More precise value analysis for "known integer or Vundef" results

[xavierleroy]

This PR introduces a new abstract value IU n, standing for "either Vint n or Vundef".

It enables more precise value analysis of comparison and selection operators, which in turn enables more compile-time evaluation of these operators.

Example:

struct s { char c[4]; char d[2]; };

int f(void)
{ struct s x; return (x.c + 6 == x.d); }

int g(void)
{ struct s x; if (x.c + 6 == x.d) return 1; else return 0; }

The current CompCert optimizes g into return 0, but fails to optimize f, as the result of the pointer comparison can be Vundef, so its abstract value cannot be I Int.zero.

With this PR, the expression x.c + 6 == x.d is analyzed as IU Int.zero, and constant-propagated to the constant 0. Hence, f and g produce the same return 0; code.

[xavierleroy]

Note: we could further reduce Uns p 0 to IU Int.zero, since 0 is the only 0-bit unsigned integer. But this needs a bit of work on the lub operator. I'll look into this later.