WIP: RFC: cache parse results while loading files

[vtjnash]

This is my attempt at partial caching of the results of loading .jl files. It is functional, but since it does not detect file changes yet, it is a little unsafe right now.

On the julia machine, it reduced the load time of Tk from 4.3 seconds to 2.9 seconds. On my Mac, it reduced the time from ~10 s to ~6 s.

TODO items:

• stop printing "cache file no found" when loading files
• address compiler warning regarding sprintf declaration
• store timestamp in the cache file (don't use old cache)
• get Jeff to merge pull requests
• improve efficiency of GC roots for cache file operations (teach the GC about a ptrhash, use weakrefs)

[StefanKarpinski]

Would full pre-compilation be better?

[vtjnash]

Would full pre-compilation be better?

I'm not sure how that is relevant.

[lindahua]

Great to see the efforts are underway.

[timholy]

From some other discussion I can't find right now, probably the main bottleneck here is macros: if package A uses macros from package B, and package B updates, then you also want to re-generate the cached versions of package A.

[timholy]

Ah, here's the link I was missing: #7280 (comment).

[vtjnash]

replace by https://github.com/vtjnash/Speed.jl

[timholy]

Was it premature to close this, or does this not offer any advantages over Speed.jl?

[vtjnash]

Speed.jl was faster (primarily because I could emit better caches) and more advanced (again, because I could emit better caches, and also because it is much easier to test preconditions in Julia than C)

[timholy]

Did you see the issues I opened?

[vtjnash]

no, thanks for pointing that out. I forgot I turned off notifications so I would stop getting messages from the JuliaLang/ hosted repos, and needed to turn them on manually for the vtjnash/ hosted repos