json 2.10.2 → 2.13.2

data/ext/json/ext/generator/generator.c

@@ -1,9 +1,12 @@
 #include "ruby.h"
 #include "../fbuffer/fbuffer.h"
+#include "../vendor/fpconv.c"
 
 #include <math.h>
 #include <ctype.h>
 
+#include "../simd/simd.h"
+
 /* ruby api and some helpers */
 
 typedef struct JSON_Generator_StateStruct {
@@ -44,7 +47,7 @@ static VALUE sym_indent, sym_space, sym_space_before, sym_object_nl, sym_array_n
 
 struct generate_json_data;
 
-typedef void (*generator_func)(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
+typedef void (*generator_func)(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
 
 struct generate_json_data {
     FBuffer *buffer;
@@ -56,20 +59,20 @@ struct generate_json_data {
 
 static VALUE cState_from_state_s(VALUE self, VALUE opts);
 static VALUE cState_partial_generate(VALUE self, VALUE obj, generator_func, VALUE io);
-static void generate_json(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_object(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_array(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_string(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_null(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_false(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_true(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
+static void generate_json(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_object(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_array(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_string(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_null(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_false(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_true(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
 #ifdef RUBY_INTEGER_UNIFICATION
-static void generate_json_integer(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
+static void generate_json_integer(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
 #endif
-static void generate_json_fixnum(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_bignum(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_float(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_fragment(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
+static void generate_json_fixnum(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_bignum(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_float(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_fragment(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
 
 static int usascii_encindex, utf8_encindex, binary_encindex;
 
@@ -108,12 +111,40 @@ typedef struct _search_state {
     const char *end;
     const char *cursor;
     FBuffer *buffer;
+
+#ifdef HAVE_SIMD
+    const char *chunk_base;
+    const char *chunk_end;
+    bool has_matches;
+
+#if defined(HAVE_SIMD_NEON)
+    uint64_t matches_mask;
+#elif defined(HAVE_SIMD_SSE2)
+    int matches_mask;
+#else
+#error "Unknown SIMD Implementation."
+#endif /* HAVE_SIMD_NEON */
+#endif /* HAVE_SIMD */
 } search_state;
 
-static inline void search_flush(search_state *search)
-{
-    fbuffer_append(search->buffer, search->cursor, search->ptr - search->cursor);
-    search->cursor = search->ptr;
+#if (defined(__GNUC__ ) || defined(__clang__))
+#define FORCE_INLINE __attribute__((always_inline))
+#else
+#define FORCE_INLINE
+#endif
+
+static inline FORCE_INLINE void search_flush(search_state *search)
+{
+    // Do not remove this conditional without profiling, specifically escape-heavy text.
+    // escape_UTF8_char_basic will advance search->ptr and search->cursor (effectively a search_flush).
+    // For back-to-back characters that need to be escaped, specifcally for the SIMD code paths, this method
+    // will be called just before calling escape_UTF8_char_basic. There will be no characers to append for the
+    // consecutive characters that need to be escaped. While the fbuffer_append is a no-op if
+    // nothing needs to be flushed, we can save a few memory references with this conditional.
+    if (search->ptr > search->cursor) {
+        fbuffer_append(search->buffer, search->cursor, search->ptr - search->cursor);
+        search->cursor = search->ptr;
+    }
 }
 
 static const unsigned char escape_table_basic[256] = {
@@ -129,6 +160,8 @@ static const unsigned char escape_table_basic[256] = {
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
 };
 
+static unsigned char (*search_escape_basic_impl)(search_state *);
+
 static inline unsigned char search_escape_basic(search_state *search)
 {
     while (search->ptr < search->end) {
@@ -143,7 +176,8 @@ static inline unsigned char search_escape_basic(search_state *search)
     return 0;
 }
 
-static inline void escape_UTF8_char_basic(search_state *search) {
+static inline FORCE_INLINE void escape_UTF8_char_basic(search_state *search)
+{
     const unsigned char ch = (unsigned char)*search->ptr;
     switch (ch) {
         case '"':  fbuffer_append(search->buffer, "\\\"", 2); break;
@@ -185,12 +219,13 @@ static inline void escape_UTF8_char_basic(search_state *search) {
  */
 static inline void convert_UTF8_to_JSON(search_state *search)
 {
-    while (search_escape_basic(search)) {
+    while (search_escape_basic_impl(search)) {
         escape_UTF8_char_basic(search);
     }
 }
 
-static inline void escape_UTF8_char(search_state *search, unsigned char ch_len) {
+static inline void escape_UTF8_char(search_state *search, unsigned char ch_len)
+{
     const unsigned char ch = (unsigned char)*search->ptr;
     switch (ch_len) {
         case 1: {
@@ -226,6 +261,228 @@ static inline void escape_UTF8_char(search_state *search, unsigned char ch_len)
     search->cursor = (search->ptr += ch_len);
 }
 
+#ifdef HAVE_SIMD
+
+static inline FORCE_INLINE char *copy_remaining_bytes(search_state *search, unsigned long vec_len, unsigned long len)
+{
+    // Flush the buffer so everything up until the last 'len' characters are unflushed.
+    search_flush(search);
+
+    FBuffer *buf = search->buffer;
+    fbuffer_inc_capa(buf, vec_len);
+
+    char *s = (buf->ptr + buf->len);
+
+    // Pad the buffer with dummy characters that won't need escaping.
+    // This seem wateful at first sight, but memset of vector length is very fast.
+    memset(s, 'X', vec_len);
+
+    // Optimistically copy the remaining 'len' characters to the output FBuffer. If there are no characters
+    // to escape, then everything ends up in the correct spot. Otherwise it was convenient temporary storage.
+    MEMCPY(s, search->ptr, char, len);
+
+    return s;
+}
+
+#ifdef HAVE_SIMD_NEON
+
+static inline FORCE_INLINE unsigned char neon_next_match(search_state *search)
+{
+    uint64_t mask = search->matches_mask;
+    uint32_t index = trailing_zeros64(mask) >> 2;
+
+    // It is assumed escape_UTF8_char_basic will only ever increase search->ptr by at most one character.
+    // If we want to use a similar approach for full escaping we'll need to ensure:
+    //     search->chunk_base + index >= search->ptr
+    // However, since we know escape_UTF8_char_basic only increases search->ptr by one, if the next match
+    // is one byte after the previous match then:
+    //     search->chunk_base + index == search->ptr
+    search->ptr = search->chunk_base + index;
+    mask &= mask - 1;
+    search->matches_mask = mask;
+    search_flush(search);
+    return 1;
+}
+
+static inline unsigned char search_escape_basic_neon(search_state *search)
+{
+    if (RB_UNLIKELY(search->has_matches)) {
+        // There are more matches if search->matches_mask > 0.
+        if (search->matches_mask > 0) {
+            return neon_next_match(search);
+        } else {
+            // neon_next_match will only advance search->ptr up to the last matching character.
+            // Skip over any characters in the last chunk that occur after the last match.
+            search->has_matches = false;
+            search->ptr = search->chunk_end;
+        }
+    }
+
+    /*
+    * The code below implements an SIMD-based algorithm to determine if N bytes at a time
+    * need to be escaped.
+    *
+    * Assume the ptr = "Te\sting!" (the double quotes are included in the string)
+    *
+    * The explanation will be limited to the first 8 bytes of the string for simplicity. However
+    * the vector insructions may work on larger vectors.
+    *
+    * First, we load three constants 'lower_bound', 'backslash' and 'dblquote" in vector registers.
+    *
+    * lower_bound: [20 20 20 20 20 20 20 20]
+    * backslash:   [5C 5C 5C 5C 5C 5C 5C 5C]
+    * dblquote:    [22 22 22 22 22 22 22 22]
+    *
+    * Next we load the first chunk of the ptr:
+    * [22 54 65 5C 73 74 69 6E] ("  T  e  \  s  t  i  n)
+    *
+    * First we check if any byte in chunk is less than 32 (0x20). This returns the following vector
+    * as no bytes are less than 32 (0x20):
+    * [0 0 0 0 0 0 0 0]
+    *
+    * Next, we check if any byte in chunk is equal to a backslash:
+    * [0 0 0 FF 0 0 0 0]
+    *
+    * Finally we check if any byte in chunk is equal to a double quote:
+    * [FF 0 0 0 0 0 0 0]
+    *
+    * Now we have three vectors where each byte indicates if the corresponding byte in chunk
+    * needs to be escaped. We combine these vectors with a series of logical OR instructions.
+    * This is the needs_escape vector and it is equal to:
+    * [FF 0 0 FF 0 0 0 0]
+    *
+    * Next we compute the bitwise AND between each byte and 0x1 and compute the horizontal sum of
+    * the values in the vector. This computes how many bytes need to be escaped within this chunk.
+    *
+    * Finally we compute a mask that indicates which bytes need to be escaped. If the mask is 0 then,
+    * no bytes need to be escaped and we can continue to the next chunk. If the mask is not 0 then we
+    * have at least one byte that needs to be escaped.
+    */
+
+    if (string_scan_simd_neon(&search->ptr, search->end, &search->matches_mask)) {
+        search->has_matches = true;
+        search->chunk_base = search->ptr;
+        search->chunk_end = search->ptr + sizeof(uint8x16_t);
+        return neon_next_match(search);
+    }
+
+    // There are fewer than 16 bytes left.
+    unsigned long remaining = (search->end - search->ptr);
+    if (remaining >= SIMD_MINIMUM_THRESHOLD) {
+        char *s = copy_remaining_bytes(search, sizeof(uint8x16_t), remaining);
+
+        uint64_t mask = compute_chunk_mask_neon(s);
+
+        if (!mask) {
+            // Nothing to escape, ensure search_flush doesn't do anything by setting
+            // search->cursor to search->ptr.
+            fbuffer_consumed(search->buffer, remaining);
+            search->ptr = search->end;
+            search->cursor = search->end;
+            return 0;
+        }
+
+        search->matches_mask = mask;
+        search->has_matches = true;
+        search->chunk_end = search->end;
+        search->chunk_base = search->ptr;
+        return neon_next_match(search);
+    }
+
+    if (search->ptr < search->end) {
+        return search_escape_basic(search);
+    }
+
+    search_flush(search);
+    return 0;
+}
+#endif /* HAVE_SIMD_NEON */
+
+#ifdef HAVE_SIMD_SSE2
+
+static inline FORCE_INLINE unsigned char sse2_next_match(search_state *search)
+{
+    int mask = search->matches_mask;
+    int index = trailing_zeros(mask);
+
+    // It is assumed escape_UTF8_char_basic will only ever increase search->ptr by at most one character.
+    // If we want to use a similar approach for full escaping we'll need to ensure:
+    //     search->chunk_base + index >= search->ptr
+    // However, since we know escape_UTF8_char_basic only increases search->ptr by one, if the next match
+    // is one byte after the previous match then:
+    //     search->chunk_base + index == search->ptr
+    search->ptr = search->chunk_base + index;
+    mask &= mask - 1;
+    search->matches_mask = mask;
+    search_flush(search);
+    return 1;
+}
+
+#if defined(__clang__) || defined(__GNUC__)
+#define TARGET_SSE2 __attribute__((target("sse2")))
+#else
+#define TARGET_SSE2
+#endif
+
+static inline TARGET_SSE2 FORCE_INLINE unsigned char search_escape_basic_sse2(search_state *search)
+{
+    if (RB_UNLIKELY(search->has_matches)) {
+        // There are more matches if search->matches_mask > 0.
+        if (search->matches_mask > 0) {
+            return sse2_next_match(search);
+        } else {
+            // sse2_next_match will only advance search->ptr up to the last matching character.
+            // Skip over any characters in the last chunk that occur after the last match.
+            search->has_matches = false;
+            if (RB_UNLIKELY(search->chunk_base + sizeof(__m128i) >= search->end)) {
+                search->ptr = search->end;
+            } else {
+                search->ptr = search->chunk_base + sizeof(__m128i);
+            }
+        }
+    }
+
+    if (string_scan_simd_sse2(&search->ptr, search->end, &search->matches_mask)) {
+        search->has_matches = true;
+        search->chunk_base = search->ptr;
+        search->chunk_end = search->ptr + sizeof(__m128i);
+        return sse2_next_match(search);
+    }
+
+    // There are fewer than 16 bytes left.
+    unsigned long remaining = (search->end - search->ptr);
+    if (remaining >= SIMD_MINIMUM_THRESHOLD) {
+        char *s = copy_remaining_bytes(search, sizeof(__m128i), remaining);
+
+        int needs_escape_mask = compute_chunk_mask_sse2(s);
+
+        if (needs_escape_mask == 0) {
+            // Nothing to escape, ensure search_flush doesn't do anything by setting
+            // search->cursor to search->ptr.
+            fbuffer_consumed(search->buffer, remaining);
+            search->ptr = search->end;
+            search->cursor = search->end;
+            return 0;
+        }
+
+        search->has_matches = true;
+        search->matches_mask = needs_escape_mask;
+        search->chunk_base = search->ptr;
+        return sse2_next_match(search);
+    }
+
+    if (search->ptr < search->end) {
+        return search_escape_basic(search);
+    }
+
+    search_flush(search);
+    return 0;
+}
+
+#endif /* HAVE_SIMD_SSE2 */
+
+#endif /* HAVE_SIMD */
+
 static const unsigned char script_safe_escape_table[256] = {
     // ASCII Control Characters
      9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
@@ -329,7 +586,8 @@ static inline unsigned char search_ascii_only_escape(search_state *search, const
     return 0;
 }
 
-static inline void full_escape_UTF8_char(search_state *search, unsigned char ch_len) {
+static inline void full_escape_UTF8_char(search_state *search, unsigned char ch_len)
+{
     const unsigned char ch = (unsigned char)*search->ptr;
     switch (ch_len) {
         case 1: {
@@ -359,7 +617,7 @@ static inline void full_escape_UTF8_char(search_state *search, unsigned char ch_
 
             uint32_t wchar = 0;
 
-            switch(ch_len) {
+            switch (ch_len) {
                 case 2:
                     wchar = ch & 0x1F;
                     break;
@@ -519,7 +777,8 @@ static VALUE mHash_to_json(int argc, VALUE *argv, VALUE self)
  * _state_ is a JSON::State object, that can also be used to configure the
  * produced JSON string output further.
  */
-static VALUE mArray_to_json(int argc, VALUE *argv, VALUE self) {
+static VALUE mArray_to_json(int argc, VALUE *argv, VALUE self)
+{
     rb_check_arity(argc, 0, 1);
     VALUE Vstate = cState_from_state_s(cState, argc == 1 ? argv[0] : Qnil);
     return cState_partial_generate(Vstate, self, generate_json_array, Qfalse);
@@ -581,7 +840,8 @@ static VALUE mFloat_to_json(int argc, VALUE *argv, VALUE self)
  *
  * Extends _modul_ with the String::Extend module.
  */
-static VALUE mString_included_s(VALUE self, VALUE modul) {
+static VALUE mString_included_s(VALUE self, VALUE modul)
+{
     VALUE result = rb_funcall(modul, i_extend, 1, mString_Extend);
     rb_call_super(1, &modul);
     return result;
@@ -788,6 +1048,21 @@ struct hash_foreach_arg {
     int iter;
 };
 
+static VALUE
+convert_string_subclass(VALUE key)
+{
+    VALUE key_to_s = rb_funcall(key, i_to_s, 0);
+
+    if (RB_UNLIKELY(!RB_TYPE_P(key_to_s, T_STRING))) {
+        VALUE cname = rb_obj_class(key);
+        rb_raise(rb_eTypeError,
+                 "can't convert %"PRIsVALUE" to %s (%"PRIsVALUE"#%s gives %"PRIsVALUE")",
+                 cname, "String", cname, "to_s", rb_obj_class(key_to_s));
+    }
+
+    return key_to_s;
+}
+
 static int
 json_object_i(VALUE key, VALUE val, VALUE _arg)
 {
@@ -801,22 +1076,22 @@ json_object_i(VALUE key, VALUE val, VALUE _arg)
     int j;
 
     if (arg->iter > 0) fbuffer_append_char(buffer, ',');
-    if (RB_UNLIKELY(state->object_nl)) {
-        fbuffer_append_str(buffer, state->object_nl);
+    if (RB_UNLIKELY(data->state->object_nl)) {
+        fbuffer_append_str(buffer, data->state->object_nl);
     }
-    if (RB_UNLIKELY(state->indent)) {
+    if (RB_UNLIKELY(data->state->indent)) {
         for (j = 0; j < depth; j++) {
-            fbuffer_append_str(buffer, state->indent);
+            fbuffer_append_str(buffer, data->state->indent);
         }
     }
 
     VALUE key_to_s;
-    switch(rb_type(key)) {
+    switch (rb_type(key)) {
         case T_STRING:
             if (RB_LIKELY(RBASIC_CLASS(key) == rb_cString)) {
                 key_to_s = key;
             } else {
-                key_to_s = rb_funcall(key, i_to_s, 0);
+                key_to_s = convert_string_subclass(key);
             }
             break;
         case T_SYMBOL:
@@ -828,21 +1103,22 @@ json_object_i(VALUE key, VALUE val, VALUE _arg)
     }
 
     if (RB_LIKELY(RBASIC_CLASS(key_to_s) == rb_cString)) {
-        generate_json_string(buffer, data, state, key_to_s);
+        generate_json_string(buffer, data, key_to_s);
     } else {
-        generate_json(buffer, data, state, key_to_s);
+        generate_json(buffer, data, key_to_s);
     }
-    if (RB_UNLIKELY(state->space_before)) fbuffer_append_str(buffer, state->space_before);
+    if (RB_UNLIKELY(state->space_before)) fbuffer_append_str(buffer, data->state->space_before);
     fbuffer_append_char(buffer, ':');
-    if (RB_UNLIKELY(state->space)) fbuffer_append_str(buffer, state->space);
-    generate_json(buffer, data, state, val);
+    if (RB_UNLIKELY(state->space)) fbuffer_append_str(buffer, data->state->space);
+    generate_json(buffer, data, val);
 
     arg->iter++;
     return ST_CONTINUE;
 }
 
-static inline long increase_depth(JSON_Generator_State *state)
+static inline long increase_depth(struct generate_json_data *data)
 {
+    JSON_Generator_State *state = data->state;
     long depth = ++state->depth;
     if (RB_UNLIKELY(depth > state->max_nesting && state->max_nesting)) {
         rb_raise(eNestingError, "nesting of %ld is too deep", --state->depth);
@@ -850,14 +1126,14 @@ static inline long increase_depth(JSON_Generator_State *state)
     return depth;
 }
 
-static void generate_json_object(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_object(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     int j;
-    long depth = increase_depth(state);
+    long depth = increase_depth(data);
 
     if (RHASH_SIZE(obj) == 0) {
         fbuffer_append(buffer, "{}", 2);
-        --state->depth;
+        --data->state->depth;
         return;
     }
 
@@ -869,49 +1145,49 @@ static void generate_json_object(FBuffer *buffer, struct generate_json_data *dat
     };
     rb_hash_foreach(obj, json_object_i, (VALUE)&arg);
 
-    depth = --state->depth;
-    if (RB_UNLIKELY(state->object_nl)) {
-        fbuffer_append_str(buffer, state->object_nl);
-        if (RB_UNLIKELY(state->indent)) {
+    depth = --data->state->depth;
+    if (RB_UNLIKELY(data->state->object_nl)) {
+        fbuffer_append_str(buffer, data->state->object_nl);
+        if (RB_UNLIKELY(data->state->indent)) {
             for (j = 0; j < depth; j++) {
-                fbuffer_append_str(buffer, state->indent);
+                fbuffer_append_str(buffer, data->state->indent);
             }
         }
     }
     fbuffer_append_char(buffer, '}');
 }
 
-static void generate_json_array(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_array(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     int i, j;
-    long depth = increase_depth(state);
+    long depth = increase_depth(data);
 
     if (RARRAY_LEN(obj) == 0) {
         fbuffer_append(buffer, "[]", 2);
-        --state->depth;
+        --data->state->depth;
         return;
     }
 
     fbuffer_append_char(buffer, '[');
-    if (RB_UNLIKELY(state->array_nl)) fbuffer_append_str(buffer, state->array_nl);
-    for(i = 0; i < RARRAY_LEN(obj); i++) {
+    if (RB_UNLIKELY(data->state->array_nl)) fbuffer_append_str(buffer, data->state->array_nl);
+    for (i = 0; i < RARRAY_LEN(obj); i++) {
         if (i > 0) {
             fbuffer_append_char(buffer, ',');
-            if (RB_UNLIKELY(state->array_nl)) fbuffer_append_str(buffer, state->array_nl);
+            if (RB_UNLIKELY(data->state->array_nl)) fbuffer_append_str(buffer, data->state->array_nl);
         }
-        if (RB_UNLIKELY(state->indent)) {
+        if (RB_UNLIKELY(data->state->indent)) {
             for (j = 0; j < depth; j++) {
-                fbuffer_append_str(buffer, state->indent);
+                fbuffer_append_str(buffer, data->state->indent);
             }
         }
-        generate_json(buffer, data, state, RARRAY_AREF(obj, i));
+        generate_json(buffer, data, RARRAY_AREF(obj, i));
     }
-    state->depth = --depth;
-    if (RB_UNLIKELY(state->array_nl)) {
-        fbuffer_append_str(buffer, state->array_nl);
-        if (RB_UNLIKELY(state->indent)) {
+    data->state->depth = --depth;
+    if (RB_UNLIKELY(data->state->array_nl)) {
+        fbuffer_append_str(buffer, data->state->array_nl);
+        if (RB_UNLIKELY(data->state->indent)) {
             for (j = 0; j < depth; j++) {
-                fbuffer_append_str(buffer, state->indent);
+                fbuffer_append_str(buffer, data->state->indent);
             }
         }
     }
@@ -960,7 +1236,7 @@ static inline VALUE ensure_valid_encoding(VALUE str)
     return str;
 }
 
-static void generate_json_string(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_string(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     obj = ensure_valid_encoding(obj);
 
@@ -973,12 +1249,18 @@ static void generate_json_string(FBuffer *buffer, struct generate_json_data *dat
     search.cursor = search.ptr;
     search.end = search.ptr + len;
 
-    switch(rb_enc_str_coderange(obj)) {
+#ifdef HAVE_SIMD
+    search.matches_mask = 0;
+    search.has_matches = false;
+    search.chunk_base = NULL;
+#endif /* HAVE_SIMD */
+
+    switch (rb_enc_str_coderange(obj)) {
         case ENC_CODERANGE_7BIT:
         case ENC_CODERANGE_VALID:
-            if (RB_UNLIKELY(state->ascii_only)) {
-                convert_UTF8_to_ASCII_only_JSON(&search, state->script_safe ? script_safe_escape_table : ascii_only_escape_table);
-            } else if (RB_UNLIKELY(state->script_safe)) {
+            if (RB_UNLIKELY(data->state->ascii_only)) {
+                convert_UTF8_to_ASCII_only_JSON(&search, data->state->script_safe ? script_safe_escape_table : ascii_only_escape_table);
+            } else if (RB_UNLIKELY(data->state->script_safe)) {
                 convert_UTF8_to_script_safe_JSON(&search);
             } else {
                 convert_UTF8_to_JSON(&search);
@@ -991,7 +1273,7 @@ static void generate_json_string(FBuffer *buffer, struct generate_json_data *dat
     fbuffer_append_char(buffer, '"');
 }
 
-static void generate_json_fallback(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_fallback(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     VALUE tmp;
     if (rb_respond_to(obj, i_to_json)) {
@@ -1001,100 +1283,116 @@ static void generate_json_fallback(FBuffer *buffer, struct generate_json_data *d
     } else {
         tmp = rb_funcall(obj, i_to_s, 0);
         Check_Type(tmp, T_STRING);
-        generate_json_string(buffer, data, state, tmp);
+        generate_json_string(buffer, data, tmp);
     }
 }
 
-static inline void generate_json_symbol(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static inline void generate_json_symbol(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
-    if (state->strict) {
-        generate_json_string(buffer, data, state, rb_sym2str(obj));
+    if (data->state->strict) {
+        generate_json_string(buffer, data, rb_sym2str(obj));
     } else {
-        generate_json_fallback(buffer, data, state, obj);
+        generate_json_fallback(buffer, data, obj);
     }
 }
 
-static void generate_json_null(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_null(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     fbuffer_append(buffer, "null", 4);
 }
 
-static void generate_json_false(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_false(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     fbuffer_append(buffer, "false", 5);
 }
 
-static void generate_json_true(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_true(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     fbuffer_append(buffer, "true", 4);
 }
 
-static void generate_json_fixnum(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_fixnum(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     fbuffer_append_long(buffer, FIX2LONG(obj));
 }
 
-static void generate_json_bignum(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_bignum(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     VALUE tmp = rb_funcall(obj, i_to_s, 0);
     fbuffer_append_str(buffer, tmp);
 }
 
 #ifdef RUBY_INTEGER_UNIFICATION
-static void generate_json_integer(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_integer(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     if (FIXNUM_P(obj))
-        generate_json_fixnum(buffer, data, state, obj);
+        generate_json_fixnum(buffer, data, obj);
     else
-        generate_json_bignum(buffer, data, state, obj);
+        generate_json_bignum(buffer, data, obj);
 }
 #endif
 
-static void generate_json_float(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_float(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     double value = RFLOAT_VALUE(obj);
-    char allow_nan = state->allow_nan;
-    if (!allow_nan) {
-        if (isinf(value) || isnan(value)) {
-            if (state->strict && state->as_json) {
-                VALUE casted_obj = rb_proc_call_with_block(state->as_json, 1, &obj, Qnil);
+    char allow_nan = data->state->allow_nan;
+    if (isinf(value) || isnan(value)) {
+        /* for NaN and Infinity values we either raise an error or rely on Float#to_s. */
+        if (!allow_nan) {
+            if (data->state->strict && data->state->as_json) {
+                VALUE casted_obj = rb_proc_call_with_block(data->state->as_json, 1, &obj, Qnil);
                 if (casted_obj != obj) {
-                    increase_depth(state);
-                    generate_json(buffer, data, state, casted_obj);
-                    state->depth--;
+                    increase_depth(data);
+                    generate_json(buffer, data, casted_obj);
+                    data->state->depth--;
                     return;
                 }
             }
             raise_generator_error(obj, "%"PRIsVALUE" not allowed in JSON", rb_funcall(obj, i_to_s, 0));
         }
+
+        VALUE tmp = rb_funcall(obj, i_to_s, 0);
+        fbuffer_append_str(buffer, tmp);
+        return;
     }
-    fbuffer_append_str(buffer, rb_funcall(obj, i_to_s, 0));
+
+    /* This implementation writes directly into the buffer. We reserve
+     * the 28 characters that fpconv_dtoa states as its maximum.
+     */
+    fbuffer_inc_capa(buffer, 28);
+    char* d = buffer->ptr + buffer->len;
+    int len = fpconv_dtoa(value, d);
+
+    /* fpconv_dtoa converts a float to its shortest string representation,
+     * but it adds a ".0" if this is a plain integer.
+     */
+    fbuffer_consumed(buffer, len);
 }
 
-static void generate_json_fragment(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_fragment(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     VALUE fragment = RSTRUCT_GET(obj, 0);
     Check_Type(fragment, T_STRING);
     fbuffer_append_str(buffer, fragment);
 }
 
-static void generate_json(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     bool as_json_called = false;
 start:
     if (obj == Qnil) {
-        generate_json_null(buffer, data, state, obj);
+        generate_json_null(buffer, data, obj);
     } else if (obj == Qfalse) {
-        generate_json_false(buffer, data, state, obj);
+        generate_json_false(buffer, data, obj);
     } else if (obj == Qtrue) {
-        generate_json_true(buffer, data, state, obj);
+        generate_json_true(buffer, data, obj);
     } else if (RB_SPECIAL_CONST_P(obj)) {
         if (RB_FIXNUM_P(obj)) {
-            generate_json_fixnum(buffer, data, state, obj);
+            generate_json_fixnum(buffer, data, obj);
         } else if (RB_FLONUM_P(obj)) {
-            generate_json_float(buffer, data, state, obj);
+            generate_json_float(buffer, data, obj);
         } else if (RB_STATIC_SYM_P(obj)) {
-            generate_json_symbol(buffer, data, state, obj);
+            generate_json_symbol(buffer, data, obj);
         } else {
             goto general;
         }
@@ -1102,43 +1400,43 @@ start:
         VALUE klass = RBASIC_CLASS(obj);
         switch (RB_BUILTIN_TYPE(obj)) {
             case T_BIGNUM:
-                generate_json_bignum(buffer, data, state, obj);
+                generate_json_bignum(buffer, data, obj);
                 break;
             case T_HASH:
                 if (klass != rb_cHash) goto general;
-                generate_json_object(buffer, data, state, obj);
+                generate_json_object(buffer, data, obj);
                 break;
             case T_ARRAY:
                 if (klass != rb_cArray) goto general;
-                generate_json_array(buffer, data, state, obj);
+                generate_json_array(buffer, data, obj);
                 break;
             case T_STRING:
                 if (klass != rb_cString) goto general;
-                generate_json_string(buffer, data, state, obj);
+                generate_json_string(buffer, data, obj);
                 break;
             case T_SYMBOL:
-                generate_json_symbol(buffer, data, state, obj);
+                generate_json_symbol(buffer, data, obj);
                 break;
             case T_FLOAT:
                 if (klass != rb_cFloat) goto general;
-                generate_json_float(buffer, data, state, obj);
+                generate_json_float(buffer, data, obj);
                 break;
             case T_STRUCT:
                 if (klass != cFragment) goto general;
-                generate_json_fragment(buffer, data, state, obj);
+                generate_json_fragment(buffer, data, obj);
                 break;
             default:
             general:
-                if (state->strict) {
-                    if (RTEST(state->as_json) && !as_json_called) {
-                        obj = rb_proc_call_with_block(state->as_json, 1, &obj, Qnil);
+                if (data->state->strict) {
+                    if (RTEST(data->state->as_json) && !as_json_called) {
+                        obj = rb_proc_call_with_block(data->state->as_json, 1, &obj, Qnil);
                         as_json_called = true;
                         goto start;
                     } else {
                         raise_generator_error(obj, "%"PRIsVALUE" not allowed in JSON", CLASS_OF(obj));
                     }
                 } else {
-                    generate_json_fallback(buffer, data, state, obj);
+                    generate_json_fallback(buffer, data, obj);
                 }
         }
     }
@@ -1148,7 +1446,7 @@ static VALUE generate_json_try(VALUE d)
 {
     struct generate_json_data *data = (struct generate_json_data *)d;
 
-    data->func(data->buffer, data, data->state, data->obj);
+    data->func(data->buffer, data, data->obj);
 
     return Qnil;
 }
@@ -1609,15 +1907,30 @@ static VALUE cState_buffer_initial_length_set(VALUE self, VALUE buffer_initial_l
     return Qnil;
 }
 
+struct configure_state_data {
+    JSON_Generator_State *state;
+    VALUE vstate;  // Ruby object that owns the state, or Qfalse if stack-allocated
+};
+
+static inline void state_write_value(struct configure_state_data *data, VALUE *field, VALUE value)
+{
+    if (RTEST(data->vstate)) {
+        RB_OBJ_WRITE(data->vstate, field, value);
+    } else {
+        *field = value;
+    }
+}
+
 static int configure_state_i(VALUE key, VALUE val, VALUE _arg)
 {
-    JSON_Generator_State *state = (JSON_Generator_State *)_arg;
+    struct configure_state_data *data = (struct configure_state_data *)_arg;
+    JSON_Generator_State *state = data->state;
 
-         if (key == sym_indent)                { state->indent = string_config(val); }
-    else if (key == sym_space)                 { state->space = string_config(val); }
-    else if (key == sym_space_before)          { state->space_before = string_config(val); }
-    else if (key == sym_object_nl)             { state->object_nl = string_config(val); }
-    else if (key == sym_array_nl)              { state->array_nl = string_config(val); }
+         if (key == sym_indent)                { state_write_value(data, &state->indent, string_config(val)); }
+    else if (key == sym_space)                 { state_write_value(data, &state->space, string_config(val)); }
+    else if (key == sym_space_before)          { state_write_value(data, &state->space_before, string_config(val)); }
+    else if (key == sym_object_nl)             { state_write_value(data, &state->object_nl, string_config(val)); }
+    else if (key == sym_array_nl)              { state_write_value(data, &state->array_nl, string_config(val)); }
     else if (key == sym_max_nesting)           { state->max_nesting = long_config(val); }
     else if (key == sym_allow_nan)             { state->allow_nan = RTEST(val); }
     else if (key == sym_ascii_only)            { state->ascii_only = RTEST(val); }
@@ -1626,11 +1939,14 @@ static int configure_state_i(VALUE key, VALUE val, VALUE _arg)
     else if (key == sym_script_safe)           { state->script_safe = RTEST(val); }
     else if (key == sym_escape_slash)          { state->script_safe = RTEST(val); }
     else if (key == sym_strict)                { state->strict = RTEST(val); }
-    else if (key == sym_as_json)               { state->as_json = RTEST(val) ? rb_convert_type(val, T_DATA, "Proc", "to_proc") : Qfalse; }
+    else if (key == sym_as_json)               {
+        VALUE proc = RTEST(val) ? rb_convert_type(val, T_DATA, "Proc", "to_proc") : Qfalse;
+        state_write_value(data, &state->as_json, proc);
+    }
     return ST_CONTINUE;
 }
 
-static void configure_state(JSON_Generator_State *state, VALUE config)
+static void configure_state(JSON_Generator_State *state, VALUE vstate, VALUE config)
 {
     if (!RTEST(config)) return;
 
@@ -1638,15 +1954,20 @@ static void configure_state(JSON_Generator_State *state, VALUE config)
 
     if (!RHASH_SIZE(config)) return;
 
+    struct configure_state_data data = {
+        .state = state,
+        .vstate = vstate
+    };
+
     // We assume in most cases few keys are set so it's faster to go over
     // the provided keys than to check all possible keys.
-    rb_hash_foreach(config, configure_state_i, (VALUE)state);
+    rb_hash_foreach(config, configure_state_i, (VALUE)&data);
 }
 
 static VALUE cState_configure(VALUE self, VALUE opts)
 {
     GET_STATE(self);
-    configure_state(state, opts);
+    configure_state(state, self, opts);
     return self;
 }
 
@@ -1654,7 +1975,7 @@ static VALUE cState_m_generate(VALUE klass, VALUE obj, VALUE opts, VALUE io)
 {
     JSON_Generator_State state = {0};
     state_init(&state);
-    configure_state(&state, opts);
+    configure_state(&state, Qfalse, opts);
 
     char stack_buffer[FBUFFER_STACK_SIZE];
     FBuffer buffer = {
@@ -1819,4 +2140,23 @@ void Init_generator(void)
     binary_encindex = rb_ascii8bit_encindex();
 
     rb_require("json/ext/generator/state");
+
+
+    switch (find_simd_implementation()) {
+#ifdef HAVE_SIMD
+#ifdef HAVE_SIMD_NEON
+        case SIMD_NEON:
+            search_escape_basic_impl = search_escape_basic_neon;
+            break;
+#endif /* HAVE_SIMD_NEON */
+#ifdef HAVE_SIMD_SSE2
+        case SIMD_SSE2:
+            search_escape_basic_impl = search_escape_basic_sse2;
+            break;
+#endif /* HAVE_SIMD_SSE2 */
+#endif /* HAVE_SIMD */
+        default:
+            search_escape_basic_impl = search_escape_basic;
+            break;
+    }
 }