change IOBuffer to use Memory internally

[vtjnash]

An Array is often still allocated on output, but this gives the compiler a chance to potentially elide that in certain cases.

For measurement, it seems about 10% faster as a string builder:

julia> @btime repr("hello\nworld"^10);
  1.096 μs (10 allocations: 640 bytes) # master
  973.000 ns (9 allocations: 608 bytes) # PR
  994.000 ns (8 allocations: 576 bytes) # also PR, after Revise-ing Base.wrap

[oscardssmith]

This looks good to me, but it is a lot of tricky code.

[vtjnash]

Yeah, looks like we trigger a CSV bug added in JuliaData/CSV.jl@953636a, which somehow gets involved with the Pkg tests (Pkg maybe is setting up a bad environment, as I don't think it looks like CSV should be installed there). As well as a problem exposed with the REPL precompile test coverage which should be an easy fix.

[IanButterworth]

CSV was added pinned as a test for a specific issue JuliaLang/Pkg.jl#3361

[vtjnash]

@nanosoldier runbenchmarks(!"scalar", vs=":master")

[oscardssmith]

the CSV issue should be fixed by JuliaData/CSV.jl#1121

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[vtjnash]

several good runtime improvements, no runtime regressions

["inference", "optimization", "Base.init_stdio(::Ptr{Cvoid})"]	0.96 (5%)	0.97 (1%) ✅
["io", "array_limit", ("display", "Matrix{Float64}(10000, 10000)")]	0.99 (5%)	0.97 (1%) ✅
["io", "array_limit", ("display", "Matrix{Float64}(100000000, 1)")]	1.00 (5%)	0.99 (1%) ✅
["io", "array_limit", ("display", "Vector{Float64}(100000000,)")]	1.00 (5%)	0.99 (1%) ✅
["dates", "string", "Date"]	0.90 (5%) ✅	0.94 (1%) ✅
["dates", "string", "DateTime"]	0.91 (5%) ✅	0.92 (1%) ✅
["io", "read", "readstring"]	0.70 (5%) ✅	0.50 (1%) ✅
["micro", "parseint"]	0.88 (5%) ✅	0.66 (1%) ✅
["problem", "imdb", "centrality"]	1.00 (5%)	0.99 (1%) ✅

[vtjnash]

Still encountering some sort of error in the REPL package, but seems possible to work around this

      From worker 14:	┌ Info: REPL: trace compile output:
2024-02-10 04:00:40 UTC	      From worker 14:	└ precompile(Tuple{typeof(Base.write), Base.GenericIOBuffer{GenericMemory{:not_atomic, UInt8, Core.AddrSpace{Core}(0x00)}}, String})

[IanButterworth]

Fixes would be adding that precompile statement or relax the test. Currently allows no compilation during interactive startup