[RFC] Doubts about digit grouping corner cases

[Martin Sebor]

Paolo Carlini wrote:

> Nathan Myers wrote:
> 
>> Besides the description of the phases of numeric parsing, there
>> is also the BNF descriptions of number formats found in the
>> num_punct section. The idea is, the syntax tells you what
>> characters to extract, and then the rest tells how to construct a
>> number from the characters extracted, or to fail.
> 
> 
> Perhaps now I get your points. I take it like this: there is a set of,
> so to speak, "general requirements" on the groupings, stated in the
> num_punct section, to which *any* number must conform. Those must
> be checked during Stage2. Then, there are the specific requirements
> encoded in do_grouping, which must checked toward the end of Stage3.
> 
> Well, this is definitely reasonable, but the description of Stage2
> (and Stage3) in 22.2.2.1.2 doesn't mention at all the first part!!
> In particular, in 22.2.2.1.2, p8-9 there is absolutely no mention
> of the fact that discard may become true or false depending also on
> the last parsed char! Not just so, something much stronger: Stage2
> may terminate *only* because in == end, not for different reasons,
> like two consecutive true discard!!!
> 
> Perhaps the Standard should be clarified?!?!

Issue 358 clarifies Stage 2 to terminate the algorithm when
a thousands_sep has been seen after a decimal_point has been
extracted:
http://anubis.dkuug.dk/jtc1/sc22/wg21/docs/lwg-defects.html#358

There is a similar problem with 'e' and 'E' (e.g., the string
"123,456e+789,012" should be interpreted as 123456.0e+789,
and the subsequent ',' should terminate input; the returned
iterator should point at the ',' just prior to '0'). An issue
is yet to be filed for it.

Martin