Compiler design#


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 an instruction sequence (Python/compile.c)

  4. Construct a Control Flow Graph and apply optimizations to it (Python/flowgraph.c)

  5. Emit bytecode based on the Control Flow Graph (Python/assemble.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.


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 Tools/peg_generator/pegen/

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/ 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;
    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/ 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 b, then a must occur at the end of its basic block, and b must occur at the start of its basic block.

As an example, consider the following code snippet:

if x < 10:

The x < 10 guard is represented by its own basic block that compares x with 10 and then ends in a conditional jump based on the result of the comparison. This conditional jump allows the block to point to both the body of the if and the body of the else. The if basic block contains the f1() and f2() calls and points to the end() basic block. The else basic block contains the g() call and similarly points to the end() block.

Note that more complex code in the guard, the if body, or the else body may be represented by multiple basic blocks. For instance, short-circuiting boolean logic in a guard like if x or y: will produce one basic block that tests the truth value of x and then points both (1) to the start of the if body and (2) to a different basic block that tests the truth value of y.

CFGs are usually one step away from final code output. Code is directly generated from the basic blocks (with jump targets adjusted based on the output order) by doing a post-order depth-first search on the CFG following the edges.

AST to CFG to bytecode#

With the AST created, the next step is to create the CFG. The first step is to convert the AST to Python bytecode without having jump targets resolved to specific offsets (this is calculated when the CFG goes to final bytecode). Essentially, this transforms the AST into Python bytecode with control flow represented by the edges of the CFG.

Conversion is done in two passes. The first creates the namespace (variables can be classified as local, free/cell for closures, or global). With that done, the second pass essentially flattens the CFG into a list and calculates jump offsets for final output of bytecode.

The conversion process is initiated by a call to the function _PyAST_Compile() in Python/compile.c. This function does both the conversion of the AST to a CFG and outputting final bytecode from the CFG. The AST to CFG step is handled mostly by two functions called by _PyAST_Compile(); _PySymtable_Build() and compiler_mod(). The former is in Python/symtable.c while the latter is Python/compile.c.

_PySymtable_Build() begins by entering the starting code block for the AST (passed-in) and then calling the proper symtable_visit_xx function (with xx being the AST node type). Next, the AST tree is walked with the various code blocks that delineate the reach of a local variable as blocks are entered and exited using symtable_enter_block() and symtable_exit_block(), respectively.

Once the symbol table is created, it is time for CFG creation, whose code is in Python/compile.c. This is handled by several functions that break the task down by various AST node types. The functions are all named compiler_visit_xx where xx is the name of the node type (such as stmt, expr, etc.). Each function receives a struct compiler * and xx_ty where xx is the AST node type. Typically these functions consist of a large ‘switch’ statement, branching based on the kind of node type passed to it. Simple things are handled inline in the ‘switch’ statement with more complex transformations farmed out to other functions named compiler_xx with xx being a descriptive name of what is being handled.

When transforming an arbitrary AST node, use the VISIT() macro. The appropriate compiler_visit_xx function is called, based on the value passed in for <node type> (so VISIT(c, expr, node) calls compiler_visit_expr(c, node)). The VISIT_SEQ() macro is very similar, but is called on AST node sequences (those values that were created as arguments to a node that used the ‘*’ modifier). There is also VISIT_SLICE() just for handling slices.

Emission of bytecode is handled by the following macros:

ADDOP(struct compiler *, int)

add a specified opcode

ADDOP_NOLINE(struct compiler *, int)

like ADDOP without a line number; used for artificial opcodes without no corresponding token in the source code

ADDOP_IN_SCOPE(struct compiler *, int)

like ADDOP, but also exits current scope; used for adding return value opcodes in lambdas and closures

ADDOP_I(struct compiler *, int, Py_ssize_t)

add an opcode that takes an integer argument

ADDOP_O(struct compiler *, int, PyObject *, TYPE)

add an opcode with the proper argument based on the position of the specified PyObject in PyObject sequence object, but with no handling of mangled names; used for when you need to do named lookups of objects such as globals, consts, or parameters where name mangling is not possible and the scope of the name is known; TYPE is the name of PyObject sequence (names or varnames)

ADDOP_N(struct compiler *, int, PyObject *, TYPE)

just like ADDOP_O, but steals a reference to PyObject

ADDOP_NAME(struct compiler *, int, PyObject *, TYPE)

just like ADDOP_O, but name mangling is also handled; used for attribute loading or importing based on name

ADDOP_LOAD_CONST(struct compiler *, PyObject *)

add the LOAD_CONST opcode with the proper argument based on the position of the specified PyObject in the consts table.

ADDOP_LOAD_CONST_NEW(struct compiler *, PyObject *)

just like ADDOP_LOAD_CONST_NEW, but steals a reference to PyObject

ADDOP_JUMP(struct compiler *, int, basicblock *)

create a jump to a basic block

ADDOP_JUMP_NOLINE(struct compiler *, int, basicblock *)

like ADDOP_JUMP without a line number; used for artificial jumps without no corresponding token in the source code.

ADDOP_JUMP_COMPARE(struct compiler *, cmpop_ty)

depending on the second argument, add an ADDOP_I with either an IS_OP, CONTAINS_OP, or COMPARE_OP opcode.

Several helper functions that will emit bytecode and are named compiler_xx() where xx is what the function helps with (list, boolop, etc.). A rather useful one is compiler_nameop(). This function looks up the scope of a variable and, based on the expression context, emits the proper opcode to load, store, or delete the variable.

As for handling the line number on which a statement is defined, this is handled by compiler_visit_stmt() and thus is not a worry.

Once the CFG is created, it must be flattened and then final emission of bytecode occurs. Flattening is handled using a post-order depth-first search. Once flattened, jump offsets are backpatched based on the flattening and then a PyCodeObject is created. All of this is handled by calling assemble().

Introducing new bytecode#

Sometimes a new feature requires a new opcode. But adding new bytecode is not as simple as just suddenly introducing new bytecode in the AST -> bytecode step of the compiler. Several pieces of code throughout Python depend on having correct information about what bytecode exists.

First, you must choose a name, implement the bytecode in Python/bytecodes.c, and add a documentation entry in Doc/library/dis.rst. Then run make regen-cases to assign a number for it (see Include/opcode_ids.h) and regenerate a number of files with the actual implementation of the bytecodes (Python/generated_cases.c.h) and additional files with metadata about them.

With a new bytecode you must also change what is called the magic number for .pyc files. The variable MAGIC_NUMBER in Lib/importlib/ contains the number. Changing this number will lead to all .pyc files with the old MAGIC_NUMBER to be recompiled by the interpreter on import. Whenever MAGIC_NUMBER is changed, the ranges in the magic_values array in PC/launcher.c must also be updated. Changes to Lib/importlib/ will take effect only after running make regen-importlib. Running this command before adding the new bytecode target to Python/bytecodes.c (followed by make regen-cases) will result in an error. You should only run make regen-importlib after the new bytecode target has been added.


On Windows, running the ./build.bat script will automatically regenerate the required files without requiring additional arguments.

Finally, you need to introduce the use of the new bytecode. Altering Python/compile.c, Python/bytecodes.c will be the primary places to change. Optimizations in Python/flowgraph.c may also need to be updated. If the new opcode affects a control flow or the block stack, you may have to update the frame_setlineno() function in Objects/frameobject.c. Lib/ may need an update if the new opcode interprets its argument in a special way (like FORMAT_VALUE or MAKE_FUNCTION).

If you make a change here that can affect the output of bytecode that is already in existence and you do not change the magic number constantly, make sure to delete your old .py(c|o) files! Even though you will end up changing the magic number if you change the bytecode, while you are debugging your work you will be changing the bytecode output without constantly bumping up the magic number. This means you end up with stale .pyc files that will not be recreated. Running find . -name '*.py[co]' -exec rm -f '{}' + should delete all .pyc files you have, forcing new ones to be created and thus allow you test out your new bytecode properly. Run make regen-importlib for updating the bytecode of frozen importlib files. You have to run make again after this for recompiling generated C files.

Code objects#

The result of PyAST_CompileObject() is a PyCodeObject which is defined in Include/cpython/code.h. And with that you now have executable Python bytecode!

The code objects (byte code) are executed in Python/ceval.c. This file will also need a new case statement for the new opcode in the big switch statement in _PyEval_EvalFrameDefault().

Important files#



Daniel C. Wang, Andrew W. Appel, Jeff L. Korn, and Chris S. Serra. The Zephyr Abstract Syntax Description Language. In Proceedings of the Conference on Domain-Specific Languages, pp. 213–227, 1997.