The PEG parser produces locations for an AST that span the entire AST node. This is well defined, and has a number of desirable properties such as the span of a node including all its children's spans.
Unfortunately, this does not produce good locations for tracing and debugging, so we transform some locations in the compiler.
This is problematic as it means that tools producing as AST (e.g. pytest) need to second guess what transformations that the compiler will make.
I proposed moving these transformations in the parser in https://github.com/python/cpython/issues/94694, but as @pablogsal points out, this means that tools which consume the AST (e.g. type checkers) need to undo the transformations to get the original spans back.
The core problem is that there are tools that consume the AST produced by the parser, and tools that produce ASTs for consumption by the compiler, and they want subtly different ASTs.
One possible fix for this is to add a keyword argument to compile() to specify whether it should modify the locations. This would allow pytest, etc. to pass location information through compile() unmodified when needed.
brandtbucher
commented
2 years ago
The PEG parser produces locations for an AST that span the entire AST node. This is well defined, and has a number of desirable properties such as the span of a node including all its children's spans.
Unfortunately, this does not produce good locations for tracing and debugging, so we transform some locations in the compiler. This is problematic as it means that tools producing as AST (e.g. pytest) need to second guess what transformations that the compiler will make.
I proposed moving these transformations in the parser in https://github.com/python/cpython/issues/94694, but as @pablogsal points out, this means that tools which consume the AST (e.g. type checkers) need to undo the transformations to get the original spans back.
The core problem is that there are tools that consume the AST produced by the parser, and tools that produce ASTs for consumption by the compiler, and they want subtly different ASTs.
One possible fix for this is to add a keyword argument to
compile()to specify whether it should modify the locations. This would allow pytest, etc. to pass location information throughcompile()unmodified when needed.