Simple closure defnode call inlining

Cython/Compiler/ExprNodes.py

@@ -3864,6 +3864,106 @@ def generate_result_code(self, code):
                 code.funcstate.release_temp(self.opt_arg_struct)
 
 
+class InlinedDefNodeCallNode(CallNode):
+    #  Inline call to defnode
+    #
+    #  function       PyCFunctionNode
+    #  function_name  NameNode
+    #  args           [ExprNode]
+
+    subexprs = ['args', 'function_name']
+    is_temp = 1
+    type = py_object_type
+    function = None
+    function_name = None
+
+    def can_be_inlined(self):
+        func_type= self.function.def_node
+        if func_type.star_arg or func_type.starstar_arg:
+            return False
+        if len(func_type.args) != len(self.args):
+            return False
+        return True
+
+    def analyse_types(self, env):
+        self.function_name.analyse_types(env)
+
+        for arg in self.args:
+            arg.analyse_types(env)
+
+        func_type = self.function.def_node
+        actual_nargs = len(self.args)
+
+        # Coerce arguments
+        some_args_in_temps = False
+        for i in xrange(actual_nargs):
+            formal_type = func_type.args[i].type
+            arg = self.args[i].coerce_to(formal_type, env)
+            if arg.is_temp:
+                if i > 0:
+                    # first argument in temp doesn't impact subsequent arguments
+                    some_args_in_temps = True
+            elif arg.type.is_pyobject and not env.nogil:
+                if arg.nonlocally_immutable():
+                    # plain local variables are ok
+                    pass
+                else:
+                    # we do not safely own the argument's reference,
+                    # but we must make sure it cannot be collected
+                    # before we return from the function, so we create
+                    # an owned temp reference to it
+                    if i > 0: # first argument doesn't matter
+                        some_args_in_temps = True
+                    arg = arg.coerce_to_temp(env)
+            self.args[i] = arg
+
+        if some_args_in_temps:
+            # if some args are temps and others are not, they may get
+            # constructed in the wrong order (temps first) => make
+            # sure they are either all temps or all not temps (except
+            # for the last argument, which is evaluated last in any
+            # case)
+            for i in xrange(actual_nargs-1):
+                arg = self.args[i]
+                if arg.nonlocally_immutable():
+                    # locals, C functions, unassignable types are safe.
+                    pass
+                elif arg.type.is_cpp_class:
+                    # Assignment has side effects, avoid.
+                    pass
+                elif env.nogil and arg.type.is_pyobject:
+                    # can't copy a Python reference into a temp in nogil
+                    # env (this is safe: a construction would fail in
+                    # nogil anyway)
+                    pass
+                else:
+                    #self.args[i] = arg.coerce_to_temp(env)
+                    # instead: issue a warning
+                    if i > 0:
+                        warning(arg.pos, "Argument evaluation order in C function call is undefined and may not be as expected", 0)
+                        break
+
+    def generate_result_code(self, code):
+        arg_code = [self.function_name.py_result()]
+        func_type = self.function.def_node
+        for arg, proto_arg in zip(self.args, func_type.args):
+            if arg.type.is_pyobject:
+                if proto_arg.hdr_type:
+                    arg_code.append(arg.result_as(proto_arg.hdr_type))
+                else:
+                    arg_code.append(arg.result_as(proto_arg.type))
+            else:
+                arg_code.append(arg.result())
+        arg_code = ', '.join(arg_code)
+        code.putln(
+            "%s = %s(%s); %s" % (
+                self.result(),
+                self.function.def_node.entry.pyfunc_cname,
+                arg_code,
+                code.error_goto_if_null(self.result(), self.pos)))
+        code.put_gotref(self.py_result())
+
+
 class PythonCapiFunctionNode(ExprNode):
     subexprs = []
     def __init__(self, pos, py_name, cname, func_type, utility_code = None):

Cython/Compiler/Optimize.py

@@ -1643,6 +1643,28 @@ def _handle_general_function_dict(self, node, pos_args, kwargs):
             return node
         return kwargs
 
+class InlineDefNodeCalls(Visitor.CythonTransform):
+    visit_Node = Visitor.VisitorTransform.recurse_to_children
+
+    def visit_SimpleCallNode(self, node):
+        self.visitchildren(node)
+        if not self.current_directives.get('optimize.inline_defnode_calls'):
+            return node
+        function_name = node.function
+        if not function_name.is_name:
+            return node
+        if len(function_name.cf_state) != 1:
+            return node
+        function = list(function_name.cf_state)[0].rhs
+        if not isinstance(function, ExprNodes.PyCFunctionNode):
+            return node
+        inlined = ExprNodes.InlinedDefNodeCallNode(
+            node.pos, function_name=function_name,
+            function=function, args=node.args)
+        if inlined.can_be_inlined():
+            return inlined
+        return node
+
 
 class OptimizeBuiltinCalls(Visitor.EnvTransform):
     """Optimize some common methods calls and instantiation patterns

Cython/Compiler/Options.py

@@ -106,6 +106,9 @@
     'warn.unused_arg': False,
     'warn.unused_result': False,
 
+# optimizations
+    'optimize.inline_defnode_calls': False,
+
 # remove unreachable code
     'remove_unreachable': True,
 

Cython/Compiler/Pipeline.py

@@ -137,6 +137,7 @@ def create_pipeline(context, mode, exclude_classes=()):
     from AutoDocTransforms import EmbedSignature
     from Optimize import FlattenInListTransform, SwitchTransform, IterationTransform
     from Optimize import EarlyReplaceBuiltinCalls, OptimizeBuiltinCalls
+    from Optimize import InlineDefNodeCalls
     from Optimize import ConstantFolding, FinalOptimizePhase
     from Optimize import DropRefcountingTransform
     from Buffer import IntroduceBufferAuxiliaryVars
@@ -185,6 +186,7 @@ def create_pipeline(context, mode, exclude_classes=()):
         MarkOverflowingArithmetic(context),
         IntroduceBufferAuxiliaryVars(context),
         _check_c_declarations,
+        InlineDefNodeCalls(context),
         AnalyseExpressionsTransform(context),
         FindInvalidUseOfFusedTypes(context),
         CreateClosureClasses(context),  ## After all lookups and type inference

tests/run/closure_inlining.pyx

@@ -0,0 +1,77 @@
+# cython: optimize.inline_defnode_calls=True
+# mode: run
+cimport cython
+
+@cython.test_fail_if_path_exists('//SimpleCallNode')
+@cython.test_assert_path_exists('//InlinedDefNodeCallNode')
+def simple_noargs():
+    """
+    >>> simple_noargs()
+    123
+    """
+    def inner():
+        return 123
+    return inner()
+
+
+@cython.test_fail_if_path_exists('//SimpleCallNode')
+@cython.test_assert_path_exists('//InlinedDefNodeCallNode')
+def test_coerce(a, int b):
+    """
+    >>> test_coerce(2, 2)
+    4
+    """
+    def inner(int a, b):
+        return a * b
+    return inner(a, b)
+
+
+cdef class Foo(object):
+    def __repr__(self):
+        return '<Foo>'
+
+
+@cython.test_fail_if_path_exists('//SimpleCallNode')
+@cython.test_assert_path_exists('//InlinedDefNodeCallNode')
+def test_func_signature(a):
+    """
+    >>> test_func_signature(Foo())
+    <Foo>
+    """
+
+    def inner(Foo a):
+        return a
+    return inner(a)
+
+@cython.test_fail_if_path_exists('//SimpleCallNode')
+@cython.test_assert_path_exists('//InlinedDefNodeCallNode')
+def test_func_signature2(a, b):
+    """
+    >>> test_func_signature2(Foo(), 123)
+    (<Foo>, 123)
+    """
+
+    def inner(Foo a, b):
+        return a, b
+    return inner(a, b)
+
+# Starred args and default values are not yet supported for inlining
+@cython.test_assert_path_exists('//SimpleCallNode')
+def test_defaults(a, b):
+    """
+    >>> test_defaults(1, 2)
+    (1, 2, 123)
+    """
+    def inner(a, b=b, c=123):
+        return a, b, c
+    return inner(a)
+
+@cython.test_assert_path_exists('//SimpleCallNode')
+def test_starred(a):
+    """
+    >>> test_starred(123)
+    (123, (), {})
+    """
+    def inner(a, *args, **kwargs):
+        return a, args, kwargs
+    return inner(a)