AK: Increase LittleEndianInputBitStream's buffer size and remove loops

We current buffer one byte of data from the underlying stream. And when
we pull bits off that buffer, we do so 1 or 8 bits at a time (depending
on whether the buffer is byte aligned). The 1-bit-at-a-time loop is by
far the most common during e.g. GZIP decompression.

This replaces the u8 buffer with a u64. And instead of looping at all,
we perform bitwise operations to extract the desired number of bits.

Using the "enwik8" file as a test (100MB uncompressed, commonly used in
benchmarks: https://www.mattmahoney.net/dc/enwik8.zip), decompression
time decreases from:

    242s to 35s on Serenity
    11.125s to 3.527s on Linux

Note that BigEndianInputBitStream can also use the same techniques,
and some of the methods here may make sense to live in an endianness-
agnostic base class. The focus is GZIP right now though, which only
uses the little endian stream.
This commit is contained in:
Timothy Flynn 2023-03-28 11:58:33 -04:00 committed by Andreas Kling
parent 20aaab47f9
commit 8e834d4bb2

View file

@ -8,6 +8,7 @@
#include <AK/ByteBuffer.h>
#include <AK/MaybeOwned.h>
#include <AK/NumericLimits.h>
#include <AK/OwnPtr.h>
#include <AK/Stream.h>
@ -46,6 +47,7 @@ public:
{
return read_bits<bool>(1);
}
/// Depending on the number of bits to read, the return type can be chosen appropriately.
/// This avoids a bunch of static_cast<>'s for the user.
// TODO: Support u128, u256 etc. as well: The concepts would be quite complex.
@ -127,16 +129,29 @@ public:
// ^Stream
virtual ErrorOr<Bytes> read_some(Bytes bytes) override
{
if (m_current_byte.has_value() && is_aligned_to_byte_boundary()) {
bytes[0] = m_current_byte.release_value();
// FIXME: This accidentally slices off the first byte of the returned span.
return m_stream->read_some(bytes.slice(1));
}
align_to_byte_boundary();
return m_stream->read_some(bytes);
size_t bytes_read = 0;
auto buffer = bytes;
if (m_bit_count > 0) {
auto bits_to_read = min(buffer.size() * bits_per_byte, m_bit_count);
auto result = TRY(read_bits(bits_to_read));
bytes_read = bits_to_read / bits_per_byte;
buffer.overwrite(0, &result, bytes_read);
buffer = buffer.slice(bytes_read);
}
buffer = TRY(m_stream->read_some(buffer));
bytes_read += buffer.size();
return bytes.trim(bytes_read);
}
virtual ErrorOr<size_t> write_some(ReadonlyBytes bytes) override { return m_stream->write_some(bytes); }
virtual bool is_eof() const override { return m_stream->is_eof() && !m_current_byte.has_value(); }
virtual bool is_eof() const override { return m_stream->is_eof() && m_bit_count == 0; }
virtual bool is_open() const override { return m_stream->is_open(); }
virtual void close() override
{
@ -148,71 +163,103 @@ public:
{
return read_bits<bool>(1);
}
/// Depending on the number of bits to read, the return type can be chosen appropriately.
/// This avoids a bunch of static_cast<>'s for the user.
// TODO: Support u128, u256 etc. as well: The concepts would be quite complex.
template<Unsigned T = u64>
ErrorOr<T> read_bits(size_t count)
{
if constexpr (IsSame<bool, T>) {
VERIFY(count == 1);
}
T result = 0;
size_t nread = 0;
while (nread < count) {
if (m_current_byte.has_value()) {
if constexpr (!IsSame<bool, T> && !IsSame<u8, T>) {
// read as many bytes as possible directly
if (((count - nread) >= 8) && is_aligned_to_byte_boundary()) {
// shift existing data over
result |= (m_current_byte.value() << nread);
nread += 8;
m_current_byte.clear();
} else {
auto const bit = (m_current_byte.value() >> m_bit_offset) & 1;
result |= (bit << nread);
++nread;
if (m_bit_offset++ == 7)
m_current_byte.clear();
}
} else {
// Always take this branch for booleans or u8: there's no purpose in reading more than a single bit
auto const bit = (m_current_byte.value() >> m_bit_offset) & 1;
if constexpr (IsSame<bool, T>)
result = bit;
else
result |= (bit << nread);
++nread;
if (m_bit_offset++ == 7)
m_current_byte.clear();
}
} else {
m_current_byte = TRY(m_stream->read_value<u8>());
m_bit_offset = 0;
}
}
auto result = TRY(peek_bits<T>(count));
discard_previously_peeked_bits(count);
return result;
}
template<Unsigned T = u64>
ErrorOr<T> peek_bits(size_t count)
{
if (count > m_bit_count)
TRY(refill_buffer_from_stream());
return lsb_aligned_buffer() & lsb_mask<T>(min(count, m_bit_count));
}
ALWAYS_INLINE void discard_previously_peeked_bits(u8 count)
{
m_bit_offset += count;
m_bit_count -= count;
}
/// Discards any sub-byte stream positioning the input stream may be keeping track of.
/// Non-bitwise reads will implicitly call this.
u8 align_to_byte_boundary()
{
u8 remaining_bits = m_current_byte.value_or(0) >> m_bit_offset;
m_current_byte.clear();
u8 remaining_bits = 0;
m_bit_buffer = lsb_aligned_buffer();
m_bit_offset = 0;
if (auto offset = m_bit_count % bits_per_byte; offset != 0) {
remaining_bits = m_bit_buffer & lsb_mask<u8>(offset);
discard_previously_peeked_bits(offset);
}
return remaining_bits;
}
/// Whether we are (accidentally or intentionally) at a byte boundary right now.
ALWAYS_INLINE bool is_aligned_to_byte_boundary() const { return m_bit_offset == 0; }
ALWAYS_INLINE bool is_aligned_to_byte_boundary() const { return m_bit_count % bits_per_byte == 0; }
private:
Optional<u8> m_current_byte;
size_t m_bit_offset { 0 };
using BufferType = u64;
template<Unsigned T>
static constexpr T lsb_mask(T bits)
{
constexpr auto max = NumericLimits<T>::max();
constexpr auto digits = NumericLimits<T>::digits();
return bits == 0 ? 0 : max >> (digits - bits);
}
ALWAYS_INLINE BufferType lsb_aligned_buffer() const
{
return m_bit_offset == bit_buffer_size ? 0 : m_bit_buffer >> m_bit_offset;
}
ErrorOr<void> refill_buffer_from_stream()
{
size_t bits_to_read = bit_buffer_size - m_bit_count;
size_t bytes_to_read = bits_to_read / bits_per_byte;
BufferType buffer = 0;
Bytes bytes { &buffer, bytes_to_read };
size_t bytes_read = 0;
// FIXME: When the underlying stream is buffered, `read_some` seems to stop before EOF.
do {
auto result = TRY(m_stream->read_some(bytes));
bytes = bytes.slice(result.size());
bytes_read += result.size();
} while (!bytes.is_empty() && !m_stream->is_eof());
m_bit_buffer = (buffer << m_bit_count) | lsb_aligned_buffer();
m_bit_count += bytes_read * bits_per_byte;
m_bit_offset = 0;
return {};
}
static constexpr size_t bits_per_byte = 8u;
static constexpr size_t bit_buffer_size = sizeof(BufferType) * bits_per_byte;
MaybeOwned<Stream> m_stream;
BufferType m_bit_buffer { 0 };
u8 m_bit_offset { 0 };
u8 m_bit_count { 0 };
};
/// A stream wrapper class that allows you to write arbitrary amounts of bits