Regression in string readuntil benchmark

[gbaraldi]

run(BaseBenchmarks.SUITE[["string", "readuntil", "target length 2"]]) regressed in #48273 @stevengj it gained some allocations and became 2x slower.

[stevengj]

This is apparently benchmarking:

buffer = IOBuffer("A" ^ 50000)

g = addgroup!(SUITE, "readuntil")
for len in (1, 2, 1000, 50000)
    g["target length $len"] = @benchmarkable readuntil(seekstart($buffer), $("A" ^ len))
end

Reading 2-byte strings from an IOBuffer probably got slower because of this diff, which traded off worse performance for reading extremely short lines for better performance reading lines with 20 characters or more. (See the comment "A single loop is 2x faster for nout=5.") Actually that portion of the PR sholdn't be relevant here because that is only for delim::UInt8, whereas this function uses a string delimiter, which is a completely different part of the codebase.

For string delimiters, one change in that PR is that it changed the initial allocated size to 70 bytes, to match the initial allocation in other parts of the code: https://github.com/JuliaLang/julia/pull/48273/files#diff-74f71402c994a78b21ded0ac040485d12bfc27e9ffaf538fa0f71f6d284e2991R1038 — maybe that is slowing us down for reading 2-byte strings? Someone could try changing that line back to StringVector(0) to see if it helps. (Of course, that might slow things down for reading longer strings.)

(In general, my feeling is that if you care about performance for reading lots of very short strings, you are better off using the new functionality with something like StringViews.jl to read the strings in-place into a re-used buffer.)