chansey97 / clprosette-miniKanren

CLP(Rosette) on top of miniKanren

CLP(Rosette)-miniKanren

A faster-miniKanren variant integrated with Rosette, inspired by CLP(SMT)-miniKanren.

API

• (rosette-typeo <var> <type) Constraint a miniKanren variable as a rosette type.
• (rosette-asserto <formula>) Check a SMT formula satisfiable.

Running

It is not available on the Racket package server currently. Just clone the repository and play with tests.

Usage

#lang racket
(require "../mk.rkt")
(require "../rosette-bridge.rkt")

(define (faco n out)
  (conde
    ((== n 0) (== out 1))
    ((rosette-typeo n r/@integer?)
     (rosette-asserto `(,r/@! (,r/@= ,n 0)))
     (fresh (n-1 r)
       (rosette-typeo n-1 r/@integer?)
       (rosette-typeo r r/@integer?)
       (rosette-typeo out r/@integer?)
       (rosette-asserto `(,r/@= (,r/@- ,n 1) ,n-1))
       (rosette-asserto `(,r/@= (,r/@* ,n ,r) ,out))
       (faco n-1 r)))))

(run 7 (q)
  (fresh (n out)
    (faco n out)
    (== q `(,n ,out))))
;; '((0 1) (1 1) (2 2) (3 6) (4 24) (5 120) (6 720))

Why CLP(Rosette)-miniKanren?

CLP(Rosette)-miniKanren is inspired by CLP(SMT)-miniKanren, but provides better interactivity, extensibility, and refutation power. For example,

1. It supports true SMT type constraints (as Rosette solvable types) instead of raw (z/ `(declare-const ...)).

   • These SMT type constraints can interact with each other. In CLP(SMT)-miniKanren, you must use tags to distinguish, which is difficult to write (consider tower of types and coercion).
   • These SMT type constraints can be propagated among variables, which prevents many "unknown constant" error in CLP(SMT)-miniKanren.

2. It can combine Rosette constraints with miniKanren symbolic constraints (e.g. integero, symbolo, =/=, absento ) in a meaningful way. For example, the following code diverges in CLP(SMT)-miniKanren, but returns '() in CLP(Rosette)-miniKanren.

   (define (nevero)
     (conde
       [(== 1 2)]
       [(nevero)]))
       
   (run 1 (q)
     (fresh (a b)
       (rosette-typeo a r/@integer?)
       (rosette-asserto `(,r/@= ,a 5))
       (rosette-typeo b r/@integer?)
       (rosette-asserto `(,r/@= ,b 5))
       (=/= a b) ; <-- promote the `=/=` to SMT assert 
       (nevero)))
   ;; '()
   

3. Better reification representation.

   Due to the lack of intermediate layers, CLP(SMT)-miniKanren cannot digest complex types (e.g. bitvector, uninterpreted function) well. You can do format conversion back forth, but it is annoying. CLP(Rosette)-miniKanren uses Rosette, so it becomes trivial.

4. It can use all the Rosette components. For example,

   • Define a Rosette function and invoke it in rosette-asserto.
   • Easy to switch different SMT solvers.

TODO

• Add labling for r/purge.
• Add tabling (two steps).
  1. Naive tabling
  2. TCLP
• Add Rosette unsolvable types support (currently only partially supported).
• Add more SMT types via extending Rosette solvable types.
• Try different search strategies with SMT (e.g. DFS with push/pop).
• Try different SMT assertion strategies (e.g. record SMT internal state? or check-smt-assumption?).
• Find killer apps and add more tests.
• Update README.

Background

This project was inspired by Nada Amin's CLP(SMT)-miniKanren and included the improvement of smt-assumptions-full-integration which was based on smt-assumptions. Most of the tests in the tests directory initially were written in CLP(SMT)-miniKanren by Nada Amin and William Byrd.

Thank Nada Amin for bring me into CLP(SMT)-miniKanren. Thank Michael Ballantyne for his insightful suggestions (especially about how disequality constraint interacts with SMT constraints). Thank William Byrd joined the meeting and gave a lot of useful comments.