Eric Michael Timmons <etimmons@mit.edu> writes:

> Fix by telling `perform-replace' to operate over the entire
> buffer. Could potentially be further be optimized by saving the point
> in the scan-function and using it as the start point in the
> operate-function.

Since the current code is trying to save the point in the scan function,
it's better to keep that optimization.  See patch below.  Should this go
to emacs-27 or master?  The assumption that point would be preserved
seems to be long-standing, but I guess the change in the default of
switch-to-buffer-preserve-window-point is what surfaces the bug and
makes it more recent...