C++: how to get fprintf results as a std::string w/o sprintf

Here’s the idiom I like for making functionality identical to ‘sprintf’, but returning a std::string, and immune to buffer overflow problems. This code is part of an open source project that I’m writing (BSD license), so everybody feel free to use this as you wish.

#include <string>
#include <cstdarg>
#include <vector>
#include <string>

std::string
format (const char *fmt, ...)
{
    va_list ap;
    va_start (ap, fmt);
    std::string buf = vformat (fmt, ap);
    va_end (ap);
    return buf;
}



std::string
vformat (const char *fmt, va_list ap)
{
    // Allocate a buffer on the stack that's big enough for us almost
    // all the time.
    size_t size = 1024;
    char buf[size];

    // Try to vsnprintf into our buffer.
    va_list apcopy;
    va_copy (apcopy, ap);
    int needed = vsnprintf (&buf[0], size, fmt, ap);
    // NB. On Windows, vsnprintf returns -1 if the string didn't fit the
    // buffer.  On Linux & OSX, it returns the length it would have needed.

    if (needed <= size && needed >= 0) {
        // It fit fine the first time, we're done.
        return std::string (&buf[0]);
    } else {
        // vsnprintf reported that it wanted to write more characters
        // than we allotted.  So do a malloc of the right size and try again.
        // This doesn't happen very often if we chose our initial size
        // well.
        std::vector <char> buf;
        size = needed;
        buf.resize (size);
        needed = vsnprintf (&buf[0], size, fmt, apcopy);
        return std::string (&buf[0]);
    }
}

EDIT: when I wrote this code, I had no idea that this required C99 conformance and that Windows (as well as older glibc) had different vsnprintf behavior, in which it returns -1 for failure, rather than a definitive measure of how much space is needed. Here is my revised code, could everybody look it over and if you think it’s ok, I will edit again to make that the only cost listed:

std::string
Strutil::vformat (const char *fmt, va_list ap)
{
    // Allocate a buffer on the stack that's big enough for us almost
    // all the time.  Be prepared to allocate dynamically if it doesn't fit.
    size_t size = 1024;
    char stackbuf[1024];
    std::vector<char> dynamicbuf;
    char *buf = &stackbuf[0];
    va_list ap_copy;

    while (1) {
        // Try to vsnprintf into our buffer.
        va_copy(ap_copy, ap);
        int needed = vsnprintf (buf, size, fmt, ap);
        va_end(ap_copy);

        // NB. C99 (which modern Linux and OS X follow) says vsnprintf
        // failure returns the length it would have needed.  But older
        // glibc and current Windows return -1 for failure, i.e., not
        // telling us how much was needed.

        if (needed <= (int)size && needed >= 0) {
            // It fit fine so we're done.
            return std::string (buf, (size_t) needed);
        }

        // vsnprintf reported that it wanted to write more characters
        // than we allotted.  So try again using a dynamic buffer.  This
        // doesn't happen very often if we chose our initial size well.
        size = (needed > 0) ? (needed+1) : (size*2);
        dynamicbuf.resize (size);
        buf = &dynamicbuf[0];
    }
}