information to be lost on converting int to float

[David Hough]

In mixed expressions involving integer and floating types, it would be
better to slightly amend ANSI-C for NCEG by requiring that, regardless
of expression evalution mode, operands of different types always be
promoted AT LEAST to what I call the "smallest encompassing" floating-point
type (but not to exceed long double!), which for IEEE systems is...

float & 16-bit int		float
float & 32-bit int		double
double & 32 bit int		double
double & 64 bit int		long double
long double & int		long double

Thus "promotion" never entails rounding except possibly after you've 
promoted everything to long double.

I don't think this would cause any trouble for IEEE, VAX, or 370
architectures, but it might be a problem for Cray if there are
peculiarities of integer arithmetic like those on the CDC-6600.
I'd be happy to limit this requirement to IEEE 754 implementations.