- Frequently Asked Questions
- What is PyPy?
- Is PyPy a drop in replacement for CPython?
- Module xyz does not work with PyPy: ImportError
- Module xyz does not work in the sandboxed PyPy?
- Do CPython Extension modules work with PyPy?
- On which platforms does PyPy run?
- Which Python version (2.x?) does PyPy implement?
- Does PyPy have a GIL? Why?
- What about numpy, numpypy, micronumpy?
- Should I install numpy or numpypy?
- Is PyPy more clever than CPython about Tail Calls?
- How do I write extension modules for PyPy?
- How fast is PyPy?
- I wrote a 3-lines benchmark and it’s not faster than CPython. Why?
- Couldn’t the JIT dump and reload already-compiled machine code?
- Would type annotations help PyPy’s performance?
- Can I use PyPy’s translation toolchain for other languages besides Python?
- How do I get into PyPy development? Can I come to sprints?
- OSError: … cannot restore segment prot after reloc… Help?
- How should I report a bug?
- Why doesn’t PyPy use Git and move to GitHub?
- What is needed for Windows 64 support of PyPy?
- How long will PyPy support Python2?
See also: Frequently ask questions about RPython.
PyPy is a reimplementation of Python in Python, using the RPython translation toolchain.
PyPy tries to find new answers about ease of creation, flexibility, maintainability and speed trade-offs for language implementations. For further details see our goal and architecture document.
The most likely stumbling block for any given project is support for extension modules. PyPy supports a continually growing number of extension modules, but so far mostly only those found in the standard library.
The language features (including builtin types and functions) are very refined and well tested, so if your project doesn’t use many extension modules there is a good chance that it will work with PyPy.
We list the known differences in cpython differences.
A module installed for CPython is not automatically available for PyPy — just like a module installed for CPython 2.6 is not automatically available for CPython 2.7 if you installed both. In other words, you need to install the module xyz specifically for PyPy.
On Linux, this means that you cannot use
apt-get or some similar
package manager: these tools are only meant for the version of CPython
provided by the same package manager. So forget about them for now
and read on.
It is quite common nowadays that xyz is available on PyPI and
<pypy> -mpip install xyz. The simplest solution is to
use virtualenv (as documented here). Then enter (activate) the virtualenv
pypy -mpip install xyz. If you don’t know or don’t want
virtualenv, you can also use
pip locally after
pypy -m ensurepip.
The ensurepip module is built-in to the PyPy downloads we provide.
Best practices with
pip is to always call it as
<python> -mpip ...,
but if you wish to be able to call
pip directly from the command line, you
pypy -mensurepip --default-pip.
If you get errors from the C compiler, the module is a CPython C Extension module using unsupported features. See below.
Alternatively, if either the module xyz is not available on PyPI or you
don’t want to use virtualenv, then download the source code of xyz,
decompress the zip/tarball, and run the standard command:
setup.py install. (Note: pypy here instead of python.) As usual
you may need to run the command with sudo for a global installation.
The other commands of
setup.py are available too, like
You cannot import any extension module in a sandboxed PyPy, sorry. Even the built-in modules available are very limited. Sandboxing in PyPy is a good proof of concept, and is without a doubt safe IMHO, however it is only a proof of concept. It currently requires some work from a motivated developer. However, until then it can only be used for “pure Python” example: programs that import mostly nothing (or only pure Python modules, recursively).
First note that some Linux distributions (e.g. Ubuntu, Debian) split PyPy into several packages. If you installed a package called “pypy”, then you may also need to install “pypy-dev” for the following to work.
We have experimental support for CPython extension modules, so
they run with minor changes. This has been a part of PyPy since
the 1.4 release, but support is still in beta phase. CPython
extension modules in PyPy are often much slower than in CPython due to
the need to emulate refcounting. It is often faster to take out your
CPython extension and replace it with a pure python version that the
JIT can see. If trying to install module xyz, and the module has both
a C and a Python version of the same code, try first to disable the C
version; this is usually easily done by changing some line in
We fully support ctypes-based extensions. But for best performance, we recommend that you use the cffi module to interface with C code.
For more information about how we manage refcounting semamtics see rawrefcount
PyPy currently supports:
- x86 machines on most common operating systems (Linux 32/64 bits, Mac OS X 64 bits, Windows 32 bits, OpenBSD, FreeBSD),
- 64-bit AArch, also known as ARM64,
- ARM hardware (ARMv6 or ARMv7, with VFPv3) running Linux (we no longer provide prebuilt binaries for these),
- big- and little-endian variants of PPC64 running Linux,
- s390x running Linux
PyPy is regularly and extensively tested on Linux machines. It works on Mac and Windows: it is tested there, but most of us are running Linux so fixes may depend on 3rd-party contributions.
To bootstrap from sources, PyPy can use either CPython 2.7 or another (e.g. older) PyPy. Cross-translation is not really supported: e.g. to build a 32-bit PyPy, you need to have a 32-bit environment. Cross-translation is only explicitly supported between a 32-bit Intel Linux and ARM Linux (see here).
PyPy comes in two versions:
- one is fully compatible with Python 2.7;
- the other is fully compatible with one 3.x version. At the time of this writing, this is 3.6.
Yes, PyPy has a GIL. Removing the GIL is very hard. On top of CPython, you have two problems: (1) GC, in this case reference counting; (2) the whole Python language.
For PyPy, the hard issue is (2): by that I mean issues like what occurs if a mutable object is changed from one thread and read from another concurrently. This is a problem for any mutable type: it needs careful review and fixes (fine-grained locks, mostly) through the whole Python interpreter. It is a major effort, although not completely impossible, as Jython/IronPython showed. This includes subtle decisions about whether some effects are ok or not for the user (i.e. the Python programmer).
CPython has additionally the problem (1) of reference counting. With PyPy, this sub-problem is simpler: we need to make our GC multithread-aware. This is easier to do efficiently in PyPy than in CPython. It doesn’t solve the issue (2), though.
Note that since 2012 there is work going on on a still very experimental Software Transactional Memory (STM) version of PyPy. This should give an alternative PyPy which works without a GIL, while at the same time continuing to give the Python programmer the complete illusion of having one. This work is currently a bit stalled because of its own technical difficulties.
Way back in 2011, the PyPy team started to reimplement numpy in PyPy. It has two pieces:
- the builtin module pypy/module/micronumpy: this is written in RPython and roughly covers the content of the
numpy.core.multiarraymodule. Confusingly enough, this is available in PyPy under the name
_numpypy. It is included by default in all the official releases of PyPy (but it might be dropped in the future).
- a fork of the official numpy repository maintained by us and informally called
numpypy: even more confusing, the name of the repo on bitbucket is
numpy. The main difference with the upstream numpy, is that it is based on the micronumpy module written in RPython, instead of of
numpy.core.multiarraywhich is written in C.
TL;DR version: you should use numpy. You can install it by doing
pypy -m pip
install numpy. You might also be interested in using the experimental PyPy
binary wheels to save compilation time.
numpy is written in C, and runs under the cpyext
compatibility layer. Nowadays, cpyext is mature enough that you can simply
use the upstream
numpy, since it passes the test suite. At the
moment of writing (October 2017) the main drawback of
numpy is that cpyext
is infamously slow, and thus it has worse performance compared to
numpypy. However, we are actively working on improving it, as we expect to
reach the same speed when HPy can be used.
On the other hand,
numpypy is more JIT-friendly and very fast to call,
since it is written in RPython: but it is a reimplementation, and it’s hard to
be completely compatible: over the years the project slowly matured and
eventually it was able to call out to the LAPACK and BLAS libraries to speed
matrix calculations, and reached around an 80% parity with the upstream
numpy. However, 80% is far from 100%. Since cpyext/numpy compatibility is
progressing fast, we have discontinued support for
No. PyPy follows the Python language design, including the built-in debugger features. This prevents tail calls, as summarized by Guido van Rossum in two blog posts. Moreover, neither the JIT nor Stackless change anything to that.
This really depends on your code. For pure Python algorithmic code, it is very fast. For more typical Python programs we generally are 3 times the speed of CPython 2.7. You might be interested in our benchmarking site and our jit documentation.
Your tests are not a benchmark: tests tend to be slow under PyPy because they run exactly once; if they are good tests, they exercise various corner cases in your code. This is a bad case for JIT compilers. Note also that our JIT has a very high warm-up cost, meaning that any program is slow at the beginning. If you want to compare the timings with CPython, even relatively simple programs need to run at least one second, preferrably at least a few seconds. Large, complicated programs need even more time to warm-up the JIT.
Three-lines benchmarks are benchmarks that either do absolutely nothing (in which case PyPy is probably a lot faster than CPython), or more likely, they are benchmarks that spend most of their time doing things in C.
For example, a loop that repeatedly issues one complex SQL operation will only measure how performant the SQL database is. Similarly, computing many elements from the Fibonacci series builds very large integers, so it only measures how performant the long integer library is. This library is written in C for CPython, and in RPython for PyPy, but that boils down to the same thing.
PyPy speeds up the code written in Python.
No, we found no way of doing that. The JIT generates machine code containing a large number of constant addresses — constant at the time the machine code is generated. The vast majority is probably not at all constants that you find in the executable, with a nice link name. E.g. the addresses of Python classes are used all the time, but Python classes don’t come statically from the executable; they are created anew every time you restart your program. This makes saving and reloading machine code completely impossible without some very advanced way of mapping addresses in the old (now-dead) process to addresses in the new process, including checking that all the previous assumptions about the (now-dead) object are still true about the new object.
Cython types are, by construction, similar to C declarations. For
example, a local variable or an instance attribute can be declared
"cdef int" to force a machine word to be used. This changes the
usual Python semantics (e.g. no overflow checks, and errors when
trying to write other types of objects there). It gives some extra
performance, but the exact benefits are unclear: right now
(January 2015) for example we are investigating a technique that would
store machine-word integers directly on instances, giving part of the
benefits without the user-supplied
PEP 484 - Type Hints, on the other hand, is almost entirely useless if you’re looking at performance. First, as the name implies, they are hints: they must still be checked at runtime, like PEP 484 says. Or maybe you’re fine with a mode in which you get very obscure crashes when the type annotations are wrong; but even in that case the speed benefits would be extremely minor.
There are several reasons for why. One of them is that annotations
are at the wrong level (e.g. a PEP 484 “int” corresponds to Python 3’s
int type, which does not necessarily fits inside one machine word;
even worse, an “int” annotation allows arbitrary int subclasses).
Another is that a lot more information is needed to produce good code
f() called here really means this function there, and
will never be monkey-patched” – same with
btw). The third reason is that some “guards” in PyPy’s JIT traces
don’t really have an obvious corresponding type (e.g. “this dict is so
far using keys which don’t override
__hash__ so a more efficient
implementation was used”). Many guards don’t even have any correspondence
with types at all (“this class attribute was not modified”; “the loop
counter did not reach zero so we don’t need to release the GIL”; and
As PyPy works right now, it is able to derive far more useful information than can ever be given by PEP 484, and it works automatically. As far as we know, this is true even if we would add other techniques to PyPy, like a fast first-pass JIT.
Yes. The toolsuite that translates the PyPy interpreter is quite general and can be used to create optimized versions of interpreters for any language, not just Python. Of course, these interpreters can make use of the same features that PyPy brings to Python: translation to various languages, stackless features, garbage collection, implementation of various things like arbitrarily long integers, etc.
Certainly you can come to sprints! We always welcome newcomers and try to help them as much as possible to get started with the project. We provide tutorials and pair them with experienced PyPy developers. Newcomers should have some Python experience and read some of the PyPy documentation before coming to a sprint.
Coming to a sprint is usually the best way to get into PyPy development. If you get stuck or need advice, contact us. IRC is the most immediate way to get feedback (at least during some parts of the day; most PyPy developers are in Europe) and the mailing list is better for long discussions.
On Linux, if SELinux is enabled, you may get errors along the lines of “OSError: externmod.so: cannot restore segment prot after reloc: Permission denied.” This is caused by a slight abuse of the C compiler during configuration, and can be disabled by running the following command with root privileges:
# setenforce 0
This will disable SELinux’s protection and allow PyPy to configure correctly. Be sure to enable it again if you need it!
Our bug tracker is here: https://foss.heptapod.net/pypy/pypy/issues/
Missing features or incompatibilities with CPython are considered bugs, and they are welcome. (See also our list of known incompatibilities.)
For bugs of the kind “I’m getting a PyPy crash or a strange exception”, please note that: We can’t do anything without reproducing the bug ourselves. We cannot do anything with tracebacks from gdb, or core dumps. This is not only because the standard PyPy is compiled without debug symbols. The real reason is that a C-level traceback is usually of no help at all in PyPy. Debugging PyPy can be annoying.
This is a clear and useful bug report. (Admittedly, sometimes the problem is really hard to reproduce, but please try to.)
In more details:
- First, please give the exact PyPy version, and the OS.
- It might help focus our search if we know if the bug can be
reproduced on a “
pypy --jit off” or not. If “
pypy --jit off” always works, then the problem might be in the JIT. Otherwise, we know we can ignore that part.
- If you got the bug using only Open Source components, please give a step-by-step guide that we can follow to reproduce the problem ourselves. Don’t assume we know anything about any program other than PyPy. We would like a guide that we can follow point by point (without guessing or having to figure things out) on a machine similar to yours, starting from a bare PyPy, until we see the same problem. (If you can, you can try to reduce the number of steps and the time it needs to run, but that is not mandatory.)
- If the bug involves Closed Source components, or just too many Open Source components to install them all ourselves, then maybe you can give us some temporary ssh access to a machine where the bug can be reproduced. Or, maybe we can download a VirtualBox or VMWare virtual machine where the problem occurs.
- If giving us access would require us to use tools other than ssh, make appointments, or sign a NDA, then we can consider a commerical support contract for a small sum of money.
- If even that is not possible for you, then sorry, we can’t help.
Of course, you can try to debug the problem yourself, and we can help you get started if you ask on the #pypy IRC channel, but be prepared: debugging an annoying PyPy problem usually involves quite a lot of gdb in auto-generated C code, and at least some knowledge about the various components involved, from PyPy’s own RPython source code to the GC and possibly the JIT.
We discussed it during the switch away from bitbucket. We concluded that (1) the Git workflow is not as well suited as the Mercurial workflow for our style, and (2) moving to github “just because everybody else does” is a argument on thin grounds.
For (1), there are a few issues, but maybe the most important one is that the PyPy repository has got thousands of named branches. Git has no equivalent concept. Git has branches, of course, which in Mercurial are called bookmarks. We’re not talking about bookmarks.
The difference between git branches and named branches is not that important in a repo with 10 branches (no matter how big). But in the case of PyPy, we have at the moment 1840 branches. Most are closed by now, of course. But we would really like to retain (both now and in the future) the ability to look at a commit from the past, and know in which branch it was made. Please make sure you understand the difference between the Git and the Mercurial branches to realize that this is not always possible with Git— we looked hard, and there is no built-in way to get this workflow.
Still not convinced? Consider this git repo with three commits: commit #2 with parent #1 and head of git branch “A”; commit #3 with also parent #1 but head of git branch “B”. When commit #1 was made, was it in the branch “A” or “B”? (It could also be yet another branch whose head was also moved forward, or even completely deleted.)
First, please note that the Windows 32 PyPy binary works just fine on Windows 64. The only problem is that it only supports up to 4GB of heap per process.
As to real Windows 64 support: Currently we don’t have an active PyPy developer whose main development platform is Windows. So if you are interested in getting Windows 64 support, we encourage you to volunteer to make it happen! Another option would be to pay some PyPy developers to implement Windows 64 support, but so far there doesn’t seem to be an overwhelming commercial interest in it.