Is there an FLS implementation floating around?
If not, it's probably something that should be considered for std.thread.

On Friday, 26 July 2013 at 04:30:18 UTC, Manu wrote:
> Is there an FLS implementation floating around?
> If not, it's probably something that should be considered for std.thread.

Maybe something like this:

----------------------------------

private struct FLSEntry
{
    void** ptr;
    size_t start;
    immutable void[] initData;
}

private __gshared FLSEntry[] flsEntries;
private __gshared size_t flsSize;

void registerFLS(void** ptr, immutable void[] initData)
{
    enum ALIGN = size_t.sizeof;
    auto size = (initData.length + ALIGN-1) % ALIGN;
    flsEntries ~= FLSEntry(ptr, flsSize, initData);
    flsSize += size;
}

// called by core.thread on fiber creation
void initFLS(Fiber fiber)
{
    fiber.fls = new void[flsSize];
    void* p = fiber.fls.ptr;
    foreach (entry; flsEntries)
        with (entry)
            p[start..start+initData.length] = initData[];
}

// called by core.thread when switching to a fiber
void switchFLS(Fiber to)
{
    void* p = to.fls.ptr;
    foreach (entry; flsEntries)
        *entry.ptr = p + entry.start;
}

struct FLS(T)
{
    shared static this()
    {
        //registerFLS(&data, (&T.init)[0..1]);
        static immutable T init;
        registerFLS(cast(void**)&data, (&init)[0..1]);
    }

    static T* data;

    alias data this;
}

// ***************************************

struct MyData
{
    int i;
}

FLS!MyData fiberData;

----------------------------------

A linker segment would be better, but not sure how feasible that would be.

On Friday, 26 July 2013 at 04:30:18 UTC, Manu wrote:
> Is there an FLS implementation floating around?
> If not, it's probably something that should be considered for std.thread.

I'd also love to have one ! (in fact I do think it should be the default).

On 07/26/2013 08:06 AM, deadalnix wrote:
> On Friday, 26 July 2013 at 04:30:18 UTC, Manu wrote:
>> Is there an FLS implementation floating around?
>> If not, it's probably something that should be considered for std.thread.
>
> I'd also love to have one ! (in fact I do think it should be the default).

What do you mean by that, FLS instead of TLS and everything runs as Fiber?

On Friday, 26 July 2013 at 04:30:18 UTC, Manu wrote:
> Is there an FLS implementation floating around?
> If not, it's probably something that should be considered for std.thread.

Nice to see this asked. For a simulation library I was developing I was thinking that fiber support had to be done by asking the library user to swap a struct every time a fiber context-switched. If that could be automated then that would be great!

On Monday, 29 July 2013 at 15:25:49 UTC, Martin Nowak wrote:
> On 07/26/2013 08:06 AM, deadalnix wrote:
>> On Friday, 26 July 2013 at 04:30:18 UTC, Manu wrote:
>>> Is there an FLS implementation floating around?
>>> If not, it's probably something that should be considered for std.thread.
>>
>> I'd also love to have one ! (in fact I do think it should be the default).
>
> What do you mean by that, FLS instead of TLS and everything runs as Fiber?

Yes, with a scheduler int the runtime.

On Tuesday, 30 July 2013 at 01:45:45 UTC, deadalnix wrote:
> On Monday, 29 July 2013 at 15:25:49 UTC, Martin Nowak wrote:
>> What do you mean by that, FLS instead of TLS and everything runs as Fiber?
>
> Yes, with a scheduler int the runtime.

I don't think FLS can be as fast as TLS because the latter benefits from many optimizations due to linker and OS support.

On Friday, 26 July 2013 at 04:30:18 UTC, Manu wrote:
> Is there an FLS implementation floating around?
> If not, it's probably something that should be considered for std.thread.

- There is WorkerLocalStorage in std.parallelism Tasks.
- Apparently you can easily capture variables in the context of the delegate.
- If you use thread affinity when executing Fibers you can access TLS.