[mypyc] feat: __mypyc_empty_tuple__ constant (#19654)

I realized that any time a user has a kwarg-only call expression like
`fn(abc=123, ...)` in their compiled code, and `func` is not a native
function, a new empty tuple is created every time

This is not really necessary, we can just hold the same empty tuple in
memory as a constant and pass it around. It's immutable, and that's
already what we're already doing, since `tuple() is tuple()` but our
current method involves more steps.

This should slightly improve the speed of kwarg-only python func
calling.

Author: BobTheBuidler

mypyc/codegen/emit.py

@@ -1036,17 +1036,21 @@ def emit_box(
             self.emit_line(f"{declaration}{dest} = PyFloat_FromDouble({src});")
         elif isinstance(typ, RTuple):
             self.declare_tuple_struct(typ)
-            self.emit_line(f"{declaration}{dest} = PyTuple_New({len(typ.types)});")
-            self.emit_line(f"if (unlikely({dest} == NULL))")
-            self.emit_line("    CPyError_OutOfMemory();")
-            # TODO: Fail if dest is None
-            for i in range(len(typ.types)):
-                if not typ.is_unboxed:
-                    self.emit_line(f"PyTuple_SET_ITEM({dest}, {i}, {src}.f{i}")
-                else:
-                    inner_name = self.temp_name()
-                    self.emit_box(f"{src}.f{i}", inner_name, typ.types[i], declare_dest=True)
-                    self.emit_line(f"PyTuple_SET_ITEM({dest}, {i}, {inner_name});")
+            if not typ.types:
+                self.emit_line(f"{declaration}{dest} = CPyTuple_LoadEmptyTupleConstant();")
+            else:
+                self.emit_line(f"{declaration}{dest} = PyTuple_New({len(typ.types)});")
+                self.emit_line(f"if (unlikely({dest} == NULL))")
+                self.emit_line("    CPyError_OutOfMemory();")
+
+                # TODO: Fail if dest is None
+                for i in range(len(typ.types)):
+                    if not typ.is_unboxed:
+                        self.emit_line(f"PyTuple_SET_ITEM({dest}, {i}, {src}.f{i}")
+                    else:
+                        inner_name = self.temp_name()
+                        self.emit_box(f"{src}.f{i}", inner_name, typ.types[i], declare_dest=True)
+                        self.emit_line(f"PyTuple_SET_ITEM({dest}, {i}, {inner_name});")
         else:
             assert not typ.is_unboxed
             # Type is boxed -- trivially just assign.

mypyc/irbuild/ll_builder.py

@@ -189,6 +189,7 @@
 )
 from mypyc.primitives.tuple_ops import (
     list_tuple_op,
+    load_empty_tuple_constant_op,
     new_tuple_op,
     new_tuple_with_length_op,
     sequence_tuple_op,
@@ -2362,8 +2363,11 @@ def builtin_len(self, val: Value, line: int, use_pyssize_t: bool = False) -> Val
             return self.call_c(generic_len_op, [val], line)
 
     def new_tuple(self, items: list[Value], line: int) -> Value:
-        size: Value = Integer(len(items), c_pyssize_t_rprimitive)
-        return self.call_c(new_tuple_op, [size] + items, line)
+        if items:
+            size: Value = Integer(len(items), c_pyssize_t_rprimitive)
+            return self.call_c(new_tuple_op, [size] + items, line)
+        else:
+            return self.call_c(load_empty_tuple_constant_op, [], line)
 
     def new_tuple_with_length(self, length: Value, line: int) -> Value:
         """This function returns an uninitialized tuple.

mypyc/lib-rt/CPy.h

@@ -64,6 +64,15 @@ typedef struct tuple_T4CIOO {
 } tuple_T4CIOO;
 #endif
 
+// System-wide empty tuple constant
+extern PyObject * __mypyc_empty_tuple__;
+
+static inline PyObject *CPyTuple_LoadEmptyTupleConstant() {
+#if !CPY_3_12_FEATURES
+    Py_INCREF(__mypyc_empty_tuple__);
+#endif
+    return __mypyc_empty_tuple__;
+}
 
 // Native object operations
 

mypyc/lib-rt/init.c

@@ -4,10 +4,21 @@
 struct ExcDummyStruct _CPy_ExcDummyStruct = { PyObject_HEAD_INIT(NULL) };
 PyObject *_CPy_ExcDummy = (PyObject *)&_CPy_ExcDummyStruct;
 
+// System-wide empty tuple constant
+PyObject * __mypyc_empty_tuple__ = NULL;
+
 // Because its dynamic linker is more restricted than linux/OS X,
 // Windows doesn't allow initializing globals with values from
 // other dynamic libraries. This means we need to initialize
 // things at load time.
 void CPy_Init(void) {
     _CPy_ExcDummyStruct.ob_base.ob_type = &PyBaseObject_Type;
+
+    // Initialize system-wide empty tuple constant
+    if (__mypyc_empty_tuple__ == NULL) {
+        __mypyc_empty_tuple__ = PyTuple_New(0);
+        if (!__mypyc_empty_tuple__) {
+            CPyError_OutOfMemory();
+        }
+    }
 }

mypyc/primitives/tuple_ops.py

@@ -55,6 +55,13 @@
     error_kind=ERR_MAGIC,
 )
 
+load_empty_tuple_constant_op = custom_op(
+    arg_types=[],
+    return_type=tuple_rprimitive,
+    c_function_name="CPyTuple_LoadEmptyTupleConstant",
+    error_kind=ERR_NEVER,
+)
+
 # PyTuple_SET_ITEM does no error checking,
 # and should only be used to fill in brand new tuples.
 new_tuple_set_item_op = custom_op(

mypyc/test-data/irbuild-basic.test

@@ -1750,7 +1750,7 @@ L0:
     r7 = __main__.globals :: static
     r8 = 'f'
     r9 = CPyDict_GetItem(r7, r8)
-    r10 = PyTuple_Pack(0)
+    r10 = CPyTuple_LoadEmptyTupleConstant()
     r11 = PyDict_Copy(r6)
     r12 = PyObject_Call(r9, r10, r11)
     r13 = unbox(tuple[int, int, int], r12)

mypyc/test-data/irbuild-classes.test

@@ -297,7 +297,7 @@ L2:
     r27 = CPyType_FromTemplate(r26, r24, r25)
     r28 = C_trait_vtable_setup()
     r29 = '__mypyc_attrs__'
-    r30 = PyTuple_Pack(0)
+    r30 = CPyTuple_LoadEmptyTupleConstant()
     r31 = PyObject_SetAttr(r27, r29, r30)
     r32 = r31 >= 0 :: signed
     __main__.C = r27 :: type
@@ -310,7 +310,7 @@ L2:
     r39 = __main__.S_template :: type
     r40 = CPyType_FromTemplate(r39, r37, r38)
     r41 = '__mypyc_attrs__'
-    r42 = PyTuple_Pack(0)
+    r42 = CPyTuple_LoadEmptyTupleConstant()
     r43 = PyObject_SetAttr(r40, r41, r42)
     r44 = r43 >= 0 :: signed
     __main__.S = r40 :: type