Optimise English implementation

[jakewilliami]

Coming back to this, there are changes I'd like to make regarding optimisation. I started these a while ago, but never pushed anything. Figured I should open a PR to track these changes.

[jakewilliami]

I'm worried that these optimisations aren't actually optimising anything! I have had success in the past with using IOBuffer instead of string concatenation, as this can be slow. However, it appears that these IOBuffer changes have slowed it down and caused it to allocate more:

julia> @btime spelled_out($123456789, lang = $:en_UK)  # master at dc74774
  8.797 μs (204 allocations: 4.39 KiB)
"one hundred and twenty-three million, four hundred and fifty-six thousand, seven hundred and eighty-nine"

julia> @btime spelled_out($123456789, lang = $:en_UK)  # optimisation at 22e2773
  9.906 μs (220 allocations: 5.83 KiB)
"one hundred and twenty-three million, four hundred and fifty-six thousand, seven hundred and eighty-nine"

[jakewilliami]

Since last commit, I figured that creating IOBuffers within each function would indeed be slow, as we are often recursing or calling functions multiple times in a single spelled_out call. Hence, I thought that the best way around this would be to only work with one IOBuffer at the top level, and have modifying functions on the buffer. So small_convert_en becomes _small_convert_en!, and large_convert_en becomes _large_convert_en!, and their first argument—an IOBuffer—is only constructed once and is modified by these functions.

The benchmarking results are as follows. Indeed, it uses fewer allocations. However, I wonder where it is spending the most time? It is not as fast as the others (in fact, it might even be slower than all of the other versions).

julia> @btime spelled_out(123456789, lang = :en_GB)  # optimisation at 3175b96
  10.275 μs (190 allocations: 3.92 KiB)
"one hundred and twenty-three million, four hundred and fifty-six thousand, seven hundred and eighty-nine"

[jakewilliami]

For some reason, on a whim I changed write(io, ...) to print(io, ...), and it is considerably faster (with the same allocations). I'm feeling good about this branch again!

julia> @btime spelled_out(123456789, lang = :en_GB)  # optimisation at a9403ed
  10.119 μs (190 allocations: 3.92 KiB)
"one hundred and twenty-three million, four hundred and fifty-six thousand, seven hundred and eighty-nine"

julia> @btime spelled_out(123456789, lang = :en_GB)  # optimisation at 00f458a
  9.296 μs (190 allocations: 3.92 KiB)
"one hundred and twenty-three million, four hundred and fifty-six thousand, seven hundred and eighty-nine"

[jakewilliami]

I don't know why, but I was being awfully liberal with using BigInts in this code. Reducing when I use these has greatly optimised this

julia> @btime spelled_out(123456789, lang = :en_GB)  # optimisation at 7fb5af2
  6.499 μs (120 allocations: 2.70 KiB)
"one hundred and twenty-three million, four hundred and fifty-six thousand, seven hundred and eighty-nine"

[jakewilliami]

I think this is probably as optimised as I'm going to get it. The attached jlprof file for profiling information (see ProfileView.jl; you will have to rename the file as GitHub wouldn't allow me to attach the jlprof file directly). At the top of the call graph are predominantly implementations of things that I cannot change nor optimise.

profview.jlprof.txt