[Bug tree-optimization/98028] [8/9/10/11 Regression] __builtin_sub_overflow_p not folded to const when some constraints are known

["amacleod at redhat dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98028

--- Comment #4 from Andrew Macleod <amacleod at redhat dot com> ---
I would expect in gcc 12 we'll be handling it differently than that.

I have equivalences/relations up and running and will eventually get to
applying those to general statements... which I haven't gotten to yet.

the gimple_fold() routine uses fold_range () via range-ops to calculate a
result. I plan to additionally check if there is a relation between the
operands, and if so, invoke another range_ops routine to see if that relation
makes any further adjustments.

in this case, we'd have VARYING - VARYING, so ranges would resolve as expected
to VARYING.

A query if there is a relation between j_1 and i_2
=========== BB 4 ============
j_1(D)  unsigned int VARYING
i_2(D)  unsigned int VARYING
Equivalence set : [j_1(D), i_2(D)]
    <bb 4> :
    _3 = i_2(D) - j_1(D);
    return _3;
shows they are in an equivalence set, so that would be trivial to fold to 0 for
MINUS_EXPR.

as well, if we see other relations
 if (j > i) __builtin_unreachable();
  return i - j;

=========== BB 4 ============
j_1(D)  unsigned int VARYING
i_2(D)  unsigned int VARYING
Relational : (j_1(D) <= i_2(D))
    <bb 4> :
    _3 = i_2(D) - j_1(D);
    return _3;

it would likewise be pretty straightforward to add the restriction that _3 will
be [0, +INF]  (or [1, +INF] if it were j_1 < i_2)
If the relation is !=, we could also provide ~[0,0]

I'd expect the routine for MINUS_EXPR to simply switch on the possible
relations and intersect the current result with this reduced possible outcome.

I hope to get most of this in early in stage 1 and we can begin
enhancing/finding opportunities.
.