Add files using upload-large-folder tool

.gitattributes

@@ -196,3 +196,4 @@ tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/_inductor/_
 .venv/lib/python3.11/site-packages/google/ai/generativelanguage_v1alpha/services/retriever_service/transports/__pycache__/rest.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
 .venv/lib/python3.11/site-packages/google/ai/generativelanguage_v1beta/services/retriever_service/transports/__pycache__/rest.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
 .venv/lib/python3.11/site-packages/google/ai/generativelanguage_v1beta/services/retriever_service/__pycache__/client.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
+.venv/lib/python3.11/site-packages/google/_upb/_message.abi3.so filter=lfs diff=lfs merge=lfs -text

.venv/lib/python3.11/site-packages/google/_upb/_message.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34d5e690ce9cc9436d760939742d6df1a0afb258df8a839fe7ff48c498095f30
+size 390920

.venv/lib/python3.11/site-packages/google/api/auth_pb2.py

@@ -0,0 +1,52 @@
+# -*- coding: utf-8 -*-
+
+# Copyright 2024 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Generated by the protocol buffer compiler.  DO NOT EDIT!
+# source: google/api/auth.proto
+"""Generated protocol buffer code."""
+from google.protobuf import descriptor as _descriptor
+from google.protobuf import descriptor_pool as _descriptor_pool
+from google.protobuf import symbol_database as _symbol_database
+from google.protobuf.internal import builder as _builder
+
+# @@protoc_insertion_point(imports)
+
+_sym_db = _symbol_database.Default()
+
+
+DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(
+    b'\n\x15google/api/auth.proto\x12\ngoogle.api"l\n\x0e\x41uthentication\x12-\n\x05rules\x18\x03 \x03(\x0b\x32\x1e.google.api.AuthenticationRule\x12+\n\tproviders\x18\x04 \x03(\x0b\x32\x18.google.api.AuthProvider"\xa9\x01\n\x12\x41uthenticationRule\x12\x10\n\x08selector\x18\x01 \x01(\t\x12,\n\x05oauth\x18\x02 \x01(\x0b\x32\x1d.google.api.OAuthRequirements\x12 \n\x18\x61llow_without_credential\x18\x05 \x01(\x08\x12\x31\n\x0crequirements\x18\x07 \x03(\x0b\x32\x1b.google.api.AuthRequirement"^\n\x0bJwtLocation\x12\x10\n\x06header\x18\x01 \x01(\tH\x00\x12\x0f\n\x05query\x18\x02 \x01(\tH\x00\x12\x10\n\x06\x63ookie\x18\x04 \x01(\tH\x00\x12\x14\n\x0cvalue_prefix\x18\x03 \x01(\tB\x04\n\x02in"\x9a\x01\n\x0c\x41uthProvider\x12\n\n\x02id\x18\x01 \x01(\t\x12\x0e\n\x06issuer\x18\x02 \x01(\t\x12\x10\n\x08jwks_uri\x18\x03 \x01(\t\x12\x11\n\taudiences\x18\x04 \x01(\t\x12\x19\n\x11\x61uthorization_url\x18\x05 \x01(\t\x12.\n\rjwt_locations\x18\x06 \x03(\x0b\x32\x17.google.api.JwtLocation"-\n\x11OAuthRequirements\x12\x18\n\x10\x63\x61nonical_scopes\x18\x01 \x01(\t"9\n\x0f\x41uthRequirement\x12\x13\n\x0bprovider_id\x18\x01 \x01(\t\x12\x11\n\taudiences\x18\x02 \x01(\tBk\n\x0e\x63om.google.apiB\tAuthProtoP\x01ZEgoogle.golang.org/genproto/googleapis/api/serviceconfig;serviceconfig\xa2\x02\x04GAPIb\x06proto3'
+)
+
+_globals = globals()
+_builder.BuildMessageAndEnumDescriptors(DESCRIPTOR, _globals)
+_builder.BuildTopDescriptorsAndMessages(DESCRIPTOR, "google.api.auth_pb2", _globals)
+if _descriptor._USE_C_DESCRIPTORS == False:
+    DESCRIPTOR._options = None
+    DESCRIPTOR._serialized_options = b"\n\016com.google.apiB\tAuthProtoP\001ZEgoogle.golang.org/genproto/googleapis/api/serviceconfig;serviceconfig\242\002\004GAPI"
+    _globals["_AUTHENTICATION"]._serialized_start = 37
+    _globals["_AUTHENTICATION"]._serialized_end = 145
+    _globals["_AUTHENTICATIONRULE"]._serialized_start = 148
+    _globals["_AUTHENTICATIONRULE"]._serialized_end = 317
+    _globals["_JWTLOCATION"]._serialized_start = 319
+    _globals["_JWTLOCATION"]._serialized_end = 413
+    _globals["_AUTHPROVIDER"]._serialized_start = 416
+    _globals["_AUTHPROVIDER"]._serialized_end = 570
+    _globals["_OAUTHREQUIREMENTS"]._serialized_start = 572
+    _globals["_OAUTHREQUIREMENTS"]._serialized_end = 617
+    _globals["_AUTHREQUIREMENT"]._serialized_start = 619
+    _globals["_AUTHREQUIREMENT"]._serialized_end = 676
+# @@protoc_insertion_point(module_scope)

.venv/lib/python3.11/site-packages/google/api/system_parameter.proto

@@ -0,0 +1,96 @@
+// Copyright 2024 Google LLC
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto3";
+
+package google.api;
+
+option go_package = "google.golang.org/genproto/googleapis/api/serviceconfig;serviceconfig";
+option java_multiple_files = true;
+option java_outer_classname = "SystemParameterProto";
+option java_package = "com.google.api";
+option objc_class_prefix = "GAPI";
+
+// ### System parameter configuration
+//
+// A system parameter is a special kind of parameter defined by the API
+// system, not by an individual API. It is typically mapped to an HTTP header
+// and/or a URL query parameter. This configuration specifies which methods
+// change the names of the system parameters.
+message SystemParameters {
+  // Define system parameters.
+  //
+  // The parameters defined here will override the default parameters
+  // implemented by the system. If this field is missing from the service
+  // config, default system parameters will be used. Default system parameters
+  // and names is implementation-dependent.
+  //
+  // Example: define api key for all methods
+  //
+  //     system_parameters
+  //       rules:
+  //         - selector: "*"
+  //           parameters:
+  //             - name: api_key
+  //               url_query_parameter: api_key
+  //
+  //
+  // Example: define 2 api key names for a specific method.
+  //
+  //     system_parameters
+  //       rules:
+  //         - selector: "/ListShelves"
+  //           parameters:
+  //             - name: api_key
+  //               http_header: Api-Key1
+  //             - name: api_key
+  //               http_header: Api-Key2
+  //
+  // **NOTE:** All service configuration rules follow "last one wins" order.
+  repeated SystemParameterRule rules = 1;
+}
+
+// Define a system parameter rule mapping system parameter definitions to
+// methods.
+message SystemParameterRule {
+  // Selects the methods to which this rule applies. Use '*' to indicate all
+  // methods in all APIs.
+  //
+  // Refer to [selector][google.api.DocumentationRule.selector] for syntax
+  // details.
+  string selector = 1;
+
+  // Define parameters. Multiple names may be defined for a parameter.
+  // For a given method call, only one of them should be used. If multiple
+  // names are used the behavior is implementation-dependent.
+  // If none of the specified names are present the behavior is
+  // parameter-dependent.
+  repeated SystemParameter parameters = 2;
+}
+
+// Define a parameter's name and location. The parameter may be passed as either
+// an HTTP header or a URL query parameter, and if both are passed the behavior
+// is implementation-dependent.
+message SystemParameter {
+  // Define the name of the parameter, such as "api_key" . It is case sensitive.
+  string name = 1;
+
+  // Define the HTTP header name to use for the parameter. It is case
+  // insensitive.
+  string http_header = 2;
+
+  // Define the URL query parameter name to use for the parameter. It is case
+  // sensitive.
+  string url_query_parameter = 3;
+}

.venv/lib/python3.11/site-packages/google/api/usage.proto

@@ -0,0 +1,96 @@
+// Copyright 2024 Google LLC
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto3";
+
+package google.api;
+
+option go_package = "google.golang.org/genproto/googleapis/api/serviceconfig;serviceconfig";
+option java_multiple_files = true;
+option java_outer_classname = "UsageProto";
+option java_package = "com.google.api";
+option objc_class_prefix = "GAPI";
+
+// Configuration controlling usage of a service.
+message Usage {
+  // Requirements that must be satisfied before a consumer project can use the
+  // service. Each requirement is of the form <service.name>/<requirement-id>;
+  // for example 'serviceusage.googleapis.com/billing-enabled'.
+  //
+  // For Google APIs, a Terms of Service requirement must be included here.
+  // Google Cloud APIs must include "serviceusage.googleapis.com/tos/cloud".
+  // Other Google APIs should include
+  // "serviceusage.googleapis.com/tos/universal". Additional ToS can be
+  // included based on the business needs.
+  repeated string requirements = 1;
+
+  // A list of usage rules that apply to individual API methods.
+  //
+  // **NOTE:** All service configuration rules follow "last one wins" order.
+  repeated UsageRule rules = 6;
+
+  // The full resource name of a channel used for sending notifications to the
+  // service producer.
+  //
+  // Google Service Management currently only supports
+  // [Google Cloud Pub/Sub](https://cloud.google.com/pubsub) as a notification
+  // channel. To use Google Cloud Pub/Sub as the channel, this must be the name
+  // of a Cloud Pub/Sub topic that uses the Cloud Pub/Sub topic name format
+  // documented in https://cloud.google.com/pubsub/docs/overview.
+  string producer_notification_channel = 7;
+}
+
+// Usage configuration rules for the service.
+//
+// NOTE: Under development.
+//
+//
+// Use this rule to configure unregistered calls for the service. Unregistered
+// calls are calls that do not contain consumer project identity.
+// (Example: calls that do not contain an API key).
+// By default, API methods do not allow unregistered calls, and each method call
+// must be identified by a consumer project identity. Use this rule to
+// allow/disallow unregistered calls.
+//
+// Example of an API that wants to allow unregistered calls for entire service.
+//
+//     usage:
+//       rules:
+//       - selector: "*"
+//         allow_unregistered_calls: true
+//
+// Example of a method that wants to allow unregistered calls.
+//
+//     usage:
+//       rules:
+//       - selector: "google.example.library.v1.LibraryService.CreateBook"
+//         allow_unregistered_calls: true
+message UsageRule {
+  // Selects the methods to which this rule applies. Use '*' to indicate all
+  // methods in all APIs.
+  //
+  // Refer to [selector][google.api.DocumentationRule.selector] for syntax
+  // details.
+  string selector = 1;
+
+  // If true, the selected method allows unregistered calls, e.g. calls
+  // that don't identify any user or application.
+  bool allow_unregistered_calls = 2;
+
+  // If true, the selected method should skip service control and the control
+  // plane features, such as quota and billing, will not be available.
+  // This flag is used by Google Cloud Endpoints to bypass checks for internal
+  // methods, such as service health check methods.
+  bool skip_service_control = 3;
+}

.venv/lib/python3.11/site-packages/jinja2/async_utils.py

@@ -0,0 +1,99 @@
+import inspect
+import typing as t
+from functools import WRAPPER_ASSIGNMENTS
+from functools import wraps
+
+from .utils import _PassArg
+from .utils import pass_eval_context
+
+if t.TYPE_CHECKING:
+    import typing_extensions as te
+
+V = t.TypeVar("V")
+
+
+def async_variant(normal_func):  # type: ignore
+    def decorator(async_func):  # type: ignore
+        pass_arg = _PassArg.from_obj(normal_func)
+        need_eval_context = pass_arg is None
+
+        if pass_arg is _PassArg.environment:
+
+            def is_async(args: t.Any) -> bool:
+                return t.cast(bool, args[0].is_async)
+
+        else:
+
+            def is_async(args: t.Any) -> bool:
+                return t.cast(bool, args[0].environment.is_async)
+
+        # Take the doc and annotations from the sync function, but the
+        # name from the async function. Pallets-Sphinx-Themes
+        # build_function_directive expects __wrapped__ to point to the
+        # sync function.
+        async_func_attrs = ("__module__", "__name__", "__qualname__")
+        normal_func_attrs = tuple(set(WRAPPER_ASSIGNMENTS).difference(async_func_attrs))
+
+        @wraps(normal_func, assigned=normal_func_attrs)
+        @wraps(async_func, assigned=async_func_attrs, updated=())
+        def wrapper(*args, **kwargs):  # type: ignore
+            b = is_async(args)
+
+            if need_eval_context:
+                args = args[1:]
+
+            if b:
+                return async_func(*args, **kwargs)
+
+            return normal_func(*args, **kwargs)
+
+        if need_eval_context:
+            wrapper = pass_eval_context(wrapper)
+
+        wrapper.jinja_async_variant = True  # type: ignore[attr-defined]
+        return wrapper
+
+    return decorator
+
+
+_common_primitives = {int, float, bool, str, list, dict, tuple, type(None)}
+
+
+async def auto_await(value: t.Union[t.Awaitable["V"], "V"]) -> "V":
+    # Avoid a costly call to isawaitable
+    if type(value) in _common_primitives:
+        return t.cast("V", value)
+
+    if inspect.isawaitable(value):
+        return await t.cast("t.Awaitable[V]", value)
+
+    return value
+
+
+class _IteratorToAsyncIterator(t.Generic[V]):
+    def __init__(self, iterator: "t.Iterator[V]"):
+        self._iterator = iterator
+
+    def __aiter__(self) -> "te.Self":
+        return self
+
+    async def __anext__(self) -> V:
+        try:
+            return next(self._iterator)
+        except StopIteration as e:
+            raise StopAsyncIteration(e.value) from e
+
+
+def auto_aiter(
+    iterable: "t.Union[t.AsyncIterable[V], t.Iterable[V]]",
+) -> "t.AsyncIterator[V]":
+    if hasattr(iterable, "__aiter__"):
+        return iterable.__aiter__()
+    else:
+        return _IteratorToAsyncIterator(iter(iterable))
+
+
+async def auto_to_list(
+    value: "t.Union[t.AsyncIterable[V], t.Iterable[V]]",
+) -> t.List["V"]:
+    return [x async for x in auto_aiter(value)]

.venv/lib/python3.11/site-packages/jinja2/compiler.py

@@ -0,0 +1,1998 @@
+"""Compiles nodes from the parser into Python code."""
+
+import typing as t
+from contextlib import contextmanager
+from functools import update_wrapper
+from io import StringIO
+from itertools import chain
+from keyword import iskeyword as is_python_keyword
+
+from markupsafe import escape
+from markupsafe import Markup
+
+from . import nodes
+from .exceptions import TemplateAssertionError
+from .idtracking import Symbols
+from .idtracking import VAR_LOAD_ALIAS
+from .idtracking import VAR_LOAD_PARAMETER
+from .idtracking import VAR_LOAD_RESOLVE
+from .idtracking import VAR_LOAD_UNDEFINED
+from .nodes import EvalContext
+from .optimizer import Optimizer
+from .utils import _PassArg
+from .utils import concat
+from .visitor import NodeVisitor
+
+if t.TYPE_CHECKING:
+    import typing_extensions as te
+
+    from .environment import Environment
+
+F = t.TypeVar("F", bound=t.Callable[..., t.Any])
+
+operators = {
+    "eq": "==",
+    "ne": "!=",
+    "gt": ">",
+    "gteq": ">=",
+    "lt": "<",
+    "lteq": "<=",
+    "in": "in",
+    "notin": "not in",
+}
+
+
+def optimizeconst(f: F) -> F:
+    def new_func(
+        self: "CodeGenerator", node: nodes.Expr, frame: "Frame", **kwargs: t.Any
+    ) -> t.Any:
+        # Only optimize if the frame is not volatile
+        if self.optimizer is not None and not frame.eval_ctx.volatile:
+            new_node = self.optimizer.visit(node, frame.eval_ctx)
+
+            if new_node != node:
+                return self.visit(new_node, frame)
+
+        return f(self, node, frame, **kwargs)
+
+    return update_wrapper(new_func, f)  # type: ignore[return-value]
+
+
+def _make_binop(op: str) -> t.Callable[["CodeGenerator", nodes.BinExpr, "Frame"], None]:
+    @optimizeconst
+    def visitor(self: "CodeGenerator", node: nodes.BinExpr, frame: Frame) -> None:
+        if (
+            self.environment.sandboxed and op in self.environment.intercepted_binops  # type: ignore
+        ):
+            self.write(f"environment.call_binop(context, {op!r}, ")
+            self.visit(node.left, frame)
+            self.write(", ")
+            self.visit(node.right, frame)
+        else:
+            self.write("(")
+            self.visit(node.left, frame)
+            self.write(f" {op} ")
+            self.visit(node.right, frame)
+
+        self.write(")")
+
+    return visitor
+
+
+def _make_unop(
+    op: str,
+) -> t.Callable[["CodeGenerator", nodes.UnaryExpr, "Frame"], None]:
+    @optimizeconst
+    def visitor(self: "CodeGenerator", node: nodes.UnaryExpr, frame: Frame) -> None:
+        if (
+            self.environment.sandboxed and op in self.environment.intercepted_unops  # type: ignore
+        ):
+            self.write(f"environment.call_unop(context, {op!r}, ")
+            self.visit(node.node, frame)
+        else:
+            self.write("(" + op)
+            self.visit(node.node, frame)
+
+        self.write(")")
+
+    return visitor
+
+
+def generate(
+    node: nodes.Template,
+    environment: "Environment",
+    name: t.Optional[str],
+    filename: t.Optional[str],
+    stream: t.Optional[t.TextIO] = None,
+    defer_init: bool = False,
+    optimized: bool = True,
+) -> t.Optional[str]:
+    """Generate the python source for a node tree."""
+    if not isinstance(node, nodes.Template):
+        raise TypeError("Can't compile non template nodes")
+
+    generator = environment.code_generator_class(
+        environment, name, filename, stream, defer_init, optimized
+    )
+    generator.visit(node)
+
+    if stream is None:
+        return generator.stream.getvalue()  # type: ignore
+
+    return None
+
+
+def has_safe_repr(value: t.Any) -> bool:
+    """Does the node have a safe representation?"""
+    if value is None or value is NotImplemented or value is Ellipsis:
+        return True
+
+    if type(value) in {bool, int, float, complex, range, str, Markup}:
+        return True
+
+    if type(value) in {tuple, list, set, frozenset}:
+        return all(has_safe_repr(v) for v in value)
+
+    if type(value) is dict:  # noqa E721
+        return all(has_safe_repr(k) and has_safe_repr(v) for k, v in value.items())
+
+    return False
+
+
+def find_undeclared(
+    nodes: t.Iterable[nodes.Node], names: t.Iterable[str]
+) -> t.Set[str]:
+    """Check if the names passed are accessed undeclared.  The return value
+    is a set of all the undeclared names from the sequence of names found.
+    """
+    visitor = UndeclaredNameVisitor(names)
+    try:
+        for node in nodes:
+            visitor.visit(node)
+    except VisitorExit:
+        pass
+    return visitor.undeclared
+
+
+class MacroRef:
+    def __init__(self, node: t.Union[nodes.Macro, nodes.CallBlock]) -> None:
+        self.node = node
+        self.accesses_caller = False
+        self.accesses_kwargs = False
+        self.accesses_varargs = False
+
+
+class Frame:
+    """Holds compile time information for us."""
+
+    def __init__(
+        self,
+        eval_ctx: EvalContext,
+        parent: t.Optional["Frame"] = None,
+        level: t.Optional[int] = None,
+    ) -> None:
+        self.eval_ctx = eval_ctx
+
+        # the parent of this frame
+        self.parent = parent
+
+        if parent is None:
+            self.symbols = Symbols(level=level)
+
+            # in some dynamic inheritance situations the compiler needs to add
+            # write tests around output statements.
+            self.require_output_check = False
+
+            # inside some tags we are using a buffer rather than yield statements.
+            # this for example affects {% filter %} or {% macro %}.  If a frame
+            # is buffered this variable points to the name of the list used as
+            # buffer.
+            self.buffer: t.Optional[str] = None
+
+            # the name of the block we're in, otherwise None.
+            self.block: t.Optional[str] = None
+
+        else:
+            self.symbols = Symbols(parent.symbols, level=level)
+            self.require_output_check = parent.require_output_check
+            self.buffer = parent.buffer
+            self.block = parent.block
+
+        # a toplevel frame is the root + soft frames such as if conditions.
+        self.toplevel = False
+
+        # the root frame is basically just the outermost frame, so no if
+        # conditions.  This information is used to optimize inheritance
+        # situations.
+        self.rootlevel = False
+
+        # variables set inside of loops and blocks should not affect outer frames,
+        # but they still needs to be kept track of as part of the active context.
+        self.loop_frame = False
+        self.block_frame = False
+
+        # track whether the frame is being used in an if-statement or conditional
+        # expression as it determines which errors should be raised during runtime
+        # or compile time.
+        self.soft_frame = False
+
+    def copy(self) -> "te.Self":
+        """Create a copy of the current one."""
+        rv = object.__new__(self.__class__)
+        rv.__dict__.update(self.__dict__)
+        rv.symbols = self.symbols.copy()
+        return rv
+
+    def inner(self, isolated: bool = False) -> "Frame":
+        """Return an inner frame."""
+        if isolated:
+            return Frame(self.eval_ctx, level=self.symbols.level + 1)
+        return Frame(self.eval_ctx, self)
+
+    def soft(self) -> "te.Self":
+        """Return a soft frame.  A soft frame may not be modified as
+        standalone thing as it shares the resources with the frame it
+        was created of, but it's not a rootlevel frame any longer.
+
+        This is only used to implement if-statements and conditional
+        expressions.
+        """
+        rv = self.copy()
+        rv.rootlevel = False
+        rv.soft_frame = True
+        return rv
+
+    __copy__ = copy
+
+
+class VisitorExit(RuntimeError):
+    """Exception used by the `UndeclaredNameVisitor` to signal a stop."""
+
+
+class DependencyFinderVisitor(NodeVisitor):
+    """A visitor that collects filter and test calls."""
+
+    def __init__(self) -> None:
+        self.filters: t.Set[str] = set()
+        self.tests: t.Set[str] = set()
+
+    def visit_Filter(self, node: nodes.Filter) -> None:
+        self.generic_visit(node)
+        self.filters.add(node.name)
+
+    def visit_Test(self, node: nodes.Test) -> None:
+        self.generic_visit(node)
+        self.tests.add(node.name)
+
+    def visit_Block(self, node: nodes.Block) -> None:
+        """Stop visiting at blocks."""
+
+
+class UndeclaredNameVisitor(NodeVisitor):
+    """A visitor that checks if a name is accessed without being
+    declared.  This is different from the frame visitor as it will
+    not stop at closure frames.
+    """
+
+    def __init__(self, names: t.Iterable[str]) -> None:
+        self.names = set(names)
+        self.undeclared: t.Set[str] = set()
+
+    def visit_Name(self, node: nodes.Name) -> None:
+        if node.ctx == "load" and node.name in self.names:
+            self.undeclared.add(node.name)
+            if self.undeclared == self.names:
+                raise VisitorExit()
+        else:
+            self.names.discard(node.name)
+
+    def visit_Block(self, node: nodes.Block) -> None:
+        """Stop visiting a blocks."""
+
+
+class CompilerExit(Exception):
+    """Raised if the compiler encountered a situation where it just
+    doesn't make sense to further process the code.  Any block that
+    raises such an exception is not further processed.
+    """
+
+
+class CodeGenerator(NodeVisitor):
+    def __init__(
+        self,
+        environment: "Environment",
+        name: t.Optional[str],
+        filename: t.Optional[str],
+        stream: t.Optional[t.TextIO] = None,
+        defer_init: bool = False,
+        optimized: bool = True,
+    ) -> None:
+        if stream is None:
+            stream = StringIO()
+        self.environment = environment
+        self.name = name
+        self.filename = filename
+        self.stream = stream
+        self.created_block_context = False
+        self.defer_init = defer_init
+        self.optimizer: t.Optional[Optimizer] = None
+
+        if optimized:
+            self.optimizer = Optimizer(environment)
+
+        # aliases for imports
+        self.import_aliases: t.Dict[str, str] = {}
+
+        # a registry for all blocks.  Because blocks are moved out
+        # into the global python scope they are registered here
+        self.blocks: t.Dict[str, nodes.Block] = {}
+
+        # the number of extends statements so far
+        self.extends_so_far = 0
+
+        # some templates have a rootlevel extends.  In this case we
+        # can safely assume that we're a child template and do some
+        # more optimizations.
+        self.has_known_extends = False
+
+        # the current line number
+        self.code_lineno = 1
+
+        # registry of all filters and tests (global, not block local)
+        self.tests: t.Dict[str, str] = {}
+        self.filters: t.Dict[str, str] = {}
+
+        # the debug information
+        self.debug_info: t.List[t.Tuple[int, int]] = []
+        self._write_debug_info: t.Optional[int] = None
+
+        # the number of new lines before the next write()
+        self._new_lines = 0
+
+        # the line number of the last written statement
+        self._last_line = 0
+
+        # true if nothing was written so far.
+        self._first_write = True
+
+        # used by the `temporary_identifier` method to get new
+        # unique, temporary identifier
+        self._last_identifier = 0
+
+        # the current indentation
+        self._indentation = 0
+
+        # Tracks toplevel assignments
+        self._assign_stack: t.List[t.Set[str]] = []
+
+        # Tracks parameter definition blocks
+        self._param_def_block: t.List[t.Set[str]] = []
+
+        # Tracks the current context.
+        self._context_reference_stack = ["context"]
+
+    @property
+    def optimized(self) -> bool:
+        return self.optimizer is not None
+
+    # -- Various compilation helpers
+
+    def fail(self, msg: str, lineno: int) -> "te.NoReturn":
+        """Fail with a :exc:`TemplateAssertionError`."""
+        raise TemplateAssertionError(msg, lineno, self.name, self.filename)
+
+    def temporary_identifier(self) -> str:
+        """Get a new unique identifier."""
+        self._last_identifier += 1
+        return f"t_{self._last_identifier}"
+
+    def buffer(self, frame: Frame) -> None:
+        """Enable buffering for the frame from that point onwards."""
+        frame.buffer = self.temporary_identifier()
+        self.writeline(f"{frame.buffer} = []")
+
+    def return_buffer_contents(
+        self, frame: Frame, force_unescaped: bool = False
+    ) -> None:
+        """Return the buffer contents of the frame."""
+        if not force_unescaped:
+            if frame.eval_ctx.volatile:
+                self.writeline("if context.eval_ctx.autoescape:")
+                self.indent()
+                self.writeline(f"return Markup(concat({frame.buffer}))")
+                self.outdent()
+                self.writeline("else:")
+                self.indent()
+                self.writeline(f"return concat({frame.buffer})")
+                self.outdent()
+                return
+            elif frame.eval_ctx.autoescape:
+                self.writeline(f"return Markup(concat({frame.buffer}))")
+                return
+        self.writeline(f"return concat({frame.buffer})")
+
+    def indent(self) -> None:
+        """Indent by one."""
+        self._indentation += 1
+
+    def outdent(self, step: int = 1) -> None:
+        """Outdent by step."""
+        self._indentation -= step
+
+    def start_write(self, frame: Frame, node: t.Optional[nodes.Node] = None) -> None:
+        """Yield or write into the frame buffer."""
+        if frame.buffer is None:
+            self.writeline("yield ", node)
+        else:
+            self.writeline(f"{frame.buffer}.append(", node)
+
+    def end_write(self, frame: Frame) -> None:
+        """End the writing process started by `start_write`."""
+        if frame.buffer is not None:
+            self.write(")")
+
+    def simple_write(
+        self, s: str, frame: Frame, node: t.Optional[nodes.Node] = None
+    ) -> None:
+        """Simple shortcut for start_write + write + end_write."""
+        self.start_write(frame, node)
+        self.write(s)
+        self.end_write(frame)
+
+    def blockvisit(self, nodes: t.Iterable[nodes.Node], frame: Frame) -> None:
+        """Visit a list of nodes as block in a frame.  If the current frame
+        is no buffer a dummy ``if 0: yield None`` is written automatically.
+        """
+        try:
+            self.writeline("pass")
+            for node in nodes:
+                self.visit(node, frame)
+        except CompilerExit:
+            pass
+
+    def write(self, x: str) -> None:
+        """Write a string into the output stream."""
+        if self._new_lines:
+            if not self._first_write:
+                self.stream.write("\n" * self._new_lines)
+                self.code_lineno += self._new_lines
+                if self._write_debug_info is not None:
+                    self.debug_info.append((self._write_debug_info, self.code_lineno))
+                    self._write_debug_info = None
+            self._first_write = False
+            self.stream.write("    " * self._indentation)
+            self._new_lines = 0
+        self.stream.write(x)
+
+    def writeline(
+        self, x: str, node: t.Optional[nodes.Node] = None, extra: int = 0
+    ) -> None:
+        """Combination of newline and write."""
+        self.newline(node, extra)
+        self.write(x)
+
+    def newline(self, node: t.Optional[nodes.Node] = None, extra: int = 0) -> None:
+        """Add one or more newlines before the next write."""
+        self._new_lines = max(self._new_lines, 1 + extra)
+        if node is not None and node.lineno != self._last_line:
+            self._write_debug_info = node.lineno
+            self._last_line = node.lineno
+
+    def signature(
+        self,
+        node: t.Union[nodes.Call, nodes.Filter, nodes.Test],
+        frame: Frame,
+        extra_kwargs: t.Optional[t.Mapping[str, t.Any]] = None,
+    ) -> None:
+        """Writes a function call to the stream for the current node.
+        A leading comma is added automatically.  The extra keyword
+        arguments may not include python keywords otherwise a syntax
+        error could occur.  The extra keyword arguments should be given
+        as python dict.
+        """
+        # if any of the given keyword arguments is a python keyword
+        # we have to make sure that no invalid call is created.
+        kwarg_workaround = any(
+            is_python_keyword(t.cast(str, k))
+            for k in chain((x.key for x in node.kwargs), extra_kwargs or ())
+        )
+
+        for arg in node.args:
+            self.write(", ")
+            self.visit(arg, frame)
+
+        if not kwarg_workaround:
+            for kwarg in node.kwargs:
+                self.write(", ")
+                self.visit(kwarg, frame)
+            if extra_kwargs is not None:
+                for key, value in extra_kwargs.items():
+                    self.write(f", {key}={value}")
+        if node.dyn_args:
+            self.write(", *")
+            self.visit(node.dyn_args, frame)
+
+        if kwarg_workaround:
+            if node.dyn_kwargs is not None:
+                self.write(", **dict({")
+            else:
+                self.write(", **{")
+            for kwarg in node.kwargs:
+                self.write(f"{kwarg.key!r}: ")
+                self.visit(kwarg.value, frame)
+                self.write(", ")
+            if extra_kwargs is not None:
+                for key, value in extra_kwargs.items():
+                    self.write(f"{key!r}: {value}, ")
+            if node.dyn_kwargs is not None:
+                self.write("}, **")
+                self.visit(node.dyn_kwargs, frame)
+                self.write(")")
+            else:
+                self.write("}")
+
+        elif node.dyn_kwargs is not None:
+            self.write(", **")
+            self.visit(node.dyn_kwargs, frame)
+
+    def pull_dependencies(self, nodes: t.Iterable[nodes.Node]) -> None:
+        """Find all filter and test names used in the template and
+        assign them to variables in the compiled namespace. Checking
+        that the names are registered with the environment is done when
+        compiling the Filter and Test nodes. If the node is in an If or
+        CondExpr node, the check is done at runtime instead.
+
+        .. versionchanged:: 3.0
+            Filters and tests in If and CondExpr nodes are checked at
+            runtime instead of compile time.
+        """
+        visitor = DependencyFinderVisitor()
+
+        for node in nodes:
+            visitor.visit(node)
+
+        for id_map, names, dependency in (
+            (self.filters, visitor.filters, "filters"),
+            (
+                self.tests,
+                visitor.tests,
+                "tests",
+            ),
+        ):
+            for name in sorted(names):
+                if name not in id_map:
+                    id_map[name] = self.temporary_identifier()
+
+                # add check during runtime that dependencies used inside of executed
+                # blocks are defined, as this step may be skipped during compile time
+                self.writeline("try:")
+                self.indent()
+                self.writeline(f"{id_map[name]} = environment.{dependency}[{name!r}]")
+                self.outdent()
+                self.writeline("except KeyError:")
+                self.indent()
+                self.writeline("@internalcode")
+                self.writeline(f"def {id_map[name]}(*unused):")
+                self.indent()
+                self.writeline(
+                    f'raise TemplateRuntimeError("No {dependency[:-1]}'
+                    f' named {name!r} found.")'
+                )
+                self.outdent()
+                self.outdent()
+
+    def enter_frame(self, frame: Frame) -> None:
+        undefs = []
+        for target, (action, param) in frame.symbols.loads.items():
+            if action == VAR_LOAD_PARAMETER:
+                pass
+            elif action == VAR_LOAD_RESOLVE:
+                self.writeline(f"{target} = {self.get_resolve_func()}({param!r})")
+            elif action == VAR_LOAD_ALIAS:
+                self.writeline(f"{target} = {param}")
+            elif action == VAR_LOAD_UNDEFINED:
+                undefs.append(target)
+            else:
+                raise NotImplementedError("unknown load instruction")
+        if undefs:
+            self.writeline(f"{' = '.join(undefs)} = missing")
+
+    def leave_frame(self, frame: Frame, with_python_scope: bool = False) -> None:
+        if not with_python_scope:
+            undefs = []
+            for target in frame.symbols.loads:
+                undefs.append(target)
+            if undefs:
+                self.writeline(f"{' = '.join(undefs)} = missing")
+
+    def choose_async(self, async_value: str = "async ", sync_value: str = "") -> str:
+        return async_value if self.environment.is_async else sync_value
+
+    def func(self, name: str) -> str:
+        return f"{self.choose_async()}def {name}"
+
+    def macro_body(
+        self, node: t.Union[nodes.Macro, nodes.CallBlock], frame: Frame
+    ) -> t.Tuple[Frame, MacroRef]:
+        """Dump the function def of a macro or call block."""
+        frame = frame.inner()
+        frame.symbols.analyze_node(node)
+        macro_ref = MacroRef(node)
+
+        explicit_caller = None
+        skip_special_params = set()
+        args = []
+
+        for idx, arg in enumerate(node.args):
+            if arg.name == "caller":
+                explicit_caller = idx
+            if arg.name in ("kwargs", "varargs"):
+                skip_special_params.add(arg.name)
+            args.append(frame.symbols.ref(arg.name))
+
+        undeclared = find_undeclared(node.body, ("caller", "kwargs", "varargs"))
+
+        if "caller" in undeclared:
+            # In older Jinja versions there was a bug that allowed caller
+            # to retain the special behavior even if it was mentioned in
+            # the argument list.  However thankfully this was only really
+            # working if it was the last argument.  So we are explicitly
+            # checking this now and error out if it is anywhere else in
+            # the argument list.
+            if explicit_caller is not None:
+                try:
+                    node.defaults[explicit_caller - len(node.args)]
+                except IndexError:
+                    self.fail(
+                        "When defining macros or call blocks the "
+                        'special "caller" argument must be omitted '
+                        "or be given a default.",
+                        node.lineno,
+                    )
+            else:
+                args.append(frame.symbols.declare_parameter("caller"))
+            macro_ref.accesses_caller = True
+        if "kwargs" in undeclared and "kwargs" not in skip_special_params:
+            args.append(frame.symbols.declare_parameter("kwargs"))
+            macro_ref.accesses_kwargs = True
+        if "varargs" in undeclared and "varargs" not in skip_special_params:
+            args.append(frame.symbols.declare_parameter("varargs"))
+            macro_ref.accesses_varargs = True
+
+        # macros are delayed, they never require output checks
+        frame.require_output_check = False
+        frame.symbols.analyze_node(node)
+        self.writeline(f"{self.func('macro')}({', '.join(args)}):", node)
+        self.indent()
+
+        self.buffer(frame)
+        self.enter_frame(frame)
+
+        self.push_parameter_definitions(frame)
+        for idx, arg in enumerate(node.args):
+            ref = frame.symbols.ref(arg.name)
+            self.writeline(f"if {ref} is missing:")
+            self.indent()
+            try:
+                default = node.defaults[idx - len(node.args)]
+            except IndexError:
+                self.writeline(
+                    f'{ref} = undefined("parameter {arg.name!r} was not provided",'
+                    f" name={arg.name!r})"
+                )
+            else:
+                self.writeline(f"{ref} = ")
+                self.visit(default, frame)
+            self.mark_parameter_stored(ref)
+            self.outdent()
+        self.pop_parameter_definitions()
+
+        self.blockvisit(node.body, frame)
+        self.return_buffer_contents(frame, force_unescaped=True)
+        self.leave_frame(frame, with_python_scope=True)
+        self.outdent()
+
+        return frame, macro_ref
+
+    def macro_def(self, macro_ref: MacroRef, frame: Frame) -> None:
+        """Dump the macro definition for the def created by macro_body."""
+        arg_tuple = ", ".join(repr(x.name) for x in macro_ref.node.args)
+        name = getattr(macro_ref.node, "name", None)
+        if len(macro_ref.node.args) == 1:
+            arg_tuple += ","
+        self.write(
+            f"Macro(environment, macro, {name!r}, ({arg_tuple}),"
+            f" {macro_ref.accesses_kwargs!r}, {macro_ref.accesses_varargs!r},"
+            f" {macro_ref.accesses_caller!r}, context.eval_ctx.autoescape)"
+        )
+
+    def position(self, node: nodes.Node) -> str:
+        """Return a human readable position for the node."""
+        rv = f"line {node.lineno}"
+        if self.name is not None:
+            rv = f"{rv} in {self.name!r}"
+        return rv
+
+    def dump_local_context(self, frame: Frame) -> str:
+        items_kv = ", ".join(
+            f"{name!r}: {target}"
+            for name, target in frame.symbols.dump_stores().items()
+        )
+        return f"{{{items_kv}}}"
+
+    def write_commons(self) -> None:
+        """Writes a common preamble that is used by root and block functions.
+        Primarily this sets up common local helpers and enforces a generator
+        through a dead branch.
+        """
+        self.writeline("resolve = context.resolve_or_missing")
+        self.writeline("undefined = environment.undefined")
+        self.writeline("concat = environment.concat")
+        # always use the standard Undefined class for the implicit else of
+        # conditional expressions
+        self.writeline("cond_expr_undefined = Undefined")
+        self.writeline("if 0: yield None")
+
+    def push_parameter_definitions(self, frame: Frame) -> None:
+        """Pushes all parameter targets from the given frame into a local
+        stack that permits tracking of yet to be assigned parameters.  In
+        particular this enables the optimization from `visit_Name` to skip
+        undefined expressions for parameters in macros as macros can reference
+        otherwise unbound parameters.
+        """
+        self._param_def_block.append(frame.symbols.dump_param_targets())
+
+    def pop_parameter_definitions(self) -> None:
+        """Pops the current parameter definitions set."""
+        self._param_def_block.pop()
+
+    def mark_parameter_stored(self, target: str) -> None:
+        """Marks a parameter in the current parameter definitions as stored.
+        This will skip the enforced undefined checks.
+        """
+        if self._param_def_block:
+            self._param_def_block[-1].discard(target)
+
+    def push_context_reference(self, target: str) -> None:
+        self._context_reference_stack.append(target)
+
+    def pop_context_reference(self) -> None:
+        self._context_reference_stack.pop()
+
+    def get_context_ref(self) -> str:
+        return self._context_reference_stack[-1]
+
+    def get_resolve_func(self) -> str:
+        target = self._context_reference_stack[-1]
+        if target == "context":
+            return "resolve"
+        return f"{target}.resolve"
+
+    def derive_context(self, frame: Frame) -> str:
+        return f"{self.get_context_ref()}.derived({self.dump_local_context(frame)})"
+
+    def parameter_is_undeclared(self, target: str) -> bool:
+        """Checks if a given target is an undeclared parameter."""
+        if not self._param_def_block:
+            return False
+        return target in self._param_def_block[-1]
+
+    def push_assign_tracking(self) -> None:
+        """Pushes a new layer for assignment tracking."""
+        self._assign_stack.append(set())
+
+    def pop_assign_tracking(self, frame: Frame) -> None:
+        """Pops the topmost level for assignment tracking and updates the
+        context variables if necessary.
+        """
+        vars = self._assign_stack.pop()
+        if (
+            not frame.block_frame
+            and not frame.loop_frame
+            and not frame.toplevel
+            or not vars
+        ):
+            return
+        public_names = [x for x in vars if x[:1] != "_"]
+        if len(vars) == 1:
+            name = next(iter(vars))
+            ref = frame.symbols.ref(name)
+            if frame.loop_frame:
+                self.writeline(f"_loop_vars[{name!r}] = {ref}")
+                return
+            if frame.block_frame:
+                self.writeline(f"_block_vars[{name!r}] = {ref}")
+                return
+            self.writeline(f"context.vars[{name!r}] = {ref}")
+        else:
+            if frame.loop_frame:
+                self.writeline("_loop_vars.update({")
+            elif frame.block_frame:
+                self.writeline("_block_vars.update({")
+            else:
+                self.writeline("context.vars.update({")
+            for idx, name in enumerate(sorted(vars)):
+                if idx:
+                    self.write(", ")
+                ref = frame.symbols.ref(name)
+                self.write(f"{name!r}: {ref}")
+            self.write("})")
+        if not frame.block_frame and not frame.loop_frame and public_names:
+            if len(public_names) == 1:
+                self.writeline(f"context.exported_vars.add({public_names[0]!r})")
+            else:
+                names_str = ", ".join(map(repr, sorted(public_names)))
+                self.writeline(f"context.exported_vars.update(({names_str}))")
+
+    # -- Statement Visitors
+
+    def visit_Template(
+        self, node: nodes.Template, frame: t.Optional[Frame] = None
+    ) -> None:
+        assert frame is None, "no root frame allowed"
+        eval_ctx = EvalContext(self.environment, self.name)
+
+        from .runtime import async_exported
+        from .runtime import exported
+
+        if self.environment.is_async:
+            exported_names = sorted(exported + async_exported)
+        else:
+            exported_names = sorted(exported)
+
+        self.writeline("from jinja2.runtime import " + ", ".join(exported_names))
+
+        # if we want a deferred initialization we cannot move the
+        # environment into a local name
+        envenv = "" if self.defer_init else ", environment=environment"
+
+        # do we have an extends tag at all?  If not, we can save some
+        # overhead by just not processing any inheritance code.
+        have_extends = node.find(nodes.Extends) is not None
+
+        # find all blocks
+        for block in node.find_all(nodes.Block):
+            if block.name in self.blocks:
+                self.fail(f"block {block.name!r} defined twice", block.lineno)
+            self.blocks[block.name] = block
+
+        # find all imports and import them
+        for import_ in node.find_all(nodes.ImportedName):
+            if import_.importname not in self.import_aliases:
+                imp = import_.importname
+                self.import_aliases[imp] = alias = self.temporary_identifier()
+                if "." in imp:
+                    module, obj = imp.rsplit(".", 1)
+                    self.writeline(f"from {module} import {obj} as {alias}")
+                else:
+                    self.writeline(f"import {imp} as {alias}")
+
+        # add the load name
+        self.writeline(f"name = {self.name!r}")
+
+        # generate the root render function.
+        self.writeline(
+            f"{self.func('root')}(context, missing=missing{envenv}):", extra=1
+        )
+        self.indent()
+        self.write_commons()
+
+        # process the root
+        frame = Frame(eval_ctx)
+        if "self" in find_undeclared(node.body, ("self",)):
+            ref = frame.symbols.declare_parameter("self")
+            self.writeline(f"{ref} = TemplateReference(context)")
+        frame.symbols.analyze_node(node)
+        frame.toplevel = frame.rootlevel = True
+        frame.require_output_check = have_extends and not self.has_known_extends
+        if have_extends:
+            self.writeline("parent_template = None")
+        self.enter_frame(frame)
+        self.pull_dependencies(node.body)
+        self.blockvisit(node.body, frame)
+        self.leave_frame(frame, with_python_scope=True)
+        self.outdent()
+
+        # make sure that the parent root is called.
+        if have_extends:
+            if not self.has_known_extends:
+                self.indent()
+                self.writeline("if parent_template is not None:")
+            self.indent()
+            if not self.environment.is_async:
+                self.writeline("yield from parent_template.root_render_func(context)")
+            else:
+                self.writeline("agen = parent_template.root_render_func(context)")
+                self.writeline("try:")
+                self.indent()
+                self.writeline("async for event in agen:")
+                self.indent()
+                self.writeline("yield event")
+                self.outdent()
+                self.outdent()
+                self.writeline("finally: await agen.aclose()")
+            self.outdent(1 + (not self.has_known_extends))
+
+        # at this point we now have the blocks collected and can visit them too.
+        for name, block in self.blocks.items():
+            self.writeline(
+                f"{self.func('block_' + name)}(context, missing=missing{envenv}):",
+                block,
+                1,
+            )
+            self.indent()
+            self.write_commons()
+            # It's important that we do not make this frame a child of the
+            # toplevel template.  This would cause a variety of
+            # interesting issues with identifier tracking.
+            block_frame = Frame(eval_ctx)
+            block_frame.block_frame = True
+            undeclared = find_undeclared(block.body, ("self", "super"))
+            if "self" in undeclared:
+                ref = block_frame.symbols.declare_parameter("self")
+                self.writeline(f"{ref} = TemplateReference(context)")
+            if "super" in undeclared:
+                ref = block_frame.symbols.declare_parameter("super")
+                self.writeline(f"{ref} = context.super({name!r}, block_{name})")
+            block_frame.symbols.analyze_node(block)
+            block_frame.block = name
+            self.writeline("_block_vars = {}")
+            self.enter_frame(block_frame)
+            self.pull_dependencies(block.body)
+            self.blockvisit(block.body, block_frame)
+            self.leave_frame(block_frame, with_python_scope=True)
+            self.outdent()
+
+        blocks_kv_str = ", ".join(f"{x!r}: block_{x}" for x in self.blocks)
+        self.writeline(f"blocks = {{{blocks_kv_str}}}", extra=1)
+        debug_kv_str = "&".join(f"{k}={v}" for k, v in self.debug_info)
+        self.writeline(f"debug_info = {debug_kv_str!r}")
+
+    def visit_Block(self, node: nodes.Block, frame: Frame) -> None:
+        """Call a block and register it for the template."""
+        level = 0
+        if frame.toplevel:
+            # if we know that we are a child template, there is no need to
+            # check if we are one
+            if self.has_known_extends:
+                return
+            if self.extends_so_far > 0:
+                self.writeline("if parent_template is None:")
+                self.indent()
+                level += 1
+
+        if node.scoped:
+            context = self.derive_context(frame)
+        else:
+            context = self.get_context_ref()
+
+        if node.required:
+            self.writeline(f"if len(context.blocks[{node.name!r}]) <= 1:", node)
+            self.indent()
+            self.writeline(
+                f'raise TemplateRuntimeError("Required block {node.name!r} not found")',
+                node,
+            )
+            self.outdent()
+
+        if not self.environment.is_async and frame.buffer is None:
+            self.writeline(
+                f"yield from context.blocks[{node.name!r}][0]({context})", node
+            )
+        else:
+            self.writeline(f"gen = context.blocks[{node.name!r}][0]({context})")
+            self.writeline("try:")
+            self.indent()
+            self.writeline(
+                f"{self.choose_async()}for event in gen:",
+                node,
+            )
+            self.indent()
+            self.simple_write("event", frame)
+            self.outdent()
+            self.outdent()
+            self.writeline(
+                f"finally: {self.choose_async('await gen.aclose()', 'gen.close()')}"
+            )
+
+        self.outdent(level)
+
+    def visit_Extends(self, node: nodes.Extends, frame: Frame) -> None:
+        """Calls the extender."""
+        if not frame.toplevel:
+            self.fail("cannot use extend from a non top-level scope", node.lineno)
+
+        # if the number of extends statements in general is zero so
+        # far, we don't have to add a check if something extended
+        # the template before this one.
+        if self.extends_so_far > 0:
+            # if we have a known extends we just add a template runtime
+            # error into the generated code.  We could catch that at compile
+            # time too, but i welcome it not to confuse users by throwing the
+            # same error at different times just "because we can".
+            if not self.has_known_extends:
+                self.writeline("if parent_template is not None:")
+                self.indent()
+            self.writeline('raise TemplateRuntimeError("extended multiple times")')
+
+            # if we have a known extends already we don't need that code here
+            # as we know that the template execution will end here.
+            if self.has_known_extends:
+                raise CompilerExit()
+            else:
+                self.outdent()
+
+        self.writeline("parent_template = environment.get_template(", node)
+        self.visit(node.template, frame)
+        self.write(f", {self.name!r})")
+        self.writeline("for name, parent_block in parent_template.blocks.items():")
+        self.indent()
+        self.writeline("context.blocks.setdefault(name, []).append(parent_block)")
+        self.outdent()
+
+        # if this extends statement was in the root level we can take
+        # advantage of that information and simplify the generated code
+        # in the top level from this point onwards
+        if frame.rootlevel:
+            self.has_known_extends = True
+
+        # and now we have one more
+        self.extends_so_far += 1
+
+    def visit_Include(self, node: nodes.Include, frame: Frame) -> None:
+        """Handles includes."""
+        if node.ignore_missing:
+            self.writeline("try:")
+            self.indent()
+
+        func_name = "get_or_select_template"
+        if isinstance(node.template, nodes.Const):
+            if isinstance(node.template.value, str):
+                func_name = "get_template"
+            elif isinstance(node.template.value, (tuple, list)):
+                func_name = "select_template"
+        elif isinstance(node.template, (nodes.Tuple, nodes.List)):
+            func_name = "select_template"
+
+        self.writeline(f"template = environment.{func_name}(", node)
+        self.visit(node.template, frame)
+        self.write(f", {self.name!r})")
+        if node.ignore_missing:
+            self.outdent()
+            self.writeline("except TemplateNotFound:")
+            self.indent()
+            self.writeline("pass")
+            self.outdent()
+            self.writeline("else:")
+            self.indent()
+
+        def loop_body() -> None:
+            self.indent()
+            self.simple_write("event", frame)
+            self.outdent()
+
+        if node.with_context:
+            self.writeline(
+                f"gen = template.root_render_func("
+                "template.new_context(context.get_all(), True,"
+                f" {self.dump_local_context(frame)}))"
+            )
+            self.writeline("try:")
+            self.indent()
+            self.writeline(f"{self.choose_async()}for event in gen:")
+            loop_body()
+            self.outdent()
+            self.writeline(
+                f"finally: {self.choose_async('await gen.aclose()', 'gen.close()')}"
+            )
+        elif self.environment.is_async:
+            self.writeline(
+                "for event in (await template._get_default_module_async())"
+                "._body_stream:"
+            )
+            loop_body()
+        else:
+            self.writeline("yield from template._get_default_module()._body_stream")
+
+        if node.ignore_missing:
+            self.outdent()
+
+    def _import_common(
+        self, node: t.Union[nodes.Import, nodes.FromImport], frame: Frame
+    ) -> None:
+        self.write(f"{self.choose_async('await ')}environment.get_template(")
+        self.visit(node.template, frame)
+        self.write(f", {self.name!r}).")
+
+        if node.with_context:
+            f_name = f"make_module{self.choose_async('_async')}"
+            self.write(
+                f"{f_name}(context.get_all(), True, {self.dump_local_context(frame)})"
+            )
+        else:
+            self.write(f"_get_default_module{self.choose_async('_async')}(context)")
+
+    def visit_Import(self, node: nodes.Import, frame: Frame) -> None:
+        """Visit regular imports."""
+        self.writeline(f"{frame.symbols.ref(node.target)} = ", node)
+        if frame.toplevel:
+            self.write(f"context.vars[{node.target!r}] = ")
+
+        self._import_common(node, frame)
+
+        if frame.toplevel and not node.target.startswith("_"):
+            self.writeline(f"context.exported_vars.discard({node.target!r})")
+
+    def visit_FromImport(self, node: nodes.FromImport, frame: Frame) -> None:
+        """Visit named imports."""
+        self.newline(node)
+        self.write("included_template = ")
+        self._import_common(node, frame)
+        var_names = []
+        discarded_names = []
+        for name in node.names:
+            if isinstance(name, tuple):
+                name, alias = name
+            else:
+                alias = name
+            self.writeline(
+                f"{frame.symbols.ref(alias)} ="
+                f" getattr(included_template, {name!r}, missing)"
+            )
+            self.writeline(f"if {frame.symbols.ref(alias)} is missing:")
+            self.indent()
+            # The position will contain the template name, and will be formatted
+            # into a string that will be compiled into an f-string. Curly braces
+            # in the name must be replaced with escapes so that they will not be
+            # executed as part of the f-string.
+            position = self.position(node).replace("{", "{{").replace("}", "}}")
+            message = (
+                "the template {included_template.__name__!r}"
+                f" (imported on {position})"
+                f" does not export the requested name {name!r}"
+            )
+            self.writeline(
+                f"{frame.symbols.ref(alias)} = undefined(f{message!r}, name={name!r})"
+            )
+            self.outdent()
+            if frame.toplevel:
+                var_names.append(alias)
+                if not alias.startswith("_"):
+                    discarded_names.append(alias)
+
+        if var_names:
+            if len(var_names) == 1:
+                name = var_names[0]
+                self.writeline(f"context.vars[{name!r}] = {frame.symbols.ref(name)}")
+            else:
+                names_kv = ", ".join(
+                    f"{name!r}: {frame.symbols.ref(name)}" for name in var_names
+                )
+                self.writeline(f"context.vars.update({{{names_kv}}})")
+        if discarded_names:
+            if len(discarded_names) == 1:
+                self.writeline(f"context.exported_vars.discard({discarded_names[0]!r})")
+            else:
+                names_str = ", ".join(map(repr, discarded_names))
+                self.writeline(
+                    f"context.exported_vars.difference_update(({names_str}))"
+                )
+
+    def visit_For(self, node: nodes.For, frame: Frame) -> None:
+        loop_frame = frame.inner()
+        loop_frame.loop_frame = True
+        test_frame = frame.inner()
+        else_frame = frame.inner()
+
+        # try to figure out if we have an extended loop.  An extended loop
+        # is necessary if the loop is in recursive mode if the special loop
+        # variable is accessed in the body if the body is a scoped block.
+        extended_loop = (
+            node.recursive
+            or "loop"
+            in find_undeclared(node.iter_child_nodes(only=("body",)), ("loop",))
+            or any(block.scoped for block in node.find_all(nodes.Block))
+        )
+
+        loop_ref = None
+        if extended_loop:
+            loop_ref = loop_frame.symbols.declare_parameter("loop")
+
+        loop_frame.symbols.analyze_node(node, for_branch="body")
+        if node.else_:
+            else_frame.symbols.analyze_node(node, for_branch="else")
+
+        if node.test:
+            loop_filter_func = self.temporary_identifier()
+            test_frame.symbols.analyze_node(node, for_branch="test")
+            self.writeline(f"{self.func(loop_filter_func)}(fiter):", node.test)
+            self.indent()
+            self.enter_frame(test_frame)
+            self.writeline(self.choose_async("async for ", "for "))
+            self.visit(node.target, loop_frame)
+            self.write(" in ")
+            self.write(self.choose_async("auto_aiter(fiter)", "fiter"))
+            self.write(":")
+            self.indent()
+            self.writeline("if ", node.test)
+            self.visit(node.test, test_frame)
+            self.write(":")
+            self.indent()
+            self.writeline("yield ")
+            self.visit(node.target, loop_frame)
+            self.outdent(3)
+            self.leave_frame(test_frame, with_python_scope=True)
+
+        # if we don't have an recursive loop we have to find the shadowed
+        # variables at that point.  Because loops can be nested but the loop
+        # variable is a special one we have to enforce aliasing for it.
+        if node.recursive:
+            self.writeline(
+                f"{self.func('loop')}(reciter, loop_render_func, depth=0):", node
+            )
+            self.indent()
+            self.buffer(loop_frame)
+
+            # Use the same buffer for the else frame
+            else_frame.buffer = loop_frame.buffer
+
+        # make sure the loop variable is a special one and raise a template
+        # assertion error if a loop tries to write to loop
+        if extended_loop:
+            self.writeline(f"{loop_ref} = missing")
+
+        for name in node.find_all(nodes.Name):
+            if name.ctx == "store" and name.name == "loop":
+                self.fail(
+                    "Can't assign to special loop variable in for-loop target",
+                    name.lineno,
+                )
+
+        if node.else_:
+            iteration_indicator = self.temporary_identifier()
+            self.writeline(f"{iteration_indicator} = 1")
+
+        self.writeline(self.choose_async("async for ", "for "), node)
+        self.visit(node.target, loop_frame)
+        if extended_loop:
+            self.write(f", {loop_ref} in {self.choose_async('Async')}LoopContext(")
+        else:
+            self.write(" in ")
+
+        if node.test:
+            self.write(f"{loop_filter_func}(")
+        if node.recursive:
+            self.write("reciter")
+        else:
+            if self.environment.is_async and not extended_loop:
+                self.write("auto_aiter(")
+            self.visit(node.iter, frame)
+            if self.environment.is_async and not extended_loop:
+                self.write(")")
+        if node.test:
+            self.write(")")
+
+        if node.recursive:
+            self.write(", undefined, loop_render_func, depth):")
+        else:
+            self.write(", undefined):" if extended_loop else ":")
+
+        self.indent()
+        self.enter_frame(loop_frame)
+
+        self.writeline("_loop_vars = {}")
+        self.blockvisit(node.body, loop_frame)
+        if node.else_:
+            self.writeline(f"{iteration_indicator} = 0")
+        self.outdent()
+        self.leave_frame(
+            loop_frame, with_python_scope=node.recursive and not node.else_
+        )
+
+        if node.else_:
+            self.writeline(f"if {iteration_indicator}:")
+            self.indent()
+            self.enter_frame(else_frame)
+            self.blockvisit(node.else_, else_frame)
+            self.leave_frame(else_frame)
+            self.outdent()
+
+        # if the node was recursive we have to return the buffer contents
+        # and start the iteration code
+        if node.recursive:
+            self.return_buffer_contents(loop_frame)
+            self.outdent()
+            self.start_write(frame, node)
+            self.write(f"{self.choose_async('await ')}loop(")
+            if self.environment.is_async:
+                self.write("auto_aiter(")
+            self.visit(node.iter, frame)
+            if self.environment.is_async:
+                self.write(")")
+            self.write(", loop)")
+            self.end_write(frame)
+
+        # at the end of the iteration, clear any assignments made in the
+        # loop from the top level
+        if self._assign_stack:
+            self._assign_stack[-1].difference_update(loop_frame.symbols.stores)
+
+    def visit_If(self, node: nodes.If, frame: Frame) -> None:
+        if_frame = frame.soft()
+        self.writeline("if ", node)
+        self.visit(node.test, if_frame)
+        self.write(":")
+        self.indent()
+        self.blockvisit(node.body, if_frame)
+        self.outdent()
+        for elif_ in node.elif_:
+            self.writeline("elif ", elif_)
+            self.visit(elif_.test, if_frame)
+            self.write(":")
+            self.indent()
+            self.blockvisit(elif_.body, if_frame)
+            self.outdent()
+        if node.else_:
+            self.writeline("else:")
+            self.indent()
+            self.blockvisit(node.else_, if_frame)
+            self.outdent()
+
+    def visit_Macro(self, node: nodes.Macro, frame: Frame) -> None:
+        macro_frame, macro_ref = self.macro_body(node, frame)
+        self.newline()
+        if frame.toplevel:
+            if not node.name.startswith("_"):
+                self.write(f"context.exported_vars.add({node.name!r})")
+            self.writeline(f"context.vars[{node.name!r}] = ")
+        self.write(f"{frame.symbols.ref(node.name)} = ")
+        self.macro_def(macro_ref, macro_frame)
+
+    def visit_CallBlock(self, node: nodes.CallBlock, frame: Frame) -> None:
+        call_frame, macro_ref = self.macro_body(node, frame)
+        self.writeline("caller = ")
+        self.macro_def(macro_ref, call_frame)
+        self.start_write(frame, node)
+        self.visit_Call(node.call, frame, forward_caller=True)
+        self.end_write(frame)
+
+    def visit_FilterBlock(self, node: nodes.FilterBlock, frame: Frame) -> None:
+        filter_frame = frame.inner()
+        filter_frame.symbols.analyze_node(node)
+        self.enter_frame(filter_frame)
+        self.buffer(filter_frame)
+        self.blockvisit(node.body, filter_frame)
+        self.start_write(frame, node)
+        self.visit_Filter(node.filter, filter_frame)
+        self.end_write(frame)
+        self.leave_frame(filter_frame)
+
+    def visit_With(self, node: nodes.With, frame: Frame) -> None:
+        with_frame = frame.inner()
+        with_frame.symbols.analyze_node(node)
+        self.enter_frame(with_frame)
+        for target, expr in zip(node.targets, node.values):
+            self.newline()
+            self.visit(target, with_frame)
+            self.write(" = ")
+            self.visit(expr, frame)
+        self.blockvisit(node.body, with_frame)
+        self.leave_frame(with_frame)
+
+    def visit_ExprStmt(self, node: nodes.ExprStmt, frame: Frame) -> None:
+        self.newline(node)
+        self.visit(node.node, frame)
+
+    class _FinalizeInfo(t.NamedTuple):
+        const: t.Optional[t.Callable[..., str]]
+        src: t.Optional[str]
+
+    @staticmethod
+    def _default_finalize(value: t.Any) -> t.Any:
+        """The default finalize function if the environment isn't
+        configured with one. Or, if the environment has one, this is
+        called on that function's output for constants.
+        """
+        return str(value)
+
+    _finalize: t.Optional[_FinalizeInfo] = None
+
+    def _make_finalize(self) -> _FinalizeInfo:
+        """Build the finalize function to be used on constants and at
+        runtime. Cached so it's only created once for all output nodes.
+
+        Returns a ``namedtuple`` with the following attributes:
+
+        ``const``
+            A function to finalize constant data at compile time.
+
+        ``src``
+            Source code to output around nodes to be evaluated at
+            runtime.
+        """
+        if self._finalize is not None:
+            return self._finalize
+
+        finalize: t.Optional[t.Callable[..., t.Any]]
+        finalize = default = self._default_finalize
+        src = None
+
+        if self.environment.finalize:
+            src = "environment.finalize("
+            env_finalize = self.environment.finalize
+            pass_arg = {
+                _PassArg.context: "context",
+                _PassArg.eval_context: "context.eval_ctx",
+                _PassArg.environment: "environment",
+            }.get(
+                _PassArg.from_obj(env_finalize)  # type: ignore
+            )
+            finalize = None
+
+            if pass_arg is None:
+
+                def finalize(value: t.Any) -> t.Any:  # noqa: F811
+                    return default(env_finalize(value))
+
+            else:
+                src = f"{src}{pass_arg}, "
+
+                if pass_arg == "environment":
+
+                    def finalize(value: t.Any) -> t.Any:  # noqa: F811
+                        return default(env_finalize(self.environment, value))
+
+        self._finalize = self._FinalizeInfo(finalize, src)
+        return self._finalize
+
+    def _output_const_repr(self, group: t.Iterable[t.Any]) -> str:
+        """Given a group of constant values converted from ``Output``
+        child nodes, produce a string to write to the template module
+        source.
+        """
+        return repr(concat(group))
+
+    def _output_child_to_const(
+        self, node: nodes.Expr, frame: Frame, finalize: _FinalizeInfo
+    ) -> str:
+        """Try to optimize a child of an ``Output`` node by trying to
+        convert it to constant, finalized data at compile time.
+
+        If :exc:`Impossible` is raised, the node is not constant and
+        will be evaluated at runtime. Any other exception will also be
+        evaluated at runtime for easier debugging.
+        """
+        const = node.as_const(frame.eval_ctx)
+
+        if frame.eval_ctx.autoescape:
+            const = escape(const)
+
+        # Template data doesn't go through finalize.
+        if isinstance(node, nodes.TemplateData):
+            return str(const)
+
+        return finalize.const(const)  # type: ignore
+
+    def _output_child_pre(
+        self, node: nodes.Expr, frame: Frame, finalize: _FinalizeInfo
+    ) -> None:
+        """Output extra source code before visiting a child of an
+        ``Output`` node.
+        """
+        if frame.eval_ctx.volatile:
+            self.write("(escape if context.eval_ctx.autoescape else str)(")
+        elif frame.eval_ctx.autoescape:
+            self.write("escape(")
+        else:
+            self.write("str(")
+
+        if finalize.src is not None:
+            self.write(finalize.src)
+
+    def _output_child_post(
+        self, node: nodes.Expr, frame: Frame, finalize: _FinalizeInfo
+    ) -> None:
+        """Output extra source code after visiting a child of an
+        ``Output`` node.
+        """
+        self.write(")")
+
+        if finalize.src is not None:
+            self.write(")")
+
+    def visit_Output(self, node: nodes.Output, frame: Frame) -> None:
+        # If an extends is active, don't render outside a block.
+        if frame.require_output_check:
+            # A top-level extends is known to exist at compile time.
+            if self.has_known_extends:
+                return
+
+            self.writeline("if parent_template is None:")
+            self.indent()
+
+        finalize = self._make_finalize()
+        body: t.List[t.Union[t.List[t.Any], nodes.Expr]] = []
+
+        # Evaluate constants at compile time if possible. Each item in
+        # body will be either a list of static data or a node to be
+        # evaluated at runtime.
+        for child in node.nodes:
+            try:
+                if not (
+                    # If the finalize function requires runtime context,
+                    # constants can't be evaluated at compile time.
+                    finalize.const
+                    # Unless it's basic template data that won't be
+                    # finalized anyway.
+                    or isinstance(child, nodes.TemplateData)
+                ):
+                    raise nodes.Impossible()
+
+                const = self._output_child_to_const(child, frame, finalize)
+            except (nodes.Impossible, Exception):
+                # The node was not constant and needs to be evaluated at
+                # runtime. Or another error was raised, which is easier
+                # to debug at runtime.
+                body.append(child)
+                continue
+
+            if body and isinstance(body[-1], list):
+                body[-1].append(const)
+            else:
+                body.append([const])
+
+        if frame.buffer is not None:
+            if len(body) == 1:
+                self.writeline(f"{frame.buffer}.append(")
+            else:
+                self.writeline(f"{frame.buffer}.extend((")
+
+            self.indent()
+
+        for item in body:
+            if isinstance(item, list):
+                # A group of constant data to join and output.
+                val = self._output_const_repr(item)
+
+                if frame.buffer is None:
+                    self.writeline("yield " + val)
+                else:
+                    self.writeline(val + ",")
+            else:
+                if frame.buffer is None:
+                    self.writeline("yield ", item)
+                else:
+                    self.newline(item)
+
+                # A node to be evaluated at runtime.
+                self._output_child_pre(item, frame, finalize)
+                self.visit(item, frame)
+                self._output_child_post(item, frame, finalize)
+
+                if frame.buffer is not None:
+                    self.write(",")
+
+        if frame.buffer is not None:
+            self.outdent()
+            self.writeline(")" if len(body) == 1 else "))")
+
+        if frame.require_output_check:
+            self.outdent()
+
+    def visit_Assign(self, node: nodes.Assign, frame: Frame) -> None:
+        self.push_assign_tracking()
+
+        # ``a.b`` is allowed for assignment, and is parsed as an NSRef. However,
+        # it is only valid if it references a Namespace object. Emit a check for
+        # that for each ref here, before assignment code is emitted. This can't
+        # be done in visit_NSRef as the ref could be in the middle of a tuple.
+        seen_refs: t.Set[str] = set()
+
+        for nsref in node.find_all(nodes.NSRef):
+            if nsref.name in seen_refs:
+                # Only emit the check for each reference once, in case the same
+                # ref is used multiple times in a tuple, `ns.a, ns.b = c, d`.
+                continue
+
+            seen_refs.add(nsref.name)
+            ref = frame.symbols.ref(nsref.name)
+            self.writeline(f"if not isinstance({ref}, Namespace):")
+            self.indent()
+            self.writeline(
+                "raise TemplateRuntimeError"
+                '("cannot assign attribute on non-namespace object")'
+            )
+            self.outdent()
+
+        self.newline(node)
+        self.visit(node.target, frame)
+        self.write(" = ")
+        self.visit(node.node, frame)
+        self.pop_assign_tracking(frame)
+
+    def visit_AssignBlock(self, node: nodes.AssignBlock, frame: Frame) -> None:
+        self.push_assign_tracking()
+        block_frame = frame.inner()
+        # This is a special case.  Since a set block always captures we
+        # will disable output checks.  This way one can use set blocks
+        # toplevel even in extended templates.
+        block_frame.require_output_check = False
+        block_frame.symbols.analyze_node(node)
+        self.enter_frame(block_frame)
+        self.buffer(block_frame)
+        self.blockvisit(node.body, block_frame)
+        self.newline(node)
+        self.visit(node.target, frame)
+        self.write(" = (Markup if context.eval_ctx.autoescape else identity)(")
+        if node.filter is not None:
+            self.visit_Filter(node.filter, block_frame)
+        else:
+            self.write(f"concat({block_frame.buffer})")
+        self.write(")")
+        self.pop_assign_tracking(frame)
+        self.leave_frame(block_frame)
+
+    # -- Expression Visitors
+
+    def visit_Name(self, node: nodes.Name, frame: Frame) -> None:
+        if node.ctx == "store" and (
+            frame.toplevel or frame.loop_frame or frame.block_frame
+        ):
+            if self._assign_stack:
+                self._assign_stack[-1].add(node.name)
+        ref = frame.symbols.ref(node.name)
+
+        # If we are looking up a variable we might have to deal with the
+        # case where it's undefined.  We can skip that case if the load
+        # instruction indicates a parameter which are always defined.
+        if node.ctx == "load":
+            load = frame.symbols.find_load(ref)
+            if not (
+                load is not None
+                and load[0] == VAR_LOAD_PARAMETER
+                and not self.parameter_is_undeclared(ref)
+            ):
+                self.write(
+                    f"(undefined(name={node.name!r}) if {ref} is missing else {ref})"
+                )
+                return
+
+        self.write(ref)
+
+    def visit_NSRef(self, node: nodes.NSRef, frame: Frame) -> None:
+        # NSRef is a dotted assignment target a.b=c, but uses a[b]=c internally.
+        # visit_Assign emits code to validate that each ref is to a Namespace
+        # object only. That can't be emitted here as the ref could be in the
+        # middle of a tuple assignment.
+        ref = frame.symbols.ref(node.name)
+        self.writeline(f"{ref}[{node.attr!r}]")
+
+    def visit_Const(self, node: nodes.Const, frame: Frame) -> None:
+        val = node.as_const(frame.eval_ctx)
+        if isinstance(val, float):
+            self.write(str(val))
+        else:
+            self.write(repr(val))
+
+    def visit_TemplateData(self, node: nodes.TemplateData, frame: Frame) -> None:
+        try:
+            self.write(repr(node.as_const(frame.eval_ctx)))
+        except nodes.Impossible:
+            self.write(
+                f"(Markup if context.eval_ctx.autoescape else identity)({node.data!r})"
+            )
+
+    def visit_Tuple(self, node: nodes.Tuple, frame: Frame) -> None:
+        self.write("(")
+        idx = -1
+        for idx, item in enumerate(node.items):
+            if idx:
+                self.write(", ")
+            self.visit(item, frame)
+        self.write(",)" if idx == 0 else ")")
+
+    def visit_List(self, node: nodes.List, frame: Frame) -> None:
+        self.write("[")
+        for idx, item in enumerate(node.items):
+            if idx:
+                self.write(", ")
+            self.visit(item, frame)
+        self.write("]")
+
+    def visit_Dict(self, node: nodes.Dict, frame: Frame) -> None:
+        self.write("{")
+        for idx, item in enumerate(node.items):
+            if idx:
+                self.write(", ")
+            self.visit(item.key, frame)
+            self.write(": ")
+            self.visit(item.value, frame)
+        self.write("}")
+
+    visit_Add = _make_binop("+")
+    visit_Sub = _make_binop("-")
+    visit_Mul = _make_binop("*")
+    visit_Div = _make_binop("/")
+    visit_FloorDiv = _make_binop("//")
+    visit_Pow = _make_binop("**")
+    visit_Mod = _make_binop("%")
+    visit_And = _make_binop("and")
+    visit_Or = _make_binop("or")
+    visit_Pos = _make_unop("+")
+    visit_Neg = _make_unop("-")
+    visit_Not = _make_unop("not ")
+
+    @optimizeconst
+    def visit_Concat(self, node: nodes.Concat, frame: Frame) -> None:
+        if frame.eval_ctx.volatile:
+            func_name = "(markup_join if context.eval_ctx.volatile else str_join)"
+        elif frame.eval_ctx.autoescape:
+            func_name = "markup_join"
+        else:
+            func_name = "str_join"
+        self.write(f"{func_name}((")
+        for arg in node.nodes:
+            self.visit(arg, frame)
+            self.write(", ")
+        self.write("))")
+
+    @optimizeconst
+    def visit_Compare(self, node: nodes.Compare, frame: Frame) -> None:
+        self.write("(")
+        self.visit(node.expr, frame)
+        for op in node.ops:
+            self.visit(op, frame)
+        self.write(")")
+
+    def visit_Operand(self, node: nodes.Operand, frame: Frame) -> None:
+        self.write(f" {operators[node.op]} ")
+        self.visit(node.expr, frame)
+
+    @optimizeconst
+    def visit_Getattr(self, node: nodes.Getattr, frame: Frame) -> None:
+        if self.environment.is_async:
+            self.write("(await auto_await(")
+
+        self.write("environment.getattr(")
+        self.visit(node.node, frame)
+        self.write(f", {node.attr!r})")
+
+        if self.environment.is_async:
+            self.write("))")
+
+    @optimizeconst
+    def visit_Getitem(self, node: nodes.Getitem, frame: Frame) -> None:
+        # slices bypass the environment getitem method.
+        if isinstance(node.arg, nodes.Slice):
+            self.visit(node.node, frame)
+            self.write("[")
+            self.visit(node.arg, frame)
+            self.write("]")
+        else:
+            if self.environment.is_async:
+                self.write("(await auto_await(")
+
+            self.write("environment.getitem(")
+            self.visit(node.node, frame)
+            self.write(", ")
+            self.visit(node.arg, frame)
+            self.write(")")
+
+            if self.environment.is_async:
+                self.write("))")
+
+    def visit_Slice(self, node: nodes.Slice, frame: Frame) -> None:
+        if node.start is not None:
+            self.visit(node.start, frame)
+        self.write(":")
+        if node.stop is not None:
+            self.visit(node.stop, frame)
+        if node.step is not None:
+            self.write(":")
+            self.visit(node.step, frame)
+
+    @contextmanager
+    def _filter_test_common(
+        self, node: t.Union[nodes.Filter, nodes.Test], frame: Frame, is_filter: bool
+    ) -> t.Iterator[None]:
+        if self.environment.is_async:
+            self.write("(await auto_await(")
+
+        if is_filter:
+            self.write(f"{self.filters[node.name]}(")
+            func = self.environment.filters.get(node.name)
+        else:
+            self.write(f"{self.tests[node.name]}(")
+            func = self.environment.tests.get(node.name)
+
+        # When inside an If or CondExpr frame, allow the filter to be
+        # undefined at compile time and only raise an error if it's
+        # actually called at runtime. See pull_dependencies.
+        if func is None and not frame.soft_frame:
+            type_name = "filter" if is_filter else "test"
+            self.fail(f"No {type_name} named {node.name!r}.", node.lineno)
+
+        pass_arg = {
+            _PassArg.context: "context",
+            _PassArg.eval_context: "context.eval_ctx",
+            _PassArg.environment: "environment",
+        }.get(
+            _PassArg.from_obj(func)  # type: ignore
+        )
+
+        if pass_arg is not None:
+            self.write(f"{pass_arg}, ")
+
+        # Back to the visitor function to handle visiting the target of
+        # the filter or test.
+        yield
+
+        self.signature(node, frame)
+        self.write(")")
+
+        if self.environment.is_async:
+            self.write("))")
+
+    @optimizeconst
+    def visit_Filter(self, node: nodes.Filter, frame: Frame) -> None:
+        with self._filter_test_common(node, frame, True):
+            # if the filter node is None we are inside a filter block
+            # and want to write to the current buffer
+            if node.node is not None:
+                self.visit(node.node, frame)
+            elif frame.eval_ctx.volatile:
+                self.write(
+                    f"(Markup(concat({frame.buffer}))"
+                    f" if context.eval_ctx.autoescape else concat({frame.buffer}))"
+                )
+            elif frame.eval_ctx.autoescape:
+                self.write(f"Markup(concat({frame.buffer}))")
+            else:
+                self.write(f"concat({frame.buffer})")
+
+    @optimizeconst
+    def visit_Test(self, node: nodes.Test, frame: Frame) -> None:
+        with self._filter_test_common(node, frame, False):
+            self.visit(node.node, frame)
+
+    @optimizeconst
+    def visit_CondExpr(self, node: nodes.CondExpr, frame: Frame) -> None:
+        frame = frame.soft()
+
+        def write_expr2() -> None:
+            if node.expr2 is not None:
+                self.visit(node.expr2, frame)
+                return
+
+            self.write(
+                f'cond_expr_undefined("the inline if-expression on'
+                f" {self.position(node)} evaluated to false and no else"
+                f' section was defined.")'
+            )
+
+        self.write("(")
+        self.visit(node.expr1, frame)
+        self.write(" if ")
+        self.visit(node.test, frame)
+        self.write(" else ")
+        write_expr2()
+        self.write(")")
+
+    @optimizeconst
+    def visit_Call(
+        self, node: nodes.Call, frame: Frame, forward_caller: bool = False
+    ) -> None:
+        if self.environment.is_async:
+            self.write("(await auto_await(")
+        if self.environment.sandboxed:
+            self.write("environment.call(context, ")
+        else:
+            self.write("context.call(")
+        self.visit(node.node, frame)
+        extra_kwargs = {"caller": "caller"} if forward_caller else None
+        loop_kwargs = {"_loop_vars": "_loop_vars"} if frame.loop_frame else {}
+        block_kwargs = {"_block_vars": "_block_vars"} if frame.block_frame else {}
+        if extra_kwargs:
+            extra_kwargs.update(loop_kwargs, **block_kwargs)
+        elif loop_kwargs or block_kwargs:
+            extra_kwargs = dict(loop_kwargs, **block_kwargs)
+        self.signature(node, frame, extra_kwargs)
+        self.write(")")
+        if self.environment.is_async:
+            self.write("))")
+
+    def visit_Keyword(self, node: nodes.Keyword, frame: Frame) -> None:
+        self.write(node.key + "=")
+        self.visit(node.value, frame)
+
+    # -- Unused nodes for extensions
+
+    def visit_MarkSafe(self, node: nodes.MarkSafe, frame: Frame) -> None:
+        self.write("Markup(")
+        self.visit(node.expr, frame)
+        self.write(")")
+
+    def visit_MarkSafeIfAutoescape(
+        self, node: nodes.MarkSafeIfAutoescape, frame: Frame
+    ) -> None:
+        self.write("(Markup if context.eval_ctx.autoescape else identity)(")
+        self.visit(node.expr, frame)
+        self.write(")")
+
+    def visit_EnvironmentAttribute(
+        self, node: nodes.EnvironmentAttribute, frame: Frame
+    ) -> None:
+        self.write("environment." + node.name)
+
+    def visit_ExtensionAttribute(
+        self, node: nodes.ExtensionAttribute, frame: Frame
+    ) -> None:
+        self.write(f"environment.extensions[{node.identifier!r}].{node.name}")
+
+    def visit_ImportedName(self, node: nodes.ImportedName, frame: Frame) -> None:
+        self.write(self.import_aliases[node.importname])
+
+    def visit_InternalName(self, node: nodes.InternalName, frame: Frame) -> None:
+        self.write(node.name)
+
+    def visit_ContextReference(
+        self, node: nodes.ContextReference, frame: Frame
+    ) -> None:
+        self.write("context")
+
+    def visit_DerivedContextReference(
+        self, node: nodes.DerivedContextReference, frame: Frame
+    ) -> None:
+        self.write(self.derive_context(frame))
+
+    def visit_Continue(self, node: nodes.Continue, frame: Frame) -> None:
+        self.writeline("continue", node)
+
+    def visit_Break(self, node: nodes.Break, frame: Frame) -> None:
+        self.writeline("break", node)
+
+    def visit_Scope(self, node: nodes.Scope, frame: Frame) -> None:
+        scope_frame = frame.inner()
+        scope_frame.symbols.analyze_node(node)
+        self.enter_frame(scope_frame)
+        self.blockvisit(node.body, scope_frame)
+        self.leave_frame(scope_frame)
+
+    def visit_OverlayScope(self, node: nodes.OverlayScope, frame: Frame) -> None:
+        ctx = self.temporary_identifier()
+        self.writeline(f"{ctx} = {self.derive_context(frame)}")
+        self.writeline(f"{ctx}.vars = ")
+        self.visit(node.context, frame)
+        self.push_context_reference(ctx)
+
+        scope_frame = frame.inner(isolated=True)
+        scope_frame.symbols.analyze_node(node)
+        self.enter_frame(scope_frame)
+        self.blockvisit(node.body, scope_frame)
+        self.leave_frame(scope_frame)
+        self.pop_context_reference()
+
+    def visit_EvalContextModifier(
+        self, node: nodes.EvalContextModifier, frame: Frame
+    ) -> None:
+        for keyword in node.options:
+            self.writeline(f"context.eval_ctx.{keyword.key} = ")
+            self.visit(keyword.value, frame)
+            try:
+                val = keyword.value.as_const(frame.eval_ctx)
+            except nodes.Impossible:
+                frame.eval_ctx.volatile = True
+            else:
+                setattr(frame.eval_ctx, keyword.key, val)
+
+    def visit_ScopedEvalContextModifier(
+        self, node: nodes.ScopedEvalContextModifier, frame: Frame
+    ) -> None:
+        old_ctx_name = self.temporary_identifier()
+        saved_ctx = frame.eval_ctx.save()
+        self.writeline(f"{old_ctx_name} = context.eval_ctx.save()")
+        self.visit_EvalContextModifier(node, frame)
+        for child in node.body:
+            self.visit(child, frame)
+        frame.eval_ctx.revert(saved_ctx)
+        self.writeline(f"context.eval_ctx.revert({old_ctx_name})")