Disambiguate get_world_counter and get_inference_world

[vchuravy]

@gbaraldi and I were looking through the code today and we were surprised by the usages of get_world_counter.
Inference is primarily executed in a frozen world-age so get_world_counter is often fixed to be jl_typeinf_world,
in most cases it seems we should be using get_world_counter(interp).

Add brief docs to Base.get_world_counter and add Base.tls_wold_age. Furthermore to disambiguate get_world_counter from the inference world rename the abstract interpreter accessor to get_inference_world

[vchuravy]

I just noticed that get_world_counter during compilation will return max world.

julia/src/gf.c

Line 32 in 1442303

return ~(size_t)0;

that makes the logic even weirder in places :)

[vtjnash]

Why does the interp have a get_world_counter function? This isn't supposed to be a fixed value, but rather the maximum observable value of jl_world_counter at the start of inference (preventing a data race on that value from leaking into making incorrect optimization results)

[vchuravy]

Why does the interp have a get_world_counter function? This isn't supposed to be a fixed value, but rather the maximum observable value of jl_world_counter at the start of inference

To enable that exactly? get_world_counter is frozen during inference so we enter inference with the last observed value and need to store it somewhere? So that's get_world_counter(interp) is for IIUC

[vchuravy]

@vtjnash explained to me yesterday that get_world_counter is like a sequence lock and so the uses in Compiler are to check if someone invalidated the global world.

I confused it with current_task->world which gives you the world you currently execute within.