30 September 2010
Nuitka Release 0.3.1
This is to inform you about the new stable release of Nuitka. It is the extremely compatible Python compiler, “download now”.
This release of Nuitka continues the focus on performance and contains only cleanups and optimization. Most go into the direction of more readable code, some aim at making the basic things faster, with good results as to performance as you can see below.
Optimization
Constants in conditions of conditional expressions (
a if cond else d),if/eliforwhileare now evaluated totrueorfalsedirectly. Before there would be temporary python object created from it which was then checked if it had a truth value.All of that is obviously overhead only. And it hurts the typically
while 1:infinite loop case badly.Do not generate code to catch
BreakExceptionorContinueExceptionunless abreakorcontinuestatement being in atry: finally:block inside that loop actually require this.Even while uncaught exceptions are cheap, it is still an improvement worthwhile and it clearly improves the readability for the normal case.
The compiler more aggressively prepares tuples, lists and dicts from the source code as constants if their contents is “immutable” instead of building at run time. An example of a “mutable” tuple would be
({},)which is not safe to share, and therefore will still be built at run time.For dictionaries and lists, copies will be made, under the assumption that copying a dictionary will always be faster, than making it from scratch.
The parameter parsing code was dynamically building the tuple of argument names to check if an argument name was allowed by checking the equivalent of
name in argument_names. This was of course wasteful and now a pre-built constant is used for this, so it should be much faster to call functions with keyword arguments.There are new templates files and also actual templates now for the
whileandforloop code generation. And I started work on having a template for assignments.
Cleanups
Do not generate code for the else of
whileandforloops if there is no such branch. This uncluttered the generated code somewhat.The indentation of the generated C++ was not very good and whitespace was often trailing, or e.g. a real tab was used instead of “t”. Some things didn’t play well together here.
Now much of the generated C++ code is much more readable and white space cleaner. For optimization to be done, the humans need to be able to read the generated code too. Mind you, the aim is not to produce usable C++, but on the other hand, it must be possible to understand it.
To the same end of readability, the empty
else {}branches are avoided forif,whileandforloops. While the C++ compiler can be expected to remove these, they seriously cluttered up things.The constant management code in
Contextwas largely simplified. Now the code is using theConstantclass to find its way around the problem that dicts, sets, etc. are not hashable, or thatcomplexis not being ordered; this was necessary to allow deeply nested constants, but it is also a simpler code now.The C++ code generated for functions now has two entry points, one for Python calls (arguments as a list and dictionary for parsing) and one where this has happened successfully. In the future this should allow for faster function calls avoiding the building of argument tuples and dictionaries all-together.
For every function there was a “traceback adder” which was only used in the C++ exception handling before exit to CPython to add to the traceback object. This was now in-lined, as it won’t be shared ever.
Numbers
python 2.6:
Pystone(1.1) time for 50000 passes = 0.65
This machine benchmarks at 76923.1 pystones/second
Nuitka 0.3.1:
Pystone(1.1) time for 50000 passes = 0.41
This machine benchmarks at 121951 pystones/second
This is 58% for 0.3.1, up from the 25% before. So it’s getting somewhere. As always you will find its latest version here.