Nuitka Release 0.4.4
Posted: | More posts about Python compiler Nuitka |
This release marks the point, where Nuitka for the first time supports all major current Python versions and all major features. It adds Python 3.3 support and it adds support for threading. And then there is a massive amount of fixes that improve compatibility even further.
Aside of that, there is major performance work. One side is the optimization of call performance (to CPython non-compiled functions) and to compiled functions, both. This gave a serious improvement to performance.
Then of course, we are making other, long term performance progress, as in "--experimental" mode, the SSA code starts to optimize unused code away. That code is not yet ready for prime time yet, but the trace structure will hold.
The test suite of CPython3.3 passes now too. The yield from is now supported, but the improved argument parsing error messages are not implemented yet.
Tracing user provided constants, now Nuitka warns about too large constants produced during optimization.
Line numbers of expressions are now updates as evaluation progresses. This almost corrects.
Finally improves Issue#9. Now only expression parts that cannot raise, do not update, which can still cause difference, but much less often, and then definitely useless.
Experimental support for threads.
Threading appears to work just fine in the most cases. It's not as optimal as I wanted it to be, but that's going to change with time.
Previous corrections for ==, !=, and <=, caused a performance regression for these operations in case of handling identical objects.
For built-in objects of sane types (not float), these operations are now accelerated again. The overreaching acceleration of >= was still there (bug, see below) and has been adapted too.
Calling non-compiled Python functions from compiled functions was slower than in CPython. It is now just as fast.
Calling compiled functions without keyword arguments has been accelerated with a dedicated entry point that may call the implementation directly and avoid parameter parsing almost entirely.
Making calls to compiled and non-compiled Python functions no longer requires to build a temporary tuple and therefore is much faster.
Parameter parsing code is now more compact, and re-uses error raises, or creates them on the fly, instead of hard coding it. Saves binary size and should be more cache friendly.
Corrected false optimization of a >= a on C++ level.
When it's not done during Nuitka compile time optimization, the rich comparison helper still contained short cuts for >=. This is now the same for all the comparison operators.
Calling a function with default values, not providing it, and not providing a value for a value without default, was not properly detecting the error, and instead causing a run time crash.
def f(a, b = 2): pass f(b = 2)
This now properly raises the TypeError exception.
Constants created with + could become larger than the normally enforced limits. Not as likely to become huge, but still potentially an issue.
The vars built-in, when used on something without __dict__ attribute, was giving AttributeError instead of TypeError.
When re-cursing to modules at compile time, script directory and current directory were used last, while at run time, it was the other way around, which caused overloaded standard library modules to not be embedded. Corrects Issue#94.
Thanks for the patch to James Michael DuPont.
Super without arguments was not raising the correct RuntimeError exception in functions that cannot be methods, but UnboundLocalError instead.
def f(): super() # Error, cannot refer to first argument of f
Generators no longer use raise StopIteration for return statements, because that one is not properly handled in try/except clauses, where it's not supposed to trigger, while try/finally should be honored.
Exception error message when throwing non-exceptions into generators was not compatible.
The use of return with value in generators is a SyntaxError before Python3.3, but that was not raised.
Variable names of the "__var" style need to be mangled. This was only done for classes, but not for functions contained in classes, there they are now mangled too.
Python3: Exceptions raised with causes were not properly chaining.
Python3: Specifying the file encoding corrupted line numbers, making them all of by one.
For containers (tuple, list, set, dict) defined on the source code level, Nuitka immediately created constant references from them.
For function calls, class creations, slice objects, this code is now re-used, and its dictionaries and tuples, may now become constants immediately, reducing noise in optimization steps.
The parameter parsing code got cleaned up. There were a lot of relics from previously explored paths. And error raises were part of the templates, but now are external code.
Global variable management moved to module objects and out of "Variables" module.
Make sure, nodes in the tree are not shared by accident.
This helped to find a case of duplicate use in the complex call helpers functions. Code generation will now notice this kind of duplication in debug mode.
The complex call helper functions were manually taking variable closure, which made these functions inconsistent to other functions, e.g. no variable version was allocated to assignments.
Removing the manual setting of variables allowed a huge reduction of code volume, as it became more generic code.
Converting user provided constants to create containers into constants immediately, to avoid noise from doing this in optimization.
The site module is now imported explicitly in the __main__ module, so it can be handled by the recursion code as well. This will help portable mode.
Many line length 80 changes, improved comments.
- The CPython3.3 test suite was added, and run with both Python3.2 and Python3.3, finding new bugs.
- The doctest to code generation didn't successfully handle all tests, most notably, "test_generators.py" was giving a SyntaxError and therefore not actually active. Correcting that improved the coverage of generator testing.
The portable code is still delayed.
Support for Python3.3 was a higher priority, but the intention is to get it into shape for Europython still.
Added notes about it being disabled it in the "User Manual" documentation.
This release is in preparation for Europython 2013. Wanted to get this much out, as it changes the status slides quite a bit, and all of that was mostly done in my Cyprus holiday a while ago.
The portable code has not seen progress. The idea here is to get this into a development version later.