-
Notifications
You must be signed in to change notification settings - Fork 0
/
ph_map.go
633 lines (522 loc) · 17.5 KB
/
ph_map.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
package finnishtable
import (
"encoding/binary"
"fmt"
"math"
"math/bits"
"unsafe"
)
// TODO: Specialize assembly for the needs for this data structure. That would
// require modifying my lovely assembly routines so I won't bother for now. I
// prefer waiting the 10+ years until Go has SIMD intrinsics.
func fixTophashForPh(hash uint8) tophash {
// We have no reserved values
return hash
}
type kv[K comparable, V any] struct {
key K
value V
}
type phtrieEntry struct {
// Really trie entries need to only carry the bucketsOffset but I decided
// to hide the hash rotation here as well just because it fits here so
// smoothly
_bucketsOffset_hashrot uint32 // 26b + 6b
}
func (e *phtrieEntry) bucketsOffset() uint32 {
return e._bucketsOffset_hashrot >> 6
}
func (e *phtrieEntry) hashrot() uint32 {
// So... bits.RotateLeft/shifts only looks at the lower 6 bits. So
// basically I'm just gonna not mask off the upper bits. It's kinda nice
// that bucketsOffset is just a shr instruction and this one is nothing.
return e._bucketsOffset_hashrot
}
// A perfect hashing Finnish hash table, or phfish for short. Stores key-value pairs.
type PhFishTable[K comparable, V any] struct {
hasher func(key unsafe.Pointer, seed uintptr) uintptr
trie []phtrieEntry
// TODO: In a decent language you could have the metas and kvs be
// interleaved into a single array and have the trie point to the start of
// the metas for that block. It would also make the phbuckets smaller as
// they wouldn't need to carry a pointer to the kvs.
//
// eg. {HDR0, tb0, tb1, tb2, tb3, tb4, kv0, kv1, kv2, kv3, kv4, HDR1, tb0,
// tb1, tb2, kv0, kv1, kv2, HDR2, tb0, tb1, ...}
//
// Less cache misses all around, l1, l2 ... tlb, etc...
allKvs []kv[K, V]
allMetas []uint8 // "encoded" (bucket header + bucket tophashes)s. See Get() for how to decode.
seed uintptr
}
func (m *PhFishTable[K, V]) Len() int {
return len(m.allKvs)
}
func (m *PhFishTable[K, V]) Get(k K) (V, bool) {
trie := m.trie
if len(trie) > 0 {
hash := m.hasher(noescape(unsafe.Pointer(&k)), m.seed)
sm := &trie[hash&uintptr(len(trie)-1)]
metastart := (*[4 + bucketSize]byte)(m.allMetas[sm.bucketsOffset():])
hdr := binary.LittleEndian.Uint32(metastart[0:])
kvsOffset := hdr
bucketmetas := (*[bucketSize]byte)(metastart[4 : 4+bucketSize])
tophash8 := fixTophashForPh(uint8(hash >> (sm.hashrot() % 64)))
tophashProbe, _ := makeHashProbes(tophash8, 0)
// NOTE: The nice thing about having to find the matching tophash is
// the case when there's none - because the key is not in the map! So
// with a good likelihood we don't have to waste our time comparing the
// keys. Goddamn is this shit practical and nice and stuff.
finder := phbucketFinderFrom(bucketmetas)
hashMatches := finder.ProbeHashMatches(tophashProbe)
if hashMatches.HasCurrent() {
// NOTE: We are potentially probing to the next bucket, but here we
// take the earliest match so there's no room for failure. Well...
// There's room for failure if we didn't also check the key
// equality, but we do. If we don't want to check for the key then
// we might have to include the size of the bucket in the bucket
// metadata so that we know not to match beyond that.
slotInBucket := hashMatches.Current()
idx := uint(kvsOffset) + uint(slotInBucket)
if idx < uint(len(m.allKvs)) { // rare branch miss only for last bucket and if tophash8(k) == 0
kv := &m.allKvs[idx]
if kv.key == k {
return kv.value, true
}
}
}
}
var zerov V
return zerov, false
}
// Returns the unique integer for this key in [0, len(kvs)). Returns -1 if k is
// not in the set
func (m *PhFishTable[K, V]) GetInt(k K) int {
trie := m.trie
if len(trie) > 0 {
hash := m.hasher(noescape(unsafe.Pointer(&k)), m.seed)
sm := &trie[hash&uintptr(len(trie)-1)]
metastart := (*[4 + bucketSize]byte)(m.allMetas[sm.bucketsOffset():])
hdr := binary.LittleEndian.Uint32(metastart[0:])
kvsOffset := hdr
bucketmetas := (*[bucketSize]byte)(metastart[4 : 4+bucketSize])
tophash8 := fixTophashForPh(uint8(hash >> (sm.hashrot() % 64)))
tophashProbe, _ := makeHashProbes(tophash8, 0)
// NOTE: The nice thing about having to find the matching tophash is
// the case when there's none - because the key is not in the map! So
// with a good likelihood we don't have to waste our time comparing the
// keys. Goddamn is this shit practical and nice and stuff.
finder := phbucketFinderFrom(bucketmetas)
hashMatches := finder.ProbeHashMatches(tophashProbe)
if hashMatches.HasCurrent() {
// NOTE: We are potentially probing to the next bucket, but here we
// take the earliest match so there's no room for failure. Well...
// There's room for failure if we didn't also check the key
// equality, but we do. If we don't want to check for the key then
// we might have to include the size of the bucket in the bucket
// metadata so that we know not to match beyond that.
slotInBucket := hashMatches.Current()
idx := uint(kvsOffset) + uint(slotInBucket)
if idx < uint(len(m.allKvs)) { // rare branch miss only for last bucket and if tophash8(k) == 0
kv := &m.allKvs[idx]
if kv.key == k {
return int(idx)
}
}
}
}
return -1
}
func MakePerfectWithUnsafeHasher[K comparable, V any](hasher func(key unsafe.Pointer, seed uintptr) uintptr, kvs []kv[K, V]) *PhFishTable[K, V] {
if len(kvs) == 0 {
return new(PhFishTable[K, V])
}
if len(kvs) > math.MaxUint32 {
panic("too many keys")
}
builder := new(phBuilder[K, V])
builder.kvs = kvs
builder.fullHashes = make([]uintptr, len(kvs))
builder.hasher = hasher
newSeed:
for {
builder.seed = uintptr(runtime_fastrand64())
{
const maxEntriesPerMap = bucketSize
adjustedSize := len(kvs) * 16 / 11 // 11/16 is the load-factor that we hit consistently
numMaps := adjustedSize / maxEntriesPerMap
numMaps = 1 << bits.Len(uint(numMaps))
initialDepth := uint8(bits.TrailingZeros(uint(numMaps)))
builder.trie = make([]*phBuilderSmolMap[K, V], numMaps)
// TODO: WHERE'S MY __REAL__ BULK ALLOCATION?
bulk := make([]phBuilderSmolMap[K, V], numMaps)
for i := range builder.trie {
builder.trie[i] = &bulk[i]
builder.trie[i].depth = initialDepth
builder.trie[i].hashrot = 56 // initially use the top 8 bits
}
}
imNotGivingUpOnYou := builder.PutAll(kvs)
if !imNotGivingUpOnYou {
goto newSeed
}
// Optimization: because we pre-allocated the trie that may have
// created some smolmaps that are not actually used or are so
// under-utilized that they could be merged with their sibling map.
//
// This doesn't really make the lookups faster but in the best case
// reduces metadata size by like ~5%. Worth it? No, I don't think så.
// The main benefit would come from it reducing the size of the trie
// but that's very unlikely to happen.
builder.compactTheMaps()
m := new(PhFishTable[K, V])
m.hasher = builder.hasher
m.seed = builder.seed
var numTotalBuckets int
var lastBucketPop int
builder.iterateMapsWithRevisiting(func(i, firstInstance int, sm *phBuilderSmolMap[K, V]) {
if i != firstInstance {
return
}
numTotalBuckets++
lastBucketPop = int(sm.pop)
})
// In the final trie we pack bucket offset to a 26-bit integer.
if numTotalBuckets >= 1<<26 {
// On average one bucket seems to contain >8 kv pairs so we get
// to store 2^26 * 2^3 = 2^29 kv pairs with a breddy gud
// likelihood.
//
// ~512 million kvs should be enough for everyone! t. Will
// Rates.
panic("too many kv pairs, sorry :-|")
}
// Calculate size for allmetas and ensure that we can do a 16-byte read
// even for the last bucket
encodedSize := numTotalBuckets*4 + len(kvs)*1
encodedSize += bucketSize - lastBucketPop
m.trie = make([]phtrieEntry, len(builder.trie))
m.allMetas = make([]uint8, encodedSize)
m.allKvs = make([]kv[K, V], len(kvs))
var bucketsOff int
var kvsOff int
builder.iterateMapsWithRevisiting(func(i, firstInstance int, sm *phBuilderSmolMap[K, V]) {
if i != firstInstance {
// we are supposed to point to the same thing as the first instance
m.trie[i] = m.trie[firstInstance]
return
}
if sm.hashrot >= 64 {
panic("oops")
}
m.trie[i] = phtrieEntry{
_bucketsOffset_hashrot: uint32(bucketsOff)<<6 | uint32(sm.hashrot),
}
if kvsOff >= 1<<27 {
panic("oops")
}
dstMetas := m.allMetas[bucketsOff:]
// first comes the 32-bits of header data
binary.LittleEndian.PutUint32(dstMetas, uint32(kvsOff))
// then the tophashes
dstHashes := dstMetas[4:][:sm.pop]
for slotInBucket := range sm.buckets[:sm.pop] {
kvIndex := sm.buckets[slotInBucket]
hash := builder.fullHashes[kvIndex]
tophash8 := fixTophashForPh(uint8(hash >> (sm.hashrot % 64)))
dstHashes[slotInBucket] = tophash8
m.allKvs[kvsOff] = builder.kvs[kvIndex]
kvsOff++
}
bucketsOff += 4 + int(sm.pop)
})
return m
}
}
func dumpPhDepths[K comparable, V any](m *phBuilder[K, V]) {
depths := make(map[uint8]int, 64)
m.iterateMapsWithRevisiting(func(i, f int, sm *phBuilderSmolMap[K, V]) {
if i == f {
depths[sm.depth]++
}
})
var maxDepth uint8
var minDepth uint8 = math.MaxUint8
for depth := range depths {
if depth > maxDepth {
maxDepth = depth
}
if depth < minDepth {
minDepth = depth
}
}
println("depths: level maps")
fmt.Printf(" [..] 0\n")
for i := minDepth; i <= maxDepth; i++ {
fmt.Printf(" [%02v] %v\n", i, depths[i])
}
fmt.Printf(" [..] 0\n")
}
type phBuilder[K comparable, V any] struct {
kvs []kv[K, V]
fullHashes []uintptr
hasher func(key unsafe.Pointer, seed uintptr) uintptr
trie []*phBuilderSmolMap[K, V] // TODO: Maybe store (smolMap.depth) here to get better use of the cache-lines
seed uintptr
// Real languages have bulk free() so freeing all of the the little maps
// should be fast.
}
type bitvector [4]uint64
func (bv *bitvector) Toggle(i uint8) {
bit := uint64(1) << (i % 64)
bv[i>>6] ^= bit
}
func (bv *bitvector) IsSet(i uint8) bool {
bit := uint64(1) << (i % 64)
return bv[i>>6]&bit != 0
}
type phBuilderSmolMap[K comparable, V any] struct {
buckets [bucketSize]uint32
// keeps track of the present tophashes in [0, 255]
//
// NOTE: With proper SIMD you probably just want to store the 16 tophashes
// with pop count. It would use less space and probably would be faster.
// Especially considering that we wouldn't have to do work to find the
// intruder when we have a tophash collision.
bitvector bitvector
pop uint8 // <= 16
depth uint8 // <= 64
hashrot uint8 // <= 56
}
func (h *phBuilder[K, V]) compactTheMaps() {
h.iterateMapsWithRevisiting(func(i, f int, sm *phBuilderSmolMap[K, V]) {
if i != f {
// already visited this
return
}
// Merging will continue as long as morale improves
for h.trie[i].depth > 0 {
hibi := uintptr(1) << uintptr(sm.depth-1)
step := hibi
siblingPos := uintptr(i) & (hibi - 1)
firstMap := h.trie[siblingPos]
secondMap := h.trie[siblingPos+step]
if firstMap.depth != secondMap.depth {
return
}
if totalPop := firstMap.pop + secondMap.pop; totalPop > bucketSize {
// No room for merging
return
}
smallerMap, biggerMap := firstMap, secondMap
if secondMap.pop < firstMap.pop {
smallerMap, biggerMap = secondMap, firstMap
}
// create a copy that we can safely mutate
mergedMap := *biggerMap
// TODO: With proper SIMD there's likely a better algorithm for
// this. Real 256-bit bitvector would make certain things easier.
insertloop:
for slotInBucket := range smallerMap.buckets[:smallerMap.pop] {
kvIndex := smallerMap.buckets[slotInBucket]
hash := h.fullHashes[kvIndex]
tophash8 := fixTophashForPh(uint8(hash >> (mergedMap.hashrot % 64)))
if !mergedMap.bitvector.IsSet(tophash8) {
if mergedMap.pop >= bucketSize {
panic("oops")
}
idx := mergedMap.pop
mergedMap.buckets[idx] = kvIndex
mergedMap.bitvector.Toggle(tophash8)
mergedMap.pop++
continue insertloop
}
if ok := h.findWorkingHashrotAndAdd(&mergedMap, hash, kvIndex); ok {
continue insertloop
}
// above loop failed to merge :-(
return
}
*firstMap = mergedMap
firstMap.depth--
// Update the trie
for i := siblingPos; i < uintptr(len(h.trie)); i += step {
h.trie[i] = firstMap
}
}
})
// try to reduce the trie size
reduceloop:
for len(h.trie) > 1 {
firstHalf, secondHalf := h.trie[:len(h.trie)/2], h.trie[len(h.trie)/2:]
for i := range firstHalf {
if firstHalf[i] != secondHalf[i] {
break reduceloop
}
}
h.trie = firstHalf
}
}
func (h *phBuilder[K, V]) findWorkingHashrotAndAdd(m *phBuilderSmolMap[K, V], frenHash uintptr, frenKvIndex uint32) bool {
newrot := m.hashrot
nextrot:
for i := 0; i < 8; i++ {
newrot -= 4
if newrot <= m.depth {
// Don't delve too deep into the triehash bits as they are equal
// for all entries in the smol map
newrot = 56
}
var bv bitvector
// Try to re-slot the original values with the new hashrot
for slotInBucket := range m.buckets[:m.pop] {
hash := h.fullHashes[m.buckets[slotInBucket]]
tophash8 := fixTophashForPh(uint8(hash >> (newrot % 64)))
// NOTE: One would assume that here the compiler would use the BTS
// instruction and check the carry flag to see the previous value
// for the bit. But apparently all compilers shit on the
// (BTS+carry) combo. How come? I want to see some of that BTS+JC
// action.
if bv.IsSet(tophash8) {
continue nextrot
}
bv.Toggle(tophash8)
}
// Then try to add the new one
{
tophash8 := fixTophashForPh(uint8(frenHash >> (newrot % 64)))
if bv.IsSet(tophash8) {
continue nextrot
}
// take the first free slot for it
freeSlotForInsert := m.pop
bv.Toggle(tophash8)
m.buckets[freeSlotForInsert] = frenKvIndex
m.pop++
}
m.bitvector = bv
m.hashrot = newrot
return true
}
return false
}
func (m *phBuilder[K, V]) PutAll(kvs []kv[K, V]) bool {
kvsLoop:
for _kvIndex := range kvs {
kvIndex := uint32(_kvIndex)
k := &kvs[kvIndex].key
hash := m.hasher(noescape(unsafe.Pointer(k)), m.seed)
m.fullHashes[kvIndex] = hash
retryloop:
for {
trie := m.trie
trieSlot := hash & uintptr(len(trie)-1)
if trieSlot >= uintptr(len(trie)) {
panic("oops")
}
sm := trie[trieSlot]
tophash8 := fixTophashForPh(uint8(hash >> (sm.hashrot % 64)))
if sm.bitvector.IsSet(tophash8) {
// oh nyoooo... A tophash collision! Perhaps even a full hash
// collision! Test it.
// Loop through all entries to find the intruder
for slotInBucket := range sm.buckets[:sm.pop] {
otherhash := m.fullHashes[sm.buckets[slotInBucket]]
if otherhash == hash {
if m.kvs[sm.buckets[slotInBucket]].key == *k {
panic("TODO?")
}
// Two keys share the same hash... Shit!
return false
}
}
if sm.pop < bucketSize {
// WE WILL NOT TOLERATE SPLITTING IF WE HAVE EMPTY SLOTS
//
// ... also, I would really like to have my 128-bit hash
// values right now.
if ok := m.findWorkingHashrotAndAdd(sm, hash, kvIndex); ok {
continue kvsLoop
}
}
// Just a top-hash collision. SPLIT and try again
m.splitForHash(sm, hash)
continue retryloop
}
if sm.pop < bucketSize {
idx := sm.pop
sm.buckets[idx] = kvIndex
sm.bitvector.Toggle(tophash8)
sm.pop++
continue kvsLoop
}
// No empty space? SPLIT and try again!
m.splitForHash(sm, hash)
}
}
return true
}
func (m *phBuilder[K, V]) splitForHash(sm *phBuilderSmolMap[K, V], hash uintptr) {
// split it
left, right := m.split(sm)
if left != sm {
// the trie update loop below requires this
panic("oops")
}
// But maybe the trie needs to grow as well...
oldDepth := left.depth - 1
if 1<<oldDepth == len(m.trie) {
oldTrie := m.trie
m.trie = make([]*phBuilderSmolMap[K, V], len(oldTrie)*2)
copy(m.trie[:len(oldTrie)], oldTrie)
copy(m.trie[len(oldTrie):], oldTrie)
}
hibi := uintptr(1) << uintptr(oldDepth)
step := hibi
for i := uintptr(hash&(hibi-1)) + step; i < uintptr(len(m.trie)); i += (step * 2) {
m.trie[i] = right
}
}
func (h *phBuilder[K, V]) split(m *phBuilderSmolMap[K, V]) (*phBuilderSmolMap[K, V], *phBuilderSmolMap[K, V]) {
if m.depth == 64 {
// 64-bit hash values have only 64 bits :-(
panic("depth overflow")
}
oldDepthBit := uint64(1) << (m.depth % 64)
m.depth++
// Re-use m as left
left := m
right := new(phBuilderSmolMap[K, V])
right.depth = m.depth
right.hashrot = m.hashrot
// Move righties to the right map
for slotInBucket := range left.buckets[:left.pop] {
hash := h.fullHashes[left.buckets[slotInBucket]]
tophash8 := fixTophashForPh(uint8(hash >> (left.hashrot % 64)))
if goesRight := hash & uintptr(oldDepthBit); goesRight == 0 {
// we want to keep our entries at the head of the array, so as we
// remove righties we leave gaps, here we fill them
moveto := slotInBucket - int(right.pop)
left.buckets[moveto] = left.buckets[slotInBucket]
continue
}
nextFreeInRight := right.pop
right.buckets[nextFreeInRight] = left.buckets[slotInBucket]
right.pop++
right.bitvector.Toggle(tophash8)
// Zero the slot left behind. Only setting the hashes is strictly necessary
left.bitvector.Toggle(tophash8)
}
left.pop -= right.pop
return left, right
}
func (m *phBuilder[K, V]) iterateMapsWithRevisiting(iter func(int, int, *phBuilderSmolMap[K, V])) {
// Oh... god... modifying iterators... pain...
for i, sm := range m.trie {
triehash := i
// Did we visit this map already? Eerily similar to the logic elsewhere ;)
hibi := uint64(1) << (sm.depth % 64)
firstInstance := int((uint64(triehash) & (hibi - 1)))
iter(i, firstInstance, sm)
}
}