exprtk could be useful as part of a CAS; however, in terms of reducing the number of AST nodes for purposes of perf improvements, this is not always desirable, as these optimizations may lead to accuracy degradations. Given IEEE 754 floating-point format, operations which are theoretically commutative are not always commutative in practice. So anything which performs simplifications, reorderings, etc, may violate an algorithm's assumptions and result in decreased accuracy. It is not uncommon for transcendental function algorithms to assume (and exploit) certain properties of IEEE 754 to achieve results to a desired precision.

This should also be a more general warning when translating/porting numerical algorithms: the order of operations matters and any change to that order can result in significant deviations in accuracy.