mirror of
https://github.com/LadybirdBrowser/ladybird.git
synced 2024-08-17 09:20:28 +03:00
While watching Andreas' most recent video, I noticed that this function only worked with 32 bit values, but was a serious performance bottleneck for the kernel. As such, I reworked it to use `size_t`, so it now can switch to 64-bit sweeps on 64-bit platforms. This caused test-js to go from 12.5 seconds hot to 11.5 seconds hot on my machine when running on KVM x86_64.
373 lines
13 KiB
C++
373 lines
13 KiB
C++
/*
|
|
* Copyright (c) 2018-2021, Andreas Kling <kling@serenityos.org>
|
|
*
|
|
* SPDX-License-Identifier: BSD-2-Clause
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include <AK/Array.h>
|
|
#include <AK/BuiltinWrappers.h>
|
|
#include <AK/Optional.h>
|
|
#include <AK/StdLibExtras.h>
|
|
#include <AK/Types.h>
|
|
|
|
namespace AK {
|
|
|
|
static constexpr Array bitmask_first_byte = { 0xFF, 0xFE, 0xFC, 0xF8, 0xF0, 0xE0, 0xC0, 0x80 };
|
|
static constexpr Array bitmask_last_byte = { 0x00, 0x1, 0x3, 0x7, 0xF, 0x1F, 0x3F, 0x7F };
|
|
|
|
class BitmapView {
|
|
public:
|
|
BitmapView() = default;
|
|
|
|
BitmapView(u8* data, size_t size)
|
|
: m_data(data)
|
|
, m_size(size)
|
|
{
|
|
}
|
|
|
|
[[nodiscard]] size_t size() const { return m_size; }
|
|
[[nodiscard]] size_t size_in_bytes() const { return ceil_div(m_size, static_cast<size_t>(8)); }
|
|
[[nodiscard]] bool get(size_t index) const
|
|
{
|
|
VERIFY(index < m_size);
|
|
return 0 != (m_data[index / 8] & (1u << (index % 8)));
|
|
}
|
|
|
|
[[nodiscard]] size_t count_slow(bool value) const
|
|
{
|
|
return count_in_range(0, m_size, value);
|
|
}
|
|
|
|
[[nodiscard]] size_t count_in_range(size_t start, size_t len, bool value) const
|
|
{
|
|
VERIFY(start < m_size);
|
|
VERIFY(start + len <= m_size);
|
|
if (len == 0)
|
|
return 0;
|
|
|
|
size_t count;
|
|
const u8* first = &m_data[start / 8];
|
|
const u8* last = &m_data[(start + len) / 8];
|
|
u8 byte = *first;
|
|
byte &= bitmask_first_byte[start % 8];
|
|
if (first == last) {
|
|
byte &= bitmask_last_byte[(start + len) % 8];
|
|
count = __builtin_popcount(byte);
|
|
} else {
|
|
count = __builtin_popcount(byte);
|
|
// Don't access *last if it's out of bounds
|
|
if (last < &m_data[size_in_bytes()]) {
|
|
byte = *last;
|
|
byte &= bitmask_last_byte[(start + len) % 8];
|
|
count += __builtin_popcount(byte);
|
|
}
|
|
if (++first < last) {
|
|
const u32* ptr32 = (const u32*)(((FlatPtr)first + sizeof(u32) - 1) & ~(sizeof(u32) - 1));
|
|
if ((const u8*)ptr32 > last)
|
|
ptr32 = (const u32*)last;
|
|
while (first < (const u8*)ptr32) {
|
|
count += __builtin_popcount(*first);
|
|
first++;
|
|
}
|
|
const u32* last32 = (const u32*)((FlatPtr)last & ~(sizeof(u32) - 1));
|
|
while (ptr32 < last32) {
|
|
count += __builtin_popcountl(*ptr32);
|
|
ptr32++;
|
|
}
|
|
for (first = (const u8*)ptr32; first < last; first++)
|
|
count += __builtin_popcount(*first);
|
|
}
|
|
}
|
|
|
|
if (!value)
|
|
count = len - count;
|
|
return count;
|
|
}
|
|
|
|
[[nodiscard]] bool is_null() const { return m_data == nullptr; }
|
|
|
|
[[nodiscard]] const u8* data() const { return m_data; }
|
|
|
|
template<bool VALUE>
|
|
Optional<size_t> find_one_anywhere(size_t hint = 0) const
|
|
{
|
|
VERIFY(hint < m_size);
|
|
const u8* end = &m_data[m_size / 8];
|
|
|
|
for (;;) {
|
|
// We will use hint as what it is: a hint. Because we try to
|
|
// scan over entire 32 bit words, we may start searching before
|
|
// the hint!
|
|
const u32* ptr32 = (const u32*)((FlatPtr)&m_data[hint / 8] & ~(sizeof(u32) - 1));
|
|
if ((const u8*)ptr32 < &m_data[0]) {
|
|
ptr32++;
|
|
|
|
// m_data isn't aligned, check first bytes
|
|
size_t start_ptr32 = (const u8*)ptr32 - &m_data[0];
|
|
size_t i = 0;
|
|
u8 byte = VALUE ? 0x00 : 0xff;
|
|
while (i < start_ptr32 && m_data[i] == byte)
|
|
i++;
|
|
if (i < start_ptr32) {
|
|
byte = m_data[i];
|
|
if constexpr (!VALUE)
|
|
byte = ~byte;
|
|
VERIFY(byte != 0);
|
|
return i * 8 + __builtin_ffs(byte) - 1;
|
|
}
|
|
}
|
|
|
|
u32 val32 = VALUE ? 0x0 : 0xffffffff;
|
|
const u32* end32 = (const u32*)((FlatPtr)end & ~(sizeof(u32) - 1));
|
|
while (ptr32 < end32 && *ptr32 == val32)
|
|
ptr32++;
|
|
|
|
if (ptr32 == end32) {
|
|
// We didn't find anything, check the remaining few bytes (if any)
|
|
u8 byte = VALUE ? 0x00 : 0xff;
|
|
size_t i = (const u8*)ptr32 - &m_data[0];
|
|
size_t byte_count = m_size / 8;
|
|
VERIFY(i <= byte_count);
|
|
while (i < byte_count && m_data[i] == byte)
|
|
i++;
|
|
if (i == byte_count) {
|
|
if (hint <= 8)
|
|
return {}; // We already checked from the beginning
|
|
|
|
// Try scanning before the hint
|
|
end = (const u8*)((FlatPtr)&m_data[hint / 8] & ~(sizeof(u32) - 1));
|
|
hint = 0;
|
|
continue;
|
|
}
|
|
byte = m_data[i];
|
|
if constexpr (!VALUE)
|
|
byte = ~byte;
|
|
VERIFY(byte != 0);
|
|
return i * 8 + __builtin_ffs(byte) - 1;
|
|
}
|
|
|
|
// NOTE: We don't really care about byte ordering. We found *one*
|
|
// free bit, just calculate the position and return it
|
|
val32 = *ptr32;
|
|
if constexpr (!VALUE)
|
|
val32 = ~val32;
|
|
VERIFY(val32 != 0);
|
|
return ((const u8*)ptr32 - &m_data[0]) * 8 + __builtin_ffsl(val32) - 1;
|
|
}
|
|
}
|
|
|
|
Optional<size_t> find_one_anywhere_set(size_t hint = 0) const
|
|
{
|
|
return find_one_anywhere<true>(hint);
|
|
}
|
|
|
|
Optional<size_t> find_one_anywhere_unset(size_t hint = 0) const
|
|
{
|
|
return find_one_anywhere<false>(hint);
|
|
}
|
|
|
|
template<bool VALUE>
|
|
Optional<size_t> find_first() const
|
|
{
|
|
size_t byte_count = m_size / 8;
|
|
size_t i = 0;
|
|
|
|
u8 byte = VALUE ? 0x00 : 0xff;
|
|
while (i < byte_count && m_data[i] == byte)
|
|
i++;
|
|
if (i == byte_count)
|
|
return {};
|
|
|
|
byte = m_data[i];
|
|
if constexpr (!VALUE)
|
|
byte = ~byte;
|
|
VERIFY(byte != 0);
|
|
return i * 8 + __builtin_ffs(byte) - 1;
|
|
}
|
|
|
|
Optional<size_t> find_first_set() const { return find_first<true>(); }
|
|
Optional<size_t> find_first_unset() const { return find_first<false>(); }
|
|
|
|
// The function will return the next range of unset bits starting from the
|
|
// @from value.
|
|
// @from: the position from which the search starts. The var will be
|
|
// changed and new value is the offset of the found block.
|
|
// @min_length: minimum size of the range which will be returned.
|
|
// @max_length: maximum size of the range which will be returned.
|
|
// This is used to increase performance, since the range of
|
|
// unset bits can be long, and we don't need the while range,
|
|
// so we can stop when we've reached @max_length.
|
|
inline Optional<size_t> find_next_range_of_unset_bits(size_t& from, size_t min_length = 1, size_t max_length = max_size) const
|
|
{
|
|
if (min_length > max_length) {
|
|
return {};
|
|
}
|
|
|
|
size_t bit_size = 8 * sizeof(size_t);
|
|
|
|
size_t* bitmap = (size_t*)m_data;
|
|
|
|
// Calculating the start offset.
|
|
size_t start_bucket_index = from / bit_size;
|
|
size_t start_bucket_bit = from % bit_size;
|
|
|
|
size_t* start_of_free_chunks = &from;
|
|
size_t free_chunks = 0;
|
|
|
|
for (size_t bucket_index = start_bucket_index; bucket_index < m_size / bit_size; ++bucket_index) {
|
|
if (bitmap[bucket_index] == NumericLimits<size_t>::max()) {
|
|
// Skip over completely full bucket of size bit_size.
|
|
if (free_chunks >= min_length) {
|
|
return min(free_chunks, max_length);
|
|
}
|
|
free_chunks = 0;
|
|
start_bucket_bit = 0;
|
|
continue;
|
|
}
|
|
if (bitmap[bucket_index] == 0x0) {
|
|
// Skip over completely empty bucket of size bit_size.
|
|
if (free_chunks == 0) {
|
|
*start_of_free_chunks = bucket_index * bit_size;
|
|
}
|
|
free_chunks += bit_size;
|
|
if (free_chunks >= max_length) {
|
|
return max_length;
|
|
}
|
|
start_bucket_bit = 0;
|
|
continue;
|
|
}
|
|
|
|
size_t bucket = bitmap[bucket_index];
|
|
u8 viewed_bits = start_bucket_bit;
|
|
u32 trailing_zeroes = 0;
|
|
|
|
bucket >>= viewed_bits;
|
|
start_bucket_bit = 0;
|
|
|
|
while (viewed_bits < bit_size) {
|
|
if (bucket == 0) {
|
|
if (free_chunks == 0) {
|
|
*start_of_free_chunks = bucket_index * bit_size + viewed_bits;
|
|
}
|
|
free_chunks += bit_size - viewed_bits;
|
|
viewed_bits = bit_size;
|
|
} else {
|
|
trailing_zeroes = count_trailing_zeroes(bucket);
|
|
bucket >>= trailing_zeroes;
|
|
|
|
if (free_chunks == 0) {
|
|
*start_of_free_chunks = bucket_index * bit_size + viewed_bits;
|
|
}
|
|
free_chunks += trailing_zeroes;
|
|
viewed_bits += trailing_zeroes;
|
|
|
|
if (free_chunks >= min_length) {
|
|
return min(free_chunks, max_length);
|
|
}
|
|
|
|
// Deleting trailing ones.
|
|
u32 trailing_ones = count_trailing_zeroes(~bucket);
|
|
bucket >>= trailing_ones;
|
|
viewed_bits += trailing_ones;
|
|
free_chunks = 0;
|
|
}
|
|
}
|
|
}
|
|
|
|
if (free_chunks < min_length) {
|
|
size_t first_trailing_bit = (m_size / bit_size) * bit_size;
|
|
size_t trailing_bits = size() % bit_size;
|
|
for (size_t i = 0; i < trailing_bits; ++i) {
|
|
if (!get(first_trailing_bit + i)) {
|
|
if (free_chunks == 0)
|
|
*start_of_free_chunks = first_trailing_bit + i;
|
|
if (++free_chunks >= min_length)
|
|
return min(free_chunks, max_length);
|
|
} else {
|
|
free_chunks = 0;
|
|
}
|
|
}
|
|
return {};
|
|
}
|
|
|
|
return min(free_chunks, max_length);
|
|
}
|
|
|
|
Optional<size_t> find_longest_range_of_unset_bits(size_t max_length, size_t& found_range_size) const
|
|
{
|
|
size_t start = 0;
|
|
size_t max_region_start = 0;
|
|
size_t max_region_size = 0;
|
|
|
|
while (true) {
|
|
// Look for the next block which is bigger than currunt.
|
|
auto length_of_found_range = find_next_range_of_unset_bits(start, max_region_size + 1, max_length);
|
|
if (length_of_found_range.has_value()) {
|
|
max_region_start = start;
|
|
max_region_size = length_of_found_range.value();
|
|
start += max_region_size;
|
|
} else {
|
|
// No ranges which are bigger than current were found.
|
|
break;
|
|
}
|
|
}
|
|
|
|
found_range_size = max_region_size;
|
|
if (max_region_size != 0) {
|
|
return max_region_start;
|
|
}
|
|
return {};
|
|
}
|
|
|
|
Optional<size_t> find_first_fit(size_t minimum_length) const
|
|
{
|
|
size_t start = 0;
|
|
auto length_of_found_range = find_next_range_of_unset_bits(start, minimum_length, minimum_length);
|
|
if (length_of_found_range.has_value()) {
|
|
return start;
|
|
}
|
|
return {};
|
|
}
|
|
|
|
Optional<size_t> find_best_fit(size_t minimum_length) const
|
|
{
|
|
size_t start = 0;
|
|
size_t best_region_start = 0;
|
|
size_t best_region_size = max_size;
|
|
bool found = false;
|
|
|
|
while (true) {
|
|
// Look for the next block which is bigger than requested length.
|
|
auto length_of_found_range = find_next_range_of_unset_bits(start, minimum_length, best_region_size);
|
|
if (length_of_found_range.has_value()) {
|
|
if (best_region_size > length_of_found_range.value() || !found) {
|
|
best_region_start = start;
|
|
best_region_size = length_of_found_range.value();
|
|
found = true;
|
|
}
|
|
start += length_of_found_range.value();
|
|
} else {
|
|
// There are no ranges which can fit requested length.
|
|
break;
|
|
}
|
|
}
|
|
|
|
if (found) {
|
|
return best_region_start;
|
|
}
|
|
return {};
|
|
}
|
|
|
|
static constexpr size_t max_size = 0xffffffff;
|
|
|
|
protected:
|
|
u8* m_data { nullptr };
|
|
size_t m_size { 0 };
|
|
};
|
|
|
|
}
|
|
|
|
using AK::BitmapView;
|