Enforce consistency between shape_id and FL_EXIVAR

The FL_EXIVAR is a bit redundant with the shape_id.
Now that the `shape_id` is embedded in all objects on all archs,
we can cheaply check if an object has any fields with a simple
bitmask.

Author: byroot

gc.c

@@ -2072,7 +2072,6 @@ rb_gc_obj_free_vm_weak_references(VALUE obj)
 
     if (FL_TEST_RAW(obj, FL_EXIVAR)) {
         rb_free_generic_ivar((VALUE)obj);
-        FL_UNSET_RAW(obj, FL_EXIVAR);
     }
 
     switch (BUILTIN_TYPE(obj)) {

shape.c

@@ -1266,6 +1266,13 @@ rb_shape_verify_consistency(VALUE obj, shape_id_t shape_id)
         }
     }
 
+    if (FL_TEST_RAW(obj, FL_EXIVAR)) {
+        RUBY_ASSERT(rb_obj_has_exivar(obj));
+    }
+    else {
+        RUBY_ASSERT(!rb_obj_has_exivar(obj));
+    }
+
     return true;
 }
 #endif

shape.h

@@ -136,15 +136,14 @@ RBASIC_SET_SHAPE_ID(VALUE obj, shape_id_t shape_id)
 {
     RUBY_ASSERT(!RB_SPECIAL_CONST_P(obj));
     RUBY_ASSERT(!RB_TYPE_P(obj, T_IMEMO) || IMEMO_TYPE_P(obj, imemo_class_fields));
-    RUBY_ASSERT(rb_shape_verify_consistency(obj, shape_id));
-
 #if RBASIC_SHAPE_ID_FIELD
     RBASIC(obj)->shape_id = (VALUE)shape_id;
 #else
     // Object shapes are occupying top bits
     RBASIC(obj)->flags &= SHAPE_FLAG_MASK;
     RBASIC(obj)->flags |= ((VALUE)(shape_id) << SHAPE_FLAG_SHIFT);
 #endif
+    RUBY_ASSERT(rb_shape_verify_consistency(obj, shape_id));
 }
 
 static inline rb_shape_t *
@@ -343,6 +342,34 @@ rb_shape_obj_has_ivars(VALUE obj)
     return rb_shape_has_ivars(RBASIC_SHAPE_ID(obj));
 }
 
+static inline bool
+rb_shape_has_fields(shape_id_t shape_id)
+{
+    return shape_id & (SHAPE_ID_OFFSET_MASK | SHAPE_ID_FL_TOO_COMPLEX);
+}
+
+static inline bool
+rb_shape_obj_has_fields(VALUE obj)
+{
+    return rb_shape_has_fields(RBASIC_SHAPE_ID(obj));
+}
+
+static inline bool
+rb_obj_has_exivar(VALUE obj)
+{
+    switch (TYPE(obj)) {
+        case T_NONE:
+        case T_OBJECT:
+        case T_CLASS:
+        case T_MODULE:
+        case T_IMEMO:
+          return false;
+        default:
+          break;
+    }
+    return rb_shape_obj_has_fields(obj);
+}
+
 // For ext/objspace
 RUBY_SYMBOL_EXPORT_BEGIN
 typedef void each_shape_callback(shape_id_t shape_id, void *data);

variable.c

@@ -1270,6 +1270,7 @@ rb_free_generic_ivar(VALUE obj)
             xfree(fields_tbl);
         }
     }
+    FL_UNSET_RAW(obj, FL_EXIVAR);
 }
 
 size_t
@@ -1542,23 +1543,30 @@ rb_ivar_delete(VALUE obj, ID id, VALUE undef)
 
     RUBY_ASSERT(removed_shape_id != INVALID_SHAPE_ID);
 
-    attr_index_t new_fields_count = RSHAPE_LEN(next_shape_id);
-
     attr_index_t removed_index = RSHAPE_INDEX(removed_shape_id);
     val = fields[removed_index];
-    size_t trailing_fields = new_fields_count - removed_index;
 
-    MEMMOVE(&fields[removed_index], &fields[removed_index + 1], VALUE, trailing_fields);
-
-    if (RB_TYPE_P(obj, T_OBJECT) &&
-        !RB_FL_TEST_RAW(obj, ROBJECT_EMBED) &&
-        rb_obj_embedded_size(new_fields_count) <= rb_gc_obj_slot_size(obj)) {
-        // Re-embed objects when instances become small enough
-        // This is necessary because YJIT assumes that objects with the same shape
-        // have the same embeddedness for efficiency (avoid extra checks)
-        RB_FL_SET_RAW(obj, ROBJECT_EMBED);
-        MEMCPY(ROBJECT_FIELDS(obj), fields, VALUE, new_fields_count);
-        xfree(fields);
+    attr_index_t new_fields_count = RSHAPE_LEN(next_shape_id);
+    if (new_fields_count) {
+        size_t trailing_fields = new_fields_count - removed_index;
+
+        MEMMOVE(&fields[removed_index], &fields[removed_index + 1], VALUE, trailing_fields);
+
+        if (RB_TYPE_P(obj, T_OBJECT) &&
+            !RB_FL_TEST_RAW(obj, ROBJECT_EMBED) &&
+            rb_obj_embedded_size(new_fields_count) <= rb_gc_obj_slot_size(obj)) {
+            // Re-embed objects when instances become small enough
+            // This is necessary because YJIT assumes that objects with the same shape
+            // have the same embeddedness for efficiency (avoid extra checks)
+            RB_FL_SET_RAW(obj, ROBJECT_EMBED);
+            MEMCPY(ROBJECT_FIELDS(obj), fields, VALUE, new_fields_count);
+            xfree(fields);
+        }
+    }
+    else {
+        if (FL_TEST_RAW(obj, FL_EXIVAR)) {
+            rb_free_generic_ivar(obj);
+        }
     }
     rb_obj_set_shape_id(obj, next_shape_id);
 
@@ -1881,8 +1889,8 @@ generic_ivar_set_set_shape_id(VALUE obj, shape_id_t shape_id, void *data)
 static shape_id_t
 generic_ivar_set_transition_too_complex(VALUE obj, void *_data)
 {
-    shape_id_t new_shape_id = rb_evict_fields_to_hash(obj);
     FL_SET_RAW(obj, FL_EXIVAR);
+    shape_id_t new_shape_id = rb_evict_fields_to_hash(obj);
     return new_shape_id;
 }
 
@@ -2407,9 +2415,9 @@ rb_copy_generic_ivar(VALUE dest, VALUE obj)
 
   clear:
     if (FL_TEST(dest, FL_EXIVAR)) {
-        RBASIC_SET_SHAPE_ID(dest, ROOT_SHAPE_ID);
         rb_free_generic_ivar(dest);
         FL_UNSET(dest, FL_EXIVAR);
+        RBASIC_SET_SHAPE_ID(dest, ROOT_SHAPE_ID);
     }
 }