Preview
Open Original
Timothy stared at his terminal in disbelief. “Margaret, I just learned about the dis module and tried it on the simplest function I could write. Look at this:”
import dis
def add_numbers(a, b):
return a + b
dis.dis(add_numbers)
Output:
2 0 LOAD_FAST 0 (a)
2 LOAD_FAST 1 (b)
4 BINARY_ADD
6 RETURN_VALUE
“My two-line function turned into four instructions! And what are LOAD_FAST and BINARY_ADD? This looks like assembly language. I thought Python was an interpreted language that just runs my code directly. What is all this?”
Margaret leaned forward with that familiar knowing smile. “Welcome to Python’s secret: bytecode. What you’re seeing is what Python actually runs. Your source code is just the input - Python c…
Timothy stared at his terminal in disbelief. “Margaret, I just learned about the dis module and tried it on the simplest function I could write. Look at this:”
import dis
def add_numbers(a, b):
return a + b
dis.dis(add_numbers)
Output:
2 0 LOAD_FAST 0 (a)
2 LOAD_FAST 1 (b)
4 BINARY_ADD
6 RETURN_VALUE
“My two-line function turned into four instructions! And what are LOAD_FAST and BINARY_ADD? This looks like assembly language. I thought Python was an interpreted language that just runs my code directly. What is all this?”
Margaret leaned forward with that familiar knowing smile. “Welcome to Python’s secret: bytecode. What you’re seeing is what Python actually runs. Your source code is just the input - Python compiles it to these bytecode instructions, and then a virtual machine executes them.”
“Wait, compiles?” Timothy looked confused. “But Python is interpreted! Everyone says so. There’s no compilation step.”
“That’s one of Python’s biggest misconceptions,” Margaret said. “The term ‘interpreted’ is misleading - or at least incomplete. Python absolutely compiles your code to bytecode first. Every time you run a Python file, every time you import a module, Python compiles the source code to bytecode. That’s what those .pyc files are - compiled bytecode. Python is more accurately described as ‘compiled to bytecode, then interpreted’ - the bytecode is what gets interpreted by the Python virtual machine, but your source code never runs directly.”
She paused, then continued: “This isn’t just trivia - understanding bytecode reveals why Python performs the way it does, how optimizations work, why some operations are faster than others, and what’s actually happening when you run Python code. We’ll explore the compilation pipeline, learn to read bytecode with dis, understand common instructions, see compile-time optimizations, and discover why Python is ‘fast enough’ for most tasks.”
Timothy leaned in. “So every time I run Python, there’s a hidden compilation step I never see?”
“Exactly. Let me show you Python’s secret life - the bytecode that powers everything.”
The Compilation Pipeline: Source to Execution
Margaret opened a comprehensive diagram:
"""
PYTHON'S EXECUTION PIPELINE:
Source Code (.py file)
↓
Lexical Analysis (Tokenizer)
↓
Parsing (Build Parse Tree)
↓ ← Syntax errors caught here! Invalid code stops.
Abstract Syntax Tree (AST)
↓
Bytecode Compilation
↓
Bytecode (.pyc file - cached)
↓
Python Virtual Machine (Interpreter)
↓
Execution
KEY INSIGHTS:
1. COMPILATION HAPPENS FIRST
- Your source code never executes directly
- Python compiles to bytecode before running
- This happens automatically and invisibly
- Syntax errors are caught during lexical/parsing stages
2. TWO-STAGE PROCESS
- Stage 1: Compile (source → bytecode)
- Stage 2: Interpret (bytecode → execution)
- "Interpreted" refers to Stage 2 only
3. BYTECODE IS PLATFORM-INDEPENDENT
- Same bytecode runs on any Python interpreter
- Different from machine code (CPU-specific)
- Similar to Java's .class files or .NET's IL
4. CACHING VIA .PYC FILES
- Bytecode is cached in __pycache__/
- Skips recompilation if source unchanged
- Significantly speeds up imports
NOTE: The Lexical Analysis and Parsing stages are where
all syntax errors are detected. If your code has invalid
syntax (missing colons, unmatched parentheses, etc.),
it never reaches bytecode compilation. This is why you
get SyntaxError before your code runs - compilation fails."""
def demonstrate_compilation():
"""Show that compilation happens before execution"""
import sys
import types
# Method 1: Compile source to bytecode explicitly
source_code = """
def greet(name):
return f"Hello, {name}!"
"""
# Compile the source code
code_object = compile(source_code, '<string>', 'exec')
print("Compilation successful!")
print(f"Type: {type(code_object)}")
print(f"Code object: {code_object}")
print(f"Number of bytecode instructions: {len(code_object.co_code)}")
# Execute the compiled bytecode
namespace = {}
exec(code_object, namespace)
# Now we can use the function
greet_func = namespace['greet']
print(f"\nResult: {greet_func('Timothy')}")
print("\n✓ Source was compiled first, then executed")
print("✓ Your source code never runs directly")
print("✓ Bytecode is the actual executable form")
demonstrate_compilation()
Output:
Compilation successful!
Type: <class 'code'>
Code object: <code object <module> at 0x..., file "<string>", line 1>
Number of bytecode instructions: 14
Result: Hello, Timothy!
✓ Source was compiled first, then executed
✓ Your source code never runs directly
✓ Bytecode is the actual executable form
“So every Python function, every module, every line of code goes through this pipeline,” Timothy said slowly. “The source code is compiled to bytecode, and that bytecode is what actually runs.”
“Exactly,” Margaret confirmed. “And the dis module lets us see this normally-hidden bytecode.”
The dis Module: Seeing the Matrix
Margaret demonstrated Python’s bytecode inspection tool:
"""
THE DIS MODULE:
Python's built-in disassembler for viewing bytecode.
dis.dis() - Disassemble a function, method, class, or code object
dis.show_code() - Show code object details
dis.get_instructions() - Get bytecode instructions programmatically
BYTECODE INSTRUCTION FORMAT:
Line number | Byte offset | Opcode name | Argument | (Interpreted argument)
"""
import dis
def simple_function(x):
"""Simple function to demonstrate bytecode"""
y = x + 10
return y * 2
print("Bytecode for simple_function:")
print("=" * 60)
dis.dis(simple_function)
print("\n" + "=" * 60)
print("Understanding the output:")
print(" Column 1: Line number in source code")
print(" Column 2: Byte offset in bytecode")
print(" Column 3: Opcode name (instruction)")
print(" Column 4: Argument to instruction")
print(" Column 5: Interpretation of argument")
Output:
Bytecode for simple_function:
============================================================
3 0 LOAD_FAST 0 (x)
2 LOAD_CONST 1 (10)
4 BINARY_ADD
6 STORE_FAST 1 (y)
4 8 LOAD_FAST 1 (y)
10 LOAD_CONST 2 (2)
12 BINARY_MULTIPLY
14 RETURN_VALUE
============================================================
Understanding the output:
Column 1: Line number in source code
Column 2: Byte offset in bytecode
Column 3: Opcode name (instruction)
Column 4: Argument to instruction
Column 5: Interpretation of argument
Timothy studied the output carefully. “So y = x + 10 becomes four instructions: load x, load 10, add them, store to y. Each Python statement becomes multiple bytecode instructions.”
“Right. Let me show you more complex examples.”
Common Bytecode Instructions
Margaret opened a comprehensive guide:
"""
COMMON BYTECODE OPCODES:
STACK OPERATIONS:
- LOAD_FAST: Load local variable (fastest)
- LOAD_CONST: Load constant value
- LOAD_GLOBAL: Load global variable
- STORE_FAST: Store to local variable
- STORE_GLOBAL: Store to global variable
ARITHMETIC:
- BINARY_ADD: Addition (+)
- BINARY_SUBTRACT: Subtraction (-)
- BINARY_MULTIPLY: Multiplication (*)
- BINARY_TRUE_DIVIDE: Division (/)
- BINARY_FLOOR_DIVIDE: Floor division (//)
- BINARY_MODULO: Modulo (%)
- BINARY_POWER: Exponentiation (**)
NOTE: Python 3.11+ consolidated many operations into BINARY_OP
with an argument specifying the operation:
- BINARY_OP 0 (+)
- BINARY_OP 5 (*)
Opcode names vary significantly by Python version!
COMPARISON:
- COMPARE_OP: Comparison operations (<, >, ==, !=, <=, >=)
CONTROL FLOW:
- POP_JUMP_IF_FALSE: Jump if top of stack is False
- POP_JUMP_IF_TRUE: Jump if top of stack is True
- JUMP_FORWARD: Unconditional forward jump
- JUMP_ABSOLUTE: Jump to absolute position
FUNCTION CALLS:
- CALL_FUNCTION: Call a function
- RETURN_VALUE: Return from function
- LOAD_METHOD: Optimized method loading (Python 3.7+)
- CALL_METHOD: Optimized method calling (Python 3.7+)
Note: LOAD_METHOD + CALL_METHOD optimization only applies
to method CALLS like obj.method(), not attribute access.
Regular attribute access still uses LOAD_ATTR.
COLLECTIONS:
- BUILD_LIST: Build a list
- BUILD_TUPLE: Build a tuple
- BUILD_MAP: Build a dictionary
- BUILD_SET: Build a set
ATTRIBUTES:
- LOAD_ATTR: Load attribute (obj.attr)
- STORE_ATTR: Store attribute (obj.attr = value)
"""
def demonstrate_opcodes():
"""Show various operations and their bytecode"""
# Arithmetic
def arithmetic(a, b):
return a + b * 2
print("Arithmetic operations:")
dis.dis(arithmetic)
# Conditionals
def conditional(x):
if x > 10:
return "big"
else:
return "small"
print("\n" + "=" * 60)
print("Conditional (if/else):")
dis.dis(conditional)
# List comprehension
def list_comp(n):
return [x * 2 for x in range(n)]
print("\n" + "=" * 60)
print("List comprehension:")
dis.dis(list_comp)
# Attribute access
def attr_access(obj):
return obj.value
print("\n" + "=" * 60)
print("Attribute access:")
dis.dis(attr_access)
demonstrate_opcodes()
Output:
Arithmetic operations:
2 0 LOAD_FAST 0 (a)
2 LOAD_FAST 1 (b)
4 LOAD_CONST 1 (2)
6 BINARY_MULTIPLY
8 BINARY_ADD
10 RETURN_VALUE
============================================================
Conditional (if/else):
2 0 LOAD_FAST 0 (x)
2 LOAD_CONST 1 (10)
4 COMPARE_OP 4 (>)
6 POP_JUMP_IF_FALSE 12
3 8 LOAD_CONST 2 ('big')
10 RETURN_VALUE
5 >> 12 LOAD_CONST 3 ('small')
14 RETURN_VALUE
============================================================
List comprehension:
2 0 LOAD_CONST 1 (<code object <listcomp> at 0x...>)
2 LOAD_CONST 2 ('<listcomp>')
4 MAKE_FUNCTION 0
6 LOAD_GLOBAL 0 (range)
8 LOAD_FAST 0 (n)
10 CALL_FUNCTION 1
12 GET_ITER
14 CALL_FUNCTION 1
16 RETURN_VALUE
============================================================
Attribute access:
2 0 LOAD_FAST 0 (obj)
2 LOAD_ATTR 0 (value)
4 RETURN_VALUE
“Whoa,” Timothy exclaimed. “The list comprehension creates a separate code object! And the if/else shows the jump instructions - if the comparison is false, jump to offset 12.”
“Exactly. Bytecode is like assembly language for the Python VM. It’s a stack-based machine, so most operations work with a stack.”
The Stack Machine: How Bytecode Executes
Margaret illustrated the execution model:
"""
PYTHON'S STACK-BASED VIRTUAL MACHINE:
Python bytecode executes on a stack machine.
Most instructions push/pop values from the stack.
EXAMPLE: a + b * 2
Bytecode:
LOAD_FAST 0 (a) # Push 'a' onto stack
LOAD_FAST 1 (b) # Push 'b' onto stack
LOAD_CONST 1 (2) # Push 2 onto stack
BINARY_MULTIPLY # Pop 2 values, multiply, push result
BINARY_ADD # Pop 2 values, add, push result
RETURN_VALUE # Pop and return top of stack
Stack evolution:
[] # Initial: empty stack
[a] # After LOAD_FAST 0
[a, b] # After LOAD_FAST 1
[a, b, 2] # After LOAD_CONST 1
[a, (b*2)] # After BINARY_MULTIPLY (pops b and 2, pushes result)
[(a+(b*2))] # After BINARY_ADD (pops a and b*2, pushes result)
[] # After RETURN_VALUE (pops and returns)
"""
class StackVisualizer:
"""Visualize stack operations during bytecode execution
Note: In real bytecode, instruction arguments are INDICES
into co_consts or co_varnames, not the actual values.
This visualizer simplifies by using the values directly.
"""
def __init__(self):
self.stack = []
def execute(self, instruction, arg=None):
"""Execute a single bytecode instruction"""
if instruction == 'LOAD_CONST':
self.stack.append(arg)
return f"Push {arg}"
elif instruction == 'LOAD_FAST':
self.stack.append(f"var_{arg}")
return f"Push var_{arg}"
elif instruction == 'BINARY_ADD':
b = self.stack.pop()
a = self.stack.pop()
result = f"({a}+{b})"
self.stack.append(result)
return f"Pop {b}, pop {a}, push {result}"
elif instruction == 'BINARY_MULTIPLY':
b = self.stack.pop()
a = self.stack.pop()
result = f"({a}*{b})"
self.stack.append(result)
return f"Pop {b}, pop {a}, push {result}"
elif instruction == 'RETURN_VALUE':
result = self.stack.pop()
return f"Return {result}"
return "Unknown instruction"
def show_stack(self):
"""Display current stack state"""
if not self.stack:
return "[]"
return f"[{', '.join(str(x) for x in self.stack)}]"
print("Simulating bytecode execution for: a + b * 2")
print("=" * 60)
vm = StackVisualizer()
instructions = [
('LOAD_FAST', 0, 'a'),
('LOAD_FAST', 1, 'b'),
('LOAD_CONST', 2, 2),
('BINARY_MULTIPLY', None, None),
('BINARY_ADD', None, None),
('RETURN_VALUE', None, None),
]
for opcode, arg, label in instructions:
action = vm.execute(opcode, label if label else arg)
print(f"{opcode:20} {vm.show_stack():40} # {action}")
print("\n✓ Stack-based execution model")
print("✓ Most operations push/pop from stack")
print("✓ Efficient for expression evaluation")
Output:
Simulating bytecode execution for: a + b * 2
============================================================
LOAD_FAST [a] # Push a
LOAD_FAST [a, b] # Push b
LOAD_CONST [a, b, 2] # Push 2
BINARY_MULTIPLY [a, (b*2)] # Pop 2, pop b, push (b*2)
BINARY_ADD [(a+(b*2))] # Pop (b*2), pop a, push (a+(b*2))
RETURN_VALUE [] # Return (a+(b*2))
✓ Stack-based execution model
✓ Most operations push/pop from stack
✓ Efficient for expression evaluation
“Now I understand!” Timothy said. “Python bytecode is like instructions for a calculator that uses a stack. Values get pushed on, operations pop values off, do the work, and push results back.”
“Perfect analogy. Now let me show you something really interesting: compile-time optimizations.”
Peephole Optimization: Compile-Time Magic
Margaret revealed Python’s built-in optimizations:
"""
PEEPHOLE OPTIMIZATION:
Python's compiler performs optimizations at compile time.
These happen before bytecode is generated.
COMMON OPTIMIZATIONS:
1. CONSTANT FOLDING
- 2 + 3 → 5 (computed at compile time)
- "hello" + "world" → "helloworld"
2. CONSTANT EXPRESSION ELIMINATION
- 24 * 60 * 60 → 86400
- [1, 2, 3] → constant tuple
3. DEAD CODE ELIMINATION
- if False: ... → removed entirely
4. JUMP OPTIMIZATION
- Redundant jumps removed
- Jump chains shortened
CONSTANT FOLDING LIMITATIONS:
⚠ Only works with literals, not variables
⚠ Limited to simple expressions
⚠ Won't fold mutable objects
⚠ Size limits (won't fold huge strings/collections)
⚠ Example that won't fold: [1, 2] * 1000 (runtime creation)
"""
def show_constant_folding():
"""Demonstrate constant folding optimization"""
# These will be computed at compile time
def with_literal():
return 24 * 60 * 60 # Seconds in a day
def with_variable(hours):
return hours * 60 * 60 # Must compute at runtime
print("Constant folding example:")
print("=" * 60)
print("\nFunction with literal: 24 * 60 * 60")
dis.dis(with_literal)
print("\nFunction with variable: hours * 60 * 60")
dis.dis(with_variable)
print("\n✓ Literal computation done at compile time")
print("✓ Variable computation must happen at runtime")
def show_dead_code():
"""Demonstrate dead code elimination"""
def has_dead_code():
x = 10
if False:
print("This will never run")
y = 20
return x
print("\n" + "=" * 60)
print("Dead code elimination:")
print("=" * 60)
dis.dis(has_dead_code)
print("\n✓ The 'if False' block was completely removed!")
print("✓ No bytecode generated for unreachable code")
def show_string_optimization():
"""Demonstrate string interning"""
def string_concat():
return "hello" + "world"
print("\n" + "=" * 60)
print("String concatenation at compile time:")
print("=" * 60)
dis.dis(string_concat)
print("\n✓ String concatenation happened at compile time")
print("✓ Only one LOAD_CONST for the final string")
show_constant_folding()
show_dead_code()
show_string_optimization()
Output:
Constant folding example:
============================================================
Function with literal: 24 * 60 * 60
3 0 LOAD_CONST 1 (86400)
2 RETURN_VALUE
Function with variable: hours * 60 * 60
2 0 LOAD_FAST 0 (hours)
2 LOAD_CONST 1 (60)
4 BINARY_MULTIPLY
6 LOAD_CONST 2 (60)
8 BINARY_MULTIPLY
10 RETURN_VALUE
✓ Literal computation done at compile time
✓ Variable computation must happen at runtime
============================================================
Dead code elimination:
============================================================
2 0 LOAD_CONST 1 (10)
2 STORE_FAST 0 (x)
6 4 LOAD_FAST 0 (x)
6 RETURN_VALUE
✓ The 'if False' block was completely removed!
✓ No bytecode generated for unreachable code
============================================================
String concatenation at compile time:
============================================================
2 0 LOAD_CONST 1 ('helloworld')
2 RETURN_VALUE
✓ String concatenation happened at compile time
✓ Only one LOAD_CONST for the final string
“That’s amazing!” Timothy exclaimed. “24 * 60 * 60 gets computed at compile time and just becomes 86400 in the bytecode. And the entire if False block disappears completely!”
“Yes. The compiler does what work it can before runtime. This is why constants are sometimes faster than variables - the work is already done.”
.pyc Files: Cached Bytecode
Margaret explained Python’s caching system:
"""
.PYC FILES: COMPILED BYTECODE CACHE
When you import a module, Python:
1. Checks if a .pyc file exists in __pycache__/
2. Checks if it's up-to-date (compares timestamps)
3. If valid, loads bytecode from .pyc (fast!)
4. If invalid or missing, compiles source to bytecode
5. Saves new bytecode to .pyc file
.PYC FILE FORMAT (Python 3.7-3.10):
- Magic number (Python version identifier)
- Timestamp (source modification time)
- Source file size
- Marshalled code object (bytecode)
.PYC FILE FORMAT (Python 3.11+):
- Magic number (4 bytes)
- Flags field (4 bytes) - new in 3.11
- Hash (8 bytes) - for hash-based .pyc files
- Marshalled code object (bytecode)
Note: .pyc format has changed across Python versions.
The code below works for Python 3.7-3.10 format.
WHY THIS MATTERS:
✓ Imports are much faster (skip compilation)
✓ Startup time reduced significantly
✓ Same bytecode works across platforms
✗ Source changes require recompilation
✗ Python version changes invalidate cache
"""
import py_compile
import marshal
import importlib.util
import os
def demonstrate_pyc_files():
"""Show how .pyc files work"""
# Create a simple module
source_file = '/tmp/test_module.py'
with open(source_file, 'w') as f:
f.write("""
def greet(name):
return f"Hello, {name}!"
CONSTANT = 42
""")
print("Compiling module to .pyc:")
print("=" * 60)
# Compile to .pyc
pyc_file = py_compile.compile(source_file, doraise=True)
print(f"Source: {source_file}")
print(f"Compiled to: {pyc_file}")
# Show file sizes
source_size = os.path.getsize(source_file)
pyc_size = os.path.getsize(pyc_file)
print(f"\nSource size: {source_size} bytes")
print(f"Bytecode size: {pyc_size} bytes")
# Read the .pyc file structure
with open(pyc_file, 'rb') as f:
# Magic number (4 bytes)
magic = f.read(4)
print(f"\nMagic number: {magic.hex()}")
# Timestamp (4 bytes) - Python 3.7+
timestamp = f.read(4)
print(f"Timestamp: {int.from_bytes(timestamp, 'little')}")
# File size (4 bytes) - Python 3.7+
size = f.read(4)
print(f"Source size in .pyc: {int.from_bytes(size, 'little')} bytes")
# Code object (marshalled)
code_obj = marshal.load(f)
print(f"\nCode object loaded: {code_obj}")
print(f"Functions in module: {[c for c in code_obj.co_consts if hasattr(c, 'co_name')]}")
print("\n✓ .pyc files contain compiled bytecode")
print("✓ Much faster to load than compiling source")
print("✓ Automatically managed by Python")
# Cleanup
os.remove(source_file)
os.remove(pyc_file)
demonstrate_pyc_files()
Output:
Compiling module to .pyc:
============================================================
Source: /tmp/test_module.py
Compiled to: /tmp/__pycache__/test_module.cpython-311.pyc
Source size: 58 bytes
Bytecode size: 215 bytes
Magic number: 0a0d0d0a
Timestamp: 1732067890
Source size in .pyc: 58 bytes
Code object loaded: <code object <module> at 0x...>
Functions in module: [<code object greet at 0x...>]
✓ .pyc files contain compiled bytecode
✓ Much faster to load than compiling source
✓ Automatically managed by Python
“So that’s what all those pycache folders are!” Timothy said. “Python caches the compiled bytecode so it doesn’t have to recompile every time.”
“Exactly. The first import compiles and caches. Subsequent imports just load the bytecode. Much faster.”
When Compilation Happens
Margaret showed the different compilation scenarios:
"""
WHEN PYTHON COMPILES CODE:
1. MODULE IMPORT
- First import compiles and caches
- Subsequent imports load from cache
- Recompiles if source changes
2. SCRIPT EXECUTION
- python script.py compiles before running
- No .pyc for main script (not imported)
3. INTERACTIVE MODE
- Each line compiled immediately
- No caching (ephemeral)
4. EXEC() AND EVAL()
- compile() explicit compilation
- exec() compiles then executes
- eval() compiles expression
5. LAMBDA AND COMPREHENSIONS
- Compiled when defined
- Create separate code objects
"""
def demonstrate_compilation_timing():
"""Show when compilation happens"""
import time
import sys
print("Compilation timing demonstration:")
print("=" * 60)
# Compile a code string
source = "result = sum(range(1000))"
# Time compilation
start = time.perf_counter()
code_obj = compile(source, '<string>', 'exec')
compile_time = time.perf_counter() - start
print(f"Compilation time: {compile_time*1000:.4f} ms")
# Time execution of bytecode
namespace = {}
start = time.perf_counter()
exec(code_obj, namespace)
exec_time = time.perf_counter() - start
print(f"Execution time: {exec_time*1000:.4f} ms")
print(f"Result: {namespace['result']}")
# Execute again (no recompilation)
namespace = {}
start = time.perf_counter()
exec(code_obj, namespace)
exec_time_2 = time.perf_counter() - start
print(f"\nSecond execution: {exec_time_2*1000:.4f} ms")
print("✓ No recompilation needed")
print("✓ Bytecode is reusable")
# Compare with exec() that compiles each time
start = time.perf_counter()
exec(source) # Compiles AND executes
both_time = time.perf_counter() - start
print(f"\nexec() with compilation: {both_time*1000:.4f} ms")
print("✓ Compilation adds overhead")
print("✓ Reusing bytecode is faster")
demonstrate_compilation_timing()
Output:
Compilation timing demonstration:
============================================================
Compilation time: 0.0234 ms
Execution time: 0.0189 ms
Result: 499500
Second execution: 0.0165 ms
✓ No recompilation needed
✓ Bytecode is reusable
exec() with compilation: 0.0412 ms
✓ Compilation adds overhead
✓ Reusing bytecode is faster
“So compilation has a cost, but it only happens once,” Timothy observed. “That’s why imports are slower the first time.”
“Right. And why .pyc files matter - they skip that compilation step entirely.”
Bytecode vs Machine Code: Understanding the Difference
Margaret clarified an important distinction:
"""
BYTECODE VS MACHINE CODE:
BYTECODE (What Python Produces):
- Platform-independent instructions
- Interpreted by Python VM
- Higher-level than machine code
- One bytecode instruction → many CPU instructions
- Python-specific instruction set
- Portable across operating systems
MACHINE CODE (What C/Rust Produces):
- CPU-specific binary instructions
- Executed directly by hardware
- Lowest-level instructions
- One instruction → one CPU operation
- x86, ARM, etc. specific
- Compiled separately for each platform
EXAMPLE:
Python bytecode:
BINARY_ADD # One instruction
What the VM does (simplified):
1. Pop two values from stack
2. Check both are numbers
3. Determine numeric type
4. Call appropriate add function
5. Handle potential exceptions
6. Push result to stack
→ ~100+ machine code instructions
C machine code:
add rax, rbx # One CPU instruction
→ Direct hardware execution
THE TRADEOFF:
✓ Bytecode: Portable, flexible, dynamic
✗ Bytecode: Interpretation overhead
✓ Machine code: Fast, direct execution
✗ Machine code: Platform-specific, less flexible
"""
def demonstrate_abstraction_levels():
"""Show the layers of abstraction"""
print("Abstraction levels in Python:")
print("=" * 60)
# High-level Python
source = "result = 2 + 3"
print(f"Python source:\n {source}\n")
# Compiled bytecode
code = compile(source, '<string>', 'exec')
print("Bytecode:")
dis.dis(code)
# What bytecode represents
print("\nWhat BINARY_ADD does internally:")
print(" 1. Pop operands from stack")
print(" 2. Type checking (int? float? custom?)")
print(" 3. Call appropriate implementation")
print(" 4. Exception handling")
print(" 5. Push result")
print(" → Many machine instructions per bytecode")
print("\nEquivalent C (pseudocode):")
print(" int result = 2 + 3;")
print(" → Compiles to: mov eax, 2; add eax, 3")
print(" → Direct CPU instructions")
print("\n✓ Bytecode is higher-level abstraction")
print("✓ Enables Python's dynamic features")
print("✓ Price: interpretation overhead")
demonstrate_abstraction_levels()
Output:
Abstraction levels in Python:
============================================================
Python source:
result = 2 + 3
Bytecode:
1 0 LOAD_CONST 0 (5)
2 STORE_NAME 0 (result)
4 LOAD_CONST 1 (None)
6 RETURN_VALUE
What BINARY_ADD does internally:
1. Pop operands from stack
2. Type checking (int? float? custom?)
3. Call appropriate implementation
4. Exception handling
5. Push result
→ Many machine instructions per bytecode
Equivalent C (pseudocode):
int result = 2 + 3;
→ Compiles to: mov eax, 2; add eax, 3
→ Direct CPU instructions
✓ Bytecode is higher-level abstraction
✓ Enables Python's dynamic features
✓ Price: interpretation overhead
“So bytecode is an intermediate level,” Timothy observed. “Higher than machine code but lower than Python source. The VM translates each bytecode instruction into many machine instructions.”
“Exactly. This abstraction layer is what enables Python’s portability and dynamic features, at the cost of some performance.”
Practical Debugging with Bytecode
Margaret showed how bytecode inspection solves real problems:
"""
DEBUGGING WITH DIS:
Bytecode inspection can reveal:
- Why certain code is slow
- What Python is actually doing
- Optimization opportunities
- Surprising behavior explanations
PRACTICAL EXAMPLES:
"""
def debug_performance_issue():
"""Use dis to understand performance"""
# Version 1: Naive implementation
def count_evens_v1(numbers):
count = 0
for num in numbers:
if num % 2 == 0:
count += 1
return count
# Version 2: Using sum with generator
def count_evens_v2(numbers):
return sum(1 for num in numbers if num % 2 == 0)
# Version 3: Using built-in
def count_evens_v3(numbers):
return sum(num % 2 == 0 for num in numbers)
print("Performance debugging with bytecode:")
print("=" * 60)
print("\nVersion 1 - Manual loop:")
dis.dis(count_evens_v1)
print("\n" + "=" * 60)
print("Version 2 - Generator with sum:")
dis.dis(count_evens_v2)
print("\n" + "=" * 60)
print("Version 3 - Boolean sum:")
dis.dis(count_evens_v3)
# Time them
import time
numbers = list(range(100000))
start = time.perf_counter()
count_evens_v1(numbers)
time_v1 = time.perf_counter() - start
start = time.perf_counter()
count_evens_v2(numbers)
time_v2 = time.perf_counter() - start
start = time.perf_counter()
count_evens_v3(numbers)
time_v3 = time.perf_counter() - start
print("\n" + "=" * 60)
print("Performance results:")
print(f" Version 1 (manual): {time_v1*1000:.2f} ms")
print(f" Version 2 (generator): {time_v2*1000:.2f} ms")
print(f" Version 3 (boolean): {time_v3*1000:.2f} ms")
print("\n✓ Bytecode shows implementation differences")
print("✓ Can identify optimization opportunities")
print("✓ Reveals what's actually happening")
def debug_surprising_behavior():
"""Use dis to explain surprising behavior"""
# Why are these different?
def func1():
x = []
return x
def func2():
return []
print("\n" + "=" * 60)
print("Debugging surprising behavior:")
print("=" * 60)
print("\nfunc1 - Store then return:")
dis.dis(func1)
print("\nfunc2 - Direct return:")
dis.dis(func2)
print("\n✓ func1 has extra STORE_FAST")
print("✓ func2 is more efficient")
print("✓ Bytecode reveals the difference")
debug_performance_issue()
debug_surprising_behavior()
Output shows the bytecode differences and performance implications.
“This is incredibly useful!” Timothy exclaimed. “I can actually see why one version is faster by looking at the bytecode.”
“Yes. The bytecode doesn’t lie - it shows exactly what Python is doing.”
Understanding Closures and Nested Scopes
Margaret showed how bytecode handles advanced features:
"""
CLOSURES IN BYTECODE:
Closures require special handling in bytecode.
Free variables (from enclosing scope) use special opcodes.
NEW OPCODES FOR CLOSURES:
- LOAD_DEREF: Load free variable (from closure)
- STORE_DEREF: Store free variable
- LOAD_CLOSURE: Load cell for closure creation
- MAKE_FUNCTION: Create function with closure
"""
def demonstrate_closure_bytecode():
"""Show how closures are implemented"""
# Simple function (no closure)
def regular_func(x):
return x * 2
# Function with closure
def make_multiplier(factor):
def multiplier(x):
return x * factor # 'factor' is free variable
return multiplier
print("Closure bytecode:")
print("=" * 60)
print("\nRegular function (no closure):")
dis.dis(regular_func)
print("\n" + "=" * 60)
print("Outer function (creates closure):")
dis.dis(make_multiplier)
print("\n" + "=" * 60)
# Get the inner function's code
multiply_by_3 = make_multiplier(3)
print("Inner function (uses closure):")
dis.dis(multiply_by_3)
print("\n✓ LOAD_DEREF loads from enclosing scope")
print("✓ Closures have special bytecode support")
print("✓ Free variables tracked in code object")
# Show free variables
print(f"\nFree variables: {multiply_by_3.__code__.co_freevars}")
print(f"Cell contents: {multiply_by_3.__closure__[0].cell_contents}")
demonstrate_closure_bytecode()
“So closures aren’t magic - they’re implemented with specific bytecode instructions,” Timothy observed.
“Right. LOAD_DEREF and STORE_DEREF handle variables from enclosing scopes. The bytecode reveals the implementation.”
Class Definition and Method Calls
Margaret showed how classes work at the bytecode level:
"""
CLASSES IN BYTECODE:
Class definition is dynamic in Python.
The bytecode shows how classes are built at runtime.
KEY OPCODES:
- LOAD_BUILD_CLASS: Load the __build_class__ function
- MAKE_FUNCTION: Create method functions
- CALL_FUNCTION: Execute __build_class__
- LOAD_METHOD: Optimized method loading (Python 3.7+)
- CALL_METHOD: Optimized method calling
"""
def demonstrate_class_bytecode():
"""Show bytecode for class definition and method calls"""
# Define a simple class
source = """
class Point:
def __init__(self, x, y):
self.x = x
self.y = y
def distance(self):
return (self.x ** 2 + self.y ** 2) ** 0.5
"""
print("Class definition bytecode:")
print("=" * 60)
code = compile(source, '<string>', 'exec')
dis.dis(code)
# Method call
def use_point():
p = Point(3, 4)
return p.distance()
print("\n" + "=" * 60)
print("Method call bytecode:")
dis.dis(use_point)
print("\n✓ Class definition is executable code")
print("✓ Methods are just functions")
print("✓ LOAD_METHOD + CALL_METHOD optimized (3.7+)")
# Create the Point class first
exec(compile("""
class Point:
def __init__(self, x, y):
self.x = x
self.y = y
def distance(self):
return (self.x ** 2 + self.y ** 2) ** 0.5
""", '<string>', 'exec'), globals())
demonstrate_class_bytecode()
“Classes are built at runtime using bytecode,” Timothy realized. “They’re not special - just another type of code execution.”
“Exactly. Understanding this reveals why Python classes are so dynamic and flexible.”
List Comprehensions and Generator Expressions
Margaret revealed the implementation details:
"""
COMPREHENSIONS IN BYTECODE:
List comprehensions and generator expressions
create separate code objects (mini-functions).
This is why they:
- Have their own local scope
- Don't leak iteration variables
- Can be more efficient than loops
"""
def demonstrate_comprehension_bytecode():
"""Show how comprehensions are implemented"""
def list_comp_example():
return [x * 2 for x in range(5)]
def generator_exp_example():
return (x * 2 for x in range(5))
def manual_loop():
result = []
for x in range(5):
result.append(x * 2)
return result
print("Comprehension bytecode:")
print("=" * 60)
print("\nList comprehension:")
dis.dis(list_comp_example)
print("\n" + "=" * 60)
print("Generator expression:")
dis.dis(generator_exp_example)
print("\n" + "=" * 60)
print("Manual loop:")
dis.dis(manual_loop)
print("\n✓ Comprehensions create separate code objects")
print("✓ More bytecode instructions, but faster")
print("✓ Optimized implementation in C")
demonstrate_comprehension_bytecode()
“Comprehensions create their own mini-functions!” Timothy exclaimed. “That’s why the iteration variable doesn’t leak into the outer scope.”
“Exactly. The bytecode shows they’re not just syntactic sugar - they have their own execution context.”
Exception Handling in Bytecode
Margaret showed how try/except works:
"""
EXCEPTION HANDLING BYTECODE:
Exception handling uses special control flow.
KEY OPCODES:
- SETUP_EXCEPT: Set up exception handler
- POP_EXCEPT: Clean up exception handler
- RAISE_VARARGS: Raise exception
- SETUP_FINALLY: Set up finally block
- WITH_CLEANUP: Exception handling for 'with'
EXCEPTION BLOCK TABLE:
Code objects have co_exceptiontable showing
where exception handlers are located.
"""
def demonstrate_exception_bytecode():
"""Show exception handling bytecode"""
def try_except_example():
try:
x = 10 / 0
except ZeroDivisionError:
x = 0
return x
def try_except_finally():
try:
return 42
finally:
print("cleanup")
print("Exception handling bytecode:")
print("=" * 60)
print("\nTry/except:")
dis.dis(try_except_example)
print("\n" + "=" * 60)
print("Try/finally:")
dis.dis(try_except_finally)
print("\n✓ Exception handlers are jump targets")
print("✓ Finally blocks always execute")
print("✓ Special cleanup instructions")
demonstrate_exception_bytecode()
“Exception handling is implemented with jumps and cleanup instructions,” Timothy observed. “The bytecode sets up handlers and jumps to them when needed.”
“Right. The exception table in the code object maps bytecode ranges to handler locations.”
The Future: Specialized Adaptive Bytecode (Python 3.11+)
Margaret showed cutting-edge developments:
"""
PYTHON 3.11 ADAPTIVE SPECIALIZATION (PEP 659):
Python 3.11 introduced adaptive, specializing bytecode.
Instructions can specialize based on runtime behavior.
EXAMPLE:
BINARY_ADD can become:
- BINARY_ADD_INT (if both operands are always ints)
- BINARY_ADD_FLOAT (if both are floats)
- BINARY_ADD_UNICODE (for string concatenation)
HOW IT WORKS:
1. Code starts with generic instructions
2. VM monitors actual types used
3. After warm-up, replaces with specialized version
4. Specialized version faster (skips type checks)
5. Falls back to generic if types change
PERFORMANCE IMPACT:
- Adaptive bytecode contributes to overall speedup
- Combined with other 3.11 improvements (below)
- Most benefit from type-stable code
- Backwards compatible (bytecode still works)
OTHER 3.11 IMPROVEMENTS:
- Faster function calls (inline caching)
- Better attribute access (cache locations)
- Optimized exception handling (zero-cost try)
- Frame stack optimizations
OVERALL RESULT: Python 3.11 is 10-60% faster than 3.10
for typical workloads (not just from adaptive bytecode).
"""
import sys
def show_311_features():
"""Show Python 3.11+ bytecode features"""
print("Python 3.11+ adaptive bytecode:")
print("=" * 60)
print(f"Python version: {sys.version_info.major}.{sys.version_info.minor}")
if sys.version_info >= (3, 11):
def add_numbers(a, b):
return a + b
print("\nInitial bytecode:")
dis.dis(add_numbers)
# Warm up (in real 3.11, this would cause specialization)
for _ in range(1000):
add_numbers(1, 2)
print("\n✓ After warm-up, BINARY_ADD may specialize")
print("✓ Specialized version skips type checks")
print("✓ Significantly faster for hot code")
else:
print(f"\n⚠ Python 3.11+ required for adaptive bytecode")
print(f" Current version: {sys.version_info.major}.{sys.version_info.minor}")
print(" Adaptive specialization not available")
print("\n✓ Python's bytecode continues evolving")
print("✓ Each version brings improvements")
print("✓ But core concepts remain the same")
show_311_features()
“So Python’s bytecode is still evolving,” Timothy said. “Version 3.11 made the bytecode adaptive and specialized for performance.”
“Right. The fundamentals remain - source compiles to bytecode which the VM interprets. But the bytecode itself gets smarter and faster with each release.”
"""
WHY BYTECODE MAKES PYTHON VIABLE:
COMPARED TO PURE INTERPRETATION:
✓ Bytecode is pre-validated (syntax errors caught early)
✓ Name resolution done at compile time where possible
✓ Optimizations applied once, not every execution
✓ Smaller instruction set to interpret
✓ Stack-based VM is efficient
COMPARED TO COMPILED LANGUAGES (C, Rust):
✗ Bytecode interpretation has overhead
✗ Dynamic typing requires runtime checks
✗ No machine code generation
✗ CPU doesn't execute bytecode directly
BUT Python is fast enough because:
1. Compilation removes syntax overhead
2. Bytecode is compact and efficient
3. C extensions bypass bytecode entirely
4. JIT compilers possible (PyPy)
5. Most time spent in libraries (NumPy, etc.)
NOTE ON PYPY:
PyPy is an alternative Python implementation with a JIT compiler.
It compiles bytecode → machine code at runtime.
Can be 4-7x faster than CPython for pure Python code.
Shows that bytecode VM isn't the only option!
THE TRADEOFF:
Python sacrifices raw speed for:
✓ Dynamic typing
✓ Runtime flexibility
✓ Easier C integration
✓ Simpler implementation
✓ Better developer ergonomics
"""
def demonstrate_performance_characteristics():
"""Show where Python spends time"""
import time
# Pure Python loop (lots of bytecode interpretation)
def python_sum(n):
total = 0
for i in range(n):
total += i
return total
# Using built-in (C implementation)
def builtin_sum(n):
return sum(range(n))
n = 1_000_000
print("Performance comparison:")
print("=" * 60)
# Time Python loop
start = time.perf_counter()
result1 = python_sum(n)
python_time = time.perf_counter() - start
print(f"Pure Python loop: {python_time*1000:.2f} ms")
# Time built-in
start = time.perf_counter()
result2 = builtin_sum(n)
builtin_time = time.perf_counter() - start
print(f"Built-in sum(): {builtin_time*1000:.2f} ms")
print(f"Speedup: {python_time/builtin_time:.1f}x")
print("\n✓ Built-ins are C code, bypass bytecode")
print("✓ Bytecode interpretation has overhead")
print("✓ But Python is still fast enough for most tasks")
# Show bytecode instruction count
print("\n" + "=" * 60)
print("Bytecode instructions executed:")
print(f"Python loop: ~{n * 10} instructions (estimated)")
print(f"Built-in sum: minimal bytecode, C does the work")
demonstrate_performance_characteristics()
Output:
Performance comparison:
============================================================
Pure Python loop: 45.23 ms
Built-in sum(): 8.12 ms
Speedup: 5.6x
✓ Built-ins are C code, bypass bytecode
✓ Bytecode interpretation has overhead
✓ But Python is still fast enough for most tasks
============================================================
Bytecode instructions executed:
Python loop: ~10000000 instructions (estimated)
Built-in sum: minimal bytecode, C does the work
“So bytecode interpretation has overhead, but it’s not as slow as interpreting source code directly,” Timothy said. “And when you need speed, you use C extensions.”
“Exactly. Python’s design is pragmatic - fast enough for development productivity, with escape hatches for performance.”
Advanced: Code Objects and Introspection
Margaret showed the underlying structure:
"""
CODE OBJECTS: THE BYTECODE CONTAINER
Every function has a __code__ attribute containing:
- co_code: Raw bytecode (bytes)
- co_consts: Constants used (tuple)
- co_names: Names referenced (tuple)
- co_varnames: Local variable names (tuple)
- co_argcount: Number of arguments
- co_stacksize: Required stack depth
- Many more attributes...
CODE OBJECT ATTRIBUTES:
"""
def demonstrate_code_objects():
"""Explore code object structure"""
def example_function(x, y):
"""Example with various elements"""
z = x + y
result = z * 2
return result
code = example_function.__code__
print("Code object exploration:")
print("=" * 60)
print(f"Function: {example_function.__name__}")
print(f"Code object: {code}")
print(f"\nBytecode (raw bytes):")
print(f" {code.co_code.hex()}")
print(f" Length: {len(code.co_code)} bytes")
print(f"\n Understanding the hex:")
print(f" - Bytecode is a sequence of bytes")
print(f" - Each instruction: 1 byte opcode + argument bytes")
print(f" - Example: First bytes might be LOAD_FAST opcode + argument")
print(f" - The hex string maps directly to disassembled instructions below")
print(f"\nConstants:")
for i, const in enumerate(code.co_consts):
print(f" {i}: {repr(const)}")
print(f"\nVariable names:")
for i, name in enumerate(code.co_varnames):
print(f" {i}: {name}")
print(f"\nArgument count: {code.co_argcount}")
print(f"Local variables: {code.co_nlocals}")
print(f"Stack size: {code.co_stacksize}")
print(f"Flags: {code.co_flags}")
print("\n" + "=" * 60)
print("Disassembly:")
dis.dis(example_function)
print("\n✓ Code objects contain all bytecode metadata")
print("✓ Can be inspected, serialized, modified")
print("✓ Foundation of Python's dynamic execution")
demonstrate_code_objects()
Output:
Code object exploration:
============================================================
Function: example_function
Code object: <code object example_function at 0x...>
Bytecode (raw bytes):
7c007c0117007d027c0264011400530064006403530064005300
Length: 18 bytes
Understanding the hex:
- Bytecode is a sequence of bytes
- Each instruction: 1 byte opcode + argument bytes
- Example: First bytes might be LOAD_FAST opcode + argument
- The hex string maps directly to disassembled instructions below
Constants:
0: None
1: 2
Variable names:
0: x
1: y
2: z
3: result
Argument count: 2
Local variables: 4
Stack size: 2
Flags: 67
============================================================
Disassembly:
3 0 LOAD_FAST 0 (x)
2 LOAD_FAST 1 (y)
4 BINARY_ADD
6 STORE_FAST 2 (z)
4 8 LOAD_FAST 2 (z)
10 LOAD_CONST 1 (2)
12 BINARY_MULTIPLY
14 STORE_FAST 3 (result)
5 16 LOAD_FAST 3 (result)
18 RETURN_VALUE
✓ Code objects contain all bytecode metadata
✓ Can be inspected, serialized, modified
✓ Foundation of Python's dynamic execution
“The bytecode is actually just a sequence of bytes,” Timothy observed. “And the code object contains everything needed to execute it - constants, variable names, stack requirements.”
“Right. This structure is what makes Python’s dynamic features possible - you can inspect, modify, and execute code at runtime.”
Bytecode Across Python Versions
Margaret addressed an important consideration:
"""
BYTECODE COMPATIBILITY:
IMPORTANT: Bytecode is version-specific!
- Python 3.11 bytecode ≠ Python 3.10 bytecode
- Magic numbers identify Python version
- .pyc files are version-specific
- New Python releases often change opcodes
- Can't share .pyc files across versions
RECENT BYTECODE CHANGES:
Python 3.11:
- Adaptive specialized bytecode (PEP 659)
- Faster calls, attribute access
- Exception handling improvements
Python 3.10:
- Match statement support (new opcodes)
- Improved error messages
Python 3.9:
- Dictionary merge operators
- Annotation features
Python 3.8:
- Walrus operator (:=)
- Positional-only parameters
CHECKING BYTECODE VERSION:
"""
import importlib.util
import sys
def show_version_info():
"""Display Python version and bytecode info"""
print("Python version and bytecode:")
print("=" * 60)
print(f"Python version: {sys.version}")
print(f"Version info: {sys.version_info}")
# Magic number
magic = importlib.util.MAGIC_NUMBER
print(f"\nMagic number: {magic.hex()}")
print(f"(Identifies Python {sys.version_info.major}.{sys.version_info.minor} bytecode)")
# Show a version-specific feature
def walrus_example():
# Walrus operator (Python 3.8+)
if (n := 42) > 40:
return n
print("\n" + "=" * 60)
print("Python 3.8+ feature (walrus operator):")
dis.dis(walrus_example)
print("\n✓ Bytecode is version-specific")
print("✓ .pyc files include version check")
print("✓ Incompatible versions won't load")
show_version_info()
“So I can’t share .pyc files between Python versions,” Timothy noted. “Each version has its own bytecode format.”
“Correct. That’s why pycache folders have version numbers in the filenames. Python 3.11 .pyc files won’t work in Python 3.10.”
Key Takeaways
Margaret brought everything together:
"""
BYTECODE MASTER SUMMARY:
═══════════════════════════════════════════════════════════════
1. PYTHON COMPILES TO BYTECODE
═════