printf scientific notation performance regression

[simonster]

 function f()
    io = IOBuffer()
    for i = 1:100000
        @printf io "%e" 28.0
    end
end

On Julia 0.3:

julia> @time f()
elapsed time: 0.078705771 seconds (4195572 bytes allocated)

On Julia 0.4:

julia> @time f()
elapsed time: 2.91291863 seconds (290595572 bytes allocated, 45.76% gc time)

@profile suggests the time is spent in grisu's GMP codepath, so I'm guessing #7291 is the culprit. cc @quinnj

[quinnj]

Hmmmm.......I'll try to look into it. A big difference here is that the double conversion library was rolling its own minimal BigInt, while grisu.jl utilized GMP. I believe some of their code makes use of reusing allocated BigInts, which we probably aren't and would explain all the extra allocation.

[simonster]

It looks like the slow path is used for most (all?) integers, e.g.:

julia> function f(x)
          io = IOBuffer()
          for i = 1:100000
              @printf io "%e" x
          end
       end;

julia> @time f(1.)
elapsed time: 2.313745457 seconds (264195572 bytes allocated, 52.41% gc time)

julia> @time f(1.+eps())
elapsed time: 0.044512516 seconds (17819740 bytes allocated)

I haven't looked into the code to see how @printf uses grisu, but it seems like it should be possible to print integers in scientific notation without resorting to GMP.

[JeffBezanson]

Bump. This is pretty serious.

[quinnj]

Status update: I've gotten a 5x improvement so far after a first pass. I'm still seeing some gc action in profiling so I'm going to keep digging.

[JeffBezanson]

Awesome, thanks.

[JeffBezanson]

@quinnj would it make sense to commit your 5x improvement? As it stands, this regression is probably bad enough to block 0.4.

[quinnj]

I want to run the full grisu testing suite on my changes first; I'll make a PR tomorrow after it runs.

[quinnj]

PR coming later tonight! (once I fix a few tests)