Output performance for numbers

[stevebleazard]

When converting numbers to strings string_of_float has a significant performance impact when the number is an INT. Detecting INTs and using string_of_int significantly improves performance. Thus, using

let f2s f = 
  (* Check value is in the JSON integer range, maybe not necessary *)
  if f > 9007199254740991. || f < -9007199254740991. then string_of_float f 
  else
    let i = int_of_float f in
    let f1 = float_of_int i in
    if f1 = f then string_of_int i 
    else string_of_float f 

Has a cost of around 6.5% for floats but results in around 4x performance improvement for INTs. Here are comparisons of straight string_of_float and f2s for 1,000,000 conversions:

string_of_float

1,552,011,783      cycles
2,958,207,469      instructions

f2s

  363,014,703      cycles                   
  945,099,124      instructions

Is it worth integrating this into the encoder?

[mjambon]

This is interesting. Why did you close the issue?

I was about to suggest to push for the change in the standard library at https://caml.inria.fr/mantis.

[stevebleazard]

Confusion on my part, I have re-opened

[mjambon]

It's a bit more complicated on 32-bit platforms (31-bit ints) because the conversion from float to int works on a restricted range (float min_int to float max_int). Here are options I can think of:

• use a fixed range regardless of the platform, defined as the range of 31-bit ints;
• detect the platform (e.g. using Sys.word_size) and run one function or the other accordingly.

Code for the second option could be like this:

let fast_string_of_float =
  match Sys.word_size with
  | 64 -> fast_string_of_float_64
  | 32 -> fast_string_of_float_32
  | _ -> string_of_float

There's also the question of whether this optimization benefits everyone. In some applications, integers may be extremely rare, in which case this "optimization" would slow them down by 6.5% (based on your report).

Atdgen could support a global (command-line) or type-specific option that turns this on or off.