# Copyright (c) Meta Platforms, Inc. and affiliates. # # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from dataclasses import dataclass, field from typing import List, Optional, Sequence from libcst import BaseStatement, CSTNode, Module from libcst._add_slots import add_slots from libcst._nodes.internal import CodegenState from libcst.metadata import BaseMetadataProvider class CodegenPartial: """ Provided by :class:`ExperimentalReentrantCodegenProvider`. Stores enough information to generate either a small patch (:meth:`get_modified_code_range`) or a new file (:meth:`get_modified_code`) by replacing the old node at this position. """ __slots__ = [ "start_offset", "end_offset", "has_trailing_newline", "_indent_tokens", "_prev_codegen_state", ] def __init__(self, state: "_ReentrantCodegenState") -> None: # store a frozen copy of these values, since they change over time self.start_offset: int = state.start_offset_stack[-1] self.end_offset: int = state.char_offset self.has_trailing_newline: bool = True # this may get updated to False later self._indent_tokens: Sequence[str] = tuple(state.indent_tokens) # everything else can be accessed from the codegen state object self._prev_codegen_state: _ReentrantCodegenState = state def get_original_module_code(self) -> str: """ Equivalent to :meth:`libcst.Module.bytes` on the top-level module that contains this statement, except that it uses the cached result from our previous code generation pass, so it's faster. """ return self._prev_codegen_state.get_code() def get_original_module_bytes(self) -> bytes: """ Equivalent to :meth:`libcst.Module.bytes` on the top-level module that contains this statement, except that it uses the cached result from our previous code generation pass, so it's faster. """ return self.get_original_module_code().encode(self._prev_codegen_state.encoding) def get_original_statement_code(self) -> str: """ Equivalent to :meth:`libcst.Module.code_for_node` on the current statement, except that it uses the cached result from our previous code generation pass, so it's faster. """ return self._prev_codegen_state.get_code()[self.start_offset : self.end_offset] def get_modified_statement_code(self, node: BaseStatement) -> str: """ Gets the new code for ``node`` as if it were in same location as the old statement being replaced. This means that it inherits details like the old statement's indentation. """ new_codegen_state = CodegenState( default_indent=self._prev_codegen_state.default_indent, default_newline=self._prev_codegen_state.default_newline, indent_tokens=list(self._indent_tokens), ) node._codegen(new_codegen_state) if not self.has_trailing_newline: new_codegen_state.pop_trailing_newline() return "".join(new_codegen_state.tokens) def get_modified_module_code(self, node: BaseStatement) -> str: """ Gets the new code for the module at the root of this statement's tree, but with the supplied replacement ``node`` in its place. """ original = self.get_original_module_code() patch = self.get_modified_statement_code(node) return f"{original[:self.start_offset]}{patch}{original[self.end_offset:]}" def get_modified_module_bytes(self, node: BaseStatement) -> bytes: """ Gets the new bytes for the module at the root of this statement's tree, but with the supplied replacement ``node`` in its place. """ return self.get_modified_module_code(node).encode( self._prev_codegen_state.encoding ) @add_slots @dataclass(frozen=False) class _ReentrantCodegenState(CodegenState): provider: BaseMetadataProvider[CodegenPartial] encoding: str = "utf-8" indent_size: int = 0 char_offset: int = 0 start_offset_stack: List[int] = field(default_factory=list) cached_code: Optional[str] = None trailing_partials: List[CodegenPartial] = field(default_factory=list) def increase_indent(self, value: str) -> None: super(_ReentrantCodegenState, self).increase_indent(value) self.indent_size += len(value) def decrease_indent(self) -> None: self.indent_size -= len(self.indent_tokens[-1]) super(_ReentrantCodegenState, self).decrease_indent() def add_indent_tokens(self) -> None: super(_ReentrantCodegenState, self).add_indent_tokens() self.char_offset += self.indent_size def add_token(self, value: str) -> None: super(_ReentrantCodegenState, self).add_token(value) self.char_offset += len(value) self.trailing_partials.clear() def before_codegen(self, node: CSTNode) -> None: if not isinstance(node, BaseStatement): return self.start_offset_stack.append(self.char_offset) def after_codegen(self, node: CSTNode) -> None: if not isinstance(node, BaseStatement): return partial = CodegenPartial(self) self.provider.set_metadata(node, partial) self.start_offset_stack.pop() self.trailing_partials.append(partial) def pop_trailing_newline(self) -> None: """ :class:`libcst.Module` contains a hack where it removes the last token (a newline) if the original file didn't have a newline. If this happens, we need to go back through every node at the end of the file, and fix their `end_offset`. """ for tp in self.trailing_partials: tp.end_offset -= len(self.tokens[-1]) tp.has_trailing_newline = False super(_ReentrantCodegenState, self).pop_trailing_newline() def get_code(self) -> str: # Ideally this would use functools.cached_property, but that's only in # Python 3.8+. # # This is a little ugly to make pyre's attribute refinement checks happy. cached_code = self.cached_code if cached_code is not None: return cached_code cached_code = "".join(self.tokens) self.cached_code = cached_code return cached_code class ExperimentalReentrantCodegenProvider(BaseMetadataProvider[CodegenPartial]): """ An experimental API that allows fast generation of modified code by recording an initial code-generation pass, and incrementally applying updates. It is a performance optimization for a few niche use-cases and is not user-friendly. **This API may change at any time without warning (including in minor releases).** This is rarely useful. Instead you should make multiple modifications to a single syntax tree, and generate the code once. However, we can think of a few use-cases for this API (hence, why it exists): - When linting a file, you might generate multiple independent patches that a user can accept or reject. Depending on your architecture, it may be advantageous to avoid regenerating the file when computing each patch. - You might want to call out to an external utility (e.g. a typechecker, such as pyre or mypy) to validate a small change. You may need to generate and test lots of these patches. Restrictions: - For safety and sanity reasons, the smallest/only level of granularity is a statement. If you need to patch part of a statement, you regenerate the entire statement. If you need to regenerate an entire module, just call :meth:`libcst.Module.code`. - This does not (currently) operate recursively. You can patch an unpatched piece of code multiple times, but you can't layer additional patches on an already patched piece of code. """ def _gen_impl(self, module: Module) -> None: state = _ReentrantCodegenState( default_indent=module.default_indent, default_newline=module.default_newline, provider=self, encoding=module.encoding, ) module._codegen(state)