Recently we found that in some loops with scaled array accesses, array range checks are not hoisted by loop predication. A typical case is below.

 for (int i = 0; i < size; i++) {
     b[3 * i] = a[3 * i];
 }

Initial investigation found that GVN optimizes the array index "3 * i" into "i << 1 + i" so it's regarded as non-linear.