Is it possible to read infinity or NaN values using input streams?

Write a function with a signature like this:

std::istream & ReadDouble(std::istream & is, double & d);

Inside it, you:

1. Read a string from the stream using operator>>
2. Try to convert the string to a double using one of various methods. std::stod, boost::lexical_cast, etc...
3. If the conversion succeeds, set the double and return the stream.
4. If the conversion fails, test the string for equality with "inf" or "INF" or whatever.
5. If the test passes, set the double to infinity and return the stream, else:
6. If the test fails, set the fail bit on the stream and return it.

Edit: To avoid the use of a wrapper structure around a double, I enclose an istream within a wrapper class instead.

Unfortunately, I am unable to figure out how to avoid the ambiguity created by adding another input method for double. For the implementation below, I created a wrapper structure around an istream, and the wrapper class implements the input method. The input method determines negativity, then tries to extract a double. If that fails, it starts a parse.

Edit: Thanks to sehe for getting me to check for error conditions better.

struct double_istream {
    std::istream ∈

    double_istream (std::istream &i) : in(i) {}

    double_istream & parse_on_fail (double &x, bool neg);

    double_istream & operator >> (double &x) {
        bool neg = false;
        char c;
        if (!in.good()) return *this;
        while (isspace(c = in.peek())) in.get();
        if (c == '-') { neg = true; }
        in >> x;
        if (! in.fail()) return *this;
        return parse_on_fail(x, neg);
    }
};

The parsing routine was a little trickier to implement than I first thought it would be, but I wanted to avoid trying to putback an entire string.

double_istream &
double_istream::parse_on_fail (double &x, bool neg) {
    const char *exp[] = { "", "inf", "NaN" };
    const char *e = exp[0];
    int l = 0;
    char inf[4];
    char *c = inf;
    if (neg) *c++ = '-';
    in.clear();
    if (!(in >> *c).good()) return *this;
    switch (*c) {
    case 'i': e = exp[l=1]; break;
    case 'N': e = exp[l=2]; break;
    }
    while (*c == *e) {
        if ((e-exp[l]) == 2) break;
        ++e; if (!(in >> *++c).good()) break;
    }
    if (in.good() && *c == *e) {
        switch (l) {
        case 1: x = std::numeric_limits<double>::infinity(); break;
        case 2: x = std::numeric_limits<double>::quiet_NaN(); break;
        }
        if (neg) x = -x;
        return *this;
    } else if (!in.good()) {
        if (!in.fail()) return *this;
        in.clear(); --c;
    }
    do { in.putback(*c); } while (c-- != inf);
    in.setstate(std::ios_base::failbit);
    return *this;
}

One difference in behavior this routine will have over the the default double input is that the - character is not consumed if the input was, for example "-inp". On failure, "-inp" will still be in the stream for double_istream, but for a regular istream only "inp" will be left in the the stream.

std::istringstream iss("1.0 -NaN inf -inf NaN 1.2");
double_istream in(iss);
double u, v, w, x, y, z;
in >> u >> v >> w >> x >> y >> z;
std::cout << u << " " << v << " " << w << " "
          << x << " " << y << " " << z << std::endl;

The output of the above snippet on my system is:

1 nan inf -inf nan 1.2

Edit: Adding a "iomanip" like helper class. A double_imanip object will act like a toggle when it appears more than once in the >> chain.

struct double_imanip {
    mutable std::istream *in;
    const double_imanip & operator >> (double &x) const {
        double_istream(*in) >> x;
        return *this;
    }
    std::istream & operator >> (const double_imanip &) const {
        return *in;
    }
};

const double_imanip &
operator >> (std::istream &in, const double_imanip &dm) {
    dm.in = ∈
    return dm;
}

And then the following code to try it out:

std::istringstream iss("1.0 -NaN inf -inf NaN 1.2 inf");
double u, v, w, x, y, z, fail_double;
std::string fail_string;
iss >> double_imanip()
    >> u >> v >> w >> x >> y >> z
    >> double_imanip()
    >> fail_double;
std::cout << u << " " << v << " " << w << " "
          << x << " " << y << " " << z << std::endl;
if (iss.fail()) {
    iss.clear();
    iss >> fail_string;
    std::cout << fail_string << std::endl;
} else {
    std::cout << "TEST FAILED" << std::endl;
}

The output of the above is:

1 nan inf -inf nan 1.2
inf

Edit from Drise: I made a few edits to accept variations such as Inf and nan that wasn't originally included. I also made it into a compiled demonstration, which can be viewed at http://ideone.com/qIFVo.

Just read your variables into string and parse them. You can't put string into double variables and expect them to be outputted like a string, because if it worked, strings wouldn't be necessary.

Seomthing like:

string text;
double d;
while(cin >> text)
{
    if(text == "Inf")       //you could also add it with negative infinity
    {
         d = std::numeric_limits<double>::infinity();
    }
    else
    {
        d = atof(text.c_str());
    }
}

Update Provided a simple test case that shows that Boost Spirit is capable to handle all varieties of special values in this area. See below: Boost Spirit (FTW).

The standard

The only normative information in this area that I've been able to find is in sections 7.19.6.1/7.19.6.2 of the C99 standard.

Sadly, the corresponding sections of the latest C++ standard document (n3337.pdf) doesn't appear to specify support for infinity, inf and or NaN in the same way. (Perhaps I'm missing a footnote that refers to the C99/C11 spec?)

The library implementors

In 2000, the Apache libstdcxx received a bug report stating

The num_get<> facet's do_get() members fail to take the special strings [-]inf[inity] and [-]nan into account. The facet reports an error when it encounters such strings. See 7.19.6.1 and 7.19.6.2 of C99 for a list of allowed strings.

However the subsequent discussion yielded that (at least with named locale-s) it would actually be illegal for an implementation to parse special values:

The characters in the lookup table are "0123456789abcdefABCDEF+-". Library issue 221 would amend that to "0123456789abcdefxABCDEFX+-". "N" isn't present in the lookup table, so stage 2 of num_get<>::do_get() is not permitted to read the character sequence "NaN".

Other resources

securecoding.cert.org clearly states that the following 'Compliant Code' is required to avoid parsing infinity or NaN. This implies, that some implementations actually support that - assuming the author ever tested the published code.

#include <cmath>

float currentBalance; /* User's cash balance */

void doDeposit() {
  float val;

  std::cin >> val;
  if (std::isinf(val)) {
    // handle infinity error
  }
  if (std::isnan(val)) {
    // handle NaN error
  }
  if (val >= MaxValue - currentBalance) {
    // Handle range error
  }

  currentBalance += val;
}

Boost Spirit (FTW)

The following trivial example has the desired output:

#include <boost/spirit/include/qi.hpp>
namespace qi = boost::spirit::qi;

int main()
{
    const std::string input = "3.14 -inf +inf NaN -NaN +NaN 42";

    std::vector<double> data;
    std::string::const_iterator f(input.begin()), l(input.end());

    bool ok = qi::parse(f,l,qi::double_ % ' ',data);

    for(auto d : data)
        std::cout << d << '\n';
}

Output:

3.14
-inf
inf
nan
-nan
nan
42

Summary/TL;DR

I'm inclined to say that C99 specifies the behaviour for *printf/*scanf to include infinity and NaN. C++11, sadly appears to not specify it (or even prohibit it, in the presence of named locales).