diff --git a/jedi/common/__init__.py b/jedi/common/__init__.py
new file mode 100644
index 00000000..ba2e2f44
--- /dev/null
+++ b/jedi/common/__init__.py
@@ -0,0 +1,2 @@
+from jedi.common.context import ContextSet, NO_CONTEXTS, \
+    iterator_to_context_set
diff --git a/jedi/common/context.py b/jedi/common/context.py
new file mode 100644
index 00000000..9a1f476c
--- /dev/null
+++ b/jedi/common/context.py
@@ -0,0 +1,58 @@
+class ContextSet(object):
+    def __init__(self, *args):
+        self._set = set(args)
+
+    @classmethod
+    def from_iterable(cls, iterable):
+        return cls.from_set(set(iterable))
+
+    @classmethod
+    def from_set(cls, set_):
+        self = cls()
+        self._set = set_
+        return self
+
+    @classmethod
+    def from_sets(cls, sets):
+        """
+        Used to work with an iterable of set.
+        """
+        aggregated = set()
+        for set_ in sets:
+            print(set_)
+            if isinstance(set_, ContextSet):
+                aggregated |= set_._set
+            else:
+                aggregated |= set_
+        return cls.from_set(aggregated)
+
+    def __or__(self, other):
+        return ContextSet.from_set(self._set | other._set)
+
+    def __iter__(self):
+        for element in self._set:
+            yield element
+
+    def __repr__(self):
+        return '%s(%s)' % (self.__class__.__name__, ', '.join(str(s) for s in self._set))
+
+    def filter(self, filter_func):
+        return ContextSet(filter(filter_func, self._set))
+
+    def __getattr__(self, name):
+        def mapper(*args, **kwargs):
+            return ContextSet.from_sets(
+                getattr(context, name)(*args, **kwargs)
+                for context in self._set
+            )
+        return mapper
+
+
+def iterator_to_context_set(func):
+    def wrapper(*args, **kwargs):
+        return ContextSet.from_iterable(func(*args, **kwargs))
+
+    return wrapper
+
+
+NO_CONTEXTS = ContextSet()
diff --git a/jedi/evaluate/__init__.py b/jedi/evaluate/__init__.py
index 346f3528..7cff98d4 100644
--- a/jedi/evaluate/__init__.py
+++ b/jedi/evaluate/__init__.py
@@ -83,6 +83,7 @@ from jedi.evaluate import pep0484
 from jedi.evaluate.filters import TreeNameDefinition, ParamName
 from jedi.evaluate.instance import AnonymousInstance, BoundMethod
 from jedi.evaluate.context import ContextualizedName, ContextualizedNode
+from jedi.common import ContextSet, NO_CONTEXTS
 from jedi import parser_utils
 
 
@@ -101,7 +102,7 @@ def _limit_context_infers(func):
             evaluator.inferred_element_counts[n] += 1
             if evaluator.inferred_element_counts[n] > 300:
                 debug.warning('In context %s there were too many inferences.', n)
-                return set()
+                return NO_CONTEXTS
         except KeyError:
             evaluator.inferred_element_counts[n] = 1
         return func(evaluator, context, *args, **kwargs)
@@ -163,7 +164,7 @@ class Evaluator(object):
         with recursion.execution_allowed(self, stmt) as allowed:
             if allowed or context.get_root_context() == self.BUILTINS:
                 return self._eval_stmt(context, stmt, seek_name)
-        return set()
+        return NO_CONTEXTS
 
     #@evaluator_function_cache(default=[])
     @debug.increase_indent
@@ -178,11 +179,11 @@ class Evaluator(object):
         """
         debug.dbg('eval_statement %s (%s)', stmt, seek_name)
         rhs = stmt.get_rhs()
-        types = self.eval_element(context, rhs)
+        context_set = self.eval_element(context, rhs)
 
         if seek_name:
             c_node = ContextualizedName(context, seek_name)
-            types = finder.check_tuple_assignments(self, c_node, types)
+            context_set = finder.check_tuple_assignments(self, c_node, context_set)
 
         first_operator = next(stmt.yield_operators(), None)
         if first_operator not in ('=', None) and first_operator.type == 'operator':
@@ -194,7 +195,7 @@ class Evaluator(object):
                 name, position=stmt.start_pos, search_global=True)
 
             for_stmt = tree.search_ancestor(stmt, 'for_stmt')
-            if for_stmt is not None and for_stmt.type == 'for_stmt' and types \
+            if for_stmt is not None and for_stmt.type == 'for_stmt' and context_set \
                     and parser_utils.for_stmt_defines_one_name(for_stmt):
                 # Iterate through result and add the values, that's possible
                 # only in for loops without clutter, because they are
@@ -208,11 +209,11 @@ class Evaluator(object):
                     with helpers.predefine_names(context, for_stmt, dct):
                         t = self.eval_element(context, rhs)
                         left = precedence.calculate(self, context, left, operator, t)
-                types = left
+                context_set = left
             else:
-                types = precedence.calculate(self, context, left, operator, types)
-        debug.dbg('eval_statement result %s', types)
-        return types
+                context_set = precedence.calculate(self, context, left, operator, context_set)
+        debug.dbg('eval_statement result %s', context_set)
+        return context_set
 
     def eval_element(self, context, element):
         if isinstance(context, iterable.CompForContext):
@@ -261,14 +262,14 @@ class Evaluator(object):
                                 new_name_dicts = list(original_name_dicts)
                                 for i, name_dict in enumerate(new_name_dicts):
                                     new_name_dicts[i] = name_dict.copy()
-                                    new_name_dicts[i][if_name.value] = set([definition])
+                                    new_name_dicts[i][if_name.value] = ContextSet(definition)
 
                                 name_dicts += new_name_dicts
                         else:
                             for name_dict in name_dicts:
                                 name_dict[if_name.value] = definitions
             if len(name_dicts) > 1:
-                result = set()
+                result = ContextSet()
                 for name_dict in name_dicts:
                     with helpers.predefine_names(context, if_stmt, name_dict):
                         result |= self._eval_element_not_cached(context, element)
@@ -293,7 +294,7 @@ class Evaluator(object):
                 return self._eval_element_not_cached(context, element)
         return self._eval_element_cached(context, element)
 
-    @evaluator_function_cache(default=set())
+    @evaluator_function_cache(default=NO_CONTEXTS)
     def _eval_element_cached(self, context, element):
         return self._eval_element_not_cached(context, element)
 
@@ -301,63 +302,66 @@ class Evaluator(object):
     @_limit_context_infers
     def _eval_element_not_cached(self, context, element):
         debug.dbg('eval_element %s@%s', element, element.start_pos)
-        types = set()
         typ = element.type
         if typ in ('name', 'number', 'string', 'atom'):
-            types = self.eval_atom(context, element)
+            return self.eval_atom(context, element)
         elif typ == 'keyword':
             # For False/True/None
             if element.value in ('False', 'True', 'None'):
-                types.add(compiled.builtin_from_name(self, element.value))
+                return ContextSet(compiled.builtin_from_name(self, element.value))
             # else: print e.g. could be evaluated like this in Python 2.7
+            return NO_CONTEXTS
         elif typ == 'lambdef':
-            types = set([er.FunctionContext(self, context, element)])
+            return ContextSet(er.FunctionContext(self, context, element))
         elif typ == 'expr_stmt':
-            types = self.eval_statement(context, element)
+            return self.eval_statement(context, element)
         elif typ in ('power', 'atom_expr'):
             first_child = element.children[0]
             if not (first_child.type == 'keyword' and first_child.value == 'await'):
-                types = self.eval_atom(context, first_child)
+                context_set = self.eval_atom(context, first_child)
                 for trailer in element.children[1:]:
                     if trailer == '**':  # has a power operation.
                         right = self.eval_element(context, element.children[2])
-                        types = set(precedence.calculate(self, context, types, trailer, right))
+                        context_set = precedence.calculate(
+                            self,
+                            context,
+                            context_set,
+                            trailer,
+                            right
+                        )
                         break
-                    types = self.eval_trailer(context, types, trailer)
+                    context_set = self.eval_trailer(context, context_set, trailer)
+            return context_set
         elif typ in ('testlist_star_expr', 'testlist',):
             # The implicit tuple in statements.
-            types = set([iterable.SequenceLiteralContext(self, context, element)])
+            return ContextSet(iterable.SequenceLiteralContext(self, context, element))
         elif typ in ('not_test', 'factor'):
-            types = self.eval_element(context, element.children[-1])
+            context_set = self.eval_element(context, element.children[-1])
             for operator in element.children[:-1]:
-                types = set(precedence.factor_calculate(self, types, operator))
+                context_set = precedence.factor_calculate(self, context_set, operator)
+            return context_set
         elif typ == 'test':
             # `x if foo else y` case.
-            types = (self.eval_element(context, element.children[0]) |
-                     self.eval_element(context, element.children[-1]))
+            return (self.eval_element(context, element.children[0]) |
+                    self.eval_element(context, element.children[-1]))
         elif typ == 'operator':
             # Must be an ellipsis, other operators are not evaluated.
             # In Python 2 ellipsis is coded as three single dot tokens, not
             # as one token 3 dot token.
             assert element.value in ('.', '...')
-            types = set([compiled.create(self, Ellipsis)])
+            return ContextSet(compiled.create(self, Ellipsis))
         elif typ == 'dotted_name':
-            types = self.eval_atom(context, element.children[0])
+            context_set = self.eval_atom(context, element.children[0])
             for next_name in element.children[2::2]:
                 # TODO add search_global=True?
-                types = unite(
-                    typ.py__getattribute__(next_name, name_context=context)
-                    for typ in types
-                )
-            types = types
+                context_set.py__getattribute__(next_name, name_context=context)
+            return context_set
         elif typ == 'eval_input':
-            types = self._eval_element_not_cached(context, element.children[0])
+            return self._eval_element_not_cached(context, element.children[0])
         elif typ == 'annassign':
-            types = pep0484._evaluate_for_annotation(context, element.children[1])
+            return pep0484._evaluate_for_annotation(context, element.children[1])
         else:
-            types = precedence.calculate_children(self, context, element.children)
-        debug.dbg('eval_element result %s', types)
-        return types
+            return precedence.calculate_children(self, context, element.children)
 
     def eval_atom(self, context, atom):
         """
@@ -377,18 +381,19 @@ class Evaluator(object):
                 position=stmt.start_pos,
                 search_global=True
             )
+
         elif isinstance(atom, tree.Literal):
             string = parser_utils.safe_literal_eval(atom.value)
-            return set([compiled.create(self, string)])
+            return ContextSet(compiled.create(self, string))
         else:
             c = atom.children
             if c[0].type == 'string':
                 # Will be one string.
-                types = self.eval_atom(context, c[0])
+                context_set = self.eval_atom(context, c[0])
                 for string in c[1:]:
                     right = self.eval_atom(context, string)
-                    types = precedence.calculate(self, context, types, '+', right)
-                return types
+                    context_set = precedence.calculate(self, context, context_set, '+', right)
+                return context_set
             # Parentheses without commas are not tuples.
             elif c[0] == '(' and not len(c) == 2 \
                     and not(c[1].type == 'testlist_comp' and
@@ -408,7 +413,7 @@ class Evaluator(object):
                         pass
 
                 if comp_for.type == 'comp_for':
-                    return set([iterable.Comprehension.from_atom(self, context, atom)])
+                    return ContextSet(iterable.Comprehension.from_atom(self, context, atom))
 
             # It's a dict/list/tuple literal.
             array_node = c[1]
@@ -420,28 +425,28 @@ class Evaluator(object):
                 context = iterable.DictLiteralContext(self, context, atom)
             else:
                 context = iterable.SequenceLiteralContext(self, context, atom)
-            return set([context])
+            return ContextSet(context)
 
     def eval_trailer(self, context, types, trailer):
         trailer_op, node = trailer.children[:2]
         if node == ')':  # `arglist` is optional.
             node = ()
 
-        new_types = set()
         if trailer_op == '[':
-            new_types |= iterable.py__getitem__(self, context, types, trailer)
+            return ContextSet(iterable.py__getitem__(self, context, types, trailer))
         else:
+            context_set = ContextSet()
             for typ in types:
                 debug.dbg('eval_trailer: %s in scope %s', trailer, typ)
                 if trailer_op == '.':
-                    new_types |= typ.py__getattribute__(
+                    context_set |= typ.py__getattribute__(
                         name_context=context,
                         name_or_str=node
                     )
                 elif trailer_op == '(':
                     arguments = param.TreeArguments(self, context, node, trailer)
-                    new_types |= self.execute(typ, arguments)
-        return new_types
+                    context_set |= self.execute(typ, arguments)
+            return context_set
 
     @debug.increase_indent
     def execute(self, obj, arguments):
@@ -460,11 +465,11 @@ class Evaluator(object):
             func = obj.py__call__
         except AttributeError:
             debug.warning("no execution possible %s", obj)
-            return set()
+            return NO_CONTEXTS
         else:
-            types = func(arguments)
-            debug.dbg('execute result: %s in %s', types, obj)
-            return types
+            context_set = func(arguments)
+            debug.dbg('execute result: %s in %s', context_set, obj)
+            return context_set
 
     def goto_definitions(self, context, name):
         def_ = name.get_definition(import_name_always=True)
@@ -509,25 +514,25 @@ class Evaluator(object):
                 return module_names
 
         par = name.parent
-        typ = par.type
-        if typ == 'argument' and par.children[1] == '=' and par.children[0] == name:
+        node_type = par.type
+        if node_type == 'argument' and par.children[1] == '=' and par.children[0] == name:
             # Named param goto.
             trailer = par.parent
             if trailer.type == 'arglist':
                 trailer = trailer.parent
             if trailer.type != 'classdef':
                 if trailer.type == 'decorator':
-                    types = self.eval_element(context, trailer.children[1])
+                    context_set = self.eval_element(context, trailer.children[1])
                 else:
                     i = trailer.parent.children.index(trailer)
                     to_evaluate = trailer.parent.children[:i]
-                    types = self.eval_element(context, to_evaluate[0])
+                    context_set = self.eval_element(context, to_evaluate[0])
                     for trailer in to_evaluate[1:]:
-                        types = self.eval_trailer(context, types, trailer)
+                        context_set = self.eval_trailer(context, context_set, trailer)
                 param_names = []
-                for typ in types:
+                for context in context_set:
                     try:
-                        get_param_names = typ.get_param_names
+                        get_param_names = context.get_param_names
                     except AttributeError:
                         pass
                     else:
@@ -535,7 +540,7 @@ class Evaluator(object):
                             if param_name.string_name == name.value:
                                 param_names.append(param_name)
                 return param_names
-        elif typ == 'dotted_name':  # Is a decorator.
+        elif node_type == 'dotted_name':  # Is a decorator.
             index = par.children.index(name)
             if index > 0:
                 new_dotted = helpers.deep_ast_copy(par)
@@ -546,7 +551,7 @@ class Evaluator(object):
                     for value in values
                 )
 
-        if typ == 'trailer' and par.children[0] == '.':
+        if node_type == 'trailer' and par.children[0] == '.':
             values = helpers.evaluate_call_of_leaf(context, name, cut_own_trailer=True)
             return unite(
                 value.py__getattribute__(name, name_context=context, is_goto=True)
diff --git a/jedi/evaluate/compiled/__init__.py b/jedi/evaluate/compiled/__init__.py
index cd59b6b3..091d07d4 100644
--- a/jedi/evaluate/compiled/__init__.py
+++ b/jedi/evaluate/compiled/__init__.py
@@ -15,6 +15,7 @@ from jedi.evaluate.filters import AbstractFilter, AbstractNameDefinition, \
     ContextNameMixin
 from jedi.evaluate.context import Context, LazyKnownContext
 from jedi.evaluate.compiled.getattr_static import getattr_static
+from jedi.common import ContextSet
 from . import fake
 
 
@@ -83,9 +84,9 @@ class CompiledObject(Context):
     def py__call__(self, params):
         if inspect.isclass(self.obj):
             from jedi.evaluate.instance import CompiledInstance
-            return set([CompiledInstance(self.evaluator, self.parent_context, self, params)])
+            return ContextSet(CompiledInstance(self.evaluator, self.parent_context, self, params))
         else:
-            return set(self._execute_function(params))
+            return ContextSet.from_iterable(self._execute_function(params))
 
     @CheckAttribute
     def py__class__(self):
@@ -221,9 +222,9 @@ class CompiledObject(Context):
     def py__getitem__(self, index):
         if type(self.obj) not in (str, list, tuple, unicode, bytes, bytearray, dict):
             # Get rid of side effects, we won't call custom `__getitem__`s.
-            return set()
+            return ContextSet()
 
-        return set([create(self.evaluator, self.obj[index])])
+        return ContextSet(create(self.evaluator, self.obj[index]))
 
     @CheckAttribute
     def py__iter__(self):
@@ -278,7 +279,7 @@ class CompiledObject(Context):
         return []  # Builtins don't have imports
 
     def dict_values(self):
-        return set(create(self.evaluator, v) for v in self.obj.values())
+        return ContextSet(create(self.evaluator, v) for v in self.obj.values())
 
 
 class CompiledName(AbstractNameDefinition):
@@ -301,7 +302,9 @@ class CompiledName(AbstractNameDefinition):
     @underscore_memoization
     def infer(self):
         module = self.parent_context.get_root_context()
-        return [_create_from_name(self._evaluator, module, self.parent_context, self.string_name)]
+        return ContextSet(_create_from_name(
+            self._evaluator, module, self.parent_context, self.string_name
+        ))
 
 
 class SignatureParamName(AbstractNameDefinition):
@@ -318,13 +321,13 @@ class SignatureParamName(AbstractNameDefinition):
     def infer(self):
         p = self._signature_param
         evaluator = self.parent_context.evaluator
-        types = set()
+        contexts = ContextSet()
         if p.default is not p.empty:
-            types.add(create(evaluator, p.default))
+            contexts = ContextSet(create(evaluator, p.default))
         if p.annotation is not p.empty:
             annotation = create(evaluator, p.annotation)
-            types |= annotation.execute_evaluated()
-        return types
+            contexts |= annotation.execute_evaluated()
+        return contexts
 
 
 class UnresolvableParamName(AbstractNameDefinition):
@@ -335,7 +338,7 @@ class UnresolvableParamName(AbstractNameDefinition):
         self.string_name = name
 
     def infer(self):
-        return set()
+        return ContextSet()
 
 
 class CompiledContextName(ContextNameMixin, AbstractNameDefinition):
@@ -356,7 +359,7 @@ class EmptyCompiledName(AbstractNameDefinition):
         self.string_name = name
 
     def infer(self):
-        return []
+        return ContextSet()
 
 
 class CompiledObjectFilter(AbstractFilter):
diff --git a/jedi/evaluate/context.py b/jedi/evaluate/context.py
index 8532cdb1..df73194f 100644
--- a/jedi/evaluate/context.py
+++ b/jedi/evaluate/context.py
@@ -1,7 +1,7 @@
 from jedi._compatibility import Python3Method
-from jedi.evaluate.utils import unite
 from parso.python.tree import ExprStmt, CompFor
 from jedi.parser_utils import clean_scope_docstring, get_doc_with_call_signature
+from jedi.common import ContextSet, NO_CONTEXTS
 
 
 class Context(object):
@@ -111,11 +111,11 @@ class AbstractLazyContext(object):
 class LazyKnownContext(AbstractLazyContext):
     """data is a context."""
     def infer(self):
-        return set([self.data])
+        return ContextSet(self.data)
 
 
 class LazyKnownContexts(AbstractLazyContext):
-    """data is a set of contexts."""
+    """data is a ContextSet."""
     def infer(self):
         return self.data
 
@@ -125,7 +125,7 @@ class LazyUnknownContext(AbstractLazyContext):
         super(LazyUnknownContext, self).__init__(None)
 
     def infer(self):
-        return set()
+        return NO_CONTEXTS
 
 
 class LazyTreeContext(AbstractLazyContext):
@@ -155,7 +155,7 @@ def get_merged_lazy_context(lazy_contexts):
 class MergedLazyContexts(AbstractLazyContext):
     """data is a list of lazy contexts."""
     def infer(self):
-        return unite(l.infer() for l in self.data)
+        return ContextSet.from_sets(l.infer() for l in self.data)
 
 
 class ContextualizedNode(object):
diff --git a/jedi/evaluate/docstrings.py b/jedi/evaluate/docstrings.py
index c552d5e3..43c950ed 100644
--- a/jedi/evaluate/docstrings.py
+++ b/jedi/evaluate/docstrings.py
@@ -25,6 +25,7 @@ from jedi.evaluate.utils import unite, indent_block
 from jedi.evaluate import context
 from jedi.evaluate.cache import evaluator_method_cache
 from jedi.evaluate.iterable import SequenceLiteralContext, FakeSequence
+from jedi.common import iterator_to_context_set, ContextSet, NO_CONTEXTS
 
 
 DOCSTRING_PARAM_PATTERNS = [
@@ -245,7 +246,7 @@ def infer_param(execution_context, param):
     from jedi.evaluate.instance import AnonymousInstanceFunctionExecution
 
     def eval_docstring(docstring):
-        return set(
+        return ContextSet.from_iterable(
             p
             for param_str in _search_param_in_docstr(docstring, param.name.value)
             for p in _evaluate_for_statement_string(module_context, param_str)
@@ -253,7 +254,7 @@ def infer_param(execution_context, param):
     module_context = execution_context.get_root_context()
     func = param.get_parent_function()
     if func.type == 'lambdef':
-        return set()
+        return NO_CONTEXTS
 
     types = eval_docstring(execution_context.py__doc__())
     if isinstance(execution_context, AnonymousInstanceFunctionExecution) and \
@@ -265,6 +266,7 @@ def infer_param(execution_context, param):
 
 
 @evaluator_method_cache()
+@iterator_to_context_set
 def infer_return_types(function_context):
     def search_return_in_docstr(code):
         for p in DOCSTRING_RETURN_PATTERNS:
@@ -278,4 +280,3 @@ def infer_return_types(function_context):
     for type_str in search_return_in_docstr(function_context.py__doc__()):
         for type_eval in _evaluate_for_statement_string(function_context.get_root_context(), type_str):
             yield type_eval
-
diff --git a/jedi/evaluate/finder.py b/jedi/evaluate/finder.py
index 38a1b357..3e5e2ef5 100644
--- a/jedi/evaluate/finder.py
+++ b/jedi/evaluate/finder.py
@@ -32,6 +32,7 @@ from jedi.evaluate import param
 from jedi.evaluate import helpers
 from jedi.evaluate.filters import get_global_filters, TreeNameDefinition
 from jedi.evaluate.context import ContextualizedName, ContextualizedNode
+from jedi.common import ContextSet
 from jedi.parser_utils import is_scope, get_parent_scope
 
 
@@ -62,7 +63,7 @@ class NameFinder(object):
             check = flow_analysis.reachability_check(
                 self._context, self._context.tree_node, self._name)
             if check is flow_analysis.UNREACHABLE:
-                return set()
+                return ContextSet()
             return self._found_predefined_types
 
         types = self._names_to_types(names, attribute_lookup)
@@ -158,22 +159,20 @@ class NameFinder(object):
         return inst.execute_function_slots(names, name)
 
     def _names_to_types(self, names, attribute_lookup):
-        types = set()
+        contexts = ContextSet.from_sets(name.infer() for name in names)
 
-        types = unite(name.infer() for name in names)
-
-        debug.dbg('finder._names_to_types: %s -> %s', names, types)
+        debug.dbg('finder._names_to_types: %s -> %s', names, contexts)
         if not names and isinstance(self._context, AbstractInstanceContext):
             # handling __getattr__ / __getattribute__
             return self._check_getattr(self._context)
 
         # Add isinstance and other if/assert knowledge.
-        if not types and isinstance(self._name, tree.Name) and \
+        if not contexts and isinstance(self._name, tree.Name) and \
                 not isinstance(self._name_context, AbstractInstanceContext):
             flow_scope = self._name
             base_node = self._name_context.tree_node
             if base_node.type == 'comp_for':
-                return types
+                return contexts
             while True:
                 flow_scope = get_parent_scope(flow_scope, include_flows=True)
                 n = _check_flow_information(self._name_context, flow_scope,
@@ -182,7 +181,7 @@ class NameFinder(object):
                     return n
                 if flow_scope == base_node:
                     break
-        return types
+        return contexts
 
 
 def _name_to_types(evaluator, context, tree_name):
@@ -210,6 +209,7 @@ def _name_to_types(evaluator, context, tree_name):
         types = pep0484.find_type_from_comment_hint_with(context, node, tree_name)
         if types:
             return types
+
     if typ in ('for_stmt', 'comp_for'):
         try:
             types = context.predefined_names[node][tree_name.value]
@@ -222,11 +222,8 @@ def _name_to_types(evaluator, context, tree_name):
         types = _remove_statements(evaluator, context, node, tree_name)
     elif typ == 'with_stmt':
         context_managers = context.eval_node(node.get_test_node_from_name(tree_name))
-        enter_methods = unite(
-            context_manager.py__getattribute__('__enter__')
-            for context_manager in context_managers
-        )
-        types = unite(method.execute_evaluated() for method in enter_methods)
+        enter_methods = context_managers.py__getattribute__('__enter__')
+        return enter_methods.execute_evaluated()
     elif typ in ('import_from', 'import_name'):
         types = imports.infer_import(context, tree_name)
     elif typ in ('funcdef', 'classdef'):
@@ -262,7 +259,7 @@ def _apply_decorators(evaluator, context, node):
             parent_context=context,
             funcdef=node
         )
-    initial = values = set([decoratee_context])
+    initial = values = ContextSet(decoratee_context)
     for dec in reversed(node.get_decorators()):
         debug.dbg('decorator: %s %s', dec, values)
         dec_values = context.eval_node(dec.children[1])
@@ -294,20 +291,12 @@ def _remove_statements(evaluator, context, stmt, name):
     Due to lazy evaluation, statements like a = func; b = a; b() have to be
     evaluated.
     """
-    types = set()
-    check_instance = None
-
-    pep0484types = \
+    pep0484_contexts = \
         pep0484.find_type_from_comment_hint_assign(context, stmt, name)
-    if pep0484types:
-        return pep0484types
-    types |= context.eval_stmt(stmt, seek_name=name)
+    if pep0484_contexts:
+        return pep0484_contexts
 
-    if check_instance is not None:
-        # class renames
-        types = set([er.get_instance_el(evaluator, check_instance, a, True)
-                     if isinstance(a, er.Function) else a for a in types])
-    return types
+    return context.eval_stmt(stmt, seek_name=name)
 
 
 def _check_flow_information(context, flow, search_name, pos):
@@ -377,26 +366,26 @@ def _check_isinstance_type(context, element, search_name):
     except AssertionError:
         return None
 
-    result = set()
+    context_set = ContextSet()
     for cls_or_tup in lazy_context_cls.infer():
         if isinstance(cls_or_tup, iterable.AbstractSequence) and \
                 cls_or_tup.array_type == 'tuple':
             for lazy_context in cls_or_tup.py__iter__():
                 for context in lazy_context.infer():
-                    result |= context.execute_evaluated()
+                    context_set |= context.execute_evaluated()
         else:
-            result |= cls_or_tup.execute_evaluated()
-    return result
+            context_set |= cls_or_tup.execute_evaluated()
+    return context_set
 
 
-def check_tuple_assignments(evaluator, contextualized_name, types):
+def check_tuple_assignments(evaluator, contextualized_name, context_set):
     """
     Checks if tuples are assigned.
     """
     lazy_context = None
     for index, node in contextualized_name.assignment_indexes():
         cn = ContextualizedNode(contextualized_name.context, node)
-        iterated = iterable.py__iter__(evaluator, types, cn)
+        iterated = iterable.py__iter__(evaluator, context_set, cn)
         for _ in range(index + 1):
             try:
                 lazy_context = next(iterated)
@@ -405,6 +394,6 @@ def check_tuple_assignments(evaluator, contextualized_name, types):
                 # would allow this loop to run for a very long time if the
                 # index number is high. Therefore break if the loop is
                 # finished.
-                return set()
-        types = lazy_context.infer()
-    return types
+                return ContextSet()
+        context_set = lazy_context.infer()
+    return context_set
diff --git a/jedi/evaluate/imports.py b/jedi/evaluate/imports.py
index 12ed630a..75139d15 100644
--- a/jedi/evaluate/imports.py
+++ b/jedi/evaluate/imports.py
@@ -31,11 +31,12 @@ from jedi.evaluate import compiled
 from jedi.evaluate import analysis
 from jedi.evaluate.cache import evaluator_method_cache
 from jedi.evaluate.filters import AbstractNameDefinition
+from jedi.common import ContextSet, NO_CONTEXTS
 
 
 # This memoization is needed, because otherwise we will infinitely loop on
 # certain imports.
-@evaluator_method_cache(default=set())
+@evaluator_method_cache(default=NO_CONTEXTS)
 def infer_import(context, tree_name, is_goto=False):
     module_context = context.get_root_context()
     import_node = search_ancestor(tree_name, 'import_name', 'import_from')
@@ -63,7 +64,7 @@ def infer_import(context, tree_name, is_goto=False):
     #    scopes = [NestedImportModule(module, import_node)]
 
     if not types:
-        return set()
+        return NO_CONTEXTS
 
     if from_import_name is not None:
         types = unite(
@@ -270,7 +271,7 @@ class Importer(object):
 
     def follow(self):
         if not self.import_path:
-            return set()
+            return NO_CONTEXTS
         return self._do_import(self.import_path, self.sys_path_with_modifications())
 
     def _do_import(self, import_path, sys_path):
@@ -296,7 +297,7 @@ class Importer(object):
 
         module_name = '.'.join(import_parts)
         try:
-            return set([self._evaluator.modules[module_name]])
+            return ContextSet(self._evaluator.modules[module_name])
         except KeyError:
             pass
 
@@ -305,7 +306,7 @@ class Importer(object):
             # the module cache.
             bases = self._do_import(import_path[:-1], sys_path)
             if not bases:
-                return set()
+                return NO_CONTEXTS
             # We can take the first element, because only the os special
             # case yields multiple modules, which is not important for
             # further imports.
@@ -323,7 +324,7 @@ class Importer(object):
             except AttributeError:
                 # The module is not a package.
                 _add_error(self.module_context, import_path[-1])
-                return set()
+                return NO_CONTEXTS
             else:
                 paths = method()
                 debug.dbg('search_module %s in paths %s', module_name, paths)
@@ -340,7 +341,7 @@ class Importer(object):
                         module_path = None
                 if module_path is None:
                     _add_error(self.module_context, import_path[-1])
-                    return set()
+                    return NO_CONTEXTS
         else:
             parent_module = None
             try:
@@ -356,7 +357,7 @@ class Importer(object):
             except ImportError:
                 # The module is not a package.
                 _add_error(self.module_context, import_path[-1])
-                return set()
+                return NO_CONTEXTS
 
         code = None
         if is_pkg:
@@ -383,10 +384,10 @@ class Importer(object):
         if module is None:
             # The file might raise an ImportError e.g. and therefore not be
             # importable.
-            return set()
+            return NO_CONTEXTS
 
         self._evaluator.modules[module_name] = module
-        return set([module])
+        return ContextSet(module)
 
     def _generate_name(self, name, in_module=None):
         # Create a pseudo import to be able to follow them.
diff --git a/jedi/evaluate/instance.py b/jedi/evaluate/instance.py
index 52852d1e..f8df7a1b 100644
--- a/jedi/evaluate/instance.py
+++ b/jedi/evaluate/instance.py
@@ -11,6 +11,7 @@ from jedi.evaluate.param import AbstractArguments, AnonymousArguments
 from jedi.cache import memoize_method
 from jedi.evaluate import representation as er
 from jedi.evaluate import iterable
+from jedi.common import ContextSet, iterator_to_context_set
 from jedi.parser_utils import get_parent_scope
 
 
@@ -250,6 +251,7 @@ class CompiledInstanceName(compiled.CompiledName):
         super(CompiledInstanceName, self).__init__(evaluator, parent_context, name)
         self._instance = instance
 
+    @iterator_to_context_set
     def infer(self):
         for result_context in super(CompiledInstanceName, self).infer():
             if isinstance(result_context, er.FunctionContext):
@@ -311,9 +313,7 @@ class CompiledBoundMethod(compiled.CompiledObject):
 
 class InstanceNameDefinition(filters.TreeNameDefinition):
     def infer(self):
-        contexts = super(InstanceNameDefinition, self).infer()
-        for context in contexts:
-            yield context
+        return super(InstanceNameDefinition, self).infer()
 
 
 class LazyInstanceName(filters.TreeNameDefinition):
@@ -331,6 +331,7 @@ class LazyInstanceName(filters.TreeNameDefinition):
 
 
 class LazyInstanceClassName(LazyInstanceName):
+    @iterator_to_context_set
     def infer(self):
         for result_context in super(LazyInstanceClassName, self).infer():
             if isinstance(result_context, er.FunctionContext):
diff --git a/jedi/evaluate/iterable.py b/jedi/evaluate/iterable.py
index 6e51894d..cee0879b 100644
--- a/jedi/evaluate/iterable.py
+++ b/jedi/evaluate/iterable.py
@@ -35,6 +35,7 @@ from jedi.evaluate import recursion
 from jedi.evaluate.cache import evaluator_method_cache
 from jedi.evaluate.filters import DictFilter, AbstractNameDefinition, \
     ParserTreeFilter
+from jedi.common import ContextSet, NO_CONTEXTS
 from jedi.parser_utils import get_comp_fors
 
 
@@ -87,7 +88,7 @@ class SpecialMethodFilter(DictFilter):
             # always only going to be one name. The same is true for the
             # inferred values.
             builtin_func = next(iter(filter.get(self.string_name)[0].infer()))
-            return set([BuiltinMethod(self.parent_context, self._callable, builtin_func)])
+            return ContextSet(BuiltinMethod(self.parent_context, self._callable, builtin_func))
 
     def __init__(self, context, dct, builtin_context):
         super(SpecialMethodFilter, self).__init__(dct)
@@ -304,7 +305,7 @@ class ListComprehension(ArrayMixin, Comprehension):
 
     def py__getitem__(self, index):
         if isinstance(index, slice):
-            return set([self])
+            return ContextSet(self)
 
         all_types = list(self.py__iter__())
         return all_types[index].infer()
@@ -339,11 +340,11 @@ class DictComprehension(ArrayMixin, Comprehension):
     @register_builtin_method('values')
     def _imitate_values(self):
         lazy_context = context.LazyKnownContexts(self.dict_values())
-        return set([FakeSequence(self.evaluator, 'list', [lazy_context])])
+        return ContextSet(FakeSequence(self.evaluator, 'list', [lazy_context]))
 
     @register_builtin_method('items')
     def _imitate_items(self):
-        items = set(
+        items = ContextSet.from_iterable(
             FakeSequence(
                 self.evaluator, 'tuple'
                 (context.LazyKnownContexts(keys), context.LazyKnownContexts(values))
@@ -385,7 +386,7 @@ class SequenceLiteralContext(ArrayMixin, AbstractSequence):
 
         # Can raise an IndexError
         if isinstance(index, slice):
-            return set([self])
+            return ContextSet(self)
         else:
             return self._defining_context.eval_node(self._items()[index])
 
@@ -396,7 +397,7 @@ class SequenceLiteralContext(ArrayMixin, AbstractSequence):
         """
         if self.array_type == 'dict':
             # Get keys.
-            types = set()
+            types = ContextSet()
             for k, _ in self._items():
                 types |= self._defining_context.eval_node(k)
             # We don't know which dict index comes first, therefore always
@@ -470,7 +471,7 @@ class DictLiteralContext(SequenceLiteralContext):
     @register_builtin_method('values')
     def _imitate_values(self):
         lazy_context = context.LazyKnownContexts(self.dict_values())
-        return set([FakeSequence(self.evaluator, 'list', [lazy_context])])
+        return ContextSet(FakeSequence(self.evaluator, 'list', [lazy_context]))
 
     @register_builtin_method('items')
     def _imitate_items(self):
@@ -482,7 +483,7 @@ class DictLiteralContext(SequenceLiteralContext):
             )) for key_node, value_node in self._items()
         ]
 
-        return set([FakeSequence(self.evaluator, 'list', lazy_contexts)])
+        return ContextSet(FakeSequence(self.evaluator, 'list', lazy_contexts))
 
 
 class _FakeArray(SequenceLiteralContext):
@@ -506,7 +507,7 @@ class FakeSequence(_FakeArray):
         return self._context_list
 
     def py__getitem__(self, index):
-        return set(self._lazy_context_list[index].infer())
+        return self._lazy_context_list[index].infer()
 
     def py__iter__(self):
         return self._lazy_context_list
@@ -641,7 +642,7 @@ def py__iter__types(evaluator, types, contextualized_node=None):
 def py__getitem__(evaluator, context, types, trailer):
     from jedi.evaluate.representation import ClassContext
     from jedi.evaluate.instance import TreeInstance
-    result = set()
+    result = ContextSet()
 
     trailer_op, node, trailer_cl = trailer.children
     assert trailer_op == "["
@@ -685,7 +686,7 @@ def py__getitem__(evaluator, context, types, trailer):
                 try:
                     result |= getitem(index)
                 except IndexError:
-                    result |= py__iter__types(evaluator, set([typ]))
+                    result |= py__iter__types(evaluator, ContextSet(typ))
                 except KeyError:
                     # Must be a dict. Lists don't raise KeyErrors.
                     result |= typ.dict_values()
@@ -696,12 +697,12 @@ def check_array_additions(context, sequence):
     """ Just a mapper function for the internal _check_array_additions """
     if sequence.array_type not in ('list', 'set'):
         # TODO also check for dict updates
-        return set()
+        return NO_CONTEXTS
 
     return _check_array_additions(context, sequence)
 
 
-@evaluator_method_cache(default=set())
+@evaluator_method_cache(default=NO_CONTEXTS)
 @debug.increase_indent
 def _check_array_additions(context, sequence):
     """
@@ -716,11 +717,11 @@ def _check_array_additions(context, sequence):
     module_context = context.get_root_context()
     if not settings.dynamic_array_additions or isinstance(module_context, compiled.CompiledObject):
         debug.dbg('Dynamic array search aborted.', color='MAGENTA')
-        return set()
+        return ContextSet()
 
     def find_additions(context, arglist, add_name):
         params = list(param.TreeArguments(context.evaluator, context, arglist).unpack())
-        result = set()
+        result = ContextSet()
         if add_name in ['insert']:
             params = params[1:]
         if add_name in ['append', 'add', 'insert']:
@@ -728,7 +729,9 @@ def _check_array_additions(context, sequence):
                 result.add(lazy_context)
         elif add_name in ['extend', 'update']:
             for key, lazy_context in params:
-                result |= set(py__iter__(context.evaluator, lazy_context.infer()))
+                result |= ContextSet.from_iterable(
+                    py__iter__(context.evaluator, lazy_context.infer())
+                )
         return result
 
     temp_param_add, settings.dynamic_params_for_other_modules = \
@@ -737,7 +740,7 @@ def _check_array_additions(context, sequence):
     is_list = sequence.name.string_name == 'list'
     search_names = (['append', 'extend', 'insert'] if is_list else ['add', 'update'])
 
-    added_types = set()
+    added_types = NO_CONTEXTS()
     for add_name in search_names:
         try:
             possible_names = module_context.tree_node.get_used_names()[add_name]
@@ -870,7 +873,7 @@ def create_index_types(evaluator, context, index):
     """
     if index == ':':
         # Like array[:]
-        return set([Slice(context, None, None, None)])
+        return ContextSet(Slice(context, None, None, None))
 
     elif index.type == 'subscript' and not index.children[0] == '.':
         # subscript basically implies a slice operation, except for Python 2's
@@ -888,7 +891,7 @@ def create_index_types(evaluator, context, index):
                 result.append(el)
         result += [None] * (3 - len(result))
 
-        return set([Slice(context, *result)])
+        return ContextSet(Slice(context, *result))
 
     # No slices
     return context.eval_node(index)
diff --git a/jedi/evaluate/param.py b/jedi/evaluate/param.py
index 5490df1f..04d93d42 100644
--- a/jedi/evaluate/param.py
+++ b/jedi/evaluate/param.py
@@ -237,7 +237,7 @@ class ExecutedParam(object):
         pep0484_hints = pep0484.infer_param(self._execution_context, self._param_node)
         doc_params = docstrings.infer_param(self._execution_context, self._param_node)
         if pep0484_hints or doc_params:
-            return list(set(pep0484_hints) | set(doc_params))
+            return pep0484_hints | doc_params
 
         return self._lazy_context.infer()
 
diff --git a/jedi/evaluate/pep0484.py b/jedi/evaluate/pep0484.py
index f9a5c5ed..8e741fb0 100644
--- a/jedi/evaluate/pep0484.py
+++ b/jedi/evaluate/pep0484.py
@@ -19,7 +19,6 @@ x support for type hint comments for functions, `# type: (int, str) -> int`.
     See comment from Guido https://github.com/davidhalter/jedi/issues/662
 """
 
-import itertools
 import os
 import re
 
@@ -30,6 +29,7 @@ from jedi.evaluate.utils import unite
 from jedi.evaluate.cache import evaluator_method_cache
 from jedi.evaluate import compiled
 from jedi.evaluate.context import LazyTreeContext
+from jedi.common import NO_CONTEXTS
 from jedi import debug
 from jedi import _compatibility
 from jedi import parser_utils
@@ -42,16 +42,15 @@ def _evaluate_for_annotation(context, annotation, index=None):
     and we're interested in that index
     """
     if annotation is not None:
-        definitions = context.eval_node(
-            _fix_forward_reference(context, annotation))
+        context_set = context.eval_node(_fix_forward_reference(context, annotation))
         if index is not None:
-            definitions = list(itertools.chain.from_iterable(
-                definition.py__getitem__(index) for definition in definitions
-                if definition.array_type == 'tuple' and
-                len(list(definition.py__iter__())) >= index))
-        return unite(d.execute_evaluated() for d in definitions)
+            context_set = context_set.filter(
+                lambda context: context.array_type == 'tuple' \
+                                and len(list(context.py__iter__())) >= index
+            ).py__getitem__(index)
+        return context_set.execute_evaluated()
     else:
-        return set()
+        return NO_CONTEXTS
 
 
 def _fix_forward_reference(context, node):
diff --git a/jedi/evaluate/precedence.py b/jedi/evaluate/precedence.py
index 44335217..5d6aef13 100644
--- a/jedi/evaluate/precedence.py
+++ b/jedi/evaluate/precedence.py
@@ -7,6 +7,7 @@ from jedi._compatibility import unicode
 from jedi import debug
 from jedi.evaluate.compiled import CompiledObject, create, builtin_from_name
 from jedi.evaluate import analysis
+from jedi.common import ContextSet
 
 # Maps Python syntax to the operator module.
 COMPARISON_OPERATORS = {
@@ -33,7 +34,7 @@ def literals_to_types(evaluator, result):
             new_result |= cls.execute_evaluated()
         else:
             new_result.add(typ)
-    return new_result
+    return ContextSet.from_set(new_result)
 
 
 def calculate_children(evaluator, context, children):
@@ -49,7 +50,7 @@ def calculate_children(evaluator, context, children):
 
         # handle lazy evaluation of and/or here.
         if operator in ('and', 'or'):
-            left_bools = set([left.py__bool__() for left in types])
+            left_bools = ContextSet(left.py__bool__() for left in types)
             if left_bools == set([True]):
                 if operator == 'and':
                     types = context.eval_node(right)
@@ -65,22 +66,22 @@ def calculate_children(evaluator, context, children):
 
 
 def calculate(evaluator, context, left_result, operator, right_result):
-    result = set()
     if not left_result or not right_result:
         # illegal slices e.g. cause left/right_result to be None
         result = (left_result or set()) | (right_result or set())
-        result = literals_to_types(evaluator, result)
+        return literals_to_types(evaluator, result)
     else:
         # I don't think there's a reasonable chance that a string
         # operation is still correct, once we pass something like six
         # objects.
         if len(left_result) * len(right_result) > 6:
-            result = literals_to_types(evaluator, left_result | right_result)
+            return literals_to_types(evaluator, left_result | right_result)
         else:
-            for left in left_result:
-                for right in right_result:
-                    result |= _element_calculate(evaluator, context, left, operator, right)
-    return result
+            return ContextSet.from_sets(
+                _element_calculate(evaluator, context, left, operator, right)
+                for left in left_result
+                for right in right_result
+            )
 
 
 def factor_calculate(evaluator, types, operator):
@@ -131,21 +132,21 @@ def _element_calculate(evaluator, context, left, operator, right):
     if operator == '*':
         # for iterables, ignore * operations
         if isinstance(left, iterable.AbstractSequence) or is_string(left):
-            return set([left])
+            return ContextSet(left)
         elif isinstance(right, iterable.AbstractSequence) or is_string(right):
-            return set([right])
+            return ContextSet(right)
     elif operator == '+':
         if l_is_num and r_is_num or is_string(left) and is_string(right):
-            return set([create(evaluator, left.obj + right.obj)])
+            return ContextSet(create(evaluator, left.obj + right.obj))
         elif _is_tuple(left) and _is_tuple(right) or _is_list(left) and _is_list(right):
-            return set([iterable.MergedArray(evaluator, (left, right))])
+            return ContextSet(iterable.MergedArray(evaluator, (left, right)))
     elif operator == '-':
         if l_is_num and r_is_num:
-            return set([create(evaluator, left.obj - right.obj)])
+            return ContextSet(create(evaluator, left.obj - right.obj))
     elif operator == '%':
         # With strings and numbers the left type typically remains. Except for
         # `int() % float()`.
-        return set([left])
+        return ContextSet(left)
     elif operator in COMPARISON_OPERATORS:
         operation = COMPARISON_OPERATORS[operator]
         if isinstance(left, CompiledObject) and isinstance(right, CompiledObject):
@@ -157,9 +158,9 @@ def _element_calculate(evaluator, context, left, operator, right):
             result = operation(left, right)
         except TypeError:
             # Could be True or False.
-            return set([create(evaluator, True), create(evaluator, False)])
+            return ContextSet(create(evaluator, True), create(evaluator, False))
         else:
-            return set([create(evaluator, result)])
+            return ContextSet(create(evaluator, result))
     elif operator == 'in':
         return set()
 
@@ -175,4 +176,4 @@ def _element_calculate(evaluator, context, left, operator, right):
         analysis.add(context, 'type-error-operation', operator,
                      message % (left, right))
 
-    return set([left, right])
+    return ContextSet(left, right)
diff --git a/jedi/evaluate/representation.py b/jedi/evaluate/representation.py
index 8a0d616c..8ccfc7b2 100644
--- a/jedi/evaluate/representation.py
+++ b/jedi/evaluate/representation.py
@@ -64,6 +64,7 @@ from jedi.evaluate.filters import ParserTreeFilter, FunctionExecutionFilter, \
     ContextNameMixin
 from jedi.evaluate import context
 from jedi.evaluate.context import ContextualizedNode
+from jedi.common import NO_CONTEXTS, ContextSet, iterator_to_context_set
 from jedi import parser_utils
 from jedi.evaluate.parser_cache import get_yield_exprs
 
@@ -83,6 +84,7 @@ class ClassName(TreeNameDefinition):
         super(ClassName, self).__init__(parent_context, tree_name)
         self._name_context = name_context
 
+    @iterator_to_context_set
     def infer(self):
         # TODO this _name_to_types might get refactored and be a part of the
         # parent class. Once it is, we can probably just overwrite method to
@@ -162,7 +164,7 @@ class ClassContext(use_metaclass(CachedMetaClass, context.TreeContext)):
 
     def py__call__(self, params):
         from jedi.evaluate.instance import TreeInstance
-        return set([TreeInstance(self.evaluator, self.parent_context, self, params)])
+        return ContextSet(TreeInstance(self.evaluator, self.parent_context, self, params))
 
     def py__class__(self):
         return compiled.create(self.evaluator, type)
@@ -227,7 +229,7 @@ class LambdaName(AbstractNameDefinition):
         return self._lambda_context.tree_node.start_pos
 
     def infer(self):
-        return set([self._lambda_context])
+        return ContextSet(self._lambda_context)
 
 
 class FunctionContext(use_metaclass(CachedMetaClass, context.TreeContext)):
@@ -260,7 +262,7 @@ class FunctionContext(use_metaclass(CachedMetaClass, context.TreeContext)):
         """
         yield_exprs = get_yield_exprs(self.evaluator, self.tree_node)
         if yield_exprs:
-            return set([iterable.Generator(self.evaluator, function_execution)])
+            return ContextSet(iterable.Generator(self.evaluator, function_execution))
         else:
             return function_execution.get_return_values()
 
@@ -312,7 +314,7 @@ class FunctionExecutionContext(context.TreeContext):
         self.tree_node = function_context.tree_node
         self.var_args = var_args
 
-    @evaluator_method_cache(default=set())
+    @evaluator_method_cache(default=NO_CONTEXTS)
     @recursion.execution_recursion_decorator()
     def get_return_values(self, check_yields=False):
         funcdef = self.tree_node
@@ -320,12 +322,12 @@ class FunctionExecutionContext(context.TreeContext):
             return self.evaluator.eval_element(self, funcdef.children[-1])
 
         if check_yields:
-            types = set()
+            context_set = NO_CONTEXTS
             returns = get_yield_exprs(self.evaluator, funcdef)
         else:
             returns = funcdef.iter_return_stmts()
-            types = set(docstrings.infer_return_types(self.function_context))
-            types |= set(pep0484.infer_return_types(self.function_context))
+            context_set = docstrings.infer_return_types(self.function_context)
+            context_set |= pep0484.infer_return_types(self.function_context)
 
         for r in returns:
             check = flow_analysis.reachability_check(self, funcdef, r)
@@ -333,18 +335,18 @@ class FunctionExecutionContext(context.TreeContext):
                 debug.dbg('Return unreachable: %s', r)
             else:
                 if check_yields:
-                    types |= set(self._eval_yield(r))
+                    context_set |= ContextSet(self._eval_yield(r))
                 else:
                     try:
                         children = r.children
                     except AttributeError:
-                        types.add(compiled.create(self.evaluator, None))
+                        context_set |= ContextSet(compiled.create(self.evaluator, None))
                     else:
-                        types |= self.eval_node(children[1])
+                        context_set |= self.eval_node(children[1])
             if check is flow_analysis.REACHABLE:
                 debug.dbg('Return reachable: %s', r)
                 break
-        return types
+        return context_set
 
     def _eval_yield(self, yield_expr):
         if yield_expr.type == 'keyword':
@@ -430,8 +432,10 @@ class ModuleAttributeName(AbstractNameDefinition):
         self.string_name = string_name
 
     def infer(self):
-        return compiled.create(self.parent_context.evaluator, str).execute(
-            param.ValuesArguments([])
+        return ContextSet(
+            compiled.create(self.parent_context.evaluator, str).execute(
+                param.ValuesArguments([])
+            )
         )
 
 
@@ -628,7 +632,7 @@ class ImplicitNSName(AbstractNameDefinition):
         self.string_name = string_name
 
     def infer(self):
-        return []
+        return NO_CONTEXTS
 
     def get_root_context(self):
         return self.implicit_ns_context