Binary feedback-directed optimization (BFDO) is a novel technique for optimizing program binaries by leveraging dynamic profiles, showing promising potential in various domains including data centers, clouds, and embedded systems. However, state-of-the-art BFDO approaches primarily focus on specific instruction set architectures and programming languages, while overlooking the architectural discrepancies, language diversity, and compiler optimization options. Consequently, a comprehensive study of BFDO is still lacking, hindering its broader applications. In this paper, to address this gap, we present the first systematic study of BFDO, providing insights into its current capabilities, key grand challenges, and future research opportunities. Specifically, we first conduct an empirical study using a novel end-to-end evaluation tool to assess the effectiveness of BFDO. Our analysis identifies three root causes of BFDO failures and highlights three factors influencing its optimization effectiveness. Based on these findings, we propose four best practices and outline three promising research opportunities, offering guidance for future advancements. Applying our best practices to RISC-V benchmarks improves performance by an average of 4.57%, effectively mitigating existing relocation failures.