[Continuing from https://lists.gnu.org/archive/html/emacs-devel/2019-09/msg00048.html]

Here is a new rx implementation (faster, easier to work with, fewer bugs, better tests), and, as a separate patch, an rx extension mechanism adding the macros `rx-define', `rx-let' and `rx-let-eval'.

The first patch is a ground-up rewrite of rx. It should be completely compatible.

The second patch adds

(rx-define NAME [ARGS] RX)
(rx-let ((NAME [ARGS] RX) ...) BODY)
(rx-let-eval ((NAME [ARGS] RX) ...) BODY)

as mentioned in the emacs-devel thread earlier. Additions to the manual are included.

Although I believe this to be a consistent and useful design that could be used as-is, some points worth thinking about are:

* Allow for multiple RXs in the definitions, making an implicit (seq ...). This could be done with the Schemeish syntax

(rx-define NAME RX...)
(rx-define (NAME ARGS...) RX...)

which is quite readable as "definition mirrors use". Should then the &rest parameter be declared using a dotted list, as

(rx-define (NAME ARG1 ARG2 . ARG-REST) RX...)

?

* There is some disagreement regarding whether function-like definitions should be standard Lisp expressions instead of the restricted substitution-based macros in this patch, as in

(rx-define whole (x) `(seq bos ,x eos))

I believe the usability of the chosen design is better, but see the point of not reinventing the wheel.

* Not entirely satisfied with the name `rx-let-eval', but unless someone comes up with something better, it stands.