diff --git a/llvm_cbuilder/builder.py b/llvm_cbuilder/builder.py
index 661c2b1..3fce585 100644
--- a/llvm_cbuilder/builder.py
+++ b/llvm_cbuilder/builder.py
@@ -149,6 +149,13 @@ class CBuilder(object):
         func = mod.add_function(functype, name=name)
         return CBuilder(func)
 
+    def depends(self, cdef, **kwargs):
+        try:
+            func = cdef.define(self.function.module, **kwargs)
+        except NameError as e:
+            (func,) = e
+        return CFunc(self, func)
+
     def printf(self, fmt, *args):
         mod = self.function.module
         int_t = lc.Type.int()
@@ -429,7 +436,7 @@ class CDefinition(CBuilder):
         functype = lc.Type.function(cls._retty_, [v for k, v in cls._argtys_])
         func = module.get_or_insert_function(functype, name=cls._name_)
         if not func.is_declaration: # already defined?
-            raise NameError('Function %s has already been defined' % cls._name_)
+            raise NameError(func)
 
         # Name all arguments
         for i, (name, _) in enumerate(cls._argtys_):
diff --git a/parallel_vectorize.py b/parallel_vectorize.py
index 25176b8..42aea26 100644
--- a/parallel_vectorize.py
+++ b/parallel_vectorize.py
@@ -87,8 +87,15 @@ class Context(CStruct):
     ]
 
 class ParallelUFunc(CDefinition):
+    '''
+    Platform dependent threading function is implemented in
+
+    def _dispatch_worker(self, worker, contexts, num_thread):
+        ...
+
+    which should be implemented in subclass or mixin.
+    '''
     _name_   = 'parallel_ufunc'
-    _retty_  = C.void
     _argtys_ = [
         ('func',       C.void_p),
         ('worker',     C.void_p),
@@ -182,13 +189,7 @@ class ParallelUFunc(CDefinition):
             cur_ctxt.completed.assign(
                                     self.constant_null(cur_ctxt.completed.type))
 
-    def _dispatch_worker(self, worker, contexts, num_thread):
-        '''
-        Dispatch worker threads.
-        '''
-        raise NotImplementedError
-
-class ParallelUFuncPosix(ParallelUFunc):
+class ParallelUFuncPosixMixin(object):
     '''
     Implements _dispatch_worker to use pthread.
     '''
@@ -316,6 +317,16 @@ class UFuncCore(CDefinition):
         '''
         raise NotImplementedError
 
+class _SpecializedParallelUFunc(ParallelUFunc, ParallelUFuncPosixMixin):
+    _argtys_ = [
+        ('args',       C.pointer(C.char_p)),
+        ('dimensions', C.pointer(C.intp)),
+        ('steps',      C.pointer(C.intp)),
+        ('data',       C.void_p),
+    ]
+
+    def body(self, args, dimensions, steps, data, ThreadCount=1):
+        pass
 
 class PThreadAPI(CExternal):
     pthread_t = C.void_p
diff --git a/test_parallel_vectorize.py b/test_parallel_vectorize.py
index fd7cc52..4fde372 100644
--- a/test_parallel_vectorize.py
+++ b/test_parallel_vectorize.py
@@ -30,6 +30,8 @@ class UFuncCore_D_D(UFuncCore):
         res = ufunc_ptr(indata.load())
         outdata.store(res)
 
+class ParallelUFuncPosix(ParallelUFunc, ParallelUFuncPosixMixin):
+    pass
 
 class Tester(CDefinition):
     '''
@@ -45,10 +47,10 @@ class Tester(CDefinition):
         ArgCount = 2
         WorkCount = 10000
 
-        parallel_ufunc = CFunc(self, ParallelUFuncPosix.define(module,
-                               ThreadCount=ThreadCount))
-        ufunc_core = CFunc(self, UFuncCore_D_D.define(module))
-        worker = CFunc(self, Work_D_D.define(module))
+        parallel_ufunc = self.depends(ParallelUFuncPosix,
+                                      ThreadCount=ThreadCount)
+        ufunc_core = self.depends(UFuncCore_D_D)
+        worker = self.depends(Work_D_D)
 
         # real work
         NULL = self.constant_null(C.void_p)