Wrong results with PredicateStores and RoundUp combinations

[TH3CHARLie]

A repro:

#include "Halide.h"
#include <iostream>
using namespace Halide;

int main() {
    Func input("input");
    Func local_sum("local_sum");
    Func blurry("blurry");
    Var x("x"), y("y");
    RVar yryf;
    input(x, y) = 2 * x + 5 * y;
    RDom r(-2, 5, -2, 5, "rdom_r");
    local_sum(x, y) = 0;
    local_sum(x, y) += input(x + r.x, y + r.y);
    blurry(x, y) = cast<int32_t>(local_sum(x, y) / 25);
    Buffer<int> buf1, buf2;
    {
        Pipeline p({blurry});
        buf1 = p.realize({128, 128});
    }
    {
        Var xo, xi, xoo, xoi, yo, yi;
        local_sum.vectorize(x).split(x, xo, xi, 2, TailStrategy::PredicateStores).split(xo, xoo, xoi, 4, TailStrategy::RoundUp).unroll(xoi);
        local_sum.update(0).unscheduled();
        Pipeline p({blurry});
        buf2 = p.realize({128, 128});
    }
    for (int i = 0; i < 128; ++i) {
        for (int j = 0; j < 128; ++j) {
            if (buf1(i, j) != buf2(i, j)) {
                std::cout << "Error incorrect result at i: " << i << " j: " << j << " expected: " << buf1(i, j) << " actual: " << buf2(i, j) << std::endl;
            }
        }
    }
    return 0;
}

[abadams]

Somehow the simplifier is being presented with unpredicated out-of-bounds stores after vectorization and dead-stripping the whole pipeline. Possibly a bug in vectorization.

[abadams]

Nope, subtle simplifer bug.