Design of CPython’s Compiler#

Abstract#

In CPython, the compilation from source code to bytecode involves several steps:

  1. Tokenize the source code (Parser/tokenizer.c)

  2. Parse the stream of tokens into an Abstract Syntax Tree (Parser/parser.c)

  3. Transform AST into a Control Flow Graph (Python/compile.c)

  4. Emit bytecode based on the Control Flow Graph (Python/compile.c)

The purpose of this document is to outline how these steps of the process work.

This document does not touch on how parsing works beyond what is needed to explain what is needed for compilation. It is also not exhaustive in terms of the how the entire system works. You will most likely need to read some source to have an exact understanding of all details.

Parsing#

As of Python 3.9, Python’s parser is a PEG parser of a somewhat unusual design (since its input is a stream of tokens rather than a stream of characters as is more common with PEG parsers).

The grammar file for Python can be found in Grammar/python.gram. The definitions for literal tokens (such as :, numbers, etc.) can be found in Grammar/Tokens. Various C files, including Parser/parser.c are generated from these (see Changing CPython’s Grammar).

Abstract Syntax Trees (AST)#

The abstract syntax tree (AST) is a high-level representation of the program structure without the necessity of containing the source code; it can be thought of as an abstract representation of the source code. The specification of the AST nodes is specified using the Zephyr Abstract Syntax Definition Language (ASDL) [Wang97].

The definition of the AST nodes for Python is found in the file Parser/Python.asdl.

Each AST node (representing statements, expressions, and several specialized types, like list comprehensions and exception handlers) is defined by the ASDL. Most definitions in the AST correspond to a particular source construct, such as an ‘if’ statement or an attribute lookup. The definition is independent of its realization in any particular programming language.

The following fragment of the Python ASDL construct demonstrates the approach and syntax:

module Python
{
    stmt = FunctionDef(identifier name, arguments args, stmt* body,
                       expr* decorators)
           | Return(expr? value) | Yield(expr? value)
           attributes (int lineno)
}

The preceding example describes two different kinds of statements and an expression: function definitions, return statements, and yield expressions. All three kinds are considered of type stmt as shown by | separating the various kinds. They all take arguments of various kinds and amounts.

Modifiers on the argument type specify the number of values needed; ? means it is optional, * means 0 or more, while no modifier means only one value for the argument and it is required. FunctionDef, for instance, takes an identifier for the name, arguments for args, zero or more stmt arguments for body, and zero or more expr arguments for decorators.

Do notice that something like ‘arguments’, which is a node type, is represented as a single AST node and not as a sequence of nodes as with stmt as one might expect.

All three kinds also have an ‘attributes’ argument; this is shown by the fact that ‘attributes’ lacks a ‘|’ before it.

The statement definitions above generate the following C structure type:

typedef struct _stmt *stmt_ty;

struct _stmt {
      enum { FunctionDef_kind=1, Return_kind=2, Yield_kind=3 } kind;
      union {
              struct {
                      identifier name;
                      arguments_ty args;
                      asdl_seq *body;
              } FunctionDef;

              struct {
                      expr_ty value;
              } Return;

              struct {
                      expr_ty value;
              } Yield;
      } v;
      int lineno;
 }

Also generated are a series of constructor functions that allocate (in this case) a stmt_ty struct with the appropriate initialization. The kind field specifies which component of the union is initialized. The FunctionDef() constructor function sets ‘kind’ to FunctionDef_kind and initializes the name, args, body, and attributes fields.

Memory Management#

Before discussing the actual implementation of the compiler, a discussion of how memory is handled is in order. To make memory management simple, an arena is used. This means that a memory is pooled in a single location for easy allocation and removal. What this gives us is the removal of explicit memory deallocation. Because memory allocation for all needed memory in the compiler registers that memory with the arena, a single call to free the arena is all that is needed to completely free all memory used by the compiler.

In general, unless you are working on the critical core of the compiler, memory management can be completely ignored. But if you are working at either the very beginning of the compiler or the end, you need to care about how the arena works. All code relating to the arena is in either Include/Internal/pycore_pyarena.h or Python/pyarena.c.

PyArena_New() will create a new arena. The returned PyArena structure will store pointers to all memory given to it. This does the bookkeeping of what memory needs to be freed when the compiler is finished with the memory it used. That freeing is done with PyArena_Free(). This only needs to be called in strategic areas where the compiler exits.

As stated above, in general you should not have to worry about memory management when working on the compiler. The technical details have been designed to be hidden from you for most cases.

The only exception comes about when managing a PyObject. Since the rest of Python uses reference counting, there is extra support added to the arena to cleanup each PyObject that was allocated. These cases are very rare. However, if you’ve allocated a PyObject, you must tell the arena about it by calling PyArena_AddPyObject().

Source Code to AST#

The AST is generated from source code using the function _PyParser_ASTFromString() or _PyParser_ASTFromFile() (from Parser/peg_api.c) depending on the input type.

After some checks, a helper function in Parser/parser.c begins applying production rules on the source code it receives; converting source code to tokens and matching these tokens recursively to their corresponding rule. The rule’s corresponding rule function is called on every match. These rule functions follow the format xx_rule. Where xx is the grammar rule that the function handles and is automatically derived from Grammar/python.gram by Tools/peg_generator/pegen/c_generator.py.

Each rule function in turn creates an AST node as it goes along. It does this by allocating all the new nodes it needs, calling the proper AST node creation functions for any required supporting functions and connecting them as needed. This continues until all nonterminal symbols are replaced with terminals. If an error occurs, the rule functions backtrack and try another rule function. If there are no more rules, an error is set and the parsing ends.

The AST node creation helper functions have the name _PyAST_xx where xx is the AST node that the function creates. These are defined by the ASDL grammar and contained in Python/Python-ast.c (which is generated by Parser/asdl_c.py from Parser/Python.asdl). This all leads to a sequence of AST nodes stored in asdl_seq structs.

To demonstrate everything explained so far, here’s the rule function responsible for a simple named import statement such as import sys. Note that error-checking and debugging code has been omitted. Removed parts are represented by .... Furthermore, some comments have been added for explanation. These comments may not be present in the actual code.

// This is the production rule (from python.gram) the rule function
// corresponds to:
// import_name: 'import' dotted_as_names
static stmt_ty
import_name_rule(Parser *p)
{
    ...
    stmt_ty _res = NULL;
    { // 'import' dotted_as_names
        ...
        Token * _keyword;
        asdl_alias_seq* a;
        // The tokenizing steps.
        if (
            (_keyword = _PyPegen_expect_token(p, 513))  // token='import'
            &&
            (a = dotted_as_names_rule(p))  // dotted_as_names
        )
        {
            ...
            // Generate an AST for the import statement.
            _res = _PyAST_Import ( a , ...);
            ...
            goto done;
        }
        ...
    }
    _res = NULL;
  done:
    ...
    return _res;
}

To improve backtracking performance, some rules (chosen by applying a (memo) flag in the grammar file) are memoized. Each rule function checks if a memoized version exists and returns that if so, else it continues in the manner stated in the previous paragraphs.

There are macros for creating and using asdl_xx_seq * types, where xx is a type of the ASDL sequence. Three main types are defined manually – generic, identifier and int. These types are found in Python/asdl.c and its corresponding header file Include/Internal/pycore_asdl.h. Functions and macros for creating asdl_xx_seq * types are as follows:

_Py_asdl_generic_seq_new(Py_ssize_t, PyArena *)

Allocate memory for an asdl_generic_seq of the specified length

_Py_asdl_identifier_seq_new(Py_ssize_t, PyArena *)

Allocate memory for an asdl_identifier_seq of the specified length

_Py_asdl_int_seq_new(Py_ssize_t, PyArena *)

Allocate memory for an asdl_int_seq of the specified length

In addition to the three types mentioned above, some ASDL sequence types are automatically generated by Parser/asdl_c.py and found in Include/Internal/pycore_ast.h. Macros for using both manually defined and automatically generated ASDL sequence types are as follows:

asdl_seq_GET(asdl_xx_seq *, int)

Get item held at a specific position in an asdl_xx_seq

asdl_seq_SET(asdl_xx_seq *, int, stmt_ty)

Set a specific index in an asdl_xx_seq to the specified value

Untyped counterparts exist for some of the typed macros. These are useful when a function needs to manipulate a generic ASDL sequence:

asdl_seq_GET_UNTYPED(asdl_seq *, int)

Get item held at a specific position in an asdl_seq

asdl_seq_SET_UNTYPED(asdl_seq *, int, stmt_ty)

Set a specific index in an asdl_seq to the specified value

asdl_seq_LEN(asdl_seq *)

Return the length of an asdl_seq or asdl_xx_seq

Note that typed macros and functions are recommended over their untyped counterparts. Typed macros carry out checks in debug mode and aid debugging errors caused by incorrectly casting from void *.

If you are working with statements, you must also worry about keeping track of what line number generated the statement. Currently the line number is passed as the last parameter to each stmt_ty function.

Changed in version 3.9: The new PEG parser generates an AST directly without creating a parse tree. Python/ast.c is now only used to validate the AST for debugging purposes.

See also

PEP 617 (PEP 617 – New PEG parser for CPython)

Control Flow Graphs#

A control flow graph (often referenced by its acronym, CFG) is a directed graph that models the flow of a program. A node of a CFG is not an individual bytecode instruction, but instead represents a sequence of bytecode instructions that always execute sequentially. Each node is called a basic block and must always execute from start to finish, with a single entry point at the beginning and a single exit point at the end. If some bytecode instruction a needs to jump to some other bytecode instruction