A shared-memory stencil with no barrier between write and read

Each thread writes its own value into shared memory, then immediately reads its right neighbor's slot, with no synchronization between the write and the read. What does this kernel produce?

kernel.cu
__global__ void stencilNoSync(float* out, const float* in) {
    __shared__ float smem[256];
    int tid = threadIdx.x;
    smem[tid] = in[blockIdx.x * blockDim.x + tid];
    // no __syncthreads() here
    float neighbor = smem[(tid + 1) % blockDim.x];
    out[blockIdx.x * blockDim.x + tid] = neighbor;
}

14 of 46 in GPU / CUDA