[Bug tree-optimization/68963] [4.9/5/6 Regression] O3 vs. O2 discards part of loop and terminates early

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68963

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, I think the issue is really in record_nonwrapping_iv doing

      wide_int min, max;
      extreme = fold_convert (unsigned_type, high);
      if (TREE_CODE (orig_base) == SSA_NAME
          && TREE_CODE (low) == INTEGER_CST
          && INTEGRAL_TYPE_P (TREE_TYPE (orig_base))
          && get_range_info (orig_base, &min, &max) == VR_RANGE
          && wi::gts_p (min, low))
        base = wide_int_to_tree (unsigned_type, min);
      else if (TREE_CODE (base) != INTEGER_CST)
        base = fold_convert (unsigned_type, low);
      delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base);

and for the non-constant (no-range-info) base case simply assuming it's
'low' (the array types low bound).  We can't simply treat low - base as
zero and ignore those upfront iterations.  If base were -1 then we'd correctly
add 1 to the upper bound.

Only if the stmt is always executed we can drop base to low (and we could
do that always, not just if base is an INTEGER_CST, so we can have
base = MIN (base, low) in that case - not at this time though).

So, different patch in testing.