Re-work indirect dependencies

mypy/build.py

@@ -47,7 +47,7 @@
 from mypy.graph_utils import prepare_sccs, strongly_connected_components, topsort
 from mypy.indirection import TypeIndirectionVisitor
 from mypy.messages import MessageBuilder
-from mypy.nodes import Import, ImportAll, ImportBase, ImportFrom, MypyFile, SymbolTable, TypeInfo
+from mypy.nodes import Import, ImportAll, ImportBase, ImportFrom, MypyFile, SymbolTable
 from mypy.partially_defined import PossiblyUndefinedVariableVisitor
 from mypy.semanal import SemanticAnalyzer
 from mypy.semanal_pass1 import SemanticAnalyzerPreAnalysis
@@ -1765,26 +1765,24 @@ def delete_cache(id: str, path: str, manager: BuildManager) -> None:
 
 For single nodes, processing is simple.  If the node was cached, we
 deserialize the cache data and fix up cross-references.  Otherwise, we
-do semantic analysis followed by type checking.  We also handle (c)
-above; if a module has valid cache data *but* any of its
-dependencies was processed from source, then the module should be
-processed from source.
-
-A relatively simple optimization (outside SCCs) we might do in the
-future is as follows: if a node's cache data is valid, but one or more
-of its dependencies are out of date so we have to re-parse the node
-from source, once we have fully type-checked the node, we can decide
-whether its symbol table actually changed compared to the cache data
-(by reading the cache data and comparing it to the data we would be
-writing).  If there is no change we can declare the node up to date,
-and any node that depends (and for which we have cached data, and
-whose other dependencies are up to date) on it won't need to be
-re-parsed from source.
+do semantic analysis followed by type checking.  Once we (re-)processed
+an SCC we check whether its interface (symbol table) is still fresh
+(matches previous cached value). If it is not, we consider dependent SCCs
+stale so that they need to be re-parsed as well.
+
+Note on indirect dependencies: normally dependencies are determined from
+imports, but since our interfaces are "opaque" (i.e. symbol tables can
+contain cross-references as well as types identified by name), these are not
+enough.  We *must* also add "indirect" dependencies from symbols and types to
+their definitions.  For this purpose, we record all accessed symbols during
+semantic analysis, and after we finished processing a module, we traverse its
+type map, and for each type we find (transitively) on which named types it
+depends.
 
 Import cycles
 -------------
 
-Finally we have to decide how to handle (c), import cycles.  Here
+Finally we have to decide how to handle (b), import cycles.  Here
 we'll need a modified version of the original state machine
 (build.py), but we only need to do this per SCC, and we won't have to
 deal with changes to the list of nodes while we're processing it.
@@ -2409,21 +2407,15 @@ def finish_passes(self) -> None:
 
             # We should always patch indirect dependencies, even in full (non-incremental) builds,
             # because the cache still may be written, and it must be correct.
-            # TODO: find a more robust way to traverse *all* relevant types?
-            all_types = list(self.type_map().values())
-            for _, sym, _ in self.tree.local_definitions():
-                if sym.type is not None:
-                    all_types.append(sym.type)
-                if isinstance(sym.node, TypeInfo):
-                    # TypeInfo symbols have some extra relevant types.
-                    all_types.extend(sym.node.bases)
-                    if sym.node.metaclass_type:
-                        all_types.append(sym.node.metaclass_type)
-                    if sym.node.typeddict_type:
-                        all_types.append(sym.node.typeddict_type)
-                    if sym.node.tuple_type:
-                        all_types.append(sym.node.tuple_type)
-            self._patch_indirect_dependencies(self.type_checker().module_refs, all_types)
+            self._patch_indirect_dependencies(
+                # Two possible sources of indirect dependencies:
+                # * Symbols not directly imported in this module but accessed via an attribute
+                #   or via a re-export (vast majority of these recorded in semantic analysis).
+                # * For each expression type we need to record definitions of type components
+                #   since "meaning" of the type may be updated when definitions are updated.
+                self.tree.module_refs | self.type_checker().module_refs,
+                set(self.type_map().values()),
+            )
 
             if self.options.dump_inference_stats:
                 dump_type_stats(
@@ -2452,7 +2444,7 @@ def free_state(self) -> None:
             self._type_checker.reset()
             self._type_checker = None
 
-    def _patch_indirect_dependencies(self, module_refs: set[str], types: list[Type]) -> None:
+    def _patch_indirect_dependencies(self, module_refs: set[str], types: set[Type]) -> None:
         assert None not in types
         valid = self.valid_references()
 

mypy/checker.py

@@ -378,11 +378,9 @@ class TypeChecker(NodeVisitor[None], TypeCheckerSharedApi):
     inferred_attribute_types: dict[Var, Type] | None = None
     # Don't infer partial None types if we are processing assignment from Union
     no_partial_types: bool = False
-
-    # The set of all dependencies (suppressed or not) that this module accesses, either
-    # directly or indirectly.
+    # Extra module references not detected during semantic analysis (these are rare cases
+    # e.g. access to class-level import via instance).
     module_refs: set[str]
-
     # A map from variable nodes to a snapshot of the frame ids of the
     # frames that were active when the variable was declared. This can
     # be used to determine nearest common ancestor frame of a variable's

mypy/checkexpr.py

@@ -198,7 +198,6 @@
 )
 from mypy.typestate import type_state
 from mypy.typevars import fill_typevars
-from mypy.util import split_module_names
 from mypy.visitor import ExpressionVisitor
 
 # Type of callback user for checking individual function arguments. See
@@ -248,36 +247,6 @@ def allow_fast_container_literal(t: Type) -> bool:
     )
 
 
-def extract_refexpr_names(expr: RefExpr, output: set[str]) -> None:
-    """Recursively extracts all module references from a reference expression.
-
-    Note that currently, the only two subclasses of RefExpr are NameExpr and
-    MemberExpr."""
-    while isinstance(expr.node, MypyFile) or expr.fullname:
-        if isinstance(expr.node, MypyFile) and expr.fullname:
-            # If it's None, something's wrong (perhaps due to an
-            # import cycle or a suppressed error).  For now we just
-            # skip it.
-            output.add(expr.fullname)
-
-        if isinstance(expr, NameExpr):
-            is_suppressed_import = isinstance(expr.node, Var) and expr.node.is_suppressed_import
-            if isinstance(expr.node, TypeInfo):
-                # Reference to a class or a nested class
-                output.update(split_module_names(expr.node.module_name))
-            elif "." in expr.fullname and not is_suppressed_import:
-                # Everything else (that is not a silenced import within a class)
-                output.add(expr.fullname.rsplit(".", 1)[0])
-            break
-        elif isinstance(expr, MemberExpr):
-            if isinstance(expr.expr, RefExpr):
-                expr = expr.expr
-            else:
-                break
-        else:
-            raise AssertionError(f"Unknown RefExpr subclass: {type(expr)}")
-
-
 class Finished(Exception):
     """Raised if we can terminate overload argument check early (no match)."""
 
@@ -370,7 +339,6 @@ def visit_name_expr(self, e: NameExpr) -> Type:
 
         It can be of any kind: local, member or global.
         """
-        extract_refexpr_names(e, self.chk.module_refs)
         result = self.analyze_ref_expr(e)
         narrowed = self.narrow_type_from_binder(e, result)
         self.chk.check_deprecated(e.node, e)
@@ -3344,7 +3312,6 @@ def check_union_call(
 
     def visit_member_expr(self, e: MemberExpr, is_lvalue: bool = False) -> Type:
         """Visit member expression (of form e.id)."""
-        extract_refexpr_names(e, self.chk.module_refs)
         result = self.analyze_ordinary_member_access(e, is_lvalue)
         narrowed = self.narrow_type_from_binder(e, result)
         self.chk.warn_deprecated(e.node, e)

mypy/checkmember.py

@@ -543,6 +543,8 @@ def analyze_member_var_access(
     if isinstance(v, FuncDef):
         assert False, "Did not expect a function"
     if isinstance(v, MypyFile):
+        # Special case: accessing module on instances is allowed, but will not
+        # be recorded by semantic analyzer.
         mx.chk.module_refs.add(v.fullname)
 
     if isinstance(vv, (TypeInfo, TypeAlias, MypyFile, TypeVarLikeExpr)):

mypy/fixup.py

@@ -441,4 +441,4 @@ def missing_info(modules: dict[str, MypyFile]) -> TypeInfo:
 
 def missing_alias() -> TypeAlias:
     suggestion = _SUGGESTION.format("alias")
-    return TypeAlias(AnyType(TypeOfAny.special_form), suggestion, line=-1, column=-1)
+    return TypeAlias(AnyType(TypeOfAny.special_form), suggestion, "<missing>", line=-1, column=-1)

mypy/indirection.py

@@ -4,17 +4,6 @@
 
 import mypy.types as types
 from mypy.types import TypeVisitor
-from mypy.util import split_module_names
-
-
-def extract_module_names(type_name: str | None) -> list[str]:
-    """Returns the module names of a fully qualified type name."""
-    if type_name is not None:
-        # Discard the first one, which is just the qualified name of the type
-        possible_module_names = split_module_names(type_name)
-        return possible_module_names[1:]
-    else:
-        return []
 
 
 class TypeIndirectionVisitor(TypeVisitor[None]):
@@ -23,49 +12,57 @@ class TypeIndirectionVisitor(TypeVisitor[None]):
     def __init__(self) -> None:
         # Module references are collected here
         self.modules: set[str] = set()
-        # User to avoid infinite recursion with recursive type aliases
-        self.seen_aliases: set[types.TypeAliasType] = set()
-        # Used to avoid redundant work
-        self.seen_fullnames: set[str] = set()
+        # User to avoid infinite recursion with recursive types
+        self.seen_types: set[types.TypeAliasType | types.Instance] = set()
 
     def find_modules(self, typs: Iterable[types.Type]) -> set[str]:
         self.modules = set()
-        self.seen_fullnames = set()
-        self.seen_aliases = set()
+        self.seen_types = set()
         for typ in typs:
             self._visit(typ)
         return self.modules
 
     def _visit(self, typ: types.Type) -> None:
-        if isinstance(typ, types.TypeAliasType):
-            # Avoid infinite recursion for recursive type aliases.
-            self.seen_aliases.add(typ)
+        # Note: instances are needed for `class str(Sequence[str]): ...`
+        if (
+            isinstance(typ, types.TypeAliasType)
+            or isinstance(typ, types.ProperType)
+            and isinstance(typ, types.Instance)
+        ):
+            # Avoid infinite recursion for recursive types.
+            if typ in self.seen_types:
+                return
+            self.seen_types.add(typ)
         typ.accept(self)
 
     def _visit_type_tuple(self, typs: tuple[types.Type, ...]) -> None:
         # Micro-optimization: Specialized version of _visit for lists
         for typ in typs:
-            if isinstance(typ, types.TypeAliasType):
-                # Avoid infinite recursion for recursive type aliases.
-                if typ in self.seen_aliases:
+            if (
+                isinstance(typ, types.TypeAliasType)
+                or isinstance(typ, types.ProperType)
+                and isinstance(typ, types.Instance)
+            ):
+                # Avoid infinite recursion for recursive types.
+                if typ in self.seen_types:
                     continue
-                self.seen_aliases.add(typ)
+                self.seen_types.add(typ)
             typ.accept(self)
 
     def _visit_type_list(self, typs: list[types.Type]) -> None:
         # Micro-optimization: Specialized version of _visit for tuples
         for typ in typs:
-            if isinstance(typ, types.TypeAliasType):
-                # Avoid infinite recursion for recursive type aliases.
-                if typ in self.seen_aliases:
+            if (
+                isinstance(typ, types.TypeAliasType)
+                or isinstance(typ, types.ProperType)
+                and isinstance(typ, types.Instance)
+            ):
+                # Avoid infinite recursion for recursive types.
+                if typ in self.seen_types:
                     continue
-                self.seen_aliases.add(typ)
+                self.seen_types.add(typ)
             typ.accept(self)
 
-    def _visit_module_name(self, module_name: str) -> None:
-        if module_name not in self.modules:
-            self.modules.update(split_module_names(module_name))
-
     def visit_unbound_type(self, t: types.UnboundType) -> None:
         self._visit_type_tuple(t.args)
 
@@ -105,27 +102,36 @@ def visit_parameters(self, t: types.Parameters) -> None:
         self._visit_type_list(t.arg_types)
 
     def visit_instance(self, t: types.Instance) -> None:
+        # Instance is named, record its definition and continue digging into
+        # components that constitute semantic meaning of this type: bases, metaclass,
+        # tuple type, and typeddict type.
+        # Note: we cannot simply record the MRO, in case an intermediate base contains
+        # a reference to type alias, this affects meaning of map_instance_to_supertype(),
+        # see e.g. testDoubleReexportGenericUpdated.
         self._visit_type_tuple(t.args)
         if t.type:
-            # Uses of a class depend on everything in the MRO,
-            # as changes to classes in the MRO can add types to methods,
-            # change property types, change the MRO itself, etc.
+            # Important optimization: instead of simply recording the definition and
+            # recursing into bases, record the MRO and only traverse generic bases.
             for s in t.type.mro:
-                self._visit_module_name(s.module_name)
-            if t.type.metaclass_type is not None:
-                self._visit_module_name(t.type.metaclass_type.type.module_name)
+                self.modules.add(s.module_name)
+                for base in s.bases:
+                    if base.args:
+                        self._visit_type_tuple(base.args)
+            if t.type.metaclass_type:
+                self._visit(t.type.metaclass_type)
+            if t.type.typeddict_type:
+                self._visit(t.type.typeddict_type)
+            if t.type.tuple_type:
+                self._visit(t.type.tuple_type)
 
     def visit_callable_type(self, t: types.CallableType) -> None:
         self._visit_type_list(t.arg_types)
         self._visit(t.ret_type)
-        if t.definition is not None:
-            fullname = t.definition.fullname
-            if fullname not in self.seen_fullnames:
-                self.modules.update(extract_module_names(t.definition.fullname))
-                self.seen_fullnames.add(fullname)
+        self._visit_type_tuple(t.variables)
 
     def visit_overloaded(self, t: types.Overloaded) -> None:
-        self._visit_type_list(list(t.items))
+        for item in t.items:
+            self._visit(item)
         self._visit(t.fallback)
 
     def visit_tuple_type(self, t: types.TupleType) -> None:
@@ -149,4 +155,9 @@ def visit_type_type(self, t: types.TypeType) -> None:
         self._visit(t.item)
 
     def visit_type_alias_type(self, t: types.TypeAliasType) -> None:
-        self._visit(types.get_proper_type(t))
+        # Type alias is named, record its definition and continue digging into
+        # components that constitute semantic meaning of this type: target and args.
+        if t.alias:
+            self.modules.add(t.alias.module)
+            self._visit(t.alias.target)
+        self._visit_type_list(t.args)