vieter-v
/
v
2
0
Fork 0
Code Issues Pull Requests Releases Activity

map: new memory layout and optimizations

pull/4002/head
ka-weihe 2020-03-12 18:36:05 +00:00 committed by GitHub
parent 26b9987880
commit 35dd7855a0
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 338 additions and 231 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

561
vlib/builtin/map.v
View File

@ -9,6 +9,44 @@ import (
hash.wyhash hash.wyhash
) )
/*
This is a very fast hashmap implementation. It has several properties
that in combination makes it very fast. Here is a short explanation of
each property. After reading this you should have a basic understanding
of how it works:
1. Hash-function (Wyhash). Wyhash is the fastest hash-function
passing SMHasher, so it was an easy choice.
2. Open addressing (Robin Hood Hashing). With this method a hash
collision is resolved by probing. As opposed to linear probing,
Robin Hood hashing has simple but clever twist: As new keys are
inserted, old keys are shifted around in a way such that all keys
stay reasonably close to the slot they originally hash to.
3. Memory layout. Key-value pairs are stored in a "DenseArray",
with an average of rougly 6.25% unused memory as opposed to
most other dynamic array implementation with a growth factor
of 1.5 or 2. The key-values keep their index in the array -
they and are not probed. Instead, this implementation uses another
array "metas" storing "metas" (meta-data). Each Key-value has
a corresponding meta. A meta stores a reference to its key-value, and
its index in "metas" is determined by the hash of the key and probing.
A meta also stores bits from the hash (for faster rehashing etc.)
and how far away it is from the index it was originally hashed to
(probe count).
meta (64 bit) = probe_count (8 bits) | hashbits (24 bits) | kv_index (32 bit)
metas = [meta, 0, meta, 0, meta, meta, meta, 0, ...]
key_values = [kv, kv, kv, kv, kv, ...]
4. Power of two. TODO: explain
5. Extra metas. TODO: explain
6. Cached rehashing TODO: explain
7. Load-factor. TODO: explain
8. Deletion. TODO: explain
*/
const ( const (
// Number of bits from the hash stored for each entry // Number of bits from the hash stored for each entry
hashbits = 24 hashbits = 24
@ -17,11 +55,17 @@ const (
// Initial log-number of buckets in the hashtable // Initial log-number of buckets in the hashtable
init_log_capicity = 5 init_log_capicity = 5
// Initial number of buckets in the hashtable // Initial number of buckets in the hashtable
init_capicity = 1<<init_log_capicity init_capicity = 1 << init_log_capicity
// Initial load-factor // Initial max load-factor
init_load_factor = 0.8 init_max_load_factor = 0.8
// Minimum Load-factor.
// Number is picked to make delete O(1) amortized
min_load_factor = 0.3
// Initial range cap // Initial range cap
init_range_cap = init_capicity - 1 init_cap = init_capicity - 2
// Used for incrementing `extra_metas` when max
// probe count is too high, to avoid overflow
extra_metas_inc = 4
// Bitmask to select all the hashbits // Bitmask to select all the hashbits
hash_mask = u32(0x00FFFFFF) hash_mask = u32(0x00FFFFFF)
// Used for incrementing the probe-count // Used for incrementing the probe-count
@ -30,29 +74,6 @@ const (
max_probe = u32(0xFF000000) max_probe = u32(0xFF000000)
) )
pub struct map {
// Byte size of value
value_bytes int
mut:
// Index of the highest index in the hashtable
range_cap u32
// Number of cached hashbits left for rehasing
window byte
// Used for right-shifting out used hashbits
shift byte
// Pointer to Key-value memory
key_values &KeyValue
// Pointer to probe_hash memory. Each Key-value has a
// corresponding probe_hash-DWORD. Upper-bits are the
// probe-count and lower-bits are bits from the hash.
probe_hash &u32
// Measure that decides when to increase the capacity
load_factor f32
pub mut:
// Number of key-values currently in the hashmap
size int
}
struct KeyValue { struct KeyValue {
key string key string
mut: mut:
@ -67,18 +88,82 @@ mut:
size u32 size u32
} }
[inline]
fn new_dense_array() DenseArray {
unsafe {
return DenseArray {
data: &KeyValue(malloc(8 * sizeof(KeyValue)))
cap: 8
size: 0
}
}
}
// Push element to array and return index
// The growth-factor is roughly 12.5 `(x + (x >> 3))`
[inline]
fn (d mut DenseArray) push(kv KeyValue) u32 {
if d.cap == d.size {
d.cap += d.cap >> 3
d.data = &KeyValue(realloc(d.data, sizeof(KeyValue) * d.cap))
}
push_index := d.size
d.data[push_index] = kv
d.size++
return push_index
}
// Move all zeros to the end of the array
// and resize array
fn (d mut DenseArray) zeros_to_end() {
mut count := u32(0)
for i in 0..d.size {
if d.data[i].key.str != 0 {
tmp := d.data[count]
d.data[count] = d.data[i]
d.data[i] = tmp
count++
}
}
d.size = count
d.cap = if count < 8 {8} else {count}
d.data = &KeyValue(realloc(d.data, sizeof(KeyValue) * d.cap))
}
pub struct map {
// Byte size of value
value_bytes int
mut:
// Index of the highest index in the hashtable
cap u32
// Number of cached hashbits left for rehasing
window byte
// Used for right-shifting out used hashbits
shift byte
// Pointer to Key-value memory
key_values DenseArray
// Pointer to meta-data
metas &u32
// Measure that decides when to increase the capacity
max_load_factor f32
// Extra metas that allows for no ranging when incrementing
// index in the hashmap
extra_metas u32
pub mut:
// Number of key-values currently in the hashmap
size int
}
fn new_map(n, value_bytes int) map { fn new_map(n, value_bytes int) map {
probe_hash_bytes := sizeof(u32) * init_capicity
key_value_bytes := sizeof(KeyValue) * init_capicity
memory := vcalloc(key_value_bytes + probe_hash_bytes)
return map{ return map{
value_bytes: value_bytes value_bytes: value_bytes
range_cap: init_range_cap cap: init_cap
shift: init_log_capicity
window: cached_hashbits window: cached_hashbits
key_values: &KeyValue(memory) shift: init_log_capicity
probe_hash: &u32(memory + key_value_bytes) key_values: new_dense_array()
load_factor: init_load_factor metas: &u32(vcalloc(sizeof(u32) * (init_capicity + extra_metas_inc)))
max_load_factor: init_max_load_factor
extra_metas: extra_metas_inc
size: 0 size: 0
} }
} }
@ -92,279 +177,301 @@ fn new_map_init(n, value_bytes int, keys &string, values voidptr) map {
} }
fn (m mut map) set(key string, value voidptr) { fn (m mut map) set(key string, value voidptr) {
// load_factor can be adjusted. if (f32(m.size << 1) / f32(m.cap)) > m.max_load_factor {
if (f32(m.size) / f32(m.range_cap)) > m.load_factor {
m.expand() m.expand()
} }
hash := wyhash.wyhash_c(key.str, u64(key.len), 0) hash := wyhash.wyhash_c(key.str, u64(key.len), 0)
mut probe_hash := u32(((hash>>m.shift) & hash_mask) | probe_inc) mut meta := u32(((hash >> m.shift) & hash_mask) | probe_inc)
mut index := hash & m.range_cap mut index := hash & m.cap
// While probe count is less // While probe count is less
for probe_hash < m.probe_hash[index] { for meta < m.metas[index] {
index = (index + 1) & m.range_cap index += 2
probe_hash += probe_inc meta += probe_inc
} }
// While we might have a match // While we might have a match
for probe_hash == m.probe_hash[index] { for meta == m.metas[index] {
if key == m.key_values[index].key { kv_index := m.metas[index + 1]
C.memcpy(m.key_values[index].value, value, m.value_bytes) if key == m.key_values.data[kv_index].key {
C.memcpy(m.key_values.data[kv_index].value, value, m.value_bytes)
return return
} }
index = (index + 1) & m.range_cap index += 2
probe_hash += probe_inc meta += probe_inc
} }
// Match is not possible anymore. // Match is not possible anymore.
// Probe until an empty index is found. // Probe until an empty index is found.
// Swap when probe count is higher/richer (Robin Hood). // Swap when probe count is higher/richer (Robin Hood).
mut current_kv := KeyValue{ kv := KeyValue{
key:key key: key
value:malloc(m.value_bytes) value: malloc(m.value_bytes)
} }
C.memcpy(current_kv.value, value, m.value_bytes) C.memcpy(kv.value, value, m.value_bytes)
for m.probe_hash[index] != 0 { mut kv_index := m.key_values.push(kv)
if probe_hash > m.probe_hash[index] { for m.metas[index] != 0 {
// Swap probe_hash if meta > m.metas[index] {
tmp_probe_hash := m.probe_hash[index] tmp_meta := m.metas[index]
m.probe_hash[index] = probe_hash m.metas[index] = meta
probe_hash = tmp_probe_hash meta = tmp_meta
// Swap KeyValue tmp_index := m.metas[index + 1]
tmp_kv := m.key_values[index] m.metas[index + 1] = kv_index
m.key_values[index] = current_kv kv_index = tmp_index
current_kv = tmp_kv
} }
index = (index + 1) & m.range_cap index += 2
probe_hash += probe_inc meta += probe_inc
} }
// Should almost never happen probe_count := (meta >> hashbits) - 1
if (probe_hash & max_probe) == max_probe { if (probe_count << 1) == m.extra_metas {
m.expand() // Should almost never happen
m.set(current_kv.key, current_kv.value) if (meta & max_probe) == max_probe {
return m.expand()
m.set(kv.key, kv.value)
return
}
m.extra_metas += extra_metas_inc
mem_size := (m.cap + 2 + m.extra_metas)
m.metas = &u32(realloc(m.metas, sizeof(u32) * mem_size))
memset(m.metas + mem_size - extra_metas_inc, 0, sizeof(u32) * extra_metas_inc)
} }
m.probe_hash[index] = probe_hash m.metas[index] = meta
m.key_values[index] = current_kv m.metas[index + 1] = kv_index
m.size++ m.size++
} }
// Doubles the size of the hashmap
fn (m mut map) expand() { fn (m mut map) expand() {
old_range_cap := m.range_cap old_cap := m.cap
// double the size of the hashmap m.cap = ((m.cap + 2)<<1) - 2
m.range_cap = ((m.range_cap + 1)<<1) - 1 // Check if any hashbits are left
// check if no hashbits are left
if m.window == 0 { if m.window == 0 {
m.shift += cached_hashbits m.shift += cached_hashbits
m.rehash(old_range_cap) m.rehash()
m.window = cached_hashbits m.window = cached_hashbits
} }
else { else {
m.cached_rehash(old_range_cap) m.cached_rehash(old_cap)
} }
m.window-- m.window--
} }
fn (m mut map) rehash(old_range_cap u32) { // Halves the size of the hashmap
probe_hash_bytes := sizeof(u32) * (m.range_cap + 1) fn (m mut map) shrink() {
key_value_bytes := sizeof(KeyValue) * (m.range_cap + 1) m.key_values.zeros_to_end()
memory := vcalloc(probe_hash_bytes + key_value_bytes) m.cap = ((m.cap + 2)>>1) - 2
mut new_key_values := &KeyValue(memory) if m.window == 16 {
mut new_probe_hash := &u32(memory + key_value_bytes) m.shift -= cached_hashbits
for i := u32(0); i < old_range_cap + 1; i++ { m.window = 0
if m.probe_hash[i] != 0 { }
mut kv := m.key_values[i] m.rehash()
hash := wyhash.wyhash_c(kv.key.str, u64(kv.key.len), 0) m.window++
mut probe_hash := u32(((hash>>m.shift) & hash_mask) | probe_inc) }
mut index := hash & m.range_cap
// While probe count is less fn (m mut map) rehash() {
for probe_hash < new_probe_hash[index] { meta_bytes := sizeof(u32) * (m.cap + 2 + m.extra_metas)
index = (index + 1) & m.range_cap m.metas = &u32(realloc(m.metas, meta_bytes))
probe_hash += probe_inc C.memset(m.metas, 0, meta_bytes)
} for i := u32(0); i < m.key_values.size; i++ {
// Probe until an empty index is found. if m.key_values.data[i].key.str == 0 {
// Swap when probe count is higher/richer (Robin Hood). continue
for new_probe_hash[index] != 0 { }
if probe_hash > new_probe_hash[index] { kv := m.key_values.data[i]
// Swap probe_hash hash := wyhash.wyhash_c(kv.key.str, u64(kv.key.len), 0)
tmp_probe_hash := new_probe_hash[index] mut meta := u32(((hash>>m.shift) & hash_mask) | probe_inc)
new_probe_hash[index] = probe_hash mut index := hash & m.cap
probe_hash = tmp_probe_hash // While probe count is less
// Swap KeyValue for meta < m.metas[index] {
tmp_kv := new_key_values[index] index += 2
new_key_values[index] = kv meta += probe_inc
kv = tmp_kv }
} // Match is not possible anymore.
index = (index + 1) & m.range_cap // Probe until an empty index is found.
probe_hash += probe_inc // Swap when probe count is higher/richer (Robin Hood).
mut kv_index := i
for m.metas[index] != 0 {
if meta > m.metas[index] {
tmp_meta := m.metas[index]
m.metas[index] = meta
meta = tmp_meta
tmp_index := m.metas[index + 1]
m.metas[index + 1] = kv_index
kv_index = tmp_index
} }
index += 2
meta += probe_inc
}
probe_count := (meta >> hashbits) - 1
if (probe_count << 1) == m.extra_metas {
// Should almost never happen // Should almost never happen
if (probe_hash & max_probe) == max_probe { if (meta & max_probe) == max_probe {
m.expand() m.expand()
m.set(kv.key, kv.value)
return return
} }
new_probe_hash[index] = probe_hash m.extra_metas += extra_metas_inc
new_key_values[index] = kv mem_size := (m.cap + 2 + m.extra_metas)
m.metas = &u32(realloc(m.metas, sizeof(u32) * mem_size))
memset(m.metas + mem_size - extra_metas_inc, 0, sizeof(u32) * extra_metas_inc)
} }
m.metas[index] = meta
m.metas[index + 1] = kv_index
} }
unsafe{
free(m.key_values)
}
m.key_values = new_key_values
m.probe_hash = new_probe_hash
} }
fn (m mut map) cached_rehash(old_range_cap u32) { fn (m mut map) cached_rehash(old_cap u32) {
probe_hash_bytes := sizeof(u32) * (m.range_cap + 1) mut new_meta := &u32(vcalloc(sizeof(u32) * (m.cap + 2 + m.extra_metas)))
key_value_bytes := sizeof(KeyValue) * (m.range_cap + 1) old_extra_metas := m.extra_metas
memory := vcalloc(probe_hash_bytes + key_value_bytes) for i := 0; i <= old_cap + old_extra_metas; i += 2 {
mut new_probe_hash := &u32(memory + key_value_bytes) if m.metas[i] == 0 {
mut new_key_values := &KeyValue(memory) continue
for i := u32(0); i < old_range_cap + 1; i++ { }
if m.probe_hash[i] != 0 { old_meta := m.metas[i]
mut kv := m.key_values[i] old_probe_count := u64((old_meta>>hashbits) - 1) << 1
mut probe_hash := m.probe_hash[i] old_index := (i - old_probe_count) & (m.cap >> 1)
original := u64(i - ((probe_hash>>hashbits) - 1)) & (m.range_cap>>1) mut index := u64(old_index) | (old_meta << m.shift) & m.cap
hash := original | (probe_hash<<m.shift) mut meta := (old_meta & hash_mask) | probe_inc
probe_hash = (probe_hash & hash_mask) | probe_inc // While probe count is less
mut index := hash & m.range_cap for meta < new_meta[index] {
// While probe count is less index += 2
for probe_hash < new_probe_hash[index] { meta += probe_inc
index = (index + 1) & m.range_cap }
probe_hash += probe_inc // Match is not possible anymore.
} // Probe until an empty index is found.
// Probe until an empty index is found. // Swap when probe count is higher/richer (Robin Hood).
// Swap when probe count is higher/richer (Robin Hood). mut kv_index := m.metas[i + 1]
for new_probe_hash[index] != 0 { for new_meta[index] != 0 {
if probe_hash > new_probe_hash[index] { if meta > new_meta[index] {
// Swap probe_hash tmp_meta := new_meta[index]
tmp_probe_hash := new_probe_hash[index] new_meta[index] = meta
new_probe_hash[index] = probe_hash meta = tmp_meta
probe_hash = tmp_probe_hash tmp_index := new_meta[index + 1]
// Swap KeyValue new_meta[index + 1] = kv_index
tmp_kv := new_key_values[index] kv_index = tmp_index
new_key_values[index] = kv
kv = tmp_kv
}
index = (index + 1) & m.range_cap
probe_hash += probe_inc
} }
index += 2
meta += probe_inc
}
probe_count := (meta >> hashbits) - 1
if (probe_count << 1) == m.extra_metas {
// Should almost never happen // Should almost never happen
if (probe_hash & max_probe) == max_probe { if (meta & max_probe) == max_probe {
free(new_meta)
m.expand() m.expand()
m.set(kv.key, kv.value)
return return
} }
new_probe_hash[index] = probe_hash m.extra_metas += extra_metas_inc
new_key_values[index] = kv mem_size := (m.cap + 2 + m.extra_metas)
new_meta = &u32(realloc(new_meta, sizeof(u32) * mem_size))
memset(new_meta + mem_size - extra_metas_inc, 0, sizeof(u32) * extra_metas_inc)
} }
new_meta[index] = meta
new_meta[index + 1] = kv_index
} }
unsafe{ unsafe{
free(m.key_values) free(m.metas)
}
m.key_values = new_key_values
m.probe_hash = new_probe_hash
}
pub fn (m mut map) delete(key string) {
hash := wyhash.wyhash_c(key.str, u64(key.len), 0)
mut index := hash & m.range_cap
mut probe_hash := u32(((hash>>m.shift) & hash_mask) | probe_inc)
for probe_hash < m.probe_hash[index] {
index = (index + 1) & m.range_cap
probe_hash += probe_inc
}
// Perform backwards shifting
for probe_hash == m.probe_hash[index] {
if key == m.key_values[index].key {
mut old_index := index
index = (index + 1) & m.range_cap
mut current_probe_hash := m.probe_hash[index]
for (current_probe_hash>>hashbits) > 1 {
m.probe_hash[old_index] = current_probe_hash - probe_inc
m.key_values[old_index] = m.key_values[index]
old_index = index
index = (index + 1) & m.range_cap
current_probe_hash = m.probe_hash[index]
}
m.probe_hash[old_index] = 0
m.size--
return
}
index = (index + 1) & m.range_cap
probe_hash += probe_inc
} }
m.metas = new_meta
} }
[inline]
fn (m map) get(key string, out voidptr) bool { fn (m map) get(key string, out voidptr) bool {
hash := wyhash.wyhash_c(key.str, u64(key.len), 0) hash := wyhash.wyhash_c(key.str, u64(key.len), 0)
mut index := hash & m.range_cap mut index := hash & m.cap
mut probe_hash := u32(((hash>>m.shift) & hash_mask) | probe_inc) mut meta := u32(((hash>>m.shift) & hash_mask) | probe_inc)
for probe_hash < m.probe_hash[index] { for meta < m.metas[index] {
index = (index + 1) & m.range_cap index += 2
probe_hash += probe_inc meta += probe_inc
} }
for probe_hash == m.probe_hash[index] { for meta == m.metas[index] {
if key == m.key_values[index].key { kv_index := m.metas[index + 1]
C.memcpy(out, m.key_values[index].value, m.value_bytes) if key == m.key_values.data[kv_index].key {
C.memcpy(out, m.key_values.data[kv_index].value, m.value_bytes)
return true return true
} }
index = (index + 1) & m.range_cap index += 2
probe_hash += probe_inc meta += probe_inc
} }
return false return false
} }
[inline]
fn (m map) exists(key string) bool { fn (m map) exists(key string) bool {
if m.value_bytes == 0 { if m.value_bytes == 0 {
return false return false
} }
hash := wyhash.wyhash_c(key.str, u64(key.len), 0) hash := wyhash.wyhash_c(key.str, u64(key.len), 0)
mut index := hash & m.range_cap mut index := hash & m.cap
mut probe_hash := u32(((hash>>m.shift) & hash_mask) | probe_inc) mut meta := u32(((hash>>m.shift) & hash_mask) | probe_inc)
for probe_hash < m.probe_hash[index] { for meta < m.metas[index] {
index = (index + 1) & m.range_cap index += 2
probe_hash += probe_inc meta += probe_inc
} }
for probe_hash == m.probe_hash[index] { for meta == m.metas[index] {
if key == m.key_values[index].key { kv_index := m.metas[index + 1]
if key == m.key_values.data[kv_index].key {
return true return true
} }
index = (index + 1) & m.range_cap index += 2
probe_hash += probe_inc meta += probe_inc
} }
return false return false
} }
pub fn (m mut map) delete(key string) {
hash := wyhash.wyhash_c(key.str, u64(key.len), 0)
mut index := hash & m.cap
mut meta := u32(((hash>>m.shift) & hash_mask) | probe_inc)
for meta < m.metas[index] {
index += 2
meta += probe_inc
}
// Perform backwards shifting
for meta == m.metas[index] {
kv_index := m.metas[index + 1]
if key == m.key_values.data[kv_index].key {
memset(&m.key_values.data[kv_index], 0, sizeof(KeyValue))
mut old_index := index
index += 2
mut cur_meta := m.metas[index]
mut cur_index := m.metas[index + 1]
for (cur_meta >> hashbits) > 1 {
m.metas[old_index] = cur_meta - probe_inc
m.metas[old_index + 1] = cur_index
old_index = index
index += 2
cur_meta = m.metas[index]
cur_index = m.metas[index + 1]
}
m.metas[old_index] = 0
m.size--
if m.cap == 30 {return}
if (f32(m.size << 1) / f32(m.cap)) < min_load_factor {
m.shrink()
}
return
}
index += 2
meta += probe_inc
}
}
pub fn (m &map) keys() []string { pub fn (m &map) keys() []string {
mut keys := [''].repeat(m.size) mut keys := [''].repeat(m.size)
//mut keys := []string{len: m.size}
if m.value_bytes == 0 { if m.value_bytes == 0 {
return keys return keys
} }
mut j := 0 mut j := 0
for i := u32(0); i < m.range_cap + 1; i++ { for i := u32(0); i < m.key_values.size; i++ {
if m.probe_hash[i] != 0 { if m.key_values.data[i].key.str == 0 {
keys[j] = m.key_values[i].key continue
j++
} }
keys[j] = m.key_values.data[i].key
j++
} }
return keys return keys
} }
pub fn (m mut map) set_load_factor(new_load_factor f32) { pub fn (m map) free() {
if new_load_factor > 1.0 { unsafe {
m.load_factor = 1.0 free(m.metas)
} free(m.key_values.data)
else if new_load_factor < 0.1 {
m.load_factor = 0.1
}
else {
m.load_factor = new_load_factor
}
}
pub fn (m mut map) free() {
unsafe{
free(m.key_values)
} }
} }