With the release of vvvv gamma 2019.2 we introduced a new backend compiling patches in real time using Roslyn. This blog post is primarily intended for a technical audience, if you're solely interested what new features it brings to the table have a look at the before mentioned blog post.
In the past VL (the language behind vvvv gamma) compiled in-memory directly to CIL using CCI. With the recent changes in the .NET ecosystem and CCI being superseded by Roslyn it became more and more apparent that at some point we'd also have to make the switch to keep up with the latest developments happening at Microsoft.
What finally pushed us into making the switch was two-folded:
Until this year in march @sebl came to the rescue by randomly dropping us a link in the chat pointing to a neat little "trick" which suddenly made it possible to translate our adaptive nodes feature directly.
After initial tests in March and April and having the patched tooltip feature still pending for the final release, we decided to let myself jump into the rabbit hole which I've finally crawled out of after more than half a year ;)
Let's go back to the example of the LERP node and let's further try to write it down in C#:
Looks neat but sadly won't work, C# will tell us that the operators +, - and * and the constant 1 are not available on T.
The trick to make it work is to outsource those operators to a so-called "witness" which in turn will provide the implementation when the LERP gets instantiated with say a vector. So let's see how the actual needed C# code is gonna look like:
and when applying it with say float we need to define a witness implementing the needed interfaces
which finally allows us to call
Fancy, no? The beauty is that when the JIT compiler hits such a code path it will be smart enough to inline all calls so in the end for the CPU the code to execute is the same as the initial naive attempt. But don't worry, this is all happening behind the scenes. In the patching world, it is as simple as it was before to patch generic numeric algorithms.
So now that we're able to translate patches directly to C# what are the implications apart from being able to export an application?
Well for us as developers it will be much easier to bring in new language features, because the code we generate will be be checked by the C# compiler and more important, we can fully debug the generated code with Visual Studio. That by the way is not only restricted to us, anyone can now attach a debugger to vvvv (or the exported app) and debug the patches.
The generated C# code will make full use of .NET generics. So when building a generic patch the generated class will also be generic in the pure .NET world. As an example let's consider the Changed node, while in the CCI based backend a Changed class was emitted for each instantiation (Changed_Float, Changed_Vector, etc.), the new Roslyn based backend will only emit one Changed<T> class and it is left to the JIT compiler of .NET to create the different versions of the needed target code. This should lead to much less code the CPU needs to execute as the JIT compiler is much smarter on when to generate new code and when not.
But what's even more important is the fact that it opens up the world of compiling VL patches as pure .NET libraries. So we can finally pre-compile our libraries (like VL.CoreLib, VL.Skia, etc.) which in turn reduces the general overhead and leads to much quicker loading times and less memory usage. As an example loading the Elementa "All At Once" help patch takes ~15 seconds the first time (compared to ~33 seconds in the old backend) and thanks to caching to disk only ~8 seconds when opening at a later time.
Apart from better loading times, it also gives the patcher the ability to instantiate any VL patch during runtime. In the previous backend, one had to use a hack and put all possible instantiations into a non-executing if-region. This is not necessary anymore as all the patches get compiled. However, I should mention here that this is only true for non-generic patches. Generic patches usually require a witness which is not so straight forward to provide.
Sadly the new backend also required some major internal changes in the frontend so it wasn't possible to guarantee existing patches would work the same way as they did before. Here follows a list of potential breaking changes:
anonymous user login