This document is the recommended first read if you are interested in using Nuitka, understand its use cases, check what you can expect, license, requirements, credits, etc.

Nuitka is the Python compiler. It is written in Python. It is a seamless replacement or extension to the Python interpreter and compiles every construct that CPython 2.6, 2.7, 3.3, 3.4, 3.5, 3.6, 3.7, 3.8, 3.9, 3.10, 3.11 have, when itself run with that Python version.

It then executes uncompiled code and compiled code together in an extremely compatible manner.

You can use all Python library modules and all extension modules freely.

Nuitka translates the Python modules into a C level program that then uses libpython and static C files of its own to execute in the same way as CPython does.

All optimization is aimed at avoiding overhead, where it's unnecessary. None is aimed at removing compatibility, although slight improvements will occasionally be done, where not every bug of standard Python is emulated, e.g. more complete error messages are given, but there is a full compatibility mode to disable even that.

• C Compiler: You need a compiler with support for C11 or alternatively for C++03 [1]

  Currently this means, you need to use one of these compilers:

  • The MinGW64 C11 compiler on Windows, must be based on gcc 11.2 or higher. It will be automatically downloaded if no usable C compiler is found, which is the recommended way of installing it, as Nuitka will also upgrade it for you.
  • Visual Studio 2022 or higher on Windows [2], older versions will work but only supported for commercial users. Configure to use the English language pack for best results (Nuitka filters away garbage outputs, but only for English language). It will be used by default if installed.
  • On all other platforms, the gcc compiler of at least version 5.1, and below that the g++ compiler of at least version 4.4 as an alternative.
  • The clang compiler on macOS X and most FreeBSD architectures.
  • On Windows the clang-cl compiler on Windows can be used if provided by the Visual Studio installer.

• Python: Version 2.6, 2.7 or 3.3, 3.4, 3.5, 3.6, 3.7, 3.8, 3.9, 3.10, 3.11

  Important

  For Python 3.3/3.4 and only those, we need other Python version as a compile time dependency.

  Nuitka itself is fully compatible with all listed versions, but Scons as an internally used tool is not.

  For these versions, you need a Python2 or Python 3.5 or higher installed as well, but only during the compile time only. That is for use with Scons (which orchestrates the C compilation), which does not support the same Python versions as Nuitka.

  In addition, on Windows, Python2 cannot be used because clcache does not work with it, there a Python 3.5 or higher needs to be installed.

  Nuitka finds these needed Python versions (e.g. on Windows via registry) and you shouldn't notice it as long as they are installed.

  Increasingly, other functionality is available when another Python has a certain package installed. For example, onefile compression will work for a Python 2.x when another Python is found that has the zstandard package installed.

  Moving binaries to other machines

  The created binaries can be made executable independent of the Python installation, with --standalone and --onefile options.

  Binary filename suffix

  The created binaries have an .exe suffix on Windows. On other platforms they have no suffix for standalone mode, or .bin suffix, that you are free to remove or change, or specify with the -o option.

  The suffix for acceleration mode is added just to be sure that the original script name and the binary name do not ever collide, so we can safely do an overwrite without destroying the original source file.

  It has to be CPython, Anaconda Python, or Homebrew

  You need the standard Python implementation, called "CPython", to execute Nuitka, because it is closely tied to implementation details of it.

  It cannot be from Windows app store

  It is known that Windows app store Python definitely does not work, it's checked against.

  It cannot be pyenv on macOS

  It is know that macOS "pyenv" does not work. Use Homebrew instead for self compiled Python installations. But note that standalone mode will be worse on these platforms and not be as backward compatible with older macOS versions.

• Operating System: Linux, FreeBSD, NetBSD, macOS X, and Windows (32bits/64 bits/ARM).

  Others may work as well. The portability is expected to be generally good, but the e.g. Scons usage may have to be adapted. Make sure to match Windows Python and C compiler architecture, or else you will get cryptic error messages.

• Architectures: x86, x86_64 (amd64), and arm, likely many more

  Other architectures are expected to also work, out of the box, as Nuitka is generally not using any hardware specifics. These are just the ones tested and known to be good. Feedback is welcome. Generally, the architectures that Debian supports can be considered good and tested too.

The recommended way of executing Nuitka is <the_right_python> -m nuitka to be absolutely certain which Python interpreter you are using, so it is easier to match with what Nuitka has.

The next best way of executing Nuitka bare that is from a source checkout or archive, with no environment variable changes, most noteworthy, you do not have to mess with PYTHONPATH at all for Nuitka. You just execute the nuitka and nuitka-run scripts directly without any changes to the environment. You may want to add the bin directory to your PATH for your convenience, but that step is optional.

Moreover, if you want to execute with the right interpreter, in that case, be sure to execute <the_right_python> bin/nuitka and be good.

Nuitka has a --help option to output what it can do:

nuitka --help

The nuitka-run command is the same as nuitka, but with a different default. It tries to compile and directly execute a Python script:

nuitka-run --help

This option that is different is --run, and passing on arguments after the first non-option to the created binary, so it is somewhat more similar to what plain python will do.

For most systems, there will be packages on the download page of Nuitka. But you can also install it from source code as described above, but also like any other Python program it can be installed via the normal python setup.py install routine.

Nuitka is licensed under the Apache License, Version 2.0; you may not use it except in compliance with the License.

You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

This is basic steps if you have nothing installed, of course if you have any of the parts, just skip it.

• Download and install Python from https://www.python.org/downloads/windows
• Select one of Windows x86-64 web-based installer (64 bits Python, recommended) or x86 executable (32 bits Python) installer.
• Verify it's working using command python --version.

• python -m pip install nuitka
• Verify using command python -m nuitka --version

• mkdir HelloWorld
• make a python file named hello.py

def talk(message):
    return "Talk " + message


def main():
    print(talk("Hello World"))


if __name__ == "__main__":
    main()

Do as you normally would. Running Nuitka on code that works incorrectly is not easier to debug.

python hello.py

python -m nuitka hello.py

Execute the hello.exe created near hello.py.

To distribute, build with --standalone option, which will not output a single executable, but a whole folder. Copy the resulting hello.dist folder to the other machine and run it.

You may also try --onefile which does create a single file, but make sure that the mere standalone is working, before turning to it, as it will make the debugging only harder, e.g. in case of missing data files.

If you want to compile a whole program recursively, and not only the single file that is the main program, do it like this:

python -m nuitka --follow-imports program.py

In case you have a source directory with dynamically loaded files, i.e. one which cannot be found by recursing after normal import statements via the PYTHONPATH (which would be the recommended way), you can always require that a given directory shall also be included in the executable:

python -m nuitka --follow-imports --include-plugin-directory=plugin_dir program.py

If you want to compile a single extension module, all you have to do is this:

python -m nuitka --module some_module.py

The resulting file some_module.so can then be used instead of some_module.py.

If you need to compile a whole package and embed all modules, that is also feasible, use Nuitka like this:

python -m nuitka --module some_package --include-package=some_package

For distribution to other systems, there is the standalone mode which produces a folder for which you can specify --standalone.

python -m nuitka --standalone program.py

Following all imports is default in this mode. You can selectively exclude modules by specifically saying --nofollow-import-to, but then an ImportError will be raised when import of it is attempted at program run time. This may cause different behavior, but it may also improve your compile time if done wisely.

For data files to be included, use the option --include-data-files=<source>=<target> where the source is a file system path, but target has to be specified relative. For standalone you can also copy them manually, but this can do extra checks, and for onefile mode, there is no manual copying possible.

To copy some or all file in a directory, use the option --include-data-files=/etc/*.txt=etc/ where you get to specify shell patterns for the files, and a subdirectory where to put them, indicated by the trailing slash.

To copy a whole folder with all files, you can use --include-data-dir=/path/to/images=images which will copy all files including a potential subdirectory structure. You cannot filter here, i.e. if you want only a partial copy, remove the files beforehand.

For package data, there is a better way, using --include-package-data which detects data files of packages automatically and copies them over. It even accepts patterns in shell style. It spares you the need to find the package directory yourself and should be preferred whenever available.

With data files, you are largely on your own. Nuitka keeps track of ones that are needed by popular packages, but it might be incomplete. Raise issues if you encounter something in these.

When that is working, you can use the onefile mode if you so desire.

python -m nuitka --onefile program.py

This will create a single binary, that extracts itself on the target, before running the program. But notice, that accessing files relative to your program is impacted, make sure to read the section Onefile: Finding files as well.

# Create a binary that unpacks into a temporary folder
python -m nuitka --onefile program.py

For the unpacking, by default a unique user temporary path one is used, and then deleted, however this default --onefile-tempdir-spec="%TEMP%/onefile_%PID%_%TIME%" can be overridden with a path specification that is using then using a cached path, avoiding repeated unpacking, e.g. with --onefile-tempdir-spec="%CACHE_DIR%/%COMPANY%/%PRODUCT%/%VERSION%" which uses version information, and user specific cache directory.

Currently these expanded tokens are available:

If you have a setup.py, setup.cfg or pyproject.toml driven creation of wheels for your software in place, putting Nuitka to use is extremely easy.

Lets start with the most common setuptools approach, you can - having Nuitka installed of course, simply execute the target bdist_nuitka rather than the bdist_wheel. It takes all the options and allows you to specify some more, that are specific to Nuitka.

# For setup.py if not you't use other build systems:
setup(
   # Data files are to be handled by setuptools and not Nuitka
   package_data={"some_package": ["some_file.txt"]},
   ...,
   # This is to pass Nuitka options.
   command_options={
      'nuitka': {
         # boolean option, e.g. if you cared for C compilation commands
         '--show-scons': True,
         # options without value, e.g. enforce using Clang
         '--clang': None,
         # options with single values, e.g. enable a plugin of Nuitka
         '--enable-plugin': "pyside2",
         # options with several values, e.g. avoiding including modules
         '--nofollow-import-to' : ["*.tests", "*.distutils"],
      },
   },
)

# For setup.py with other build systems:
# The tuple nature of the arguments is required by the dark nature of
# "setuptools" and plugins to it, that insist on full compatibility,
# e.g. "setuptools_rust"

setup(
   # Data files are to be handled by setuptools and not Nuitka
   package_data={"some_package": ["some_file.txt"]},
   ...,
   # This is to pass Nuitka options.
   ...,
   command_options={
      'nuitka': {
         # boolean option, e.g. if you cared for C compilation commands
         '--show-scons': ("setup.py", True),
         # options without value, e.g. enforce using Clang
         '--clang': ("setup.py", None),
         # options with single values, e.g. enable a plugin of Nuitka
         '--enable-plugin': ("setup.py", "pyside2"),
         # options with several values, e.g. avoiding including modules
         '--nofollow-import-to' : ("setup.py", ["*.tests", "*.distutils"]),
      }
   },
)

If for some reason, you cannot or do not what to change the target, you can add this to your setup.py.

# For setup.py
setup(
   ...,
   build_with_nuitka=True
)

Or you could put it in your setup.cfg

[metadata]
build_with_nuitka = True

And last, but not least, Nuitka also supports the new build meta, so when you have a pyproject.toml already, simple replace or add this value:

[build-system]
requires = ["setuptools>=42", "wheel", "nuitka", "toml"]
build-backend = "nuitka.distutils.Build"

# Data files are to be handled by setuptools and not Nuitka
[tool.setuptools.package-data]
some_package = ['data_file.txt']

[nuitka]
# These are not recommended, but they make it obvious to have effect.

# boolean option, e.g. if you cared for C compilation commands, leading
# dashes are omitted
show-scons = true

# options with single values, e.g. enable a plugin of Nuitka
enable-plugin = pyside2

# options with several values, e.g. avoiding including modules, accepts
# list argument.
nofollow-import-to = ["*.tests", "*.distutils"]

If you have multiple programs, that each should be executable, in the past you had to compile multiple times, and deploy all of these. With standalone mode, this of course meant that you were fairly wasteful, as sharing the folders could be done, but wasn't really supported by Nuitka.

Enter Multidist. There is an option --main that replaces or adds to the positional argument given. And it can be given multiple times. When given multiple times, Nuitka will create a binary that contains the code of all the programs given, but sharing modules used in them. They therefore do not have to be distributed multiple times.

Lets call the basename of the main path, and entry point. The names of these must of course be different. Then the created binary can execute either entry point, and will react to what sys.argv[0] appears to it. So if executed in the right way (with something like subprocess or OS API you can control this name), or by renaming or copying the binary, or symlinking to it, you can then achieve the miracle.

This allows to combine very different programs into one.

This mode works with standalone, onefile, and mere acceleration. It does not work with module mode.

For good looks, you may specify icons. On Windows, you can provide an icon file, a template executable, or a PNG file. All of these will work and may even be combined:

# These create binaries with icons on Windows
python -m nuitka --onefile --windows-icon-from-ico=your-icon.png program.py
python -m nuitka --onefile --windows-icon-from-ico=your-icon.ico program.py
python -m nuitka --onefile --windows-icon-template-exe=your-icon.ico program.py

# These create application bundles with icons on macOS
python -m nuitka --macos-create-app-bundle --macos-app-icon=your-icon.png program.py
python -m nuitka --macos-create-app-bundle --macos-app-icon=your-icon.icns program.py

Entitlements for an macOS application bundle can be added with the option, --macos-app-protected-resource, all values are listed on this page from Apple

An example value would be --macos-app-protected-resource=NSMicrophoneUsageDescription:Microphone access for requesting access to a Microphone. After the colon, the descriptive text is to be given.

On Windows, the console is opened by programs unless you say so. Nuitka defaults to this, effectively being only good for terminal programs, or programs where the output is requested to be seen. There is a difference in pythonw.exe and python.exe along those lines. This is replicated in Nuitka with the option --disable-console. Nuitka recommends you to consider this in case you are using PySide6 e.g. and other GUI packages, e.g. wx, but it leaves the decision up to you. In case, you know your program is console application, just using --enable-console which will get rid of these kinds of outputs from Nuitka.

Splash screens are useful when program startup is slow. Onefile startup itself is not slow, but your program may be, and you cannot really know how fast the computer used will be, so it might be a good idea to have them. Luckily with Nuitka, they are easy to add for Windows.

For splash screen, you need to specify it as an PNG file, and then make sure to disable the splash screen when your program is ready, e.g. has complete the imports, prepared the window, connected to the database, and wants the splash screen to go away. Here we are using the project syntax to combine the code with the creation, compile this:

# nuitka-project: --onefile
# nuitka-project: --onefile-windows-splash-screen-image={MAIN_DIRECTORY}/Splash-Screen.png

# Whatever this is obviously
print("Delaying startup by 10s...")
import time, tempfile, os
time.sleep(10)

# Use this code to signal the splash screen removal.
if "NUITKA_ONEFILE_PARENT" in os.environ:
   splash_filename = os.path.join(
      tempfile.gettempdir(),
      "onefile_%d_splash_feedback.tmp" % int(os.environ["NUITKA_ONEFILE_PARENT"]),
   )

   if os.path.exists(splash_filename):
      os.unlink(splash_filename)

print("Done... splash should be gone.")
...

# Rest of your program goes here.

For analysis of your program and Nuitka packaging, there is the Compilation Report available. You can also make custom reports providing your own template, with a few of them built-in to Nuitka. These reports carry all the detail information, e.g. when a module was attempted to be imported, but not found, you can see where that happens. For bug reporting, it is very much recommended to provide the report.

You can attach copyright and trademark information, company name, product name, and so on to your compilation. This is then used in version information for the created binary on Windows, or application bundle on macOS. If you find something that it's lacking, let us know.

Binaries compiled on Windows with default settings of Nuitka and no further actions taken might be recognized by some AV vendors as malware. This is avoidable, but only in Nuitka commercial there is actual support and instructions for how to do it, seeing this as a typical commercial only need. https://nuitka.net/doc/commercial.html

Sometimes the C compilers will crash saying they cannot allocate memory or that some input was truncated, or similar error messages, clearly from it. There are several options you can explore here:

There is a dedicated option --low-memory which influences decisions of Nuitka, such that it avoids high usage of memory during compilation at the cost of increased compile time.

Do not use a 32 bits compiler, but a 64 bit one. If you are using Python with 32 bits on Windows, you most definitely ought to use MSVC as the C compiler, and not MinGW64. The MSVC is a cross compiler, and can use more memory than gcc on that platform. If you are not on Windows, that is not an option of course. Also using the 64 bits Python will work.

When you compile from a living installation, that may well have many optional dependencies of your software installed. Some software, will then have imports on these, and Nuitka will compile them as well. Not only may these be just the trouble makers, they also require more memory, so get rid of that. Of course you do have to check that your program has all needed dependencies before you attempt to compile, or else the compiled program will equally not run.

With --lto=yes or --lto=no you can switch the C compilation to only produce bytecode, and not assembler code and machine code directly, but make a whole program optimization at the end. This will change the memory usage pretty dramatically, and if you error is coming from the assembler, using LTO will most definitely avoid that.

People have reported that programs that fail to compile with gcc due to its bugs or memory usage work fine with clang on Linux. On Windows, this could still be an option, but it needs to be implemented first for the automatic downloaded gcc, that would contain it. Since MSVC is known to be more memory effective anyway, you should go there, and if you want to use Clang, there is support for the one contained in MSVC.

On systems with not enough RAM, you need to use swap space. Running out of it is possibly a cause, and adding more swap space, or one at all, might solve the issue, but beware that it will make things extremely slow when the compilers swap back and forth, so consider the next tip first or on top of it.

With the --jobs option of Nuitka, it will not start many C compiler instances at once, each competing for the scarce resource of RAM. By picking a value of one, only one C compiler instance will be running, and on a 8 core system, that reduces the amount of memory by factor 8, so that's a natural choice right there.

If your script modifies sys.path to e.g. insert directories with source code relative to it, Nuitka will not be able to see those. However, if you set the PYTHONPATH to the resulting value, it will be able to compile it and find the used modules from these paths as well.

A very frequent pattern with private code is that it scans plugin directories of some kind, and e.g. uses os.listdir, then considers Python filenames, and then opens a file and does exec on them. This approach is working for Python code, but for compiled code, you should use this much cleaner approach, that works for pure Python code and is a lot less vulnerable.

# Using a package name, to locate the plugins. This is also a sane
# way to organize them into a directory.
scan_path = scan_package.__path__

for item in pkgutil.iter_modules(scan_path):
   importlib.import_module(scan_package.__name__ + "." + item.name)

   # You may want to do it recursively, but we don't do this here in
   # this example. If you want to, handle that in this kind of branch.
   if item.ispkg:
      ...

If your program fails to file data, it can cause all kinds of different behaviors, e.g. a package might complain it is not the right version, because a VERSION file check defaulted to unknown. The absence of icon files or help texts, may raise strange errors.

Often the error paths for files not being present are even buggy and will reveal programming errors like unbound local variables. Please look carefully at these exceptions keeping in mind that this can be the cause. If you program works without standalone, chances are data files might be cause.

The most common error indicating file absence is of course an uncaught FileNotFoundError with a filename. You should figure out what package is missing files and then use --include-package-data (preferably), or --include-data-dir/--include-data-files to include them.

Nuitka has plugins that deal with copying DLLs. For NumPy, SciPy, Tkinter, etc.

These need special treatment to be able to run on other systems. Manually copying them is not enough and will given strange errors. Sometimes newer version of packages, esp. NumPy can be unsupported. In this case you will have to raise an issue, and use the older one.

If you want to manually add a DLL or an EXE, because it is your project only, you will have to use user Yaml files describing where they can be found. This is described in detail with examples in the Nuitka Package Configuration page.

Some packages are a single import, but to Nuitka mean that more than a thousand packages (literally) are to be included. The prime example of Pandas, which does want to plug and use just about everything you can imagine. Multiple frameworks for syntax highlighting everything imaginable take time.

Nuitka will have to learn effective caching to deal with this in the future. Right now, you will have to deal with huge compilation times for these.

A major weapon in fighting dependency creep should be applied, namely the anti-bloat plugin, which offers interesting abilities, that can be put to use and block unneeded imports, giving an error for where they occur. Use it e.g. like this --noinclude-pytest-mode=nofollow --noinclude-setuptools-mode=nofollow and e.g. also --noinclude-custom-mode=setuptools:error to get the compiler to error out for a specific package. Make sure to check its help output. It can take for each module of your choice, e.g. forcing also that e.g. PyQt5 is considered uninstalled for standalone mode.

It's also driven by a configuration file, anti-bloat.yml that you can contribute to, removing typical bloat from packages. Feel free to enhance it and make PRs towards Nuitka with it.

The standard code that normally works, also works, you should refer to os.path.dirname(__file__) or use all the packages like pkgutil, pkg_resources, importlib.resources to locate data files near the standalone binary.

There is a difference between sys.argv[0] and __file__ of the main module for onefile mode, that is caused by using a bootstrap to a temporary location. The first one will be the original executable path, where as the second one will be the temporary or permanent path the bootstrap executable unpacks to. Data files will be in the later location, your original environment files will be in the former location.

Given 2 files, one which you expect to be near your executable, and one which you expect to be inside the onefile binary, access them like this.

# This will find a file *near* your onefile.exe
open(os.path.join(os.path.dirname(sys.argv[0]), "user-provided-file.txt"))
# This will find a file *inside* your onefile.exe
open(os.path.join(os.path.dirname(__file__), "user-provided-file.txt"))

For debugging purposes, remove --disable-console or use the options --force-stdout-spec and --force-stderr-spec with paths as documented for --onefile-tempdir-spec above. These can be relative to the program or absolute, so you can see the outputs given.

Sometimes people use this kind of code, which for packages on PyPI, we deal with by doing source code patches on the fly. If this is in your own code, here is what you can do:

def binder(func, name):
   result = types.FunctionType(func.__code__, func.__globals__, name=func.__name__, argdefs=func.__defaults__, closure=func.__closure__)
   result = functools.update_wrapper(result, func)
   result.__kwdefaults__ = func.__kwdefaults__
   result.__name__ = name
   return result

Compiled functions cannot be used to create uncompiled ones from, so the above code will not work. However, there is a dedicated clone method, that is specific to them, so use this instead.

def binder(func, name):
   try:
      result = func.clone()
   except AttributeError:
      result = types.FunctionType(func.__code__, func.__globals__, name=func.__name__, argdefs=func.__defaults__, closure=func.__closure__)
      result = functools.update_wrapper(result, func)
      result.__kwdefaults__ = func.__kwdefaults__

   result.__name__ = name
   return result

A package can be compiled with Nuitka, no problem, but when it comes to executing it, python -m compiled_module is not going to work and give the error No code object available for AssertsTest because the compiled module is not source code, and Python will not just load it. The closest would be python -c "import compile_module" and you might have to call the main function yourself.

To support this, the CPython runpy and/or ExtensionFileLoader would need improving such that Nuitka could supply its compiled module object for Python to use.

There is support for conditional options, and options using pre-defined variables, this is an example:

# Compilation mode, support OS specific.
# nuitka-project-if: {OS} in ("Windows", "Linux", "Darwin", "FreeBSD"):
#    nuitka-project: --onefile
# nuitka-project-if: {OS} not in ("Windows", "Linux", "Darwin", "FreeBSD"):
#    nuitka-project: --standalone

# The PySide2 plugin covers qt-plugins
# nuitka-project: --enable-plugin=pyside2
# nuitka-project: --include-qt-plugins=sensible,qml

The comments must be a start of line, and indentation is to be used, to end a conditional block, much like in Python. There are currently no other keywords than the used ones demonstrated above.

You can put arbitrary Python expressions there, and if you wanted to e.g. access a version information of a package, you could simply use __import__("module_name").__version__ if that would be required to e.g. enable or disable certain Nuitka settings. The only thing Nuitka does that makes this not Python expressions, is expanding {variable} for a pre-defined set of variables:

Table with supported variables:

The use of {MAIN_DIRECTORY} is recommended when you want to specify a filename relative to the main script, e.g. for use in data file options or user package configuration yaml files,

# nuitka-project: --include-data-files={MAIN_DIRECTORY}/my_icon.png=my_icon.png
# nuitka-project: --user-package-configuration-file={MAIN_DIRECTORY}/user.nuitka-package.config.yml

For passing things like -O or -S to Python, to your compiled program, there is a command line option name --python-flag= which makes Nuitka emulate these options.

The most important ones are supported, more can certainly be added.

The C compiler, when invoked with the same input files, will take a long time and much CPU to compile over and over. Make sure you are having ccache installed and configured when using gcc (even on Windows). It will make repeated compilations much faster, even if things are not yet not perfect, i.e. changes to the program can cause many C files to change, requiring a new compilation instead of using the cached result.

On Windows, with gcc Nuitka supports using ccache.exe which it will offer to download from an official source and it automatically. This is the recommended way of using it on Windows, as other versions can e.g. hang.

Nuitka will pick up ccache if it's in found in system PATH, and it will also be possible to provide if by setting NUITKA_CCACHE_BINARY to the full path of the binary, this is for use in CI systems where things might be non-standard.

For the MSVC compilers and ClangCL setups, using the clcache is automatic and included in Nuitka.

On macOS and Intel, there is an automatic download of a ccache binary from our site, for arm64 arches, it's recommended to use this setup, which installs Homebrew and ccache in there. Nuitka picks that one up automatically if it on that kind of machine. You need and should not use Homebrew with Nuitka otherwise, it's not the best for standalone deployments, but we can take ccache from there.

export HOMEBREW_INSTALL_FROM_API=1
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install.sh)"
eval $(/opt/homebrew/bin/brew shellenv)
brew install ccache

The storage for cache results of all kinds, downloads, cached compilation results from C and Nuitka, is done in a platform dependent directory as determined by the appdirs package. However, you can override it with setting the environment variable NUITKA_CACHE_DIR to a base directory. This is for use in environments where the home directory is not persisted, but other paths are.

Avoid running the nuitka binary, doing python -m nuitka will make a 100% sure you are using what you think you are. Using the wrong Python will make it give you SyntaxError for good code or ImportError for installed modules. That is happening, when you run Nuitka with Python2 on Python3 code and vice versa. By explicitly calling the same Python interpreter binary, you avoid that issue entirely.

The fastest binaries of pystone.exe on Windows with 64 bits Python proved to be significantly faster with MinGW64, roughly 20% better score. So it is recommended for use over MSVC. Using clang-cl.exe of Clang7 was faster than MSVC, but still significantly slower than MinGW64, and it will be harder to use, so it is not recommended.

On Linux for pystone.bin the binary produced by clang6 was faster than gcc-6.3, but not by a significant margin. Since gcc is more often already installed, that is recommended to use for now.

Differences in C compilation times have not yet been examined.

Using the Python DLL, like standard CPython does can lead to unexpected slowdowns, e.g. in uncompiled code that works with Unicode strings. This is because calling to the DLL rather than residing in the DLL causes overhead, and this even happens to the DLL with itself, being slower, than a Python all contained in one binary.

So if feasible, aim at static linking, which is currently only possible with Anaconda Python on non-Windows, Debian Python2, self compiled Pythons (do not activate --enable-shared, not needed), and installs created with pyenv.

The process of making standalone executables for Windows traditionally involves using an external dependency walker in order to copy necessary libraries along with the compiled executables to the distribution folder.

There is plenty of ways to find that something is missing. Do not manually copy things into the folder, esp. not DLLs, as that's not going to work. Instead make bug reports to get these handled by Nuitka properly.

On Windows, the Windows Defender tool and the Windows Indexing Service both scan the freshly created binaries, while Nuitka wants to work with it, e.g. adding more resources, and then preventing operations randomly due to holding locks. Make sure to exclude your compilation stage from these services.

Whether compiling with MingW or MSVC, the standalone programs have external dependencies to Visual C Runtime libraries. Nuitka tries to ship those dependent DLLs by copying them from your system.

Beginning with Microsoft Windows 10, Microsoft ships ucrt.dll (Universal C Runtime libraries) which handles calls to api-ms-crt-*.dll.

With earlier Windows platforms (and wine/ReactOS), you should consider installing Visual C runtime libraries before executing a Nuitka standalone compiled program.

Depending on the used C compiler, you'll need the following redist versions on the target machines. However notice that compilation using the 14.3 based version is recommended.

When using MingGW64, you'll need the following redist versions:

Once the corresponding runtime libraries are installed on the target system, you may remove all api-ms-crt-*.dll files from your Nuitka compiled dist folder.

Nuitka does not sys.frozen unlike other tools, because it usually triggers inferior code for no reason. For Nuitka, we have the module attribute __compiled__ to test if a specific module was compiled, and the function attribute __compiled__ to test if a specific function was compiled.

Nuitka will apply values from the environment variables CCFLAGS, LDFLAGS during the compilation on top of what it determines to be necessary. Beware of course, that is this is only useful if you know what you are doing, so should this pose an issues, raise them only with perfect information.

Nuitka will automatically target the architecture of the Python you are using. If this is 64 bits, it will create a 64 bits binary, if it is 32 bits, it will create a 32 bits binary. You have the option to select the bits when you download the Python. In the output of python -m nuitka --version there is a line for the architecture. It Arch: x86_64 for 64 bits, and just Arch: x86 for 32 bits.

The C compiler will be picked to match that more or less automatically. If you specify it explicitly and it mismatches, you will get a warning about the mismatch and informed that you compiler choice was rejected.

When you use --report=compilation-report.xml Nuitka will create an XML file with detailed information about the compilation and packaging process. This is growing in completeness with very release and exposes module usage attempts, timings of the compilation, plugin influences, data file paths, DLLs, and reasons why things are included or not.

At this time, the report contains absolute paths in some places, with your private information. The goal is to make this blended out by default, because we also want to become able to compare compilation reports from different setups, e.g. with updated packages, and see the changes to Nuitka. The report is however recommended for your bug reporting.

Also, another form is available, where the report is free form and according to a Jinja2 template of yours, and one that is included in Nuitka. The same information as used to produce the XML file is accessible. However, right now this is not yet documented, but we plan to add a table with the data. For reader of the source code that is familiar with Jinja2, however, it will be easy to do it now already.

If you have a template, you can use it like this --report-template=your_template.rst.j2:your_report.rst and of course, the usage of restructured text, is only an example. You can use markdown, your own XML, or whatever you see fit. Nuitka will just expand the template with the compilation report data.

Currently the follow reports are included in Nuitka. You just use the name as a filename, and Nuitka will pick that one instead.

This chapter gives an overview, of what to currently expect in terms of performance from Nuitka. It's a work in progress and is updated as we go. The current focus for performance measurements is Python 2.7, but 3.x is going to follow later.

The results are the top value from this kind of output, running pystone 1000 times and taking the minimal value. The idea is that the fastest run is most meaningful, and eliminates usage spikes.

echo "Uncompiled Python2"
for i in {1..100}; do BENCH=1 python2 tests/benchmarks/pystone.py ; done | sort -rn | head -n 1
python2 -m nuitka --lto=yes --pgo tests/benchmarks/pystone.py
echo "Compiled Python2"
for i in {1..100}; do BENCH=1 ./pystone.bin ; done | sort -n | head -rn 1

echo "Uncompiled Python3"
for i in {1..100}; do BENCH=1 python3 tests/benchmarks/pystone3.py ; done | sort -rn | head -n 1
python3 -m nuitka --lto=yes --pgo tests/benchmarks/pystone3.py
echo "Compiled Python3"
for i in {1..100}; do BENCH=1 ./pystone3.bin ; done | sort -rn | head -n 1

Remember, this project needs constant work. Although the Python compatibility is insanely high, and test suite works near perfectly, there is still more work needed, esp. to make it do more optimization. Try it out, and when popular packages do not work, please make reports on GitHub.

Nuitka announcements and interesting stuff is pointed to on both the Mastodon and Twitter accounts, but obviously with not too many details, usually pointing to the website, but sometimes I also ask questions there.

@KayHayen on Mastodon. @KayHayen on Twitter.

Should you encounter any issues, bugs, or ideas, please visit the Nuitka bug tracker and report them.

Best practices for reporting bugs:

• Please always include the following information in your report, for the underlying Python version. You can easily copy&paste this into your report. It does contain more information that you think. Do not write something manually. You may always add of course.

  python -m nuitka --version

• Try to make your example minimal. That is, try to remove code that does not contribute to the issue as much as possible. Ideally come up with a small reproducing program that illustrates the issue, using print with different results when that programs runs compiled or native.

• If the problem occurs spuriously (i.e. not each time), try to set the environment variable PYTHONHASHSEED to 0, disabling hash randomization. If that makes the problem go away, try increasing in steps of 1 to a hash seed value that makes it happen every time, include it in your report.

• Do not include the created code in your report. Given proper input, it's redundant, and it's not likely that I will look at it without the ability to change the Python or Nuitka source and re-run it.

• Do not send screenshots of text, that is bad and lazy. Instead, capture text outputs from the console.

Consider using this software with caution. Even though many tests are applied before releases, things are potentially breaking. Your feedback and patches to Nuitka are very welcome.

You are more than welcome to join Nuitka development and help to complete the project in all minor and major ways.

The development of Nuitka occurs in git. We currently have these 3 branches:

• main

  This branch contains the stable release to which only hotfixes for bugs will be done. It is supposed to work at all times and is supported.

• develop

  This branch contains the ongoing development. It may at times contain little regressions, but also new features. On this branch, the integration work is done, whereas new features might be developed on feature branches.

• factory

  This branch contains unfinished and incomplete work. It is very frequently subject to git rebase and the public staging ground, where my work for develop branch lives first. It is intended for testing only and recommended to base any of your own development on. When updating it, you very often will get merge conflicts. Simply resolve those by doing git fetch && git reset --hard origin/factory and switch to the latest version.

Should you feel that you cannot help Nuitka directly, but still want to support, please consider making a donation and help this way.

The code objects are empty for native compiled functions. There is no bytecode with Nuitka's compiled function objects, so there is no way to provide it.

There is no tracing of compiled functions to attach a debugger to.

The most important form of optimization is the constant folding. This is when an operation can be fully predicted at compile time. Currently, Nuitka does these for some built-ins (but not all yet, somebody to look at this more closely will be very welcome!), and it does it e.g. for binary/unary operations and comparisons.

Constants currently recognized:

5 + 6  # binary operations
not 7  # unary operations
5 < 6  # comparisons
range(3)  # built-ins

Literals are the one obvious source of constants, but also most likely other optimization steps like constant propagation or function inlining will be. So this one should not be underestimated and a very important step of successful optimizations. Every option to produce a constant may impact the generated code quality a lot.

At the core of optimizations, there is an attempt to determine the values of variables at run time and predictions of assignments. It determines if their inputs are constants or of similar values. An expression, e.g. a module variable access, an expensive operation, may be constant across the module of the function scope and then there needs to be none or no repeated module variable look-up.

Consider e.g. the module attribute __name__ which likely is only ever read, so its value could be predicted to a constant string known at compile time. This can then be used as input to the constant folding.

if __name__ == "__main__":
    # Your test code might be here
    use_something_not_use_by_program()

Also, built-in exception name references are optimized if they are used as a module level read-only variables:

try:
    something()
except ValueError:  # The ValueError is a slow global name lookup normally.
    pass

For built-in calls like type, len, or range it is often possible to predict the result at compile time, esp. for constant inputs the resulting value often can be precomputed by Nuitka. It can simply determine the result or the raised exception and replace the built-in call with that value, allowing for more constant folding or code path reduction.

type("string")  # predictable result, builtin type str.
len([1, 2])  # predictable result
range(3, 9, 2)  # predictable result
range(3, 9, 0)  # predictable exception, range raises due to 0.

Sometimes the result of a built-in should not be predicted when the result is big. A range() call e.g. may give too big values to include the result in the binary. Then it is not done.

range(100000)  # We do not want this one to be expanded

For conditional statements, some branches may not ever be taken, because of the condition truth value being possible to predict. In these cases, the branch not taken and the condition check is removed.

This can typically predict code like this:

if __name__ == "__main__":
    # Your test code might be here
    use_something_not_use_by_program()

or

if False:
    # Your deactivated code might be here
    use_something_not_use_by_program()

It will also benefit from constant propagations, or enable them because once some branches have been removed, other things may become more predictable, so this can trigger other optimization to become possible.

Every branch removed makes optimization more likely. With some code branches removed, access patterns may be more friendly. Imagine e.g. that a function is only called in a removed branch. It may be possible to remove it entirely, and that may have other consequences too.

For exceptions that are determined at compile time, there is an expression that will simply do raise the exception. These can be propagated upwards, collecting potentially "side effects", i.e. parts of expressions that were executed before it occurred, and still have to be executed.

Consider the following code:

print(side_effect_having() + (1 / 0))
print(something_else())

The (1 / 0) can be predicted to raise a ZeroDivisionError exception, which will be propagated through the + operation. That part is just Constant Propagation as normal.

The call side_effect_having() will have to be retained though, but the print does not and can be turned into an explicit raise. The statement sequence can then be aborted and as such the something_else call needs no code generation or consideration anymore.

To that end, Nuitka works with a special node that raises an exception and is wrapped with a so-called "side_effects" expression, but yet can be used in the code as an expression having a value.

Consider the following code:

try:
    b = 8
    print(range(3, b, 0))
    print("Will not be executed")
except ValueError as e:
    print(e)

The try block is bigger than it needs to be. The statement b = 8 cannot cause a ValueError to be raised. As such it can be moved to outside the try without any risk.

b = 8
try:
    print(range(3, b, 0))
    print("Will not be executed")
except ValueError as e:
    print(e)

With the exception propagation, it then becomes possible to transform this code:

try:
    b = 8
    print(range(3, b, 0))
    print("Will not be executed!")
except ValueError as e:
    print(e)

try:
    raise ValueError("range() step argument must not be zero")
except ValueError as e:
    print(e)

Which then can be lowered in complexity by avoiding the raise and catch of the exception, making it:

e = ValueError("range() step argument must not be zero")
print(e)

For loops and conditional statements that contain only code without effect, it should be possible to remove the whole construct:

for i in range(1000):
    pass

The loop could be removed, at maximum, it should be considered an assignment of variable i to 999 and no more.

Another example:

if side_effect_free:
    pass

The condition check should be removed in this case, as its evaluation is not needed. It may be difficult to predict that side_effect_free has no side effects, but many times this might be possible.

When the length of the right-hand side of an assignment to a sequence can be predicted, the unpacking can be replaced with multiple assignments.

a, b, c = 1, side_effect_free(), 3

a = 1
b = side_effect_free()
c = 3

This is of course only really safe if the left-hand side cannot raise an exception while building the assignment targets.

We do this now, but only for constants, because we currently have no ability to predict if an expression can raise an exception or not.

When a construct like in xrange() or in range() is used, it is possible to know what the iteration does and represent that so that iterator users can use that instead.

I consider that:

for i in xrange(1000):
    something(i)

could translate xrange(1000) into an object of a special class that does the integer looping more efficiently. In case i is only assigned from there, this could be a nice case for a dedicated class.

Functions are structured so that their parameter parsing and tp_call interface is separate from the actual function code. This way the call can be optimized away. One problem is that the evaluation order can differ.

def f(a, b, c):
    return a, b, c


f(c=get1(), b=get2(), a=get3())

This will have to evaluate first get1(), then get2() and only then get3() and then make the function call with these values.

Therefore it will be necessary to have a staging of the parameters before making the actual call, to avoid a re-ordering of the calls to get1(), get2(), and get3().

In some cases, accesses to list constants can become tuple constants instead.

Consider that:

for x in [a, b, c]:
    something(x)

Can be optimized into this:

for x in (a, b, c):
    something(x)

This allows for simpler, faster code to be generated, and fewer checks needed, because e.g. the tuple is clearly immutable, whereas the list needs a check to assert that. This is also possible for sets.

In theory, something similar is also possible for dict. For the later, it will be non-trivial though to maintain the order of execution without temporary values introduced. The same thing is done for pure constants of these types, they change to tuple values when iterated.

Nuitka does not include metadata in the distribution. It's rather large, and the goal is to use it at compile time. Therefore information about entry points, version checks, etc. are all done at compile time rather than at run time. Not only is that faster, it also recognized problems sooner.

pkg_resources.require("lxml")
importlib.metadata.version("lxml")
...

This document is written in REST. That is an ASCII format which is readable to human, but easily used to generate PDF or HTML documents.

You will find the current version at: https://nuitka.net/doc/user-manual.html