Taint analysis method in binary code plays an important role in reverse engineering, malicious code detecting and vulnerabilities analysis. Currently, most of taint analysis methods fail to operate float point instruction, and they do not propagate taints accurately and efficiently enough. In the paper, a taint analysis method is implemented based on offline indices of instruction trace, which are byte-grained and utilize taint tags. A generation and query algorithm of offline indices is also presented. Instructions unrelated to taint data are skipped with offline indices, which improves the efficiency of taint analysis. The taint loss problem resulted from real time translation is described and solved for the first time. Taint tags are utilized to denote where the taint data is derived. A more complete taint propagation algorithm, which could operate float point instructions and insure the taint data flow from source operands into the destination operands precisely, is also presented. Flexible user-configuration mechanism is implemented to produce taint data on the fly with black list. The proposed method is applied in vulnerabilities detecting and evaluated with 12 vulnerabilities as test cases. The experimental result shows that this taint analysis method is able to detect more vulnerabilities than TEMU, and is 5 times faster in average.