- Communication vs. Coding
- Google Summer of Code for Nuitka
- Website Overhaul
- Performance Work
- Improved Annotations
- Release or not
- Standalone Improvements
- Caching Examined
- Python3 deprecation warnings
- Help Wanted
My new communication strategy is a full success, engagement with Nuitka is on an all time high.
But the recent weeks more than ever highlighted why I have to force myself to do it. I do not like to talk about unfinished stuff. And right now, there is really a lot of it, almost only it. Also I was ill, and otherwise busy, so this is now late by a week.
But I am keeping it up, and will give an update, despite the feeling that it would be better to just finish a few of those things and then talk about it, but then it will take forever and leave you in the dark. And that is not what is supposed to be.
Bear in mind, that this is supposed to be a quick, not too polished, and straight from top of my head, even if really a lot of content. But I feel that esp. the optimization parts are worth reading.
So the 0.6.0 release was a huge success, but it definitely wasn't perfect, and hotfixes were necessary. The latest one 0.6.0.5 was done just yesterday and actually contains one for an important mis-optimization being done, and you ought to update to it from any prior 0.6.0 release.
There are also a few remaining compatibility issues fixed for 3.7 and generally using the latest hotfix is always a good idea.
Kind of what one has to expect from a 0 release, this one also had more expose than usual is seems.
I need more people to work on Nuitka. One way of doing this could be to participate in Google Summer of Code under the Python umbrella. To make that possible, I need you to volunteer as a mentor. So please, please, do.
I know you will feel not qualified. But I just need a backup that will help a student around obstacles in case I go missing. Contact me and I will be very happy.
I updated the website to recent Nikola and dropped the tag cloud that I was using. Should have cleaner and better looks. Also integrated privacy aware sharing links, where two clicks are necessary to share a page or article like this one on Twitter, Facebook, etc.
Also the download page saw some structural updates and polishing. It should easier to overview now.
The feedback for performance and the work on 0.6.1 are fully ongoing, and there are many major points that are ongoing. I want to briefly cover each one of them now, but many of them will only have full effect, once everything is in place, which each one is very critical.
So, with the type tracing, objects have known types, and short of using a C type, knowing e.g. that an object is an int, and the other one too, doing + for them can take a lot of advantage avoiding unrelated checks and code paths, even if still using PyObject * at the end of the day.
And even we are only knowing it's not an int, but say one value is a tuple and the other an unknown, that allows to remove checks for int shortcuts as they can no longer apply. These are tiny optimizations then, but still worthwhile.
To further this, first the inplace operations for a couple of more or less randomly selected types, list, tuple, int, long, str, unicode, bytes, and float, have been looked at and have gotten their own special object based helpers if one or both types are known to be of that kind.
A report has been added, that will tell when such an operation could have been used, but was not available. This uncovered where typical stuff goes non optimized, a nice principle to see what is actually happening.
So adding list and str would now give a warning, although of course, the optimization phase ought to catch the static raise that is and never let it get there, so this report also addresses missing optimization in an earlier phase.
So the in-place operations were then covered, so this was extended to mere + operations too, the ones that are not in-place. Sometimes, esp. for immutable types, there was already code for that, e.g. int doesn't really do it, in other cases, list + list code for a quicker concat was added.
And again a report for where it's missing was added and basic coverage for most of the types. However, in some instances, the optimization doesn't use the full knowledge yet. But where it does, it will shove off quite a few cycles.
To apply these things effectively, optimization and value tracing need to know types in the first place. I have found two obstacles for that. One are branch merges. If a branch or both assign to the same type or original type, well the type is changed. Previously it became "unknown" which is treated as object for code generation, and allows nothing really. But now that is better on develop now, and was actually a trivial missing thing.
The other area is loops. Loops put values to unknown when entering loop body, and again when leaving. Essentially making type tracing not effective where it is needed the most to achieve actual performance. Also this was limiting the knowledge for all function to one type to not happening for these kinds of variables that were assigned inside a loop at all.
Took me a while, but I figured out how to build type tracing for loops that works. It currently is still unfinished in my private repo, but passes all tests, I would just like to make it use dedicated interfaces, and clean it up.
I will most likely have that for 0.6.1 too and that should expand the cases where types are known in code generation by a fair amount.
The effect of that will be that more often C code generation will actually see types. Currently e.g. a boolean variable that is assigned in a loop, cannot use the C target type in code generation. Once loop code is merged, it will however take advantage there too. And only then I think adding "C int" as a C type makes sense at all.
Then another area is performance regressions. So one thing I did early on in the 0.6.1 cycle was using the "module var C target type" to get in-place working for those too. Doing string concatenations on module variables could be slower by an order of magnitude, as could be other operations.
I still need to do it for closure variables too. Then Nuitka will do at least as many of them perfectly as CPython does. It also will be better at it them, because e.g. it doesn't have to delete from the module dictionary first, due to it never taking a reference, and same applies to the cell. Should be faster for that too.
But strings in-place on these if not optimized, it will look very ugly in terms of worse performance, so 0.6.0 was still pretty bad for some users. This will however hopefully be addressed in 0.6.1 then.
Another field was in-place string add for the already optimized case, it was still slower than CPython, and I finally found out what causes this. And that is the using of libpython where PyUnicode_Append is far worse than in the python binary that you normally use, I have see that at least for 3.5 and higher CPython. Analysis showed that e.g. MiniConda had the issue to a much smaller extent, and was being much faster anyway, but probably just has better libpython compilation flags.
So what to do. Ultimately that was to be solved by including a clone of that function, dubbed UNICODE_APPEND that behaves the same, and can even shove off a couple of cycles, by indicating the Python error status without extra checks, and specializing it for the pure unicode += unicode case that we see most often, same for UNICODE_CONCAT for mere +.
Right now the benchmarks to show it do not exist yet. Again something that typically wants me to delay stuff. But as you can imagine, tracking down these hard issues, writing that much code to replace the unicode resizing, is hard enough by itself.
But I hope to convince myself that this will allow to show that for compiled code, things are going to be faster only now.
In fact, speedcenter as a whole is currently broken, mostly due to Nikola changes that I am trying to work around, but it will take more time apparently and isn't finished as I write this.
Another optimization end, is the type shapes of the + operation itself. Right now what is being done is that the shape is derived from the shape of the left argument with the right shape to be considered by it. These also have reports now, for cases where they are missing. So saying e.g. that int + float results in float and these kinds of things, are stuff being encoded there right now.
This is necessary step to e.g. know that int + int -> int_or_long, to make effective loop variable optimization.
Without these, and again, that is a lot of code to write, there is no way to hope for wide spread type knowledge in code generation.
Something missing there, is to also make it known that + unlike it currently is now, should not in all cases lead to "control flow escape" with the consequence of removing all stuff, and expecting an exception possible, but instead to let the int type also make known that + int ont it not only gives an int_or_long result shape, but also while doing so, that it will never raise an exception (bare MemoryError), and therefore allow more optimization to happen and less and therefore faster code generated.
Until this is done, what is actually going to happen is that while the + result is known, Nuitka will assume control flow escape.
And speaking of that, I think this puts too many variables to a too unknown state. You can to distrust all values, but not the types in this case, so that could be better, but right now it is not. Something else to look into.
So 0.6.1 is in full swing in terms of optimization. All these ends need a completion, and then I can expect to use advantage of things in a loop, and ultimately to generate C performance code for one example of loop. esp. if we add a C int target type, which currently isn't yet started, because I think it would barely be used yet.
But we are getting there and I wouldn't even say we are making small steps, this is all just work to be completed, nothing fundamental about it. But it may take more than one release for sure.
Mind you, there is not only +, there is also -, *, %, and many more operators, all of them will require work. Granted, loop variables tend to use + more often, but any un-optimized operation will immediately loose a lot of type knowledge.
There are two kinds of annotations, ones for classes and modules, which actually are stored in a __annotations__ variable, and everything else is mostly just ignored.
So Nuitka got the criterion wrong, and did one thing for functions, and the other for everything else. So that annotations in generators, coroutines and asyncgen ended up with wrong, crashing, and slower code, due to it updating the module __annotations__, so that one is important too if you have to do those.
To release or not. There is at least one bug about star imports that affects numpy that is solved in develop, and wasn't back ported, and I was thinking it only applies to develop, but in fact does to stable. It makes me want to release even before all these optimization things happen and are polished, and I might well decide to go with that.
Maybe I only add the closure in-place stuff and the polish the loop SSA stuff, and then call it a release. It already will solve a lot of performance issues that exist right now, while staging the ground for more.
Standalone work is also improving. Using pyi files got more apt, and a few things were added, all of which make sense to be used by people.
But I also have a backlog of issues there however. I will schedule one sprint for those I guess, where I focus on these. I am neglecting those somewhat recently.
For the static code, I now noticed that it's compiled for each target name, due to the build directory being part of the object file for debug. For gcc 8 there is an option to allow pointing at the original static C file location, and then ccache is more effective, because object files will be the same.
That's actually pretty bad, as most of my machines are on gcc-6 and makes me think that libnuitka.a is really more of an requirement than ever. I might take some time to get this sorted out.
So Nuitka supports the no_warnings Python flag, and for a long time I have been annoyed at how it was not working for Python3 in some cases. The code was manually settign filters, but these would get overridden by CPython test suites testing warnings. And the code said that there is no CPython C-API to control it, which is just plain wrong.
So I changed that and it became possible to remove lots of ignore_stderr annotations in CPython test suites, and more importantly, I can stop adding them for when running older/newer CPython version with a suite.
I continue to be very active there.
And lets not forget, having followers make me happy. So do re-tweets.
Adding Twitter more prominently to the web site is something that is also going to happen.
If you are interested, I am tagging issues help wanted and there is a bunch, and very likely at least one you can help with.
Nuitka definitely needs more people to work on it.