Peeking Behind the Bytecode Curtain

When the CPython interpreter executes your program, it first translates it into a sequence of bytecode instructions. Bytecode is an intermediate language for the Python virtual machine that’s used as a performance optimization.

Instead of directly executing the human-readable source code, compact numeric codes, constants, and references are used that represent the result of compiler parsing and semantic analysis.

This saves time and memory for repeated executions of programs or parts of programs. For example, the bytecode resulting from this compilation step is cached on disk in .pyc and .pyo files so that executing the same Python file is faster the second time around.
意思就是偏向于机器码的速度更快

All of this is completely transparent to the programmer. You don’t have to be aware that this intermediate translation step happens, or how the Python virtual machine deals with the bytecode. In fact, the bytecode format is deemed an implementation detail and not guaranteed to remain stable or compatible between Python versions.

And yet, I find it very enlightening to see how the sausage is made and to peek behind the abstractions provided by the CPython interpreter. Understanding at least some of the inner workings can help you write more performant code (when that’s important). And it’s also a lot of fun.

Let’s take this simple greet() function as a lab sample we can play with and use to understand Python’s bytecode:

def greet(name):
  return 'Hello, ' + name + '!'

>>> greet('Guido')
'Hello, Guido!'

Remember how I said that CPython first translates our source code into an intermediate language before it “runs” it? Well, if that’s true, we should be able to see the results of this compilation step. And we can.

Each function has a __code__ attribute (in Python 3) that we can use to get at the virtual machine instructions, constants, and variables used by our greet function:

>>> greet.__code__.co_code
b'dx01|x00x17x00dx02x17x00Sx00'
>>> greet.__code__.co_consts
(None, 'Hello, ', '!')
>>> greet.__code__.co_varnames
('name',)

You can see co_consts contains parts of the greeting string our function assembles. Constants and code are kept separate to save memory space. Constants are, well, constant-meaning they can never be modified and are used interchangeably in multiple places.

So instead of repeating the actual constant values in the co_code instruction stream, Python stores constants separately in a lookup table. The instruction stream can then refer to a constant with an index into the lookup table. The same is true for variables stored in the co_varnames field.
python将常量分别存储在查询表里。我们可以通过索引来通过查询表查找常量。我们可以通过co_varnames来查找存储的变量。

I hope this general concept is starting to become more clear. But looking at the co_code instruction stream still makes me feel a little queasy. This intermediate language is clearly meant to be easy to work with for the Python virtual machine, not humans. After all, that’s what the text-based source code is for.

The developers working on CPython realized that too. So they gave us another tool called a disassembler to make inspecting the bytecode easier.

Python’s bytecode disassembler lives in the dis module that’s part of the standard library. So we can just import it and call dis.dis() on our greet function to get a slightly easier-to-read representation of its bytecode:

>>> import dis
>>> dis.dis(greet)
  2   0 LOAD_CONST 1 ('Hello, ')
        2 LOAD_FAST 0 (name)
        4 BINARY_ADD
        6 LOAD_CONST 2 ('!')
        8 BINARY_ADD
        10 RETURN_VALUE

The main thing disassembling did was split up the instruction stream and give each opcode in it a human-readable name like LOAD_CONST.

You can also see how constant and variable references are now interleaved with the bytecode and printed in full to spare us the mental gymnastics of a co_const or co_varnames table lookup. Neat!

Looking at the human-readable opcodes, we can begin to understand how CPython represents and executes the 'Hello, ' + name + '!' expression in the original greet() function.

It first retrieves the constant at index 1 ('Hello, ') and puts it on the stack. It then loads the contents of the name variable and also puts them on the stack.

The stack is the data structure used as internal working storage for the virtual machine. There are different classes of virtual machines and one of them is called a stack machine. CPython’s virtual machine is an implementation of such a stack machine. If the whole thing is named after the stack, you can imagine what a central role this data structure plays.

By the way—I’m only touching the surface here. If you’re interested in this topic you’ll find a book recommendation at the end of this chapter. Reading up on virtual machine theory is enlightening (and a ton of fun).

What’s interesting about a stack as an abstract data structure is that, at the bare minimum, it only supports two operations: push and pop. Push adds a value to the top of the stack and pop removes and returns the topmost value. Unlike an array, there’s no way to access elements “below” the top level.

I find it fascinating that such a simple data structure has so many uses. But I’m getting carried away again…

Let’s assume the stack starts out empty. After the first two opcodes have been executed, this is what the contents of the VM stack look like (0 is the topmost element):

0: 'Guido' (contents of "name")
1: 'Hello, '

Then there’s another LOAD_CONST to get the exclamation mark string on the stack:

0: '!'
1: 'Hello, Guido'

The next BINARY_ADD opcode again combines the two to generate the final greeting string:

0: 'Hello, Guido!'

The last bytecode instruction is RETURN_VALUE which tells the virtual machine that what’s currently on top of the stack is the return value for this function so it can be passed on to the caller.

And voila, we just traced back how our greet() function gets executed internally by the CPython virtual machine. Isn’t that cool?

There’s much more to say about virtual machines, and this isn’t the book for it. But if this got you interested, I highly recommend that you do some more reading on this fascinating subject.

It can be a lot of fun to define your own bytecode languages and to build little virtual machine experiments for them. A book on this topic that I’d recommend is Compiler Design: Virtual Machines by Wilhelm and Seidl.

Key Takeaways

• CPython executes programs by first translating them into an intermediate bytecode and then running the bytecode on a stackbased virtual machine.
• You can use the built-in dis module to peek behind the scenes and inspect the bytecode.
• Study up on virtual machines—it’s worth it.