In real-world automotive control, there are many constraints to be considered. In order to explicitly treat the constraints, we introduce a model-prediction-based algorithm called a reference governor (RG). The RG generates modified references so that predicted future variables in a closed-loop system satisfy their constraints. One merit of introducing the RG is that effort required in control development and calibration would be reduced. In the preceding research work by Nakada et al., only a single reference case was considered. However, it is difficult to extend the previous work to more complicated systems with multiple references such as the air path control of a diesel engine due to interference between the boosting and exhaust gas recirculation (EGR) systems. Moreover, in the air path control, multiple constraints need to be considered to ensure hardware limits. Hence, it is quite beneficial to cultivate RG methodologies to deal with multiple references and constraints. In this paper, we develop the RG algorithm based on gradient descent method to allow for multiple references and constraints. We demonstrate the effectiveness of the presented RG algorithm in a transient driving cycle experiment using a real engine, in which constraints are enforced on maximal boost pressure, turbine speed, compressor surge and maximal and minimal EGR rates. The experiment implies that we have expanded the applicability of an RG to system with multiple references compared to the previous work for only a single reference.