__syncthreads() called inside one side of a branch

A block of 256 threads (8 warps) runs this kernel. What happens?

kernel.cu
__global__ void badBarrier(float* data) {
    int tid = threadIdx.x;
    if (tid < 128) {
        data[tid] = data[tid] * 2.0f;
        __syncthreads();
        data[tid] += 1.0f;
    } else {
        data[tid] = data[tid] * 3.0f;
    }
}

13 of 46 in GPU / CUDA