Number to/from string API

[huonw]

This is an umbrella issue following on from/working with #4819. The number <-> string conversion needs work and thought.

For example, the current to_str version allocates, and the current from_str doesn't work with the compiler optimisations very well: none of the (for example) exponent parsing code is removed when the function is specialised for ints.

Another important thing to consider is how non-primitive numbers (e.g. bigint/rational) are to be supported/helped (including not at all, other than the normal to/from_str traits)

[huonw]

For comparison, glibc's scanf: floating point, and integers (these are entirely separate).

And printf: general number printing and specifically floating point (it actually uses (stack-allocated) GMP bignums). Notably the general number printing is one huge (500 line) macro that gets called twice.

[Aatch]

Also, I have been working on a generic parser for printf-like format strings. The current grammar is this:

 Placeholder := Indicator Position? Flag* Width? Precision? Specifier
 Position := '{' '-'? [0-9]+ '}'
 Width := [1-9] | '[' ('-'? [0-9]+ | '*') ']'
 Precision := '.' ([1-9] | '[' ('-'? [0-9]+ | '*' ) ']')
 Flag := <From supplied set>
 Specifier := <From supplied set>

So thinks are specified roughly like: %{1} [2].9f, which is where the values is in argument 1 and the width is retrieved from argument 2.

[huonw]

A possibility for code sharing would be having a munch_number<Num: Integer>(&str) -> (Option<Num>, uint) function that parses a number at the start of the string, so from_str would do this then check that the whole string has been parsed.

This would allow float parsing to use the int parsing code to read subsections.

(This function might be actually useful for any from_str value.)

[Kimundi]

I started on writing specialised replacements for ints, uints and floats. Currently the new (u)int versions use a stack allocated vector and a callback function. For the float versions, I started looking into papers and the links here for an optimal algorithm. I will also look into making the number munching thing work.

[huonw]

Two papers about printing floats I saw on Reddit recently:

• "How to Print Floating-Point Numbers Accurately" Steele & White 1990
• "Printing Floating-Point Numbers Quickly and Accurately with Integers" Loitsch 2004