Incorrect behavior of reversed_range with -O2 g++ option

[mkatliar]

Example code:

#include "boost/iterator/counting_iterator.hpp"
#include "boost/range/iterator_range.hpp"
#include "boost/range/adaptor/reversed.hpp"

#include <iostream>

int main(int, char **)
{
    using iterator = boost::counting_iterator<int>;
    using iterator_range = boost::iterator_range<iterator>;
    using reversed_range = boost::reversed_range<iterator_range>;

    iterator first(0), last(3);
    iterator_range range(first, last);
    reversed_range r_range = boost::adaptors::reverse(range);

    reversed_range::const_iterator it = r_range.begin();
    for (; it != r_range.end(); ++it)
        std::cout << *it << " ";
    std::cout << std::endl;
}

Compile with: -O2

Output: 3 integer numbers, different from run to run, similar to

22007 22007 22007

(with g++) or

-1807142200 0 0

(with clang++)

When compiled with -O0 produces the correct output:

2 1 0

Reproducible with: Boost 1.67.0, 1.68.0 g++ 6.4.0, 7.3.0, 8.2.0; clang++ 7.0.0-3 CPU architecture: x86-64 (Intel Xeon)

[mkatliar]

Works correctly with +O1 in 8.0.0-3, but not in g++ 8.3.0

[morinmorin]

Please don't use counting_iterator/counting_range. Use irange(0, 3), instead.

[mkatliar]

@morinmorin first, I can't avoid using counting_iterator. I am using boost::make_iterator_range() with a third-party library, which returns an iterator which can be a counting_iterator or some other type, depending on template parameters. Second, as long as the iterator type models the required concept, the code should work.

[neilgroves]

On Fri, 20 Sep 2019 at 13:28, Mikhail Katliar notifications@github.com wrote:

@morinmorin https://github.com/morinmorin first, I can't avoid using counting_iterator. I am using boost::make_iterator_range() with a third-party library, which returns an iterator which can be a counting_iterator or some other type, depending on template parameters. Second, as long as the iterator type models the required concept, the code should work.

This isn't morinmorin's defect. It is mine, or at least more mine since I wrote the reversed adaptor. The counting_iterator returns a value. This isn't a valid model of the (Legacy) random-access iterator concept. Thus it isn't totally unreasonable that Boost.Range cannot operate with it in all cases.

I may look at this again when I get a little free time. In the interim you can have a full refund. Please don't complain to morinmorin who is helping you and I, often clearing up my mistakes and my frequent inability to spare as much time as I would like. Patches are always welcome.

Regards, Neil Groves

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/boostorg/range/issues/82?email_source=notifications&email_token=ABTIJ2RHSEHDO6CJEJ4ACGLQKS6VTA5CNFSM4GF7OQW2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD7GQ6YA#issuecomment-533532512, or mute the thread https://github.com/notifications/unsubscribe-auth/ABTIJ2R3IE2743C7J53Z5VTQKS6VTANCNFSM4GF7OQWQ .

[morinmorin]

@mkatliar If you want to use them with reversed adaptor (among others), the option would be convert them into integer_range, which is defined in irange.hpp.

counting_iterator is a beast; deferencing it returns a reference to the data it owns. It cannot work even with std::reverse_iterator. Unfortunately, changing its reference type to a value type would not be an option now, since that would demote the iterator category to InputIterator (and then break some users' code).

@neilgroves Hey, it's just that counting_iterator is a beast 😄

Regards, Michel Morin