[mypyc] feat: PyObject_CallObject op for fn(*args) fastpath (#19631)

BobTheBuidler · web-flow · commit d1dffe275b9e · 2025-08-26T18:36:23.000+01:00
This PR adds a new custom_op for PyObject_CallObject which is more efficient than PyObject_Call in cases where there are no kwargs. posarg-only use cases are already optimized but this is helpful for patterns such as `fn(*args)` or `fn(a1, a2, *args)` This PR extends #19623 and #19629 , as this change will not be helpful until those PRs are merged.
diff --git a/mypyc/annotate.py b/mypyc/annotate.py
@@ -77,6 +77,7 @@ def __init__(self, message: str, priority: int = 1) -> None:
     "PyNumber_Rshift": Annotation('Generic ">>" operation.'),
     "PyNumber_Invert": Annotation('Generic "~" operation.'),
     "PyObject_Call": Annotation("Generic call operation."),
+    "PyObject_CallObject": Annotation("Generic call operation."),
     "PyObject_RichCompare": Annotation("Generic comparison operation."),
     "PyObject_GetItem": Annotation("Generic indexing operation."),
     "PyObject_SetItem": Annotation("Generic indexed assignment."),
diff --git a/mypyc/irbuild/ll_builder.py b/mypyc/irbuild/ll_builder.py
@@ -141,6 +141,7 @@
     generic_ssize_t_len_op,
     py_call_op,
     py_call_with_kwargs_op,
+    py_call_with_posargs_op,
     py_getattr_op,
     py_method_call_op,
     py_vectorcall_method_op,
@@ -805,7 +806,7 @@ def _construct_varargs(
                             value.type, RTuple
                         ):
                             value = self.primitive_op(sequence_tuple_op, [value], line)
-                        return value, self._create_dict([], [], line)
+                        return value, None
                     elif len(args) == 2 and args[1][1] == ARG_STAR2:
                         # fn(*args, **kwargs)
                         # TODO: extend to cover(*args, **k, **w, **a, **r, **g, **s)
@@ -938,7 +939,7 @@ def _construct_varargs(
             elif not is_tuple_rprimitive(star_result.type):
                 # if star_result is a tuple we took the fast path
                 star_result = self.primitive_op(list_tuple_op, [star_result], line)
-        if has_star2 and star2_result is None:
+        if has_star2 and star2_result is None and len(star2_keys) > 0:
             # TODO: use dict_copy_op for simple cases of **kwargs
             star2_result = self._create_dict(star2_keys, star2_values, line)
 
@@ -964,13 +965,16 @@ def py_call(
         if arg_kinds is None or all(kind == ARG_POS for kind in arg_kinds):
             return self.call_c(py_call_op, [function] + arg_values, line)
 
-        # Otherwise fallback to py_call_with_kwargs_op.
+        # Otherwise fallback to py_call_with_posargs_op or py_call_with_kwargs_op.
         assert arg_names is not None
 
         pos_args_tuple, kw_args_dict = self._construct_varargs(
             list(zip(arg_values, arg_kinds, arg_names)), line, has_star=True, has_star2=True
         )
-        assert pos_args_tuple and kw_args_dict
+        assert pos_args_tuple
+
+        if kw_args_dict is None:
+            return self.call_c(py_call_with_posargs_op, [function, pos_args_tuple], line)
 
         return self.call_c(py_call_with_kwargs_op, [function, pos_args_tuple, kw_args_dict], line)
 
@@ -1169,8 +1173,7 @@ def native_args_to_positional(
                 assert star_arg
                 output_arg = star_arg
             elif arg.kind == ARG_STAR2:
-                assert star2_arg
-                output_arg = star2_arg
+                output_arg = star2_arg or self._create_dict([], [], line)
             elif not lst:
                 if is_fixed_width_rtype(arg.type):
                     output_arg = Integer(0, arg.type)
diff --git a/mypyc/primitives/generic_ops.py b/mypyc/primitives/generic_ops.py
@@ -308,6 +308,15 @@
     error_kind=ERR_MAGIC,
 )
 
+# Call callable object with positional args only: func(*args)
+# Arguments are (func, *args tuple).
+py_call_with_posargs_op = custom_op(
+    arg_types=[object_rprimitive, object_rprimitive],
+    return_type=object_rprimitive,
+    c_function_name="PyObject_CallObject",
+    error_kind=ERR_MAGIC,
+)
+
 # Call method with positional arguments: obj.method(arg1, ...)
 # Arguments are (object, attribute name, arg1, ...).
 py_method_call_op = custom_op(
diff --git a/mypyc/test-data/irbuild-basic.test b/mypyc/test-data/irbuild-basic.test
@@ -1674,20 +1674,17 @@ def g():
     r0 :: tuple[int, int, int]
     r1 :: dict
     r2 :: str
-    r3 :: object
-    r4 :: dict
-    r5, r6 :: object
-    r7 :: tuple[int, int, int]
+    r3, r4, r5 :: object
+    r6 :: tuple[int, int, int]
 L0:
     r0 = (2, 4, 6)
     r1 = __main__.globals :: static
     r2 = 'f'
     r3 = CPyDict_GetItem(r1, r2)
-    r4 = PyDict_New()
-    r5 = box(tuple[int, int, int], r0)
-    r6 = PyObject_Call(r3, r5, r4)
-    r7 = unbox(tuple[int, int, int], r6)
-    return r7
+    r4 = box(tuple[int, int, int], r0)
+    r5 = PyObject_CallObject(r3, r4)
+    r6 = unbox(tuple[int, int, int], r5)
+    return r6
 def h():
     r0 :: tuple[int, int]
     r1 :: dict
@@ -1698,9 +1695,8 @@ def h():
     r6 :: ptr
     r7, r8 :: object
     r9 :: tuple
-    r10 :: dict
-    r11 :: object
-    r12 :: tuple[int, int, int]
+    r10 :: object
+    r11 :: tuple[int, int, int]
 L0:
     r0 = (4, 6)
     r1 = __main__.globals :: static
@@ -1714,10 +1710,9 @@ L0:
     r7 = box(tuple[int, int], r0)
     r8 = CPyList_Extend(r4, r7)
     r9 = PyList_AsTuple(r4)
-    r10 = PyDict_New()
-    r11 = PyObject_Call(r3, r9, r10)
-    r12 = unbox(tuple[int, int, int], r11)
-    return r12
+    r10 = PyObject_CallObject(r3, r9)
+    r11 = unbox(tuple[int, int, int], r10)
+    return r11
 
 [case testStar2Args]
 from typing import Tuple
@@ -3562,15 +3557,12 @@ def wrapper_deco_obj.__call__(__mypyc_self__, args):
     __mypyc_self__ :: __main__.wrapper_deco_obj
     args :: tuple
     r0 :: __main__.deco_env
-    r1 :: object
-    r2 :: dict
-    r3 :: object
+    r1, r2 :: object
 L0:
     r0 = __mypyc_self__.__mypyc_env__
     r1 = r0.fn
-    r2 = PyDict_New()
-    r3 = PyObject_Call(r1, args, r2)
-    return r3
+    r2 = PyObject_CallObject(r1, args)
+    return r2
 def deco(fn):
     fn :: object
     r0 :: __main__.deco_env
@@ -3613,15 +3605,13 @@ def wrapper_deco_obj.__call__(__mypyc_self__, args):
     r0 :: __main__.deco_env
     r1 :: object
     r2 :: tuple
-    r3 :: dict
-    r4 :: object
+    r3 :: object
 L0:
     r0 = __mypyc_self__.__mypyc_env__
     r1 = r0.fn
     r2 = PyList_AsTuple(args)
-    r3 = PyDict_New()
-    r4 = PyObject_Call(r1, r2, r3)
-    return r4
+    r3 = PyObject_CallObject(r1, r2)
+    return r3
 def deco(fn):
     fn :: object
     r0 :: __main__.deco_env
@@ -3716,15 +3706,13 @@ def wrapper_deco_obj.__call__(__mypyc_self__, args):
     r0 :: __main__.deco_env
     r1 :: object
     r2 :: tuple
-    r3 :: dict
-    r4 :: object
+    r3 :: object
 L0:
     r0 = __mypyc_self__.__mypyc_env__
     r1 = r0.fn
     r2 = PySequence_Tuple(args)
-    r3 = PyDict_New()
-    r4 = PyObject_Call(r1, r2, r3)
-    return r4
+    r3 = PyObject_CallObject(r1, r2)
+    return r3
 def deco(fn):
     fn :: object
     r0 :: __main__.deco_env
