Visualization of JITted code
Hello.
We're proud to announce the first public release of the jitviewer. As of now, jitviewer is a slightly internal tool that helps understanding how your Python source code is compiled by the PyPy's JIT all the way down to machine code.
To install it, you need a very recent version of PyPy (newer than 9th of August), for example one of the nightly builds:
- install pip and distribute either by creating a PyPy virtualenv or by following the installation instructions.
- make sure to have a source code checkout of PyPy and put it in your PYTHONPATH.
- pip install jitviewer. Note that you need to run the pip executable which belongs to PyPy, not the globally installed one.
Have a look at the README for how to start it, or try the online demo if you just want to play with it.
The jitviewer is a web application written with flask and jinja2. If you have experience with web development and you want to help PyPy, don't hesitate to contact us, there are plenty of things to improve in it :-).
What does the jitviewer really do?¶
At the top of the page, you will see the list of pieces of code which has been compiled by the JIT. You will see entries for both normal loops and for "entry bridges". This is not the right place to discuss the difference between those, but you most probably want to look at loops, because usually it's where most of the time is spent.
Note that for each loop, you will see the name of the function which contains the first instruction of the loop. However, thanks to the inlining done by the JIT, it will contain also the code for other functions.
Once you select a loop, the jitviewer shows how the JIT has compiled the Python source code into assembler in a hierarchical way. It displays four levels:
-
Python source code: only the lines shown in azure have been compiled for this particular loop, the ones in gray have not.
-
Python bytecode, the one you would get by doing:
def f(a, b): return a + b import dis dis.dis(f)
The opcodes are e.g. LOAD_FAST, LOAD_GLOBAL etc. The opcodes which are not in bold have been completely optimized aways by the JIT.
-
Intermediate representation of jit code (IR). This is a combination of operations (like integer addition, reading fields out of structures) and guards (which check that the assumptions we made are actually true). Guards are in red. These operations are "at the same level as C": so, for example, + takes two unboxed integers which can be stored into the register of the CPU.
-
Assembler: you can see it by clicking on "Show assembler" in the menu on the right.
Sometimes you'll find that a guard fails often enough that a new piece of assembler is required to be compiled. This is an alternative path through the code and it's called a bridge. You can see bridges in the jitviewer when there is a link next to a guard. For more information about purpose look up the jit documentation.
Comments
I'm getting a TemplateNotFound jinja2 exception when I run the jitviewer.py as shown in the README.
I think you have to python setup.py install it in a virtualenv. It might not work from the checkout any more.
That fixed it, thanks.
Would it be possible to get some screenshots of jitviewer, as the online demo is currently down?
The demo doesn't work
Please, could you put it back?
Thanks a lot!
I'm developing a programming languaje based on mindmaps and I would like to know if my code works with pypy...
jitviewer repository - https://bitbucket.org/pypy/jitviewer