value range propagation for _bitwise_ OR

[Rainer Deyke]

On 4/11/2010 20:51, Steven Schveighoffer wrote:
> Range propagation is needed to determine if you can put a value into a
> smaller type.  At that point, all that is needed is the min and max. 

Technically, you don't even need min and max at that point.  A bit mask
would be enough.

> However, you are correct that without more data, the compiler might
> reject legitimate implicit casts that are easily proven.
> 
> Would (min, max, mask) be enough to allow this?  Or do we need an exact
> list...

I was think about two masks (min_mask and max_mask) such that
'(min_mask & x) == min_mask' and '(max_mask | x) == max_mask'.  In other
words, 'min_mask' contains the bits that are definitely 1 in 'x' while
'max_mask' contains the bits that /could/ be 1 in 'x'.  This avoids loss
of information in the presence of the bitwise inverse operator (unary '~').


-- 
Rainer Deyke - rainerd at eldwood.com