Boost.Spirit bug when mixing "alternates" with "optionals"?

Question

I've only been working with Boost.Spirit (from Boost 1.44) for three days, trying to parse raw e-mail messages by way of the exact grammar in RFC2822. I thought I was starting to understand it and get somewhere, but then I ran into a problem:

#include <iostream>
#include <boost/spirit/include/qi.hpp>

namespace qi = boost::spirit::qi;
using qi::omit;
using qi::repeat;
using std::cout;
using std::endl;

typedef qi::rule<std::string::const_iterator, std::string()> strrule_t;

void test(const std::string input, strrule_t rule) {
    std::string target;
    std::string::const_iterator i = input.begin(), ie = input.end();

    if (qi::parse(i, ie, rule, target)) {
        cout << "Success: '" << target << "'" << endl;
    } else {
        cout << "Failed to match." << endl;
    }
}

int main() {
    strrule_t obsolete_year = omit[-qi::char_(" \t")] >> repeat(2)[qi::digit] >>
        omit[-qi::char_(" \t")];
    strrule_t correct_year = repeat(4)[qi::digit];

    test("1776", correct_year | repeat(2)[qi::digit]); // 1: Works, reports 1776.
    test("76",   obsolete_year);                       // 2: Works, reports 76.
    test("76",   obsolete_year | correct_year);        // 3: Works, reports 76.
    test(" 76",  correct_year | obsolete_year);        // 4: Works, reports 76.
    test("76",   correct_year | obsolete_year);        // 5: Fails.
    test("76",   correct_year | repeat(2)[qi::digit]); // 6: Also fails.
}

If test #3 works, then why does test #5 -- the exact same test with the two alternatives reversed -- fail?

By the same token, if you'll pardon the expression: if test #4 works, and the space at the beginning is marked as optional, then why does test #5 (the exact same test with the exact same input, save that there's no leading space in the input) fail?

And finally, if this is a bug in Boost.Spirit (as I suspect it must be), how can I work around it?

Answer 1

That is because you hit a bug in Spirit's directive repeat[]. Thanks for the report, I fixed this problem in SVN (rev. [66167]) and it will be available in Boost V1.45. At the same time I would like to add your small test as a regression test to Spirit's test suite. I hope you don't mind me doing so.