Add a best effort heap-wide object property slot cache

config/config-options/DUK_USE_HOBJECT_HASH_PROP_LIMIT.yaml

@@ -1,6 +1,6 @@
 define: DUK_USE_HOBJECT_HASH_PROP_LIMIT
 introduced: 2.1.0
-default: 8
+default: 64
 tags:
   - performance
   - lowmemory
@@ -18,3 +18,11 @@ description: >
 
   A lower value improves performance (a value as low a 4-8 can be useful)
   while a higher value conserves memory.
+
+  The heap-wide object property slot cache accelerates repeated property
+  accesses to small and medium size objects with no upfront cost, so the
+  default limit has been raised to a relatively high value; the hash part
+  still makes sense for large objects where slot cache misses lead to very
+  large linear scans.  The hash part also makes sense for long term objects
+  which are accessed often, e.g. prototype objects, but these's currently not
+  enough state to ensure they get and keep a hash part.

src-input/duk_forwdecl.h

@@ -45,6 +45,7 @@ struct duk_breakpoint;
 struct duk_activation;
 struct duk_catcher;
 struct duk_strcache;
+struct duk_slotcache_entry;
 struct duk_ljstate;
 struct duk_strtab_entry;
 
@@ -105,6 +106,7 @@ typedef struct duk_breakpoint duk_breakpoint;
 typedef struct duk_activation duk_activation;
 typedef struct duk_catcher duk_catcher;
 typedef struct duk_strcache duk_strcache;
+typedef struct duk_slotcache_entry duk_slotcache_entry;
 typedef struct duk_ljstate duk_ljstate;
 typedef struct duk_strtab_entry duk_strtab_entry;
 

src-input/duk_heap.h

@@ -146,6 +146,11 @@
 #define DUK_HEAP_REMOVE_FROM_FINALIZE_LIST(heap,hdr)      duk_heap_remove_from_finalize_list((heap), (hdr))
 #endif
 
+/* Slotcache is used to speeding up object property lookups without
+ * caching a property value.
+ */
+#define DUK_HEAP_SLOTCACHE_SIZE                           256
+
 /*
  *  Built-in strings
  */
@@ -285,6 +290,15 @@ struct duk_strcache {
 	duk_uint32_t cidx;
 };
 
+/*
+ *  Property slot cache.  Best effort data structure to cache property
+ *  table slots for recently accessed object/key combinations.
+ */
+
+struct duk_slotcache_entry {
+	duk_uint8_t slot;
+};
+
 /*
  *  Longjmp state, contains the information needed to perform a longjmp.
  *  Longjmp related values are written to value1, value2, and iserror.
@@ -535,6 +549,11 @@ struct duk_heap {
 	 */
 	duk_strcache strcache[DUK_HEAP_STRCACHE_SIZE];
 
+	/* Property slot cache, global structure for speeding up commonly
+	 * referenced property access.
+	 */
+	duk_slotcache_entry slotcache[DUK_HEAP_SLOTCACHE_SIZE];
+
 	/* Built-in strings. */
 #if defined(DUK_USE_ROM_STRINGS)
 	/* No field needed when strings are in ROM. */
@@ -595,6 +614,11 @@ DUK_INTERNAL void duk_heap_strtable_dump(duk_heap *heap);
 DUK_INTERNAL_DECL void duk_heap_strcache_string_remove(duk_heap *heap, duk_hstring *h);
 DUK_INTERNAL_DECL duk_uint_fast32_t duk_heap_strcache_offset_char2byte(duk_hthread *thr, duk_hstring *h, duk_uint_fast32_t char_offset);
 
+/* FIXME: duk_small_uint_t? */
+DUK_INTERNAL_DECL duk_uint32_t duk_heap_slotcache_lookup(duk_heap *heap, duk_hobject *obj, duk_hstring *key);
+DUK_INTERNAL_DECL void duk_heap_slotcache_insert(duk_heap *heap, duk_hobject *obj, duk_hstring *key, duk_uint32_t slot);
+DUK_INTERNAL_DECL duk_uint32_t duk_heap_slotcache_getkey(duk_hobject *obj, duk_hstring *key);
+
 #if defined(DUK_USE_PROVIDE_DEFAULT_ALLOC_FUNCTIONS)
 DUK_INTERNAL_DECL void *duk_default_alloc_function(void *udata, duk_size_t size);
 DUK_INTERNAL_DECL void *duk_default_realloc_function(void *udata, void *ptr, duk_size_t newsize);

src-input/duk_heap_alloc.c

@@ -680,6 +680,7 @@ DUK_LOCAL void duk__dump_type_sizes(void) {
 	DUK__DUMPSZ(duk_activation);
 	DUK__DUMPSZ(duk_catcher);
 	DUK__DUMPSZ(duk_strcache);
+	DUK__DUMPSZ(duk_slotcache_entry);
 	DUK__DUMPSZ(duk_ljstate);
 	DUK__DUMPSZ(duk_fixedbuffer);
 	DUK__DUMPSZ(duk_bitdecoder_ctx);
@@ -1051,6 +1052,12 @@ duk_heap *duk_heap_alloc(duk_alloc_function alloc_func,
 	}
 #endif
 
+	/*
+	 *  Init slotcache
+	 */
+
+	/* Nothing to init now. */
+
 	/* XXX: error handling is incomplete.  It would be cleanest if
 	 * there was a setjmp catchpoint, so that all init code could
 	 * freely throw errors.  If that were the case, the return code

src-input/duk_heap_slotcache.c

@@ -0,0 +1,52 @@
+/*
+ *  Object property table slot cache.
+ */
+
+#include "duk_internal.h"
+
+/* FIXME: check if it's useful to compute the slotcache key only once in property code
+ * (live range is not trivial so maybe not).
+ */
+
+DUK_LOCAL DUK_ALWAYS_INLINE duk_uint32_t duk__slotcache_compute_key(duk_hobject *obj, duk_hstring *key) {
+	DUK_ASSERT(obj != NULL);
+	DUK_ASSERT(key != NULL);
+
+	return (((duk_uint32_t) obj) ^ DUK_HSTRING_GET_HASH(key)) & (DUK_HEAP_SLOTCACHE_SIZE - 1);
+}
+
+DUK_INTERNAL duk_uint32_t duk_heap_slotcache_getkey(duk_hobject *obj, duk_hstring *key) {
+	DUK_ASSERT(obj != NULL);
+	DUK_ASSERT(key != NULL);
+
+	return duk__slotcache_compute_key(obj, key);
+}
+
+DUK_INTERNAL duk_uint32_t duk_heap_slotcache_lookup(duk_heap *heap, duk_hobject *obj, duk_hstring *key) {
+	duk_uint32_t idx;
+	duk_slotcache_entry *ent;
+
+	DUK_ASSERT(heap != NULL);
+	DUK_ASSERT(obj != NULL);
+	DUK_ASSERT(key != NULL);
+
+	/* FIXME: assert: hashlimit <= 256 */
+
+	idx = duk__slotcache_compute_key(obj, key);
+	ent = heap->slotcache + idx;
+	return ent->slot;
+}
+
+DUK_INTERNAL void duk_heap_slotcache_insert(duk_heap *heap, duk_hobject *obj, duk_hstring *key, duk_uint32_t slot) {
+	duk_uint32_t idx;
+	duk_slotcache_entry *ent;
+
+	/* FIXME: assert: hashlimit <= 256 */
+	/* FIXME: skip check if hash part present and hashlimit correct */
+	if (DUK_UNLIKELY(slot >= DUK_UINT8_MAX)) {
+		return;
+	}
+	idx = duk__slotcache_compute_key(obj, key);
+	ent = heap->slotcache + idx;
+	ent->slot = (duk_uint16_t) slot;
+}

src-input/duk_hobject_props.c

@@ -1125,6 +1125,8 @@ DUK_INTERNAL void duk_hobject_compact_props(duk_hthread *thr, duk_hobject *obj)
  */
 
 DUK_INTERNAL void duk_hobject_find_existing_entry(duk_heap *heap, duk_hobject *obj, duk_hstring *key, duk_int_t *e_idx, duk_int_t *h_idx) {
+	duk_uint32_t slot;
+
 	DUK_ASSERT(obj != NULL);
 	DUK_ASSERT(key != NULL);
 	DUK_ASSERT(e_idx != NULL);
@@ -1143,10 +1145,26 @@ DUK_INTERNAL void duk_hobject_find_existing_entry(duk_heap *heap, duk_hobject *o
 		duk_hstring **h_keys_base;
 		DUK_DDD(DUK_DDDPRINT("duk_hobject_find_existing_entry() using linear scan for lookup"));
 
+		slot = duk_heap_slotcache_lookup(heap, obj, key);
+		if (slot < DUK_HOBJECT_GET_ENEXT(obj)) {
+			if (DUK_HOBJECT_E_GET_KEY(heap, obj, slot) == key) {
+				DUK_D(DUK_DPRINT("slot cache hit: %p %p %!O -> %ld (slotcache lookup key %lu)",
+				                 (void *) obj, (void *) key, key, (long) slot,
+				                 (unsigned long) duk_heap_slotcache_getkey(obj, key)));
+				*e_idx = slot;
+				*h_idx = -1;
+				return;
+			}
+		}
+
 		h_keys_base = DUK_HOBJECT_E_GET_KEY_BASE(heap, obj);
 		n = DUK_HOBJECT_GET_ENEXT(obj);
 		for (i = 0; i < n; i++) {
 			if (h_keys_base[i] == key) {
+				DUK_D(DUK_DPRINT("slot cache insert: %p %p %!O -> %ld (slotcache lookup key %lu)",
+				                 (void *) obj, (void *) key, key, (long) i,
+				                 (unsigned long) duk_heap_slotcache_getkey(obj, key)));
+				duk_heap_slotcache_insert(heap, obj, key, i);
 				*e_idx = i;
 				*h_idx = -1;
 				return;

tools/configure.py

@@ -465,6 +465,7 @@ def default_from_script_path(optname, orig, alternatives):
         'duk_heap_refcount.c',
         'duk_heap_stringcache.c',
         'duk_heap_stringtable.c',
+        'duk_heap_slotcache.c',
         'duk_hnatfunc.h',
         'duk_hobject_alloc.c',
         'duk_hobject_class.c',

util/dist.py

@@ -373,6 +373,7 @@ def main():
         'duk_heap_refcount.c',
         'duk_heap_stringcache.c',
         'duk_heap_stringtable.c',
+        'duk_heap_slotcache.c',
         'duk_henv.h',
         'duk_hnatfunc.h',
         'duk_hobject_alloc.c',