jagotu / bacil Goto Github PK

I know this project is nowhere near the point of Java <-> C# interop yet, but I can't get it off my mind.

What will calling C# code from Java code look like? I'm not familiar with how namespaces, statics, etc. work in .NET libraries. I want to assume it can't be similar to JS, because unlike in JS, in C# not every identifier is a value, but I haven't yet dived as deep into Truffle as I would like.

Hi @jagotu !
Your project looks amazing :)
Any update??
What kinds of programs can it currently run?
I know it's a bit early but your project would get public attention if you published interesting benchmarks.

The instructions mention HelloWorld.dll, but it is not included in the build ?

While checking the reason for lower performance of some benchmarks, I noticed that if loops remain after TruffleTier (which they seem to do as TruffleTier just peels the bytecode loop), no later tiers unroll them even though form the graphs should fullfill rules for the LoopFullUnrollPhase that runs in all three of these tiers. This prevents other optimizations and constant folding that could be achieved by unrolling the loop.

For example, this procedure:

static int simpletest()
{
    int result = 0;
    for (int i = 0; i < 5; i++)
    {
        result += i;
    }
    return result;
}

Copmiles to this graph after TruffleTier:

And the LoopNode survives even after low tier:

I've tried to debug the LoopFullUnrollPhase and it seems the detection of counted loops happens only once, when the loop is still too complex to be recognized as counted. It could very well be a bug/imperfection in truffle/graal itself, but I don't currently have the time to investigate.

Fixing this and making sure that the loops are unrolled should provide a significant boost to the slowest of benchmarks, like nbody, which loops over an array of 6 plantes and none of the virtual calls can be constantized due to the lack of unrolling.