[Bug tree-optimization/66678] loop counter not accurately described by vrp

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66678

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to vries from comment #2)
> Created attachment 35878 [details]
> tentative patch

That single-use case is awfully special ... just add an unrelated use
to the function and we no longer optimize again.

The real issue with VRP is that given an assert-expr we have no chance to
record multiple ranges via equivalences because equivalences are still tied
to SSA defs.  The same issue pops up for other cases.

So instead of hacking in single-use special-casing we should rather fix
that deficiency...

A possible different way to represent equivalences is to not tie value-ranges
directly to SSA names (by indexing the array with their SSA_NAME_VERSION) but
instead introduce an indirection and thus allow arbitrary new value-ranges
to be stored.

So instead of

static value_range_t **vr_value;

have

vec<value_range_t> values;
static unsigned *vr_value;

and thus get_value_range doing

   unsigned ver = SSA_NAME_VERSION (var);
   return &values[vr_value[ver]];

and make equivalences a bitmap of indexes into 'values' instead of
SSA_NAME_VERSIONs.

Then you can always record both ranges from the assert (thus keep both
the symbolic and a constant range when you now have to decide for one)