Better precompilability #287
Conversation
After #259, the result of `filename(q)` is not inferrable. This uses "the kernel trick" to provide a single point where the result must be inferred, and an `invokelatest` to prevent the abstract signature from being inferred. It then precompiles the `::String` variant. Consequently, with respect to latency this should get us back to where we were before #259.
For reasons I don't fully understand, either the string interpolation or the error messes up precompilation of `applicable_loaders`. This splits it out into a separate function and calls it via `invokelatest` to prevent it from blocking precompilation.
I haven't seen this warning come up in a long time, so I don't think this code is active. It adds significant latency due to the time needed to infer the Pkg calls, so let's ditch it. xref #226.
|
The first commit seems fine to me, though I don't fully understand why |
It's because |
|
Oh, I see. It was specifically this change https://github.com/JuliaIO/FileIO.jl/pull/259/files#diff-525588e68b2421901be164965272940effa093de733a3fd36c4b9a4344b8c20cL29 that's causing the problem. The term |
|
Parametrization might be undesirable from a latency perspective: any method accepting a |
These allow more concrete inference and eliminate some weird sources of latency. Very briefly, the first commit compensates for #259 (CC @rofinn, @omus), the last implements the conclusion of discussion in #226 (CC @johnnychen94). All part of a grand scheme to reduce latency (e.g., JuliaGraphics/Gtk.jl#551 (comment)).