Performance regressions negatively impact the execution time and memory usage of software systems, leading to user dissatisfaction or even failure of the software in the market. Performance testing, such as microbenchmarks, allows developers to detect regressions early and address them before they affect overall system efficiency. In practice, however, existing microbenchmarks might not cover recent code changes or be sufficient to detect performance regressions. To address this problem, we introduce an on-demand performance testing approach that expands existing benchmarks only when code changes occur rather than exhaustively benchmarking the entire system. The approach is integrated into a continuous integration workflow. First, it analyzes code modifications to determine whether existing microbenchmarks already exercise affected methods. When this is not the case, it incrementally augments the performance microbenchmarks by either converting JUnit tests into JMH microbenchmarks or generating new microbenchmarks using large language models. We conduct a preliminary evaluation of our approach on Byte Buddy, a widely used open-source Java library, to assess the effectiveness of the existing and generated microbenchmarks in detecting performance regressions. Our results show that manually written microbenchmarks cover 10 of 26 modified methods but detect only 5 performance regressions. In turn, the augmentation approach detects 17 performance regressions, increasing the detection rate from 19.2% to 65.4% (+46.2%).