num_get::do_get incorrect digit grouping check

[llvmbot]

Bugzilla Link	28704
Resolution	FIXED
Resolved on	Jun 17, 2019 06:42
Version	unspecified
OS	All
Reporter	LLVM Bugzilla Contributor
CC	@Quuxplusone,@mclow

Extended Description

For a locale which specifies digit grouping, num_get::do_get will enforce
that the number read in contains the grouping character at the correct
locations. It does this even if the number read in didn't have any grouping
characters present.

This is shown best with a code example:

#include
#include

int main()
{
std::locale::global(std::locale("en_US.UTF-8"));
std::istringstream iss("1024");
int a = 0;
iss >> a;
std::cout << a << ", " << iss.fail() << std::endl;
}

With libc++ this prints:

1024, 1

With libstdc++ this prints:

1024, 0

http://coliru.stacked-crooked.com/a/9283df64c2bd5142

The standard says:

Digit grouping is checked. That is, the positions of discarded separators is
examined for consistency with
use_facet<numpunct >(loc).grouping(). If they are not consistent then
ios_base::failbit is assigned to err.

http://eel.is/c++draft/facet.num.get.virtuals#4

Whilst the standard could be clearer, it seems one can take it to mean that as
there are no discarded characters, there is nothing to examine for consistency.

This is the approach that libstdc++ takes, as well as boost's lexical_cast:

https://svn.boost.org/trac/boost/ticket/5585

I think it's reasonable that libc++ changes its implementation to match.

[llvmbot]

assigned to @mclow

[mclow]

Reduced test case
Reduced test case.
Passes with Visual Studio online compiler.
Passes (with a warning) with GCC 7.
Fails with libc++

[mclow]

If you've said "this is a facet that does grouping", then it's hard to argue that "no grouping" is consistent with that facet.

On the other hand, the original example would be nice to support.

[mclow]

Fixed in revision 362508.

[Quuxplusone]

Marshall, your libcxx commit changes a line in src/locale.cpp from

-    if (__grouping.size() != 0)

to

+       if (__grouping.size() >= 0 && __g_end - __g > 1)

which might have been meant to be != instead of >= in the new line.
At least, according to my compiler, __grouping.size() is unsigned, so >= 0 is invariably true. Should it have been != (or >), or did you mean to remove the now-tautological condition entirely?

[mclow]

thanks. Commit 363557