Getting Started with Python Internals

This article is a summary of what I learned from Philip Guo’s CPython internals: A ten-hour codewalk through the Python interpreter source code. I highly recommend his course. He go through great materials in his videos. You can think of this article as a companion text version of his course, so you can come back for your own references. The only major difference between Philip Guo’s course and this article is the Python version. In his course he was using 2.7.8. This article I’m using the 3.6.0 sourcecode. In this article, you’ll know the very basic things about Python internals, and, hopefully, be able to explore the Python internals on your own.

This post was originally shared on my personal blog — If you think syntax highlighting is a good thing, be sure to check out the original post.

What does CPython do to my code?

You might have heard people tell you Python is a interpreted language. You give the Python interpreter your sourcecode, and boom, the interpreter spits out the output.

However, this might be a little bit confusing, but Python does compile your code. Instead of the one step to get from the sourcecode to the output, Python compiles your sourcecode into an easier format — bytecodes, for the Python virtual machine. Then the interpreter program consumes the bytecode and act on it. We’ll be focusing on the later part where we get from bytecode to the output.

Here’s a simple diagram:

How does the CPython project look like?

Main sub-directories in the CPython project:

  1. Include/ - all the .h (header) files where the interfaces are defined
  2. Objects/ - the C code that represents Python objects
  3. Python/ - the main Python runtime

Other sub-directories:

  1. Modules/ - built-in modules implemented in C
  2. Libs/ - standard libraries implemented in Python

Python bytecode and the disassembler

Let’s define our testing module:

x = 1
y = 2
z = x + y

In Python’s interactive interpreter, we can invoke the built-in function: compile, and get the code object of this module.

c = c = compile(open('').read(), '', 'exec')
# <code object <module> at 0x..., file "", line 1>

# b'd\x00Z\x00d\x01Z\x01e\x00e\x01\x17\x00Z\x02e\x03e\x02\x83\x01\x01\x00d\x02S\x00'

[byte for byte in c.co_code]
# [100, 0, 90, 0, 100, 1, 90, 1, 101, 0, 101, 1, 23, 0, 90, 2, 101, 3, 101, 2, 131, 1, 1, 0, 100, 2, 83, 0]

The above code is the bytecode representation of our original sourcecode. However, looking at those numbers do not really help unless you are a Python guru. Fortunately, in Python’s standard library there’s a module called dis, which would translate those byte codes for us.

To disassemble the Python module, run $ python3 -m dis

  1           0 LOAD_CONST               0 (1)
2 STORE_NAME 0 (x)

2 4 LOAD_CONST 1 (2)
6 STORE_NAME 1 (y)

3 8 LOAD_NAME 0 (x)
10 LOAD_NAME 1 (y)
14 STORE_NAME 2 (z)

4 16 LOAD_NAME 3 (print)
18 LOAD_NAME 2 (z)
24 LOAD_CONST 2 (None)

The format for the disassembled code is:

line_number_in_sourcecode -> byte_offset OP_CODE -> internal_book_keeping_stuff (argument_name)

The byte code is mapped to this disassembled code, somehow, with some optimization. Thus, it’s not very obvious to me how to decode each byte code to which line of the disassembler output. So for now, we’ll be sticking with the more readable disassembler output.

If you are interested in how the disassembler works, module dis is located at /Lib/

Python is a Stack Machine

If you could quite understand the disassembler output, feel free to skip this section.

Before we jump into the disassembled code above, you need to know — Python virtual machine is a “Stack Machine.” Python internally keeps track of a “Value Stack,” which stores all the values for the upcoming operations to use. If you still don’t have any idea what a value stack is about, imaging Python is a lazy guy who always grabs the top-most T-shirt from his drawer. When he has more clean T-shirts, he simply stacks them on top of the other T-shirts. This guy sometimes may grab an extra T-shirt to go to the gym, sometimes may take several T-shirt for donation, or he may do any action with his T-shirts, but he is very discipline about picking the top-most T-shirt first. This guy is a Stack Machine.

Let’s look at the disassembled code above. Without knowing the internal, we could sort of guess what’s happening in the Python virtual machine already:

  • When a const is loaded by calling LOAD_CONST, the value is pushed onto the Value Stack.
  • When STORE_NAME is called, the last value on the Value Stack is popped and saved in the memory associated with the variable name, x for example. the 0 in line 3 STORE_NAME 0 (x) indicates that name x is the 0th variable name.
  • LOAD_NAME pushes whatever value the variable name is associated with on top of the Value Stack. (Only the reference of the value is pushed. So now the value has the Reference Counting of 2. One from the variable name, another from the Value Stack.)
  • BINARY_ADD pops the two values from the Value Stack and pushs the result on top of the Value Stack again.
  • CALL_FUNCTION calls the print function and use z as its argument. (In legacy Python, or Python 2, print was a primitive operation and has its own opcode.)
  • For completeness, the module returns a None value.

Python opcode

We sort of guess our way through the disassembled code. Let’s see how they are formally defined in CPython.

First look at /Include/opcode.h. This file defines all the Python opcodes, which Python could act upon. For example, line 78:

#define LOAD_CONST              100

This defines the opcode for LOAD_CONST.

One interesting opcode to notice is in line 68, HAVE_ARGUMENT. This opcode is just a placeholder. Any opcode include and above 90 takes argument(s).

Main Interpreter Loop

Now let’s see file /Python/ceval.c. Line 721 is the start of the gigantic interpreter main function, which ends at line 3692:

PyObject *
_PyEval_EvalFrameDefault(PyFrameObject *f, int throwflag)
} // line 3692

To understand this function:

  • Return type — Everything in Python is an object, an PyObject object in CPython code.
  • Input argument — A PyFrameObject is a piece of Python code.

One object worth noticing in this function is PyObject **stack_pointer in line 727. This object is a list of pointers pointing to the Value Stack, which we mentioned previously.

This is it, line 1108, the start of the main interpreter loop! This is the start of the infinite loop to go through the bytecode. This main interpreter loop ends at line 3614:

for (;;) {
} /* main loop */ // Line 3614

Now see line 1220, a GIANT switch case that tells you what C code to operate for each opcode case:

switch (opcode) {

Line 3528, breaking out of the main interpreter loop:

assert(why != WHY_NOT);

Line 3691, return of the main function:

return _Py_CheckFunctionResult(NULL, retval, "PyEval_EvalFrameEx");

And that is the structure of your CPython runtime! Now you know how to look at the CPython code to figure out what’s happening internally when your code is executing. For example, you can locate the switch case for LOAD_CONST in line 1244 to see what's happening when this opcode is used:

PyObject *value = GETITEM(consts, oparg);

What it does in high level is:

  1. Gets the value
  2. Increase the reference count of the value by 1
  3. Push the value on top of the Value Stack


Let’s recap some key points in this article:

  1. Python does compile your sourcecode.
  2. Use $ python3 -m dis <YOUR_PYTHON_FILE> to disassemble your code.
  3. Python is a stack machine.
  4. /Include/opcode.h has all the opcodes defined.
  5. CPython runtime’s main interpreter loop locates in /Python/ceval.c.
  6. The main interpreter loop is a giant switch case in an infinite loop.