Re: [Qemu-devel] [PATCH v9 01/37] qobject: Document more shortcomings in our number handling

[Eric Blake <eblake@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 2772 bytes --]

On 01/20/2016 11:21 PM, Markus Armbruster wrote:
>> One alternative is to always output a guaranteed unambiguous decimal
>> string (although not necessarily the shortest), by using %.17f, using
>> <float.h> DBL_DECIMAL_DIG.  (Note that DBL_DIG of 15 is NOT sufficient -
>> it is the lower limit that says that a decimal->float->decimal will not
>> change the decimal; but we want the converse where a
>> float->decimal->float will not change the float.  There are stretches of
>> numbers where the pigeonhole principle applies; you can think of it this
>> way: there is no way to map all possible 2^10 (1024) binary values
>> inside 2^3 (1000) decimal digits without at least 24 of them needing one
>> more decimal digit.  But by the same arguments, DBL_DECIMAL_DIG is an
>> upper limit and usually more than you need.)
>>
>> So, the question is whether we want to always output 17 digits, or
>> whether we want to do the poor-man's truncation scheme (easy to
>> understand, but not optimal use of the processor), or go all the way to
>> the algorithm of that paper (faster but lots harder to understand).  For
>> reference, here's the poor-man's algorithm in pseudocode:
> 
> I don't think we want to implement floating-point formatting ourselves.

Well, we already have _some_ level of floating-point support built in
for TCG emulation of floating point on various architectures.  I don't
know how much would be easily reusable for this case, though.

> 
>> if 0, inf, nan:
>>     special case
>> else:
>>     Obtain the DBL_DECIMAL_DIG string via sprintf %.17f
>>     i = 17;
>>     do {
>>         truncate the original string to i-1 decimal characters
>>         parse that with strtod()
>>         if the bit pattern differs:
>>             break;
>>     } while (--i);
>>     assert(i)
>>     use i digits of the string
> 
> That's a lot of strtod()...  May not be noticable if we write the result
> to a slowish sink.  Binary search could save a few.
> 
> Naive idea: chop off trailing '0'?

That, and rounding trailing '9'.  Any other digits in positions 1-16 are
significant (other digits in position 17 might not matter, though), so I
guess a full 17 iterations is probably not strictly necessary.  Our
current code DOES chop trailing '0', but starting from the flawed
premise that 6 digits was enough precision.

> 
>> As a separate patch, of course, but I have a pending patch that provides
>> a single place where we could drop in such an improvement:
>> https://lists.gnu.org/archive/html/qemu-devel/2015-12/msg03932.html
> 
> Definitely separate.

Okay, all thoughts for a later day :)

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 604 bytes --]