[PATCH] 4511638: Double.toString(double) sometimes produces incorrect results

[raffaello.giulietti at gmail.com]

Hi Andrew,

On 2018-09-27 15:28, Andrew Haley wrote:
> On 09/27/2018 01:23 PM, Raffaello Giulietti wrote:
> 
>> In principle I agree with you.
>>
>> However, in this case the maths underlying the algorithm to select the 
>> decimal are too involved to explain in comment form. I'm in the course 
>> of preparing a paper that explains the idea and the details. Then it 
>> should be easier to make sense out of the code.
> 
> In which case, the code must contain pointers to the relevant parts of
> the paper.
> 

Sure.


>> Since to my knowledge the algorithm is novel, it will require some time 
>> for me to translate in paper form.
> 
> Sure, but how do you expect anyone to review your code without the
> necessary explanation? Do you think that people should approve your
> patch without that paper?
> 

I've no idea on how OpenJDK code is reviewed, so I cannot honestly
answer your questions. What I can tell is that if I were a reviewer, I
wouldn't trust rather involved code like mine without an explanation,
exactly as you seem inclined to do.

On the other side, in April this year I submitted another quite fast and
supposedly correct algorithm on this mailing list and I referred to an
accompanying paper by myself that gives full explanations on that
variant. Except for a couple of persons in private, nobody cared to send
me any observation or comment, neither on the code nor on the paper.

The present algorithm is superior. I have the theory in notes, in my
head, on napkins, on paper sheets all over my desk and floors. But
rather than spending time on the paper itself, like I did almost in vain
for the April variant, I preferred investing it in coding, for several
reasons:
* Only code executes, not a paper.
* Only code gives results that can be compared against.
* Only code can give indications on performance enhancements.
* Only code is interesting to be submitted to the OpenJDK.
* Having a paper without having tried the ideas in code is half the fun
and half as useful.

It's only a matter of priorities and of limited time, my spare time.
That said, now that I have some feedback and that the code has been
exercised at least some 400 billions times on doubles without errors and
on all 2^32 floats without errors, I now know that there is some
interest and that it is worth pushing energies in the paper.



>> There are some succinct comments here that explain the expected
>> results.  I'm not the kind of programmer that comments every line
>> since here the mechanics is simple enough to follow in Java
>> directly. A good explanation would either be mathematical, which
>> requires better typography than US-ASCII, or some explanatory
>> drawings.
> 
> Sure, these can be offline somewhere. The difficulty is always the
> mapping from the mathematics onto the Java code, and this is what must
> be explained in the comments.
> 

... or in the paper.


Greetings
Raffaello