GT_ADD works bad for `ADD(ref 0, cnst int)` case

[sandreenko]

That blocks the swap:

            noway_assert(op1);
            if (op1->OperIsConst() && (op1->gtType != TYP_REF))
            {
                // TODO-Review: We used to assert here that
                // noway_assert(!op2->OperIsConst() || !opts.OptEnabled(CLFLG_CONSTANTFOLD));
                // With modifications to AddrTaken==>AddrExposed, we did more assertion propagation,
                // and would sometimes hit this assertion.  This may indicate a missed "remorph".
                // Task is to re-enable this assertion and investigate.

                /* Swap the operands */
                tree->AsOp()->gtOp1 = op2;
                tree->AsOp()->gtOp2 = op1;

                op1 = op2;
                op2 = tree->AsOp()->gtOp2;
            }

so for a tree-like ADD(ref 0, cnst_int 4) we don't transform that tree,

                        // If this addition is adding an offset to a null pointer,
                        // avoid the work and yield the null pointer immediately.
                        // Dereferencing the pointer in either case will have the
                        // same effect.

that comment is probably wrong because it expects ref 0 as the second operand, but it will be always the first.

and at the end SetEvalOrder will ignore ref and swap them. See issue 32466