As of RStan 2.21, when you compile a model RStan downloads the latest version of the Stanc3 parser (that translates Stan syntax to c++ to be compiled). You can opt out of this by running: rstan_options(javascript = FALSE) when you first load the package.
However, a 4x slowdown is pretty dramatic so we’d really like to dig into this and figure out what caused the regression. Are you able to share the model, some data, and a rough estimate of the last time the model ran at the expected speed?
That only checks whether your Stan program still parses with the next version of RStan. The C++ code that gets compiled is still generated by the old C++ parser.