The documentation for mark-sexp says:
> Set mark ARG sexps from point.
> The place mark goes is the same place C-M-<right> would
> move to with the same argument.
This is correct but misleading. Worse still is the next part:
> Interactively, if this command is repeated
> or (in Transient Mark mode) if the mark is active,
> it marks the next ARG sexps after the ones already marked.
This says it marks sexps *after* the ones already marked. This is incorrect; if point is *after* mark, and mark is active, this function marks sexps *before* the ones already marked.
> In my opinion, the description of that feature should not be part of the
> mark-sexp command documentation, because the feature is actually
> provided by C-x C-x (exchange-point-and-mark).
One way for this situation to occur is by using C-x C-x, yes, but it's not the only way. You can set point, move forward sexps, and then calling mark-sexp will mark sexps backwards.
There is specific code in mark-sexp to check which direction to move; it should be documented as such.