How to get to 100% precise tracing GC?

[Ola Fosheim Grøstad]

On Sunday, 15 November 2020 at 22:29:27 UTC, Rainer Schuetze 
wrote:
> Indeed, that could work, although the "stack map" seems a bit 
> too simple. It prevents some common optimizations, like having 
> pointers only in registers, or reusing the same stack area in 
> different scopes inside a function.

My understanding is that LLVM use the term "stack map" for two 
different solutions. The old one is based on registering gc-roots 
"manually", but the newer one that is for patch-points (points 
where you can inject code at runtime) including register 
information. It does inhibit some optimizations as that point is 
assumed to write state, but the docs says that it can be marked 
as "read only", which should work out, I think? But it isn't 
trivial to make use of it and it will have some performance 
impact compared to manual allocations.

My personal opinion is that if one reduce collection to a single 
thread and also reduce the number of pointers, then one can get 
to a good place. One way to reduce scanned pointers would be to 
allow developers to also mark pointers as non-owning, and there 
are more tricks available...