[Migrated] Language feature: barriers

まず、issue にある GroupSharedArray、GroupSharedWriter、GroupSharedReader、および barrier の例を確認します。記載されている DerefMut/IndexMut の制限と、書き込みまたは読み取りの繰り返しを許可するケースを調査します。完了条件は、現在の不完全なプロトタイプではなく、barrier を強制するための明確で健全な設計が決定されていることです。

Issue automatically imported from old repo: https://github.com/EmbarkStudios/rust-gpu/issues/8
Old labels: t: design
Originally creatd by Jasper-Bekkers on 2020-08-18T17:33:10Z

We want to have the same types of safety on the GPU as we do on the CPU. One of those areas that make it easy to introduce race conditions is with memory barriers. HLSL and GLSL require the programmer to explicitly put them in the right locations without any formal verifiction.

This is an attempt at solving that within the Rust type system (attempt still has racy bugs and is largely incomplete).

use core::ops::Deref;
use core::ops::DerefMut;
use core::ops::Index;
use core::ops::IndexMut;

struct MyFoo {
    test: GroupSharedArray<u32>,
}

struct GroupSharedArray<T>
where
    T: Default + Copy,
{
    data: Box<[T]>,
    size: usize,
}

impl<T: Default + Copy> GroupSharedArray<T> {
    fn new() -> Self {
        Self {
            size: 100,
            data: Box::new([T::default(); 100]),
        }
    }
}

struct GroupSharedWriter<T> {
    data: T,
}

impl<T> Deref for GroupSharedWriter<T> {
    type Target = T;

    fn deref(&self) -> &Self::Target {
        &self.data
    }
}

impl<T> DerefMut for GroupSharedWriter<T> {
    fn deref_mut(&mut self) -> &mut Self::Target {
        &mut self.data
    }
}

impl<T> GroupSharedWriter<T> {
    fn new(data: T) -> Self {
        Self { data }
    }
    fn barrier(self) -> GroupSharedReader<T> {
        GroupSharedReader { data: self.data }
    }
}

struct GroupSharedReader<T> {
    data: T,
}

impl<T> Deref for GroupSharedReader<T> {
    type Target = T;

    fn deref(&self) -> &Self::Target {
        &self.data
    }
}

impl<T> GroupSharedReader<T> {
    fn barrier(self) -> GroupSharedWriter<T> {
        GroupSharedWriter { data: self.data }
    }
}

impl<T: Default + Copy> Index<Uniform<u32>> for GroupSharedArray<T> {
    type Output = T;

    fn index(&self, index: Uniform<u32>) -> &Self::Output {
        &self.data[index.data as usize]
    }
}

impl<T: Default + Copy> IndexMut<Uniform<u32>> for GroupSharedArray<T> {
    fn index_mut(&mut self, index: Uniform<u32>) -> &mut Self::Output {
        &mut self.data[index.data as usize]
    }
}

struct Uniform<T> {
    data: T,
}

impl<T> Uniform<T> {
    fn new(data: T) -> Self {
        Self { data }
    }
}

fn thread_idx() -> Uniform<u32> {
    Uniform::new(0)
}

fn main() {
    let mut data = GroupSharedWriter::new(MyFoo {
        test: GroupSharedArray::new(),
    }); // new does barrier

    data.test[thread_idx()] = thread_idx().data;
    data.test[thread_idx()] = 666; // should be illegal

    let value = data.test[thread_idx()]; // should be illegal

    //data.test[0] = thread_idx().data; // what happens in this case? it's unsafe/unsound. race cnd over contents of `test`

    let other_data = data.barrier(); // `barrier` return GroupSharedReader
    
    // some race-y cases handled correctly:
    // data.test[thread_idx()] = 666; // is correctly marked illegal
    // other_data.test[thread_idx()] = 666; // is correctly marked illegal

    let test = other_data.data.test[thread_idx()];
}

// tl;dr
// - use move semantics to enforce barriers betwen reading and writing
// - broken right now because DerefMut and IndexMut require Deref and Index to be implemented
// - broken right now because we can write to the same location multiple times (data race)