Add ref-counted string slices (#1175)

Instead of copying long substring slices, store a reference to and
offset into the parent string.

Gets more profitable as strings get longer. For 64k string slices,
it's about 6.5x faster than copying.

The downside of ref-counted string slices is that they can keep the
parent string alive for longer than it otherwise would have been,
leading to memory usage that is higher than without string slices.
That's why this optimization is only applied to long-ish slices,
currently slices > 1,024 bytes.

Possible future enhancements are slicing only when the substring
is > x% of the parent string, or copying lazily when the slice
string is the only thing referencing the parent string.

Author: bnoordhuis

quickjs.c

@@ -211,6 +211,9 @@ typedef enum JSErrorEnum {
 #define JS_MAX_LOCAL_VARS 65535
 #define JS_STACK_SIZE_MAX 65534
 #define JS_STRING_LEN_MAX ((1 << 30) - 1)
+// 1,024 bytes is about the cutoff point where it starts getting
+// more profitable to ref slice than to copy
+#define JS_STRING_SLICE_LEN_MAX 1024 // in bytes
 
 #define __exception __attribute__((warn_unused_result))
 
@@ -545,7 +548,12 @@ typedef enum {
     JS_ATOM_KIND_PRIVATE,
 } JSAtomKindEnum;
 
-#define JS_ATOM_HASH_MASK  ((1 << 30) - 1)
+typedef enum {
+    JS_STRING_KIND_NORMAL,
+    JS_STRING_KIND_SLICE,
+} JSStringKind;
+
+#define JS_ATOM_HASH_MASK  ((1 << 29) - 1)
 
 struct JSString {
     JSRefCountHeader header; /* must come first, 32-bit */
@@ -554,7 +562,8 @@ struct JSString {
     /* for JS_ATOM_TYPE_SYMBOL: hash = 0, atom_type = 3,
        for JS_ATOM_TYPE_PRIVATE: hash = 1, atom_type = 3
        XXX: could change encoding to have one more bit in hash */
-    uint32_t hash : 30;
+    uint32_t hash : 29;
+    uint8_t kind : 1;
     uint8_t atom_type : 2; /* != 0 if atom, JS_ATOM_TYPE_x */
     uint32_t hash_next; /* atom_index for JS_ATOM_TYPE_SYMBOL */
     JSWeakRefRecord *first_weak_ref;
@@ -563,14 +572,39 @@ struct JSString {
 #endif
 };
 
+typedef struct JSStringSlice {
+    JSString *parent;
+    uint32_t start; // in characters, not bytes
+} JSStringSlice;
+
 static inline uint8_t *str8(JSString *p)
 {
-    return (void *)(p + 1);
+    JSStringSlice *slice;
+
+    switch (p->kind) {
+    case JS_STRING_KIND_NORMAL:
+        return (void *)&p[1];
+    case JS_STRING_KIND_SLICE:
+        slice = (void *)&p[1];
+        return str8(slice->parent) + slice->start;
+    }
+    abort();
+    return NULL;
 }
 
 static inline uint16_t *str16(JSString *p)
 {
-    return (void *)(p + 1);
+    JSStringSlice *slice;
+
+    switch (p->kind) {
+    case JS_STRING_KIND_NORMAL:
+        return (void *)&p[1];
+    case JS_STRING_KIND_SLICE:
+        slice = (void *)&p[1];
+        return str16(slice->parent) + slice->start;
+    }
+    abort();
+    return NULL;
 }
 
 typedef struct JSClosureVar {
@@ -2049,6 +2083,7 @@ static JSString *js_alloc_string_rt(JSRuntime *rt, int max_len, int is_wide_char
     str->header.ref_count = 1;
     str->is_wide_char = is_wide_char;
     str->len = max_len;
+    str->kind = JS_STRING_KIND_NORMAL;
     str->atom_type = 0;
     str->hash = 0;          /* optional but costless */
     str->hash_next = 0;     /* optional */
@@ -2069,18 +2104,28 @@ static JSString *js_alloc_string(JSContext *ctx, int max_len, int is_wide_char)
     return p;
 }
 
+static inline void js_free_string0(JSRuntime *rt, JSString *str);
+
 /* same as JS_FreeValueRT() but faster */
 static inline void js_free_string(JSRuntime *rt, JSString *str)
 {
-    if (--str->header.ref_count <= 0) {
-        if (str->atom_type) {
-            JS_FreeAtomStruct(rt, str);
-        } else {
+    if (--str->header.ref_count <= 0)
+        js_free_string0(rt, str);
+}
+
+static inline void js_free_string0(JSRuntime *rt, JSString *str)
+{
+    if (str->atom_type) {
+        JS_FreeAtomStruct(rt, str);
+    } else {
 #ifdef ENABLE_DUMPS // JS_DUMP_LEAKS
-            list_del(&str->link);
+        list_del(&str->link);
 #endif
-            js_free_rt(rt, str);
+        if (str->kind == JS_STRING_KIND_SLICE) {
+            JSStringSlice *slice = (void *)&str[1];
+            js_free_string(rt, slice->parent); // safe, recurses only 1 level
         }
+        js_free_rt(rt, str);
     }
 }
 
@@ -2962,6 +3007,7 @@ static JSAtom __JS_NewAtom(JSRuntime *rt, JSString *str, int atom_type)
             p->header.ref_count = 1;
             p->is_wide_char = str->is_wide_char;
             p->len = str->len;
+            p->kind = JS_STRING_KIND_NORMAL;
 #ifdef ENABLE_DUMPS // JS_DUMP_LEAKS
             list_add_tail(&p->link, &rt->string_list);
 #endif
@@ -2976,6 +3022,7 @@ static JSAtom __JS_NewAtom(JSRuntime *rt, JSString *str, int atom_type)
         p->header.ref_count = 1;
         p->is_wide_char = 1;    /* Hack to represent NULL as a JSString */
         p->len = 0;
+        p->kind = JS_STRING_KIND_NORMAL;
 #ifdef ENABLE_DUMPS // JS_DUMP_LEAKS
         list_add_tail(&p->link, &rt->string_list);
 #endif
@@ -3680,13 +3727,37 @@ static JSValue js_new_string_char(JSContext *ctx, uint16_t c)
 
 static JSValue js_sub_string(JSContext *ctx, JSString *p, int start, int end)
 {
-    int len = end - start;
+    JSStringSlice *slice;
+    JSString *q;
+    int len;
+
+    len = end - start;
     if (start == 0 && end == p->len) {
         return js_dup(JS_MKPTR(JS_TAG_STRING, p));
     }
     if (len <= 0) {
         return js_empty_string(ctx->rt);
     }
+    if (len > (JS_STRING_SLICE_LEN_MAX >> p->is_wide_char)) {
+        if (p->kind == JS_STRING_KIND_SLICE) {
+            slice = (void *)&p[1];
+            p = slice->parent;
+            start += slice->start;
+        }
+        // allocate as 16 bit wide string to avoid wastage;
+        // js_alloc_string allocates 1 byte extra for 8 bit strings;
+        q = js_alloc_string(ctx, sizeof(*slice)/2, /*is_wide_char*/true);
+        if (!q)
+            return JS_EXCEPTION;
+        q->is_wide_char = p->is_wide_char;
+        q->kind = JS_STRING_KIND_SLICE;
+        q->len = len;
+        slice = (void *)&q[1];
+        slice->parent = p;
+        slice->start = start;
+        p->header.ref_count++;
+        return JS_MKPTR(JS_TAG_STRING, q);
+    }
     if (p->is_wide_char) {
         JSString *str;
         int i;
@@ -5750,17 +5821,7 @@ static void js_free_value_rt(JSRuntime *rt, JSValue v)
 
     switch(tag) {
     case JS_TAG_STRING:
-        {
-            JSString *p = JS_VALUE_GET_STRING(v);
-            if (p->atom_type) {
-                JS_FreeAtomStruct(rt, p);
-            } else {
-#ifdef ENABLE_DUMPS // JS_DUMP_LEAKS
-                list_del(&p->link);
-#endif
-                js_free_rt(rt, p);
-            }
-        }
+        js_free_string0(rt, JS_VALUE_GET_STRING(v));
         break;
     case JS_TAG_OBJECT:
     case JS_TAG_FUNCTION_BYTECODE:
@@ -58038,6 +58099,13 @@ uintptr_t js_std_cmd(int cmd, ...) {
         *pv = ctx->error_back_trace;
         ctx->error_back_trace = JS_UNDEFINED;
         break;
+    case 3: // GetStringKind
+        ctx = va_arg(ap, JSContext *);
+        pv = va_arg(ap, JSValue *);
+        rv = -1;
+        if (JS_IsString(*pv))
+            rv = JS_VALUE_GET_STRING(*pv)->kind;
+        break;
     default:
         rv = -1;
     }

run-test262.c

@@ -1700,15 +1700,32 @@ void update_stats(JSRuntime *rt, const char *filename) {
     js_mutex_unlock(&stats_mutex);
 }
 
+static JSValue qjs_black_box(JSContext *ctx, JSValueConst this_val,
+                            int argc, JSValueConst argv[], int magic)
+{
+    return JS_NewInt32(ctx, js_std_cmd(magic, ctx, &argv[0]));
+}
+
+static const JSCFunctionListEntry qjs_methods[] = {
+    JS_CFUNC_MAGIC_DEF("getStringKind", 1, qjs_black_box, /*GetStringKind*/3),
+};
+
+static const JSCFunctionListEntry qjs_object =
+    JS_OBJECT_DEF("qjs", qjs_methods, countof(qjs_methods), JS_PROP_C_W_E);
+
 JSContext *JS_NewCustomContext(JSRuntime *rt)
 {
     JSContext *ctx;
+    JSValue obj;
 
     ctx = JS_NewContext(rt);
     if (ctx && local) {
         js_init_module_std(ctx, "qjs:std");
         js_init_module_os(ctx, "qjs:os");
         js_init_module_bjson(ctx, "qjs:bjson");
+        obj = JS_GetGlobalObject(ctx);
+        JS_SetPropertyFunctionList(ctx, obj, &qjs_object, 1);
+        JS_FreeValue(ctx, obj);
     }
     return ctx;
 }

tests/microbench.js

@@ -797,6 +797,39 @@ function string_build4(n)
     return n * 100;
 }
 
+function string_slice1(n)
+{
+    var i, j, s;
+    s = "x".repeat(1<<16);
+    for (i = 0; i < n; i++) {
+        for (j = 0; j < 1000; j++)
+            s.slice(-1); // too short for JSStringSlice
+    }
+    return n * 1000;
+}
+
+function string_slice2(n)
+{
+    var i, j, s;
+    s = "x".repeat(1<<16);
+    for (i = 0; i < n; i++) {
+        for (j = 0; j < 1000; j++)
+            s.slice(-1024);
+    }
+    return n * 1000;
+}
+
+function string_slice3(n)
+{
+    var i, j, s;
+    s = "x".repeat(1<<16);
+    for (i = 0; i < n; i++) {
+        for (j = 0; j < 1000; j++)
+            s.slice(1);
+    }
+    return n * 1000;
+}
+
 /* sort bench */
 
 function sort_bench(text) {
@@ -1114,6 +1147,9 @@ function main(argc, argv, g)
         string_build2,
         //string_build3,
         //string_build4,
+        string_slice1,
+        string_slice2,
+        string_slice3,
         sort_bench,
         int_to_string,
         int_toString,

tests/test_builtin.js

@@ -378,6 +378,11 @@ function test_string()
     assert(eval('"\0"'), "\0");
 
     assert("abc".padStart(Infinity, ""), "abc");
+
+    assert(qjs.getStringKind("xyzzy".slice(1)),
+           /*JS_STRING_KIND_NORMAL*/0);
+    assert(qjs.getStringKind("xyzzy".repeat(512).slice(1)),
+           /*JS_STRING_KIND_SLICE*/1);
 }
 
 function test_math()