This is a blind fix (for the superficial issue, but might be symptomatic of some other problem) for, e.g., the following kernel:

import loopy as lp

knl = lp.make_kernel(
    "{[i, j, k]: 0 <= i, j, k < n }",
    """
    z = simul_reduce(sum, [i, j, k], i + j + k)
    z2 = simul_reduce(sum, [i, j, k], i * j * k)
    """,
    lang_version=(2018, 2)
)

knl = lp.split_reduction_inward(knl, "k")
from loopy.transform.data import reduction_arg_to_subst_rule
knl = reduction_arg_to_subst_rule(knl, ["i", "j"])
knl = lp.realize_reduction(knl)

which throws AttributeError: 'TypeReader' object has no attribute 'symbols_with_unknown_types' from this line. TypeReader is simply missing a symbols_with_unknown_types attribute that its base class has (as well as other attributes TypeInferenceReader sets at __init__, incidentally). (The example is a minimal one derived roughly from test_global_parallel_reduction, for context. The resulting kernel looks fine after this fix.)