switch index optimization

[Joe Buck]

> The talk on linux-kernel about the bogus non-void function returns
> without a value warning you get from
> 
>   int foo(int x, int y) {
>     switch (x & 3) { 
>       case 0: ...; return y;
>       case 1: ...; return y;
>       case 2: ...; return y;
>       case 3: ...; return y;
>     }
>   }
> 
> got me thinking.  The following patch won't actually fix this case,
> since this is below the threshold for a tablejump, but it points the
> way, and does do some good.  

[ patch deleted ]

Not a bad idea, but kind of an unclean implementation.

If you're going to start putting in recognizers for expressions that
have restricted ranges, I think it would be better to put them
in a separate function, and write it in a clean way so you can cover
more cases.  The function would return the range (or an over-approximation
of the range) of the expression.  You could then cleanly handle bitwise
and, modulo, branching on a char, etc.  

But to do really well, range propagation needs to be done with dataflow
analysis (though in this case you can do it without).