Advanced debugging tools#

If you reached here, you want to dive into, or use, more advanced tooling. This is usually not necessary for first-time contributors and most day-to-day development. These are used more rarely, for example close to a new NumPy release, or when a large or particular complex change was made.

Some of these tools are used in NumPy’s continuous integration tests. If you see a test failure that only happens under a debugging tool, these instructions should hopefully enable you to reproduce the test failure locally.

Since not all of these tools are used on a regular basis and only available on some systems, please expect differences, issues, or quirks; we will be happy to help if you get stuck and appreciate any improvements or suggestions to these workflows.

Finding C errors with additional tooling#

Most development will not require more than a typical debugging toolchain as shown in Debugging. But for example memory leaks can be particularly subtle or difficult to narrow down.

We do not expect any of these tools to be run by most contributors. However, you can ensure that we can track down such issues more easily:

  • Tests should cover all code paths, including error paths.

  • Try to write short and simple tests. If you have a very complicated test consider creating an additional simpler test as well. This can be helpful, because often it is only easy to find which test triggers an issue and not which line of the test.

  • Never use np.empty if data is read/used. valgrind will notice this and report an error. When you do not care about values, you can generate random values instead.

This will help us catch any oversights before your change is released and means you do not have to worry about making reference counting errors, which can be intimidating.

Python debug build#

Debug builds of Python are easily available for example via the system package manager on Linux systems, but are also available on other platforms, possibly in a less convenient format. If you cannot easily install a debug build of Python from a system package manager, you can build one yourself using pyenv. For example, to install and globally activate a debug build of Python 3.13.3, one would do:

pyenv install -g 3.13.3
pyenv global 3.13.3

Note that pyenv install builds Python from source, so you must ensure that Python’s dependencies are installed before building, see the pyenv documentation for platform-specific installation instructions. You can use pip to install Python dependencies you may need for your debugging session. If there is no debug wheel available on pypi, you will need to build the dependencies from source and ensure that your dependencies are also compiled as debug builds.

Often debug builds of Python name the Python executable pythond instead of python. To check if you have a debug build of Python installed, you can run e.g. pythond -m sysconfig to get the build configuration for the Python executable. A debug build will be built with debug compiler options in CFLAGS (e.g. -g -Og).

Running the Numpy tests or an interactive terminal is usually as easy as:

python3.8d runtests.py
# or
python3.8d runtests.py --ipython

and were already mentioned in Debugging.

A Python debug build will help:

  • Find bugs which may otherwise cause random behaviour. One example is when an object is still used after it has been deleted.

  • Python debug builds allows to check correct reference counting. This works using the additional commands:

    sys.gettotalrefcount()
sys.getallocatedblocks()

  • Python debug builds allow easier debugging with gdb and other C debuggers.

Use together with pytest#

Running the test suite only with a debug python build will not find many errors on its own. An additional advantage of a debug build of Python is that it allows detecting memory leaks.

A tool to make this easier is pytest-leaks, which can be installed using pip. Unfortunately, pytest itself may leak memory, but good results can usually (currently) be achieved by removing:

@pytest.fixture(autouse=True)
def add_np(doctest_namespace):
    doctest_namespace['np'] = numpy

@pytest.fixture(autouse=True)
def env_setup(monkeypatch):
    monkeypatch.setenv('PYTHONHASHSEED', '0')

from numpy/conftest.py (This may change with new pytest-leaks versions or pytest updates).

This allows to run the test suite, or part of it, conveniently:

python3.8d runtests.py -t numpy/_core/tests/test_multiarray.py -- -R2:3 -s

where -R2:3 is the pytest-leaks command (see its documentation), the -s causes output to print and may be necessary (in some versions captured output was detected as a leak).

Note that some tests are known (or even designed) to leak references, we try to mark them, but expect some false positives.

valgrind#

Valgrind is a powerful tool to find certain memory access problems and should be run on complicated C code. Basic use of valgrind usually requires no more than:

PYTHONMALLOC=malloc valgrind python runtests.py

where PYTHONMALLOC=malloc is necessary to avoid false positives from python itself. Depending on the system and valgrind version, you may see more false positives. valgrind supports “suppressions” to ignore some of these, and Python does have a suppression file (and even a compile time option) which may help if you find it necessary.

Valgrind helps:

  • Find use of uninitialized variables/memory.

  • Detect memory access violations (reading or writing outside of allocated memory).

  • Find many memory leaks. Note that for most leaks the python debug build approach (and pytest-leaks) is much more sensitive. The reason is that valgrind can only detect if memory is definitely lost. If:

    dtype = np.dtype(np.int64)
arr.astype(dtype=dtype)

    Has incorrect reference counting for dtype, this is a bug, but valgrind cannot see it because np.dtype(np.int64) always returns the same object. However, not all dtypes are singletons, so this might leak memory for different input. In rare cases NumPy uses malloc and not the Python memory allocators which are invisible to the Python debug build. malloc should normally be avoided, but there are some exceptions (e.g. the PyArray_Dims structure is public API and cannot use the Python allocators.)

Even though using valgrind for memory leak detection is slow and less sensitive it can be a convenient: you can run most programs with valgrind without modification.

Things to be aware of:

  • Valgrind does not support the numpy longdouble, this means that tests will fail or be flagged errors that are completely fine.

  • Expect some errors before and after running your NumPy code.

  • Caches can mean that errors (specifically memory leaks) may not be detected or are only detect at a later, unrelated time.

A big advantage of valgrind is that it has no requirements aside from valgrind itself (although you probably want to use debug builds for better tracebacks).

Use together with pytest#

You can run the test suite with valgrind which may be sufficient when you are only interested in a few tests:

PYTHONMALLOC=malloc valgrind python runtests.py \
 -t numpy/_core/tests/test_multiarray.py -- --continue-on-collection-errors

Note the --continue-on-collection-errors, which is currently necessary due to missing longdouble support causing failures (this will usually not be necessary if you do not run the full test suite).

If you wish to detect memory leaks you will also require --show-leak-kinds=definite and possibly more valgrind options. Just as for pytest-leaks certain tests are known to leak cause errors in valgrind and may or may not be marked as such.

We have developed pytest-valgrind which:

  • Reports errors for each test individually

  • Narrows down memory leaks to individual tests (by default valgrind only checks for memory leaks after a program stops, which is very cumbersome).

Please refer to its README for more information (it includes an example command for NumPy).

C debuggers#

Whenever NumPy crashes or when working on changes to NumPy’s low-level C or C++ code, it’s often convenient to run Python under a C debugger to get more information. A debugger can aid in understanding an interpreter crash (e.g. due to a segmentation fault) by providing a C call stack at the site of the crash. The call stack often provides valuable context to understand the nature of a crash. C debuggers are also very useful during development, allowing interactive debugging in the C implementation of NumPy.

The NumPy developers often use both gdb and lldb to debug Numpy. As a rule of thumb, gdb is often easier to use on Linux while lldb is easier to use on a Mac environment. They have disjoint user interfaces, so you will need to learn how to use whichever one you land on. The gdb to lldb command map is a convnient reference for how to accomplish common recipes in both debuggers.

Building With Debug Symbols#

The spin development workflow tool. has built-in support for working with both gdb and lldb via the spin gdb and spin lldb commands.

For both debuggers, it’s advisable to build NumPy in either the debug or debugoptimized meson build profile. To use debug you can pass the option via spin build:

spin build -- -Dbuildtype=debug

to use debugoptimized you’re pass -Dbuildtype=debugoptimized instead.

You can pass additional arguments to meson setup besides buildtype using the same positional argument syntax for spin build.

Running a Test Script#

Let’s say you have a test script named test.py that lives in a test folder in the same directory as the NumPy source checkout. You could execute the test script using the spin build of NumPy with the following incantation:

spin gdb ../test/test.py

This will launch into gdb. If all you care about is a call stack for a crash, type “r” and hit enter. Your test script will run and if a crash happens, you type “bt” to get a traceback. For lldb, the instructions are similar, just replace spin gdb with spin lldb.

You can also set breakpoints and use other more advanced techniques. See the documentation for your debugger for more details.

One common issue with breakpoints in NumPy is that some code paths get hit repeatedly during the import of the numpy module. This can make it tricky or tedious to find the first “real” call after the NumPy import has completed and the numpy module is fully initialized.

One workaround is to use a script like this:

import os
import signal

import numpy as np

PID = os.getpid()

def do_nothing(*args):
    pass

signal.signal(signal.SIGUSR1, do_nothing)

os.kill(PID, signal.SIGUSR1)

# the code to run under a debugger follows

This example installs a signal handler for the SIGUSR1 signal that does nothing and then calls os.kill on the Python process with the SIGUSR1 signal. This causes the signal handler to fire and critically also causes both gdb and lldb to halt execution inside of the kill syscall.

If you run lldb you should see output something like this:

Process 67365 stopped
 * thread #1, queue = 'com.apple.main-thread', stop reason = signal SIGUSR1
     frame #0: 0x000000019c4b9da4 libsystem_kernel.dylib`__kill + 8
 libsystem_kernel.dylib`__kill:
 ->  0x19c4b9da4 <+8>:  b.lo   0x19c4b9dc4    ; <+40>
     0x19c4b9da8 <+12>: pacibsp
     0x19c4b9dac <+16>: stp    x29, x30, [sp, #-0x10]!
     0x19c4b9db0 <+20>: mov    x29, sp
 Target 0: (python3.13) stopped.
 (lldb) bt
 * thread #1, queue = 'com.apple.main-thread', stop reason = signal SIGUSR1
   * frame #0: 0x000000019c4b9da4 libsystem_kernel.dylib`__kill + 8
     frame #1: 0x000000010087f5c4 libpython3.13.dylib`os_kill + 104
     frame #2: 0x000000010071374c libpython3.13.dylib`cfunction_vectorcall_FASTCALL + 276
     frame #3: 0x00000001006c1e3c libpython3.13.dylib`PyObject_Vectorcall + 88
     frame #4: 0x00000001007edd1c libpython3.13.dylib`_PyEval_EvalFrameDefault + 23608
     frame #5: 0x00000001007e7e6c libpython3.13.dylib`PyEval_EvalCode + 252
     frame #6: 0x0000000100852944 libpython3.13.dylib`run_eval_code_obj + 180
     frame #7: 0x0000000100852610 libpython3.13.dylib`run_mod + 220
     frame #8: 0x000000010084fa4c libpython3.13.dylib`_PyRun_SimpleFileObject + 868
     frame #9: 0x000000010084f400 libpython3.13.dylib`_PyRun_AnyFileObject + 160
     frame #10: 0x0000000100874ab8 libpython3.13.dylib`pymain_run_file + 336
     frame #11: 0x0000000100874324 libpython3.13.dylib`Py_RunMain + 1516
     frame #12: 0x000000010087459c libpython3.13.dylib`pymain_main + 324
     frame #13: 0x000000010087463c libpython3.13.dylib`Py_BytesMain + 40
     frame #14: 0x000000019c152b98 dyld`start + 6076
(lldb)

As you can see, the C stack trace is inside of the kill syscall and an lldb prompt is active, allowing interactively setting breakpoints. Since the os.kill call happens after the numpy module is already fully initialized, this means any breakpoints set inside of kill will happen after numpy is finished initializing.

Use together with pytest#

You can also run pytest tests under a debugger. This requires using the debugger in a slightly more manual fashion, since spin does not yet automate this process. First, run spin build to ensure there is a fully built copy of NumPy managed by spin. Then, to run the tests under lldb you would do something like this:

spin lldb $(which python) $(which pytest) build-install/usr/lib/python3.13/site-packages/numpy/_core/tests/test_multiarray.py

This will execute the tests in test_multiarray.py under lldb after typing ‘r’ and hitting enter. Note that this command comes from a session using Python 3.13 on a Mac. If you are using a different Python version or operating system, the directory layout inside build-install may be slightly different.

You can set breakpoints as described above. The issue about breakpoints commonly being hit during NumPy import also applies - consider refactoring your test workflow into a test script so you can adopt the workaround using os.kill described above.

Note the use of $(which python) to ensure the debugger receives a path to a Python executable. If you are using pyenv, you may need to replace which python with pyenv which python, since pyenv relies on shim scripts that which doesn’t know about.

Compiler Sanitizers#

The compiler sanitizer suites shipped by both GCC and LLVM offer a means to detect many common programming errors at runtime. The sanitizers work by instrumenting the application code at build time so additional runtime checks fire. Typically, sanitizers are run during the course of regular testing and if a sanitizer check fails, this leads to a test failure or crash, along with a report about the nature of the failure.

While it is possible to use sanitizers with a “regular” build of CPython - it is best if you can set up a Python environment based on a from-source Python build with sanitizer instrumentation, and then use the instrumented Python to build NumPy and run the tests. If the entire Python stack is instrumented using the same sanitizer runtime, it becomes possible to identify issues that happen across the Python stack. This enables detecting memory leaks in NumPy due to misuse of memory allocated in CPython, for example.

Build Python with Sanitizer Instrumentation#

See the section in the Python developer’s guide on this topic for more information about building Python from source. To enable address sanitizer, you will need to pass --with-address-sanitizer to the configure script invocation when you build Python.

You can also use pyenv to automate the process of building Python and quickly activate or deactivate a Python installation using a command-line interface similar to virtual environments. With pyenv you could install an ASAN-instrumented build of Python 3.13 like this:

CONFIGURE_OPTS="--with-address-sanitizer" pyenv install 3.13

If you are interested in thread sanitizer, the cpython_sanity docker images might also be a quicker choice that bypasses building Python from source, although it may be annoying to do debugging work inside of a docker image.

Use together with spin#

However you build Python, once you have an instrumented Python build, you can install NumPy’s development and test dependencies and build NumPy with address sanitizer instrumentation. For example, to build NumPy with the debug profile and address sanitizer, you would pass additional build options to meson like this:

spin build -- -Dbuildtype=debug -Db_sanitize=address

Once the build is finished, you can use other spin command like spin test and spin gdb as with any other Python build.

Special considerations#

Some NumPy tests intentionally lead to malloc returning NULL. In its default configuration, some of the compiler sanitizers flag this as an error. You can disable that check by passing allocator_may_return_null=1 to the sanitizer as an option. For example, with address sanitizer:

ASAN_OPTIONS=allocator_may_return_null=1 spin test

You may see memory leaks coming from the Python interpreter, particularly on MacOS. If the memory leak reports are not useful, you can disable leak detection by passing detect_leaks=0 in ASAN_OPTIONS. You can pass more than one option using a colon-delimited list, like this:

ASAN_OPTIONS=allocator_may_return_null=1:halt_on_error=1:detect_leaks=1 spin test

The halt_on_error option can be particularly useful – it hard-crashes the Python executable whenever it detects an error, along with a report about the error that includes a stack trace.

You can also take a look at the compiler_sanitizers.yml GitHub actions workflow configuration. It describes several different CI jobs that are run as part of the NumPy tests using Thread, Address, and Undefined Behavior sanitizer.