Store insertion order using offsets

Include/internal/pycore_dict.h

@@ -329,6 +329,15 @@ get_insertion_order_array(PyDictValues *values)
     return (uint8_t *)&values->values[values->capacity];
 }
 
+/* The insertion order array stores deltas (ix - position) instead of
+   absolute indices.
+
+   A sequential store (ix == values->size) has delta 0.  The slots of the
+   array beyond values->size are kept zero (zeroed when the inline values
+   are allocated, and zeroed again in delete_index_from_values when an
+   entry is removed), so callers handling that common case can skip this
+   function entirely and just bump values->size -- see
+   _STORE_ATTR_INSTANCE_VALUE. */
 static inline void
 _PyDictValues_AddToInsertionOrder(PyDictValues *values, Py_ssize_t ix)
 {
@@ -337,7 +346,7 @@ _PyDictValues_AddToInsertionOrder(PyDictValues *values, Py_ssize_t ix)
     uint8_t *array = get_insertion_order_array(values);
     assert(size < values->capacity);
     assert(((uint8_t)ix) == ix);
-    array[size] = (uint8_t)ix;
+    array[size] = (uint8_t)(ix - size);
     values->size = size+1;
 }
 

Include/internal/pycore_optimizer.h

@@ -95,6 +95,18 @@ typedef struct _JitOptContext {
     JitOptRef stack_array[ABSTRACT_INTERP_STACK_SIZE];
     _PyJitUopBuffer out_buffer;
     _PyBloomFilter *dependencies;
+
+    /* gh-138453: static tracking of a freshly allocated object's inline
+     * values->size, used to specialise _STORE_ATTR_INSTANCE_VALUE.
+     * fresh_alloc_pending_type is set by _CHECK_OBJECT and consumed by the
+     * immediately following _ALLOCATE_OBJECT.  fresh_alloc_sym is the symbol
+     * of the object being tracked (NULL when not tracking); fresh_alloc_count
+     * is the number of sequential stores proven so far; fresh_alloc_next_offset
+     * is the byte offset of the next expected inline value slot. */
+    PyTypeObject *fresh_alloc_pending_type;
+    JitOptSymbol *fresh_alloc_sym;
+    int fresh_alloc_count;
+    uint32_t fresh_alloc_next_offset;
 } JitOptContext;