WIP: start trying to fix on julia nightly

[tkelman]

Still getting a cannot resize array with shared data, not sure how to best deal with that. Commenting out the readbytes! method that was added in JuliaLang/julia#14667 seems to work, but may reintroduce the perf issue in base. cc @vtjnash

One ugly way to fix this is to change this line to
n = invoke(readbytes!, (IO, AbstractArray{UInt8}, Any), jl_ctx, jl_msg, sz)

Or JuliaLang/julia#14699 may change the behavior here. Regardless of the fix, I could use some help writing tests in base that are representative of the way this package needs to use IO functions. First for calling readbytes! with a non-Int input nb, where widening wait_readnb(x::LibuvStream, nb::Int) and wait_readnb(s::BufferStream, nb::Int) to accept Integer should be safe as far as I can tell. I can figure out BufferStream (bufstr = BufferStream(); data = UInt8[0,0,0,0,0]; sz = UInt(5); write(bufstr, data+1); readbytes!(bufstr, data, sz)) but not sure how to do any of the other LibuvStream subtypes. From what I can tell, jl_ctx when this package executes its tests is always a TCPSocket here.

[vtjnash]

i think the issue is that i assumed resize!(b, nr) would be a no-op if length(b) == nr, but that's not true in this instance. it looks like #14699 should fix that

[tkelman]

Yes it looks that way to me as well. I think 14699 is likely too big for backporting, so what's the right approach w.r.t. backporting the original change? Does it do any good if the introduced readbytes! method is left out?

[vtjnash]

yeah, that's still most of the improvement. could also add a test before the call to resize! and skip it if it should be a no-op

[tkelman]

I think this is going to get fixed on nightly, making this unnecessary. Still need more tests in base that are representative of what this package needs.

@vtjnash can you clarify, by "that's still most of the improvement" do you mean the parts of the change other than the new method?

[vtjnash]

yes. it's still a huge speedup to add just the fastpath optimizations for the wait_nb and eof methods