>>> +                  (read-string
>>> +                   prompt
>>> +                   nil nil
>>> +                   (if read-regexp-defaults-function
>>> +                       (funcall read-regexp-defaults-function)
>>> +                     (query-replace-read-from-suggestions))
>> Isn't it too strange that read-string uses the read-regexp function?
>> Maybe better to add a new variable read-string-defaults-function.
>
> That sounds fine to me.
>
>> Then it could provide an option to use 'car' of
>> query-replace-read-from-suggestions, and allow a custom function.
>
> ...less confident about this one, because read-regexp-defaults-function
> doesn't do anything like that, right? It might be nice if the options
> mirrored each other.
>
> But if it's too much of a bother right now, I could just add two
> hand-written prompts with default to xref-find-references-and-replace,
> without the fancy "arrow". It's not like the replacements history is going
> to see too much use for this command.
>
> In the meantime, I think you should install your latest patch now (without
> the xref.el part, since it doesn't work yet).

After thinking more about this, I can't find a possible use for
read-string-defaults-function, because every call of read-string
provides own default value.  Also using read-regexp-defaults-function
in query-replace-read-from is not the right thing either - when
the users already customized it for e.g. occur, it would be too
unexpected when it will use a tag at point instead of from->to
pairs in query-replace.

Since query-replace is a very special command, the most uncontroversial
thing to do for a conservative approach would be to add two specific
variables (that later could be turned into options when needed):