This is a CPython bytecode interpreter written Python.
You can use this to:
- Learn about how the internals of CPython works since this models that
- Experiment with additional opcodes, or ways to change the run-time environment
- Use as a sandboxed environment for trying pieces of execution
- Have one Python program that runs multiple versions of Python bytecode.
- Use in a dynamic fuzzer or in coholic execution for analysis
The ability to run simple Python bytecode as far back as 2.4 from Python 3.10 or Python 3.10 from a Python 2.7 interpreter (assuming no fancy async features of newer runtimes) I find pretty neat.
Also, idea of the sandboxed environment in a debugger I find interesting. (Note: currently environments are not sandboxed that well, but I am working towards that, and it is a bit challenging.)
Since there is a separate execution, and traceback stack, inside a debugger you can try things out in the middle of a debug session without effecting the real execution. On the other hand if a sequence of executions works out, it is possible to copy this (under certain circumstances) back into CPython's execution stack.
I have hooked in trepan3k into this interpreter so you have a pdb/gdb like debugger also with the ability to step bytecode instructions.
To experiment with faster ways to support trace callbacks such as those used in a debugger, added is an instruction to support fast breakpoints and breakpointing on a particular instruction that doesn't happen to be on a line boundary. I believe this could could and should be ported back to CPython and there would be benefit. (Python 3.8 supports the ability to save additional frame information which is where the original opcode is stored. It just needs the BRKPT opcode)
Although this is far in the future, suppose you to add a race detector? It might be easier to prototype it in Python here. (This assumes the interpreter supports threading well, I suspect it doesn't)
Another unexplored avenue implied above is mixing interpretation and direct CPython execution. In fact, there are bugs so this happens right now, but it will be turned into a feature. Some functions or classes you may want to not run under a slow interpreter while others you do want to run under the interpreter.
What to know instructions get run when you write some simple code? Try this:
$ xpython -vc "x, y = 2, 3; x **= y" INFO:xpython.vm:L. 1 @ 0: LOAD_CONST (2, 3) INFO:xpython.vm: @ 2: UNPACK_SEQUENCE 2 INFO:xpython.vm: @ 4: STORE_NAME (2) x INFO:xpython.vm: @ 6: STORE_NAME (3) y INFO:xpython.vm:L. 1 @ 8: LOAD_NAME x INFO:xpython.vm: @ 10: LOAD_NAME y INFO:xpython.vm: @ 12: INPLACE_POWER (2, 3) INFO:xpython.vm: @ 14: STORE_NAME (8) x INFO:xpython.vm: @ 16: LOAD_CONST None INFO:xpython.vm: @ 18: RETURN_VALUE (None)
Option -c is the same as Python's flag (program passed in as string)
and -v is also analogus Python's flag. Here, it shows the bytecode
instructions run.
Note that the disassembly above in the dynamic trace above gives a
little more than what you'd see from a static disassembler from
Python's dis module. In particular, the STORE_NAME
instructions show the value that is store, e.g. "2" at instruction
offset 4 into name x. Similarly INPLACE_POWER shows the operands, 2 and 3, which is how the value
8 is derived as the operand of the next instruction, STORE_NAME.
Want more like the execution stack stack and block stack in addition? Add another v:
$ xpython -vvc "x, y = 2, 3; x **= y"
DEBUG:xpython.vm:make_frame: code=<code object <module> at 0x7f8018507db0, file "<string x, y = 2, 3; x **= y>", line 1>, callargs={}, f_globals=(<class 'dict'>, 140188140947488), f_locals=(<class 'NoneType'>, 93856967704000)
DEBUG:xpython.vm:<Frame at 0x7f80184c1e50: '<string x, y = 2, 3; x **= y>':1 @-1>
DEBUG:xpython.vm: frame.stack: []
DEBUG:xpython.vm: blocks : []
INFO:xpython.vm:L. 1 @ 0: LOAD_CONST (2, 3) <module> in <string x, y = 2, 3; x **= y>:1
DEBUG:xpython.vm: frame.stack: [(2, 3)]
DEBUG:xpython.vm: blocks : []
INFO:xpython.vm: @ 2: UNPACK_SEQUENCE 2 <module> in <string x, y = 2, 3; x **= y>:1
DEBUG:xpython.vm: frame.stack: [3, 2]
DEBUG:xpython.vm: blocks : []
INFO:xpython.vm: @ 4: STORE_NAME (2) x <module> in <string x, y = 2, 3; x **= y>:1
DEBUG:xpython.vm: frame.stack: [3]
DEBUG:xpython.vm: blocks : []
INFO:xpython.vm: @ 6: STORE_NAME (3) y <module> in <string x, y = 2, 3; x **= y>:1
DEBUG:xpython.vm: frame.stack: []
DEBUG:xpython.vm: blocks : []
INFO:xpython.vm:L. 1 @ 8: LOAD_NAME x <module> in <string x, y = 2, 3; x **= y>:1
DEBUG:xpython.vm: frame.stack: [2]
DEBUG:xpython.vm: blocks : []
INFO:xpython.vm: @ 10: LOAD_NAME y <module> in <string x, y = 2, 3; x **= y>:1
DEBUG:xpython.vm: frame.stack: [2, 3]
DEBUG:xpython.vm: blocks : []
INFO:xpython.vm: @ 12: INPLACE_POWER (2, 3) <module> in <string x, y = 2, 3; x **= y>:1
DEBUG:xpython.vm: frame.stack: [8]
DEBUG:xpython.vm: blocks : []
INFO:xpython.vm: @ 14: STORE_NAME (8) x <module> in <string x, y = 2, 3; x **= y>:1
DEBUG:xpython.vm: frame.stack: []
DEBUG:xpython.vm: blocks : []
INFO:xpython.vm: @ 16: LOAD_CONST None <module> in <string x, y = 2, 3; x **= y>:1
DEBUG:xpython.vm: frame.stack: [None]
DEBUG:xpython.vm: blocks : []
INFO:xpython.vm: @ 18: RETURN_VALUE (None) <module> in <string x, y = 2, 3; x **= y>:1
Want to see this colorized in a terminal? Use this via trepan-xpy -x:
Suppose you have Python 2.4 bytecode (or some other bytecode) for this, but you are running Python 3.7?
$ xpython -v test/examples/assign-2.4.pyc INFO:xpython.vm:L. 1 @ 0: LOAD_CONST (2, 3) INFO:xpython.vm: @ 3: UNPACK_SEQUENCE 2 INFO:xpython.vm: @ 6: STORE_NAME (2) x INFO:xpython.vm: @ 9: STORE_NAME (3) y INFO:xpython.vm:L. 2 @ 12: LOAD_NAME x INFO:xpython.vm: @ 15: LOAD_NAME y INFO:xpython.vm: @ 18: INPLACE_POWER (2, 3) INFO:xpython.vm: @ 19: STORE_NAME (8) x INFO:xpython.vm: @ 22: LOAD_CONST None INFO:xpython.vm: @ 25: RETURN_VALUE (None)
Not much has changed here, other then the fact that that in after 3.6 instructions are two bytes instead of 1- or 3-byte instructions.
The above examples show straight-line code, so you see all of the instructions. But don't confuse this with a disassembler like pydisasm from xdis. The below example, with conditional branching example makes this more clear:
$ xpython -vc "x = 6 if __name__ != '__main__' else 10"
INFO:xpython.vm:L. 1 @ 0: LOAD_NAME __name__
INFO:xpython.vm: @ 2: LOAD_CONST __main__
INFO:xpython.vm: @ 4: COMPARE_OP ('__main__', '__main__') !=
INFO:xpython.vm: @ 6: POP_JUMP_IF_FALSE 12
^^ Note jump below
INFO:xpython.vm: @ 12: LOAD_CONST 10
INFO:xpython.vm: @ 14: STORE_NAME (10) x
INFO:xpython.vm: @ 16: LOAD_CONST None
INFO:xpython.vm: @ 18: RETURN_VALUE (None)
Want even more status and control? See trepan-xpy.
Currently bytecode from Python versions 3.10 to 3.2, and 2.7 to 2.4 are supported. The most recent versions of Python don't have all opcodes implemented. This is only one of many interests I have, so support may be shoddy. I use funding to help me direct where my attention goes in fixing problems, which are vast in this project.
Byterun, from which this was based on, is awesome. But it cheats in subtle ways.
Want to write a very small interpreter using CPython?
# get code somehow exec(code)
This cheats in kind of a gross way, but this the kind of cheating goes
on in Byterun in a more subtle way. As in the example above which
relies on built-in function exec to do all of the work, Byterun
relies on various similar sorts of built-in functions to support
opcode interpretation. In fact, if the code you were interpreting
was the above, Byterun would use its built-in function for running
code inside the exec function call, so all of the bytecode that gets
run inside code inside code would not seen for interpretation.
Also, built-in functions like exec, and other built-in modules have an effect in the interpreter namespace. So the two namespaces then get intermingled.
One example of this that has been noted is for import. See
nedbat#26. But there are others
cases as well. While we haven't addressed the import issue
mentioned in issue 26, we have addressed similar kinds of issues like
this.
Some built-in functions and the inpsect module require built-in
types like cell, traceback, or frame objects, and they can't use the
corresponding interpreter classes. Here is an example of this in
Byterun: class __init__ functions don't get traced into, because
the built-in function __build_class__ is relied on. And
__build_class__ needs a native function, not an
interpreter-traceable function. See
nedbat#20.
Also Byterun is loose in accepting bytecode opcodes that is invalid for particular Python but may be valid for another. I suppose this is okay since you don't expect invalid opcodes appearing in valid bytecode. It can however accidentally or erronously appear code that has been obtained via some sort of extraction process, when the extraction process isn't accruate.
In contrast to Byterun, x-python is more stringent what opcodes it accepts.
Byterun needs the kind of overhaul we have here to be able to scale to support bytecode for more Pythons, and to be able to run bytecode across different versions of Python. Specifically, you can't rely on Python's dis module if you expect to run a bytecode other than the bytecode that the interpreter is running, or run newer "wordcode" bytecode on a "byte"-oriented byteocde, or vice versa.
In contrast, x-python there is a clear distinction between the version being interpreted and the version of Python that is running. There is tighter control of opcodes and an opcode's implementation is kept for each Python version. So we'll warn early when something is invalid. You can run bytecode back to Python 2.4 using Python 3.10 (largely), which is amazing give that 3.10's native byte code is 2 bytes per instruction while 2.4's is 1 or 3 bytes per instruction.
The "largely" part is, as mentioned above, because the interpreter has
always made use of Python builtins and libraries, and for the most
part these haven't changed very much. Often, since many of the
underlying builtins are the same, the interpreter can (and does) make
use interpreter internals. For example, built-in functions like
range() are supported this way.
So interpreting bytecode from a newer Python release than the release the Python interpreter is using, is often doable too. Even though Python 2.7 doesn't support keyword-only arguments or format strings, it can still interpret bytecode created from using these constructs.
That's possible here because these specific features are more
syntactic sugar rather than extensions to the runtime. For example,
format strings basically map down to using the format() function
which is available on 2.7.
But new features like asynchronous I/O and concurrency primitives are not in the older versions. So those need to be simulated, and that too is a possibility if there is interest or support.
You can run many of the tests that Python uses to test itself, and I do! And most of those work. Right now this program works best on Python up to 3.4 when life in Python was much simpler. It runs over 300 in Python's test suite for itself without problems. For Python 3.6 the number drops down to about 237; Python 3.9 is worse still.
This is a fork of Byterun. which is a pure-Python implementation of a Python bytecode execution virtual machine. Ned Batchelder started it (based on work from Paul Swartz) to get a better understanding of bytecodes so he could fix branch coverage bugs in coverage.py.