PEP: 638 Title: Syntactic Macros Author: Mark Shannon Status: Draft Type: Standards Track Content-Type: text/x-rst Created: 24-Sep-2020 Abstract ======== This PEP adds support for syntactic macros to Python. A macro is a compile-time function that transforms a part of the program to allow functionality that cannot be expressed cleanly in normal library code. The term "syntactic" means that this sort of macro operates on the program's syntax tree. This reduces the chance of mistranslation that can happen with text-based substitution macros, and allows the implementation of `hygienic macros`__. __ https://en.wikipedia.org/wiki/Hygienic_macro Syntactic macros allow libraries to modify the abstract syntax tree during compilation, providing the ability to extend the language for specific domains without adding to complexity to the language as a whole. Motivation ========== New language features can be controversial, disruptive and sometimes divisive. Python is now sufficiently powerful and complex, that many proposed additions are a net loss for the language due to the additional complexity. Although a language change may make certain patterns easy to express, it will have a cost. Each new feature makes the language larger, harder to learn and harder to understand. Python was once described as `Python Fits Your Brain`__, but that becomes less and less true as more and more features are added. Because of the high cost of adding a new feature, it is very difficult or impossible to add a feature that would benefit only some users, regardless of how many users, or how beneficial that feature would be to them. The use of Python in data science and machine learning has grown very rapidly over the last few years. However, most of the core developers of Python do not have a background in data science or machine learning. This makes it extremely difficult for the core developers to determine if a language extension for machine learning is worthwhile. By allowing language extensions to be modular and distributable, like libraries, domain specific extensions can be implemented without negatively impacting users outside of that domain. A web developer is likely to want a very different set of extensions from a data scientist. We need to let the community develop their own extensions. Without some form of user defined language extensions, there will be a constant battle between those wanting to keep the language compact and fitting their brains, and those wanting a new feature that suits their domain or programming style. __ https://www.linuxjournal.com/article/4731 Improving the expressiveness of libraries for specific domains '''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''' Many domains see repeated patterns that are difficult or impossible to express as a library. Macros can allow those patterns to be expressed in a more concise and less error prone way. Trialing new language features '''''''''''''''''''''''''''''' It is possible to demonstrate potential language extensions using macros. For example, macros would have enabled the ``with`` statement and ``yield from`` expression to have been trialed. Doing so might well have lead to a higher quality implementation at first release, by allowing more testing before those features were included in the language. It is nearly impossible to make sure that a new feature is completely reliable before it is released; bugs relating to the ``with`` and ``yield from`` features were still being fixed many years after they were released. Long term stability for the bytecode interpreter '''''''''''''''''''''''''''''''''''''''''''''''' Historically new language features have been implemented by naive compilation of the AST into new, complex bytecode instructions. Those bytecodes have often had their own internal flow-control, performing operations that that could, and should, have been done in the compiler. For example, until recently flow control within the ``try``-``finally`` and ``with`` statements was managed by complicated bytecodes with context dependent semantics. The control flow within those statements is now implemented in the compiler, making the interpreter simpler and faster. By implementing new features as AST transformations, the existing compiler can generate the bytecode for a feature without having to modify the interpreter. A stable interpreter is necessary if we are to improve the performance and portability of the CPython VM. Rationale ========= Python is both expressive and easy to learn; it is widely recognized as the easiest to learn widely-used programming language. However, it is not the most flexible. That title belongs to lisp. Because lisp is homoiconic, meaning that lisp programs are lisp data-structures, lisp programs can be manipulated by lisp programs. Thus much of the language can be defined in itself. We would like that ability in Python, without the many parentheses that characterize lisp. Fortunately, homoiconicity is not needed for a language to be able to manipulate itself, all that is needed is the ability to manipulate programs after parsing, but before translation to an executable form. Python already has the components needed. The syntax tree of Python is available through the ``ast`` module. All that is needed is a marker to tell the compiler that a macro is present, and the ability for the compiler to callback into user code to manipulate the AST. Specification ============= Syntax '''''' Lexical analysis ~~~~~~~~~~~~~~~~ Any sequence of identifier characters followed by an exclamation point (exclamation mark, UK English) will be tokenized as a ``MACRO_NAME``. Statement form ~~~~~~~~~~~~~~ :: macro_stmt = MACRO_NAME testlist [ "import" NAME ] [ "as" NAME ] [ ":" NEWLINE suite ] Expression form ~~~~~~~~~~~~~~~ :: macro_expr = MACRO_NAME "(" testlist ")" Resolving ambiguity ~~~~~~~~~~~~~~~~~~~ The statement form of a macro takes precedence, so that the code ``macro_name!(x)`` will be parsed as a macro statement, not as an expression statement containing a macro expression. Semantics ''''''''' Compilation ~~~~~~~~~~~ Upon encountering a ``macro`` during translation to bytecode, the code generator will look up the macro processor registered for the macro, and pass the AST, rooted at the macro to the processor function. The returned AST will then be substituted for the original tree. For macros with multiple names, several trees will be passed to the macro processor, but only one will be returned and substituted, shorting the enclosing block of statements. This process can be repeated, to enable macros to return AST nodes including other macros. The compiler will not look up a macro processor until that macro is reached, so that inner macros do not need to have processors registered. For example, in a ``switch`` macro, the ``case`` and ``default`` macros wouldn't need processors registered as they would be eliminated by the ``switch`` processor. To enable definition of macros to be imported, the macros ``import!`` and ``from!`` are predefined. They support the following syntax: :: "import!" dotted_name "as" name "from!" dotted_name "import" name [ "as" name ] The ``import!`` macro performs a compile time import of ``dotted_name`` to find the macro processor, then registers it under ``name`` for the scope currently being compiled. The ``from!`` macro performs a compile time import of ``dotted_name.name`` to find the macro processor, then registers it under ``name`` (using the ``name`` following "as", if present) for the scope currently being compiled. Note that, since ``import!`` and ``from!`` only define the macro for the scope in which the import is present, all uses of a macro must be preceded by an explicit ``import!`` or ``from!`` to improve clarity. For example, to import the macro "compile" from "my.compiler": :: from! my.compiler import compile Defining macro processors ~~~~~~~~~~~~~~~~~~~~~~~~~ A macro processor is defined by a four-tuple, consisting of ``(func, kind, version, additional_names)`` * ``func`` must be a callable that takes ``len(additional_names)+1`` arguments, all of which are abstract syntax trees, and returns a single abstract syntax tree. * ``kind`` must be one of the following: * ``macros.STMT_MACRO`` A statement macro where the body of the macro is indented. This is the only form which is allowed to have additional names. * ``macros.SIBLING_MACRO`` A statement macro where the body of the macro is the next statement is the same block. The following statement is moved into the macro as its body. * ``macros.EXPR_MACRO`` An expression macro. * ``version`` is used to track versions of macros, so that generated bytecodes can be correctly cached. It must be an integer. * ``additional_names`` are the names of the additional parts of the macro, and must be a tuple of strings. :: # (func, _ast.STMT_MACRO, VERSION, ()) stmt_macro!: multi_statement_body # (func, _ast.SIBLING_MACRO, VERSION, ()) sibling_macro! single_statement_body # (func, _ast.EXPR_MACRO, VERSION, ()) x = expr_macro!(...) # (func, _ast.STMT_MACRO, VERSION, ("subsequent_macro_part",)) multi_part_macro!: multi_statement_body subsequent_macro_part!: multi_statement_body The compiler will check that the syntax used matches the declared kind. For convenience, the decorator ``macro_processor`` is provided in the ``macros`` module to mark a function as a macro processor: :: def macro_processor(kind, version, *additional_names): def deco(func): return func, kind, version, additional_names return deco Which can be used to help declare macro processors, for example: :: @macros.macro_processor(macros.STMT_MACRO, 1_08) def switch(astnode): ... AST extensions ~~~~~~~~~~~~~~ Two new AST nodes will be needed to express macros, ``macro_stmt`` and ``macro_expr``. :: class macro_stmt(_ast.stmt): _fields = "name", "args", "importname", "asname", "body" class macro_expr(_ast.expr): _fields = "name", "args" In addition, macro processors will needs a means to express control flow or side effecting code, that produces a value. To support this, a new ast node, called ``stmt_expr``, that combines a statement and expression will be added. This new ast node will be a subtype of ``expr``, but include a statement to allow side effects. It will be compiled to bytecode by compiling the statement, then compiling the value. :: class stmt_expr(_ast.expr): _fields = "stmt", "value" Hygiene and debugging ~~~~~~~~~~~~~~~~~~~~~ Macros processors will often need to create new variables. Those variables need to named in such as way as to avoid contaminating the original code and other macros. No rules for naming will be enforced, but to ensure hygiene and help debugging, the following naming scheme is recommended: * All generated variable names should start with a ``$`` * Purely artificial variable names should start ``$$mname`` where ``mname`` is the name of the macro. * Variables derived from real variables should start ``$vname`` where ``vname`` is the name of the variable. * All variable names should include the line number and the column offset, separated by an underscore. Examples: * Purely generated name: ``$$macro_17_0`` * Name derived from a variable for an expression macro: ``$var_12_5`` Examples '''''''' Compile time checked data structures ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ It is common to encode tables of data in Python as large dictionaries. However, these can be hard to maintain and error prone. Macros allow this data to be written in a more readable format. Then, at compile time, it can be verified and converted to an efficient format. For example, suppose we have a two dictionary literals mapping codes to names, and vice versa. This is error prone, as the dictionaries may have duplicate keys, or one table may not be the inverse of the other. A macro could generate the two mappings from a single table and, at the same time, verify that no duplicates are present. :: color_to_code = { "red": 1, "blue": 2, "green": 3, } code_to_color = { 1: "red", 2: "blue", 3: "yellow", # error } would become: :: bijection! color_to_code, code_to_color: "red" = 1 "blue" = 2 "green" = 3 Domain specific extensions ~~~~~~~~~~~~~~~~~~~~~~~~~~ Where I see macros having real value is in specific domains, not in general purpose language features. For example, parsers. Here's part of a parser definition for Python, using macros: :: choice! single_input: NEWLINE simple_stmt sequence!: compound_stmt NEWLINE Compilers ~~~~~~~~~ Runtime compilers, such as ``numba`` have to reconstitute the Python source, or attempt to analyze the bytecode. It would be simpler and more reliable for them to get the AST directly: :: from! my.jit.library import jit jit! def func(): ... Matching symbolic expressions ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ When matching something representing syntax, such a Python ``ast`` node, or a ``sympy`` expression, it is convenient to match against the actual syntax, not the data structure representing it. For example, a calculator could be implemented using a domain specific macro for matching syntax: :: from! ast_matcher import match def calculate(node): if isinstance(node, Num): return node.n match! node: case! a + b: return calculate(a) + calculate(b) case! a - b: return calculate(a) - calculate(b) case! a * b: return calculate(a) * calculate(b) case! a / b: return calculate(a) / calculate(b) Which could be converted to: :: def calculate(node): if isinstance(node, Num): return node.n $$match_4_0 = node if isinstance($$match_4_0, _ast.Add): a, b = $$match_4_0.left, $$match_4_0.right return calculate(a) + calculate(b) elif isinstance($$match_4_0, _ast.Sub): a, b = $$match_4_0.left, $$match_4_0.right return calculate(a) - calculate(b) elif isinstance($$match_4_0, _ast.Mul): a, b = $$match_4_0.left, $$match_4_0.right return calculate(a) * calculate(b) elif isinstance($$match_4_0, _ast.Div): a, b = $$match_4_0.left, $$match_4_0.right return calculate(a) / calculate(b) Zero cost markers and annotations ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Annotations, either decorators or PEP 3107 function annotations, have a runtime cost even if they serve only as markers for checkers or as documentation. :: @do_nothing_marker def foo(...): ... can be replaced with the zero cost macro: :: do_nothing_marker!: def foo(...): ... Protyping language extensions ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Although macros would be most valuable for domain specific extensions, it is possible to demonstrate possible language extensions using macros. f-strings: .......... The f-string ``f"..."`` could be implemented as macro as ``f!("...")``. Which is not quite as nice to read, but would still be useful for experimenting with. Try finally statement: ...................... :: try_!: body finally!: closing Would be translated roughly as: :: try: body except: closing else: closing Note: Care must be taken to handle returns, breaks and continues correctly. The above code is merely illustrative. With statement: ............... :: with! open(filename) as fd: return fd.read() The above would require handling ``open`` specially. An alternative that would be more explicit, would be: :: with! open!(filename) as fd: return fd.read() Macro definition macros ~~~~~~~~~~~~~~~~~~~~~~~ Languages that have syntactic macros usually provide a macro for defining macros. This PEP intentionally does not do that, as it is not yet clear what a good design would be, and we want to allow the community to define their own macros. One possible form could be: :: macro_def! name: input: ... # input pattern, defining meta-variables output: ... # output pattern, using meta-variables Backwards Compatibility ======================= This PEP is fully backwards compatible. Performance Implications ======================== For code that doesn't use macros, there will be no effect on performance. For code that does use macros and has already been compiled to bytecode, there will be some slight overhead to check that the version of macros used to compile the code match the imported macro processors. For code that has not been compiled, or compiled with different versions of the macro processors, then there would be the usual overhead of bytecode compilation, plus any additional overhead of macro processing. It is worth noting that the speed of source to bytecode compilation is largely irrelevant for Python performance. Implementation ============== In order to allow transformation of the AST at compile time by Python code, all AST nodes in the compiler will have to be Python objects. To do that efficiently, will mean making all the nodes in the ``_ast`` module immutable, so as not degrade performance by much. They will need to be immutable to guarantee that the AST remains a *tree* to avoid having to support cyclic GC. Making them immutable means they will not have a ``__dict__`` attribute, making them compact. AST nodes in the ``ast`` module will remain mutable. Currently, all AST nodes are allocated using an arena allocator. Changing to use the standard allocator might slow compilation down a little, but has advantages in terms of maintenance, as much code can be deleted. Reference Implementation '''''''''''''''''''''''' None as yet. Copyright ========= This document is placed in the public domain or under the CC0-1.0-Universal license, whichever is more permissive. .. Local Variables: mode: indented-text indent-tabs-mode: nil sentence-end-double-space: t fill-column: 70 coding: utf-8 End: