HUF_READ_STATS_WORKSPACE_SIZE_U32 * 4
+ (HUF_SYMBOLVALUE_MAX + 1)
+ (HUF_SYMBOLVALUE_MAX + 1);
const HUF_READ_DTABLE_X2_WORKSPACE_SIZE: usize =
HUF_TABLELOG_MAX as usize * (HUF_TABLELOG_MAX as usize + 1) * 4
+ (HUF_TABLELOG_MAX as usize + 1) * 4
+ (HUF_TABLELOG_MAX as usize + 3) * 4
+ (HUF_SYMBOLVALUE_MAX + 1)
+ (HUF_SYMBOLVALUE_MAX + 1)
+ HUF_READ_STATS_WORKSPACE_SIZE_U32 * 4;
#[derive(Copy, Clone, Debug)]
struct DTableDesc {
max_table_log: u8,
table_type: u8,
table_log: u8,
reserved: u8,
}
#[repr(C)]
#[derive(Copy, Clone)]
struct HufDEltX2 {
sequence: u16,
nb_bits: u8,
length: u8,
}
#[inline]
unsafe fn get_dtable_desc(table: *const u32) -> DTableDesc {
let bytes = table.cast::<u8>();
DTableDesc {
max_table_log: bytes.read(),
table_type: bytes.add(1).read(),
table_log: bytes.add(2).read(),
reserved: bytes.add(3).read(),
}
}
#[inline]
unsafe fn set_dtable_desc(table: *mut u32, desc: DTableDesc) {
let bytes = table.cast::<u8>();
bytes.write(desc.max_table_log);
bytes.add(1).write(desc.table_type);
bytes.add(2).write(desc.table_log);
bytes.add(3).write(desc.reserved);
}
#[inline]
unsafe fn x1_entry(table: *const u32, index: usize) -> (u8, u8) {
let entry = table.add(1).cast::<u8>().add(index * 2);
(entry.read(), entry.add(1).read())
}
#[inline]
unsafe fn set_x1_entry(table: *mut u32, index: usize, nb_bits: u8, symbol: u8) {
let entry = table.add(1).cast::<u8>().add(index * 2);
entry.write(nb_bits);
entry.add(1).write(symbol);
}
#[inline]
unsafe fn x2_entry(table: *const u32, index: usize) -> HufDEltX2 {
table.add(1).cast::<HufDEltX2>().add(index).read_unaligned()
}
#[inline]
unsafe fn set_x2_raw(table: *mut u32, index: usize, raw: u32) {
table
.add(1)
.cast::<HufDEltX2>()
.add(index)
.cast::<u32>()
.write_unaligned(raw);
}
#[inline]
fn huf_rescale_stats(
huff_weight: &mut [u8; HUF_SYMBOLVALUE_MAX + 1],
rank_val: &mut [u32; HUF_TABLELOG_MAX as usize + 1],
nb_symbols: usize,
table_log: u32,
target_table_log: u32,
) -> u32 {
if table_log > target_table_log {
return table_log;
}
if table_log < target_table_log {
let scale = target_table_log - table_log;
for weight in huff_weight.iter_mut().take(nb_symbols) {
if *weight != 0 {
*weight = weight.wrapping_add(scale as u8);
}
}
for symbol in ((scale + 1)..=target_table_log).rev() {
rank_val[symbol as usize] = rank_val[(symbol - scale) as usize];
}
for symbol in (1..=scale).rev() {
rank_val[symbol as usize] = 0;
}
}
target_table_log
}
#[no_mangle]
pub unsafe extern "C" fn HUF_readDTableX1_wksp(
dtable: *mut u32,
src: *const c_void,
src_size: usize,
_work_space: *mut c_void,
wksp_size: usize,
flags: c_int,
) -> usize {
if wksp_size < HUF_READ_DTABLE_X1_WORKSPACE_SIZE {
return ERROR(ZstdErrorCode::TableLogTooLarge);
}
let mut rank_val = [0u32; HUF_TABLELOG_MAX as usize + 1];
let mut huff_weight = [0u8; HUF_SYMBOLVALUE_MAX + 1];
let mut stats_wksp = [0u32; HUF_READ_STATS_WORKSPACE_SIZE_U32];
let mut nb_symbols = 0u32;
let mut table_log = 0u32;
let i_size = HUF_readStats_wksp(
huff_weight.as_mut_ptr(),
huff_weight.len(),
rank_val.as_mut_ptr(),
&mut nb_symbols,
&mut table_log,
src,
src_size,
stats_wksp.as_mut_ptr().cast::<c_void>(),
std::mem::size_of_val(&stats_wksp),
flags,
);
if ERR_isError(i_size) {
return i_size;
}
let mut desc = get_dtable_desc(dtable);
let max_table_log = u32::from(desc.max_table_log) + 1;
let target_table_log = max_table_log.min(HUF_DECODER_FAST_TABLELOG);
table_log = huf_rescale_stats(
&mut huff_weight,
&mut rank_val,
nb_symbols as usize,
table_log,
target_table_log,
);
if table_log > max_table_log {
return ERROR(ZstdErrorCode::TableLogTooLarge);
}
desc.table_type = 0;
desc.table_log = table_log as u8;
set_dtable_desc(dtable, desc);
let mut rank_start = [0u32; HUF_TABLELOG_MAX as usize + 1];
let mut next_rank_start = 0u32;
for weight in 0..=table_log as usize {
rank_start[weight] = next_rank_start;
next_rank_start = next_rank_start.wrapping_add(rank_val[weight]);
}
let mut symbols = [0u8; HUF_SYMBOLVALUE_MAX + 1];
for (symbol, weight) in huff_weight
.iter()
.copied()
.take(nb_symbols as usize)
.enumerate()
{
let weight = weight as usize;
let index = rank_start[weight] as usize;
symbols[index] = symbol as u8;
rank_start[weight] = rank_start[weight].wrapping_add(1);
}
let mut symbol = rank_val[0] as usize;
let mut rank_start_index = 0usize;
for (weight, symbol_count) in rank_val
.iter()
.copied()
.take(table_log as usize + 1)
.enumerate()
.skip(1)
{
let symbol_count = symbol_count as usize;
let length = (1usize << weight) >> 1;
let nb_bits = (table_log as usize + 1 - weight) as u8;
for item in 0..symbol_count {
let value = symbols[symbol + item];
for entry in 0..length {
set_x1_entry(dtable, rank_start_index + entry, nb_bits, value);
}
rank_start_index += length;
}
symbol += symbol_count;
}
i_size
}
#[inline]
fn build_delt_x2_u32(symbol: u32, nb_bits: u32, base_seq: u16, level: u32) -> u32 {
if cfg!(target_endian = "little") {
let sequence = if level == 1 {
symbol
} else {
u32::from(base_seq) + (symbol << 8)
};
sequence + (nb_bits << 16) + (level << 24)
} else {
let sequence = if level == 1 {
symbol << 8
} else {
(u32::from(base_seq) << 8) + symbol
};
(sequence << 16) + (nb_bits << 8) + level
}
}
#[allow(clippy::too_many_arguments)]
unsafe fn fill_dtable_x2_for_weight(
dtable: *mut u32,
table_offset: usize,
sorted_symbols: &[u8; HUF_SYMBOLVALUE_MAX + 1],
begin: usize,
end: usize,
nb_bits: u32,
table_log: u32,
base_seq: u16,
level: u32,
) {
let length = 1usize << (table_log - nb_bits);
let mut out = table_offset;
for symbol in sorted_symbols.iter().take(end).skip(begin) {
let entry = build_delt_x2_u32(u32::from(*symbol), nb_bits, base_seq, level);
for _ in 0..length {
set_x2_raw(dtable, out, entry);
out += 1;
}
}
}
#[allow(clippy::too_many_arguments)]
unsafe fn fill_dtable_x2_level2(
dtable: *mut u32,
table_offset: usize,
target_log: u32,
consumed_bits: u32,
rank_val: &[u32; HUF_TABLELOG_MAX as usize + 1],
min_weight: usize,
max_weight1: usize,
sorted_symbols: &[u8; HUF_SYMBOLVALUE_MAX + 1],
rank_start: &[u32; HUF_TABLELOG_MAX as usize + 3],
nb_bits_baseline: u32,
base_seq: u16,
) {
if min_weight > 1 {
let skip_size = rank_val[min_weight] as usize;
let entry = build_delt_x2_u32(u32::from(base_seq), consumed_bits, 0, 1);
for index in 0..skip_size {
set_x2_raw(dtable, table_offset + index, entry);
}
}
for weight in min_weight..max_weight1 {
let begin = rank_start[weight] as usize;
let end = rank_start[weight + 1] as usize;
let nb_bits = nb_bits_baseline - weight as u32;
fill_dtable_x2_for_weight(
dtable,
table_offset + rank_val[weight] as usize,
sorted_symbols,
begin,
end,
nb_bits + consumed_bits,
target_log,
base_seq,
2,
);
}
}
unsafe fn fill_dtable_x2(
dtable: *mut u32,
target_log: u32,
sorted_symbols: &[u8; HUF_SYMBOLVALUE_MAX + 1],
rank_start: &[u32; HUF_TABLELOG_MAX as usize + 3],
rank_val_origin: &[[u32; HUF_TABLELOG_MAX as usize + 1]; HUF_TABLELOG_MAX as usize],
max_weight: u32,
nb_bits_baseline: u32,
) {
let rank_val = &rank_val_origin[0];
let scale_log = nb_bits_baseline as i32 - target_log as i32;
let min_bits = nb_bits_baseline - max_weight;
for weight in 1..=max_weight as usize {
let begin = rank_start[weight] as usize;
let end = rank_start[weight + 1] as usize;
let nb_bits = nb_bits_baseline - weight as u32;
if target_log - nb_bits >= min_bits {
let mut start = rank_val[weight] as usize;
let length = 1usize << (target_log - nb_bits);
let mut min_weight = nb_bits as i32 + scale_log;
if min_weight < 1 {
min_weight = 1;
}
for symbol in sorted_symbols.iter().take(end).skip(begin) {
fill_dtable_x2_level2(
dtable,
start,
target_log,
nb_bits,
&rank_val_origin[nb_bits as usize],
min_weight as usize,
max_weight as usize + 1,
sorted_symbols,
rank_start,
nb_bits_baseline,
u16::from(*symbol),
);
start += length;
}
} else {
fill_dtable_x2_for_weight(
dtable,
rank_val[weight] as usize,
sorted_symbols,
begin,
end,
nb_bits,
target_log,
0,
1,
);
}
}
}
#[no_mangle]
pub unsafe extern "C" fn HUF_readDTableX2_wksp(
dtable: *mut u32,
src: *const c_void,
src_size: usize,
_work_space: *mut c_void,
wksp_size: usize,
flags: c_int,
) -> usize {
if wksp_size < HUF_READ_DTABLE_X2_WORKSPACE_SIZE {
return ERROR(ZstdErrorCode::Generic);
}
let mut desc = get_dtable_desc(dtable);
let mut max_table_log = u32::from(desc.max_table_log);
if max_table_log > HUF_TABLELOG_MAX {
return ERROR(ZstdErrorCode::TableLogTooLarge);
}
let mut rank_stats = [0u32; HUF_TABLELOG_MAX as usize + 1];
let mut weight_list = [0u8; HUF_SYMBOLVALUE_MAX + 1];
let mut stats_wksp = [0u32; HUF_READ_STATS_WORKSPACE_SIZE_U32];
let mut nb_symbols = 0u32;
let mut table_log = 0u32;
let i_size = HUF_readStats_wksp(
weight_list.as_mut_ptr(),
weight_list.len(),
rank_stats.as_mut_ptr(),
&mut nb_symbols,
&mut table_log,
src,
src_size,
stats_wksp.as_mut_ptr().cast::<c_void>(),
std::mem::size_of_val(&stats_wksp),
flags,
);
if ERR_isError(i_size) {
return i_size;
}
if table_log > max_table_log {
return ERROR(ZstdErrorCode::TableLogTooLarge);
}
if table_log <= HUF_DECODER_FAST_TABLELOG && max_table_log > HUF_DECODER_FAST_TABLELOG {
max_table_log = HUF_DECODER_FAST_TABLELOG;
}
let mut max_weight = table_log as usize;
while rank_stats[max_weight] == 0 {
max_weight -= 1;
}
/* `rank_start` deliberately has the leading pad from the C workspace.
* The table builder consumes the padded view after the sorting pass. */
let mut rank_start = [0u32; HUF_TABLELOG_MAX as usize + 3];
let mut next_rank_start = 0u32;
for weight in 1..=max_weight {
let current = next_rank_start;
next_rank_start = next_rank_start.wrapping_add(rank_stats[weight]);
rank_start[weight + 1] = current;
}
rank_start[1] = next_rank_start;
rank_start[max_weight + 2] = next_rank_start;
let mut sorted_symbols = [0u8; HUF_SYMBOLVALUE_MAX + 1];
for (symbol, weight) in weight_list
.iter()
.copied()
.take(nb_symbols as usize)
.enumerate()
{
let weight = weight as usize;
let index = rank_start[weight + 1] as usize;
sorted_symbols[index] = symbol as u8;
rank_start[weight + 1] = rank_start[weight + 1].wrapping_add(1);
}
rank_start[1] = 0;
let mut rank_val = [[0u32; HUF_TABLELOG_MAX as usize + 1]; HUF_TABLELOG_MAX as usize];
let rescale = max_table_log as i32 - table_log as i32 - 1;
let mut next_rank_val = 0u32;
for weight in 1..=max_weight {
let current = next_rank_val;
let shift = (weight as i32 + rescale) as u32;
next_rank_val = next_rank_val.wrapping_add(rank_stats[weight] << shift);
rank_val[0][weight] = current;
}
let min_bits = table_log + 1 - max_weight as u32;
let base_rank_val = rank_val[0];
for consumed in min_bits..(max_table_log - min_bits + 1) {
for (target, source) in rank_val[consumed as usize]
.iter_mut()
.zip(base_rank_val)
.take(max_weight + 1)
.skip(1)
{
*target = source >> consumed;
}
}
fill_dtable_x2(
dtable,
max_table_log,
&sorted_symbols,
&rank_start,
&rank_val,
max_weight as u32,
table_log + 1,
);
desc.table_log = max_table_log as u8;
desc.table_type = 1;
set_dtable_desc(dtable, desc);
i_size
}
#[inline]
unsafe fn decode_symbol_x1(
bit_stream: *mut BIT_DStream_t,
dtable: *const u32,
table_log: u32,
) -> u8 {
let value = BIT_lookBitsFast(bit_stream, table_log);
let (nb_bits, symbol) = x1_entry(dtable, value);
BIT_skipBits(bit_stream, u32::from(nb_bits));
symbol
}
#[inline]
unsafe fn decode_x1_fast_sequence(
mut out: *mut u8,
bit_stream: *mut BIT_DStream_t,
dtable: *const u32,
table_log: u32,
is_64_bits: bool,
) -> *mut u8 {
let slots = huf_fast_sequence_slots(is_64_bits);
if slots[0] {
*out = decode_symbol_x1(bit_stream, dtable, table_log);
out = out.add(1);
}
if slots[1] {
*out = decode_symbol_x1(bit_stream, dtable, table_log);
out = out.add(1);
}
if slots[2] {
*out = decode_symbol_x1(bit_stream, dtable, table_log);
out = out.add(1);
}
if slots[3] {
*out = decode_symbol_x1(bit_stream, dtable, table_log);
out = out.add(1);
}
out
}
unsafe fn decode_stream_x1(
mut out: *mut u8,
out_end: *mut u8,
bit_stream: *mut BIT_DStream_t,
dtable: *const u32,
table_log: u32,
) -> usize {
let out_start = out;
if out_end.offset_from(out) > 3 {
while BIT_reloadDStream(bit_stream) == BIT_DStream_status::Unfinished
&& out < out_end.sub(3)
{
out = decode_x1_fast_sequence(out, bit_stream, dtable, table_log, MEM_64bits());
}
} else {
BIT_reloadDStream(bit_stream);
}
if MEM_32bits() {
while BIT_reloadDStream(bit_stream) == BIT_DStream_status::Unfinished && out < out_end {
*out = decode_symbol_x1(bit_stream, dtable, table_log);
out = out.add(1);
}
}
while out < out_end {
*out = decode_symbol_x1(bit_stream, dtable, table_log);
out = out.add(1);
}
out.offset_from(out_start) as usize
}
#[inline]
unsafe fn decode_symbol_x2(
out: *mut u8,
bit_stream: *mut BIT_DStream_t,
dtable: *const u32,
table_log: u32,
) -> usize {
let value = BIT_lookBitsFast(bit_stream, table_log);
let entry = x2_entry(dtable, value);
MEM_write16(out.cast::<c_void>(), entry.sequence);
BIT_skipBits(bit_stream, u32::from(entry.nb_bits));
entry.length as usize
}
#[inline]
unsafe fn decode_x2_fast_sequence(
mut out: *mut u8,
bit_stream: *mut BIT_DStream_t,
dtable: *const u32,
table_log: u32,
is_64_bits: bool,
) -> *mut u8 {
let slots = huf_fast_sequence_slots(is_64_bits);
if slots[0] {
out = out.add(decode_symbol_x2(out, bit_stream, dtable, table_log));
}
if slots[1] {
out = out.add(decode_symbol_x2(out, bit_stream, dtable, table_log));
}
if slots[2] {
out = out.add(decode_symbol_x2(out, bit_stream, dtable, table_log));
}
if slots[3] {
out = out.add(decode_symbol_x2(out, bit_stream, dtable, table_log));
}
out
}
#[inline]
unsafe fn decode_last_symbol_x2(
out: *mut u8,
bit_stream: *mut BIT_DStream_t,
dtable: *const u32,
table_log: u32,
) -> usize {
let value = BIT_lookBitsFast(bit_stream, table_log);
let entry = x2_entry(dtable, value);
*out = if cfg!(target_endian = "little") {
entry.sequence as u8
} else {
(entry.sequence >> 8) as u8
};
if entry.length == 1 {
BIT_skipBits(bit_stream, u32::from(entry.nb_bits));
} else if (*bit_stream).bitsConsumed < usize::BITS {
BIT_skipBits(bit_stream, u32::from(entry.nb_bits));
if (*bit_stream).bitsConsumed > usize::BITS {
(*bit_stream).bitsConsumed = usize::BITS;
}
}
1
}
unsafe fn decode_stream_x2(
mut out: *mut u8,
out_end: *mut u8,
bit_stream: *mut BIT_DStream_t,
dtable: *const u32,
table_log: u32,
) -> usize {
let out_start = out;
let word_size = std::mem::size_of::<usize>();
if out_end.offset_from(out) as usize >= word_size {
if table_log <= HUF_DECODER_FAST_TABLELOG && MEM_64bits() {
while BIT_reloadDStream(bit_stream) == BIT_DStream_status::Unfinished
&& out_end.offset_from(out) > 9
{
for _ in 0..5 {
out = out.add(decode_symbol_x2(out, bit_stream, dtable, table_log));
}
}
} else {
while BIT_reloadDStream(bit_stream) == BIT_DStream_status::Unfinished
&& out < out_end.sub(word_size - 1)
{
out = decode_x2_fast_sequence(out, bit_stream, dtable, table_log, MEM_64bits());
}
}
} else {
BIT_reloadDStream(bit_stream);
}
if out_end.offset_from(out) as usize >= 2 {
while BIT_reloadDStream(bit_stream) == BIT_DStream_status::Unfinished
&& out <= out_end.sub(2)
{
out = out.add(decode_symbol_x2(out, bit_stream, dtable, table_log));
}
while out <= out_end.sub(2) {
out = out.add(decode_symbol_x2(out, bit_stream, dtable, table_log));
}
}
if out < out_end {
out = out.add(decode_last_symbol_x2(out, bit_stream, dtable, table_log));
}
out.offset_from(out_start) as usize
}
unsafe fn decompress_1x1_using_dtable(
dst: *mut c_void,
dst_size: usize,
src: *const c_void,
src_size: usize,
dtable: *const u32,
) -> usize {
let mut bit_stream = std::mem::zeroed::<BIT_DStream_t>();
let init = BIT_initDStream(&mut bit_stream, src, src_size);
if ERR_isError(init) {
return init;
}
let out = dst.cast::<u8>();
decode_stream_x1(
out,
out.add(dst_size),
&mut bit_stream,
dtable,
u32::from(get_dtable_desc(dtable).table_log),
);
if BIT_endOfDStream(&bit_stream) == 0 {
return ERROR(ZstdErrorCode::CorruptionDetected);
}
dst_size
}
unsafe fn decompress_1x2_using_dtable(
dst: *mut c_void,
dst_size: usize,
src: *const c_void,
src_size: usize,
dtable: *const u32,
) -> usize {
let mut bit_stream = std::mem::zeroed::<BIT_DStream_t>();
let init = BIT_initDStream(&mut bit_stream, src, src_size);
if ERR_isError(init) {
return init;
}
let out = dst.cast::<u8>();
decode_stream_x2(
out,
out.add(dst_size),
&mut bit_stream,
dtable,
u32::from(get_dtable_desc(dtable).table_log),
);
if BIT_endOfDStream(&bit_stream) == 0 {
return ERROR(ZstdErrorCode::CorruptionDetected);
}
dst_size
}
unsafe fn split_4x_input(
src: *const c_void,
src_size: usize,
) -> Result<([*const u8; 4], [usize; 4]), usize> {
if src_size < 10 {
return Err(ERROR(ZstdErrorCode::CorruptionDetected));
}
let input = src.cast::<u8>();
let length1 = MEM_readLE16(input.cast::<c_void>()) as usize;
let length2 = MEM_readLE16(input.add(2).cast::<c_void>()) as usize;
let length3 = MEM_readLE16(input.add(4).cast::<c_void>()) as usize;
let header_and_streams = match length1
.checked_add(length2)
.and_then(|value| value.checked_add(length3))
.and_then(|value| value.checked_add(6))
{
Some(value) if value <= src_size => value,
_ => return Err(ERROR(ZstdErrorCode::CorruptionDetected)),
};
let length4 = src_size - header_and_streams;
let stream1 = input.add(6);
let stream2 = stream1.add(length1);
let stream3 = stream2.add(length2);
let stream4 = stream3.add(length3);
Ok((
[stream1, stream2, stream3, stream4],
[length1, length2, length3, length4],
))
}
unsafe fn decompress_4x1_using_dtable(
dst: *mut c_void,
dst_size: usize,
src: *const c_void,
src_size: usize,
dtable: *const u32,
) -> usize {
if dst_size < 6 {
return ERROR(ZstdErrorCode::CorruptionDetected);
}
let (streams, lengths) = match split_4x_input(src, src_size) {
Ok(value) => value,
Err(error) => return error,
};
let segment_size = match dst_size.checked_add(3) {
Some(size) => size / 4,
None => return ERROR(ZstdErrorCode::CorruptionDetected),
};
let out = dst.cast::<u8>();
let table_log = u32::from(get_dtable_desc(dtable).table_log);
for stream in 0..4 {
let start = match segment_size.checked_mul(stream) {
Some(offset) if offset <= dst_size => out.add(offset),
_ => return ERROR(ZstdErrorCode::CorruptionDetected),
};
let end_offset = (segment_size.saturating_mul(stream + 1)).min(dst_size);
let end = out.add(end_offset);
let mut bit_stream = std::mem::zeroed::<BIT_DStream_t>();
let init = BIT_initDStream(
&mut bit_stream,
streams[stream].cast::<c_void>(),
lengths[stream],
);
if ERR_isError(init) {
return init;
}
decode_stream_x1(start, end, &mut bit_stream, dtable, table_log);
if BIT_endOfDStream(&bit_stream) == 0 {
return ERROR(ZstdErrorCode::CorruptionDetected);
}
}
dst_size
}
unsafe fn decompress_4x2_using_dtable(
dst: *mut c_void,
dst_size: usize,
src: *const c_void,
src_size: usize,
dtable: *const u32,
) -> usize {
if dst_size < 6 {
return ERROR(ZstdErrorCode::CorruptionDetected);
}
let (streams, lengths) = match split_4x_input(src, src_size) {
Ok(value) => value,
Err(error) => return error,
};
let segment_size = match dst_size.checked_add(3) {
Some(size) => size / 4,
None => return ERROR(ZstdErrorCode::CorruptionDetected),
};
let out = dst.cast::<u8>();
let table_log = u32::from(get_dtable_desc(dtable).table_log);
for stream in 0..4 {
let start = match segment_size.checked_mul(stream) {
Some(offset) if offset <= dst_size => out.add(offset),
_ => return ERROR(ZstdErrorCode::CorruptionDetected),
};
let end_offset = (segment_size.saturating_mul(stream + 1)).min(dst_size);
let end = out.add(end_offset);
let mut bit_stream = std::mem::zeroed::<BIT_DStream_t>();
let init = BIT_initDStream(
&mut bit_stream,
streams[stream].cast::<c_void>(),
lengths[stream],
);
if ERR_isError(init) {
return init;
}
decode_stream_x2(start, end, &mut bit_stream, dtable, table_log);
if BIT_endOfDStream(&bit_stream) == 0 {
return ERROR(ZstdErrorCode::CorruptionDetected);
}
}
dst_size
}
const ALGO_TIME: [[(u32, u32); 2]; 16] = [
[(0, 0), (1, 1)],
[(0, 0), (1, 1)],
[(150, 216), (381, 119)],
[(170, 205), (514, 112)],
[(177, 199), (539, 110)],
[(197, 194), (644, 107)],
[(221, 192), (735, 107)],
[(256, 189), (881, 106)],
[(359, 188), (1167, 109)],
[(582, 187), (1570, 114)],
[(688, 187), (1712, 122)],
[(825, 186), (1965, 136)],
[(976, 185), (2131, 150)],
[(1180, 186), (2070, 175)],
[(1377, 185), (1731, 202)],
[(1412, 185), (1695, 202)],
];
#[no_mangle]
pub extern "C" fn HUF_selectDecoder(dst_size: usize, c_src_size: usize) -> c_uint {
if let Some(decoder) = huf_forced_decoder() {
return match decoder {
HufForcedDecoder::X1 => 0,
HufForcedDecoder::X2 => 1,
};
}
if dst_size == 0 {
return 0;
}
let quantization = if c_src_size >= dst_size {
15
} else {
c_src_size.saturating_mul(16) / dst_size
};
let decode_256 = (dst_size >> 8) as u32;
let (table0, decode0) = ALGO_TIME[quantization][0];
let (table1, decode1) = ALGO_TIME[quantization][1];
let time0 = table0.wrapping_add(decode0.wrapping_mul(decode_256));
let mut time1 = table1.wrapping_add(decode1.wrapping_mul(decode_256));
time1 = time1.wrapping_add(time1 >> 5);
(time1 < time0) as c_uint
}
#[no_mangle]
pub unsafe extern "C" fn HUF_decompress1X_usingDTable(
dst: *mut c_void,
max_dst_size: usize,
src: *const c_void,
src_size: usize,
dtable: *const u32,
_flags: c_int,
) -> usize {
let table_type = get_dtable_desc(dtable).table_type;
match huf_forced_decoder() {
Some(HufForcedDecoder::X1) => {
debug_assert_eq!(table_type, 0);
decompress_1x1_using_dtable(dst, max_dst_size, src, src_size, dtable)
}
Some(HufForcedDecoder::X2) => {
debug_assert_ne!(table_type, 0);
decompress_1x2_using_dtable(dst, max_dst_size, src, src_size, dtable)
}
None if table_type != 0 => {
decompress_1x2_using_dtable(dst, max_dst_size, src, src_size, dtable)
}
None => decompress_1x1_using_dtable(dst, max_dst_size, src, src_size, dtable),
}
}
#[no_mangle]
pub unsafe extern "C" fn HUF_decompress4X_usingDTable(
dst: *mut c_void,
max_dst_size: usize,
src: *const c_void,
src_size: usize,
dtable: *const u32,
_flags: c_int,
) -> usize {
let table_type = get_dtable_desc(dtable).table_type;
match huf_forced_decoder() {
Some(HufForcedDecoder::X1) => {
debug_assert_eq!(table_type, 0);
decompress_4x1_using_dtable(dst, max_dst_size, src, src_size, dtable)
}
Some(HufForcedDecoder::X2) => {
debug_assert_ne!(table_type, 0);
decompress_4x2_using_dtable(dst, max_dst_size, src, src_size, dtable)
}
None if table_type != 0 => {
decompress_4x2_using_dtable(dst, max_dst_size, src, src_size, dtable)
}
None => decompress_4x1_using_dtable(dst, max_dst_size, src, src_size, dtable),
}
}
#[no_mangle]
pub unsafe extern "C" fn HUF_decompress1X1_DCtx_wksp(
dctx: *mut u32,
dst: *mut c_void,
dst_size: usize,
src: *const c_void,
src_size: usize,
work_space: *mut c_void,
wksp_size: usize,
flags: c_int,
) -> usize {
let header_size = HUF_readDTableX1_wksp(dctx, src, src_size, work_space, wksp_size, flags);
if ERR_isError(header_size) {
return header_size;
}
if header_size >= src_size {
return ERROR(ZstdErrorCode::SrcSizeWrong);
}
decompress_1x1_using_dtable(
dst,
dst_size,
src.cast::<u8>().add(header_size).cast::<c_void>(),
src_size - header_size,
dctx,
)
}
#[no_mangle]
pub unsafe extern "C" fn HUF_decompress1X2_DCtx_wksp(
dctx: *mut u32,
dst: *mut c_void,
dst_size: usize,
src: *const c_void,
src_size: usize,
work_space: *mut c_void,
wksp_size: usize,
flags: c_int,
) -> usize {
let header_size = HUF_readDTableX2_wksp(dctx, src, src_size, work_space, wksp_size, flags);
if ERR_isError(header_size) {
return header_size;
}
if header_size >= src_size {
return ERROR(ZstdErrorCode::SrcSizeWrong);
}
decompress_1x2_using_dtable(
dst,
dst_size,
src.cast::<u8>().add(header_size).cast::<c_void>(),
src_size - header_size,
dctx,
)
}
#[allow(clippy::too_many_arguments)]
unsafe fn decompress_4x1_dctx_wksp(
dctx: *mut u32,
dst: *mut c_void,
dst_size: usize,
src: *const c_void,
src_size: usize,
work_space: *mut c_void,
wksp_size: usize,
flags: c_int,
) -> usize {
let header_size = HUF_readDTableX1_wksp(dctx, src, src_size, work_space, wksp_size, flags);
if ERR_isError(header_size) {
return header_size;
}
if header_size >= src_size {
return ERROR(ZstdErrorCode::SrcSizeWrong);
}
decompress_4x1_using_dtable(
dst,
dst_size,
src.cast::<u8>().add(header_size).cast::<c_void>(),
src_size - header_size,
dctx,
)
}
#[allow(clippy::too_many_arguments)]
unsafe fn decompress_4x2_dctx_wksp(
dctx: *mut u32,
dst: *mut c_void,
dst_size: usize,
src: *const c_void,
src_size: usize,
work_space: *mut c_void,
wksp_size: usize,
flags: c_int,
) -> usize {
let header_size = HUF_readDTableX2_wksp(dctx, src, src_size, work_space, wksp_size, flags);
if ERR_isError(header_size) {
return header_size;
}
if header_size >= src_size {
return ERROR(ZstdErrorCode::SrcSizeWrong);
}
decompress_4x2_using_dtable(
dst,
dst_size,
src.cast::<u8>().add(header_size).cast::<c_void>(),
src_size - header_size,
dctx,
)
}
#[no_mangle]
pub unsafe extern "C" fn HUF_decompress1X_DCtx_wksp(
dctx: *mut u32,
dst: *mut c_void,
dst_size: usize,
src: *const c_void,
src_size: usize,
work_space: *mut c_void,
wksp_size: usize,
flags: c_int,
) -> usize {
if dst_size == 0 {
return ERROR(ZstdErrorCode::DstSizeTooSmall);
}
if src_size > dst_size {
return ERROR(ZstdErrorCode::CorruptionDetected);
}
if src_size == dst_size {
std::ptr::copy_nonoverlapping(src.cast::<u8>(), dst.cast::<u8>(), dst_size);
return dst_size;
}
if src_size == 1 {
std::ptr::write_bytes(dst.cast::<u8>(), *src.cast::<u8>(), dst_size);
return dst_size;
}
if HUF_selectDecoder(dst_size, src_size) != 0 {
HUF_decompress1X2_DCtx_wksp(
dctx, dst, dst_size, src, src_size, work_space, wksp_size, flags,
)
} else {
HUF_decompress1X1_DCtx_wksp(
dctx, dst, dst_size, src, src_size, work_space, wksp_size, flags,
)
}
}
#[no_mangle]
pub unsafe extern "C" fn HUF_decompress4X_hufOnly_wksp(
dctx: *mut u32,
dst: *mut c_void,
dst_size: usize,
src: *const c_void,
src_size: usize,
work_space: *mut c_void,
wksp_size: usize,
flags: c_int,
) -> usize {
if dst_size == 0 {
return ERROR(ZstdErrorCode::DstSizeTooSmall);
}
if src_size == 0 {
return ERROR(ZstdErrorCode::CorruptionDetected);
}
if HUF_selectDecoder(dst_size, src_size) != 0 {
decompress_4x2_dctx_wksp(
dctx, dst, dst_size, src, src_size, work_space, wksp_size, flags,
)
} else {
decompress_4x1_dctx_wksp(
dctx, dst, dst_size, src, src_size, work_space, wksp_size, flags,
)
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::bitstream::{BIT_CStream_t, BIT_addBits, BIT_closeCStream, BIT_initCStream};
#[cfg(not(any(
feature = "huf-force-decompress-x1",
feature = "huf-force-decompress-x2"
)))]
use crate::bitstream::{BIT_DStream_t, BIT_initDStream};
#[cfg(not(feature = "huf-force-decompress-x2"))]
fn x1_table_for_binary_alphabet() -> [u32; 2] {
/* maxTableLog=1, type=X1, tableLog=1. Each X1 entry is
* { nbBits, symbol }, and the table consumes the two words after the
* descriptor exactly like HUF_CREATE_STATIC_DTABLEX1(). */
[0x0101_0001, 0x0101_0001]
}
fn encoded_binary_symbols(symbols: &[u8]) -> ([u8; 32], usize) {
let mut compressed = [0u8; 32];
let mut stream = unsafe { std::mem::zeroed::<BIT_CStream_t>() };
assert_eq!(
unsafe {
BIT_initCStream(
&mut stream,
compressed.as_mut_ptr().cast::<c_void>(),
compressed.len(),
)
},
0
);
for symbol in symbols.iter().rev() {
unsafe { BIT_addBits(&mut stream, usize::from(*symbol), 1) };
}
let compressed_size = unsafe { BIT_closeCStream(&mut stream) };
assert_ne!(compressed_size, 0);
(compressed, compressed_size)
}
#[cfg(not(feature = "huf-force-decompress-x2"))]
#[test]
fn x1_using_dtable_decodes_a_bitstream_written_by_the_c_compatible_writer() {
let dtable = x1_table_for_binary_alphabet();
let expected = [0u8, 1, 1, 0, 1, 0, 0, 1];
let (compressed, compressed_size) = encoded_binary_symbols(&expected);
let mut output = [0u8; 8];
let result = unsafe {
HUF_decompress1X_usingDTable(
output.as_mut_ptr().cast::<c_void>(),
output.len(),
compressed.as_ptr().cast::<c_void>(),
compressed_size,
dtable.as_ptr(),
0,
)
};
assert_eq!(result, output.len());
assert_eq!(output, expected);
}
#[cfg(not(feature = "huf-force-decompress-x1"))]
#[test]
fn x2_using_dtable_decodes_a_binary_stream_and_preserves_the_two_byte_layout() {
/* maxTableLog=1, type=X2, tableLog=1. The two entries each decode
* exactly one symbol, exercising the final-symbol path as well. */
let dtable = [0x0101_0101, 0x0101_0000, 0x0101_0001];
let expected = [1u8, 0, 1, 1, 0, 0, 1];
let (compressed, compressed_size) = encoded_binary_symbols(&expected);
let mut output = [0u8; 7];
let result = unsafe {
HUF_decompress1X_usingDTable(
output.as_mut_ptr().cast::<c_void>(),
output.len(),
compressed.as_ptr().cast::<c_void>(),
compressed_size,
dtable.as_ptr(),
0,
)
};
assert_eq!(result, output.len());
assert_eq!(output, expected);
}
#[test]
fn x1_table_builder_expands_a_reference_c_header() {
/* Header emitted by the C HUF table writer for weights [1, 1, 2]. */
let header = [129u8, 0x11];
let mut dtable = [0x0b00_000b_u32; 1 + (1 << 11)];
let mut workspace = [0u8; 2560];
let result = unsafe {
HUF_readDTableX1_wksp(
dtable.as_mut_ptr(),
header.as_ptr().cast::<c_void>(),
header.len(),
workspace.as_mut_ptr().cast::<c_void>(),
workspace.len(),
0,
)
};
assert_eq!(result, header.len());
let desc = unsafe { get_dtable_desc(dtable.as_ptr()) };
assert_eq!(
(desc.max_table_log, desc.table_type, desc.table_log),
(11, 0, 11)
);
assert_eq!(unsafe { x1_entry(dtable.as_ptr(), 0) }, (2, 0));
assert_eq!(unsafe { x1_entry(dtable.as_ptr(), 511) }, (2, 0));
assert_eq!(unsafe { x1_entry(dtable.as_ptr(), 512) }, (2, 1));
assert_eq!(unsafe { x1_entry(dtable.as_ptr(), 1024) }, (1, 2));
}
#[test]
fn table_builders_preserve_workspace_error_contracts() {
let header = [129u8, 0x11];
let mut x1 = [0x0b00_000b_u32; 1 + (1 << 11)];
let mut x2 = [0x0c00_000c_u32; 1 + (1 << 12)];
let mut workspace = [0u8; 2560];
let x1_error = unsafe {
HUF_readDTableX1_wksp(
x1.as_mut_ptr(),
header.as_ptr().cast::<c_void>(),
header.len(),
workspace.as_mut_ptr().cast::<c_void>(),
HUF_READ_DTABLE_X1_WORKSPACE_SIZE - 1,
0,
)
};
assert_eq!(x1_error, ERROR(ZstdErrorCode::TableLogTooLarge));
let x2_error = unsafe {
HUF_readDTableX2_wksp(
x2.as_mut_ptr(),
header.as_ptr().cast::<c_void>(),
header.len(),
workspace.as_mut_ptr().cast::<c_void>(),
HUF_READ_DTABLE_X2_WORKSPACE_SIZE - 1,
0,
)
};
assert_eq!(x2_error, ERROR(ZstdErrorCode::Generic));
}
#[cfg(not(any(
feature = "huf-force-decompress-x1",
feature = "huf-force-decompress-x2"
)))]
#[test]
fn selector_matches_the_reference_timing_table() {
assert_eq!(HUF_selectDecoder(128 * 1024, 16 * 1024), 1);
assert_eq!(HUF_selectDecoder(128 * 1024, 128 * 1024), 0);
}
#[test]
fn forced_decoder_feature_controls_the_public_selector() {
#[cfg(feature = "huf-force-decompress-x1")]
{
assert_eq!(huf_forced_decoder(), Some(HufForcedDecoder::X1));
assert_eq!(HUF_selectDecoder(128 * 1024, 16 * 1024), 0);
}
#[cfg(feature = "huf-force-decompress-x2")]
{
assert_eq!(huf_forced_decoder(), Some(HufForcedDecoder::X2));
assert_eq!(HUF_selectDecoder(128 * 1024, 128 * 1024), 1);
}
#[cfg(not(any(
feature = "huf-force-decompress-x1",
feature = "huf-force-decompress-x2"
)))]
{
assert_eq!(huf_forced_decoder(), None);
}
}
#[test]
fn portable_fast_sequences_match_the_c_32_and_64_bit_macro_guards() {
assert_eq!(huf_fast_sequence_slots(false), [false, true, false, true]);
assert_eq!(huf_fast_sequence_slots(true), [true, true, true, true]);
}
#[cfg(not(any(
feature = "huf-force-decompress-x1",
feature = "huf-force-decompress-x2"
)))]
#[test]
fn simulated_32_bit_fast_sequences_decode_only_the_c_permitted_two_symbols() {
let expected = [0u8, 1, 1, 0];
let (compressed, compressed_size) = encoded_binary_symbols(&expected);
let x1_dtable = x1_table_for_binary_alphabet();
let mut x1_stream = unsafe { std::mem::zeroed::<BIT_DStream_t>() };
assert_eq!(
unsafe {
BIT_initDStream(
&mut x1_stream,
compressed.as_ptr().cast::<c_void>(),
compressed_size,
)
},
compressed_size
);
let mut x1_output = [0xa5u8; 4];
let x1_end = unsafe {
decode_x1_fast_sequence(
x1_output.as_mut_ptr(),
&mut x1_stream,
x1_dtable.as_ptr(),
1,
false,
)
};
assert_eq!(unsafe { x1_end.offset_from(x1_output.as_mut_ptr()) }, 2);
assert_eq!(&x1_output[..2], &expected[..2]);
assert_eq!(&x1_output[2..], &[0xa5, 0xa5]);
/* Each X2 entry emits two bytes. A 32-bit fast sequence must take
* only the X1 and X0 slots, so it advances four bytes, rather than
* the eight bytes four X2 decodes would produce. */
let x2_dtable = [0x0101_0101, 0x0201_0100, 0x0201_0001];
let mut x2_stream = unsafe { std::mem::zeroed::<BIT_DStream_t>() };
assert_eq!(
unsafe {
BIT_initDStream(
&mut x2_stream,
compressed.as_ptr().cast::<c_void>(),
compressed_size,
)
},
compressed_size
);
let mut x2_output = [0xa5u8; 8];
let x2_end = unsafe {
decode_x2_fast_sequence(
x2_output.as_mut_ptr(),
&mut x2_stream,
x2_dtable.as_ptr(),
1,
false,
)
};
assert_eq!(unsafe { x2_end.offset_from(x2_output.as_mut_ptr()) }, 4);
assert_eq!(&x2_output[..4], &expected);
assert_eq!(&x2_output[4..], &[0xa5, 0xa5, 0xa5, 0xa5]);
}
}
Reference in New Issue View Git Blame Copy Permalink