remake hash_set realization. rm indexing

new_version/direct_hash/about.md

@@ -1,7 +1,31 @@
-В этой реалзации попытка хранить хэш функции напрямую, без кодирования
+В этой реализации хэши хранятся напрямую, без Golomb-Rice-кодирования.
 
-В теории это должно увеличить размер set-строки ~x2, но
+Set-строка целиком декодируется и проверяется при первом обращении. Как и в
+`set9.c`, полностью декодированные массивы хэшей сохраняются в двух LRU-кэшах
+по 512 записей — отдельно для первого и второго операнда.
 
-- позволить индексироваться по элементам хэша
-- кратно сократить время на дешифровку
-- кэшировать можно соответствия индекс - значение, чтобы не работать с битами (может и не стоит того)
+Для 32-битного формата Base64
+декодируется сразу в три `unsigned` за блок с одновременной проверкой порядка.
+Большое понижение BPP выполняется radix-сортировкой, а не отдельным проходом на
+каждый бит.
+
+Последний запуск `taskset -c 2 python3 benchmark.py`:
+
+```text
+symbols=1000 required=500 bpp=32
+implementation  set_chars  format
+set9                3994  golomb/base62
+direct              5340  D1/base64
+
+operation                 set9       direct   direct/set9
+set_fini only           147.45 us  105.09 us         0.71x
+new+add (ctypes)        611.28 us  608.24 us         1.00x
+new+add+fini (ctypes)   769.66 us  693.60 us         0.90x
+rpmsetcmp cold          129.67 us  113.04 us         0.87x
+rpmsetcmp warm            2.99 us    1.04 us         0.35x
+```
+
+Для разреженного сравнения
+(`taskset -c 2 python3 benchmark.py --required 1`) получено
+`83.22 us` на холодном кэше и `0.86 us` на прогретом: соответственно `0.88x`
+и `1.02x` от времени `set9`.

new_version/direct_hash/benchmark.py

@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+import argparse
+import ctypes
+import gc
+import os
+import statistics
+import subprocess
+import time
+from pathlib import Path
+
+HERE = Path(__file__).resolve().parent
+BUILD = HERE / "build"
+LIBC = ctypes.CDLL(None)
+LIBC.free.argtypes = [ctypes.c_void_p]
+
+
+class SetAPI:
+    def __init__(self, path: Path):
+        self.lib = ctypes.CDLL(str(path))
+        self.lib.set_new.restype = ctypes.c_void_p
+        self.lib.set_add.argtypes = [ctypes.c_void_p, ctypes.c_char_p]
+        self.lib.set_fini.argtypes = [ctypes.c_void_p, ctypes.c_int]
+        self.lib.set_fini.restype = ctypes.c_void_p
+        self.lib.set_free.argtypes = [ctypes.c_void_p]
+        self.lib.set_free.restype = ctypes.c_void_p
+        self.lib.rpmsetcmp.argtypes = [ctypes.c_char_p, ctypes.c_char_p]
+        self.lib.rpmsetcmp.restype = ctypes.c_int
+
+    def new_with_symbols(self, symbols):
+        value = self.lib.set_new()
+        if not value:
+            raise RuntimeError("set_new returned NULL")
+        for symbol in symbols:
+            self.lib.set_add(value, symbol)
+        return value
+
+    def release(self, value, result):
+        if result:
+            LIBC.free(result)
+        self.lib.set_free(value)
+
+    def encode(self, symbols, bpp):
+        value = self.new_with_symbols(symbols)
+        result = self.lib.set_fini(value, bpp)
+        if not result:
+            raise RuntimeError("set_fini returned NULL")
+        encoded = ctypes.string_at(result)
+        self.release(value, result)
+        return encoded
+
+
+def median_fini(api, symbols, bpp, calls, rounds):
+    samples = []
+    for _ in range(rounds):
+        sets = [api.new_with_symbols(symbols) for _ in range(calls)]
+        results = []
+        start = time.perf_counter_ns()
+        for value in sets:
+            results.append(api.lib.set_fini(value, bpp))
+        samples.append((time.perf_counter_ns() - start) / calls)
+        if not all(results):
+            raise RuntimeError("set_fini returned NULL")
+        encoded = [ctypes.string_at(result) for result in results]
+        if len(set(encoded)) != 1:
+            raise RuntimeError("set_fini is not deterministic")
+        for value, result in zip(sets, results):
+            api.release(value, result)
+    return statistics.median(samples)
+
+
+def median_build(api, symbols, bpp, calls, rounds):
+    samples = []
+    for _ in range(rounds):
+        sets = []
+        results = []
+        start = time.perf_counter_ns()
+        for _ in range(calls):
+            value = api.new_with_symbols(symbols)
+            result = api.lib.set_fini(value, bpp)
+            sets.append(value)
+            results.append(result)
+        samples.append((time.perf_counter_ns() - start) / calls)
+        if not all(results):
+            raise RuntimeError("set_fini returned NULL")
+        for value, result in zip(sets, results):
+            api.release(value, result)
+    return statistics.median(samples)
+
+
+def median_add(api, symbols, calls, rounds):
+    samples = []
+    for _ in range(rounds):
+        values = []
+        start = time.perf_counter_ns()
+        for _ in range(calls):
+            values.append(api.new_with_symbols(symbols))
+        samples.append((time.perf_counter_ns() - start) / calls)
+        for value in values:
+            api.lib.set_free(value)
+    return statistics.median(samples)
+
+
+def median_cmp(api, provider, requirement, calls, rounds):
+    expected = api.lib.rpmsetcmp(provider, requirement)
+    if expected != 1 or api.lib.rpmsetcmp(provider, provider) != 0:
+        raise RuntimeError(f"unexpected rpmsetcmp result: {expected}")
+    for _ in range(100):
+        api.lib.rpmsetcmp(provider, requirement)
+
+    samples = []
+    for _ in range(rounds):
+        checksum = 0
+        start = time.perf_counter_ns()
+        for _ in range(calls):
+            checksum += api.lib.rpmsetcmp(provider, requirement)
+        samples.append((time.perf_counter_ns() - start) / calls)
+        if checksum != calls:
+            raise RuntimeError("rpmsetcmp result changed during benchmark")
+    return statistics.median(samples)
+
+
+def cold_cmp_once(api, provider, requirement):
+    read_fd, write_fd = os.pipe()
+    pid = os.fork()
+    if pid == 0:
+        os.close(read_fd)
+        start = time.perf_counter_ns()
+        result = api.lib.rpmsetcmp(provider, requirement)
+        elapsed = time.perf_counter_ns() - start
+        os.write(write_fd, f"{elapsed} {result}".encode())
+        os.close(write_fd)
+        os._exit(0)
+
+    os.close(write_fd)
+    payload = b""
+    while chunk := os.read(read_fd, 128):
+        payload += chunk
+    os.close(read_fd)
+    _, status = os.waitpid(pid, 0)
+    if status != 0:
+        raise RuntimeError(f"cold rpmsetcmp child failed: status={status}")
+    elapsed, result = map(int, payload.split())
+    if result != 1:
+        raise RuntimeError(f"unexpected cold rpmsetcmp result: {result}")
+    return elapsed
+
+
+def median_cmp_cold(api, provider, requirement, calls, rounds):
+    samples = []
+    for _ in range(rounds):
+        total = sum(cold_cmp_once(api, provider, requirement) for _ in range(calls))
+        samples.append(total / calls)
+    return statistics.median(samples)
+
+
+def verify_complete_decoding(api, provider, requirement):
+    if api.lib.rpmsetcmp(provider, requirement) != 1:
+        raise RuntimeError("provider must contain requirement")
+    if api.lib.rpmsetcmp(requirement, provider) != -1:
+        raise RuntimeError("requirement must be contained by provider")
+
+    payload_position = 4 + (len(provider) - 5) * 3 // 4
+    corrupted = provider[:payload_position] + b"!" + provider[payload_position + 1 :]
+    if api.lib.rpmsetcmp(corrupted, requirement) != -3:
+        raise RuntimeError("first operand was not decoded and validated completely")
+    if api.lib.rpmsetcmp(requirement, corrupted) != -4:
+        raise RuntimeError("second operand was not decoded and validated completely")
+
+
+def format_time(ns):
+    return f"{ns / 1000:.2f} us"
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Compare set9 and direct-hash set APIs")
+    parser.add_argument("--symbols", type=int, default=1000)
+    parser.add_argument(
+        "--required",
+        type=int,
+        help="number of evenly distributed required symbols (default: every second symbol)",
+    )
+    parser.add_argument("--bpp", type=int, default=32)
+    parser.add_argument("--rounds", type=int, default=7)
+    parser.add_argument("--fini-calls", type=int, default=5)
+    parser.add_argument("--cmp-calls", type=int, default=2000)
+    parser.add_argument("--cold-calls", type=int, default=20)
+    parser.add_argument("--skip-build", action="store_true")
+    args = parser.parse_args()
+    if args.symbols < 2 or not 10 <= args.bpp <= 32:
+        parser.error("symbols must be >= 2 and bpp must be in 10..32")
+    if args.required is not None and not 1 <= args.required < args.symbols:
+        parser.error("required must be in 1..symbols-1")
+    if min(args.rounds, args.fini_calls, args.cmp_calls, args.cold_calls) < 1:
+        parser.error("rounds and call counts must be positive")
+
+    if not args.skip_build:
+        subprocess.run([str(HERE / "build.sh")], check=True)
+
+    symbols = tuple(
+        f"symbol_{i:08d}_version_ALT_{i % 97}".encode() for i in range(args.symbols)
+    )
+    required = (
+        symbols[::2]
+        if args.required is None
+        else tuple(symbols[i * args.symbols // args.required] for i in range(args.required))
+    )
+    apis = {
+        "set9": SetAPI(BUILD / "libset9.so"),
+        "direct": SetAPI(BUILD / "libdirect-hash.so"),
+    }
+
+    gc.disable()
+    try:
+        timings = {name: [[] for _ in range(5)] for name in apis}
+        lengths = {}
+        encoded = {}
+        for name, api in apis.items():
+            provider = api.encode(symbols, args.bpp)
+            requirement = api.encode(required, args.bpp)
+            encoded[name] = (provider, requirement)
+            wire_format = "D1/base64" if provider.startswith(b"D1") else "golomb/base62"
+            lengths[name] = (len(provider), wire_format)
+
+        operations = (
+            lambda name, api: median_fini(api, symbols, args.bpp, args.fini_calls, 1),
+            lambda name, api: median_add(api, symbols, args.fini_calls, 1),
+            lambda name, api: median_build(api, symbols, args.bpp, args.fini_calls, 1),
+            lambda name, api: median_cmp_cold(
+                api, encoded[name][0], encoded[name][1], args.cold_calls, 1
+            ),
+            lambda name, api: median_cmp(
+                api, encoded[name][0], encoded[name][1], args.cmp_calls, 1
+            ),
+        )
+        names = tuple(apis)
+        for operation_index, operation in enumerate(operations):
+            for round_index in range(args.rounds):
+                order = names if round_index % 2 == 0 else tuple(reversed(names))
+                for name in order:
+                    timings[name][operation_index].append(operation(name, apis[name]))
+        timings = {
+            name: tuple(statistics.median(samples) for samples in operation_samples)
+            for name, operation_samples in timings.items()
+        }
+        # Run validation after timing: forked cold samples must inherit an empty
+        # decoded-set cache from the parent process.
+        for name, api in apis.items():
+            verify_complete_decoding(api, *encoded[name])
+    finally:
+        gc.enable()
+
+    print(f"symbols={args.symbols} required={len(required)} bpp={args.bpp}")
+    print("implementation  set_chars  format")
+    for name in apis:
+        print(f"{name:<14} {lengths[name][0]:>9}  {lengths[name][1]}")
+    print("\noperation                 set9       direct   direct/set9")
+    labels = (
+        "set_fini only",
+        "new+add (ctypes)",
+        "new+add+fini (ctypes)",
+        "rpmsetcmp cold",
+        "rpmsetcmp warm",
+    )
+    for index, label in enumerate(labels):
+        old = timings["set9"][index]
+        new = timings["direct"][index]
+        print(f"{label:<22} {format_time(old):>10} {format_time(new):>10} {new / old:>12.2f}x")
+
+
+if __name__ == "__main__":
+    main()

new_version/direct_hash/build.sh

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+HERE=$(cd "$(dirname "$0")" && pwd)
+ROOT=$(cd "$HERE/../.." && pwd)
+BUILD="$HERE/build"
+mkdir -p "$BUILD"
+touch "$BUILD/rpmlib.h" "$BUILD/system.h"
+cp "$HERE/../roaring_bitmap/set.h" "$BUILD/set.h"
+
+CFLAGS=(-O2 -std=gnu11 -D_GNU_SOURCE -Wall -Wextra -Werror -I"$BUILD")
+COMPAT=(-include "$ROOT/scripts/rpmsetcmp/newset_compat.h")
+
+for tool in mkset setcmp; do
+  cc "${CFLAGS[@]}" "${COMPAT[@]}" \
+    "$ROOT/reimplement/set9.c" "$ROOT/scripts/rpmsetcmp/$tool.c" \
+    -o "$BUILD/$tool-set9"
+  cc "${CFLAGS[@]}" "${COMPAT[@]}" \
+    "$HERE/hash_set.c" "$ROOT/scripts/rpmsetcmp/$tool.c" \
+    -o "$BUILD/$tool-direct"
+done
+
+cc "${CFLAGS[@]}" -fPIC -shared "${COMPAT[@]}" \
+  "$ROOT/reimplement/set9.c" -o "$BUILD/libset9.so"
+cc "${CFLAGS[@]}" -fPIC -shared "${COMPAT[@]}" \
+  "$HERE/hash_set.c" -o "$BUILD/libdirect-hash.so"
+
+printf 'Built tools and benchmark libraries in %s\n' "$BUILD"

new_version/direct_hash/hash_set.c

@@ -0,0 +1,949 @@
+#include <assert.h>
+#include <limits.h>
+#include <stdatomic.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/types.h>
+
+#include "rpmlib.h"
+#include "set.h"
+#include "system.h"
+
+/*
+ * This is intentionally a new set-string format. It is not compatible with
+ * the Golomb-Rice/base62 strings produced by the original lib/set.c.
+ *
+ *   D1<two decimal bpp digits><RFC 4648 base64 of packed hashes>
+ *
+ * Sorted unique hashes are packed least-significant bit first, using exactly
+ * bpp bits per hash. Base64 is only a textual representation of those bytes;
+ * there is no delta or Golomb-Rice coding.
+ */
+#define FORMAT_PREFIX "D1"
+#define FORMAT_HEADER_LEN 4
+
+_Static_assert(CHAR_BIT == 8, "direct-hash format requires 8-bit bytes");
+
+struct set {
+  size_t cnt;
+  size_t symbols_cap;
+  size_t strings_len;
+  size_t strings_cap;
+  char* strings;
+  struct symbols {
+    size_t offset;
+    unsigned hash;
+  }* symbols_v;
+};
+
+struct decoded_set {
+  unsigned* hashes;
+  size_t count;
+  unsigned bpp;
+};
+
+enum {
+  DECODED_CACHE_SIZE = 512,
+  DECODED_CACHE_BUCKETS = 1024,
+  PAIR_CACHE_SIZE = 4,
+};
+
+struct decoded_cache_entry;
+
+struct pair_cache_entry {
+  uint64_t other_identity;
+  int result;
+};
+
+struct decoded_cache_entry {
+  struct decoded_cache_entry* bucket_next;
+  struct decoded_cache_entry* newer;
+  struct decoded_cache_entry* older;
+  char* str;
+  unsigned* hashes;
+  size_t len;
+  size_t count;
+  uint32_t fingerprint;
+  uint64_t identity;
+  unsigned bucket;
+  unsigned target_bpp;
+  unsigned pair_next;
+  struct pair_cache_entry pairs[PAIR_CACHE_SIZE];
+};
+
+struct set_meta {
+  const char* str;
+  size_t len;
+  unsigned bpp;
+};
+
+static unsigned decoded_cache_count[2];
+static struct decoded_cache_entry* decoded_cache_buckets[2][DECODED_CACHE_BUCKETS];
+static struct decoded_cache_entry* decoded_cache_newest[2];
+static struct decoded_cache_entry* decoded_cache_oldest[2];
+static uint64_t decoded_cache_next_identity = 1;
+/* Cached arrays remain in use until comparison completes, so lookup, eviction,
+ * and comparison share one lock. */
+static atomic_flag decoded_cache_lock = ATOMIC_FLAG_INIT;
+
+struct set* set_new(void) {
+  struct set* set = xmalloc(sizeof(*set));
+  set->cnt = 0;
+  set->symbols_cap = 0;
+  set->strings_len = 0;
+  set->strings_cap = 0;
+  set->strings = NULL;
+  set->symbols_v = NULL;
+
+  return set;
+}
+
+void set_add(struct set* set, const char* sym) {
+  if (set->cnt == set->symbols_cap) {
+    set->symbols_cap += 1024;
+    set->symbols_v = xrealloc(set->symbols_v, sizeof(*set->symbols_v) * set->symbols_cap);
+  }
+
+  size_t length = strlen(sym) + 1;
+  size_t required = set->strings_len + length;
+  if (required > set->strings_cap) {
+    size_t capacity = set->strings_cap ? set->strings_cap : 4096;
+    while (capacity < required) capacity *= 2;
+    set->strings = xrealloc(set->strings, capacity);
+    set->strings_cap = capacity;
+  }
+
+  set->symbols_v[set->cnt].offset = set->strings_len;
+  set->symbols_v[set->cnt].hash = 0;
+  memcpy(set->strings + set->strings_len, sym, length);
+  set->strings_len = required;
+  ++set->cnt;
+
+  return;
+}
+
+struct set* set_free(struct set* set) {
+  if (set) {
+    _free(set->strings);
+    _free(set->symbols_v);
+    set = _free(set);
+  }
+
+  return NULL;
+}
+
+static unsigned hash(const char* str) {
+  unsigned hash = UINT32_C(0x9e3779b9);
+  const unsigned char* p = (const unsigned char*)str;
+
+  while (*p) {
+    hash += *p++;
+    hash += hash << 10;
+    hash ^= hash >> 6;
+  }
+
+  hash += hash << 3;
+  hash ^= hash >> 11;
+  hash += hash << 15;
+
+  return hash;
+}
+
+static int compare_symbols(const void* arg1, const void* arg2) {
+  const struct symbols* s1 = arg1;
+  const struct symbols* s2 = arg2;
+
+  if (s1->hash > s2->hash) return 1;
+  if (s1->hash < s2->hash) return -1;
+
+  return 0;
+}
+
+static void sort_symbols(struct symbols* values, size_t count, unsigned bpp) {
+  if (count < 128) {
+    qsort(values, count, sizeof(*values), compare_symbols);
+
+    return;
+  }
+
+  struct symbols* temporary = xmalloc(count * sizeof(*temporary));
+  struct symbols* source = values;
+  struct symbols* destination = temporary;
+  unsigned passes = (bpp + 7) / 8;
+
+  for (unsigned pass = 0; pass < passes; ++pass) {
+    size_t offsets[256] = {0};
+    unsigned shift = pass * 8;
+    for (size_t i = 0; i < count; ++i) ++offsets[(source[i].hash >> shift) & 0xffu];
+
+    size_t position = 0;
+    for (size_t i = 0; i < 256; ++i) {
+      size_t bucket_count = offsets[i];
+      offsets[i] = position;
+      position += bucket_count;
+    }
+
+    for (size_t i = 0; i < count; ++i) {
+      unsigned bucket = (source[i].hash >> shift) & 0xffu;
+      destination[offsets[bucket]++] = source[i];
+    }
+
+    struct symbols* swap = source;
+    source = destination;
+    destination = swap;
+  }
+
+  if (source != values) memcpy(values, source, count * sizeof(*values));
+  _free(temporary);
+
+  return;
+}
+
+static const char base64_alphabet[] =
+    "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+
+static size_t base64_encoded_size(size_t byte_count) {
+  if (byte_count > SIZE_MAX - 2) abort();
+  size_t groups = (byte_count + 2) / 3;
+  if (groups > (SIZE_MAX - FORMAT_HEADER_LEN - 1) / 4) abort();
+
+  return groups * 4;
+}
+
+static void base64_encode(const unsigned char* input, size_t input_len, char* output) {
+  while (input_len >= 3) {
+    uint32_t value = ((uint32_t)input[0] << 16) | ((uint32_t)input[1] << 8) | input[2];
+    output[0] = base64_alphabet[(value >> 18) & 0x3f];
+    output[1] = base64_alphabet[(value >> 12) & 0x3f];
+    output[2] = base64_alphabet[(value >> 6) & 0x3f];
+    output[3] = base64_alphabet[value & 0x3f];
+    input += 3;
+    input_len -= 3;
+    output += 4;
+  }
+
+  if (input_len == 1) {
+    uint32_t value = (uint32_t)input[0] << 16;
+    output[0] = base64_alphabet[(value >> 18) & 0x3f];
+    output[1] = base64_alphabet[(value >> 12) & 0x3f];
+    output[2] = '=';
+    output[3] = '=';
+  } else if (input_len == 2) {
+    uint32_t value = ((uint32_t)input[0] << 16) | ((uint32_t)input[1] << 8);
+    output[0] = base64_alphabet[(value >> 18) & 0x3f];
+    output[1] = base64_alphabet[(value >> 12) & 0x3f];
+    output[2] = base64_alphabet[(value >> 6) & 0x3f];
+    output[3] = '=';
+  }
+
+  return;
+}
+
+static unsigned char* pack_hashes(const unsigned* hashes, size_t count, unsigned bpp,
+                                  size_t* byte_count) {
+  if (count > (SIZE_MAX - 7) / bpp) abort();
+  size_t bit_count = count * bpp;
+  *byte_count = (bit_count + 7) / 8;
+  unsigned char* bytes = xmalloc(*byte_count);
+  unsigned char* output = bytes;
+  uint64_t bits = 0;
+  unsigned filled = 0;
+
+  for (size_t i = 0; i < count; ++i) {
+    bits |= (uint64_t)hashes[i] << filled;
+    filled += bpp;
+
+    while (filled >= 8) {
+      *output++ = (unsigned char)bits;
+      bits >>= 8;
+      filled -= 8;
+    }
+  }
+
+  if (filled) *output++ = (unsigned char)bits;
+  assert((size_t)(output - bytes) == *byte_count);
+
+  return bytes;
+}
+
+const char* set_fini(struct set* set, int bpp) {
+  assert(set != NULL);
+  assert(set->cnt > 0);
+  assert(bpp >= 10 && bpp <= 32);
+
+  unsigned mask = bpp < 32 ? (UINT32_C(1) << bpp) - 1 : UINT32_MAX;
+  for (size_t i = 0; i < set->cnt; ++i) {
+    set->symbols_v[i].hash = hash(set->strings + set->symbols_v[i].offset) & mask;
+  }
+  sort_symbols(set->symbols_v, set->cnt, (unsigned)bpp);
+
+  for (size_t i = 0; i + 1 < set->cnt; ++i) {
+    if (set->symbols_v[i].hash != set->symbols_v[i + 1].hash) continue;
+    const char* left = set->strings + set->symbols_v[i].offset;
+    const char* right = set->strings + set->symbols_v[i + 1].offset;
+    if (strcmp(left, right) != 0) fprintf(stderr, "warning: hash collision: %s %s\n", left, right);
+  }
+
+  unsigned* unique_hashes = xmalloc(set->cnt * sizeof(*unique_hashes));
+  size_t unique_count = 0;
+  for (size_t i = 0; i < set->cnt; ++i) {
+    while (i + 1 < set->cnt && set->symbols_v[i].hash == set->symbols_v[i + 1].hash) ++i;
+    unique_hashes[unique_count++] = set->symbols_v[i].hash;
+  }
+
+  size_t byte_count;
+  unsigned char* allocated_bytes = NULL;
+  const unsigned char* bytes;
+#if UINT_MAX == UINT32_MAX && defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && \
+    __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+  if (bpp == 32 && sizeof(unsigned) == 4) {
+    byte_count = unique_count * sizeof(*unique_hashes);
+    bytes = (const unsigned char*)unique_hashes;
+  } else
+#endif
+  {
+    allocated_bytes = pack_hashes(unique_hashes, unique_count, (unsigned)bpp, &byte_count);
+    bytes = allocated_bytes;
+  }
+  size_t payload_len = base64_encoded_size(byte_count);
+  char* output = xmalloc(FORMAT_HEADER_LEN + payload_len + 1);
+  memcpy(output, FORMAT_PREFIX, sizeof(FORMAT_PREFIX) - 1);
+  output[2] = (char)('0' + bpp / 10);
+  output[3] = (char)('0' + bpp % 10);
+  base64_encode(bytes, byte_count, output + FORMAT_HEADER_LEN);
+  output[FORMAT_HEADER_LEN + payload_len] = '\0';
+
+  _free(allocated_bytes);
+  _free(unique_hashes);
+
+  return output;
+}
+
+static const unsigned char base64_values[256] = {
+    ['A'] = 1,  ['B'] = 2,  ['C'] = 3,  ['D'] = 4,  ['E'] = 5,  ['F'] = 6,  ['G'] = 7,  ['H'] = 8,
+    ['I'] = 9,  ['J'] = 10, ['K'] = 11, ['L'] = 12, ['M'] = 13, ['N'] = 14, ['O'] = 15, ['P'] = 16,
+    ['Q'] = 17, ['R'] = 18, ['S'] = 19, ['T'] = 20, ['U'] = 21, ['V'] = 22, ['W'] = 23, ['X'] = 24,
+    ['Y'] = 25, ['Z'] = 26, ['a'] = 27, ['b'] = 28, ['c'] = 29, ['d'] = 30, ['e'] = 31, ['f'] = 32,
+    ['g'] = 33, ['h'] = 34, ['i'] = 35, ['j'] = 36, ['k'] = 37, ['l'] = 38, ['m'] = 39, ['n'] = 40,
+    ['o'] = 41, ['p'] = 42, ['q'] = 43, ['r'] = 44, ['s'] = 45, ['t'] = 46, ['u'] = 47, ['v'] = 48,
+    ['w'] = 49, ['x'] = 50, ['y'] = 51, ['z'] = 52, ['0'] = 53, ['1'] = 54, ['2'] = 55, ['3'] = 56,
+    ['4'] = 57, ['5'] = 58, ['6'] = 59, ['7'] = 60, ['8'] = 61, ['9'] = 62, ['+'] = 63, ['/'] = 64,
+};
+
+static inline int base64_value(unsigned char c) { return (int)base64_values[c] - 1; }
+
+static inline int has_set_prefix(const char* str) {
+  return str[0] == 's' && str[1] == 'e' && str[2] == 't' && str[3] == ':';
+}
+
+static int set_meta_init(const char* source, struct set_meta* meta) {
+  const char* str = source;
+  if (has_set_prefix(str)) str += 4;
+  if (has_set_prefix(str)) return -1;
+
+  /* A valid direct set has at least one complete Base64 quartet.  Checking
+   * this fixed prefix makes cache hits independent of total key length. */
+  if (str[0] != FORMAT_PREFIX[0] || str[1] != FORMAT_PREFIX[1]) return -1;
+  if (str[2] < '0' || str[2] > '9' || str[3] < '0' || str[3] > '9') return -1;
+  if (str[4] == '\0' || str[5] == '\0' || str[6] == '\0' || str[7] == '\0') return -1;
+
+  unsigned bpp = (unsigned)(str[2] - '0') * 10 + (unsigned)(str[3] - '0');
+  if (bpp < 10 || bpp > 32) return -1;
+
+  meta->str = str;
+  meta->len = 0;
+  meta->bpp = bpp;
+  return 0;
+}
+
+struct decode_writer {
+  unsigned* hashes;
+  size_t capacity;
+  size_t written;
+  uint64_t bits;
+  uint64_t mask;
+  unsigned previous;
+  unsigned filled;
+  unsigned bpp;
+  int has_previous;
+};
+
+static inline int decode_writer_put(struct decode_writer* writer, uint32_t bytes,
+                                    unsigned byte_count) {
+  writer->bits |= (uint64_t)bytes << writer->filled;
+  writer->filled += byte_count * 8;
+
+  while (writer->filled >= writer->bpp) {
+    if (writer->written == writer->capacity) return -1;
+
+    unsigned current = (unsigned)(writer->bits & writer->mask);
+    writer->bits >>= writer->bpp;
+    writer->filled -= writer->bpp;
+
+    if (writer->has_previous && writer->previous >= current) return -1;
+    if (writer->hashes) writer->hashes[writer->written] = current;
+    writer->previous = current;
+    writer->has_previous = 1;
+    ++writer->written;
+  }
+
+  return 0;
+}
+
+static int decode_base64_bytes(const char* input, size_t input_len, unsigned char* output,
+                               size_t output_len) {
+  unsigned char* const output_end = output + output_len;
+  size_t offset = 0;
+  for (; offset + 4 < input_len; offset += 4) {
+    int v0 = base64_value((unsigned char)input[offset]);
+    int v1 = base64_value((unsigned char)input[offset + 1]);
+    int v2 = base64_value((unsigned char)input[offset + 2]);
+    int v3 = base64_value((unsigned char)input[offset + 3]);
+    if ((v0 | v1 | v2 | v3) < 0) return -1;
+    output[0] = (unsigned char)((v0 << 2) | (v1 >> 4));
+    output[1] = (unsigned char)(((v1 & 0x0f) << 4) | (v2 >> 2));
+    output[2] = (unsigned char)(((v2 & 0x03) << 6) | v3);
+    output += 3;
+  }
+
+  int v0 = base64_value((unsigned char)input[offset]);
+  int v1 = base64_value((unsigned char)input[offset + 1]);
+  if ((v0 | v1) < 0 || output == output_end) return -1;
+  *output++ = (unsigned char)((v0 << 2) | (v1 >> 4));
+
+  if (input[offset + 2] == '=') {
+    return input[offset + 3] == '=' && (v1 & 0x0f) == 0 && output == output_end ? 0 : -1;
+  }
+
+  int v2 = base64_value((unsigned char)input[offset + 2]);
+  if (v2 < 0 || output == output_end) return -1;
+  *output++ = (unsigned char)(((v1 & 0x0f) << 4) | (v2 >> 2));
+  if (input[offset + 3] == '=') return (v2 & 0x03) == 0 && output == output_end ? 0 : -1;
+
+  int v3 = base64_value((unsigned char)input[offset + 3]);
+  if (v3 < 0 || output == output_end) return -1;
+  *output++ = (unsigned char)(((v2 & 0x03) << 6) | v3);
+  return output == output_end ? 0 : -1;
+}
+
+static inline int decode_base64_triplet(const char* input, uint32_t* triplet) {
+  int v0 = base64_value((unsigned char)input[0]);
+  int v1 = base64_value((unsigned char)input[1]);
+  int v2 = base64_value((unsigned char)input[2]);
+  int v3 = base64_value((unsigned char)input[3]);
+  if ((v0 | v1 | v2 | v3) < 0) return -1;
+
+  *triplet = (uint32_t)((v0 << 2) | (v1 >> 4)) |
+             (uint32_t)(((v1 & 0x0f) << 4) | (v2 >> 2)) << 8 |
+             (uint32_t)(((v2 & 0x03) << 6) | v3) << 16;
+  return 0;
+}
+
+static int decode_base64_u32(const char* input, size_t input_len, unsigned* hashes,
+                             size_t count) {
+  size_t blocks = count / 3;
+  size_t written = 0;
+  unsigned previous = 0;
+  int has_previous = 0;
+
+  for (size_t block = 0; block < blocks; ++block) {
+    uint32_t t0, t1, t2, t3;
+    if (decode_base64_triplet(input, &t0) < 0 ||
+        decode_base64_triplet(input + 4, &t1) < 0 ||
+        decode_base64_triplet(input + 8, &t2) < 0 ||
+        decode_base64_triplet(input + 12, &t3) < 0)
+      return -1;
+
+    unsigned value0 = (unsigned)(t0 | ((t1 & UINT32_C(0xff)) << 24));
+    unsigned value1 = (unsigned)((t1 >> 8) | ((t2 & UINT32_C(0xffff)) << 16));
+    unsigned value2 = (unsigned)((t2 >> 16) | (t3 << 8));
+    if ((has_previous && previous >= value0) || value0 >= value1 || value1 >= value2) return -1;
+
+    hashes[written++] = value0;
+    hashes[written++] = value1;
+    hashes[written++] = value2;
+    previous = value2;
+    has_previous = 1;
+    input += 16;
+    input_len -= 16;
+  }
+
+  size_t remaining = count - written;
+  if (remaining != 0) {
+    size_t byte_count = remaining * sizeof(*hashes);
+    unsigned char tail[2 * sizeof(*hashes)];
+    if (decode_base64_bytes(input, input_len, tail, byte_count) < 0) return -1;
+    for (size_t i = 0; i < remaining; ++i, ++written) {
+      const unsigned char* bytes = tail + i * 4;
+      unsigned current = (unsigned)bytes[0] | ((unsigned)bytes[1] << 8) |
+                         ((unsigned)bytes[2] << 16) | ((unsigned)bytes[3] << 24);
+      if (has_previous && previous >= current) return -1;
+      hashes[written] = current;
+      previous = current;
+      has_previous = 1;
+    }
+  } else if (input_len != 0) {
+    return -1;
+  }
+
+  return 0;
+}
+
+static int decode_set_sized(const char* str, size_t str_len, struct decoded_set* decoded) {
+  if (str_len == 0) str_len = strlen(str);
+  if (str_len <= FORMAT_HEADER_LEN) return -1;
+  if (strncmp(str, FORMAT_PREFIX, sizeof(FORMAT_PREFIX) - 1) != 0) return -1;
+  if (str[2] < '0' || str[2] > '9' || str[3] < '0' || str[3] > '9') return -1;
+
+  unsigned bpp = (unsigned)(str[2] - '0') * 10 + (unsigned)(str[3] - '0');
+  if (bpp < 10 || bpp > 32) return -1;
+
+  const char* input = str + FORMAT_HEADER_LEN;
+  size_t input_len = str_len - FORMAT_HEADER_LEN;
+  if (input_len == 0 || input_len % 4 != 0 || input_len / 4 > SIZE_MAX / 3) return -1;
+
+  size_t byte_count = input_len / 4 * 3;
+  if (input[input_len - 1] == '=') --byte_count;
+  if (input[input_len - 2] == '=') --byte_count;
+  if (byte_count > SIZE_MAX / 8) return -1;
+
+  size_t count = byte_count * 8 / bpp;
+  if (count == 0 || count > (SIZE_MAX - 7) / bpp || count > SIZE_MAX / sizeof(unsigned) ||
+      (count * bpp + 7) / 8 != byte_count) {
+    return -1;
+  }
+
+  unsigned* hashes = xmalloc(count * sizeof(*hashes));
+#if UINT_MAX == UINT32_MAX
+  if (bpp == 32 && sizeof(unsigned) == 4) {
+    if (decode_base64_u32(input, input_len, hashes, count) < 0) goto invalid;
+    if (decoded) {
+      decoded->hashes = hashes;
+      decoded->count = count;
+      decoded->bpp = bpp;
+    }
+    return 0;
+  }
+
+  /* Byte-aligned widths can bypass the generic bit reservoir. Decode the
+   * Base64 payload into the front of the final allocation, then expand 16-
+   * and 24-bit values backwards so unread packed bytes are never overwritten. */
+  if ((bpp == 16 || bpp == 24) && sizeof(unsigned) == 4) {
+    if (decode_base64_bytes(input, input_len, (unsigned char*)hashes, byte_count) < 0) goto invalid;
+    if (bpp == 16) {
+      for (size_t i = count; i > 0; --i) {
+        const unsigned char* bytes = (const unsigned char*)hashes + (i - 1) * 2;
+        hashes[i - 1] = (unsigned)bytes[0] | ((unsigned)bytes[1] << 8);
+      }
+    } else if (bpp == 24) {
+      for (size_t i = count; i > 0; --i) {
+        const unsigned char* bytes = (const unsigned char*)hashes + (i - 1) * 3;
+        hashes[i - 1] = (unsigned)bytes[0] | ((unsigned)bytes[1] << 8) |
+                        ((unsigned)bytes[2] << 16);
+      }
+    }
+    for (size_t i = 1; i < count; ++i) {
+      if (hashes[i - 1] >= hashes[i]) goto invalid;
+    }
+    if (decoded) {
+      decoded->hashes = hashes;
+      decoded->count = count;
+      decoded->bpp = bpp;
+    }
+    return 0;
+  }
+#endif
+
+  struct decode_writer writer = {
+      .hashes = hashes,
+      .capacity = count,
+      .mask = bpp < 32 ? (UINT64_C(1) << bpp) - 1 : UINT32_MAX,
+      .bpp = bpp,
+  };
+
+  for (size_t offset = 0; offset < input_len; offset += 4) {
+    int v0 = base64_value((unsigned char)input[offset]);
+    int v1 = base64_value((unsigned char)input[offset + 1]);
+    int last = offset + 4 == input_len;
+    if (v0 < 0 || v1 < 0) goto invalid;
+
+    uint32_t bytes = (uint32_t)((v0 << 2) | (v1 >> 4));
+    if (input[offset + 2] == '=') {
+      if (!last || input[offset + 3] != '=' || (v1 & 0x0f) != 0 ||
+          decode_writer_put(&writer, bytes, 1) < 0)
+        goto invalid;
+      continue;
+    }
+
+    int v2 = base64_value((unsigned char)input[offset + 2]);
+    if (v2 < 0) goto invalid;
+    bytes |= (uint32_t)(((v1 & 0x0f) << 4) | (v2 >> 2)) << 8;
+    if (input[offset + 3] == '=') {
+      if (!last || (v2 & 0x03) != 0 || decode_writer_put(&writer, bytes, 2) < 0) goto invalid;
+      continue;
+    }
+
+    int v3 = base64_value((unsigned char)input[offset + 3]);
+    if (v3 < 0) goto invalid;
+    bytes |= (uint32_t)(((v2 & 0x03) << 6) | v3) << 16;
+    if (decode_writer_put(&writer, bytes, 3) < 0) goto invalid;
+  }
+
+  if (writer.written != count || writer.bits != 0) goto invalid;
+
+  if (decoded) {
+    decoded->hashes = hashes;
+    decoded->count = count;
+    decoded->bpp = bpp;
+  }
+
+  return 0;
+
+invalid:
+  _free(hashes);
+
+  return -1;
+}
+
+/* Reduce a sorted set of (bpp + 1)-bit values to a sorted set of bpp-bit values. */
+static size_t downsample_set(size_t count, const unsigned* hashes, unsigned* result, unsigned bpp) {
+  unsigned mask = (UINT32_C(1) << bpp) - 1;
+  size_t lower = 0;
+  size_t upper = count;
+
+  while (lower < upper) {
+    size_t middle = lower + (upper - lower) / 2;
+    if (hashes[middle] <= mask)
+      lower = middle + 1;
+    else
+      upper = middle;
+  }
+
+  unsigned* output = result;
+  const unsigned* low = hashes;
+  const unsigned* low_end = hashes + lower;
+  const unsigned* high = hashes + lower;
+  const unsigned* high_end = hashes + count;
+
+  while (low < low_end && high < high_end) {
+    unsigned low_value = *low;
+    unsigned high_value = *high & mask;
+    if (low_value < high_value) {
+      *output++ = low_value;
+      ++low;
+    } else if (high_value < low_value) {
+      *output++ = high_value;
+      ++high;
+    } else {
+      *output++ = low_value;
+      ++low;
+      ++high;
+    }
+  }
+  while (low < low_end) *output++ = *low++;
+  while (high < high_end) *output++ = *high++ & mask;
+
+  return (size_t)(output - result);
+}
+
+static void downsample_radix_to(struct decoded_set* set, unsigned target_bpp) {
+  unsigned* original = set->hashes;
+  unsigned* scratch = xmalloc(set->count * sizeof(*scratch));
+  unsigned mask = (UINT32_C(1) << target_bpp) - 1;
+  for (size_t i = 0; i < set->count; ++i) original[i] &= mask;
+
+  unsigned* source = original;
+  unsigned* destination = scratch;
+  unsigned passes = (target_bpp + 7) / 8;
+  for (unsigned pass = 0; pass < passes; ++pass) {
+    size_t offsets[256] = {0};
+    unsigned shift = pass * 8;
+    for (size_t i = 0; i < set->count; ++i) ++offsets[(source[i] >> shift) & 0xffu];
+
+    size_t position = 0;
+    for (size_t i = 0; i < 256; ++i) {
+      size_t bucket_count = offsets[i];
+      offsets[i] = position;
+      position += bucket_count;
+    }
+    for (size_t i = 0; i < set->count; ++i) {
+      unsigned value = source[i];
+      destination[offsets[(value >> shift) & 0xffu]++] = value;
+    }
+
+    unsigned* swap = source;
+    source = destination;
+    destination = swap;
+  }
+
+  size_t unique_count = 1;
+  for (size_t i = 1; i < set->count; ++i) {
+    if (source[i] != source[unique_count - 1]) source[unique_count++] = source[i];
+  }
+
+  if (source == original) {
+    _free(scratch);
+  } else {
+    _free(original);
+    set->hashes = scratch;
+  }
+  set->count = unique_count;
+  set->bpp = target_bpp;
+}
+
+static void downsample_to(struct decoded_set* set, unsigned target_bpp) {
+  if (set->bpp == target_bpp) return;
+
+  unsigned passes = (target_bpp + 7) / 8;
+  if (set->bpp - target_bpp > passes) {
+    downsample_radix_to(set, target_bpp);
+    return;
+  }
+
+  unsigned* original = set->hashes;
+  unsigned* scratch = xmalloc(set->count * sizeof(*scratch));
+  unsigned* source = original;
+  unsigned* destination = scratch;
+
+  while (set->bpp > target_bpp) {
+    --set->bpp;
+    set->count = downsample_set(set->count, source, destination, set->bpp);
+    unsigned* swap = source;
+    source = destination;
+    destination = swap;
+  }
+
+  if (source == original) {
+    _free(scratch);
+  } else {
+    _free(original);
+    set->hashes = scratch;
+  }
+
+  return;
+}
+
+static uint32_t decoded_cache_fingerprint(const struct set_meta* meta, unsigned target_bpp) {
+  const unsigned char* str = (const unsigned char*)meta->str;
+  uint32_t fingerprint = (uint32_t)str[4] | ((uint32_t)str[5] << 8) |
+                         ((uint32_t)str[6] << 16) | ((uint32_t)str[7] << 24);
+  fingerprint ^= meta->bpp * UINT32_C(0x27d4eb2d);
+  fingerprint ^= target_bpp * UINT32_C(0x85ebca6b);
+  fingerprint ^= fingerprint >> 11;
+  fingerprint *= UINT32_C(0x9e3779b1);
+  fingerprint ^= fingerprint >> 16;
+  return fingerprint;
+}
+
+static void decoded_cache_touch(struct decoded_cache_entry* entry, unsigned cache_id) {
+  if (entry == decoded_cache_newest[cache_id]) return;
+
+  if (entry->newer) entry->newer->older = entry->older;
+  if (entry->older) entry->older->newer = entry->newer;
+  if (entry == decoded_cache_oldest[cache_id]) decoded_cache_oldest[cache_id] = entry->newer;
+
+  entry->newer = NULL;
+  entry->older = decoded_cache_newest[cache_id];
+  decoded_cache_newest[cache_id]->newer = entry;
+  decoded_cache_newest[cache_id] = entry;
+}
+
+static void decoded_cache_remove(struct decoded_cache_entry* victim, unsigned cache_id) {
+  if (victim->newer)
+    victim->newer->older = victim->older;
+  else
+    decoded_cache_newest[cache_id] = victim->older;
+  if (victim->older)
+    victim->older->newer = victim->newer;
+  else
+    decoded_cache_oldest[cache_id] = victim->newer;
+
+  struct decoded_cache_entry** link = &decoded_cache_buckets[cache_id][victim->bucket];
+  while (*link && *link != victim) link = &(*link)->bucket_next;
+  assert(*link == victim);
+  *link = victim->bucket_next;
+  assert(decoded_cache_count[cache_id] > 0);
+  --decoded_cache_count[cache_id];
+
+  _free(victim->hashes);
+  _free(victim);
+}
+
+static void decoded_cache_reset_pair_identities(void) {
+  for (unsigned bucket = 0; bucket < DECODED_CACHE_BUCKETS; ++bucket) {
+    for (struct decoded_cache_entry* provider = decoded_cache_buckets[0][bucket]; provider;
+         provider = provider->bucket_next) {
+      for (unsigned i = 0; i < PAIR_CACHE_SIZE; ++i) provider->pairs[i].other_identity = 0;
+    }
+  }
+  decoded_cache_next_identity = 1;
+}
+
+static int cache_decode_set(const struct set_meta* meta, unsigned target_bpp, unsigned cache_id,
+                            const unsigned** hashes, size_t* count,
+                            struct decoded_cache_entry** cache_entry) {
+  assert(cache_id < 2);
+  assert(target_bpp <= meta->bpp);
+
+  uint32_t fingerprint = decoded_cache_fingerprint(meta, target_bpp);
+  unsigned bucket = fingerprint & (DECODED_CACHE_BUCKETS - 1);
+  for (struct decoded_cache_entry* entry = decoded_cache_buckets[cache_id][bucket]; entry;
+       entry = entry->bucket_next) {
+    if (entry->fingerprint != fingerprint || entry->target_bpp != target_bpp ||
+        strcmp(entry->str, meta->str) != 0)
+      continue;
+
+    decoded_cache_touch(entry, cache_id);
+    *hashes = entry->hashes;
+    *count = entry->count;
+    *cache_entry = entry;
+    return 0;
+  }
+
+  size_t len = strlen(meta->str);
+  size_t input_len = len - FORMAT_HEADER_LEN;
+  if (input_len == 0 || input_len % 4 != 0 || input_len / 4 > SIZE_MAX / 3) return -1;
+
+  struct decoded_set decoded;
+  if (decode_set_sized(meta->str, len, &decoded) < 0) return -1;
+  if (decoded.bpp != meta->bpp) {
+    _free(decoded.hashes);
+    return -1;
+  }
+  downsample_to(&decoded, target_bpp);
+
+  if (len > SIZE_MAX - sizeof(struct decoded_cache_entry) - 1) {
+    _free(decoded.hashes);
+    return -1;
+  }
+  struct decoded_cache_entry* entry = xmalloc(sizeof(*entry) + len + 1);
+  memset(entry, 0, sizeof(*entry));
+  entry->str = (char*)(entry + 1);
+  memcpy(entry->str, meta->str, len + 1);
+  entry->hashes = decoded.hashes;
+  entry->len = len;
+  entry->count = decoded.count;
+  entry->fingerprint = fingerprint;
+  if (decoded_cache_next_identity == 0) decoded_cache_reset_pair_identities();
+  entry->identity = decoded_cache_next_identity++;
+  entry->bucket = bucket;
+  entry->target_bpp = target_bpp;
+
+  if (decoded_cache_count[cache_id] == DECODED_CACHE_SIZE) {
+    decoded_cache_remove(decoded_cache_oldest[cache_id], cache_id);
+  }
+
+  entry->bucket_next = decoded_cache_buckets[cache_id][bucket];
+  decoded_cache_buckets[cache_id][bucket] = entry;
+  entry->older = decoded_cache_newest[cache_id];
+  if (decoded_cache_newest[cache_id]) {
+    decoded_cache_newest[cache_id]->newer = entry;
+  } else {
+    decoded_cache_oldest[cache_id] = entry;
+  }
+  decoded_cache_newest[cache_id] = entry;
+  ++decoded_cache_count[cache_id];
+
+  *hashes = entry->hashes;
+  *count = entry->count;
+  *cache_entry = entry;
+  return 0;
+}
+
+static const unsigned* step_lower_bound(const unsigned* first, const unsigned* last, unsigned value,
+                                        size_t jump) {
+  size_t count = (size_t)(last - first);
+  if (count == 0 || first[0] >= value) return first;
+  if (jump == 0) jump = 1;
+
+  size_t position = 0;
+  size_t step = jump;
+  while (step != 0) {
+    if (step > count - position - 1) {
+      step /= 2;
+      continue;
+    }
+    size_t next = position + step;
+    if (first[next] < value)
+      position = next;
+    else
+      step /= 2;
+  }
+
+  return first + position + 1;
+}
+
+static int sorted_subset(const unsigned* small, size_t small_count, const unsigned* large,
+                         size_t large_count) {
+  const unsigned* small_end = small + small_count;
+  const unsigned* large_end = large + large_count;
+  size_t jump = large_count / small_count;
+
+  if (jump < 4) {
+    while (small < small_end) {
+      unsigned value = *small++;
+      while (large < large_end && *large < value) ++large;
+      if (large == large_end || *large != value) return 0;
+      ++large;
+    }
+
+    return 1;
+  }
+
+  while (small < small_end) {
+    unsigned value = *small++;
+    large = step_lower_bound(large, large_end, value, jump);
+    if (large == large_end || *large != value) return 0;
+    ++large;
+  }
+
+  return 1;
+}
+
+static int rpmsetcmp_locked(const char* str1, const char* str2) {
+  struct set_meta meta1;
+  if (set_meta_init(str1, &meta1) < 0) return -3;
+
+  struct set_meta meta2;
+  int meta2_status = set_meta_init(str2, &meta2);
+  unsigned target_bpp =
+      meta2_status == 0 && meta2.bpp < meta1.bpp ? meta2.bpp : meta1.bpp;
+
+  const unsigned* hashes1;
+  size_t count1;
+  struct decoded_cache_entry* entry1;
+  if (cache_decode_set(&meta1, target_bpp, 0, &hashes1, &count1, &entry1) < 0) return -3;
+  if (meta2_status < 0) return -4;
+
+  const unsigned* hashes2;
+  size_t count2;
+  struct decoded_cache_entry* entry2;
+  if (cache_decode_set(&meta2, target_bpp, 1, &hashes2, &count2, &entry2) < 0) return -4;
+
+  for (unsigned i = 0; i < PAIR_CACHE_SIZE; ++i) {
+    if (entry1->pairs[i].other_identity == entry2->identity) return entry1->pairs[i].result;
+  }
+
+  int result;
+  if (count1 == count2)
+    result = memcmp(hashes1, hashes2, count1 * sizeof(*hashes1)) == 0 ? 0 : -2;
+  else if (count1 > count2)
+    result = sorted_subset(hashes2, count2, hashes1, count1) ? 1 : -2;
+  else
+    result = sorted_subset(hashes1, count1, hashes2, count2) ? -1 : -2;
+
+  struct pair_cache_entry* pair = &entry1->pairs[entry1->pair_next++ % PAIR_CACHE_SIZE];
+  pair->other_identity = entry2->identity;
+  pair->result = result;
+  return result;
+}
+
+int rpmsetcmp(const char* str1, const char* str2) {
+  while (atomic_flag_test_and_set_explicit(&decoded_cache_lock, memory_order_acquire)) {
+  }
+  int result = rpmsetcmp_locked(str1, str2);
+  atomic_flag_clear_explicit(&decoded_cache_lock, memory_order_release);
+  return result;
+}

new_version/main_ideas.md

@@ -6,6 +6,7 @@
 
 ## хранить сразу расшифрованный set
 
+- WIP [здесь](new_version/direct_hash/about.md)
 - огромная часть ресурсов уходит не на просмотр включения множеств, а на декодирование set-строк
   - при возможности хранить больше данных за более дешёвое сравнение - прекрасно
   - тупо условный формат:
@@ -37,7 +38,3 @@
 
 - если условно "отсортировать" массив provides/requires, можно получить лучшую работу с кжшом
   - (надеюсь, что под капотом оно уже и так это делает, но проверить стоит)
-
-```
-
-```