Most Effective way of developing a new GC for D

[Basile B.]

On Sunday, 30 September 2018 at 09:40:10 UTC, Per Nordlöw wrote:
> I'm gonna play around with creating a GC that alleviates some 
> of the uses described in
>
> https://olshansky.me/gc/runtime/dlang/2017/06/14/inside-d-gc.html
>
> What's the most effective way of incrementally developing a new 
> pluggable GC for druntime with regards to prevention of 
> really-hard-to-find-bugs?
>
> I'm aware of the run-time flag added in 2.072:
>
> https://dlang.org/changelog/2.072.0.html#gc-runtimeswitch-added
>
> Is it inevitable to rebuild druntime everytime I make an update 
> to a new GC?

No.

> Is so what's the preferred way of selecting a modified druntime 
> in a standard installation of dmd on a Linux system (Ubuntu 
> 18.04 in my case). Will Digger makes things easier?

The immediate idea (and maybe naive too) is to link once against 
a new GC that follows the interface but has assignable handlers. 
Once DMD build, you can just do

     dmd -unittest temp_gc_handlers.d && ./temp_gc_handlers

At each iteration. temp_gc_handlers.d would have a static module 
ctor that assigns its content to the GC handler.

Once good you put the handler code in the real GC. I think that 
it would be possible that a few runtime stuff leak until the 
devel be almost complete but that doesn't matter if stuff are 
checked, let's say, from a checkpoint that says "from here this 
amount of leak is normal, it should disappear automatically in 
the real GC".

But again, this method is very hypothetical.