- Abstract
- Features
- More Exact location information
- Debugging Python bytecode (no source available)
- Source-code Syntax Colorization
- Command Completion
- Terminal Handling
- Signal Handling
- Better Support for Thread Debugging
- Smart Eval
- Function Breakpoints
- More Stepping Control
- Event Tracing of Calls and Returns
- Debugger Macros via Python Lambda expressions
- Byte-code Instruction Introspection
- Some Debugger Command Arguments can be Variables and Expressions
- Out-of-Process Debugging
- Egg, Wheel, and Tarballs
- Modularity
- Documentation
- See Also
This is a gdb-like debugger for Python. It is a rewrite of pdb from the ground up. I was disappointed with the flakiness, imprecision, and poor quality of coding, modularity, and level of documentation when I first looked at pdb. (pdb has gotten better since then. But a full and complete debugger, is way more complex than what you'd expect from a Standard Python module; it requires a larger set of supporting packages too than is found in the Standard Python library).
trepan3k is both a high-level debugger as well as a lower-level
bytecode debugger inspector. The code understands a lot about byte
code and the Python code object. The debugger makes use of this
knowledge to get more precise and accurate results and provide more
reliable operations.
A command-line interface (CLI) is provided as well as remote access interface over TCP/IP.
See the entry-exit for the various ways you can enter the debugger.
This code supports versions of Python back to version 3.0 using different git branches. See trepan2 for the same code modified to work with Python 2.
Since this debugger is similar to other trepanning debuggers and gdb in general, knowledge gained by learning this is transferable to those debuggers and vice versa.
There's a lot of cool stuff here that's not in the stock Python debugger pdb, or any other Python debugger that I know about.
Python reports line information on the granularity of a line. For Python versions up to 3.8, To get more precise information, we can (de)parse into Python the byte code around a bytecode offset such as the place you are stopped at.
So far as I know, there is no other debugger that decompiles code at runtime to narrow position down to the specific bytecode instruction.
See the deparse command for details on getting this kind of information.
The problem with deparsing after 3.8 is that there is no decompiler that can deparse code and give associations to bytecode instructions. I am slowly working on that though.
We use information in Python's code object line number table in byte to understand which lines are breakpointable, and in which module or function the line appears in. Use info-line to see this information. Most if not all other debuggers do go to such lengths, and as a result, it is possible to request stopping on a line number that can never occur without complaint.
In the future, we may allow specifying an offset to indicate which offset to stop at when there are several choices for a given line number.
You can pass the debugger the name of Python bytecode and many times,
the debugger will merrily proceed. This debugger tries very hard to
find the source code. Either by using the current executable search
path (e.g. PATH) or for some by looking inside the bytecode for a
filename in the main code object (co_filename) and applying that
with a search path that takes into account the directory where the
bytecode lives.
Failing to find source code this way, and in other situations where source code can't be found, the debugger will decompile the bytecode and use that for showing the source text. This allows us to debug ``eval``'d or ``exec``'d code.
But if you happen to know where the source code is located, you can associate a file source code with the current name listed in the bytecode. See the set-substitute command for details here.
Terminal source code is colorized via pygments. And with that, you can set the pygments color style, e.g. "colorful", "paraiso-dark". See set-style . Furthermore, we make use of terminal bold and emphasized text in debugger output and help text. Of course, you can also turn this off. You can use your own pygments_style, provided you have a terminal that supports 256 colors. If your terminal supports the basic ANSI color sequences only, we support that too in both dark and light themes.
Command completion is available for GNU readline and
prompt_toolkit. While prompt_toolkit is new, command completion for
GNU Readline is not just a simple static list but varies depending on
the context. For example, for frame-changing commands that take
optional numbers, the list of valid numbers is considered.
In time (and perhaps with some volunteers), prompt_toolkit
completion will be as good as GNU Readline completion.
We can adjust debugger output depending on the line width of your terminal. If it changes, or you want to adjust it, see set-width.
Following gdb, we provide its rich set of signal handling. From the gdb documentation:
GDB has the ability to detect any occurrence of a signal in your program. You can tell GDB in advance what to do for each kind of signal.
When you are stopped inside a thread, the thread name is shown to make this fact more clear and you can see and switch between frames in different threads. See frame for more information.
And following gdb, you can list the threads too. See info-threads for more information.
If you want to evaluate the current source line before it is run in
the code, use eval. To evaluate the text of a common fragment of a
line, such as the expression part of an if statement, you can do
that with eval?. See eval for more information.
Many Python debuggers only allow setting a breakpoint at a line event and functions are treated like line numbers. But functions and lines are fundamentally different. If I write:
def five(): return 5
this line contains three different kinds of things. First, there is
the code in Python that defines the function five() for the first
time. Then there is the function itself, and then there is some code
inside that function.
In this debugger, you can give the name of a function by surrounding
adding () at the end:
break five()
Also five could be a method of an object that is currently defined when the
breakpoint command is given:
self.five()
Sometimes you want small steps, and sometimes large steps.
This fundamental issue is handled in a couple of ways:
There are now step event and next event commands with
aliases to s+, s>, and so on. The plus-suffixed commands force
a different line on a subsequent stop, the dash-suffixed commands
don't. Suffixes >, <, and ! specify call, return
and exception events respectively. And without a suffix, you get
the default; this is set by the set different command.
By default, the debugger stops at every event: call, return,
line, exception, c-call, c-exception. If you just want
to stop at line events (which is largely what happens in
pdb) you can. If however you just want to stop at calls and returns,
that's possible too. Or pick some combination.
In conjunction with handling all events by default, the event status is shown when stopped. The reason for stopping is also available via info program.
I'm not sure why this was not done before. Probably because of the lack of the ability to set and move by different granularities, tracing calls and returns leads to too many uninteresting stops (such as at the same place you just were at). Also, stopping on function definitions probably also added to this tedium.
Because we're really handling return events, we can stop on the return. This is a little more precise than pdb's retval command.
There are debugger macros. In gdb, there is a macro debugger command to extend debugger commands.
However, Python has its own rich programming language so it seems silly to recreate the macro language that is in gdb. Simpler and more powerful is just to use Python here. A debugger macro here is just a lambda expression that returns a string or a list of strings. Each string returned should be a debugger command.
We also have aliases for the extremely simple situation where you want to give an alias to an existing debugger command. But beware: Some commands, like step inspect command suffixes and change their behavior accordingly.
We also provide extending the debugger either through additional Python packages.
We do more in the way of looking at the byte codes to give better information. Through this, we can provide:
- a skip command. It is like the jump command, but you don't have to deal with line numbers.
- disassembly of code fragments. You can now disassemble relative to the stack frames you are currently stopped at.
- Better interpretation of where you are when inside execfile or exec. (But really though this is probably a Python compiler misfeature.)
- Check that breakpoints are set only where they make sense.
- A more accurate determination of if you are at a function-defining
def or class statements (because the caller's instruction contains
MAKE_FUNCTIONorBUILD_CLASS.)
Even without "deparsing" mentioned above, the ability to disassemble where the PC is currently located (see info-pc), by line number range or byte-offset range lets you tell exactly where you are and code is getting run.
Commands that take integer arguments like up, list, or disassemble allow you to use a Python expression which may include local or global variables that evaluate to an integer. This eliminates the need in gdb for special "dollar" debugger variables. (Note however because of shlex parsing, expressions can't have embedded blanks.)
You can now debug your program in a different process or even a different computer on a different network!
Related, is flexible support for remapping path names from the file system, e.g. the filesystem seen inside a docker container or on a remote filesystem with locally-installed files. See subst for more information.
Can be installed via the usual pip or easy_install. There is a source tarball. How To Install has full instructions and installation using git or by other means.
Because this debugger is modular, I have been able to use it as the basis for debuggers in other projects. In particular, it is used as a module in trepanxpy, a debugger for Python interpreter, x-python, written in Python.
It is also used as a module inside an experimental open-source Wolfram Mathematica interpreter, Mathics3.
Using pytracer, the Debugger plays nice with other trace hooks. You can have several debugger objects.
Many of the things listed below do not directly impact end-users, but it does eventually by way of more robust and featureful code. And keeping developers happy is a good thing.(TM)
- Commands and subcommands are individual classes now, not methods in a class. This means they now have properties like the context in which they can be run, minimum abbreviation names, or alias names. To add a new command you basically add a file in a directory.
- I/O is its own layer. This simplifies interactive readline behavior from reading commands over a TCP socket.
- An interface is its own layer. Local debugging, remote debugging, and running debugger commands from a file (
source) are different interfaces. This means, for example, that we are able to give better error reporting if a debugger command file has an error. - There is an experimental Python-friendly interface for front-ends
- more testable. Much more unit and functional tests.
Documentation: http://python3-trepan.readthedocs.org
- trepanxpy: trepan debugger for x-python, the bytecode interpreter written in Python
- https://github.com/rocky/trepan-xpy: Python debugger using this code to support x-python
- https://pypi.python.org/pypi/uncompyle6: Python decompiler
- https://pypi.python.org/pypi/decompyle3: Python 3.7 and 3.8 decompiler
- https://pypi.python.org/pypi/xdis: cross-platform disassembler
