[mypyc] Make type objects immortal if using free threading (#19538)

If they are not immortal, concurrent construction of objects by multiple
threads can cause serious contention due to reference count updates.

Making them immortal is similar to how both user-defined normal Python 
classes and built-in types in free-threaded builds are immortal.

Fix the issue for both native and non-native classes. Dataclasses still
have contention, and they may be harder to fix (this may require a fix 
in CPython).

This speeds up a few micro-benchmarks that construct instances of
classes in multiple threads by a big factor (5x+).

Author: JukkaL

mypyc/irbuild/builder.py

@@ -424,6 +424,10 @@ def new_tuple(self, items: list[Value], line: int) -> Value:
     def debug_print(self, toprint: str | Value) -> None:
         return self.builder.debug_print(toprint)
 
+    def set_immortal_if_free_threaded(self, v: Value, line: int) -> None:
+        """Make an object immortal on free-threaded builds (to avoid contention)."""
+        self.builder.set_immortal_if_free_threaded(v, line)
+
     # Helpers for IR building
 
     def add_to_non_ext_dict(
@@ -433,6 +437,10 @@ def add_to_non_ext_dict(
         key_unicode = self.load_str(key)
         self.primitive_op(dict_set_item_op, [non_ext.dict, key_unicode, val], line)
 
+        # It's important that accessing class dictionary items from multiple threads
+        # doesn't cause contention.
+        self.builder.set_immortal_if_free_threaded(val, line)
+
     def gen_import(self, id: str, line: int) -> None:
         self.imports[id] = None
 

mypyc/irbuild/classdef.py

@@ -262,6 +262,9 @@ def finalize(self, ir: ClassIR) -> None:
         non_ext_class = load_non_ext_class(self.builder, ir, self.non_ext, self.cdef.line)
         non_ext_class = load_decorated_class(self.builder, self.cdef, non_ext_class)
 
+        # Try to avoid contention when using free threading.
+        self.builder.set_immortal_if_free_threaded(non_ext_class, self.cdef.line)
+
         # Save the decorated class
         self.builder.add(
             InitStatic(non_ext_class, self.cdef.name, self.builder.module_name, NAMESPACE_TYPE)
@@ -449,6 +452,11 @@ def allocate_class(builder: IRBuilder, cdef: ClassDef) -> Value:
     )
     # Create the class
     tp = builder.call_c(pytype_from_template_op, [template, tp_bases, modname], cdef.line)
+
+    # Set type object to be immortal if free threaded, as otherwise reference count contention
+    # can cause a big performance hit.
+    builder.set_immortal_if_free_threaded(tp, cdef.line)
+
     # Immediately fix up the trait vtables, before doing anything with the class.
     ir = builder.mapper.type_to_ir[cdef.info]
     if not ir.is_trait and not ir.builtin_base:

mypyc/irbuild/ll_builder.py

@@ -6,6 +6,7 @@
 
 from __future__ import annotations
 
+import sys
 from collections.abc import Sequence
 from typing import Callable, Final, Optional
 
@@ -16,6 +17,7 @@
 from mypyc.common import (
     BITMAP_BITS,
     FAST_ISINSTANCE_MAX_SUBCLASSES,
+    IS_FREE_THREADED,
     MAX_LITERAL_SHORT_INT,
     MAX_SHORT_INT,
     MIN_LITERAL_SHORT_INT,
@@ -164,6 +166,7 @@
     fast_isinstance_op,
     none_object_op,
     not_implemented_op,
+    set_immortal_op,
     var_object_size,
 )
 from mypyc.primitives.registry import (
@@ -2322,6 +2325,11 @@ def new_tuple_with_length(self, length: Value, line: int) -> Value:
     def int_to_float(self, n: Value, line: int) -> Value:
         return self.primitive_op(int_to_float_op, [n], line)
 
+    def set_immortal_if_free_threaded(self, v: Value, line: int) -> None:
+        """Make an object immortal on free-threaded builds (to avoid contention)."""
+        if IS_FREE_THREADED and sys.version_info >= (3, 14):
+            self.primitive_op(set_immortal_op, [v], line)
+
     # Internal helpers
 
     def decompose_union_helper(

mypyc/lib-rt/CPy.h

@@ -931,6 +931,10 @@ PyObject *CPy_GetANext(PyObject *aiter);
 void CPy_SetTypeAliasTypeComputeFunction(PyObject *alias, PyObject *compute_value);
 void CPyTrace_LogEvent(const char *location, const char *line, const char *op, const char *details);
 
+#if CPY_3_14_FEATURES
+void CPy_SetImmortal(PyObject *obj);
+#endif
+
 #ifdef __cplusplus
 }
 #endif

mypyc/lib-rt/misc_ops.c

@@ -1058,7 +1058,7 @@ void CPyTrace_LogEvent(const char *location, const char *line, const char *op, c
 
 #endif
 
-#ifdef CPY_3_12_FEATURES
+#if CPY_3_12_FEATURES
 
 // Copied from Python 3.12.3, since this struct is internal to CPython. It defines
 // the structure of typing.TypeAliasType objects. We need it since compute_value is
@@ -1088,3 +1088,13 @@ void CPy_SetTypeAliasTypeComputeFunction(PyObject *alias, PyObject *compute_valu
 }
 
 #endif
+
+#if CPY_3_14_FEATURES
+
+#include "internal/pycore_object.h"
+
+void CPy_SetImmortal(PyObject *obj) {
+    _Py_SetImmortal(obj);
+}
+
+#endif

mypyc/lib-rt/mypyc_util.h

@@ -139,8 +139,9 @@ static inline CPyTagged CPyTagged_ShortFromSsize_t(Py_ssize_t x) {
     return x << 1;
 }
 
-// Are we targeting Python 3.12 or newer?
+// Are we targeting Python 3.X or newer?
 #define CPY_3_12_FEATURES (PY_VERSION_HEX >= 0x030c0000)
+#define CPY_3_14_FEATURES (PY_VERSION_HEX >= 0x030e0000)
 
 #if CPY_3_12_FEATURES
 

mypyc/primitives/misc_ops.py

@@ -311,3 +311,18 @@
     return_type=void_rtype,
     error_kind=ERR_NEVER,
 )
+
+# Mark object as immortal -- it won't be freed via reference counting, as
+# the reference count won't be updated any longer. Immortal objects support
+# fast concurrent read-only access from multiple threads when using free
+# threading, since this eliminates contention from concurrent reference count
+# updates.
+#
+# Needs at least Python 3.14.
+set_immortal_op = custom_primitive_op(
+    name="set_immmortal",
+    c_function_name="CPy_SetImmortal",
+    arg_types=[object_rprimitive],
+    return_type=void_rtype,
+    error_kind=ERR_NEVER,
+)