This work analyzes the math libraries of the two main vendor HLS toolchains (Vitis HLS and oneAPI) and identifies functions with significant accuracy issues and non-pipelined implementations. A systematic approach for the generation of function approximations is presented to address these issues. The approach works with both toolchains, has automatic function property detection and a resource-aware strategy selection that finds promising combinations of polynomial degree and number of segments for the approximation. The integrated resource model is able to predict the DSP usage exactly, and gives reasonable indications for other resource types. We show that regarding accuracy, this approach is able to fix a significant accuracy issue in the erf function while even saving resources. For functions which were identified to lack a pipelined implementation, this approach is able to provide a pipelined implementation that allows gaining two orders of magnitude in throughput, at moderate resource increases and partially reduced accuracy.