multi_compress 0.2.4

data/ext/multi_compress/multi_compress.c

@@ -0,0 +1,2912 @@
+#include <ruby.h>
+#include <ruby/encoding.h>
+#include <ruby/thread.h>
+#include <ruby/fiber/scheduler.h>
+#include <brotli/decode.h>
+#include <brotli/encode.h>
+#include <lz4.h>
+#include <lz4hc.h>
+#include <pthread.h>
+#include <stdio.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <zstd.h>
+#include <zdict.h>
+
+#define MAX_DECOMPRESS_SIZE   (512ULL * 1024 * 1024)
+#define DEFAULT_MAX_RATIO     1000ULL
+#define RATIO_MIN_INPUT_BYTES 1024ULL
+#define DICT_FILE_MAX_SIZE    (32ULL * 1024 * 1024)
+
+typedef struct {
+    size_t gvl_unlock_threshold;
+    size_t fiber_yield_chunk;
+    size_t fiber_stream_threshold;
+} algo_policy_t;
+
+static const algo_policy_t ZSTD_POLICY = {
+    .gvl_unlock_threshold = 64 * 1024,
+    .fiber_yield_chunk = 64 * 1024,
+    .fiber_stream_threshold = 32 * 1024,
+};
+
+static const algo_policy_t LZ4_POLICY = {
+    .gvl_unlock_threshold = 128 * 1024,
+    .fiber_yield_chunk = 128 * 1024,
+    .fiber_stream_threshold = 64 * 1024,
+};
+
+static const algo_policy_t BROTLI_POLICY = {
+    .gvl_unlock_threshold = 16 * 1024,
+    .fiber_yield_chunk = 16 * 1024,
+    .fiber_stream_threshold = 8 * 1024,
+};
+
+static VALUE mMultiCompress;
+static VALUE eError;
+static VALUE eDataError;
+static VALUE eMemError;
+static VALUE eStreamError;
+static VALUE eUnsupportedError;
+static VALUE eLevelError;
+static VALUE cDeflater;
+static VALUE cInflater;
+static VALUE cWriter;
+static VALUE cReader;
+static VALUE cDictionary;
+static VALUE mZstd;
+static VALUE mLZ4;
+static VALUE mBrotli;
+static rb_encoding *binary_encoding;
+static struct {
+    ID zstd, lz4, brotli;
+    ID algo, algorithm, level, dictionary, size;
+    ID max_output_size, max_ratio;
+    ID fastest, default_, best;
+    ID yield_, join;
+    ID ivar_dictionary;
+} id_cache;
+
+static struct {
+    VALUE zstd, lz4, brotli;
+    VALUE algo, algorithm, level, dictionary, size;
+    VALUE max_output_size, max_ratio;
+} sym_cache;
+
+typedef enum { ALGO_ZSTD = 0, ALGO_LZ4 = 1, ALGO_BROTLI = 2 } compress_algo_t;
+
+typedef struct dictionary_s dictionary_t;
+static const rb_data_type_t dictionary_type;
+
+static void init_id_cache(void) {
+    id_cache.zstd = rb_intern("zstd");
+    id_cache.lz4 = rb_intern("lz4");
+    id_cache.brotli = rb_intern("brotli");
+    id_cache.algo = rb_intern("algo");
+    id_cache.algorithm = rb_intern("algorithm");
+    id_cache.level = rb_intern("level");
+    id_cache.dictionary = rb_intern("dictionary");
+    id_cache.size = rb_intern("size");
+    id_cache.max_output_size = rb_intern("max_output_size");
+    id_cache.max_ratio = rb_intern("max_ratio");
+    id_cache.fastest = rb_intern("fastest");
+    id_cache.default_ = rb_intern("default");
+    id_cache.best = rb_intern("best");
+    id_cache.yield_ = rb_intern("yield");
+    id_cache.join = rb_intern("join");
+    id_cache.ivar_dictionary = rb_intern("@dictionary");
+
+    sym_cache.zstd = ID2SYM(id_cache.zstd);
+    sym_cache.lz4 = ID2SYM(id_cache.lz4);
+    sym_cache.brotli = ID2SYM(id_cache.brotli);
+    sym_cache.algo = ID2SYM(id_cache.algo);
+    sym_cache.algorithm = ID2SYM(id_cache.algorithm);
+    sym_cache.level = ID2SYM(id_cache.level);
+    sym_cache.dictionary = ID2SYM(id_cache.dictionary);
+    sym_cache.size = ID2SYM(id_cache.size);
+    sym_cache.max_output_size = ID2SYM(id_cache.max_output_size);
+    sym_cache.max_ratio = ID2SYM(id_cache.max_ratio);
+}
+
+static inline VALUE opt_get(VALUE opts, VALUE sym) {
+    return NIL_P(opts) ? Qnil : rb_hash_aref(opts, sym);
+}
+
+static inline VALUE opt_lookup2(VALUE opts, VALUE sym, VALUE default_value) {
+    return NIL_P(opts) ? default_value : rb_hash_lookup2(opts, sym, default_value);
+}
+
+static inline void reject_algorithm_keyword(VALUE opts) {
+    if (NIL_P(opts))
+        return;
+    if (rb_hash_lookup2(opts, sym_cache.algorithm, Qundef) != Qundef) {
+        rb_raise(rb_eArgError, "unknown keyword: :algorithm (use :algo)");
+    }
+}
+
+static inline dictionary_t *opt_dictionary(VALUE dict_val) {
+    dictionary_t *dict;
+
+    if (NIL_P(dict_val))
+        return NULL;
+    if (!rb_obj_is_kind_of(dict_val, cDictionary)) {
+        rb_raise(rb_eTypeError, "dictionary must be a MultiCompress::Dictionary");
+    }
+
+    TypedData_Get_Struct(dict_val, dictionary_t, &dictionary_type, dict);
+    return dict;
+}
+
+static inline void raise_if_path_has_null_byte(VALUE path) {
+    if (memchr(RSTRING_PTR(path), '\0', (size_t)RSTRING_LEN(path)) != NULL) {
+        rb_raise(rb_eArgError, "path contains null byte");
+    }
+}
+
+static inline void join_thread(VALUE thread) {
+    rb_funcall(thread, id_cache.join, 0);
+}
+
+static inline void scheduler_yield(VALUE scheduler) {
+    rb_funcall(scheduler, id_cache.yield_, 0);
+}
+
+static inline VALUE dictionary_ivar_get(VALUE self) {
+    return rb_ivar_get(self, id_cache.ivar_dictionary);
+}
+
+static inline void dictionary_ivar_set(VALUE self, VALUE dictionary) {
+    rb_ivar_set(self, id_cache.ivar_dictionary, dictionary);
+}
+
+static inline uint32_t read_le_u32(const uint8_t *p) {
+    return (uint32_t)p[0] | ((uint32_t)p[1] << 8) | ((uint32_t)p[2] << 16) | ((uint32_t)p[3] << 24);
+}
+
+static inline void write_le_u32(uint8_t *p, uint32_t v) {
+    p[0] = (uint8_t)(v & 0xFF);
+    p[1] = (uint8_t)((v >> 8) & 0xFF);
+    p[2] = (uint8_t)((v >> 16) & 0xFF);
+    p[3] = (uint8_t)((v >> 24) & 0xFF);
+}
+
+static inline const algo_policy_t *algo_policy(compress_algo_t algo) {
+    switch (algo) {
+    case ALGO_ZSTD:
+        return &ZSTD_POLICY;
+    case ALGO_LZ4:
+        return &LZ4_POLICY;
+    case ALGO_BROTLI:
+        return &BROTLI_POLICY;
+    }
+    return &ZSTD_POLICY;
+}
+
+static compress_algo_t sym_to_algo(VALUE sym) {
+    if (!SYMBOL_P(sym)) {
+        rb_raise(rb_eTypeError, "algo must be a Symbol (:zstd, :lz4, :brotli)");
+    }
+
+    ID id = SYM2ID(sym);
+    if (id == id_cache.zstd)
+        return ALGO_ZSTD;
+    if (id == id_cache.lz4)
+        return ALGO_LZ4;
+    if (id == id_cache.brotli)
+        return ALGO_BROTLI;
+    rb_raise(rb_eArgError, "Unknown algorithm: %s", rb_id2name(id));
+    return ALGO_ZSTD;
+}
+
+static inline VALUE algo_to_sym(compress_algo_t algo) {
+    switch (algo) {
+    case ALGO_ZSTD:
+        return sym_cache.zstd;
+    case ALGO_LZ4:
+        return sym_cache.lz4;
+    case ALGO_BROTLI:
+        return sym_cache.brotli;
+    }
+    return Qnil;
+}
+
+typedef struct {
+    int min, max, fastest, default_, best;
+    const char *name;
+} level_spec_t;
+
+static const level_spec_t level_spec[] = {
+    [ALGO_ZSTD] = {.min = 1, .max = 22, .fastest = 1, .default_ = 3, .best = 19, .name = "zstd"},
+    [ALGO_LZ4] = {.min = 1, .max = 16, .fastest = 1, .default_ = 1, .best = 16, .name = "lz4"},
+    [ALGO_BROTLI] =
+        {.min = 0, .max = 11, .fastest = 0, .default_ = 6, .best = 11, .name = "brotli"},
+};
+
+static int resolve_level(compress_algo_t algo, VALUE level_val) {
+    const level_spec_t *spec = &level_spec[algo];
+
+    if (NIL_P(level_val))
+        return spec->default_;
+
+    if (SYMBOL_P(level_val)) {
+        ID id = SYM2ID(level_val);
+        if (id == id_cache.fastest)
+            return spec->fastest;
+        if (id == id_cache.default_)
+            return spec->default_;
+        if (id == id_cache.best)
+            return spec->best;
+        rb_raise(eLevelError, "Unknown named level: %s", rb_id2name(id));
+    }
+
+    int level = NUM2INT(level_val);
+    if (level < spec->min || level > spec->max)
+        rb_raise(eLevelError, "%s level must be %d..%d, got %d", spec->name, spec->min, spec->max,
+                 level);
+    return level;
+}
+
+static compress_algo_t detect_algo(const uint8_t *data, size_t len) {
+    if (len >= 4) {
+        if (data[0] == 0x28 && data[1] == 0xB5 && data[2] == 0x2F && data[3] == 0xFD) {
+            return ALGO_ZSTD;
+        }
+    }
+
+    if (len >= 12) {
+        uint32_t orig = read_le_u32(data);
+        uint32_t comp = read_le_u32(data + 4);
+        if (orig > 0 && orig <= 256U * 1024 * 1024 && comp > 0 && comp <= 256U * 1024 * 1024 &&
+            orig <= (uint32_t)INT_MAX && comp <= (uint32_t)LZ4_compressBound((int)orig) &&
+            (size_t)8 + (size_t)comp + 4 == len) {
+            size_t tail = 8 + (size_t)comp;
+            if (data[tail] == 0 && data[tail + 1] == 0 && data[tail + 2] == 0 &&
+                data[tail + 3] == 0) {
+                return ALGO_LZ4;
+            }
+        }
+    }
+
+    rb_raise(eDataError, "cannot detect compression format (no magic bytes found). "
+                         "Use algo: :zstd, :lz4, or :brotli explicitly.");
+    return ALGO_ZSTD;
+}
+
+static inline VALUE rb_binary_str_new(const char *ptr, long len) {
+    VALUE str = rb_str_new(ptr, len);
+    rb_enc_associate(str, binary_encoding);
+    return str;
+}
+
+static inline VALUE rb_binary_str_buf_new(long capa) {
+    VALUE str = rb_str_buf_new(capa);
+    rb_enc_associate(str, binary_encoding);
+    return str;
+}
+
+static inline VALUE rb_binary_str_buf_reserve(long capa) {
+    VALUE str = rb_str_buf_new(capa);
+    rb_enc_associate(str, binary_encoding);
+    if (capa > 0)
+        rb_str_modify_expand(str, capa + 1);
+    return str;
+}
+
+static inline void grow_binary_str(VALUE str, size_t cur_len, size_t new_cap) {
+    size_t cur_cap = (size_t)rb_str_capacity(str);
+    if (new_cap <= cur_cap)
+        return;
+    rb_str_set_len(str, (long)cur_len);
+    rb_str_modify_expand(str, (long)(new_cap - cur_len));
+}
+
+typedef struct {
+    size_t max_output_size;
+    int max_ratio_enabled;
+    unsigned long long max_ratio;
+} limits_config_t;
+
+static void limits_config_init(limits_config_t *limits) {
+    limits->max_output_size = (size_t)MAX_DECOMPRESS_SIZE;
+    limits->max_ratio_enabled = 1;
+    limits->max_ratio = DEFAULT_MAX_RATIO;
+}
+
+static void limits_config_apply_opts(VALUE opts, limits_config_t *limits) {
+    if (NIL_P(opts))
+        return;
+
+    VALUE val = opt_lookup2(opts, sym_cache.max_output_size, Qundef);
+    if (val != Qundef && !NIL_P(val)) {
+        size_t max_output_size = NUM2SIZET(val);
+        if (max_output_size == 0)
+            rb_raise(rb_eArgError, "max_output_size must be greater than 0");
+        limits->max_output_size = max_output_size;
+    }
+
+    val = opt_lookup2(opts, sym_cache.max_ratio, Qundef);
+    if (val == Qundef)
+        return;
+    if (NIL_P(val)) {
+        limits->max_ratio_enabled = 0;
+        limits->max_ratio = 0;
+        return;
+    }
+
+    unsigned long long max_ratio = NUM2ULL(val);
+    if (max_ratio == 0)
+        rb_raise(rb_eArgError, "max_ratio must be greater than 0 or nil");
+    limits->max_ratio_enabled = 1;
+    limits->max_ratio = max_ratio;
+}
+
+static void parse_limits_from_opts(VALUE opts, limits_config_t *limits) {
+    limits_config_init(limits);
+    limits_config_apply_opts(opts, limits);
+}
+
+static size_t checked_add_size(size_t left, size_t right, const char *message) {
+    if (SIZE_MAX - left < right)
+        rb_raise(eDataError, "%s", message);
+    return left + right;
+}
+
+static size_t ratio_limit_bytes(size_t total_input, unsigned long long max_ratio) {
+    if (total_input == 0)
+        return SIZE_MAX;
+    if (max_ratio > ((unsigned long long)SIZE_MAX / (unsigned long long)total_input))
+        return SIZE_MAX;
+    return total_input * (size_t)max_ratio;
+}
+
+static void enforce_output_and_ratio_limits(size_t total_output, size_t total_input,
+                                            size_t max_output_size, int max_ratio_enabled,
+                                            unsigned long long max_ratio) {
+    if (total_output > max_output_size) {
+        rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)", max_output_size);
+    }
+
+    if (!max_ratio_enabled || total_input < RATIO_MIN_INPUT_BYTES)
+        return;
+
+    size_t ratio_limit = ratio_limit_bytes(total_input, max_ratio);
+    if (total_output > ratio_limit) {
+        size_t ratio = total_input == 0 ? 0 : (total_output / total_input);
+        rb_raise(eDataError, "decompression ratio exceeds limit (ratio=%zu, max=%llu)", ratio,
+                 max_ratio);
+    }
+}
+
+static VALUE current_fiber_scheduler(void) {
+    VALUE sched = rb_fiber_scheduler_current();
+    if (sched == Qnil || sched == Qfalse)
+        return Qnil;
+    return sched;
+}
+
+static int has_fiber_scheduler(void) {
+    return current_fiber_scheduler() != Qnil;
+}
+
+static void unblock_noop(void *arg) {
+    (void)arg;
+}
+
+static inline void run_without_gvl(void *(*func)(void *), void *arg) {
+    rb_thread_call_without_gvl(func, arg, unblock_noop, NULL);
+}
+
+typedef struct {
+    void *(*func)(void *);
+    void *arg;
+
+    VALUE scheduler;
+    VALUE blocker;
+    VALUE fiber;
+} fiber_worker_ctx_t;
+
+static void *fiber_worker_nogvl(void *arg) {
+    fiber_worker_ctx_t *c = (fiber_worker_ctx_t *)arg;
+    c->func(c->arg);
+    return NULL;
+}
+
+static VALUE fiber_worker_thread(void *arg) {
+    fiber_worker_ctx_t *c = (fiber_worker_ctx_t *)arg;
+    rb_thread_call_without_gvl(fiber_worker_nogvl, c, RUBY_UBF_PROCESS, NULL);
+    rb_fiber_scheduler_unblock(c->scheduler, c->blocker, c->fiber);
+    return Qnil;
+}
+
+static void run_via_fiber_worker(VALUE scheduler, void *(*func)(void *), void *arg) {
+    fiber_worker_ctx_t ctx = {
+        .func = func,
+        .arg = arg,
+        .scheduler = scheduler,
+        .blocker = rb_obj_alloc(rb_cObject),
+        .fiber = rb_fiber_current(),
+    };
+    VALUE th = rb_thread_create(fiber_worker_thread, &ctx);
+    rb_fiber_scheduler_block(scheduler, ctx.blocker, Qnil);
+    join_thread(th);
+}
+
+static inline size_t fiber_maybe_yield(size_t bytes_since_yield, size_t just_processed,
+                                       size_t yield_chunk, int *did_yield) {
+    *did_yield = 0;
+    bytes_since_yield += just_processed;
+    if (bytes_since_yield >= yield_chunk) {
+        VALUE scheduler = current_fiber_scheduler();
+        if (scheduler != Qnil) {
+            scheduler_yield(scheduler);
+            *did_yield = 1;
+        }
+        return 0;
+    }
+    return bytes_since_yield;
+}
+
+#define DICT_CDICT_CACHE_SIZE 4
+
+typedef struct {
+    int level;
+    ZSTD_CDict *cdict;
+} cdict_cache_entry_t;
+
+struct dictionary_s {
+    compress_algo_t algo;
+    uint8_t *data;
+    size_t size;
+
+    cdict_cache_entry_t cdict_cache[DICT_CDICT_CACHE_SIZE];
+    int cdict_cache_count;
+
+    ZSTD_DDict *ddict;
+};
+
+static void dict_free(void *ptr) {
+    dictionary_t *dict = (dictionary_t *)ptr;
+    for (int i = 0; i < dict->cdict_cache_count; i++) {
+        if (dict->cdict_cache[i].cdict)
+            ZSTD_freeCDict(dict->cdict_cache[i].cdict);
+    }
+    if (dict->ddict)
+        ZSTD_freeDDict(dict->ddict);
+    if (dict->data)
+        xfree(dict->data);
+    xfree(dict);
+}
+
+static size_t dict_memsize(const void *ptr) {
+    const dictionary_t *d = (const dictionary_t *)ptr;
+    size_t total = sizeof(dictionary_t) + d->size;
+
+    for (int i = 0; i < d->cdict_cache_count; i++) {
+        if (d->cdict_cache[i].cdict)
+            total += d->size + 4096;
+    }
+    if (d->ddict)
+        total += d->size + 4096;
+    return total;
+}
+
+static const rb_data_type_t dictionary_type = {
+    "Compress::Dictionary", {NULL, dict_free, dict_memsize}, 0, 0, RUBY_TYPED_FREE_IMMEDIATELY};
+
+static VALUE dict_alloc(VALUE klass) {
+    dictionary_t *d = ALLOC(dictionary_t);
+    memset(d, 0, sizeof(dictionary_t));
+    return TypedData_Wrap_Struct(klass, &dictionary_type, d);
+}
+
+static ZSTD_CDict *dict_get_cdict(dictionary_t *dict, int level) {
+    for (int i = 0; i < dict->cdict_cache_count; i++) {
+        if (dict->cdict_cache[i].level == level)
+            return dict->cdict_cache[i].cdict;
+    }
+
+    ZSTD_CDict *cdict = ZSTD_createCDict(dict->data, dict->size, level);
+    if (!cdict)
+        return NULL;
+
+    for (int i = 0; i < dict->cdict_cache_count; i++) {
+        if (dict->cdict_cache[i].level == level) {
+            ZSTD_freeCDict(cdict);
+            return dict->cdict_cache[i].cdict;
+        }
+    }
+
+    if (dict->cdict_cache_count < DICT_CDICT_CACHE_SIZE) {
+        dict->cdict_cache[dict->cdict_cache_count].level = level;
+        dict->cdict_cache[dict->cdict_cache_count].cdict = cdict;
+        dict->cdict_cache_count++;
+    } else {
+        ZSTD_CDict *old_cdict = dict->cdict_cache[0].cdict;
+        memmove(&dict->cdict_cache[0], &dict->cdict_cache[1],
+                sizeof(cdict_cache_entry_t) * (DICT_CDICT_CACHE_SIZE - 1));
+        dict->cdict_cache[DICT_CDICT_CACHE_SIZE - 1].level = level;
+        dict->cdict_cache[DICT_CDICT_CACHE_SIZE - 1].cdict = cdict;
+        if (old_cdict)
+            ZSTD_freeCDict(old_cdict);
+    }
+    return cdict;
+}
+
+static ZSTD_DDict *dict_get_ddict(dictionary_t *dict) {
+    if (!dict->ddict) {
+        dict->ddict = ZSTD_createDDict(dict->data, dict->size);
+    }
+    return dict->ddict;
+}
+
+typedef struct {
+    const char *src;
+    size_t src_len;
+    char *dst;
+    size_t dst_cap;
+    int level;
+    ZSTD_CDict *cdict;
+    size_t result;
+    int error;
+} zstd_compress_args_t;
+
+static void *zstd_compress_nogvl(void *arg) {
+    zstd_compress_args_t *a = (zstd_compress_args_t *)arg;
+    if (a->cdict) {
+        ZSTD_CCtx *cctx = ZSTD_createCCtx();
+        if (!cctx) {
+            a->error = 1;
+            return NULL;
+        }
+        a->result =
+            ZSTD_compress_usingCDict(cctx, a->dst, a->dst_cap, a->src, a->src_len, a->cdict);
+        ZSTD_freeCCtx(cctx);
+    } else {
+        a->result = ZSTD_compress(a->dst, a->dst_cap, a->src, a->src_len, a->level);
+    }
+    a->error = 0;
+    return NULL;
+}
+
+typedef struct {
+    const void *src;
+    size_t src_len;
+    void *dst;
+    size_t dst_cap;
+    ZSTD_DDict *ddict;
+    size_t result;
+    int error;
+} zstd_decompress_args_t;
+
+static void *zstd_decompress_nogvl(void *arg) {
+    zstd_decompress_args_t *a = (zstd_decompress_args_t *)arg;
+    if (a->ddict) {
+        ZSTD_DCtx *dctx = ZSTD_createDCtx();
+        if (!dctx) {
+            a->error = 1;
+            return NULL;
+        }
+        a->result =
+            ZSTD_decompress_usingDDict(dctx, a->dst, a->dst_cap, a->src, a->src_len, a->ddict);
+        ZSTD_freeDCtx(dctx);
+    } else {
+        a->result = ZSTD_decompress(a->dst, a->dst_cap, a->src, a->src_len);
+    }
+    a->error = 0;
+    return NULL;
+}
+
+typedef struct {
+    const uint8_t *src;
+    size_t src_len;
+    char *dst;
+    size_t out_offset;
+    int error;
+    char err_msg[64];
+} lz4_decompress_all_args_t;
+
+static void *lz4_decompress_all_nogvl(void *arg) {
+    lz4_decompress_all_args_t *a = (lz4_decompress_all_args_t *)arg;
+    const uint8_t *src = a->src;
+    size_t slen = a->src_len;
+    char *out_ptr = a->dst;
+    size_t out_offset = 0;
+    size_t pos = 0;
+
+    while (pos + 4 <= slen) {
+        uint32_t orig_size = read_le_u32(src + pos);
+        if (orig_size == 0)
+            break;
+        uint32_t comp_size = read_le_u32(src + pos + 4);
+
+        int dsize = LZ4_decompress_safe((const char *)(src + pos + 8), out_ptr + out_offset,
+                                        (int)comp_size, (int)orig_size);
+        if (dsize < 0) {
+            a->error = 1;
+            snprintf(a->err_msg, sizeof(a->err_msg), "lz4 decompress failed");
+            return NULL;
+        }
+
+        out_offset += dsize;
+        pos += 8 + comp_size;
+    }
+
+    a->out_offset = out_offset;
+    a->error = 0;
+    return NULL;
+}
+
+typedef struct {
+    const char *src;
+    int src_len;
+    char *dst;
+    int dst_cap;
+    int level;
+    int result;
+} lz4_compress_args_t;
+
+static void *lz4_compress_nogvl(void *arg) {
+    lz4_compress_args_t *a = (lz4_compress_args_t *)arg;
+    if (a->level > 1) {
+        a->result = LZ4_compress_HC(a->src, a->dst, a->src_len, a->dst_cap, a->level);
+    } else {
+        a->result = LZ4_compress_default(a->src, a->dst, a->src_len, a->dst_cap);
+    }
+    return NULL;
+}
+
+typedef struct {
+    int level;
+    size_t src_len;
+    const uint8_t *src;
+    size_t *out_len;
+    uint8_t *dst;
+    BROTLI_BOOL result;
+} brotli_compress_args_t;
+
+static void *brotli_compress_nogvl(void *arg) {
+    brotli_compress_args_t *a = (brotli_compress_args_t *)arg;
+    a->result = BrotliEncoderCompress(a->level, BROTLI_DEFAULT_WINDOW, BROTLI_DEFAULT_MODE,
+                                      a->src_len, a->src, a->out_len, a->dst);
+    return NULL;
+}
+
+typedef struct {
+    ZSTD_CStream *cstream;
+    ZSTD_outBuffer *output;
+    ZSTD_inBuffer *input;
+    size_t result;
+} zstd_stream_chunk_args_t;
+
+static void *zstd_compress_stream_chunk_nogvl(void *arg) {
+    zstd_stream_chunk_args_t *a = (zstd_stream_chunk_args_t *)arg;
+    a->result = ZSTD_compressStream(a->cstream, a->output, a->input);
+    return NULL;
+}
+
+typedef struct {
+    const char *src;
+    size_t src_len;
+    int level;
+    ZSTD_CDict *cdict;
+    char *dst;
+    size_t dst_cap;
+    size_t result;
+    int error;
+
+    VALUE scheduler;
+    VALUE blocker;
+    VALUE fiber;
+} zstd_fiber_compress_t;
+
+typedef struct {
+    ZSTD_CStream *cstream;
+    ZSTD_inBuffer *input;
+    ZSTD_outBuffer *output;
+    size_t result;
+
+    VALUE scheduler;
+    VALUE blocker;
+    VALUE fiber;
+} zstd_stream_chunk_fiber_t;
+
+static void *zstd_stream_chunk_fiber_nogvl(void *arg) {
+    zstd_stream_chunk_fiber_t *a = (zstd_stream_chunk_fiber_t *)arg;
+    a->result = ZSTD_compressStream(a->cstream, a->output, a->input);
+    return NULL;
+}
+
+static VALUE zstd_stream_chunk_fiber_thread(void *arg) {
+    zstd_stream_chunk_fiber_t *a = (zstd_stream_chunk_fiber_t *)arg;
+    rb_thread_call_without_gvl(zstd_stream_chunk_fiber_nogvl, a, RUBY_UBF_PROCESS, NULL);
+    rb_fiber_scheduler_unblock(a->scheduler, a->blocker, a->fiber);
+    return Qnil;
+}
+
+typedef struct {
+    BrotliEncoderState *enc;
+    BrotliEncoderOperation op;
+    size_t *available_in;
+    const uint8_t **next_in;
+    size_t *available_out;
+    uint8_t **next_out;
+    BROTLI_BOOL result;
+
+    VALUE scheduler;
+    VALUE blocker;
+    VALUE fiber;
+} brotli_stream_chunk_fiber_t;
+
+static void *brotli_stream_chunk_fiber_nogvl(void *arg) {
+    brotli_stream_chunk_fiber_t *a = (brotli_stream_chunk_fiber_t *)arg;
+    a->result = BrotliEncoderCompressStream(a->enc, a->op, a->available_in, a->next_in,
+                                            a->available_out, a->next_out, NULL);
+    return NULL;
+}
+
+static VALUE brotli_stream_chunk_fiber_thread(void *arg) {
+    brotli_stream_chunk_fiber_t *a = (brotli_stream_chunk_fiber_t *)arg;
+    rb_thread_call_without_gvl(brotli_stream_chunk_fiber_nogvl, a, RUBY_UBF_PROCESS, NULL);
+    rb_fiber_scheduler_unblock(a->scheduler, a->blocker, a->fiber);
+    return Qnil;
+}
+
+typedef struct {
+    size_t encoded_size;
+    const uint8_t *encoded_buffer;
+    size_t *decoded_size;
+    uint8_t *decoded_buffer;
+    BrotliDecoderResult result;
+} brotli_decompress_args_t;
+
+static void *brotli_decompress_nogvl(void *arg) {
+    brotli_decompress_args_t *a = (brotli_decompress_args_t *)arg;
+    a->result = BrotliDecoderDecompress(a->encoded_size, a->encoded_buffer, a->decoded_size,
+                                        a->decoded_buffer);
+    return NULL;
+}
+
+typedef struct {
+    ZSTD_DStream *dstream;
+    ZSTD_outBuffer *output;
+    ZSTD_inBuffer *input;
+    size_t result;
+} zstd_decompress_stream_chunk_args_t;
+
+static void *zstd_decompress_stream_chunk_nogvl(void *arg) {
+    zstd_decompress_stream_chunk_args_t *a = (zstd_decompress_stream_chunk_args_t *)arg;
+    a->result = ZSTD_decompressStream(a->dstream, a->output, a->input);
+    return NULL;
+}
+
+typedef struct {
+    BrotliDecoderState *dec;
+    size_t *available_in;
+    const uint8_t **next_in;
+    size_t *available_out;
+    uint8_t **next_out;
+    BrotliDecoderResult result;
+} brotli_decompress_stream_args_t;
+
+static void *brotli_decompress_stream_nogvl(void *arg) {
+    brotli_decompress_stream_args_t *a = (brotli_decompress_stream_args_t *)arg;
+    a->result = BrotliDecoderDecompressStream(a->dec, a->available_in, a->next_in, a->available_out,
+                                              a->next_out, NULL);
+    return NULL;
+}
+
+static void *zstd_fiber_compress_nogvl(void *arg) {
+    zstd_fiber_compress_t *a = (zstd_fiber_compress_t *)arg;
+    if (a->cdict) {
+        ZSTD_CCtx *cctx = ZSTD_createCCtx();
+        if (!cctx) {
+            a->error = 1;
+            return NULL;
+        }
+        a->result =
+            ZSTD_compress_usingCDict(cctx, a->dst, a->dst_cap, a->src, a->src_len, a->cdict);
+        ZSTD_freeCCtx(cctx);
+    } else {
+        a->result = ZSTD_compress(a->dst, a->dst_cap, a->src, a->src_len, a->level);
+    }
+    return NULL;
+}
+
+static VALUE zstd_fiber_compress_thread(void *arg) {
+    zstd_fiber_compress_t *a = (zstd_fiber_compress_t *)arg;
+    rb_thread_call_without_gvl(zstd_fiber_compress_nogvl, a, RUBY_UBF_PROCESS, NULL);
+    rb_fiber_scheduler_unblock(a->scheduler, a->blocker, a->fiber);
+    return Qnil;
+}
+
+static VALUE compress_compress(int argc, VALUE *argv, VALUE self) {
+    VALUE data, opts;
+    rb_scan_args(argc, argv, "1:", &data, &opts);
+    StringValue(data);
+    reject_algorithm_keyword(opts);
+
+    VALUE algo_sym = Qnil, level_val = Qnil, dict_val = Qnil;
+    if (!NIL_P(opts)) {
+        algo_sym = opt_get(opts, sym_cache.algo);
+        level_val = opt_get(opts, sym_cache.level);
+        dict_val = opt_get(opts, sym_cache.dictionary);
+    }
+
+    compress_algo_t algo = NIL_P(algo_sym) ? ALGO_ZSTD : sym_to_algo(algo_sym);
+    int level = resolve_level(algo, level_val);
+
+    dictionary_t *dict = NULL;
+    if (!NIL_P(dict_val)) {
+        if (algo == ALGO_LZ4) {
+            rb_raise(eUnsupportedError, "LZ4 does not support dictionaries");
+        }
+        dict = opt_dictionary(dict_val);
+    }
+
+    const char *src = RSTRING_PTR(data);
+    size_t slen = RSTRING_LEN(data);
+    const algo_policy_t *policy = algo_policy(algo);
+
+    switch (algo) {
+    case ALGO_ZSTD: {
+        size_t bound = ZSTD_compressBound(slen);
+
+        ZSTD_CDict *cdict = NULL;
+        if (dict) {
+            cdict = dict_get_cdict(dict, level);
+            if (!cdict)
+                rb_raise(eMemError, "zstd: failed to create/get cdict");
+        }
+
+        if (slen < policy->gvl_unlock_threshold) {
+            VALUE dst = rb_binary_str_buf_reserve(bound);
+            size_t csize;
+            if (cdict) {
+                ZSTD_CCtx *cctx = ZSTD_createCCtx();
+                if (!cctx)
+                    rb_raise(eMemError, "zstd: failed to create context");
+                csize = ZSTD_compress_usingCDict(cctx, RSTRING_PTR(dst), bound, src, slen, cdict);
+                ZSTD_freeCCtx(cctx);
+            } else {
+                csize = ZSTD_compress(RSTRING_PTR(dst), bound, src, slen, level);
+            }
+            if (ZSTD_isError(csize))
+                rb_raise(eError, "zstd compress: %s", ZSTD_getErrorName(csize));
+            rb_str_set_len(dst, (long)csize);
+            RB_GC_GUARD(data);
+            return dst;
+        }
+
+        {
+            VALUE scheduler = current_fiber_scheduler();
+            if (scheduler != Qnil) {
+                char *out_buf = (char *)malloc(bound);
+                if (!out_buf)
+                    rb_raise(eMemError, "zstd: malloc failed");
+
+                VALUE blocker = rb_obj_alloc(rb_cObject);
+
+                zstd_fiber_compress_t fargs = {
+                    .src = src,
+                    .src_len = slen,
+                    .level = level,
+                    .cdict = cdict,
+                    .dst = out_buf,
+                    .dst_cap = bound,
+                    .result = 0,
+                    .error = 0,
+                    .scheduler = scheduler,
+                    .blocker = blocker,
+                    .fiber = rb_fiber_current(),
+                };
+
+                VALUE rb_thread = rb_thread_create(zstd_fiber_compress_thread, &fargs);
+                rb_fiber_scheduler_block(scheduler, blocker, Qnil);
+                join_thread(rb_thread);
+
+                if (fargs.error) {
+                    free(out_buf);
+                    rb_raise(eMemError, "zstd: failed to create context");
+                }
+                if (ZSTD_isError(fargs.result)) {
+                    free(out_buf);
+                    rb_raise(eError, "zstd compress: %s", ZSTD_getErrorName(fargs.result));
+                }
+
+                VALUE result = rb_binary_str_new(out_buf, (long)fargs.result);
+                free(out_buf);
+                RB_GC_GUARD(data);
+                return result;
+            }
+        }
+
+        {
+            VALUE dst = rb_binary_str_buf_reserve(bound);
+            zstd_compress_args_t args = {
+                .src = src,
+                .src_len = slen,
+                .dst = RSTRING_PTR(dst),
+                .dst_cap = bound,
+                .level = level,
+                .cdict = cdict,
+                .result = 0,
+                .error = 0,
+            };
+            run_without_gvl(zstd_compress_nogvl, &args);
+
+            if (args.error)
+                rb_raise(eMemError, "zstd: failed to create context");
+            if (ZSTD_isError(args.result))
+                rb_raise(eError, "zstd compress: %s", ZSTD_getErrorName(args.result));
+
+            rb_str_set_len(dst, (long)args.result);
+            RB_GC_GUARD(data);
+            return dst;
+        }
+    }
+    case ALGO_LZ4: {
+        if (slen > (size_t)INT_MAX)
+            rb_raise(eError, "lz4: input too large (max 2GB)");
+        int bound = LZ4_compressBound((int)slen);
+
+        int csize;
+        if (slen >= policy->gvl_unlock_threshold) {
+            VALUE dst = rb_binary_str_buf_reserve(8 + (size_t)bound + 4);
+            char *out = RSTRING_PTR(dst);
+
+            write_le_u32((uint8_t *)out, (uint32_t)slen);
+
+            lz4_compress_args_t args = {
+                .src = src,
+                .src_len = (int)slen,
+                .dst = out + 8,
+                .dst_cap = bound,
+                .level = level,
+            };
+
+            VALUE scheduler = current_fiber_scheduler();
+            if (scheduler != Qnil) {
+                run_via_fiber_worker(scheduler, lz4_compress_nogvl, &args);
+            } else {
+                run_without_gvl(lz4_compress_nogvl, &args);
+            }
+            csize = args.result;
+
+            if (csize <= 0)
+                rb_raise(eError, "lz4 compress failed");
+
+            write_le_u32((uint8_t *)out + 4, (uint32_t)csize);
+
+            size_t total = 8 + (size_t)csize;
+            write_le_u32((uint8_t *)out + total, 0);
+
+            rb_str_set_len(dst, (long)(total + 4));
+            RB_GC_GUARD(data);
+            return dst;
+        } else {
+            VALUE dst = rb_binary_str_buf_reserve(8 + bound + 4);
+            char *out = RSTRING_PTR(dst);
+
+            write_le_u32((uint8_t *)out, (uint32_t)slen);
+
+            if (level > 1) {
+                csize = LZ4_compress_HC(src, out + 8, (int)slen, bound, level);
+            } else {
+                csize = LZ4_compress_default(src, out + 8, (int)slen, bound);
+            }
+            if (csize <= 0)
+                rb_raise(eError, "lz4 compress failed");
+
+            write_le_u32((uint8_t *)out + 4, (uint32_t)csize);
+
+            size_t total = 8 + csize;
+            write_le_u32((uint8_t *)out + total, 0);
+
+            rb_str_set_len(dst, total + 4);
+            RB_GC_GUARD(data);
+            return dst;
+        }
+    }
+    case ALGO_BROTLI: {
+        size_t out_len = BrotliEncoderMaxCompressedSize(slen);
+        if (out_len == 0)
+            out_len = slen + (slen >> 2) + 1024;
+
+        if (dict) {
+            VALUE dst = rb_binary_str_buf_reserve(out_len);
+            BrotliEncoderState *enc = BrotliEncoderCreateInstance(NULL, NULL, NULL);
+            if (!enc)
+                rb_raise(eMemError, "brotli: failed to create encoder");
+
+            BrotliEncoderPreparedDictionary *pd =
+                BrotliEncoderPrepareDictionary(BROTLI_SHARED_DICTIONARY_RAW, dict->size, dict->data,
+                                               BROTLI_MAX_QUALITY, NULL, NULL, NULL);
+            if (!pd) {
+                BrotliEncoderDestroyInstance(enc);
+                rb_raise(eMemError, "brotli: failed to prepare dictionary");
+            }
+
+            if (!BrotliEncoderSetParameter(enc, BROTLI_PARAM_QUALITY, level) ||
+                !BrotliEncoderAttachPreparedDictionary(enc, pd)) {
+                BrotliEncoderDestroyPreparedDictionary(pd);
+                BrotliEncoderDestroyInstance(enc);
+                rb_raise(eError, "brotli: failed to attach dictionary");
+            }
+
+            size_t available_in = slen;
+            const uint8_t *next_in = (const uint8_t *)src;
+            size_t available_out = out_len;
+            uint8_t *next_out = (uint8_t *)RSTRING_PTR(dst);
+            size_t initial_out = available_out;
+
+            BROTLI_BOOL ok =
+                BrotliEncoderCompressStream(enc, BROTLI_OPERATION_FINISH, &available_in, &next_in,
+                                            &available_out, &next_out, NULL);
+
+            BrotliEncoderDestroyPreparedDictionary(pd);
+            BrotliEncoderDestroyInstance(enc);
+            if (!ok)
+                rb_raise(eError, "brotli compress with dict failed");
+
+            rb_str_set_len(dst, initial_out - available_out);
+            RB_GC_GUARD(data);
+            return dst;
+        } else if (slen >= policy->gvl_unlock_threshold) {
+            VALUE dst = rb_binary_str_buf_reserve(out_len);
+            size_t actual_out_len = out_len;
+
+            brotli_compress_args_t args = {
+                .level = level,
+                .src_len = slen,
+                .src = (const uint8_t *)src,
+                .out_len = &actual_out_len,
+                .dst = (uint8_t *)RSTRING_PTR(dst),
+            };
+
+            VALUE scheduler = current_fiber_scheduler();
+            if (scheduler != Qnil) {
+                run_via_fiber_worker(scheduler, brotli_compress_nogvl, &args);
+            } else {
+                run_without_gvl(brotli_compress_nogvl, &args);
+            }
+
+            if (!args.result)
+                rb_raise(eError, "brotli compress failed");
+
+            rb_str_set_len(dst, (long)actual_out_len);
+            RB_GC_GUARD(data);
+            return dst;
+        } else {
+            VALUE dst = rb_binary_str_buf_reserve(out_len);
+            BROTLI_BOOL ok =
+                BrotliEncoderCompress(level, BROTLI_DEFAULT_WINDOW, BROTLI_DEFAULT_MODE, slen,
+                                      (const uint8_t *)src, &out_len, (uint8_t *)RSTRING_PTR(dst));
+            if (!ok)
+                rb_raise(eError, "brotli compress failed");
+            rb_str_set_len(dst, out_len);
+            RB_GC_GUARD(data);
+            return dst;
+        }
+    }
+    }
+
+    return Qnil;
+}
+
+static VALUE compress_decompress(int argc, VALUE *argv, VALUE self) {
+    VALUE data, opts;
+    rb_scan_args(argc, argv, "1:", &data, &opts);
+    StringValue(data);
+    reject_algorithm_keyword(opts);
+
+    VALUE algo_sym = Qnil, dict_val = Qnil;
+    limits_config_t limits;
+    parse_limits_from_opts(opts, &limits);
+    if (!NIL_P(opts)) {
+        algo_sym = opt_get(opts, sym_cache.algo);
+        dict_val = opt_get(opts, sym_cache.dictionary);
+    }
+
+    const uint8_t *src = (const uint8_t *)RSTRING_PTR(data);
+    size_t slen = RSTRING_LEN(data);
+
+    compress_algo_t algo;
+    if (NIL_P(algo_sym)) {
+        algo = detect_algo(src, slen);
+    } else {
+        algo = sym_to_algo(algo_sym);
+    }
+
+    const algo_policy_t *policy = algo_policy(algo);
+
+    dictionary_t *dict = NULL;
+    if (!NIL_P(dict_val)) {
+        if (algo == ALGO_LZ4) {
+            rb_raise(eUnsupportedError, "LZ4 does not support dictionaries");
+        }
+        dict = opt_dictionary(dict_val);
+    }
+
+    switch (algo) {
+    case ALGO_ZSTD: {
+        unsigned long long frame_size = ZSTD_getFrameContentSize(src, slen);
+        if (frame_size == ZSTD_CONTENTSIZE_ERROR) {
+            rb_raise(eDataError, "zstd: not valid compressed data");
+        }
+
+        if (frame_size != ZSTD_CONTENTSIZE_UNKNOWN) {
+            if (frame_size > limits.max_output_size) {
+                rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)",
+                         limits.max_output_size);
+            }
+            enforce_output_and_ratio_limits((size_t)frame_size, slen, limits.max_output_size,
+                                            limits.max_ratio_enabled, limits.max_ratio);
+        }
+
+        if (frame_size != ZSTD_CONTENTSIZE_UNKNOWN && frame_size <= limits.max_output_size) {
+            size_t dsize;
+
+            if (frame_size >= algo_policy(ALGO_ZSTD)->gvl_unlock_threshold) {
+                VALUE dst = rb_binary_str_buf_reserve((size_t)frame_size);
+
+                ZSTD_DDict *ddict = NULL;
+                if (dict) {
+                    ddict = dict_get_ddict(dict);
+                    if (!ddict)
+                        rb_raise(eMemError, "zstd: failed to create ddict");
+                }
+
+                zstd_decompress_args_t args = {
+                    .src = src,
+                    .src_len = slen,
+                    .dst = RSTRING_PTR(dst),
+                    .dst_cap = (size_t)frame_size,
+                    .ddict = ddict,
+                };
+
+                VALUE scheduler = current_fiber_scheduler();
+                if (scheduler != Qnil) {
+                    run_via_fiber_worker(scheduler, zstd_decompress_nogvl, &args);
+                } else {
+                    run_without_gvl(zstd_decompress_nogvl, &args);
+                }
+
+                if (args.error)
+                    rb_raise(eMemError, "zstd: failed to create dctx");
+                dsize = args.result;
+                if (ZSTD_isError(dsize))
+                    rb_raise(eDataError, "zstd decompress: %s", ZSTD_getErrorName(dsize));
+
+                enforce_output_and_ratio_limits(dsize, slen, limits.max_output_size,
+                                                limits.max_ratio_enabled, limits.max_ratio);
+                rb_str_set_len(dst, (long)dsize);
+                RB_GC_GUARD(data);
+                return dst;
+            } else {
+                VALUE dst = rb_binary_str_buf_reserve((size_t)frame_size);
+
+                if (dict) {
+                    ZSTD_DDict *ddict = dict_get_ddict(dict);
+                    if (!ddict)
+                        rb_raise(eMemError, "zstd: failed to create ddict");
+                    ZSTD_DCtx *dctx = ZSTD_createDCtx();
+                    if (!dctx)
+                        rb_raise(eMemError, "zstd: failed to create dctx");
+                    dsize = ZSTD_decompress_usingDDict(dctx, RSTRING_PTR(dst), (size_t)frame_size,
+                                                       src, slen, ddict);
+                    ZSTD_freeDCtx(dctx);
+                } else {
+                    dsize = ZSTD_decompress(RSTRING_PTR(dst), (size_t)frame_size, src, slen);
+                }
+
+                if (ZSTD_isError(dsize))
+                    rb_raise(eDataError, "zstd decompress: %s", ZSTD_getErrorName(dsize));
+                enforce_output_and_ratio_limits(dsize, slen, limits.max_output_size,
+                                                limits.max_ratio_enabled, limits.max_ratio);
+                rb_str_set_len(dst, dsize);
+                RB_GC_GUARD(data);
+                return dst;
+            }
+        }
+
+        ZSTD_DCtx *dctx = ZSTD_createDCtx();
+        if (!dctx)
+            rb_raise(eMemError, "zstd: failed to create dctx");
+
+        if (dict) {
+            ZSTD_DDict *ddict = dict_get_ddict(dict);
+            if (ddict) {
+                size_t r = ZSTD_DCtx_refDDict(dctx, ddict);
+                if (ZSTD_isError(r)) {
+                    ZSTD_freeDCtx(dctx);
+                    rb_raise(eError, "zstd dict ref: %s", ZSTD_getErrorName(r));
+                }
+            }
+        }
+
+        size_t alloc_size = (slen > limits.max_output_size / 8) ? limits.max_output_size : slen * 8;
+        if (alloc_size < 4096)
+            alloc_size = limits.max_output_size < 4096 ? limits.max_output_size : 4096;
+        if (alloc_size == 0)
+            alloc_size = limits.max_output_size;
+
+        VALUE dst = rb_binary_str_buf_reserve(alloc_size);
+        size_t total_out = 0;
+
+        ZSTD_inBuffer input = {src, slen, 0};
+        while (input.pos < input.size) {
+            if (total_out >= alloc_size) {
+                if (alloc_size >= limits.max_output_size) {
+                    ZSTD_freeDCtx(dctx);
+                    rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)",
+                             limits.max_output_size);
+                }
+                size_t next_cap = alloc_size * 2;
+                if (next_cap > limits.max_output_size)
+                    next_cap = limits.max_output_size;
+                alloc_size = next_cap;
+                grow_binary_str(dst, total_out, alloc_size);
+            }
+
+            size_t remaining_budget = limits.max_output_size - total_out;
+            if (remaining_budget == 0) {
+                ZSTD_freeDCtx(dctx);
+                rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)",
+                         limits.max_output_size);
+            }
+
+            size_t out_cap = alloc_size - total_out;
+            if (out_cap > remaining_budget)
+                out_cap = remaining_budget;
+
+            ZSTD_outBuffer output = {RSTRING_PTR(dst) + total_out, out_cap, 0};
+            size_t ret = ZSTD_decompressStream(dctx, &output, &input);
+            if (ZSTD_isError(ret)) {
+                ZSTD_freeDCtx(dctx);
+                rb_raise(eDataError, "zstd decompress: %s", ZSTD_getErrorName(ret));
+            }
+            total_out = checked_add_size(total_out, output.pos,
+                                         "decompressed output exceeds representable size");
+            enforce_output_and_ratio_limits(total_out, slen, limits.max_output_size,
+                                            limits.max_ratio_enabled, limits.max_ratio);
+            if (ret == 0)
+                break;
+        }
+
+        ZSTD_freeDCtx(dctx);
+        rb_str_set_len(dst, total_out);
+        RB_GC_GUARD(data);
+        return dst;
+    }
+    case ALGO_LZ4: {
+        if (slen < 4)
+            rb_raise(eDataError, "lz4: data too short");
+
+        size_t total_orig = 0;
+        size_t scan_pos = 0;
+        while (scan_pos + 4 <= slen) {
+            uint32_t orig_size = read_le_u32(src + scan_pos);
+            if (orig_size == 0)
+                break;
+            if (scan_pos + 8 > slen)
+                rb_raise(eDataError, "lz4: truncated block header");
+            uint32_t comp_size = read_le_u32(src + scan_pos + 4);
+            if (scan_pos + 8 + comp_size > slen)
+                rb_raise(eDataError, "lz4: truncated block data");
+            if (orig_size > 256 * 1024 * 1024)
+                rb_raise(eDataError, "lz4: block too large (%u)", orig_size);
+            total_orig = checked_add_size(total_orig, orig_size,
+                                          "decompressed output exceeds representable size");
+            if (total_orig > limits.max_output_size)
+                rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)",
+                         limits.max_output_size);
+            scan_pos += 8 + comp_size;
+        }
+
+        enforce_output_and_ratio_limits(total_orig, slen, limits.max_output_size,
+                                        limits.max_ratio_enabled, limits.max_ratio);
+
+        VALUE result = rb_binary_str_buf_reserve(total_orig);
+
+        lz4_decompress_all_args_t args = {
+            .src = src,
+            .src_len = slen,
+            .dst = RSTRING_PTR(result),
+            .out_offset = 0,
+            .error = 0,
+        };
+
+        if (total_orig >= algo_policy(ALGO_LZ4)->gvl_unlock_threshold) {
+            VALUE scheduler = current_fiber_scheduler();
+            if (scheduler != Qnil) {
+                run_via_fiber_worker(scheduler, lz4_decompress_all_nogvl, &args);
+            } else {
+                run_without_gvl(lz4_decompress_all_nogvl, &args);
+            }
+        } else {
+            lz4_decompress_all_nogvl(&args);
+        }
+
+        if (args.error)
+            rb_raise(eDataError, "%s", args.err_msg);
+
+        enforce_output_and_ratio_limits(args.out_offset, slen, limits.max_output_size,
+                                        limits.max_ratio_enabled, limits.max_ratio);
+        rb_str_set_len(result, args.out_offset);
+        RB_GC_GUARD(data);
+        return result;
+    }
+    case ALGO_BROTLI: {
+        size_t alloc_size = (slen > limits.max_output_size / 4) ? limits.max_output_size : slen * 4;
+        if (alloc_size < 1024)
+            alloc_size = limits.max_output_size < 1024 ? limits.max_output_size : 1024;
+        if (alloc_size == 0)
+            alloc_size = limits.max_output_size;
+
+        BrotliDecoderState *dec = BrotliDecoderCreateInstance(NULL, NULL, NULL);
+        if (!dec)
+            rb_raise(eMemError, "brotli: failed to create decoder");
+
+        if (dict) {
+            BrotliDecoderAttachDictionary(dec, BROTLI_SHARED_DICTIONARY_RAW, dict->size,
+                                          dict->data);
+        }
+
+        VALUE dst = rb_binary_str_buf_reserve(alloc_size);
+        size_t total_out = 0;
+
+        size_t available_in = slen;
+        const uint8_t *next_in = src;
+
+        VALUE scheduler = current_fiber_scheduler();
+
+        BrotliDecoderResult res = BROTLI_DECODER_RESULT_NEEDS_MORE_OUTPUT;
+        while (res == BROTLI_DECODER_RESULT_NEEDS_MORE_OUTPUT) {
+            size_t remaining_budget = limits.max_output_size - total_out;
+            if (remaining_budget == 0) {
+                BrotliDecoderDestroyInstance(dec);
+                rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)",
+                         limits.max_output_size);
+            }
+
+            size_t available_out = alloc_size - total_out;
+            if (available_out > remaining_budget)
+                available_out = remaining_budget;
+            uint8_t *next_out = (uint8_t *)RSTRING_PTR(dst) + total_out;
+
+            if (scheduler != Qnil && available_in >= policy->fiber_stream_threshold) {
+                brotli_decompress_stream_args_t sargs = {
+                    .dec = dec,
+                    .available_in = &available_in,
+                    .next_in = &next_in,
+                    .available_out = &available_out,
+                    .next_out = &next_out,
+                    .result = BROTLI_DECODER_RESULT_ERROR,
+                };
+                run_via_fiber_worker(scheduler, brotli_decompress_stream_nogvl, &sargs);
+                res = sargs.result;
+            } else {
+                res = BrotliDecoderDecompressStream(dec, &available_in, &next_in, &available_out,
+                                                    &next_out, NULL);
+            }
+
+            total_out = next_out - (uint8_t *)RSTRING_PTR(dst);
+            enforce_output_and_ratio_limits(total_out, slen, limits.max_output_size,
+                                            limits.max_ratio_enabled, limits.max_ratio);
+
+            if (res == BROTLI_DECODER_RESULT_NEEDS_MORE_OUTPUT) {
+                if (alloc_size >= limits.max_output_size) {
+                    BrotliDecoderDestroyInstance(dec);
+                    rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)",
+                             limits.max_output_size);
+                }
+                size_t next_cap = alloc_size * 2;
+                if (next_cap > limits.max_output_size)
+                    next_cap = limits.max_output_size;
+                alloc_size = next_cap;
+                grow_binary_str(dst, total_out, alloc_size);
+            }
+        }
+
+        BrotliDecoderDestroyInstance(dec);
+
+        if (res != BROTLI_DECODER_RESULT_SUCCESS) {
+            rb_raise(eDataError, "brotli decompress failed");
+        }
+        rb_str_set_len(dst, total_out);
+        RB_GC_GUARD(data);
+        return dst;
+    }
+    }
+
+    return Qnil;
+}
+
+static uint32_t crc32_tables[8][256];
+static int crc32_tables_initialized = 0;
+
+static void crc32_init_tables(void) {
+    if (crc32_tables_initialized)
+        return;
+
+    for (uint32_t i = 0; i < 256; i++) {
+        uint32_t crc = i;
+        for (int j = 0; j < 8; j++) {
+            crc = (crc >> 1) ^ (0xEDB88320 & (-(int32_t)(crc & 1)));
+        }
+        crc32_tables[0][i] = crc;
+    }
+
+    for (uint32_t i = 0; i < 256; i++) {
+        uint32_t crc = crc32_tables[0][i];
+        for (int t = 1; t < 8; t++) {
+            crc = crc32_tables[0][crc & 0xFF] ^ (crc >> 8);
+            crc32_tables[t][i] = crc;
+        }
+    }
+
+    crc32_tables_initialized = 1;
+}
+
+static uint32_t crc32_compute(const uint8_t *data, size_t len, uint32_t crc) {
+    crc = ~crc;
+
+    while (len >= 8) {
+        uint32_t val0 = crc ^ ((uint32_t)data[0] | ((uint32_t)data[1] << 8) |
+                               ((uint32_t)data[2] << 16) | ((uint32_t)data[3] << 24));
+        uint32_t val1 = (uint32_t)data[4] | ((uint32_t)data[5] << 8) | ((uint32_t)data[6] << 16) |
+                        ((uint32_t)data[7] << 24);
+
+        crc = crc32_tables[7][(val0) & 0xFF] ^ crc32_tables[6][(val0 >> 8) & 0xFF] ^
+              crc32_tables[5][(val0 >> 16) & 0xFF] ^ crc32_tables[4][(val0 >> 24) & 0xFF] ^
+              crc32_tables[3][(val1) & 0xFF] ^ crc32_tables[2][(val1 >> 8) & 0xFF] ^
+              crc32_tables[1][(val1 >> 16) & 0xFF] ^ crc32_tables[0][(val1 >> 24) & 0xFF];
+
+        data += 8;
+        len -= 8;
+    }
+
+    while (len--) {
+        crc = crc32_tables[0][(crc ^ *data++) & 0xFF] ^ (crc >> 8);
+    }
+
+    return ~crc;
+}
+
+static VALUE compress_crc32(int argc, VALUE *argv, VALUE self) {
+    VALUE data, prev;
+    rb_scan_args(argc, argv, "11", &data, &prev);
+    StringValue(data);
+
+    const uint8_t *src = (const uint8_t *)RSTRING_PTR(data);
+    size_t len = RSTRING_LEN(data);
+    uint32_t crc = NIL_P(prev) ? 0 : NUM2UINT(prev);
+
+    return UINT2NUM(crc32_compute(src, len, crc));
+}
+
+static VALUE compress_adler32(int argc, VALUE *argv, VALUE self) {
+    VALUE data, prev;
+    rb_scan_args(argc, argv, "11", &data, &prev);
+    StringValue(data);
+
+    const uint8_t *src = (const uint8_t *)RSTRING_PTR(data);
+    size_t len = RSTRING_LEN(data);
+    uint32_t adler = NIL_P(prev) ? 1 : NUM2UINT(prev);
+
+    uint32_t s1 = adler & 0xFFFF;
+    uint32_t s2 = (adler >> 16) & 0xFFFF;
+    const uint32_t BASE = 65521;
+
+    while (len > 0) {
+        size_t chunk = len > 5552 ? 5552 : len;
+        len -= chunk;
+        for (size_t i = 0; i < chunk; i++) {
+            s1 += src[i];
+            s2 += s1;
+        }
+        s1 %= BASE;
+        s2 %= BASE;
+        src += chunk;
+    }
+
+    return UINT2NUM((s2 << 16) | s1);
+}
+
+static VALUE compress_algorithms(VALUE self) {
+    VALUE ary = rb_ary_new_capa(3);
+    rb_ary_push(ary, sym_cache.zstd);
+    rb_ary_push(ary, sym_cache.lz4);
+    rb_ary_push(ary, sym_cache.brotli);
+    return ary;
+}
+
+static VALUE compress_available_p(VALUE self, VALUE algo_sym) {
+    sym_to_algo(algo_sym);
+    return Qtrue;
+}
+
+static VALUE compress_version(VALUE self, VALUE algo_sym) {
+    compress_algo_t algo = sym_to_algo(algo_sym);
+    switch (algo) {
+    case ALGO_ZSTD:
+        return rb_str_new_cstr(ZSTD_versionString());
+    case ALGO_LZ4:
+        return rb_sprintf("%d.%d.%d", LZ4_VERSION_MAJOR, LZ4_VERSION_MINOR, LZ4_VERSION_RELEASE);
+    case ALGO_BROTLI:
+        return rb_sprintf("%d.%d.%d", BrotliEncoderVersion() >> 24,
+                          (BrotliEncoderVersion() >> 12) & 0xFFF, BrotliEncoderVersion() & 0xFFF);
+    }
+    return Qnil;
+}
+
+#define LZ4_RING_BUFFER_SIZE  (64 * 1024)
+#define LZ4_RING_BUFFER_TOTAL (LZ4_RING_BUFFER_SIZE * 2)
+
+typedef struct {
+    compress_algo_t algo;
+    int level;
+    int closed;
+    int finished;
+
+    union {
+        ZSTD_CStream *zstd;
+        BrotliEncoderState *brotli;
+        LZ4_stream_t *lz4;
+    } ctx;
+
+    struct {
+        char *buf;
+        size_t ring_offset;
+        size_t pending;
+    } lz4_ring;
+} deflater_t;
+
+static void deflater_free(void *ptr) {
+    deflater_t *d = (deflater_t *)ptr;
+    if (!d->closed) {
+        switch (d->algo) {
+        case ALGO_ZSTD:
+            if (d->ctx.zstd)
+                ZSTD_freeCStream(d->ctx.zstd);
+            break;
+        case ALGO_BROTLI:
+            if (d->ctx.brotli)
+                BrotliEncoderDestroyInstance(d->ctx.brotli);
+            break;
+        case ALGO_LZ4:
+            if (d->ctx.lz4)
+                LZ4_freeStream(d->ctx.lz4);
+            break;
+        }
+    }
+    if (d->lz4_ring.buf)
+        xfree(d->lz4_ring.buf);
+    xfree(d);
+}
+
+static size_t deflater_memsize(const void *ptr) {
+    const deflater_t *d = (const deflater_t *)ptr;
+    size_t s = sizeof(deflater_t);
+    if (d->lz4_ring.buf)
+        s += LZ4_RING_BUFFER_TOTAL;
+    return s;
+}
+
+static const rb_data_type_t deflater_type = {"Compress::Deflater",
+                                             {NULL, deflater_free, deflater_memsize},
+                                             0,
+                                             0,
+                                             RUBY_TYPED_FREE_IMMEDIATELY};
+
+static VALUE deflater_alloc(VALUE klass) {
+    deflater_t *d = ALLOC(deflater_t);
+    memset(d, 0, sizeof(deflater_t));
+    return TypedData_Wrap_Struct(klass, &deflater_type, d);
+}
+
+static VALUE deflater_initialize(int argc, VALUE *argv, VALUE self) {
+    VALUE opts;
+    rb_scan_args(argc, argv, "0:", &opts);
+    reject_algorithm_keyword(opts);
+
+    deflater_t *d;
+    TypedData_Get_Struct(self, deflater_t, &deflater_type, d);
+
+    VALUE algo_sym = Qnil, level_val = Qnil, dict_val = Qnil;
+    if (!NIL_P(opts)) {
+        algo_sym = opt_get(opts, sym_cache.algo);
+        level_val = opt_get(opts, sym_cache.level);
+        dict_val = opt_get(opts, sym_cache.dictionary);
+    }
+
+    d->algo = NIL_P(algo_sym) ? ALGO_ZSTD : sym_to_algo(algo_sym);
+    d->level = resolve_level(d->algo, level_val);
+    d->closed = 0;
+    d->finished = 0;
+
+    dictionary_t *dict = NULL;
+    if (!NIL_P(dict_val)) {
+        if (d->algo == ALGO_LZ4) {
+            rb_raise(eUnsupportedError, "LZ4 does not support dictionaries");
+        }
+        dict = opt_dictionary(dict_val);
+        dictionary_ivar_set(self, dict_val);
+    }
+
+    switch (d->algo) {
+    case ALGO_ZSTD: {
+        d->ctx.zstd = ZSTD_createCStream();
+        if (!d->ctx.zstd)
+            rb_raise(eMemError, "zstd: failed to create stream");
+
+        if (dict) {
+            ZSTD_CCtx_reset(d->ctx.zstd, ZSTD_reset_session_only);
+            ZSTD_CCtx_setParameter(d->ctx.zstd, ZSTD_c_compressionLevel, d->level);
+            size_t r = ZSTD_CCtx_loadDictionary(d->ctx.zstd, dict->data, dict->size);
+            if (ZSTD_isError(r))
+                rb_raise(eError, "zstd dict load: %s", ZSTD_getErrorName(r));
+        } else {
+            size_t r = ZSTD_initCStream(d->ctx.zstd, d->level);
+            if (ZSTD_isError(r))
+                rb_raise(eError, "zstd init: %s", ZSTD_getErrorName(r));
+        }
+        break;
+    }
+    case ALGO_BROTLI: {
+        d->ctx.brotli = BrotliEncoderCreateInstance(NULL, NULL, NULL);
+        if (!d->ctx.brotli)
+            rb_raise(eMemError, "brotli: failed to create encoder");
+        if (!BrotliEncoderSetParameter(d->ctx.brotli, BROTLI_PARAM_QUALITY, d->level)) {
+            BrotliEncoderDestroyInstance(d->ctx.brotli);
+            d->ctx.brotli = NULL;
+            rb_raise(eError, "brotli: failed to set quality parameter");
+        }
+        if (dict) {
+            BrotliEncoderPreparedDictionary *pd =
+                BrotliEncoderPrepareDictionary(BROTLI_SHARED_DICTIONARY_RAW, dict->size, dict->data,
+                                               BROTLI_MAX_QUALITY, NULL, NULL, NULL);
+            if (!pd) {
+                BrotliEncoderDestroyInstance(d->ctx.brotli);
+                d->ctx.brotli = NULL;
+                rb_raise(eMemError, "brotli: failed to prepare dictionary");
+            }
+            if (!BrotliEncoderAttachPreparedDictionary(d->ctx.brotli, pd)) {
+                BrotliEncoderDestroyPreparedDictionary(pd);
+                BrotliEncoderDestroyInstance(d->ctx.brotli);
+                d->ctx.brotli = NULL;
+                rb_raise(eError, "brotli: failed to attach dictionary");
+            }
+            BrotliEncoderDestroyPreparedDictionary(pd);
+        }
+        break;
+    }
+    case ALGO_LZ4: {
+        d->ctx.lz4 = LZ4_createStream();
+        if (!d->ctx.lz4)
+            rb_raise(eMemError, "lz4: failed to create stream");
+        LZ4_resetStream(d->ctx.lz4);
+        d->lz4_ring.buf = ALLOC_N(char, LZ4_RING_BUFFER_TOTAL);
+        d->lz4_ring.ring_offset = 0;
+        d->lz4_ring.pending = 0;
+        break;
+    }
+    }
+
+    return self;
+}
+
+static VALUE lz4_compress_ring_block(deflater_t *d) {
+    if (d->lz4_ring.pending == 0)
+        return rb_binary_str_new("", 0);
+
+    char *block_start = d->lz4_ring.buf + d->lz4_ring.ring_offset - d->lz4_ring.pending;
+    int src_size = (int)d->lz4_ring.pending;
+    int bound = LZ4_compressBound(src_size);
+
+    VALUE output = rb_binary_str_buf_new(8 + bound);
+    char *out = RSTRING_PTR(output);
+
+    write_le_u32((uint8_t *)out, (uint32_t)src_size);
+
+    int csize = LZ4_compress_fast_continue(d->ctx.lz4, block_start, out + 8, src_size, bound, 1);
+    if (csize <= 0)
+        rb_raise(eError, "lz4 stream compress block failed");
+
+    write_le_u32((uint8_t *)out + 4, (uint32_t)csize);
+
+    rb_str_set_len(output, 8 + csize);
+    d->lz4_ring.pending = 0;
+
+    if (d->lz4_ring.ring_offset >= LZ4_RING_BUFFER_SIZE) {
+        d->lz4_ring.ring_offset = 0;
+    }
+
+    return output;
+}
+
+static VALUE deflater_write(VALUE self, VALUE chunk) {
+    deflater_t *d;
+    TypedData_Get_Struct(self, deflater_t, &deflater_type, d);
+    if (d->closed)
+        rb_raise(eStreamError, "stream is closed");
+    if (d->finished)
+        rb_raise(eStreamError, "stream is already finished");
+    StringValue(chunk);
+
+    const char *src = RSTRING_PTR(chunk);
+    size_t slen = RSTRING_LEN(chunk);
+    const algo_policy_t *policy = algo_policy(d->algo);
+    if (slen == 0)
+        return rb_binary_str_new("", 0);
+
+    switch (d->algo) {
+    case ALGO_ZSTD: {
+        ZSTD_inBuffer input = {src, slen, 0};
+        size_t out_cap = ZSTD_CStreamOutSize();
+        size_t result_cap = out_cap > slen ? out_cap : slen;
+        VALUE result = rb_binary_str_buf_reserve(result_cap);
+        size_t result_len = 0;
+        VALUE scheduler = current_fiber_scheduler();
+
+        while (input.pos < input.size) {
+            if (result_len + out_cap > result_cap) {
+                result_cap = result_cap * 2;
+                grow_binary_str(result, result_len, result_cap);
+            }
+
+            ZSTD_outBuffer output = {RSTRING_PTR(result) + result_len, out_cap, 0};
+
+            if (scheduler != Qnil && (input.size - input.pos) >= policy->fiber_stream_threshold) {
+                zstd_stream_chunk_fiber_t fargs = {
+                    .cstream = d->ctx.zstd,
+                    .input = &input,
+                    .output = &output,
+                    .result = 0,
+                    .scheduler = scheduler,
+                    .blocker = rb_obj_alloc(rb_cObject),
+                    .fiber = rb_fiber_current(),
+                };
+                VALUE th = rb_thread_create(zstd_stream_chunk_fiber_thread, &fargs);
+                rb_fiber_scheduler_block(scheduler, fargs.blocker, Qnil);
+                join_thread(th);
+
+                if (ZSTD_isError(fargs.result))
+                    rb_raise(eError, "zstd compress stream: %s", ZSTD_getErrorName(fargs.result));
+            } else if (scheduler == Qnil &&
+                       (input.size - input.pos) >= policy->gvl_unlock_threshold) {
+                zstd_stream_chunk_args_t args = {
+                    .cstream = d->ctx.zstd,
+                    .output = &output,
+                    .input = &input,
+                    .result = 0,
+                };
+                run_without_gvl(zstd_compress_stream_chunk_nogvl, &args);
+                if (ZSTD_isError(args.result))
+                    rb_raise(eError, "zstd compress stream: %s", ZSTD_getErrorName(args.result));
+            } else {
+                size_t ret = ZSTD_compressStream(d->ctx.zstd, &output, &input);
+                if (ZSTD_isError(ret))
+                    rb_raise(eError, "zstd compress stream: %s", ZSTD_getErrorName(ret));
+            }
+            result_len += output.pos;
+        }
+        rb_str_set_len(result, result_len);
+        RB_GC_GUARD(chunk);
+        return result;
+    }
+    case ALGO_BROTLI: {
+        size_t available_in = slen;
+        const uint8_t *next_in = (const uint8_t *)src;
+        size_t result_cap = slen;
+        if (result_cap < 1024)
+            result_cap = 1024;
+        VALUE result = rb_binary_str_buf_reserve(result_cap);
+        size_t result_len = 0;
+        VALUE scheduler = current_fiber_scheduler();
+        int use_fiber = (scheduler != Qnil);
+        size_t fiber_counter = 0;
+
+        while (available_in > 0 || BrotliEncoderHasMoreOutput(d->ctx.brotli)) {
+            size_t available_out = 0;
+            uint8_t *next_out = NULL;
+            BROTLI_BOOL ok;
+
+            if (use_fiber && available_in >= policy->fiber_stream_threshold) {
+                brotli_stream_chunk_fiber_t fargs = {
+                    .enc = d->ctx.brotli,
+                    .op = BROTLI_OPERATION_PROCESS,
+                    .available_in = &available_in,
+                    .next_in = &next_in,
+                    .available_out = &available_out,
+                    .next_out = &next_out,
+                    .result = BROTLI_FALSE,
+                    .scheduler = scheduler,
+                    .blocker = rb_obj_alloc(rb_cObject),
+                    .fiber = rb_fiber_current(),
+                };
+                VALUE th = rb_thread_create(brotli_stream_chunk_fiber_thread, &fargs);
+                rb_fiber_scheduler_block(scheduler, fargs.blocker, Qnil);
+                join_thread(th);
+                ok = fargs.result;
+            } else {
+                ok = BrotliEncoderCompressStream(d->ctx.brotli, BROTLI_OPERATION_PROCESS,
+                                                 &available_in, &next_in, &available_out, &next_out,
+                                                 NULL);
+            }
+            if (!ok)
+                rb_raise(eError, "brotli compress stream failed");
+
+            const uint8_t *out_data;
+            size_t out_size = 0;
+            out_data = BrotliEncoderTakeOutput(d->ctx.brotli, &out_size);
+            if (out_size > 0) {
+                if (result_len + out_size > result_cap) {
+                    result_cap = (result_len + out_size) * 2;
+                    grow_binary_str(result, result_len, result_cap);
+                }
+
+                memcpy(RSTRING_PTR(result) + result_len, out_data, out_size);
+                result_len += out_size;
+                if (use_fiber) {
+                    int did_yield = 0;
+                    fiber_counter = fiber_maybe_yield(fiber_counter, out_size,
+                                                      policy->fiber_yield_chunk, &did_yield);
+                    (void)did_yield;
+                }
+            }
+        }
+        rb_str_set_len(result, result_len);
+        RB_GC_GUARD(chunk);
+        return result;
+    }
+    case ALGO_LZ4: {
+        VALUE result = rb_binary_str_buf_reserve(0);
+        size_t result_len = 0;
+        size_t result_cap = 0;
+        int use_fiber = has_fiber_scheduler();
+        size_t fiber_counter = 0;
+
+        while (slen > 0) {
+            size_t space = LZ4_RING_BUFFER_SIZE - d->lz4_ring.pending;
+            size_t copy = slen < space ? slen : space;
+
+            if (d->lz4_ring.ring_offset + copy > LZ4_RING_BUFFER_TOTAL) {
+                rb_raise(eError, "lz4: ring buffer overflow");
+            }
+
+            memcpy(d->lz4_ring.buf + d->lz4_ring.ring_offset, src, copy);
+            d->lz4_ring.ring_offset += copy;
+            d->lz4_ring.pending += copy;
+            src += copy;
+            slen -= copy;
+
+            if (d->lz4_ring.pending >= (size_t)LZ4_RING_BUFFER_SIZE) {
+                VALUE block = lz4_compress_ring_block(d);
+                size_t blen = RSTRING_LEN(block);
+                if (blen > 0) {
+                    if (result_len + blen > result_cap) {
+                        result_cap = (result_len + blen) * 2;
+                        if (result_cap < 256)
+                            result_cap = 256;
+                        grow_binary_str(result, result_len, result_cap);
+                    }
+                    memcpy(RSTRING_PTR(result) + result_len, RSTRING_PTR(block), blen);
+                    result_len += blen;
+                }
+                if (use_fiber) {
+                    int did_yield = 0;
+                    fiber_counter = fiber_maybe_yield(fiber_counter, LZ4_RING_BUFFER_SIZE,
+                                                      policy->fiber_yield_chunk, &did_yield);
+                    (void)did_yield;
+                }
+            }
+        }
+        rb_str_set_len(result, result_len);
+        RB_GC_GUARD(chunk);
+        return result;
+    }
+    }
+    return rb_binary_str_new("", 0);
+}
+
+static VALUE deflater_flush(VALUE self) {
+    deflater_t *d;
+    TypedData_Get_Struct(self, deflater_t, &deflater_type, d);
+    if (d->closed)
+        rb_raise(eStreamError, "stream is closed");
+    if (d->finished)
+        rb_raise(eStreamError, "stream is already finished");
+
+    switch (d->algo) {
+    case ALGO_ZSTD: {
+        size_t out_cap = ZSTD_CStreamOutSize();
+        size_t result_cap = out_cap;
+        VALUE result = rb_binary_str_buf_reserve(result_cap);
+        size_t result_len = 0;
+        size_t ret;
+
+        do {
+            if (result_len + out_cap > result_cap) {
+                result_cap *= 2;
+                grow_binary_str(result, result_len, result_cap);
+            }
+
+            ZSTD_outBuffer output = {RSTRING_PTR(result) + result_len, out_cap, 0};
+            ret = ZSTD_flushStream(d->ctx.zstd, &output);
+            if (ZSTD_isError(ret))
+                rb_raise(eError, "zstd flush: %s", ZSTD_getErrorName(ret));
+            result_len += output.pos;
+        } while (ret > 0);
+
+        rb_str_set_len(result, result_len);
+        return result;
+    }
+    case ALGO_BROTLI: {
+        size_t available_in = 0;
+        const uint8_t *next_in = NULL;
+        size_t result_cap = 1024;
+        VALUE result = rb_binary_str_buf_reserve(result_cap);
+        size_t result_len = 0;
+
+        do {
+            size_t available_out = 0;
+            uint8_t *next_out = NULL;
+            BROTLI_BOOL ok =
+                BrotliEncoderCompressStream(d->ctx.brotli, BROTLI_OPERATION_FLUSH, &available_in,
+                                            &next_in, &available_out, &next_out, NULL);
+            if (!ok)
+                rb_raise(eError, "brotli flush failed");
+            const uint8_t *out_data;
+            size_t out_size = 0;
+            out_data = BrotliEncoderTakeOutput(d->ctx.brotli, &out_size);
+            if (out_size > 0) {
+                if (result_len + out_size > result_cap) {
+                    result_cap = (result_len + out_size) * 2;
+                    grow_binary_str(result, result_len, result_cap);
+                }
+
+                memcpy(RSTRING_PTR(result) + result_len, out_data, out_size);
+                result_len += out_size;
+            }
+        } while (BrotliEncoderHasMoreOutput(d->ctx.brotli));
+
+        rb_str_set_len(result, result_len);
+        return result;
+    }
+    case ALGO_LZ4:
+        return lz4_compress_ring_block(d);
+    }
+    return rb_binary_str_new("", 0);
+}
+
+static VALUE deflater_finish(VALUE self) {
+    deflater_t *d;
+    TypedData_Get_Struct(self, deflater_t, &deflater_type, d);
+    if (d->closed)
+        rb_raise(eStreamError, "stream is closed");
+    if (d->finished)
+        return rb_binary_str_new("", 0);
+    d->finished = 1;
+
+    switch (d->algo) {
+    case ALGO_ZSTD: {
+        size_t out_cap = ZSTD_CStreamOutSize();
+        size_t result_cap = out_cap;
+        VALUE result = rb_binary_str_buf_reserve(result_cap);
+        size_t result_len = 0;
+        size_t ret;
+
+        do {
+            if (result_len + out_cap > result_cap) {
+                result_cap *= 2;
+                grow_binary_str(result, result_len, result_cap);
+            }
+
+            ZSTD_outBuffer output = {RSTRING_PTR(result) + result_len, out_cap, 0};
+            ret = ZSTD_endStream(d->ctx.zstd, &output);
+            if (ZSTD_isError(ret))
+                rb_raise(eError, "zstd end stream: %s", ZSTD_getErrorName(ret));
+            result_len += output.pos;
+        } while (ret > 0);
+
+        rb_str_set_len(result, result_len);
+        return result;
+    }
+    case ALGO_BROTLI: {
+        size_t available_in = 0;
+        const uint8_t *next_in = NULL;
+        size_t result_cap = 1024;
+        VALUE result = rb_binary_str_buf_reserve(result_cap);
+        size_t result_len = 0;
+
+        do {
+            size_t available_out = 0;
+            uint8_t *next_out = NULL;
+            BROTLI_BOOL ok =
+                BrotliEncoderCompressStream(d->ctx.brotli, BROTLI_OPERATION_FINISH, &available_in,
+                                            &next_in, &available_out, &next_out, NULL);
+            if (!ok)
+                rb_raise(eError, "brotli finish failed");
+            const uint8_t *out_data;
+            size_t out_size = 0;
+            out_data = BrotliEncoderTakeOutput(d->ctx.brotli, &out_size);
+            if (out_size > 0) {
+                if (result_len + out_size > result_cap) {
+                    result_cap = (result_len + out_size) * 2;
+                    grow_binary_str(result, result_len, result_cap);
+                }
+
+                memcpy(RSTRING_PTR(result) + result_len, out_data, out_size);
+                result_len += out_size;
+            }
+        } while (BrotliEncoderHasMoreOutput(d->ctx.brotli) ||
+                 !BrotliEncoderIsFinished(d->ctx.brotli));
+
+        rb_str_set_len(result, result_len);
+        return result;
+    }
+    case ALGO_LZ4: {
+        size_t result_cap = 256;
+        VALUE result = rb_binary_str_buf_reserve(result_cap);
+        size_t result_len = 0;
+
+        if (d->lz4_ring.pending > 0) {
+            VALUE block = lz4_compress_ring_block(d);
+            size_t blen = RSTRING_LEN(block);
+            if (blen > 0) {
+                if (blen + 4 > result_cap) {
+                    result_cap = blen + 4;
+                    grow_binary_str(result, result_len, result_cap);
+                }
+
+                memcpy(RSTRING_PTR(result), RSTRING_PTR(block), blen);
+                result_len = blen;
+            }
+        }
+
+        if (result_len + 4 > result_cap) {
+            result_cap = result_len + 4;
+            grow_binary_str(result, result_len, result_cap);
+        }
+
+        char *out = RSTRING_PTR(result) + result_len;
+        write_le_u32((uint8_t *)out, 0);
+        result_len += 4;
+
+        rb_str_set_len(result, result_len);
+        return result;
+    }
+    }
+    return rb_binary_str_new("", 0);
+}
+
+static VALUE deflater_reset(VALUE self) {
+    deflater_t *d;
+    TypedData_Get_Struct(self, deflater_t, &deflater_type, d);
+
+    VALUE dict_val = dictionary_ivar_get(self);
+    dictionary_t *dict = NULL;
+    if (!NIL_P(dict_val)) {
+        TypedData_Get_Struct(dict_val, dictionary_t, &dictionary_type, dict);
+    }
+
+    switch (d->algo) {
+    case ALGO_ZSTD:
+        if (d->ctx.zstd) {
+            ZSTD_CCtx_reset(d->ctx.zstd, ZSTD_reset_session_only);
+            ZSTD_CCtx_setParameter(d->ctx.zstd, ZSTD_c_compressionLevel, d->level);
+            if (dict) {
+                size_t r = ZSTD_CCtx_loadDictionary(d->ctx.zstd, dict->data, dict->size);
+                if (ZSTD_isError(r))
+                    rb_raise(eError, "zstd dict reload on reset: %s", ZSTD_getErrorName(r));
+            }
+        }
+        break;
+    case ALGO_BROTLI:
+        if (d->ctx.brotli) {
+            BrotliEncoderDestroyInstance(d->ctx.brotli);
+            d->ctx.brotli = BrotliEncoderCreateInstance(NULL, NULL, NULL);
+            if (!d->ctx.brotli)
+                rb_raise(eMemError, "brotli: failed to recreate encoder");
+            if (!BrotliEncoderSetParameter(d->ctx.brotli, BROTLI_PARAM_QUALITY, d->level))
+                rb_raise(eError, "brotli: failed to set quality on reset");
+            if (dict) {
+                BrotliEncoderPreparedDictionary *pd = BrotliEncoderPrepareDictionary(
+                    BROTLI_SHARED_DICTIONARY_RAW, dict->size, dict->data, BROTLI_MAX_QUALITY, NULL,
+                    NULL, NULL);
+                if (!pd)
+                    rb_raise(eMemError, "brotli: failed to prepare dictionary on reset");
+                if (!BrotliEncoderAttachPreparedDictionary(d->ctx.brotli, pd)) {
+                    BrotliEncoderDestroyPreparedDictionary(pd);
+                    rb_raise(eError, "brotli: failed to reattach dictionary on reset");
+                }
+                BrotliEncoderDestroyPreparedDictionary(pd);
+            }
+        }
+        break;
+    case ALGO_LZ4:
+        if (d->ctx.lz4)
+            LZ4_resetStream(d->ctx.lz4);
+        d->lz4_ring.ring_offset = 0;
+        d->lz4_ring.pending = 0;
+        break;
+    }
+    d->closed = 0;
+    d->finished = 0;
+    return self;
+}
+
+static VALUE deflater_close(VALUE self) {
+    deflater_t *d;
+    TypedData_Get_Struct(self, deflater_t, &deflater_type, d);
+    if (d->closed)
+        return Qnil;
+
+    switch (d->algo) {
+    case ALGO_ZSTD:
+        if (d->ctx.zstd) {
+            ZSTD_freeCStream(d->ctx.zstd);
+            d->ctx.zstd = NULL;
+        }
+        break;
+    case ALGO_BROTLI:
+        if (d->ctx.brotli) {
+            BrotliEncoderDestroyInstance(d->ctx.brotli);
+            d->ctx.brotli = NULL;
+        }
+        break;
+    case ALGO_LZ4:
+        if (d->ctx.lz4) {
+            LZ4_freeStream(d->ctx.lz4);
+            d->ctx.lz4 = NULL;
+        }
+        break;
+    }
+    d->closed = 1;
+    return Qnil;
+}
+
+static VALUE deflater_closed_p(VALUE self) {
+    deflater_t *d;
+    TypedData_Get_Struct(self, deflater_t, &deflater_type, d);
+    return d->closed ? Qtrue : Qfalse;
+}
+
+typedef struct {
+    compress_algo_t algo;
+    int closed;
+    int finished;
+    size_t max_output_size;
+    size_t total_output;
+    size_t total_input;
+    int max_ratio_enabled;
+    unsigned long long max_ratio;
+
+    union {
+        ZSTD_DStream *zstd;
+        BrotliDecoderState *brotli;
+    } ctx;
+
+    struct {
+        char *buf;
+        size_t len;
+        size_t cap;
+        size_t offset;
+    } lz4_buf;
+} inflater_t;
+
+static void inflater_free(void *ptr) {
+    inflater_t *inf = (inflater_t *)ptr;
+    if (!inf->closed) {
+        switch (inf->algo) {
+        case ALGO_ZSTD:
+            if (inf->ctx.zstd)
+                ZSTD_freeDStream(inf->ctx.zstd);
+            break;
+        case ALGO_BROTLI:
+            if (inf->ctx.brotli)
+                BrotliDecoderDestroyInstance(inf->ctx.brotli);
+            break;
+        case ALGO_LZ4:
+            break;
+        }
+    }
+    if (inf->lz4_buf.buf)
+        xfree(inf->lz4_buf.buf);
+    xfree(inf);
+}
+
+static size_t inflater_memsize(const void *ptr) {
+    const inflater_t *inf = (const inflater_t *)ptr;
+    return sizeof(inflater_t) + inf->lz4_buf.cap;
+}
+
+static const rb_data_type_t inflater_type = {"Compress::Inflater",
+                                             {NULL, inflater_free, inflater_memsize},
+                                             0,
+                                             0,
+                                             RUBY_TYPED_FREE_IMMEDIATELY};
+
+static VALUE inflater_alloc(VALUE klass) {
+    inflater_t *inf = ALLOC(inflater_t);
+    memset(inf, 0, sizeof(inflater_t));
+    return TypedData_Wrap_Struct(klass, &inflater_type, inf);
+}
+
+static VALUE inflater_initialize(int argc, VALUE *argv, VALUE self) {
+    VALUE opts;
+    rb_scan_args(argc, argv, "0:", &opts);
+    reject_algorithm_keyword(opts);
+
+    inflater_t *inf;
+    TypedData_Get_Struct(self, inflater_t, &inflater_type, inf);
+
+    VALUE algo_sym = Qnil, dict_val = Qnil;
+    limits_config_t limits;
+    parse_limits_from_opts(opts, &limits);
+    if (!NIL_P(opts)) {
+        algo_sym = opt_get(opts, sym_cache.algo);
+        dict_val = opt_get(opts, sym_cache.dictionary);
+    }
+
+    inf->algo = NIL_P(algo_sym) ? ALGO_ZSTD : sym_to_algo(algo_sym);
+    inf->closed = 0;
+    inf->finished = 0;
+    inf->max_output_size = limits.max_output_size;
+    inf->total_output = 0;
+    inf->total_input = 0;
+    inf->max_ratio_enabled = limits.max_ratio_enabled;
+    inf->max_ratio = limits.max_ratio;
+
+    dictionary_t *dict = NULL;
+    if (!NIL_P(dict_val)) {
+        if (inf->algo == ALGO_LZ4) {
+            rb_raise(eUnsupportedError, "LZ4 does not support dictionaries");
+        }
+        dict = opt_dictionary(dict_val);
+        dictionary_ivar_set(self, dict_val);
+    }
+
+    switch (inf->algo) {
+    case ALGO_ZSTD:
+        inf->ctx.zstd = ZSTD_createDStream();
+        if (!inf->ctx.zstd)
+            rb_raise(eMemError, "zstd: failed to create dstream");
+        if (dict) {
+            ZSTD_DCtx_reset(inf->ctx.zstd, ZSTD_reset_session_only);
+            size_t r = ZSTD_DCtx_loadDictionary(inf->ctx.zstd, dict->data, dict->size);
+            if (ZSTD_isError(r))
+                rb_raise(eError, "zstd dict load: %s", ZSTD_getErrorName(r));
+        } else {
+            ZSTD_initDStream(inf->ctx.zstd);
+        }
+        break;
+    case ALGO_BROTLI:
+        inf->ctx.brotli = BrotliDecoderCreateInstance(NULL, NULL, NULL);
+        if (!inf->ctx.brotli)
+            rb_raise(eMemError, "brotli: failed to create decoder");
+        if (dict) {
+            BrotliDecoderAttachDictionary(inf->ctx.brotli, BROTLI_SHARED_DICTIONARY_RAW, dict->size,
+                                          dict->data);
+        }
+        break;
+    case ALGO_LZ4:
+        inf->lz4_buf.cap = 16 * 1024;
+        inf->lz4_buf.buf = ALLOC_N(char, inf->lz4_buf.cap);
+        inf->lz4_buf.len = 0;
+        inf->lz4_buf.offset = 0;
+        break;
+    }
+
+    return self;
+}
+
+static VALUE inflater_write(VALUE self, VALUE chunk) {
+    inflater_t *inf;
+    TypedData_Get_Struct(self, inflater_t, &inflater_type, inf);
+    if (inf->closed)
+        rb_raise(eStreamError, "stream is closed");
+    StringValue(chunk);
+
+    const char *src = RSTRING_PTR(chunk);
+    size_t slen = RSTRING_LEN(chunk);
+    const algo_policy_t *policy = algo_policy(inf->algo);
+    if (slen == 0)
+        return rb_binary_str_new("", 0);
+
+    inf->total_input =
+        checked_add_size(inf->total_input, slen, "compressed input exceeds representable size");
+
+    switch (inf->algo) {
+    case ALGO_ZSTD: {
+        ZSTD_inBuffer input = {src, slen, 0};
+        size_t out_cap = ZSTD_DStreamOutSize();
+        size_t result_cap = out_cap > slen * 2 ? out_cap : slen * 2;
+        size_t remaining_total_budget =
+            inf->max_output_size > inf->total_output ? inf->max_output_size - inf->total_output : 0;
+        if (remaining_total_budget == 0)
+            rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)",
+                     inf->max_output_size);
+        if (result_cap > remaining_total_budget)
+            result_cap = remaining_total_budget;
+        VALUE result = rb_binary_str_buf_reserve(result_cap);
+        size_t result_len = 0;
+        VALUE scheduler = current_fiber_scheduler();
+
+        while (input.pos < input.size) {
+            size_t remaining_budget = inf->max_output_size - inf->total_output - result_len;
+            if (remaining_budget == 0)
+                rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)",
+                         inf->max_output_size);
+
+            if (result_len + out_cap > result_cap) {
+                size_t next_cap = result_cap * 2;
+                if (next_cap > inf->max_output_size - inf->total_output)
+                    next_cap = inf->max_output_size - inf->total_output;
+                result_cap = next_cap;
+                rb_str_resize(result, result_cap);
+            }
+
+            size_t current_out_cap = out_cap;
+            if (current_out_cap > remaining_budget)
+                current_out_cap = remaining_budget;
+
+            ZSTD_outBuffer output = {RSTRING_PTR(result) + result_len, current_out_cap, 0};
+            size_t ret;
+
+            if (scheduler != Qnil && (input.size - input.pos) >= policy->fiber_stream_threshold) {
+                zstd_decompress_stream_chunk_args_t args = {
+                    .dstream = inf->ctx.zstd,
+                    .output = &output,
+                    .input = &input,
+                    .result = 0,
+                };
+                run_via_fiber_worker(scheduler, zstd_decompress_stream_chunk_nogvl, &args);
+                ret = args.result;
+            } else {
+                ret = ZSTD_decompressStream(inf->ctx.zstd, &output, &input);
+            }
+
+            if (ZSTD_isError(ret))
+                rb_raise(eDataError, "zstd decompress stream: %s", ZSTD_getErrorName(ret));
+            result_len = checked_add_size(result_len, output.pos,
+                                          "decompressed output exceeds representable size");
+            size_t total_output = checked_add_size(
+                inf->total_output, result_len, "decompressed output exceeds representable size");
+            enforce_output_and_ratio_limits(total_output, inf->total_input, inf->max_output_size,
+                                            inf->max_ratio_enabled, inf->max_ratio);
+            if (ret == 0)
+                break;
+        }
+        inf->total_output = checked_add_size(inf->total_output, result_len,
+                                             "decompressed output exceeds representable size");
+        rb_str_set_len(result, result_len);
+        RB_GC_GUARD(chunk);
+        return result;
+    }
+    case ALGO_BROTLI: {
+        size_t available_in = slen;
+        const uint8_t *next_in = (const uint8_t *)src;
+        size_t remaining_total_budget =
+            inf->max_output_size > inf->total_output ? inf->max_output_size - inf->total_output : 0;
+        if (remaining_total_budget == 0)
+            rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)",
+                     inf->max_output_size);
+        size_t result_cap = slen * 2;
+        if (result_cap < 1024)
+            result_cap = 1024;
+        if (result_cap > remaining_total_budget)
+            result_cap = remaining_total_budget;
+        VALUE result = rb_binary_str_buf_reserve(result_cap);
+        size_t result_len = 0;
+        VALUE scheduler = current_fiber_scheduler();
+
+        while (available_in > 0 || BrotliDecoderHasMoreOutput(inf->ctx.brotli)) {
+            size_t available_out = 0;
+            uint8_t *next_out = NULL;
+            BrotliDecoderResult res;
+
+            if (scheduler != Qnil && available_in >= policy->fiber_stream_threshold) {
+                brotli_decompress_stream_args_t sargs = {
+                    .dec = inf->ctx.brotli,
+                    .available_in = &available_in,
+                    .next_in = &next_in,
+                    .available_out = &available_out,
+                    .next_out = &next_out,
+                    .result = BROTLI_DECODER_RESULT_ERROR,
+                };
+                run_via_fiber_worker(scheduler, brotli_decompress_stream_nogvl, &sargs);
+                res = sargs.result;
+            } else {
+                res = BrotliDecoderDecompressStream(inf->ctx.brotli, &available_in, &next_in,
+                                                    &available_out, &next_out, NULL);
+            }
+            if (res == BROTLI_DECODER_RESULT_ERROR)
+                rb_raise(eDataError, "brotli decompress stream: %s",
+                         BrotliDecoderErrorString(BrotliDecoderGetErrorCode(inf->ctx.brotli)));
+            const uint8_t *out_data;
+            size_t out_size = 0;
+            out_data = BrotliDecoderTakeOutput(inf->ctx.brotli, &out_size);
+            if (out_size > 0) {
+                size_t total_output = checked_add_size(
+                    inf->total_output,
+                    checked_add_size(result_len, out_size,
+                                     "decompressed output exceeds representable size"),
+                    "decompressed output exceeds representable size");
+                enforce_output_and_ratio_limits(total_output, inf->total_input,
+                                                inf->max_output_size, inf->max_ratio_enabled,
+                                                inf->max_ratio);
+
+                if (result_len + out_size > result_cap) {
+                    result_cap = result_len + out_size;
+                    rb_str_resize(result, result_cap);
+                }
+
+                memcpy(RSTRING_PTR(result) + result_len, out_data, out_size);
+                result_len += out_size;
+            }
+            if (res == BROTLI_DECODER_RESULT_SUCCESS)
+                break;
+            if (res == BROTLI_DECODER_RESULT_NEEDS_MORE_INPUT && available_in == 0)
+                break;
+        }
+        inf->total_output = checked_add_size(inf->total_output, result_len,
+                                             "decompressed output exceeds representable size");
+        rb_str_set_len(result, result_len);
+        RB_GC_GUARD(chunk);
+        return result;
+    }
+    case ALGO_LZ4: {
+        size_t data_len = inf->lz4_buf.len - inf->lz4_buf.offset;
+        size_t needed = data_len + slen;
+        // TODO(v0.4): optional standard LZ4 frame format support via lz4frame.h
+
+        if (inf->lz4_buf.offset > 0 && needed > inf->lz4_buf.cap) {
+            if (data_len > 0)
+                memmove(inf->lz4_buf.buf, inf->lz4_buf.buf + inf->lz4_buf.offset, data_len);
+            inf->lz4_buf.offset = 0;
+            inf->lz4_buf.len = data_len;
+        } else if (inf->lz4_buf.offset > inf->lz4_buf.cap / 2) {
+            if (data_len > 0)
+                memmove(inf->lz4_buf.buf, inf->lz4_buf.buf + inf->lz4_buf.offset, data_len);
+            inf->lz4_buf.offset = 0;
+            inf->lz4_buf.len = data_len;
+        }
+
+        needed = inf->lz4_buf.len + slen;
+        if (needed > inf->lz4_buf.cap) {
+            inf->lz4_buf.cap = needed * 2;
+            REALLOC_N(inf->lz4_buf.buf, char, inf->lz4_buf.cap);
+        }
+        memcpy(inf->lz4_buf.buf + inf->lz4_buf.len, src, slen);
+        inf->lz4_buf.len += slen;
+
+        size_t remaining_total_budget =
+            inf->max_output_size > inf->total_output ? inf->max_output_size - inf->total_output : 0;
+        if (remaining_total_budget == 0)
+            rb_raise(eDataError, "decompressed output exceeds limit (%zu bytes)",
+                     inf->max_output_size);
+        size_t result_cap = slen * 4;
+        if (result_cap < 256)
+            result_cap = 256;
+        if (result_cap > remaining_total_budget)
+            result_cap = remaining_total_budget;
+        VALUE result = rb_binary_str_buf_new(result_cap);
+        size_t result_len = 0;
+        int use_fiber = has_fiber_scheduler();
+        size_t fiber_counter = 0;
+
+        size_t pos = inf->lz4_buf.offset;
+        while (pos + 4 <= inf->lz4_buf.len) {
+            const uint8_t *p = (const uint8_t *)(inf->lz4_buf.buf + pos);
+            uint32_t orig_size = read_le_u32(p);
+            if (orig_size == 0) {
+                inf->finished = 1;
+                pos += 4;
+                break;
+            }
+            if (pos + 8 > inf->lz4_buf.len)
+                break;
+            uint32_t comp_size = read_le_u32(p + 4);
+            if (pos + 8 + comp_size > inf->lz4_buf.len)
+                break;
+            if (orig_size > 64 * 1024 * 1024)
+                rb_raise(eDataError, "lz4 stream: block too large (%u)", orig_size);
+
+            size_t total_output =
+                checked_add_size(inf->total_output,
+                                 checked_add_size(result_len, orig_size,
+                                                  "decompressed output exceeds representable size"),
+                                 "decompressed output exceeds representable size");
+            enforce_output_and_ratio_limits(total_output, inf->total_input, inf->max_output_size,
+                                            inf->max_ratio_enabled, inf->max_ratio);
+
+            if (result_len + orig_size > result_cap) {
+                result_cap = result_len + orig_size;
+                rb_str_resize(result, result_cap);
+            }
+
+            int dsize =
+                LZ4_decompress_safe(inf->lz4_buf.buf + pos + 8, RSTRING_PTR(result) + result_len,
+                                    (int)comp_size, (int)orig_size);
+            if (dsize < 0)
+                rb_raise(eDataError, "lz4 stream decompress block failed");
+
+            result_len += dsize;
+            pos += 8 + comp_size;
+            if (use_fiber) {
+                int did_yield = 0;
+                fiber_counter = fiber_maybe_yield(fiber_counter, (size_t)dsize,
+                                                  policy->fiber_yield_chunk, &did_yield);
+                (void)did_yield;
+            }
+        }
+
+        inf->lz4_buf.offset = pos;
+        inf->total_output = checked_add_size(inf->total_output, result_len,
+                                             "decompressed output exceeds representable size");
+        rb_str_set_len(result, result_len);
+        RB_GC_GUARD(chunk);
+        return result;
+    }
+    }
+    return rb_binary_str_new("", 0);
+}
+
+static VALUE inflater_finish(VALUE self) {
+    inflater_t *inf;
+    TypedData_Get_Struct(self, inflater_t, &inflater_type, inf);
+    if (inf->closed)
+        rb_raise(eStreamError, "stream is closed");
+    inf->finished = 1;
+    return rb_binary_str_new("", 0);
+}
+
+static VALUE inflater_reset(VALUE self) {
+    inflater_t *inf;
+    TypedData_Get_Struct(self, inflater_t, &inflater_type, inf);
+
+    VALUE dict_val = dictionary_ivar_get(self);
+    dictionary_t *dict = NULL;
+    if (!NIL_P(dict_val)) {
+        TypedData_Get_Struct(dict_val, dictionary_t, &dictionary_type, dict);
+    }
+
+    switch (inf->algo) {
+    case ALGO_ZSTD:
+        if (inf->ctx.zstd) {
+            ZSTD_DCtx_reset(inf->ctx.zstd, ZSTD_reset_session_only);
+            if (dict) {
+                size_t r = ZSTD_DCtx_loadDictionary(inf->ctx.zstd, dict->data, dict->size);
+                if (ZSTD_isError(r))
+                    rb_raise(eError, "zstd dict reload on reset: %s", ZSTD_getErrorName(r));
+            }
+        }
+        break;
+    case ALGO_BROTLI:
+        if (inf->ctx.brotli) {
+            BrotliDecoderDestroyInstance(inf->ctx.brotli);
+            inf->ctx.brotli = BrotliDecoderCreateInstance(NULL, NULL, NULL);
+            if (!inf->ctx.brotli)
+                rb_raise(eMemError, "brotli: failed to recreate decoder");
+            if (dict) {
+                BrotliDecoderAttachDictionary(inf->ctx.brotli, BROTLI_SHARED_DICTIONARY_RAW,
+                                              dict->size, dict->data);
+            }
+        }
+        break;
+    case ALGO_LZ4:
+        inf->lz4_buf.len = 0;
+        inf->lz4_buf.offset = 0;
+        break;
+    }
+    inf->closed = 0;
+    inf->finished = 0;
+    inf->total_output = 0;
+    inf->total_input = 0;
+    return self;
+}
+
+static VALUE inflater_close(VALUE self) {
+    inflater_t *inf;
+    TypedData_Get_Struct(self, inflater_t, &inflater_type, inf);
+    if (inf->closed)
+        return Qnil;
+
+    switch (inf->algo) {
+    case ALGO_ZSTD:
+        if (inf->ctx.zstd) {
+            ZSTD_freeDStream(inf->ctx.zstd);
+            inf->ctx.zstd = NULL;
+        }
+        break;
+    case ALGO_BROTLI:
+        if (inf->ctx.brotli) {
+            BrotliDecoderDestroyInstance(inf->ctx.brotli);
+            inf->ctx.brotli = NULL;
+        }
+        break;
+    case ALGO_LZ4:
+        break;
+    }
+    inf->closed = 1;
+    return Qnil;
+}
+
+static VALUE inflater_closed_p(VALUE self) {
+    inflater_t *inf;
+    TypedData_Get_Struct(self, inflater_t, &inflater_type, inf);
+    return inf->closed ? Qtrue : Qfalse;
+}
+
+static VALUE dict_initialize(int argc, VALUE *argv, VALUE self) {
+    VALUE raw, opts;
+    rb_scan_args(argc, argv, "1:", &raw, &opts);
+    StringValue(raw);
+    reject_algorithm_keyword(opts);
+
+    dictionary_t *d;
+    TypedData_Get_Struct(self, dictionary_t, &dictionary_type, d);
+
+    VALUE algo_sym = Qnil;
+    if (!NIL_P(opts)) {
+        algo_sym = opt_get(opts, sym_cache.algo);
+    }
+    d->algo = NIL_P(algo_sym) ? ALGO_ZSTD : sym_to_algo(algo_sym);
+
+    if (d->algo == ALGO_LZ4)
+        rb_raise(eUnsupportedError, "LZ4 does not support dictionaries");
+
+    d->size = RSTRING_LEN(raw);
+    d->data = ALLOC_N(uint8_t, d->size);
+    memcpy(d->data, RSTRING_PTR(raw), d->size);
+
+    return self;
+}
+
+static VALUE train_dictionary_internal(VALUE samples, VALUE size_val, compress_algo_t algo) {
+    Check_Type(samples, T_ARRAY);
+
+    if (algo == ALGO_BROTLI) {
+        rb_raise(eUnsupportedError, "Brotli dictionary training is not supported via this API. "
+                                    "Create a raw dictionary using "
+                                    "MultiCompress::Dictionary.new(data, algo: :brotli)");
+    }
+
+    size_t dict_capacity = NIL_P(size_val) ? 112640 /* 110 KiB, zstd default */
+                                           : NUM2SIZET(size_val);
+    if (dict_capacity < 256)
+        rb_raise(rb_eArgError, "dictionary size must be at least 256 bytes");
+
+    long num_samples = RARRAY_LEN(samples);
+    if (num_samples < 1)
+        rb_raise(rb_eArgError, "need at least 1 sample for training");
+
+    size_t total_size = 0;
+    for (long i = 0; i < num_samples; i++) {
+        VALUE s = rb_ary_entry(samples, i);
+        StringValue(s);
+        size_t slen = RSTRING_LEN(s);
+        if (slen < 8)
+            rb_raise(rb_eArgError, "sample %ld is too small (%zu bytes), minimum is 8 bytes", i,
+                     slen);
+        total_size += slen;
+    }
+
+    uint8_t *dict_buf = ALLOC_N(uint8_t, dict_capacity);
+    char *concat = ALLOC_N(char, total_size);
+    size_t *sizes = ALLOC_N(size_t, (size_t)num_samples);
+
+    size_t offset = 0;
+    for (long i = 0; i < num_samples; i++) {
+        VALUE s = rb_ary_entry(samples, i);
+        StringValue(s);
+        size_t slen = RSTRING_LEN(s);
+        memcpy(concat + offset, RSTRING_PTR(s), slen);
+        sizes[i] = slen;
+        offset += slen;
+        RB_GC_GUARD(s);
+    }
+
+    size_t dict_size =
+        ZDICT_trainFromBuffer(dict_buf, dict_capacity, concat, sizes, (unsigned)num_samples);
+
+    xfree(concat);
+    xfree(sizes);
+
+    if (ZDICT_isError(dict_size)) {
+        const char *err = ZDICT_getErrorName(dict_size);
+        xfree(dict_buf);
+        rb_raise(eError,
+                 "dictionary training failed: %s "
+                 "(tip: provide more samples; total sample bytes should be "
+                 "~100x the dictionary size)",
+                 err);
+    }
+
+    VALUE dict_obj = dict_alloc(cDictionary);
+    dictionary_t *d;
+    TypedData_Get_Struct(dict_obj, dictionary_t, &dictionary_type, d);
+    memset(d, 0, sizeof(*d));
+    d->algo = algo;
+    d->data = dict_buf;
+    d->size = dict_size;
+    return dict_obj;
+}
+
+static VALUE zstd_train_dictionary(int argc, VALUE *argv, VALUE self) {
+    // #if defined(__APPLE__) && (defined(__arm64__) || defined(__aarch64__))
+    //     rb_raise(eUnsupportedError,
+    //              "Zstd dictionary training is temporarily disabled on arm64-darwin "
+    //              "because the current vendored trainer path crashes on this platform");
+    // #endif
+
+    VALUE samples, opts;
+    rb_scan_args(argc, argv, "1:", &samples, &opts);
+    reject_algorithm_keyword(opts);
+    VALUE size_val = opt_get(opts, sym_cache.size);
+    return train_dictionary_internal(samples, size_val, ALGO_ZSTD);
+}
+
+static VALUE brotli_train_dictionary(int argc, VALUE *argv, VALUE self) {
+    VALUE samples, opts;
+    rb_scan_args(argc, argv, "1:", &samples, &opts);
+    reject_algorithm_keyword(opts);
+    VALUE size_val = opt_get(opts, sym_cache.size);
+
+    return train_dictionary_internal(samples, size_val, ALGO_BROTLI);
+}
+
+static VALUE dict_load(int argc, VALUE *argv, VALUE self) {
+    VALUE path, opts;
+    rb_scan_args(argc, argv, "1:", &path, &opts);
+    StringValue(path);
+    reject_algorithm_keyword(opts);
+    raise_if_path_has_null_byte(path);
+
+    VALUE algo_sym = Qnil;
+    if (!NIL_P(opts)) {
+        algo_sym = opt_get(opts, sym_cache.algo);
+    }
+    compress_algo_t algo = NIL_P(algo_sym) ? ALGO_ZSTD : sym_to_algo(algo_sym);
+
+    if (algo == ALGO_LZ4)
+        rb_raise(eUnsupportedError, "LZ4 does not support dictionaries");
+
+    const char *cpath = RSTRING_PTR(path);
+    FILE *f = fopen(cpath, "rb");
+    if (!f)
+        rb_sys_fail(cpath);
+
+    fseek(f, 0, SEEK_END);
+    long file_size = ftell(f);
+    fseek(f, 0, SEEK_SET);
+
+    if (file_size <= 0) {
+        fclose(f);
+        rb_raise(eDataError, "dictionary file is empty: %s", cpath);
+    }
+    if ((unsigned long long)file_size > DICT_FILE_MAX_SIZE) {
+        fclose(f);
+        rb_raise(eDataError, "dictionary file too large (%ld bytes, max=%d)", file_size,
+                 (int)DICT_FILE_MAX_SIZE);
+    }
+
+    uint8_t *buf = ALLOC_N(uint8_t, file_size);
+    size_t read_bytes = fread(buf, 1, file_size, f);
+    fclose(f);
+
+    if ((long)read_bytes != file_size) {
+        xfree(buf);
+        rb_raise(eDataError, "failed to read dictionary: %s", cpath);
+    }
+
+    VALUE dict_obj = dict_alloc(cDictionary);
+    dictionary_t *d;
+    TypedData_Get_Struct(dict_obj, dictionary_t, &dictionary_type, d);
+    d->algo = algo;
+    d->data = buf;
+    d->size = (size_t)file_size;
+    return dict_obj;
+}
+
+static VALUE dict_save(VALUE self, VALUE path) {
+    dictionary_t *d;
+    TypedData_Get_Struct(self, dictionary_t, &dictionary_type, d);
+
+    const char *cpath = StringValueCStr(path);
+    FILE *f = fopen(cpath, "wb");
+    if (!f)
+        rb_sys_fail(cpath);
+
+    size_t written = fwrite(d->data, 1, d->size, f);
+    fclose(f);
+
+    if (written != d->size)
+        rb_raise(eError, "failed to write dictionary to %s", cpath);
+    return path;
+}
+
+static VALUE dict_algo(VALUE self) {
+    dictionary_t *d;
+    TypedData_Get_Struct(self, dictionary_t, &dictionary_type, d);
+    return algo_to_sym(d->algo);
+}
+
+static VALUE dict_size(VALUE self) {
+    dictionary_t *d;
+    TypedData_Get_Struct(self, dictionary_t, &dictionary_type, d);
+    return SIZET2NUM(d->size);
+}
+
+void Init_multi_compress(void) {
+    binary_encoding = rb_ascii8bit_encoding();
+    init_id_cache();
+    crc32_init_tables();
+
+    mMultiCompress = rb_define_module("MultiCompress");
+
+    eError = rb_define_class_under(mMultiCompress, "Error", rb_eStandardError);
+    eDataError = rb_define_class_under(mMultiCompress, "DataError", eError);
+    eMemError = rb_define_class_under(mMultiCompress, "MemError", eError);
+    eStreamError = rb_define_class_under(mMultiCompress, "StreamError", eError);
+    eUnsupportedError = rb_define_class_under(mMultiCompress, "UnsupportedError", eError);
+    eLevelError = rb_define_class_under(mMultiCompress, "LevelError", eError);
+
+    mZstd = rb_define_module_under(mMultiCompress, "Zstd");
+    mLZ4 = rb_define_module_under(mMultiCompress, "LZ4");
+    mBrotli = rb_define_module_under(mMultiCompress, "Brotli");
+
+    rb_define_const(mZstd, "MIN_LEVEL", INT2FIX(1));
+    rb_define_const(mZstd, "MAX_LEVEL", INT2FIX(22));
+    rb_define_const(mZstd, "DEFAULT_LEVEL", INT2FIX(3));
+    rb_define_const(mLZ4, "MIN_LEVEL", INT2FIX(1));
+    rb_define_const(mLZ4, "MAX_LEVEL", INT2FIX(16));
+    rb_define_const(mLZ4, "DEFAULT_LEVEL", INT2FIX(1));
+    rb_define_const(mBrotli, "MIN_LEVEL", INT2FIX(0));
+    rb_define_const(mBrotli, "MAX_LEVEL", INT2FIX(11));
+    rb_define_const(mBrotli, "DEFAULT_LEVEL", INT2FIX(6));
+
+    rb_define_module_function(mMultiCompress, "compress", compress_compress, -1);
+    rb_define_module_function(mMultiCompress, "decompress", compress_decompress, -1);
+    rb_define_module_function(mMultiCompress, "crc32", compress_crc32, -1);
+    rb_define_module_function(mMultiCompress, "adler32", compress_adler32, -1);
+    rb_define_module_function(mMultiCompress, "algorithms", compress_algorithms, 0);
+    rb_define_module_function(mMultiCompress, "available?", compress_available_p, 1);
+    rb_define_module_function(mMultiCompress, "version", compress_version, 1);
+
+    cDeflater = rb_define_class_under(mMultiCompress, "Deflater", rb_cObject);
+    rb_define_alloc_func(cDeflater, deflater_alloc);
+    rb_define_method(cDeflater, "initialize", deflater_initialize, -1);
+    rb_define_method(cDeflater, "write", deflater_write, 1);
+    rb_define_method(cDeflater, "flush", deflater_flush, 0);
+    rb_define_method(cDeflater, "finish", deflater_finish, 0);
+    rb_define_method(cDeflater, "reset", deflater_reset, 0);
+    rb_define_method(cDeflater, "close", deflater_close, 0);
+    rb_define_method(cDeflater, "closed?", deflater_closed_p, 0);
+
+    cInflater = rb_define_class_under(mMultiCompress, "Inflater", rb_cObject);
+    rb_define_alloc_func(cInflater, inflater_alloc);
+    rb_define_method(cInflater, "initialize", inflater_initialize, -1);
+    rb_define_method(cInflater, "write", inflater_write, 1);
+    rb_define_method(cInflater, "finish", inflater_finish, 0);
+    rb_define_method(cInflater, "reset", inflater_reset, 0);
+    rb_define_method(cInflater, "close", inflater_close, 0);
+    rb_define_method(cInflater, "closed?", inflater_closed_p, 0);
+
+    cWriter = rb_define_class_under(mMultiCompress, "Writer", rb_cObject);
+    cReader = rb_define_class_under(mMultiCompress, "Reader", rb_cObject);
+
+    cDictionary = rb_define_class_under(mMultiCompress, "Dictionary", rb_cObject);
+    rb_define_alloc_func(cDictionary, dict_alloc);
+    rb_define_method(cDictionary, "initialize", dict_initialize, -1);
+    rb_define_singleton_method(cDictionary, "load", dict_load, -1);
+    rb_define_method(cDictionary, "save", dict_save, 1);
+    rb_define_method(cDictionary, "algo", dict_algo, 0);
+    rb_define_method(cDictionary, "size", dict_size, 0);
+    rb_define_singleton_method(mZstd, "train_dictionary", zstd_train_dictionary, -1);
+    rb_define_singleton_method(mBrotli, "train_dictionary", brotli_train_dictionary, -1);
+}