Support recursive named tuples (#13371)

This is a continuation of #13297 

The main change here is that although named tuples are still stored in symbol tables as `TypeInfo`s, when type analyzer sees them, it creates a `TypeAliasType` targeting what it would return before (a `TupleType` with a fallback to an instance of that `TypeInfo`). Although it is a significant change, IMO this is the simplest but still clean way to support recursive named tuples.

Also it is very simple to extend to TypedDicts, but I wanted to make the latter in a separate PR, to minimize the scope of changes.
It would be great if someone can take a look at this PR soon.

The most code changes are to make named tuples semantic analysis idempotent, previously they were analyzed "for real" only once, when all types were ready. It is not possible anymore if we want them to be recursive. So I pass in `existing_info` everywhere, and update it instead of creating a new one every time.

mypy/checkmember.py

@@ -563,6 +563,7 @@ def analyze_descriptor_access(descriptor_type: Type, mx: MemberContext) -> Type:
  The return type of the appropriate ``__get__`` overload for the descriptor.
  """
  instance_type = get_proper_type(mx.original_type)
+ orig_descriptor_type = descriptor_type
  descriptor_type = get_proper_type(descriptor_type)
 
  if isinstance(descriptor_type, UnionType):
@@ -571,10 +572,10 @@ def analyze_descriptor_access(descriptor_type: Type, mx: MemberContext) -> Type:
  [analyze_descriptor_access(typ, mx) for typ in descriptor_type.items]
  )
  elif not isinstance(descriptor_type, Instance):
- return descriptor_type
+ return orig_descriptor_type
 
  if not descriptor_type.type.has_readable_member("__get__"):
- return descriptor_type
+ return orig_descriptor_type
 
  dunder_get = descriptor_type.type.get_method("__get__")
  if dunder_get is None:

mypy/fixup.py

@@ -75,6 +75,7 @@ def visit_type_info(self, info: TypeInfo) -> None:
  p.accept(self.type_fixer)
  if info.tuple_type:
  info.tuple_type.accept(self.type_fixer)
+ info.update_tuple_type(info.tuple_type)
  if info.typeddict_type:
  info.typeddict_type.accept(self.type_fixer)
  if info.declared_metaclass:
@@ -337,6 +338,12 @@ def lookup_fully_qualified_alias(
  node = stnode.node if stnode else None
  if isinstance(node, TypeAlias):
  return node
+ elif isinstance(node, TypeInfo):
+ if node.tuple_alias:
+ return node.tuple_alias
+ alias = TypeAlias.from_tuple_type(node)
+ node.tuple_alias = alias
+ return alias
  else:
  # Looks like a missing TypeAlias during an initial daemon load, put something there
  assert (

mypy/messages.py

@@ -2292,6 +2292,11 @@ def visit_instance(self, t: Instance) -> None:
  self.instances.append(t)
  super().visit_instance(t)
 
+ def visit_type_alias_type(self, t: TypeAliasType) -> None:
+ if t.alias and not t.is_recursive:
+ t.alias.target.accept(self)
+ super().visit_type_alias_type(t)
+
 
 def find_type_overlaps(*types: Type) -> Set[str]:
  """Return a set of fullnames that share a short name and appear in either type.

mypy/nodes.py

@@ -2656,6 +2656,7 @@ class is generic then it will be a type constructor of higher kind.
  "bases",
  "_promote",
  "tuple_type",
+ "tuple_alias",
  "is_named_tuple",
  "typeddict_type",
  "is_newtype",
@@ -2794,6 +2795,9 @@ class is generic then it will be a type constructor of higher kind.
  # It is useful for plugins to add their data to save in the cache.
  metadata: Dict[str, JsonDict]
 
+ # Store type alias representing this type (for named tuples).
+ tuple_alias: Optional["TypeAlias"]
+
  FLAGS: Final = [
  "is_abstract",
  "is_enum",
@@ -2840,6 +2844,7 @@ def __init__(self, names: "SymbolTable", defn: ClassDef, module_name: str) -> No
  self._promote = []
  self.alt_promote = None
  self.tuple_type = None
+ self.tuple_alias = None
  self.is_named_tuple = False
  self.typeddict_type = None
  self.is_newtype = False
@@ -2970,6 +2975,15 @@ def direct_base_classes(self) -> "List[TypeInfo]":
  """
  return [base.type for base in self.bases]
 
+ def update_tuple_type(self, typ: "mypy.types.TupleType") -> None:
+ """Update tuple_type and tuple_alias as needed."""
+ self.tuple_type = typ
+ alias = TypeAlias.from_tuple_type(self)
+ if not self.tuple_alias:
+ self.tuple_alias = alias
+ else:
+ self.tuple_alias.target = alias.target
+
  def __str__(self) -> str:
  """Return a string representation of the type.
 
@@ -3258,6 +3272,17 @@ def __init__(
  self.eager = eager
  super().__init__(line, column)
 
+ @classmethod
+ def from_tuple_type(cls, info: TypeInfo) -> "TypeAlias":
+ """Generate an alias to the tuple type described by a given TypeInfo."""
+ assert info.tuple_type
+ return TypeAlias(
+ info.tuple_type.copy_modified(fallback=mypy.types.Instance(info, [])),
+ info.fullname,
+ info.line,
+ info.column,
+ )
+
  @property
  def name(self) -> str:
  return self._fullname.split(".")[-1]

mypy/semanal.py

@@ -221,11 +221,11 @@
  PRIORITY_FALLBACKS,
  SemanticAnalyzerInterface,
  calculate_tuple_fallback,
+ has_placeholder,
  set_callable_name as set_callable_name,
 )
 from mypy.semanal_typeddict import TypedDictAnalyzer
 from mypy.tvar_scope import TypeVarLikeScope
-from mypy.type_visitor import TypeQuery
 from mypy.typeanal import (
  TypeAnalyser,
  TypeVarLikeList,
@@ -1425,7 +1425,12 @@ def analyze_class_body_common(self, defn: ClassDef) -> None:
 
  def analyze_namedtuple_classdef(self, defn: ClassDef) -> bool:
  """Check if this class can define a named tuple."""
- if defn.info and defn.info.is_named_tuple:
+ if (
+ defn.info
+ and defn.info.is_named_tuple
+ and defn.info.tuple_type
+ and not has_placeholder(defn.info.tuple_type)
+ ):
  # Don't reprocess everything. We just need to process methods defined
  # in the named tuple class body.
  is_named_tuple, info = True, defn.info # type: bool, Optional[TypeInfo]
@@ -1782,10 +1787,9 @@ def configure_base_classes(
  base_types: List[Instance] = []
  info = defn.info
 
- info.tuple_type = None
  for base, base_expr in bases:
  if isinstance(base, TupleType):
- actual_base = self.configure_tuple_base_class(defn, base, base_expr)
+ actual_base = self.configure_tuple_base_class(defn, base)
  base_types.append(actual_base)
  elif isinstance(base, Instance):
  if base.type.is_newtype:
@@ -1828,23 +1832,19 @@ def configure_base_classes(
  return
  self.calculate_class_mro(defn, self.object_type)
 
- def configure_tuple_base_class(
- self, defn: ClassDef, base: TupleType, base_expr: Expression
- ) -> Instance:
+ def configure_tuple_base_class(self, defn: ClassDef, base: TupleType) -> Instance:
  info = defn.info
 
  # There may be an existing valid tuple type from previous semanal iterations.
  # Use equality to check if it is the case.
- if info.tuple_type and info.tuple_type != base:
+ if info.tuple_type and info.tuple_type != base and not has_placeholder(info.tuple_type):
  self.fail("Class has two incompatible bases derived from tuple", defn)
  defn.has_incompatible_baseclass = True
- info.tuple_type = base
- if isinstance(base_expr, CallExpr):
- defn.analyzed = NamedTupleExpr(base.partial_fallback.type)
- defn.analyzed.line = defn.line
- defn.analyzed.column = defn.column
+ if info.tuple_alias and has_placeholder(info.tuple_alias.target):
+ self.defer(force_progress=True)
+ info.update_tuple_type(base)
 
- if base.partial_fallback.type.fullname == "builtins.tuple":
+ if base.partial_fallback.type.fullname == "builtins.tuple" and not has_placeholder(base):
  # Fallback can only be safely calculated after semantic analysis, since base
  # classes may be incomplete. Postpone the calculation.
  self.schedule_patch(PRIORITY_FALLBACKS, lambda: calculate_tuple_fallback(base))
@@ -2627,7 +2627,10 @@ def analyze_enum_assign(self, s: AssignmentStmt) -> bool:
  def analyze_namedtuple_assign(self, s: AssignmentStmt) -> bool:
  """Check if s defines a namedtuple."""
  if isinstance(s.rvalue, CallExpr) and isinstance(s.rvalue.analyzed, NamedTupleExpr):
- return True # This is a valid and analyzed named tuple definition, nothing to do here.
+ if s.rvalue.analyzed.info.tuple_type and not has_placeholder(
+ s.rvalue.analyzed.info.tuple_type
+ ):
+ return True # This is a valid and analyzed named tuple definition, nothing to do here.
  if len(s.lvalues) != 1 or not isinstance(s.lvalues[0], (NameExpr, MemberExpr)):
  return False
  lvalue = s.lvalues[0]
@@ -3028,6 +3031,9 @@ def check_and_set_up_type_alias(self, s: AssignmentStmt) -> bool:
  # unless using PEP 613 `cls: TypeAlias = A`
  return False
 
+ if isinstance(s.rvalue, CallExpr) and s.rvalue.analyzed:
+ return False
+
  existing = self.current_symbol_table().get(lvalue.name)
  # Third rule: type aliases can't be re-defined. For example:
  # A: Type[float] = int
@@ -3157,9 +3163,8 @@ def check_and_set_up_type_alias(self, s: AssignmentStmt) -> bool:
  self.cannot_resolve_name(lvalue.name, "name", s)
  return True
  else:
- self.progress = True
  # We need to defer so that this change can get propagated to base classes.
- self.defer(s)
+ self.defer(s, force_progress=True)
  else:
  self.add_symbol(lvalue.name, alias_node, s)
  if isinstance(rvalue, RefExpr) and isinstance(rvalue.node, TypeAlias):
@@ -5484,7 +5489,7 @@ def tvar_scope_frame(self, frame: TypeVarLikeScope) -> Iterator[None]:
  yield
  self.tvar_scope = old_scope
 
- def defer(self, debug_context: Optional[Context] = None) -> None:
+ def defer(self, debug_context: Optional[Context] = None, force_progress: bool = False) -> None:
  """Defer current analysis target to be analyzed again.
 
  This must be called if something in the current target is
@@ -5498,6 +5503,8 @@ def defer(self, debug_context: Optional[Context] = None) -> None:
  They are usually preferable to a direct defer() call.
  """
  assert not self.final_iteration, "Must not defer during final iteration"
+ if force_progress:
+ self.progress = True
  self.deferred = True
  # Store debug info for this deferral.
  line = (
@@ -5999,19 +6006,6 @@ def is_future_flag_set(self, flag: str) -> bool:
  return self.modules[self.cur_mod_id].is_future_flag_set(flag)
 
 
-class HasPlaceholders(TypeQuery[bool]):
- def __init__(self) -> None:
- super().__init__(any)
-
- def visit_placeholder_type(self, t: PlaceholderType) -> bool:
- return True
-
-
-def has_placeholder(typ: Type) -> bool:
- """Check if a type contains any placeholder types (recursively)."""
- return typ.accept(HasPlaceholders())
-
-
 def replace_implicit_first_type(sig: FunctionLike, new: Type) -> FunctionLike:
  if isinstance(sig, CallableType):
  if len(sig.arg_types) == 0:

mypy/semanal_namedtuple.py

@@ -44,6 +44,7 @@
  PRIORITY_FALLBACKS,
  SemanticAnalyzerInterface,
  calculate_tuple_fallback,
+ has_placeholder,
  set_callable_name,
 )
 from mypy.types import (
@@ -109,8 +110,11 @@ def analyze_namedtuple_classdef(
  items, types, default_items = result
  if is_func_scope and "@" not in defn.name:
  defn.name += "@" + str(defn.line)
+ existing_info = None
+ if isinstance(defn.analyzed, NamedTupleExpr):
+ existing_info = defn.analyzed.info
  info = self.build_namedtuple_typeinfo(
- defn.name, items, types, default_items, defn.line
+ defn.name, items, types, default_items, defn.line, existing_info
  )
  defn.info = info
  defn.analyzed = NamedTupleExpr(info, is_typed=True)
@@ -164,7 +168,14 @@ def check_namedtuple_classdef(
  if stmt.type is None:
  types.append(AnyType(TypeOfAny.unannotated))
  else:
- analyzed = self.api.anal_type(stmt.type)
+ # We never allow recursive types at function scope. Although it is
+ # possible to support this for named tuples, it is still tricky, and
+ # it would be inconsistent with type aliases.
+ analyzed = self.api.anal_type(
+ stmt.type,
+ allow_placeholder=self.options.enable_recursive_aliases
+ and not self.api.is_func_scope(),
+ )
  if analyzed is None:
  # Something is incomplete. We need to defer this named tuple.
  return None
@@ -226,7 +237,7 @@ def check_namedtuple(
  name += "@" + str(call.line)
  else:
  name = var_name = "namedtuple@" + str(call.line)
- info = self.build_namedtuple_typeinfo(name, [], [], {}, node.line)
+ info = self.build_namedtuple_typeinfo(name, [], [], {}, node.line, None)
  self.store_namedtuple_info(info, var_name, call, is_typed)
  if name != var_name or is_func_scope:
  # NOTE: we skip local namespaces since they are not serialized.
@@ -262,12 +273,22 @@ def check_namedtuple(
  }
  else:
  default_items = {}
- info = self.build_namedtuple_typeinfo(name, items, types, default_items, node.line)
+
+ existing_info = None
+ if isinstance(node.analyzed, NamedTupleExpr):
+ existing_info = node.analyzed.info
+ info = self.build_namedtuple_typeinfo(
+ name, items, types, default_items, node.line, existing_info
+ )
+
  # If var_name is not None (i.e. this is not a base class expression), we always
  # store the generated TypeInfo under var_name in the current scope, so that
  # other definitions can use it.
  if var_name:
  self.store_namedtuple_info(info, var_name, call, is_typed)
+ else:
+ call.analyzed = NamedTupleExpr(info, is_typed=is_typed)
+ call.analyzed.set_line(call)
  # There are three cases where we need to store the generated TypeInfo
  # second time (for the purpose of serialization):
  # * If there is a name mismatch like One = NamedTuple('Other', [...])
@@ -408,7 +429,12 @@ def parse_namedtuple_fields_with_types(
  except TypeTranslationError:
  self.fail("Invalid field type", type_node)
  return None
- analyzed = self.api.anal_type(type)
+ # We never allow recursive types at function scope.
+ analyzed = self.api.anal_type(
+ type,
+ allow_placeholder=self.options.enable_recursive_aliases
+ and not self.api.is_func_scope(),
+ )
  # Workaround #4987 and avoid introducing a bogus UnboundType
  if isinstance(analyzed, UnboundType):
  analyzed = AnyType(TypeOfAny.from_error)
@@ -428,6 +454,7 @@ def build_namedtuple_typeinfo(
  types: List[Type],
  default_items: Mapping[str, Expression],
  line: int,
+ existing_info: Optional[TypeInfo],
  ) -> TypeInfo:
  strtype = self.api.named_type("builtins.str")
  implicit_any = AnyType(TypeOfAny.special_form)
@@ -448,18 +475,23 @@ def build_namedtuple_typeinfo(
  literals: List[Type] = [LiteralType(item, strtype) for item in items]
  match_args_type = TupleType(literals, basetuple_type)
 
- info = self.api.basic_new_typeinfo(name, fallback, line)
+ info = existing_info or self.api.basic_new_typeinfo(name, fallback, line)
  info.is_named_tuple = True
  tuple_base = TupleType(types, fallback)
- info.tuple_type = tuple_base
+ if info.tuple_alias and has_placeholder(info.tuple_alias.target):
+ self.api.defer(force_progress=True)
+ info.update_tuple_type(tuple_base)
  info.line = line
  # For use by mypyc.
  info.metadata["namedtuple"] = {"fields": items.copy()}
 
  # We can't calculate the complete fallback type until after semantic
  # analysis, since otherwise base classes might be incomplete. Postpone a
  # callback function that patches the fallback.
- self.api.schedule_patch(PRIORITY_FALLBACKS, lambda: calculate_tuple_fallback(tuple_base))
+ if not has_placeholder(tuple_base):
+ self.api.schedule_patch(
+ PRIORITY_FALLBACKS, lambda: calculate_tuple_fallback(tuple_base)
+ )
 
  def add_field(
  var: Var, is_initialized_in_class: bool = False, is_property: bool = False
@@ -489,6 +521,7 @@ def add_field(
  if self.options.python_version >= (3, 10):
  add_field(Var("__match_args__", match_args_type), is_initialized_in_class=True)
 
+ assert info.tuple_type is not None # Set by update_tuple_type() above.
  tvd = TypeVarType(
  SELF_TVAR_NAME, info.fullname + "." + SELF_TVAR_NAME, -1, [], info.tuple_type
  )