Out of respect to the following posts, I'll post this article in English too.
- http://nedbatchelder.com/blog/200804/the_structure_of_pyc_files.html
- http://code.activestate.com/recipes/577880-inspect-a-pyc-file/
I tried reading pyc content in __pycache__
using the code mentioned above, to understand what the pyc structure looks like in recent days.
However, it was totally unsuccessful due to some mysterious error I don't understand.
$ python --version
Python 3.5.2
$ python show_pyc.py __pycache__/hello.cpython-35.pyc
magic b'160d0d0a'
moddate b'6a393e58' (Wed Nov 30 11:28:58 2016)
source_size: 227
Traceback (most recent call last):
File "show_pyc.py", line 74, in <module>
show_file(sys.argv[1])
File "show_pyc.py", line 70, in show_file
show_code(marshal.load(f))
ValueError: bad marshal data (unknown type code)
Actually, Ian (in the second article) appropriately mentioned in the comment.
The file format has changed slightly as of Python 3.3+, so the recipe above no longer works.
In addition to the two original four-byte fields there is a new four-byte field that encodes the size of the source file as a long.
OK, pyc header after "3.3+" now contains another 4 bytes!
Because of this slight modification, all the documents before Python 3.3 may contain misleading descriptions. For another example, take PEP 3147.
Byte code files contain two 32-bit big-endian numbers followed by the marshaled [2] code object.
The 32-bit numbers represent a magic number and a timestamp.
This is not the case anymore. Anyway, the PEP was originally released for Python 3.2, and there was no guarantee pyc format would not change over time.
Here's my modified version of pyc reader.
import binascii
import dis
import marshal
import sys
import time
import types
def get_long(s):
return s[0] + (s[1] << 8) + (s[2] << 16) + (s[3] << 24)
def show_hex(label, h, indent):
h = binascii.hexlify(h).decode('ascii')
if len(h) < 60:
print('%s%s %s' % (indent, label, h))
else:
print('%s%s' % (indent, label))
for i in range(0, len(h), 60):
print('%s %s' % (indent, h[i:i+60]))
def show_code(code, indent=''):
print('%scode' % indent)
indent += ' '
print('%sargcount %d' % (indent, code.co_argcount))
print('%snlocals %d' % (indent, code.co_nlocals))
print('%sstacksize %d' % (indent, code.co_stacksize))
print('%sflags %04x' % (indent, code.co_flags))
show_hex('code', code.co_code, indent=indent)
dis.disassemble(code)
print('%sconsts' % indent)
for const in code.co_consts:
if isinstance(const, types.CodeType):
show_code(const, indent+' ')
else:
print(' %s%r' % (indent, const))
print('%snames %r' % (indent, code.co_names))
print('%svarnames %r' % (indent, code.co_varnames))
print('%sfreevars %r' % (indent, code.co_freevars))
print('%scellvars %r' % (indent, code.co_cellvars))
print('%sfilename %r' % (indent, code.co_filename))
print('%sname %r' % (indent, code.co_name))
print('%sfirstlineno %d' % (indent, code.co_firstlineno))
show_hex('lnotab', code.co_lnotab, indent=indent)
def show_file(fname: str) -> None:
with open(fname, 'rb') as f:
magic_str = f.read(4)
mtime_str = f.read(4)
mtime = get_long(mtime_str)
modtime = time.asctime(time.localtime(mtime))
print('magic %s' % binascii.hexlify(magic_str))
print('moddate %s (%s)' % (binascii.hexlify(mtime_str), modtime))
if sys.version_info < (3, 3):
print('source_size: (unknown)')
else:
source_size = get_long(f.read(4))
print('source_size: %s' % source_size)
show_code(marshal.loads(f.read()))
if __name__ == '__main__':
show_file(sys.argv[1])
Let the new reader work on the following code.
a, b = 1, 0
if a or b:
print("Hello World")
$ python --version
Python 3.5.2
$ ls -l hello.py
-rwxr-xr-x 1 dmiyakawa 48 Nov 30 12:41 hello.py
$ python -m py_compile hello.py
$ python show_pyc.py __pycache__/hello.cpython-35.pyc
magic b'160d0d0a'
moddate b'574a3e58' (Wed Nov 30 12:41:11 2016)
source_size: 48
code
argcount 0
nlocals 0
stacksize 2
flags 0040
code
6404005c02005a00005a0100650000731800650100722200650200640200
8301000164030053
1 0 LOAD_CONST 4 ((1, 0))
3 UNPACK_SEQUENCE 2
6 STORE_NAME 0 (a)
9 STORE_NAME 1 (b)
2 12 LOAD_NAME 0 (a)
15 POP_JUMP_IF_TRUE 24
18 LOAD_NAME 1 (b)
21 POP_JUMP_IF_FALSE 34
3 >> 24 LOAD_NAME 2 (print)
27 LOAD_CONST 2 ('Hello World')
30 CALL_FUNCTION 1 (1 positional, 0 keyword pair)
33 POP_TOP
>> 34 LOAD_CONST 3 (None)
37 RETURN_VALUE
consts
1
0
'Hello World'
None
(1, 0)
names ('a', 'b', 'print')
varnames ()
freevars ()
cellvars ()
filename 'hello.py'
name '<module>'
firstlineno 1
lnotab 0c010c01
Note the size of the source file (48) is appropriately embedded in pyc too. That is the new part which is introduced in Python 3.3+ (Sorry I don't know what "+" means here).
This seemed working fine with Python 3.5.2, 3.4.3, 3.3.6, 3.2.6, and 3.6.0b3 on my environment with MacOS Sierra + pyenv. In 3.2.6, obviously, it does not answer source size because it is not embedded in pyc either.
For readers from future: do not rely on the assumption "pyc format won't change", as I did.
Note (2018-02-01)
Python 3.7 or later may have different pyc format, which will be more "deterministic". See the following PEP