IO read performance

[tkoolen]

I opened JuliaCI/BaseBenchmarks.jl#223 to track performance of reading a Float64 from an IOBuffer. This currently allocates.

I thought that maybe this would be resolved in 0.7 (elimination of the allocation of the Ref here), but I guess that's being explicitly prevented by the @noinline here.

BufferedStreams.jl has this implementation, which doesn't allocate. Is something like this acceptable for Base as well? I don't think there are any issues with GC roots or alignment in this implementation, are there?

[tkoolen]

Also, should the @noinline (comment: mark noinline to ensure ref is gc-rooted somewhere (by the caller)) just be replaced with a GC.@preserve now?

[JeffBezanson]

Yes, That noinline should be replaced with a gc-preserve now.

[vtjnash]

No, that’s there to guarantee the allocation. We can’t fix that (or enable general stack allocation) due to copy-stacks.

[JeffBezanson]

Oh, right. Well that will be fixed pretty soon too.

[tkoolen]

Is the BufferedStreams implementation OK though? I understand this should all get better soon, but I kind of need a non-allocating solution right now.

[tkoolen]

FYI, I just put together https://github.com/tkoolen/FastIOBuffers.jl.

[damiendr]

@tkoolen came across the same problem, your package works well for me so far. It would be nice to have this level of performance by default.