JTC1/SC22/WG14 N817

    WG14/N817  J11/98-016
    
    nan() function changes
    
    
    In the current C9X working draft, please change:
    
      7.7.11.2 The nan function
    
      Synopsis
    
      #include <math.h>
      double nan(const char *tagp);
    
      Description
    
      If the implementation supports quiet NaNs for the double
      type, then the call nan("n-char-sequence") is equivalent
      to strtod("NAN(n-char-sequence)", (char**) NULL); the call
      nan("") is equivalent to strtod("NAN()", (char**) NULL).
      If tagp does not point to an n-char-sequence string then
      the result NaN's content is unspecified.  If the
      implementation does not support quiet NaNs for the double
      type, a call to the nan function is unspecified.
    
      Returns
    
      The nan function returns a quiet NaN, if available, with
      content indicated through tagp.
    
    to:
    
      7.7.11.2 The nan function
    
      Synopsis
    
      #include <math.h>
      double nan(const char *tagp);
    
      Description
    
      The call nan("n-char-sequence") is equivalent to
      strtod("NAN(n-char-sequence)", (char**) NULL); the call
      nan("") is equivalent to strtod("NAN()", (char**) NULL).
    
      Returns
    
      The nan function returns a quiet NaN, if available, with
      content indicated through tagp.  If the implementation
      does not support quiet NaNs, or if tagp does not point to
      a null character or an n-char sequence, the function
      returns zero. 
    
    
    Remove from Annex K.1 Unspecified behaviour, the equivalent of:
    
      If the implementation does not support quiet NaNs for the
      double type, a call to the nan function is unspecified. 
    
    
    Add to the Rationale document, section 7.7.11.2:
    
      The nan() and scanf() family functions treat NaNs in the
      same manner.  That manner is defined in terms of the
      strtod() function.  If an implementation does not support
      NaNs, then these functions return zero (which provides a
      testable value for failure) in place of the expected NaN.