patch to make blip work in a stable way

I've changed the license to "BSD style: $(LICENSE)" as for lack of a better idea.

need to resolve outstanding questions about the replacement Atomic module. Specifically with regard to how many calls are involved. Also, should we use 'naked' on the asm where that works?

As I said I don't see where the new atomic modules makes many calls, could you be more specific?

Places where I see more than 0 calls: high level functions: get/setFlag, but there the cost of the barrier probably shadows everything. One could try to put the barrier in the asm, but it is much less flexible, and I doubt you would gain something (a good compiler should inline the barrier op).

what does disturb you exactly? Is it the boolean returning atomicCASB? that should be inlined, I switched to an atomicCas that returns the original value by default because that is more efficient and safer to do loops like atomicOp (as h3r3tic made me realize), so it is a better primitive.

atomicOp/fallbacks I think that is as efficient as it can be without writing everything in assembler.

The basic functions are templates, on dmd I use asm blocks, as we discussed on the IRC the only thing that I see could be sone is using naked asm, but I don't know if that really improves the timings, and in any case I suspect that it will disallow inlining even more than asm blocks. (the prologue/epiloque cannot be removed).

have you looked at the disassembled code yet? As I'd said before, that's where the prior issues had shown up

I join the asm for

module tstAtomic;
import blip.sync.Atomic;
private extern(C) int printf(char*,...);
void main(){
    size_t i;
    
    atomicAdd(i,cast(size_t)1);
    if (atomicCAS(i,cast(size_t)0,cast(size_t)1)==1){
        printf("set back!\n");
    }
    flagSet(i,cast(size_t)4);
    flagGet(i);
}

to me it looks exactly as I would have guessed from the source. Yes flagGet/flagSet don't inline the call to the barrier (but even it it looks simple at asm level a barrier is slow, so inlining would not gain a lot. The rest seems to me pretty standard.

Were you looking at other functions?

the assempler for the sample module on osx with dmd -O -inline -release

ehm assmbler :)

by the way compilation time of the Atomic module for me was at most 0.017s real time also on my old 1.82 intel core Duo laptop with 5400 rpm HD. (as expected, as it is mostly templates).

I have never seen that "it was the module that took longest to compile".

(In [5511]) fixes #1959 :: patch to make blip work in a stable way; big thanks to fawzi