Encoder (output stays byte-exact with libbrotlienc v1.2.1):
- H58/H68 (q5-6): match the tags 8 at a time (SWAR, like
GetMatchingTagMask's non-SSE2 code) and visit the matches with a de
Bruijn ctz; inline the bucket entry check into the H5/H6 and H58/H68
loops, caching the 4 bytes compared at best_len; a dedicated H58/H68
store and hash.
- find_last_distances caches the current byte at best_len.
- ctz64 and a branch-free log2_floor_non_zero (de Bruijn tables), and
find_match_length_with_limit uses ctz.
- load32le, load64le, peek_u16, poke_u16 and poke_u32 use one bounds
check each.
Decoder (same output and statuses):
- The literal fast path keeps pos and the counts in local variables.
- Copies of at most 16 bytes, with a distance of at least 16, copy 16
bytes (like the C code; the window gap keeps the extra bytes unread).
script/manual-test-brotli-float.c also checks log2_floor_non_zero (every
u32) and ctz64.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UZ978BikvkegL2JUkqoB1e
- encode_float.wuffs: package-private fast paths for the f32/f64 add, sub,
mul, compare and int conversion calls in the encoder's hot loops. Calling
std/softfloat is a cross-package public call (a receiver check, never
inlined). Each fast path returns exactly the softfloat result and falls
back to it for the uncommon cases. script/manual-test-brotli-float.c
cross-checks them against hardware float and double arithmetic.
- Zopfli update_nodes: compute a length's cost only when its copy code
changes, hoist the distance cache values and the continuation byte, and
use a table for small int to f32 conversions.
- FindBlocks: memoize (block_switch_bitcost + insert_cost) for capped costs
and skip the subtraction for costs at or above a per-byte threshold
(S(c) = c - min_cost is monotone, so such costs would be capped anyway).
- BrotliPopulationCost: per-call memo of each count's term and depth, and
skip the zero terms of the entropy sums.
- std/softfloat: a 64-bit f32_add_mags.
Output is unchanged (mimic tests on gcc and clang, parity sweeps).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UZ978BikvkegL2JUkqoB1e
The decoder can now report metadata blocks, like libbrotlidec's
BrotliDecoderSetMetadataCallbacks: transform_io returns
"@metadata reported", metadata_length gives the block's length and the
block's contents are written to dst. The differential fuzzer compares
the reported metadata with libbrotlidec's callbacks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UZ978Fjdsyig4v3wYAZRg5
libbrotlienc can read the uninitialized tail of a ring buffer that it grew
after a short first write, and the Wuffs encoder reads the corresponding
workbuf bytes, so both outputs can depend on memory contents (and agree
when the memory does). The fuzzer reruns a mismatch with zeroed memory for
both encoders and counts it as uninit_tail when the outputs then agree and
Wuffs' output decodes.
build_and_store_huffman_tree_fast_generic indexed the 704-entry
ZERO_REPS and NON_ZERO_REPS tables with reps & 511, so a run of more
than 511 equal code lengths (e.g. unused command codes between 75 and
688 in a quality 2 metablock of more than 128 commands) was written
wrongly and the stream did not decode. Found by the differential
fuzzer (quality 2, lgwin 10).
At qualities 0 and 1, a metadata block whose body went through the C
encoder's tiny_buf_ leaves next_out_ pointing at remaining_metadata_bytes_,
and a later flush's padding bits overwrite it, so every later
BrotliEncoderCompressStream call fails. Wuffs produces the same bytes and
does not mirror the failure. The fuzzer recognizes the corrupted C state.
Add non-coroutine read_symbol_fast and read_bits_fast helpers that are
tried before the coroutine versions, and a process_commands_fast loop
that keeps a 64-bit read-ahead bit buffer and hands back the whole
unconsumed bytes when it exits. process_commands now tracks its position
in a command (cmd_state) so that both loops can resume each other.
LZ77 copies that neither wrap nor reach the ring buffer's end are done in
chunks of at most the distance, which matches a byte-by-byte copy.
Output is unchanged. On alice29 (decoder reused) decoding is about 1.8x
to 2.1x faster for q1, q5, q9 and q11 inputs.
Add f32_compare and f64_compare, which run the cmp, eq, ne, lt, le, gt,
ge, min and max functions on every pair of 594 structured values
against the hardware's comparison operators.
The f64 and f32 comparison, min and max functions now compare the bit
patterns directly when neither operand is negative or a NaN, and call
the unchanged general code (now f64_cmp_slow and f32_cmp_slow)
otherwise. Results are unchanged: the ops harness and the old-vs-new
differential test report zero mismatches.
f64_add and f64_sub now share one fast path for normal operands
(f64_add_normal) and both return early when one operand is zero. The f32
add and sub do the same, and f32_from_u32 converts values below 2^24
directly. All results are unchanged: the sweep, the old-vs-new
differential test and the ops harness report zero mismatches.
Each bit_cost (one in the encoder, one in its metablock_builder) has a
64Ki-entry FastLog2 cache. The metablock_builder filled all of it before
building each quality 10 or 11 metablock (about 20 ms per encoder, since
each entry is an exact software log2), and the encoder's own cache was
never filled, so the quality 10 and 11 literal cost estimates computed
one software log2 per input byte.
Instead, a bitmap (zeroed when the struct is initialized) records which
cache entries are valid, and the new fast_log2_memo fills an entry the
first time it is needed. Typically only a few thousand entries are used.
bits_entropy, population_cost and their callers become impure (!) to
call it. The values are computed by the same functions, so the results
are unchanged.
On alice29 at quality 10, the encoder executes about 20% fewer
instructions, and a 1000-byte input takes 6 ms instead of 21 ms.
f32_add, f32_sub and f32_mul now compute the result directly in binary32
when both operands are normal and the result is normal (or overflows),
instead of always going through f64. f64_add and f64_sub get a similar
fast path for normal operands. Every other case takes the previous code
path, so results (including NaN payloads) are unchanged.
Add script/manual-test-softfloat-sweep.c, which checks the f32 and f64
add, sub, mul and div operations over every pair of operand exponents
with many mantissa patterns, and the f32 conversions, against the
hardware and (bit-for-bit, NaNs included) the via-f64 definitions.
libbrotlidec's internal input buffer, used after it rolls back a partially
read command, distance or block switch, makes it run out of input (and check
for an overrun metablock) at more points than the Wuffs decoder. Found by
differential fuzzing; add a mimic test that pins the behavior.
Mirror the Zopfli path's compound dictionary code in
backward_references_hq.c and hash.h: LookupAllCompoundDictionaryMatches,
FindAllCompoundDictionaryMatches, MergeMatches, the lz_matches_offset
and shadow_matches scratch space, the gap in ComputeDistanceShortcut,
UpdateNodes (including its compound dictionary last distance branch)
and BrotliZopfliCreateCommands, and the dictionary distance passed to
FindAllMatchesH10.
Remove the quality 10 and 11 dictionary gate (and its now unused
status), size the quality 11 variable region for the shadow matches
and turn the gate's test into a byte-exact check. The mimic dictionary
grid, big and rejections tests now also cover qualities 10 and 11.
At qualities 10 and 11, attaching a dictionary makes transform_io
return "#unsupported encoder quality" until the Zopfli path's compound
dictionary lookups are ported.
Add mimic_brotli_encode_with_dictionaries to the mimic library and
encoder dictionary tests: the add_dictionary API, its workbuf
accounting and rejection cases (matching libbrotlienc's default build),
a parity grid over qualities 0 to 9 with several dictionary sets, stream
offsets, large window, flushes and metadata, and dictionaries of 2 to
33 MiB, each round tripped through both decoders.
Document the encoder's dictionary API, workbuf layout and differences
from the C code.
Mirror BrotliEncoderPrepareDictionary (BROTLI_SHARED_DICTIONARY_RAW),
CreatePreparedDictionary and BrotliEncoderAttachPreparedDictionary as
encoder.add_dictionary, with the prepared hash tables (slot offsets,
heads and items) in the workbuf after the dictionary bytes, and the
compound dictionary search (FindCompoundDictionaryMatch and
LookupCompoundDictionaryMatch) for the hashers whose
CreateBackwardReferences variant enables it, the compound dictionary
gap for static dictionary distances and the compound branch of
ExtendLastCommand.
Extend the encoder parity script with a -dict grid that also checks
round trips through libbrotlidec and the Wuffs decoder.
Put the q10/q11 variable region before the ring buffer (now the last
region), growing it (and moving the ring buffer) at the first
transform_io call to use a workbuf longer than workbuf_len's minimum.
Large window now also works at qualities 10 and 11.
workbuf_len's minimum now assumes at most (n / 16) + 1 blocks per
category and (n / 64) + 64 first-phase clusters (measured maxima are
about 1/83 and 1/258 of n), and 16 BackwardMatches per byte at
quality 11 (measured maximum about 7.4). The maximum stays the C
worst case. Add q10/q11 mimic tests and workbuf_len checks, and use
the minimum in the tests' encode plan driver.
Move the ring buffer to the end of the workbuf layout (at qualities 2 to
9). A workbuf shorter than workbuf_len now works if it holds the input
so far (plus 16 bytes of slack) in the ring buffer region and the input
stays short enough for the ring buffer never to wrap. This matters most
for large-window brotli, where workbuf_len at lgwin 30 is more than
2 GiB, mostly for a ring buffer that short inputs never touch.
The parity harness now also checks its large-window configurations with
the shortest such workbuf (and that one byte less fails).
Port H10 (hash_to_binary_tree_inc.h), literal_cost.c,
backward_references_hq.c, the static dictionary match finder
(static_dict.c) and the quality 10+ parts of the metablock writer,
using the metablock_builder. The variable-size Zopfli and
metablock_builder scratch lives in a new workbuf layout region.
Port the composite hashers H35, H55 and H65 (hash_composite_inc.h) and
the rolling hashers HROLLING_FAST and HROLLING (hash_rolling_inc.h),
which ChooseHasher picks for lgwin above 24, and drop the
"#unsupported encoder quality" gate for the large-window quirk at
qualities 3 to 9. The workbuf's hasher region now includes the rolling
hasher's 64 MiB table when a composite hasher can be chosen.
Extend the encoder parity harness with a large-window grid (lgwin 10 to
30 at qualities 2 to 9) and a -lw option that runs only that grid.
This removes the encoder's copies of PrefixEncodeCopyDistance,
CommandRestoreDistanceCode, BrotliInitDistanceParams and
BrotliOptimizeHuffmanCountsForRle. The shared BrotliInitDistanceParams
also handles large-window distance params (used at quality 2).
Bound the number of block types by the number of symbols in
workbuf_len_worst_case, document the metablock_builder interface and its
workbuf layout in README.md, and add deterministic bit cost and
build_meta_block tests (with golden values from libbrotlienc v1.2.1) to
test/c/std/brotli.c.
Give the metablock builder an upper bound on the workbuf length that
build_meta_block needs, expose it through the metablock_tester and check
it against the high water mark in the parity harness. Add shard options
(-largeoff, -edge, -manytypes) and a stress phase that reaches 256 block
types to the harness.