Re: [Qemu-devel] [PATCH 3/3] Add test suite for json marshalling

[Markus Armbruster]

Jamie Lokier <address@hidden> writes:

> Anthony Liguori wrote:
>> After checking that we can demarshal, marshal again and compared to
>> the expected decoded value.  This doesn't work so well for floats
>> because they cannot be accurately represented in decimal but we try
>> our best.
>
> Good sprintf/scanf/strtod implementations do guarantee that what's
> printed and then parsed gets back the same floating point value, as
> long as you have printed sufficient decimal digits.
>
> I'm not sure if FLT_DIG/DLB_DIG are the right number of digits,
> though.  Glibc's documentation of those is confusing and they might
> mean something a little different.
>
> -- Jamie

Yes, conversion fp <-> decimal with sufficient precision can be made
exact.  However, it's non-trivial[1], and thus neither IEEE-754 nor ISO
C require it.  IIRC, GNU libc makes it exact.

No, FLT_DIG/DBL_DIG is not sufficient precision.  You need 7 decimal
digits for IEEE single, 17 for IEEE double[2].  FLT_DIG is 6 and DBL_DIG
is 15 on my system.

An alternative to all that ingenuity is C99's hexadecimal format.


[1] Steele & White: How to print floating-point numbers accurately
http://portal.acm.org/citation.cfm?id=93559
Clinger: How to Read Floating Point Numbers Accurately,
http://portal.acm.org/citation.cfm?id=93557
[2] http://docs.sun.com/source/806-3568/ncg_goldberg.html#1251